Standing balance assessment and prognosis after stroke

Executive assessment

Standing balance should be measured after stroke, but there is no single standing test, sway variable, or score threshold that answers all the clinically important questions. The strongest practical approach combines a phase-appropriate clinical assessment with a small number of explicitly defined standing tasks, adding instrumentation when it resolves a particular question about sensory dependence, interlimb control, or weight transfer. A test of what the person can do, a measure of how the person maintains equilibrium, and a model of what will happen next are different products. Their evidence should remain separate.

For routine clinical balance measurement, the third Stroke Recovery and Rehabilitation Roundtable recommends the Mini-BESTest, with the Berg Balance Scale (BBS) additionally used when Functional Ambulation Categories (FAC) is below 3. That recommendation supports standardisation of measurement across stroke studies; it does not validate a universal Mini-BESTest falls probability or make the BBS a biomechanical reference standard. At the severe end, the Postural Assessment Scale for Stroke Patients (PASS) is important because lying, sitting, transitions, and assisted standing remain measurable before advanced standing tasks are feasible. At the ambulant end, the BBS can reach its ceiling while reactive, sensory, and gait-related balance deficits persist. [1–5]

For quantitative quiet standing, centre-of-pressure (COP) velocity is generally a more defensible starting point than a large panel of poorly repeatable features. However, even high intraclass correlation coefficients (ICCs) coexist with substantial individual measurement error. Gasq and colleagues found excellent between-day reliability for several velocity measures but much larger relative errors for ellipse area and mediolateral mean COP position. Bower and colleagues demonstrated clinically feasible Wii Balance Board assessments, but their static sway minimum detectable changes (MDCs) were approximately 35–44% of the sample mean. These observations favour carefully repeated, protocol-matched measurement over declaring small numerical differences to be recovery. [6, 7]

The most consequential negative finding is that useful measurement is not necessarily useful prognosis. In Bower and colleagues' prospective post-discharge cohort, none of six quiet-standing COP velocity variables independently predicted prospectively recorded falls over 12 months; Step Test and Timed Up and Go (TUG) performance were more informative. Conversely, Mansfield and colleagues reported prospective relationships between falls rate and limb-specific balance control, interlimb synchronisation, and reactive stepping. These results are not contradictory: net sway, limb contributions, and the ability to respond to instability are distinct constructs, assessed in different samples and models. [8, 9]

A second important negative finding concerns recovery monitoring. In a 2025 multicentre cohort, body-worn sensor sway changed little in many people whose functional balance and independence improved. That does not prove that postural sway is clinically irrelevant; it does show that less sway cannot be assumed to be the universal signature of better rehabilitation. Compensation, changing movement strategies, increased confidence, and task-specific recovery can produce different trajectories. [10, 11]

For rehabtools, the immediate opportunity is a transparent measurement record: exact task, completed duration, assistance, footwear/orthosis, paretic side, raw units, repetitions, and a protocol-specific uncertainty statement. A general stroke falls score inferred from a few seconds of sway is not supported by this evidence. Any future prognostic output needs a defined endpoint and horizon, adequate events, prespecified adjustment, external validation, calibration, and evidence that it adds information beyond inexpensive clinical predictors. A transparent descriptive result is more useful than an apparently precise risk estimate built from an incompatible cutoff.

Scope and appraisal

The clinical question and its boundaries

This report considers adults after ischaemic or haemorrhagic stroke and asks two related questions. First, which clinical and instrumented assessments provide interpretable information about standing balance? Second, when measured at a defined point after stroke, what does that information contribute to prognosis for falls, balance recovery, walking, and independence?

The central constructs are quiet standing; standing with altered visual, surface, or support conditions; voluntary limits of stability and anticipatory control; reactive balance responses; and functional balance scales that contain standing tasks. Trunk and sitting assessments are included as explicitly labelled early-recovery context. Gait measures are included where they clarify the incremental value of standing assessment or the interpretation of whole-scale scores, rather than as a substitute for a separate gait report. This is not a treatment-efficacy review. Improvement during rehabilitation establishes neither that a particular treatment caused the change nor that the test is an adequate surrogate for a clinically important outcome.

Time since onset should be recorded continuously, alongside clinically useful phase labels. The working framework distinguishes acute stroke, including the first week; early subacute recovery after the first week through approximately three months; late subacute recovery through six months; and chronic stroke beyond six months. Original studies used different boundaries, including 'chronic' after three months. Such studies are described with their actual eligibility and observed timing, not silently relabelled as if all chronic samples were equivalent. Admission and discharge are service events, not biological recovery stages. A discharge assessment at 24 days after onset and one at several months are not interchangeable baselines. [1, 7, 8, 12]

The ability to attempt an assessment is part of its clinical meaning. Studies requiring 30 seconds of unsupported stance, five minutes of independent standing, or ten metres of walking exclude many people with severe stroke. The most vulnerable patient may have no valid sway file precisely because standing is unsafe. Analyses restricted to completed trials estimate performance among those able to complete them; they do not describe the entire rehabilitation population. An inaccessible test must not be represented as normal, zero sway, or a missing-at-random measurement.

Search approach and source accessibility

The focused Scopus search combined stroke with standing balance, postural sway, force-platform/force-plate, single-leg stance, or quiet-standing terms and reliability, validity, responsiveness, measurement-error, or prognosis terms. Nine pages yielded 219 records and 219 distinct source identifiers, matching the reported total. Citation-network searches used OpenAlex and Semantic Scholar, with bounded requests from the quantitative standing review, the original PASS paper, and the prospective Bower falls cohort. The exact database queries are reproduced in the search appendix below; page records and exact-query recovery files were retained for audit.

The approach is a critical narrative review with auditable search coverage, not a registered systematic review, an exhaustive review of every database, or a PRISMA review. The PubMed and Scopus queries are complementary rather than identical. Title/abstract restrictions can miss articles with poorly indexed constructs. Some retrieved studies used balance solely as an intervention outcome and were not selected for detailed appraisal. Selection prioritised original measurement studies, prospective prognostic cohorts, clinically influential scale studies, newer contradictory evidence, and studies directly relevant to practical implementation.

Source accessibility is reported for each reference. Complete article bodies and tables were retrieved for the major quantitative reliability studies, several scale comparisons, the LEAPS falls analysis, Bower's prospective falls cohort, newer sensor monitoring, and markerless agreement work. Other influential papers remain abstract-level, including selected BBS/PASS walking-prognosis cohorts. The complete Chinsongkram BESTest and 2026 diagnostic/prognostic review PDFs were subsequently examined; the review’s separate supplemental files were not included in the supplied PDF. The original PASS report was subsequently examined in complete licensed HTML, with its embedded tables and figures inspected visually. Numerical findings based only on abstracts are identified accordingly; unobserved protocols, covariate sets, calibration, or subgroup results are not inferred. The source-access appendix identifies completed upgrades and remaining limits.

How evidence is weighed

Reliability is not one property. Rater agreement when two people score the same performance differs from repeatability when the patient performs again; immediate repeated trials differ from reassessment a week later. A high ICC indicates preservation of relative ranking in the sampled population. It can be high because participants differ greatly, even when individual repeated measurements differ by a clinically substantial amount. The ICC model, number of averaged trials, confidence interval, and retest interval therefore matter at least as much as the adjective 'excellent'.

An MDC, smallest real difference (SRD), or minimum detectable difference describes change exceeding estimated error under specified conditions and a chosen confidence level. It does not say whether that change is important to a patient. An anchor-based minimal important change or minimal clinically important difference (MIC/MCID) asks a different question and depends on the anchor, who rates it, baseline severity, observation period, and analysis. Distribution-only values are not direct evidence of patient importance. A statistically responsive test can detect a group shift without reliably identifying meaningful change in an individual.

Construct validity also requires a specific question. Correlation with the BBS is useful convergent evidence for functional balance, but the BBS is not a gold standard for the forces under each foot. Poor correlation between weight distribution and BBS can reflect genuinely different constructs, compensation permitted by the scale, measurement error, or a restricted score range. Conversely, high correlation between two composite scales does not prove that their individual items or underlying systems are equivalent.

For prognosis, future falls collected prospectively are given greater weight than recalled falls or a contemporaneous 'fall-risk' scale. A model predicting a BBS cutoff predicts a balance-score category unless actual future falls were observed. A model predicting walking independence at discharge is not a falls model. Association, discrimination, calibration, and decision value are reported separately. In particular, a significant odds ratio can coexist with weak discrimination, and an apparently favourable area under the curve (AUC) can coexist with poor calibration or optimism from selecting predictors and cutoffs in the same small sample.

What existing reviews add

The quantitative standing review by Bruyneel and Dubé found a literature with generally favourable reliability but much less consistent validity, variable methodological reporting, and substantial reliance on relatively able participants. Its distinction between quantitative instruments and observation-based scales is useful. Its recommendations also depend on practical considerations such as cost, time, and portability. It should be read as a map of the earlier literature rather than proof that every device or metric is interchangeable. [13]

The complete 2026 BBS/Mini-BESTest diagnostic and prognostic review is a more recent map, with a search through 24 December 2025. It included 55 reports, a reported 6,865 participants and 23 reference standards; its flow diagram identifies 26 studies contributing to meta-analysis. The reported participant total should not be assumed to represent 6,865 demonstrably unique people across every related publication. English and Japanese articles were eligible. The review separates current-status classification from prediction and uses bivariate random-effects pooling when at least two studies are available. This is an appropriate conceptual distinction, but the available original studies remain heterogeneous in severity, setting, reference definition and threshold. [14]

For BBS prediction of independent walking, seven studies gave summary sensitivity 0.778 (95% CI 0.675–0.856), specificity 0.805 (0.703–0.878) and AUC 0.858. Nine BBS falls-prediction studies gave sensitivity 0.618 (0.538–0.692), specificity 0.711 (0.629–0.782) and AUC 0.704. These are pooled operating characteristics across study-specific cutoffs, not accuracy for one recommended BBS threshold. The Table 3 prognostic Mini-BESTest entries contain only one study each and no pooled estimate. Thus, stronger BBS evidence for some outcomes does not establish that the Mini-BESTest is an inferior prospective predictor in a sufficiently powered, matched comparison. [14]

Current walking classification has a different interpretation. BBS independence classification, based on six studies with multiple independent samples in one report, had AUC 0.917, sensitivity 0.880 (0.805–0.929) and specificity 0.851 (0.801–0.891). The two-study Mini-BESTest estimate had AUC 0.790, sensitivity 0.736 (0.547–0.865) and very imprecise specificity 0.802 (0.228–0.982). For current community-ambulator classification, reported AUCs were 0.844 for BBS and 0.882 for Mini-BESTest, but the latter came from only two studies with wide intervals: sensitivity 0.800 (0.405–0.959), specificity 0.834 (0.361–0.978). Much of this endpoint is walking-speed-defined capacity rather than observed community participation. These aggregate comparisons are not a randomised or uniformly paired head-to-head test and should not be presented as a universal scale ranking. [14]

The risk-of-bias analysis strengthens the caution. Forty-eight of 55 studies were judged high risk overall, including 45 for patient selection and 39 for flow/timing. The reviewers regarded cognitive exclusions as a patient-selection concern and removed the standard QUADAS-2 question about whether thresholds were prespecified. Consequently, the already adverse risk profile does not fully represent the usual concern about deriving an optimal threshold and assessing it in the same data. Only five reports supplied full contingency tables, and six supplied some cells; other cells were reconstructed from sample counts and rounded sensitivity/specificity or requested from authors. Subgroup analyses were post hoc. Most threshold studies lacked internal or external validation, and the reviewers could not stratify by time since stroke. No threshold-free subgroup inference can remove those design limitations. [14]

Cohort reconciliation was attempted: the larger Maeda overlapping cohort was retained, the larger Makizako overlapping threshold sample was selected, and three stated independent Ichinosawa samples were retained. This is useful but does not establish exhaustive person-level reconciliation across all 55 reports. The full article also contains unresolved reporting conflicts. Its Results says Miyata 2020 was excluded for a mixed-diagnosis population, yet Table 3 includes it in both community-ambulator pools. Table 2 lists Miyata 2016 and Inoue 2023 for Mini-BESTest future falls, while Table 3 reports only Inoue; the main text does not resolve that selection discrepancy. The review's Table 1 also assigns the Louie walking-outcome AUCs differently from the verified original abstract. Original-study estimates in this report therefore remain anchored to their own source, rather than being overwritten by the review table. [14, 15]

The review reports higher falls AUC for a two-or-more-falls subgroup than a one-or-more-falls subgroup (0.843 versus 0.672), and explores groups of studies using higher versus lower cutoffs. Those findings describe different study groups and outcomes; they do not show that raising the cutoff in a given patient population improves prediction. The authors explicitly conclude that the evidence is insufficient to recommend a specific threshold for each reference standard. The complete article PDF was examined, including its main tables; separate supplemental datasets, plots and subgroup tables were not supplied with it, and the meta-analysis was not independently rerun. Its principal contribution is endpoint-specific orientation and a warning about weak validation, not a deployable falls calculator. [14]

The older BBS review and the more recent accelerometer review remain useful contextual sources, but they cannot replace current primary-study appraisal. The accelerometer review included only four studies and did not establish an agreed sensor placement, configuration, or added value beyond comprehensive assessment. The SRRR3 consensus likewise did not establish a definitive biomechanical core set for steady-state, proactive, and reactive balance; its proposed measurement priorities should not be described as fully validated clinical prediction rules. [1, 16, 17]

Table 1 Outcome specific pooled accuracy reported by Kobayashi 2026

Complete main article examined; separate supplements unavailable. These are reported estimates across study-specific thresholds, not one transferable cutoff or uniform head-to-head comparison. Modified QUADAS-2 judged 48/55 high risk and omitted prespecified-threshold question. Main-text/table inclusion conflicts are discussed in the narrative.

Clinical assessment what the patient can do

Whole scale scores versus isolated standing constructs

The BBS is a 14-item functional balance scale, scored to 56. It samples tasks such as transfers, reaching, turning, altered base of support, and single-leg standing. Its total is therefore broader than quiet stance. The Mini-BESTest, scored to 28 in its current standard form, includes anticipatory control, reactive responses, sensory orientation, and dynamic gait. The full BESTest retains additional domains, including biomechanical constraints and stability limits/verticality. PASS, scored to 36, deliberately spans maintenance of posture and transitions at levels accessible early after stroke.

These scales answer useful but different questions. A BBS total can document functional progression and support handover. A Mini-BESTest profile can show difficult balance demands after a BBS ceiling has been reached. PASS can show gains before independent standing is achievable. None should be described as measuring a single latent 'standing stability' variable with perfectly interval-scaled units. A five-point change near the floor may comprise qualitatively different recovery from a five-point change near the ceiling.

For rehabtools, store the named instrument, version, scoring range, total, and item scores where legitimately available. Never combine 28-point and historical 32-point Mini-BESTest implementations. Preserve whether a short form was administered directly or reconstructed from a full scale. A system that extracts an isolated BBS standing item should identify that item and its own ordinal score, not inherit the reliability, responsiveness, or prognosis of the full BBS total.

Table 2 Clinical measurement the claim must match the test

These are complementary constructs, not interchangeable rankings of a single balance ability.

PASS and the early severely affected patient

The complete original PASS article separates two samples: 58 patients assessed longitudinally at days 30 and 90, plus 12 additional patients for reliability. Thirty healthy controls provided reference observations. The stroke sample comprised first unilateral supratentorial events, excluding dementia and other balance-affecting conditions; it does not directly validate the scale across every posterior-fossa or cognitively impaired presentation. Day-30 PASS correlated with concurrent FIM at 0.73 and day-90 FIM at 0.75. Those are unadjusted correlations, not a calibrated individual prognosis. [2]

Reliability involved two raters on the same day and repeat scoring by one rater three days later. Average item kappa was 0.88 between raters (range 0.64–1.00) and 0.72 within rater (0.45–1.00); total-score Pearson correlations were reported as 0.99 and 0.98 in the Results. Neither kappa nor Pearson correlation is an ICC or an MDC. The 12-person reliability sample is small, and some item agreement was weaker than the whole-scale summary suggests. The embedded item-kappa table was visually inspected and supports those distinctions. [2]

The full article also shows why phase and construct matter. Thirty-eight percent reached the maximum PASS score by day 90, so strong early utility does not eliminate a later ceiling. The instrumental comparator was an eight-second seated balance task on a laterally unstable rocking platform, completed by a selected subset; it was not a quiet-standing force-platform reference. The original scoring appendix distinguishes supported standing from unsupported standing, allows free foot position, and grades assistance and task duration. These details should remain attached to any digital implementation rather than replaced with a generic fixed-foot stance protocol. [2]

A particularly relevant comparison selected 49 people with severe balance deficits at days 14 and 30. The abstract reports a standardised response mean (SRM) of 0.79 for PASS versus 0.39 for BBS, with 42.9% versus 6.1% exceeding the respective MDC95. That supports choosing an instrument with enough low-level content when severe deficits create BBS floor limitations. The authors called changes beyond MDC 'clinically significant'; strictly, the reported criterion is detectable change, not independently established patient-important change. [3]

This is why severe stroke should not be squeezed into a standing-only digital workflow. A person may improve in rolling, unsupported sitting, moving from lying to sitting, or assisted standing while being unable to produce any safe 30-second unsupported stance. The appropriate record can include PASS and labelled sitting/trunk context, plus the maximum safely completed standing task. It should not imply that sitting balance is quiet standing or that progress in one construct guarantees recovery in another.

Lateropulsion further complicates assessment. Pushing behaviour and altered perceived verticality can change support requirements, weight distribution, and the direction of instability. A study of 44 subacute patients selected for pushing behaviour examined 137 repeated assessments and proposed a Burke Lateropulsion Scale threshold of at least 3 rather than at least 2. Repeated observations from the same people should not be counted as 137 independent patients, and that classification threshold is not a falls threshold. It is reasonable to record a dedicated lateropulsion measure when clinically present rather than force all asymmetric posture into a generic sway score. [25]

BBS reliability measurement error and changing severity

Stevenson's original study remains especially instructive because it estimated absolute error in routine inpatient practice. Forty-eight patients were assessed on consecutive days by two raters. The reported SEM was 2.49 BBS points; MDC90 was 5.8 and MDC95 was 6.9. Error differed across functional strata: MDC95 was 6.3 in independent walkers, 6.0 with standby assistance, and 8.1 with assistance. Agreement between the statistical change classification and clinicians' judgments was limited, reinforcing that detectable change and perceived clinical change are not the same. These are interrater, short-interval estimates in that inpatient sample, not a single stroke-wide rule of 'six points'. [19]

A later chronic-stroke study illustrates how different an estimate can be. Among 56 outpatients who could walk at least ten metres, generally around 22 months after onset, the BBS had ICC(2,1) 0.99 (95% CI 0.98–0.99), SEM 0.98, and MDC95 2.7 points. The study also found change during outpatient rehabilitation. Those estimates should not displace the inpatient estimates: the phase, ability range, rater arrangement, retest interval, and sources of variation differ. The article contains some reporting inconsistencies elsewhere in its tables and discussion; the specific BBS figures quoted here are those from its reliability table. [12]

Hiengkaew and colleagues explicitly examined measurement change in chronic stroke with different degrees of plantarflexor tone. Its inclusion in the evidence map is important because tone and motor severity can alter repeatability and test performance. The full text was not retrieved for this review, so subgroup-specific MDC values should not be implemented from secondary summaries. Any default uncertainty band in software should name the source population and protocol rather than choose the smallest published value. [26]

Hayashi and colleagues' complete multicentre report clarifies why the acute-stroke BBS MCID should remain anchor-specific. Seventy-five patients contributed BBS data, with baseline assessment averaging 4.2 days after onset and follow-up 12.9 days later. Participants had first supratentorial stroke, could understand the test, and could begin walking practice within five days; these criteria limit generalisation to the full severe-stroke population. A seven-level global rating treated improvement of at least +2 as meaningful, with patient and therapist ratings collected independently before physical testing. [27]

Table 3 reports a BBS ROC threshold of 12.5 points against the patient rating, AUC 0.74 (95% CI 0.62–0.85), sensitivity 0.55 and specificity 0.89. Against the therapist rating, the threshold was 6.5 points, AUC 0.78 (0.67–0.89), sensitivity 0.80 and specificity 0.72. The motor-FIM anchor gave inadequate discrimination, AUC 0.53 (0.39–0.66), so no ROC threshold was estimated. Those different operating characteristics should accompany the thresholds; the larger patient-anchored value was not a universally sensitive indicator of perceived improvement. [27]

Distribution-based results came from a different question: 64 patients had repeat BBS measurements at follow-up, yielding SEM 0.8 and MDC95 2.3 points; half the change-score SD was 4.9. The repeat-measurement interval and reliability model are not sufficiently specified in the accessible text to transplant this MDC into every clinical schedule. The article also contains a source inconsistency: its Table 2 patient-GRC mean changes, 10.7 versus 9.5, do not reproduce Table 3's stated 8.2-point change-difference estimate. That estimate should not be implemented without clarification. The independently tabulated ROC results can be described, while preserving this reporting concern and the separation of detectable from important change. [27]

Table 3 Selected clinimetric estimates retain design and units

All intervals are 95% CIs where shown. Do not substitute MDC90 for MDC95 or distribution-based change for patient importance.

The Mini BESTest full BESTest and short forms

Chinsongkram and colleagues' original 2014 BESTest study comprises a 12-person reliability sample and a separate 70-person validity sample at Prasart Neurological Institute. Its historical definition of subacute stroke was less than four months, not the contemporary seven-day-to-six-month convention. Participants had first unilateral hemispheric stroke, stable condition and capacity to follow instructions; exclusions included MMSE below 24, aphasia, relevant brainstem/cerebellar involvement and other balance-affecting disorders. No participants were diagnosed with neglect or pushing behaviour. Applicability to those commonly encountered presentations is therefore unproven despite the broad range of motor scores. [21]

Five therapists rescored the same recorded performances twice, seven days apart, after two training workshops. The authors specified ICC(2,k) for interrater and ICC(3,k) for intrarater reliability. Total-score estimates were 0.99 (95% CI 0.98–0.99) and 0.99 (0.99–1.00), respectively; interrater section ICCs ranged from 0.87 (0.72–0.96) to 0.98 (0.95–0.99). These are average-rating video-scoring estimates. They cannot be relabelled single-rater patient test–retest ICCs, and this paper supplies neither a new patient-performance SEM/MDC nor a longitudinal responsiveness estimate. A different patient's recovery assessment still includes administration, task-performance and day-to-day variation that rescoring a fixed recording holds constant. [21]

The validity sample deliberately contained 35 patients with total Fugl-Meyer motor score 0–55 and 35 above 55. The balance assessor was blinded to Fugl-Meyer scores; duplicate tasks across instruments were performed once and scored under each instrument's rules, with video review and adjudication when needed. The combined assessment took approximately 90 minutes, sometimes completed the next day; the Discussion estimates about 30–35 minutes for BESTest alone. BESTest correlations were 0.96 with BBS, PASS and Mini-BESTest and 0.91 with Community Balance and Mobility Scale. These are strong convergent associations in a wide, deliberately stratified severity spectrum. Shared tasks and common performance recordings also mean the correlations are not wholly independent confirmation of six distinct physiological balance systems. [21]

The floor/ceiling analysis is clinically more informative than a generic statement of validity. Twelve of 35 low-motor-function patients (34.3%) scored zero on Mini-BESTest, whereas no participant scored zero on the full BESTest. Only 4.3% of all participants attained a perfect BBS or PASS score, but the authors' separate 'responsive ceiling' definition, being within the highest 10% of the scale, included 45.7% and 60.0% of high-motor-function patients respectively. These near-ceiling proportions are not the same as maximum-score ceiling rates. Likewise, the full BESTest's lack of a total-score floor did not mean every task was feasible: many reactive, foam-eyes-closed and gait items scored zero in over 80% of the low-function group. Broader lower-level content preserves total-score range without demonstrating that inaccessible individual domains were measured precisely. [21]

The ROC endpoint was concurrent membership in the higher Fugl-Meyer motor group, not later walking or falls. BESTest greater than 49% had AUC 0.88 (0.80–0.95), sensitivity 71.4% (54.7–83.7) and specificity 91.4% (76.7–97.7). Mini-BESTest greater than 9 and BBS greater than 19 both had AUC 0.85. These data-derived thresholds in a 35:35 stratified sample are not stroke-wide screening cutoffs or prognostic thresholds. Table 3's printed negative likelihood ratios for Mini-BESTest and BBS do not agree with the corresponding sensitivity/specificity entries and should not be programmed without clarification. This is a source-reporting concern, not a justification to silently repair the published table. [21]

The 12/70 sample sizes resemble Winairuk's later study, but the reported sites and grouping rules differ: Lerdsin Hospital and lower-extremity Fugl-Meyer strata in Winairuk, versus Prasart and total motor-score strata here. Their reported BBS means also differ. Shared participants cannot be established from matching sample sizes, and the papers should not be labelled either confirmed duplicates or fully independent person-level replications without further documentation. [4, 21]

Winairuk and colleagues' 2019 study provides unusually useful detail about what 'reliability' meant. Five raters rescored videotaped performances from 12 people, while 70 additional people contributed validity and four-week responsiveness data. The short-form scores were largely extracted from the full BESTest. Excellent total-score agreement therefore primarily demonstrates reproducibility of scoring the same performance; it does not include all variation arising when a person repeats the test on another day. It also does not directly establish the administration burden or psychometrics of separately delivered short forms. [4]

The 70-person group was early after stroke, averaging about 16 days from onset, with BBS mean 31.24 and a very broad range. At baseline, 21.4% had the minimum Mini-BESTest score, compared with 5.7% for the full BESTest; S-BESTest and Brief-BESTest were intermediate. At four weeks, the Mini-BESTest ceiling remained low. This pattern is clinically coherent: difficult stepping, sensory, and gait tasks can be inaccessible early, while leaving useful measurement range later. High whole-scale correlation with BBS does not remove the floor problem in the subgroup most likely to need basic postural assessment. [4]

The same study estimated Mini-BESTest MDC95 at 3.35 points, based on reliability from rescoring video rather than full repeat-performance variability, but its anchor-derived change thresholds varied markedly. Using BBS change as the reference gave 7.5 points during the first two weeks and 4.5 during the next two; global-rating anchors produced different values. GRC AUC was 0.71 versus 0.68 for the BBS-change anchor in the first interval, and 0.69 versus 0.79 in the second interval. These are interval-specific point estimates, not evidence of statistically significant superiority. The BBS reference is another balance measure, not a patient-important endpoint. These values cannot all be treated as competing estimates of one universal MCID. They demonstrate the dependency of a change threshold on what the investigator defined as improvement. [4]

Tamura and colleagues' complete 2024 multicentre original clarifies the early-subacute Mini-BESTest MCID more precisely than its abstract. Fifty-three of 58 patients were analysed at two inpatient rehabilitation units. Eligibility required first supratentorial stroke within three months, stable condition, and ability to follow verbal instructions; severe visual impairment and other neurological or musculoskeletal conditions were excluded. Baseline testing occurred a mean 19.2 days after onset, with follow-up 14.0 days later. Mean baseline Mini-BESTest was 17.0/28, MMSE 25.3, and 37 patients walked without physical assistance, a category that included supervision; 16 required assistance. Thus, this is an early, selected rehabilitation cohort, not evidence for every severe or cognitively impaired presentation. [20]

The same physical therapist administered the Japanese Mini-BESTest at both assessments, with scoring-video training across sites. Patient and therapist global ratings were obtained independently before follow-up testing; patient interviews were conducted by staff not directly delivering physical therapy. The seven-level anchor defined at least +2, 'significantly better', as meaningful, deliberately excluding 'slightly better'. Thirty-three patients rated themselves improved, whereas therapists classified only 18 as improved. The therapist anchor correlated with Mini-BESTest change at ρ=0.54, exceeding the prespecified 0.50 criterion; the patient anchor correlated at only 0.19 and did not support MCID estimation. These safeguards reduce some immediate score-influence bias, but the therapist still had clinical knowledge of the patient's progress and the design was not an independent blinded outcome assessment. [20]

The walking-ability subgroup results materially restrict interpretation. Among the 37 unassisted or supervised walkers, therapist-anchor correlation with score change was 0.57; among the 16 assisted walkers it was 0.26. Patient-anchor correlations were 0.33 and −0.39 respectively. These small subgroup analyses are imprecise, but they do not validate the pooled threshold specifically for assisted walkers. Nor does failure of the patient anchor mean that patient-perceived recovery is unimportant: it indicates that this particular global question and the Mini-BESTest did not capture sufficiently aligned change in this setting. [20]

Four statistical approaches were planned, but only three yielded estimates. Among the 17 people assigned therapist GRC +2, mean change was 4.5 points (95% CI 3.2–5.8). The ROC threshold was also 4.5, with AUC 0.79 (0.67–0.91), sensitivity 0.56 and specificity 0.91; its bootstrap threshold interval was 1.5–4.5. Predictive modelling yielded 3.2 (2.6–3.8). The improvement-rate-adjusted method was not used because baseline, follow-up and change scores were nonnormally distributed. Confidence intervals for the ROC and predictive methods used 1,000 bootstrap samples. A random-effects weighted average of the three estimates gave 3.8 (2.9–5.0). The three outputs reuse the same patients and anchor; their pooled interval should not be interpreted as independent multi-cohort replication, and pooling does not remove shared recall bias or uncertainty about the anchor. [20]

A rounded four-point gain can therefore be shown as a therapist-anchored, two-week interpretive reference for comparable patients. It is not the same operating rule as the ROC cutoff of 4.5, which corresponds to at least five points on an integer-valued score. The ROC sensitivity of 0.56 also means that this more specific cutoff missed many therapist-rated improvements in the derivation sample. The Methods text misdescribes the Youden objective as sensitivity plus 1−specificity rather than sensitivity plus specificity minus one; the article also identifies published R code. This reporting ambiguity cannot establish which implementation was executed, so the reported threshold should remain a study result rather than a silently reconstructed algorithm. [20]

Crucially, Tamura did not estimate test–retest SEM or MDC in these 53 participants. The Discussion's comparison with an MDC of approximately 3.4 points imports Winairuk's estimate, which was based on rescoring video. It does not establish a matched two-week repeat-performance error threshold in Tamura's cohort. The pooled MCID interval also extends below that borrowed MDC. Neither the numerical proximity to another setting's MCID nor the authors' broad generalisability argument demonstrates phase-invariant meaning. Four points should not trigger an automatic declaration of reduced falls, independent walking, improved daily activity, or treatment efficacy: those outcomes were not validated by this anchor study, and spontaneous recovery plus usual rehabilitation were not separated experimentally. [4, 20]

A 2024 prospective comparison reinforces the need to match the scale to the sample. Among 58 subacute patients, BBS and Mini-BESTest correlated strongly, but 43.1% reached a perfect BBS score at discharge versus 5.2% for Mini-BESTest. Admission scores had only moderate discrimination for inpatient falls, with AUCs around 0.73–0.74. The paper's BBS falls sensitivity differs between narrative text and a figure caption, so that sensitivity should not be hard-coded. The broad point is clearer than the disputed number: avoiding a ceiling can improve measurement range without automatically producing a superior prognostic model. [5]

The chronic-stroke Brief-BESTest study included 50 stroke participants, 27 of whom completed reliability testing. Total-score ICC(2,1) values were 0.972 intra-rater and 0.974 interrater, but agreement for individual reactive and strength items was appreciably weaker. The reported cutoff below 18 distinguished stroke participants from controls, and below 14 distinguished outdoor aid users from nonusers. Neither was a prospective falls threshold. A short composite can be reliable overall while one clinically important component still requires careful examiner training. [28]

Timed stance single leg stance reach and the Step Test

A timed stance test and an instrumented stance test answer different questions. The former asks how long a specified posture can be maintained before a defined failure; the latter asks what happens during a defined recording interval. A person who completes 30 seconds can be at the ceiling of a timed task while still showing substantial sway. Conversely, computing a quiet-standing metric from a failed five-second attempt and comparing it with a 30-second reference conflates duration, task failure, and the signal statistic.

The 50-person chronic-stroke reproducibility study by Flansbjer and colleagues found ICC(2,1) values of 0.88 for BBS, 0.88 for nonparetic single-leg stance, and 0.92 for paretic single-leg stance. Nevertheless, individual SRD was 42% for nonparetic and 74% for paretic single-leg stance, compared with 8% for BBS. These full-manuscript findings are a useful warning against equating a high ICC with sensitivity to individual change. The protocol averaged three alternating-leg trials capped at 15 seconds, with the same assessor after seven days, comfortable shoes, and usual ankle–foot orthoses allowed. Participants could walk at least 300 m; the same 50 people also contributed the 2005 gait and knee-strength reliability studies, so these are shared-cohort measurements. [22]

Single-leg stance must identify the stance limb. 'Affected-side test' can mean standing on the affected limb or moving that limb, with opposite implications. Record the tested support limb, raised limb, arm position, footwear/orthosis, permitted support, maximum duration, failed attempts, and termination criterion. A hand touching a rail is not equivalent to an examiner standing nearby without contact. A support-assisted measurement may be useful, but it is a different condition and should not inherit unsupported reference values.

Functional Reach samples voluntary displacement without changing the base of support. It is affected by arm/trunk strategy, shoulder limitations, initial stance, and whether heel rise or stepping is permitted. In Bower's chronic/late-subacute outpatient sample, functional reach showed good repeatability but an MDC of 6.24 cm. It correlated moderately with eyes-open total COP velocity, not perfectly. Reach should therefore be treated as an anticipatory functional task rather than a direct geometric measure of the maximum stable COP envelope. [7]

The Step Test adds a different demand: repeated foot clearance and weight transfer while the other limb supports the body. It is not quiet standing, but it can be especially informative after unilateral stroke. In Bower's prospective falls cohort, performance in each stance-limb condition was associated with future falls after adjustment, whereas quiet-standing COP velocity did not. The stance-versus-moving-limb label must be explicit; otherwise a graph can reverse the clinical interpretation. A full account of gait and stepping belongs alongside, not inside, a claim about static standing sway. [8]

Reactive and anticipatory balance

Reactive balance cannot be inferred reliably from a comfortable quiet stance or from voluntary reaching. It concerns response to an unexpected disturbance and may involve a compensatory step, grasp, multiple steps, or external assistance. The direction of perturbation, paretic versus nonparetic stepping, availability of a preferred limb, and perturbation intensity all matter. Stroke-specific asymmetry makes a single undifferentiated 'pull test' score particularly liable to conceal useful information.

Handelzalts and colleagues exposed 15 people with stroke and 15 controls to multidirectional, progressively stronger surface translations. Trained observers had 100% agreement on the harness-defined fall threshold and very high agreement on step strategies. The threshold correlated with BBS and walking endurance. This supports reliable scoring of that laboratory response, not validation of actual future community falls. The harness outcome is a controlled task failure with safety equipment, not a naturally occurring fall. [24]

Anticipatory balance includes preparatory adjustments before a voluntary movement, such as raising an arm or initiating a step. Garland and colleagues followed 27 people during a month of early rehabilitation and observed improvement in functional and physiological measures, while some participants compensated with nonparetic rather than restored paretic anticipatory activation. The abstract-level evidence is sufficient to reject an oversimplified assumption that improved function always means normalised bilateral control. Detailed electromyographic thresholds and treatment interpretation require the full original paper. [11]

For practical assessment, select reactive and anticipatory tasks because they address a clinical question the quiet-stance test cannot answer. Record actual response strategy and assistance, not just the total score. Challenging perturbation testing requires trained supervision and an appropriate safety arrangement. This report does not provide an unsupervised home perturbation protocol or imply that a phone measurement can substitute for the examiner's role.

Instrumented standing what the signal represents

Net COP centre of mass and limb specific control

A force platform estimates COP from forces and moments at the support surface. COP is not the body's centre of mass (COM), and neither is directly interchangeable with trunk acceleration measured by a wearable. A person can use ankle, hip, trunk, and stepping strategies that alter these signals differently. It is consequently inappropriate to label any convenient sensor output 'sway' and apply a threshold validated for a different physical quantity.

A single plate provides net COP beneath both feet. Two plates, one per foot, additionally allow separate vertical loading and limb-specific COP behaviour. This matters after stroke because the less affected limb may carry more weight, generate a disproportionate share of corrective activity, or compensate for impaired control on the paretic side. A relatively ordinary net trajectory can conceal unequal contributions. Equally, greater asymmetry is not necessarily equivalent to greater instability: the relationship depends on task demands and the person's compensatory strategy. [6, 7, 9, 29, 30]

At least four quantities must remain separate: percentage body weight on the paretic side; an absolute asymmetry magnitude; an index of the relative contribution of each limb's COP movement; and temporal synchronisation between the limbs. A signed value preserves direction, whereas an absolute value discards whether loading is towards the paretic or nonparetic side. Two patients can have the same absolute asymmetry with opposite loading patterns. An index derived from velocity may behave differently from one derived from displacement. Software labels and formulas should make these distinctions visible rather than present a single generic 'symmetry' percentage.

Mean net COP position on one plate is also an imperfect substitute for direct load measurement under each foot. It reflects foot placement and the location of the resultant force, as well as loading. Gasq and colleagues specifically noted this limitation when interpreting mediolateral COP position. If the clinical question is paretic loading, directly measured vertical force from separate support surfaces is conceptually preferable to inferring it from a body outline or single net COP coordinate. [6]

Between day measurement error the Gasq study

Gasq and colleagues recruited 20 people at least one month after stroke who could stand independently for five minutes. Their mean time since stroke was 10.3 months, with a range extending from one to 37 months, and PASS performance was relatively high. The protocol used barefoot stance, heels 3 cm apart, toes angled out by 30 degrees, three 51.2-second trials in each visual condition, one-minute seated rests, and a seven-day retest at the same time of day without medication change. Data were sampled at 40 Hz, and the three trials were averaged. [6]

Under that exact protocol, resultant COP velocity had ICC(2,k) 0.94 (95% CI 0.84–0.98) with eyes open and 0.97 (0.91–0.99) with eyes closed. SEM/SRD were 2.1/6.1 mm/s and 3.2/9.5 mm/s, respectively. Directional velocity measures were similarly relatively reliable. In contrast, eyes-open 90% ellipse area had ICC 0.76 (0.38–0.90), and mediolateral mean COP position had ICCs of 0.78 and 0.71. Those confidence intervals matter: a high point estimate from a small sample does not establish uniformly precise measurement.

The study's error analysis is as valuable as its ICCs. Ellipse-area error increased with the magnitude of the measurement, so a constant error band in square millimetres was inappropriate. The investigators reported ratio and percentage approaches after assessing heteroscedasticity. For eyes-closed area, the SRD ratio was 2.16; for eyes-open area it was 3.68. Those are large changes relative to the original measurement. A visually impressive area plot can therefore be a less robust individual monitoring outcome than a simpler velocity statistic. The study does not establish that velocity is universally optimal; it establishes that metric selection needs absolute-error evidence under the intended protocol.

For rehabtools, this translates into three requirements. The number of averaged trials must be linked to the uncertainty estimate. Error expressed as a percentage or ratio must not be converted casually into a fixed absolute MDC. Finally, an apparent improvement smaller than the applicable error should be reported as uncertain, even when the display can show many decimal places.

A lower cost implementation Bowers Wii Balance Board study

Bower and colleagues examined 30 outpatients more than three months after noncerebellar stroke; median time since stroke was 13.5 months. Participants could stand unsupported for 30 seconds and walk independently with or without aids. Severe dysphasia, dyspraxia, and cognitive impairment below the study's MMSE criterion were excluded. The findings therefore do not automatically transfer to acute, nonambulatory, cerebellar, or substantially cognitively impaired stroke. [7]

The custom software sampled COP at 40 Hz and used an eighth-order low-pass filter at 12 Hz. Static stance was recorded for 30 seconds with eyes open and closed; weight distribution used a board beneath each foot. Dynamic lateral weight shifting and sit-to-stand force variables were also collected. Three trials were summarised by their median for most tasks, whereas one trial was used for static weight-bearing asymmetry. The week-apart retest used a fixed testing order and the same assessors. The combined board assessments took about 21 minutes on average, not simply the time occupied by one 30-second recording.

For eyes-open total COP velocity, ICC(2,k) was 0.87 (95% CI 0.73–0.94), SEM 0.19 cm/s, and MDC95 0.53 cm/s, approximately 40.8% of the study mean. Eyes-closed total velocity had ICC 0.94 (0.87–0.97), SEM 0.26 cm/s, and MDC95 0.71 cm/s, approximately 35.1%. Paretic loading showed ICC 0.82 (0.64–0.91) and MDC95 10 percentage points of body weight. Lateral weight-shift count was more repeatable, with ICC 0.98 and MDC95 1.45 shifts per 30 seconds. These are useful estimates, but they apply to the complete device, processing, trial-summary, and task protocol.

Dynamic lateral shifting related more consistently to clinical dynamic-balance tasks than did static loading asymmetry. Static asymmetry showed little relationship with the included clinical tests. This is useful differentiation rather than a reason to combine the outcomes indiscriminately. If a clinician asks whether a patient can transfer weight voluntarily, the lateral-shift task may answer that question more directly than quiet-standing asymmetry. If the question is whether spontaneous loading changes, the quiet-standing measure remains the relevant construct.

Subacute reliability and newer limb specific metrics

Aryan and colleagues' 2023 study addressed an important phase gap by evaluating 24 people, on average 41 days after stroke, using two 30-second eyes-open trials. Mean AP COP speed and directional weight-bearing asymmetry both had ICC(2,1) approximately 0.91; MDC95 values were 7.84 mm/s and 7.02 percentage points of body weight, respectively. Speed-based symmetry was more reliable than displacement-based symmetry. However, this was within-session repeatability, with a mean retest interval of about 11.5 minutes and some repetitions separated by a minute or less. It is not a between-day recovery threshold. [29]

The same study found considerably weaker repeatability for some attractive 'advanced' measures: AP displacement RMS and median AP power frequency had low ICCs, and ellipse-area MDC was large. Nonlinear or frequency-domain terminology does not itself confer superior clinical validity. Such metrics may be informative in mechanistic research, but their signal-length requirements, preprocessing, estimator parameters, and individual error need explicit justification before they enter a routine clinical dashboard.

Jagroop and colleagues studied 20 people more than six months after stroke, comparing two 30-second trials. The verified abstract reports average-measure ICC(3,2) values of 0.79–0.95, while single-trial ICC(3,1) values for symmetry and synchronisation were below 0.75. Possible proportional error was also noted. The practical implication is to preserve the distinction between a single trial and an average of trials. The full article remains necessary to implement exact confidence intervals, formulas, and metric-specific limits rather than extrapolate from a summary range. [31]

Table 4 Quantitative standing reliability protocol specific examples

EO=eyes open; EC=eyes closed. Absolute error estimates are not MICs and should not be pooled across devices or summaries.

What newer validity studies do and do not validate

Aryan and colleagues' later validity study analysed stored subacute data from a Toronto cohort. Eligibility for the original assessment required independent standing and understanding instructions; 20 of 102 original participants lacked a valid 30-second trial. Forty-eight contributed contemporaneous BBS comparisons and 75 contributed retrospective fall-history/risk comparisons. AP and ML COP speed correlated moderately with BBS; weight-bearing asymmetry and symmetry index did not. Paretic loading discriminated prior acute-care fallers only modestly, with AUC 0.67 (95% CI 0.51–0.83). This is retrospective discrimination, not prediction of new falls. [30]

The paper is especially valuable for showing why a physiological metric and a functional scale may diverge. BBS performance can improve through compensation without a normalisation of loading or limb contribution. A weak correlation does not necessarily invalidate the force signal for its intended biomechanical purpose. It does limit claims that the signal is a substitute for the scale. Likewise, the study's internally selected paretic-loading cutoff of 47.8% should not become a universal falls warning at 'less than half body weight'.

The Toronto research programme produced several related analyses of standing control, reliability, and falls. Reuse of historical datasets means that publications may not represent independent replications. The exact degree of person-level overlap is not always ascertainable from the accessible reports; it is therefore flagged rather than assumed absent. A large count of papers should not be presented as a corresponding count of independent external validation cohorts. [9, 29–31]

Sensory manipulation and the distinction between SOT and mCTSIB

Removing vision, standing on foam, or using a moving support surface challenges different aspects of postural control. These manipulations can reveal a deficit not evident in comfortable eyes-open stance. Nevertheless, an increase in sway with eyes closed is not a diagnosis of a single sensory lesion. Sensory integration, motor capacity, fear, attention, foot placement, and the difficulty of the condition all contribute. A visual ratio can be unstable when its denominator is small, and an uncompleted condition cannot be represented as a numerical ratio without an explicit rule.

The Sensory Organization Test (SOT) uses a specific apparatus and multiple combinations of visual surround and support-surface manipulation. It is not the same as a four-condition modified Clinical Test of Sensory Interaction on Balance (mCTSIB) performed on firm ground and foam. A quiet-standing phone assessment should not use SOT normative categories simply because both involve eyes-open and eyes-closed conditions.

A prospective SOT study analysed 84 first-ischaemic-stroke participants after discharge, with 32 fallers over six months. SOT composite cutoff 60 had sensitivity 71.9%, specificity 65.4%, and AUC 0.686; combining SOT, BBS, and Falls Efficacy Scale–International increased sensitivity at a substantial specificity cost. Selection favoured people with several minutes of unsupported standing and excluded severe language impairment. The report contains inconsistent standing-duration and FAC eligibility wording, further limiting precise protocol transfer. These internally derived thresholds are not externally validated decision rules. [32]

The newer instrumented mCTSIB composite study included 54 people with first stroke and reported AUC 0.84 for discriminating walking independence and moderate association with Mini-BESTest. Its verified abstract describes regression-derived weighting across conditions. The endpoint is contemporaneous walking status, not future falls or recovery, and the composite's apparent performance may be optimistic without independent validation. It is an interesting candidate metric, not an established replacement for the individual conditions or a prognostic score. [33]

Wearables repeatability is not sufficient for monitoring

Wearable sensors can improve portability and expose trunk motion that is not visible in an ordinal scale. They also introduce dependencies on anatomical placement, attachment, sensor orientation, calibration, sampling, drift correction, and the chosen acceleration/angular-velocity summary. A lower-back signal cannot be assumed equivalent to an upper-trunk signal, and a value derived from angular displacement is not directly comparable with COP velocity. The name of the physical measurement and its units should accompany the result.

Perez-Cruzado and colleagues' early single-leg-stance sensor study involved only four chronic stroke survivors. Although observer reliability estimates appeared favourable, that extremely small sample and exclusion of severe neglect make it exploratory evidence. It cannot establish broad phase-specific reference values, robust MDCs, or prospective fall prediction. The practical contribution is a clearly described instrumented task and a feasibility demonstration, not permission to treat a sensor-enhanced single-leg test as a validated clinical risk product. [23]

Geerars and colleagues provide a much stronger test of the monitoring proposition. Across five rehabilitation centres, 115 people were enrolled and 94 supplied repeated assessments. Testing averaged 3.3 weeks after stroke initially and 9.2 weeks at the final assessment. Standing tasks used an L5 sensor; eyes-open stance lasted 60 seconds and eyes-closed/foam conditions 30 seconds. The output was a specifically processed IMU 'path' feature with study-specific units, not force-platform COP path length. Mini-BESTest was selectively administered when BBS was at least 45. [10]

Only 16.5% of those with paired eyes-open standing measurements improved beyond the IMU MDC, compared with 54.4% improving beyond the BBS threshold. The corresponding improvement proportions for eyes-closed and foam standing were 4.2% and 3.2%. Overall, 67.4% of prespecified responsiveness hypotheses were rejected. Some participants showed increased sway despite improved functional scores. These findings argue against using reduced sway as a generic proxy for functional recovery.

Several qualifications matter. The study tested a particular sensor feature and task set, not every possible wearable measure. Missing standing conditions were related to ability and technical/organisational factors; selective Mini-BESTest administration produced a smaller, higher-functioning subset. Global ratings were incomplete and only partly concordant between patients and therapists. These issues limit the generality of a negative result, but they do not rescue the assumption that greater numerical sensitivity necessarily produces better clinical monitoring.

The useful product response is to show trajectories side by side: task completion, assistance, clinical scale, and the specific sensor metric. A pattern of improved BBS with unchanged sway is a finding to interpret, not an error to hide. It may motivate a focused examination of strategy, confidence, or sensory dependence; it does not independently show failed rehabilitation.

Markerless video and pressure based alternatives

A camera can estimate movement and body configuration, but it does not directly measure vertical force. Sheehy and colleagues compared a Kinect-based Weight Shift Tool with a pressure mat in 12 younger adults, 12 older adults, and 12 people with stroke. In the stroke group, average bias could be small while limits of agreement were very wide: during equal stance without an arm raise, bias was 0.6 percentage points but limits spanned approximately −27.9 to +29.1. Such agreement is inadequate for substituting the camera estimate for measured limb loading in an individual. [34]

The lesson is not that every future video method will fail. It is that a visually plausible body-centre estimate or a high correlation cannot establish interchangeability for weight distribution. Posture, body shape, trunk lean, and the way a person generates force can dissociate geometry from loading. A claimed improvement in symmetry of a few percentage points would be uninterpretable with error of that magnitude.

A 2024 pressure-plate versus dual-force-plate study is relevant technology context but was conducted in healthy adults, not a stroke cohort. Some measures correlated strongly across devices while absolute COP values were systematically lower on the pressure plate, with proportional bias. It therefore supports caution about device substitution, not stroke-specific criterion validity. A device should not inherit another instrument's normative values or MDC merely because both output a variable called COP velocity. [35]

Video-based clinical scoring is another distinct use case. It may improve documentation or facilitate rater review, but agreement with a therapist's score does not validate biomechanical force estimates, and a machine-learning classifier trained on scale labels does not automatically predict future falls. Validation must match the actual product claim: timing, task completion, ordinal scoring, joint kinematics, load estimation, and prognosis each need their own evidence.

Falls prognosis identify the event before interpreting the test

Falls are a person task environment outcome

Standing balance is only one contributor to falls after stroke. The event also depends on whether the person attempts transfers or walking, supervision, assistive-device use, cognition and attention, environmental obstacles, continence-related urgency, vision, medication effects, and the ability to recover after instability. Restricting mobility may reduce exposure while leaving capacity poor; increasing mobility may increase opportunities to fall while function improves. A simple linear assumption that better performance always means fewer observed falls is therefore unsafe.

Falls outcomes differ materially across studies. Any fall, two or more falls, total fall rate, serious injury, and a composite of recurrent or injurious falls should not be pooled conceptually. An inpatient fall during a short rehabilitation stay occurs under different exposure and supervision from a community fall over a year. A cutoff optimised for sensitivity to any event may behave poorly when used to identify repeated falls or serious injury.

Prospective diaries with reminders and follow-up provide stronger ascertainment than a single retrospective question. Even prospective studies lose participants to death, recurrent stroke, discharge changes, or nonresponse. Excluding those people without considering why data are missing may remove patients whose prognosis differs most. The number of participants originally recruited should therefore be separated from the number analysed and the number of events that support the model.

Clinical balance and future recurrent or injurious falls

The LEAPS analysis is a useful anchor because it was relatively large and explicitly prospectively ascertained falls. Four hundred eight participants were assessed at about two months after stroke and monitored for approximately ten further months using monthly calendars and follow-up calls. They had residual paresis, could walk ten feet with at most one-person assistance, followed a three-step command, and walked slower than 0.8 m/s. This is a selected rehabilitation-trial cohort rather than an unselected stroke population. [36]

There were 147 participants with multiple and/or injurious falls, 88 with a single noninjurious fall, and 173 nonfallers. The analysis considered 41 candidate variables and used classification trees with ten-fold cross-validation. BBS at or below 42 was the strongest single split for multiple/injurious falls. Apparent sensitivity/specificity were 73%/53%; cross-validated values were 78%/39%. Adding predictors improved fit within the sample but did not generalise well under cross-validation. This is a far more cautionary result than a statement that 'BBS 42 predicts falls accurately'.

The endpoint itself matters. The LEAPS split does not provide an individual probability of any fall, and the low cross-validated specificity implies many false-positive classifications. The study's intervention context also affected activity exposure and falls; this report uses those observations to understand prognosis, not to review treatment efficacy. The strongest practical inference is that impaired clinical balance remains relevant, but simple dichotomisation leaves much unexplained.

Mackintosh and colleagues followed 55 people for six months after stroke rehabilitation using prospective diaries. The abstract reports 25 fallers, including only 12 recurrent fallers. A combination of inpatient fall history and poor balance, represented by BBS below 49 or Step Test below 7, gave sensitivity and specificity above 80%. That combination is clinically plausible, but a model based on 12 recurrent events is vulnerable to unstable cutoff selection and optimism. The full paper is needed before treating the exact combination as an implementable rule. [37] Walsh et al. (2017; https://doi.org/10.1093/ageing/afw255) independently evaluated the BBS-plus-inpatient-fall rule in a cohort of 128 recruits: 117 had six-month outcomes, and 110 were included in the complete-case analysis. Discrimination was poor: AUC 0.56 (95% CI 0.46–0.67), sensitivity 18.8% and specificity 93.6%. That validation did not test the Step Test version.

Ashburn and colleagues recruited 122 people and obtained one-year fall status for 115; 48 experienced repeated falls. Several balance and functional variables screened as candidate predictors, but the final combination emphasised near-falls in hospital and upper-limb function, with sensitivity 60% and specificity 70%. The abstract-level result is important because it does not place standing balance alone at the centre of every model. The capacity to arrest or protect against a fall and observed instability during real care can matter alongside a formal scale. [38] In the independent Walsh et al. cohort (2017; https://doi.org/10.1093/ageing/afw255), 110 participants had twelve-month outcomes and 102 entered the complete-case analysis. The modified two-predictor near-fall/upper-limb rule discriminated poorly: AUC 0.55 (95% CI 0.44–0.66), sensitivity 51.9% and specificity 58.7%. Validation substituted a reweighted Motor Assessment Scale upper-limb score for the original Rivermead Motor Assessment measure; it was not an identical predictor replication or a test of every related model.

Acute postural control and later falls

In FallsGOT, 490 people entered the acute-stroke cohort, but fall outcomes were available from fewer respondents at each later time point. The verified report gives 140 fallers among 348 respondents at 12 months. Poor early postural control and walking-aid use were associated with subsequent falls, with adjusted odds ratios of 3.92 and 2.84 respectively. This supports early postural assessment as part of risk formulation; it does not establish the same absolute risk in every person with an aid or a poor score. [39]

The earlier POSTGOT study assessed 96 people in the first week and recorded falls through one year. Multiple clinical tests were associated with falls, but inability to perform the ten-metre walk was the strongest reported indicator. This is a reminder that inability is informative and that a standing-only instrument can omit important risk information. The two Gothenburg reports come from the same research setting but should not be described as independent external validation of the same model without confirming their recruitment and model relationships. [40]

Inpatient risk thresholds are particularly context-bound. A secondary record analysis of 818 stroke rehabilitation admissions found an optimal admission BBS cutoff of 29 for inpatient falls, with sensitivity 82.4% and specificity 57.4% in the verified abstract. That threshold is much lower than several post-discharge thresholds because the population, supervision, endpoint, and exposure differ. It should not be averaged with community thresholds to create a compromise cutoff. [41]

Quiet standing variables and direct prospective falls evidence

Bower and colleagues recruited 96 people before discharge and obtained 12-month falls data for 81, assessed at a median of 24 days after stroke. Twenty-three participants fell and 13 fell recurrently. Standing COP velocity was measured with eyes open and closed, in total, AP, and ML directions; clinical dynamic balance and instrumented gait were also evaluated. Candidate-specific ordinal regression models of prospective falls adjusted for country, prior falls, and assistance, with additional models including comfortable walking speed. [8]

None of the six quiet-standing COP velocity variables was statistically significant in those models. For example, eyes-open total velocity had adjusted interquartile-range odds ratio 1.26 (95% CI 0.61–2.61), or 1.29 (0.61–2.76) with gait speed included. In contrast, poorer Step Test performance and slower standard TUG remained associated with falls. The affected- and less-affected-stance-limb Step Test comparisons gave substantial odds ratios, but wide confidence intervals showed uncertainty. The study did not deliver an externally calibrated absolute-risk calculator.

The negative static finding should not be explained away simply because sway differed numerically between groups. It means that the study did not establish the proposed independent association at its sample size, outcome definition, and protocol. It also does not prove zero prognostic value for every static metric: velocity is not ellipse area, net sway is not limb coordination, and a single future fall is not a fall count. The appropriate conclusion is narrower and more useful: this clinically relevant prospective study provides no basis for treating its quiet-standing COP velocity measures as stand-alone 12-month falls predictors.

Mansfield and colleagues' complete original report provides a complementary, more specific finding. Of 419 people discharged during recruitment, 172 were eligible, 100 consented, and 95 supplied follow-up; 35 reported 83 falls over up to six months, while only 71 completed reactive stepping assessment. Eligibility required independent ambulation and home discharge. The quiet-standing protocol used 30 seconds of eyes-open dual-plate recording at 256 Hz with a 10-Hz filter. Paretic contribution was the paretic AP-COP displacement RMS divided by the sum of both limbs' RMS values; synchronisation was the correlation between left and right AP-COP signals. Neither is simply percentage body-weight loading. [9]

Reactive assessment used a forward lean-and-release with approximately 10% of body weight supported before release, a therapist and safety harness, and five trials each of usual-response and encouraged-use conditions. In the latter, the preferred stepping limb was blocked, testing whether the participant could step with the other, unblocked limb. The important failure variable was inability to step effectively with the unblocked limb, including an attempted blocked-limb step or a sliding response. It is not a general score for preferring the nonparetic limb. [9]

No candidate significantly predicted binary any-fall status after the study's multiplicity correction, so multivariable logistic models were not pursued. Positive findings came from Poisson falls-rate analyses. Adjusted rate ratios were 0.8 (95% CI 0.7–1.0) per 0.1 higher paretic contribution, 0.9 (0.8–0.9) per 0.1 higher synchronisation, and 1.2 (1.0–1.3) per 20 percentage-point increase in unblocked-limb stepping failure. Step-length variability also remained associated, at 1.4 (1.2–1.7) per centimetre. ML net-COP RMS lost significance after adjustment. These rates, increments, and endpoints must be retained; they cannot be translated into a universal probability of becoming a faller. [9]

Adjustment included age, NIHSS, BBS, and walking time, but walking time was self-reported during follow-up and averaged across three assessments. This makes the analysis useful for examining association conditional on activity exposure, but it is not a discharge-only model with all inputs known at prediction time. Predictor screening, the restricted reactive-testing subset, self-reported activity, and incomplete follow-up further constrain transportability. The article does not provide external validation, clinical calibration, or a decision threshold for implementation. Its recruitment period, October 2010 to March 2013, overlaps the historical Toronto dataset dates used in later force-platform analyses, so these papers should not be counted as independent replication without person-level reconciliation. [9]

The contrast suggests a sensible research priority: test whether limb-specific or reactive measures add clinically meaningful information beyond fall history, clinical balance, mobility, and activity exposure. It does not yet justify replacing those inexpensive predictors with a more elaborate apparatus. Incremental model value needs direct testing; finding one significant sensor coefficient in an adjusted regression is not by itself evidence of better clinical decisions.

Table 5 Falls prognosis direct future events and tempting substitutes

Thresholds refer to different outcomes and populations. Se=sensitivity; Sp=specificity. They are not recommended universal clinical cutoffs.

Why retrospective classification must remain separate

The 49-person retrospective sway study by Jonsdottir and colleagues classified falls during the previous six months. It combined clinical scales and stabilometric variables using stepwise regression. Apparent AUC was 0.68 for the clinical model and 0.74 for a selected combined model; the latter retained BBS, Barthel Index, and ML sway amplitude. Smaller rather than larger ML sway was associated with faller status in that selected model. [42]

This result cautions against imposing a simple monotonic interpretation on every sway feature. It also cannot demonstrate future prediction. Small sample size, 21 prior fallers, data-driven variable selection, uncertain temporal direction, and absence of independent validation make the apparent increment provisional. The table additionally gives a confidence interval excluding one for one selected odds ratio while the accompanying p value is 0.06, so borderline significance should not be overinterpreted.

The later Aryan paper similarly used prior acute-care falls and a contemporaneous STRATIFY risk classification. A receiver-operating curve against either label is not a prospective falls validation. A model trained to reproduce a clinician's risk classification can at best reproduce that classification's limitations unless later real events are independently collected. For rehabtools, the endpoint field should explicitly distinguish 'recalled prior fall', 'clinical risk category', 'inpatient incident fall', and 'prospectively recorded community fall'. [30]

Balance mobility exposure and apparently discordant directions

Simpson and colleagues prospectively followed stroke survivors and matched controls with monthly falls diaries. Eighty stroke participants were analysed, with 109 falls. In the stroke model, BBS was associated with fewer falls, but faster TUG performance was associated with more falls after accounting for other variables. The investigators used negative-binomial models, retaining age and cognition alongside BBS and TUG. [43]

This does not mean that slow mobility should be encouraged or that improving TUG causes falls. It illustrates the difference between capacity and exposure and the conditional meaning of a regression coefficient. When balance is held constant, someone moving more quickly or more often may have more opportunities to encounter hazards. Collinearity, model selection, residual confounding, and sample limitations can also affect coefficient direction. A single score should therefore inform a broader clinical formulation rather than override knowledge of what the patient actually does.

Patient-reported concern about falling is likewise complementary rather than interchangeable with physical performance. Confidence may reflect experience, avoidance, insight, or anxiety. A high-confidence person can still have poor reactive capacity; a cautious person can have relatively good measured balance. Clinical and patient-reported measures can legitimately disagree, and the discrepancy may itself identify a useful discussion or assessment target.

Recovery walking and independence prognosis

Functional recovery is not normalisation of every balance mechanism

Stroke recovery can improve task success through restoration of impaired control, compensatory strategies, or both. Clinical scales largely credit successful functional performance. A force-platform or wearable measure may detect a persisting asymmetry that does not prevent the task, while a patient gains independence through a safe compensatory strategy. Conversely, a small improvement in a physiological variable may be insufficient to change assistance needs or everyday function.

This distinction explains why assessment should retain both capacity and mechanism where they matter clinically. It also limits surrogate reasoning. A treatment-associated reduction in sway does not show reduced falls, and an increase in BBS does not establish normalisation of paretic contribution. Neither should be interpreted as a universal marker of neural recovery. The Garland physiological cohort and the newer Geerars sensor cohort provide different empirical examples of this separation. [10, 11]

Earlier instrumented recovery cohorts

The earlier de Haart cohort is an important counterweight to interpreting the newer wearable study too broadly. Thirty-seven postacute inpatients were first tested once they could stand unsupported for 30 seconds, on average around ten weeks after stroke, and reassessed over twelve weeks. The verified abstract describes reduced sway and visual dependence, especially in the frontal plane, while static and dynamic asymmetry did not normalise alongside functional improvement. This supports the possibility of genuine physiological change in a selected phase and protocol; it does not establish that every sway feature tracks functional recovery or that observed change was caused by the rehabilitation treatment. [44]

A related 36-person weight-shifting report used visually guided frontal-plane COP movement. Speed and precision improved, but a directional time asymmetry persisted; older age and neglect were associated with slower shifting. Moderate correlations with quiet-standing measures supported the additional information provided by a voluntary task. Full texts were not retrieved, and overlap between these closely related Dutch cohorts is not resolved here. They should not be treated as independent replications solely because they appeared in separate papers. Together, they strengthen the case for measuring both task capacity and strategy while avoiding a universal requirement for symmetrical normalisation. [45]

Longitudinal clinical balance trajectories

The 2022 longitudinal balance cohort analysed 135 people initially and 93 in its longitudinal component. It described differing BBS trajectories in groups labelled mild and moderate based on multivariable baseline clustering. Improvement was greatest during the first three months; many participants subsequently showed greater measured impairment at one year than at three months. Age and cognition were associated with the balance trajectory. The study is useful for showing that a discharge gain does not guarantee a stable long-term course. [46]

Its machine-learning analysis, however, classified whether BBS was below 45 at any assessment. Despite language connecting that category to falls risk, the target was a BBS-defined category rather than prospectively observed falls. Very high classification accuracy should therefore not be reported as 98% accuracy for predicting falls. Baseline and future scale-related information, cohort size, repeated assessments, and the definition of the target all need scrutiny before such a model is deployed.

The study also illustrates the danger of ceiling effects. A mildly affected person beginning near the BBS maximum has little available score range for improvement. Smaller measured change can coexist with recovery in difficult stepping, confidence, dual-task activity, or endurance. When the clinical question evolves, extending assessment to a more demanding scale or task is preferable to concluding that no recovery occurred because a ceiling-bound score stayed constant.

BBS and future walking independence

The Louie and Eng cohort contained 123 rehabilitation inpatients and considered two different outcomes after four weeks: reaching a community-walking speed of at least 0.8 m/s and walking without physical assistance. The abstract reports admission BBS cutoffs of 29 for the former and 12 for initially nonambulatory patients achieving the latter, with AUCs 0.88 and 0.73 respectively. These are different prediction tasks even within one cohort. 'Community ambulation' here is a laboratory walking-speed criterion, not observed unrestricted community participation. [15]

Jenkin and colleagues analysed 68 adults unable to walk independently at rehabilitation admission. BBS was retained from a suite of clinical assessments, with odds ratio 1.23 per point (95% CI 1.02–1.49); a cutoff of at least 14 had sensitivity 0.73, specificity 0.89, and AUC 0.81 (0.71–0.92) for independent walking at discharge. Available public methods identify opportunistic sampling and data-driven predictor selection; the whole article has not been retrieved. Differences from other cutoffs can reflect case mix, outcome definition, and length of stay rather than disagreement about the instrument. [47]

These studies support using admission balance as one prognostic indicator and as a basis for discussing likely support needs. They do not justify telling a patient below a cutoff that independent walking will not return. Thresholds maximise a statistical criterion in the development sample; they do not define a biological boundary. Reassessment, motor recovery, cognition, sensory status, and rehabilitation context can materially change the prognosis.

Isolated stance and transition in walking capacity

Medina-Mirapeix and colleagues followed 109 patients recruited within four months after stroke, stratified by initial ambulation class. Public original results describe 55 transitions to a higher class and eight deaths/losses before transition or discharge. Among 44 initial nonambulators, semi-tandem stance had AUC 0.850 (95% CI 0.66–1.00) for three-month transition, falling to 0.726 (0.54–0.90) at discharge. Cox analyses adjusted for age and time since stroke. The internally selected one-second threshold is exploratory; the full stance administration and assistance rules were not retrieved. Performance varied by baseline ambulation class, so the result cannot be generalised to all stroke survivors. These are capacity-category transitions, not observed community participation or future falls. Possible overlap with related Jerez chair-rise cohorts remains unconfirmed. [48]

PASS sitting context and early nonambulatory prognosis

Huang and colleagues examined 341 initially nonambulatory patients and reported useful PASS discrimination for walking at discharge. However, the official PDF has a consequential reporting conflict: narrative and abstract report 246 ambulatory outcomes, while tables label or total 95. The total-PASS AUC is reported as 0.884 and cutoff 12.5, but the prevalence and group-dependent predictive values cannot be accepted uncritically. This paper supports further evaluation of PASS prognosis; its tabulated rule should not be implemented without clarification. 'Static PASS' also includes sitting and supported-standing items, rather than isolated unsupported stance. [18]

Tyson and colleagues' Brunel Balance Assessment work followed an early stroke cohort into three-month function. The verified abstract describes 102 patients initially and 75 with follow-up; balance disability predicted subsequent activities of daily living and mobility alongside other variables. Because the BBA spans sitting, standing, and stepping ability, it is relevant to stratified early assessment, not proof that a single timed stance forecasts independence. Loss to follow-up and anterior-circulation/weakness-based selection limit generalisation. [49]

When a patient cannot safely stand, sitting and trunk measures may provide useful early information while standing capability is reassessed over time. The SRRR3 recommendation distinguishes sitting balance assessment from standing balance. For software, that separation should remain explicit: sitting measurements can contextualise prognosis in severe stroke without being merged into an unexplained standing score. [1]

Independence discharge destination and broader outcomes

An observational decision-tree analysis of 289 patients used admission motor and cognitive assessments to classify modified Barthel outcome at discharge. Sit-to-stand ability and BBS appeared in motor branches. The reported accuracy was high, but variable selection, cutpoints, and the single-service outcome require cautious interpretation; length of stay appears among some cognitive branches, which complicates use as an admission-only prognostic tool. The target was discharge ADL status, not independent standing, future falls, or a guaranteed treatment response. [50]

Discharge destination and length of stay are additionally affected by caregiver availability, housing, service organisation, funding, and local discharge practice. A balance association in one hospital cannot be transported into another setting without checking those structural differences. A person with similar physical ability may return home with support in one service and enter residential care in another. A numerical 'home discharge probability' should therefore not be generated from BBS alone on the basis of a historical association.

The selected standing-balance evidence does not establish a broadly validated static-sway model for long-term institutionalisation, mortality, or quality of life. Those outcomes should not be inferred from findings about falls or walking capacity. Broader disability and participation measures may be appropriate alongside balance, but they need endpoint-specific longitudinal evidence. Absence of a validated model in this review is not evidence that balance has no relationship with those outcomes; it is a limit on the claims the current evidence can support.

A newer time to walking model and why it remains provisional

A 2026 multicentre prospective report analysed 168 patients and modelled time to FAC 5 between six and 24 weeks after stroke. Its abstract identifies prespecified admission predictors including BBS, knee extension, age, cognition, visuospatial function, and continence, with multiple imputation and bootstrap internal validation. Reported time-dependent AUCs were 0.86–0.92 and calibration slope 0.92. These features are encouraging because they address time and calibration rather than only one optimised cutoff. External validation is explicitly still needed, and the full model was not retrieved. [51]

This is the appropriate direction for future prognostic development: define the population and starting point, handle incomplete follow-up and predictors transparently, preserve continuous information where possible, examine calibration, and test transportability. It is still a multivariable walking model. Its reported performance does not isolate the contribution of standing balance or show that adding a sway sensor would improve the model.

Choosing a practical standing assessment

Begin with the decision then select the task

The starting question should be explicit. Is the clinician documenting assistance needs, detecting change in functional balance, investigating asymmetric loading, examining sensory dependence, or estimating a specified future outcome? A single protocol cannot be assumed to optimise all five. Choosing a test because it produces an attractive graph before deciding what the graph must mean reverses the measurement process.

For an early nonambulatory patient, prioritise safe clinical assessment spanning sitting, transitions, and supported/unsupported standing as feasible. Record the highest safely achieved condition and the assistance required. PASS is often informative at this end of the spectrum; a dedicated trunk/sitting assessment can provide context. Do not insist on a platform trial merely to create a complete dashboard. The absence of a valid unsupported-standing signal should remain visible as a clinically meaningful limitation.

For a patient who stands independently but has substantial functional limitations, BBS and a specified standing task can offer complementary information. If quantitative loading or sway will affect interpretation, collect it under a reproducible protocol. The reason to use dual plates is the limb-specific question, not a presumption that more sensors make a better outcome measure.

For an independently ambulant person near the BBS ceiling, Mini-BESTest or selected sufficiently challenging standing and stepping tasks can preserve measurement range. Difficult conditions should be chosen because they address an identified question, with appropriate guarding. An eyes-closed-on-foam task that is almost universally uncompleted in a severe cohort, or an eyes-open wide-base task that everyone completes easily, may contribute little discrimination despite being standardised.

A reproducible quiet standing protocol

There is no single stroke protocol proven optimal for every metric and severity. A defensible service protocol should nevertheless standardise a small number of fundamentals: visual condition, support surface, foot coordinates, arm position, footwear and orthosis, instructions, recording duration, repetitions, rest, and safety/termination rules. Preserve enough information to reproduce the test rather than merely label it 'static balance'.

SRRR3 prioritises limb-specific assessment using dual force plates and eyes-open/eyes-closed standing of at least 30 seconds per condition, with three-trial averages, while acknowledging the lack of a definitive biomechanical core set. This is a sensible standardisation direction where feasible. It is not permission to apply a three-trial-average MDC to one trial, or a 51.2-second barefoot protocol's threshold to a 30-second shod protocol. [1, 6, 7, 29, 31]

Foot position has a large conceptual effect because it changes the support geometry. Fixed placement aids comparison, but the chosen width and toe angle may be unattainable or atypical for someone with severe impairment, contracture, or orthotic needs. A comfortable-stance alternative can be clinically useful if precisely recorded. It should remain a separately labelled protocol, not quietly merged with a fixed-position reference.

Instructions also change the task. 'Stand naturally', 'stand as still as possible', 'load both feet equally', and 'watch this feedback target' are different conditions. Equal-loading instructions and real-time feedback convert an observation of spontaneous strategy into a performance task. Both may be useful, but they answer different questions. In particular, spontaneous asymmetry and ability to correct asymmetry on request should be stored separately.

Before the recording, allow a reproducible settling period and define when the analysed window starts. Retain trial-level duration and quality flags. A step, rail contact, examiner assistance, conversation, or device artefact may invalidate the intended quiet-standing condition. The event should be recorded, not simply filtered out until a cleaner-looking trajectory remains. The number of attempts and the reason for an excluded trial are important evidence about feasibility.

Table 6 Practical standing protocol record

This is an implementation recommendation, not a new validated stroke score.

Assistance shoes orthoses and safety

A nearby examiner without physical contact, light touch, active hand support, manual assistance, and a harness providing load support are not equivalent conditions. A no-contact safety harness can differ from one bearing weight. An aid may provide mechanical stabilisation, alter sensory information, change loading, and permit a task that was otherwise impossible. The result is still interpretable if the support condition is explicit, but it must not be compared indiscriminately with unsupported data.

Footwear and ankle–foot orthoses affect the tested system. This is not a reason to remove a clinically necessary orthosis for a nominally purer measurement. It is a reason to specify whether the question concerns usual functional performance or a particular impairment under a controlled condition. Repeat visits should match the intended condition whenever possible, and a changed orthosis or footwear should create a visible protocol-change marker.

Medical stability, orthostatic symptoms, pain, fatigue, comprehension, and recent falls should be considered before challenging stance. A sudden deterioration should prompt clinical assessment rather than automatic progression to a harder balance condition. The report's protocol discussion is for supervised professional measurement, not a directive to challenge unsafe standing at home.

Cognition aphasia neglect and lesion heterogeneity

Many studies excluded people unable to follow instructions, with substantial cognitive impairment, severe aphasia/dysphasia, or severe neglect. Some excluded cerebellar strokes or specific brainstem lesions. These exclusions improve experimental feasibility but restrict the population to which reliability and thresholds apply. An instrument that performs well in cognitively intact ambulant survivors is not thereby validated for a patient with severe neglect and impaired awareness.

Language difficulty should be distinguished from inability to understand the task. A person with expressive aphasia may perform a demonstrated motor task successfully, whereas a verbally fluent person may have impaired attention or insight. Record the communication adaptation and whether it departs from a scale's standard administration. If nonstandard cueing materially changes the task, interpret the result as adapted performance rather than silently treating it as a normative score.

A longitudinal study of 53 postacute patients found that neglect remained associated with BBS after accounting for lower-limb paresis, while its association with FAC did not. This abstract-verified result supports assessing neglect as a relevant covariate, without treating it as the sole explanation for mobility prognosis. [52]

Neglect and lateropulsion deserve dedicated recording because they can influence visual orientation, spatial reference, loading direction, and protective responses. Side of brain lesion, side of motor impairment, and side of the task are different fields. Bilateral or poorly lateralised impairments should not be forced into a simple left-versus-right model. Stroke subtype and phase should remain available for interpretation even if the primary graph is intentionally simple.

Processing and metric selection

Retain device model, calibration procedure, sampling frequency, filtering, coordinate system, analysis-window duration, and metric formulas. COP path length, mean speed, RMS displacement, confidence-ellipse area, spectral parameters, and entropy are not interchangeable measures of the same quantity. Summing AP and ML path length is not generally the same as resultant path length. Different confidence ellipses and different entropy parameters can yield different values from the same raw signal.

Sampling should be sufficient for the intended outcome, with appropriate anti-aliasing and filtering. A general balance sampling recommendation does not establish adequacy for very rapid force-development analysis. Changing the filter cutoff or smoothing algorithm can change a velocity estimate even without biological change. Processing versions should therefore be retained and, if changed, either reprocess the historical raw data consistently or clearly mark that the series is no longer directly comparable.

Start with a small prespecified set rather than dozens of exploratory variables. For a standing platform workflow, total and directional COP velocity, task completion, and clearly defined limb loading may be a reasonable initial set when their reliability is established for the protocol. Add limb-specific coordination, limits of stability, or perturbation outcomes when they answer an additional question. Advanced features belong in an exploratory layer until repeatability, validity, and incremental clinical value are demonstrated.

Implications for rehabtools

A minimum transparent assessment record

Each record should identify the patient context relevant to interpretation: days since onset; stroke type and affected side; current walking/standing assistance; relevant orthosis or aid; communication adaptations; and the selected clinical instrument/version. The task record should include eyes, surface, foot placement, arms, instructions, feedback, support, duration, trials, rest, and termination. The technical record should preserve device and software versions, acquisition/processing settings, and raw units.

The result should distinguish four states: not attempted for safety or other reason; attempted but not completed; completed with altered support/protocol; and completed validly under the target protocol. These are recommendations for data design rather than a validated new scale. They prevent clinically important information from disappearing into a generic missing-value cell and make later analysis less vulnerable to treating selected successful trials as representative of all patients.

Where a scale is used, preserve item-level information as well as the total. An unchanged total can hide an improvement in one task and deterioration in another. When an isolated item is displayed, label it as such and avoid implying that the full scale's predictive model applies to that item. For bilateral tasks, show which limb bears weight and which moves; the word 'affected' alone is insufficient.

Interpreting the result and its change

A clinically useful display can separate observed performance, uncertainty, and interpretation. For example, it can show the measured COP velocity, the actual three-trial spread, whether the protocol matches the previous visit, and a source-specific MDC or a statement that no validated matching MDC is available. The explanatory text should say that MDC reflects measurement error, not a guaranteed clinically important improvement.

If the protocol changes, do not conceal the discontinuity in a smooth trend line. A patient progressing from hand-supported to unsupported standing has improved in an important functional way even if sway increases. The appropriate display shows the change in condition and treats the unsupported series as a new comparison context. Similarly, a patient completing a previously impossible sensory condition has achieved something that a paired-completer analysis alone would miss.

Avoid a universal green/red interpretation for symmetry. Fifty–fifty loading is a geometric reference, not a validated universal safety target or proof of restored paretic control. A patient may achieve symmetry by a different strategy, and another may function safely with persistent asymmetry. Show direction, magnitude, and task context, then relate the finding to functional performance and the clinical question.

The software should not treat disagreement between BBS, Mini-BESTest, and sway as a defect. They sample different constructs and have different floors, ceilings, and measurement error. A useful report might state that functional performance improved beyond the applicable error estimate while the particular sway measure remained uncertain. That is more faithful to the evidence than choosing one number to declare the entire balance system 'better'.

Requirements before a future risk output

A future falls or recovery model should start with a locked target population, prediction time, horizon, and outcome. Examples include any inpatient fall before discharge, two or more community falls within six months, or achieving FAC 5 by 20 weeks. These are separate models, not interchangeable labels on one risk score.

Predictors should be available at the stated prediction time. Information known only at discharge cannot be used in an admission model without changing the prediction question. Repeated assessments should be analysed at the patient level, with training/validation splits that prevent the same person's data from appearing in both sets. If model development uses data from a published cohort already contributing to other analyses, those reports should not be counted as independent external validation.

The comparison model should include cheap, credible predictors such as previous falls, near-falls, clinical balance, mobility/assistance, relevant cognition, and activity exposure where feasible. A sensor model must demonstrate added value over that baseline, not just good discrimination in isolation. Examine absolute calibration, subgroup performance, missing-predictor handling, and whether a decision threshold changes useful care at acceptable false-positive and false-negative costs.

External validation should first preserve the intended prediction time, outcome, and population while testing performance in independent patients and services. Wider changes, such as acute versus rehabilitation discharge or inpatient versus community use, require explicit transportability evaluation and may constitute a new intended use. Relevant variation includes assistance levels, lesion patterns, language/cognitive impairment, and service structure. Recalibration may be needed even if discrimination remains similar. A model that cannot be evaluated in the groups most often excluded from development should state that restriction prominently.

Finally, an individual prediction should support clinical judgment rather than determine access to rehabilitation or a fixed ceiling on recovery. Uncertainty is particularly important early after stroke, when change is rapid and a single baseline can be unstable. A threshold from an observational cohort is not a justified reason to deny opportunities for reassessment or rehabilitation.

What is ready now and what remains research

Ready for carefully qualified clinical use are standardised clinical-scale recording; transparent timed-task completion; repeatable, protocol-matched quantitative standing measurement; and side-specific descriptions of loading, sway, and response strategy. The exact choice depends on phase, ability, and the assessment purpose. Interpretation should combine the measure with the person's functional goals and clinical context.

Promising but not yet sufficiently established for general deployment are new weighted sensory composites, universal wearable recovery indices, camera-derived weight distribution, and stand-alone instrumented falls probabilities. These need stronger criterion agreement, repeatability over clinically relevant intervals, prospective endpoints, independent validation, and evidence of useful added information.

The most useful research agenda is therefore not to discover an ever larger set of statistically significant features. It is to determine which small set of feasible measurements changes interpretation or decisions, for which patients, at what point after stroke, and with what uncertainty. That agenda includes negative findings, inability to complete tasks, and discordance between function and physiology. Those are central results, not inconvenient exclusions.

Table 7 What a sensor result can legitimately support

Source priorities and remaining uncertainty

Some influential secondary sources remain available here only through abstracts or partial public material, especially selected BBS walking-prognosis cohorts, Mackintosh's recurrent-falls model, FallsGOT and certain metric-specific reliability papers. Their access limitations remain beside the relevant claims. The priority BESTest and 2026 BBS/Mini-BESTest review originals have been fully examined; remaining uncertainty includes study reporting and transportability, not simply access.

Those gaps primarily limit precise replication of protocols, predictor scaling, model diagnostics, and selected subgroup thresholds. They do not justify filling in missing values from another disease or an adjacent test. Parkinson’s disease cutoffs, general older-adult thresholds, and healthy-device validation are not substitutes for stroke-specific evidence.

The conclusions most securely supported by directly examined original sources are the importance of phase and ability selection; the gap between high ICC and individual change sensitivity; the distinction between whole scales and standing mechanisms; the lack of independent prospective falls associations for Bower's quiet-standing velocity measures; the limited responsiveness of the tested wearable sway feature; and the inadequate individual agreement of the evaluated markerless weight-distribution method. These findings support a practical, transparent measurement tool while placing clear boundaries around automated prognosis.

Abbreviations

ABC: Activities-specific Balance Confidence scale; ADL: activities of daily living; AP: anteroposterior; AUC: area under the receiver-operating-characteristic curve; BBA: Brunel Balance Assessment; BBS: Berg Balance Scale; BESTest: Balance Evaluation Systems Test; CI: confidence interval; COM: centre of mass; COP: centre of pressure; FAC: Functional Ambulation Categories; FIM: Functional Independence Measure; FMA: Fugl-Meyer Assessment; GRC: Global Rating of Change; ICC: intraclass correlation coefficient; IMU: inertial measurement unit; MCID/MIC: minimal clinically important difference/minimal important change; mCTSIB: modified Clinical Test of Sensory Interaction on Balance; MDC: minimal detectable change; ML: mediolateral; MMSE: Mini-Mental State Examination; PASS: Postural Assessment Scale for Stroke Patients; RMS: root mean square; SEM: standard error of measurement; SOT: Sensory Organization Test; SRD: smallest real difference; SRM: standardised response mean; SRRR3: third Stroke Recovery and Rehabilitation Roundtable; TUG: Timed Up and Go.

Appendix reproducible searches and source priorities

Search date: 2 October 2026. PubMed query, executed in native syntax:

(stroke[Title/Abstract] OR poststroke[Title/Abstract] OR "post-stroke"[Title/Abstract]) AND (balance[Title/Abstract] OR postural[Title/Abstract] OR stance[Title/Abstract]) AND (reliability[Title/Abstract] OR validity[Title/Abstract] OR responsiveness[Title/Abstract] OR prognos*[Title/Abstract] OR predict*[Title/Abstract] OR "minimal detectable"[Title/Abstract]) AND ("Berg"[Title/Abstract] OR "BESTest"[Title/Abstract] OR "Postural Assessment Scale"[Title/Abstract] OR "standing"[Title/Abstract] OR "force platform"[Title/Abstract] OR "force plate"[Title/Abstract] OR "sway"[Title/Abstract])

Reported result: 601 records. Connector retrieval: 13 pages, 601 rows, but 402 unique PMIDs. Reconciliation: official NCBI ESearch returned 601 identifiers for the same query; EFetch retrieved 599 journal records and two book chapters, accounting for all identifiers. The book chapters were contextual material, not original stroke studies. This audit distinguishes a terminal page from an actually reconciled identifier set.

Scopus query, executed in native syntax:

TITLE-ABS-KEY((stroke OR poststroke OR "post-stroke") AND ("standing balance" OR "postural sway" OR "force platform" OR "force plate" OR "single-leg stance" OR "quiet standing") AND (reliab* OR valid* OR predict* OR prognos* OR responsiv* OR "measurement error")) AND PUBYEAR < 2027

Reported result: 219 records. Nine pages returned 219 rows and 219 unique provider identifiers. These are search-result counts, not the number of studies included in this narrative review. Cross-database duplicates and multiple reports of the same cohort should not be counted as independent patients or independent validations.

Citation chasing used backward references from the quantitative standing review and forward citations from the original PASS and Bower prospective-falls papers, through OpenAlex and Semantic Scholar, requesting up to 100 records per source for each seed. These bounded network results were used for targeted discovery; they do not establish exhaustive citation coverage. Further targeted searches addressed missing full texts, newer sensor and markerless studies, and current consensus. No claim is made that all eligible studies in all databases were included.

Source accessibility and remaining limits

The priority original-source upgrades are complete: the Flansbjer accepted manuscript, Mansfield reactive-balance publisher HTML, original PASS HTML with embedded images, Hayashi acute BBS/walking-speed MCID HTML, Tamura early-subacute Mini-BESTest MCID HTML, and Chinsongkram BESTest and Kobayashi review PDFs were examined. Their findings and limitations have been updated together in the narrative, comparison tables and reference access notes. The main Kobayashi article was available, but its separate supplementary datasets and subgroup plots were not included; unresolved main-text/table discrepancies remain explicit. [2, 9, 14, 20–22, 27]

Some secondary sources still support only abstract-level interpretation. Particularly useful future source checks would be Mackintosh's recurrent-falls cohort (DOI 10.1016/j.apmr.2006.09.004), FallsGOT (10.1177/0269215518819701), and Jagroop's unconventional COP reliability study (10.1016/j.gaitpost.2023.03.021). Their exact protocol, adjustment or metric-specific precision has not been inferred where it was not available. These residual limits do not change the report's central recommendations, but they rule out implementing unverified numerical details. [31, 37, 39]

Appendix primary study comparisons

The following comparisons preserve the distinction between measurement properties, current-status discrimination, longitudinal recovery, and future outcomes. Full-source access is identified separately from abstract or partial access. In particular, a retrospective faller classifier is not a prospective falls model, and a healthy-device study is not stroke validation.

Clinical scale measurement studies

Table 8 Clinical scale measurement studies primary study comparison

Original studies are compared by their actual target and design. Contextual healthy-device work is explicitly labelled. Cohort overlap is flagged where known or unresolved. These rows support critical comparison rather than a pooled effect estimate.

Force platform measurement studies

Table 9 Force platform measurement studies primary study comparison

Original studies are compared by their actual target and design. Contextual healthy-device work is explicitly labelled. Cohort overlap is flagged where known or unresolved. These rows support critical comparison rather than a pooled effect estimate.

Sensor video and reactive assessment studies

Table 10 Sensor video and reactive assessment studies primary study comparison

Original studies are compared by their actual target and design. Contextual healthy-device work is explicitly labelled. Cohort overlap is flagged where known or unresolved. These rows support critical comparison rather than a pooled effect estimate.

Studies of actual falls and prior fall status

Table 11 Studies of actual falls and prior fall status primary study comparison

Original studies are compared by their actual target and design. Contextual healthy-device work is explicitly labelled. Cohort overlap is flagged where known or unresolved. These rows support critical comparison rather than a pooled effect estimate.

Walking and independence prognostic studies

Table 12 Walking and independence prognostic studies primary study comparison

Original studies are compared by their actual target and design. Contextual healthy-device work is explicitly labelled. Cohort overlap is flagged where known or unresolved. These rows support critical comparison rather than a pooled effect estimate.

Recovery severe impairment and contextual studies

Table 13 Recovery severe impairment and contextual studies primary study comparison

Original studies are compared by their actual target and design. Contextual healthy-device work is explicitly labelled. Cohort overlap is flagged where known or unresolved. These rows support critical comparison rather than a pooled effect estimate.

References

References are numbered in first citation order. Study specific source descriptions identify the material examined and do not constitute a study quality rating. Links identify the original publication or the explicitly named primary source version.

1. Van Criekinge, Tamaya, Heremans, Charlotte, Burridge, Jane, Deutsch, Judith E, Hammerbeck, Ulrike, Hollands, Kristen, et al. Standardized measurement of balance and mobility post-stroke: Consensus-based core recommendations from the third Stroke Recovery and Rehabilitation Roundtable. International journal of stroke : official journal of the International Stroke Society. 2024. DOI 10.1177/17474930231205207 Source examined: Consensus full text verified; co-publication 10.1177/15459683231209154 counted once.

Source note: SRC-ed2e0d6af8e3 Van Criekinge 2024

2. C Benaim, D A Pérennou, J Villy, M Rousseaux, J Y Pelissier. Validation of a standardized assessment of postural control in stroke patients: the Postural Assessment Scale for Stroke Patients (PASS). Stroke. 1999. DOI 10.1161/01.str.30.9.1862 Source examined: Complete licensed publisher HTML examined, including scoring appendix; four embedded tables and four figures visually inspected at supplied resolution. 58 longitudinal patients plus separate 12-person reliability sample. Instrumental comparator was seated rocking-platform balance, not quiet standing. Image-only tables were read visually; no inferential numeric extraction from plotted limits.

Source note: SRC-e6e0ecad858b C Benaim 1999

3. Yi-Jing Huang, Gong-Hong Lin, Shih-Chieh Lee, Ching-Lin Hsieh. A Comparison of the Responsiveness of the Postural Assessment Scale for Stroke and the Berg Balance Scale in Patients With Severe Balance Deficits After Stroke. Journal of geriatric physical therapy (2001). 2020. DOI 10.1519/jpt.0000000000000247 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-2905fa92df9d Yi-Jing Huang 2020

4. Thitimard Winairuk, Marco Y C Pang, Vitoon Saengsirisuwan, Fay B Horak, Rumpa Boonsinsukh. Comparison of measurement properties of three shortened versions of the balance evaluation system test (BESTest) in people with subacute stroke. Journal of rehabilitation medicine. 2019. DOI 10.2340/16501977-2589 Source examined: Full article body and tables examined via PubMed Central.

Source note: SRC-f0d593a723bd Thitimard Winairuk 2019

5. Seigo Inoue, Hideyuki Takagi, Emiko Tan, Chisato Oyama, Eri Otaka, Kunitsugu Kondo, et al. Comparison of usefulness between the Mini-Balance Evaluation Systems Test and the Berg Balance Scale for measuring balance in patients with subacute stroke: a prospective cohort study. Frontiers in rehabilitation sciences. 2023. DOI 10.3389/fresc.2023.1308706 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-86d6ba34331a Seigo Inoue 2023

6. David Gasq, Marc Labrunée, David Amarantini, Philippe Dupui, Richard Montoya, Philippe Marque. Between-day reliability of centre of pressure measures for balance assessment in hemiplegic stroke patients. Journal of neuroengineering and rehabilitation. 2014. DOI 10.1186/1743-0003-11-39 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-694e3138fe03 David Gasq 2014

7. Kelly J Bower, Jennifer L McGinley, Kimberly J Miller, Ross A Clark. Instrumented static and dynamic balance assessment after stroke using Wii Balance Boards: reliability and association with clinical tests. PloS one. 2014. DOI 10.1371/journal.pone.0115282 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-e37bc164bdf9 Bower Kelly J 2014

8. Kelly Bower, Shamala Thilarajah, Yong-Hao Pua, Gavin Williams, Dawn Tan, Benjamin Mentiplay, et al. Dynamic balance and instrumented gait variables are independent predictors of falls following stroke. Journal of neuroengineering and rehabilitation. 2019. DOI 10.1186/s12984-018-0478-4 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-94f1a4507ae0 Bower K 2019

9. A Mansfield, J S Wong, W E McIlroy, L Biasin, K Brunton, M Bayley, et al. Do measures of reactive balance control predict falls in people with stroke returning to the community? Physiotherapy. 2015. DOI 10.1016/j.physio.2015.01.009 Source examined: Complete licensed publisher HTML examined, including all sections, three tables and references; not a publisher PDF. Binary any-fall analyses null after multiplicity correction; positive adjusted findings are falls-rate associations. Activity covariate measured during follow-up, not available at discharge.

Source note: SRC-378b7aac3f15 A Mansfield 2015

10. Marieke Geerars, Natasja C Wouda, Richard A W Felius, Johanna M A Visser-Meily, Martijn F Pisters, Michiel Punt. Postural Sway Measurement Using a Body-Worn Movement Sensor in Clinical Stroke Rehabilitation: Exploring Sensitivity to Change and Responsiveness. Physical therapy. 2025. DOI 10.1093/ptj/pzaf021 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-444d6bc4525b Marieke Geerars 2025

11. S Jayne Garland, Deborah A Willems, Tanya D Ivanova, Kimberly J Miller. Recovery of standing balance and functional mobility after stroke. Archives of physical medicine and rehabilitation. 2003. DOI 10.1016/j.apmr.2003.03.002 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-57c8590e8db9 S Jayne Garland 2003

12. Ahmad H Alghadir, Einas S Al-Eisa, Shahnawaz Anwer, Bibhuti Sarkar. Reliability, validity, and responsiveness of three scales for measuring balance in patients with chronic stroke. BMC neurology. 2018. DOI 10.1186/s12883-018-1146-9 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-93198acef60c Ahmad H Alghadir 2018

13. Bruyneel, Anne-Violette, Dubé, François. Best Quantitative Tools for Assessing Static and Dynamic Standing Balance after Stroke: A Systematic Review. Physiotherapy Canada. Physiotherapie Canada. 2021. DOI 10.3138/ptc-2020-0005 Source examined: Full text read on public PMC page.

Source note: SRC-51f799063cb8 Bruyneel 2021

14. Sota Kobayashi, Tomohiko Kamo, Hirofumi Ogihara, Shuntaro Tamura, Hiroki Kubo, Tatsuya Igarashi, et al. Diagnostic and prognostic accuracy of Berg Balance Scale and Mini-Balance Evaluation Systems Test in individuals with stroke: A systematic review and meta-analysis. Physiotherapy theory and practice. 2026. DOI 10.1080/09593985.2026.2685324 Source examined: Complete licensed original article PDF examined, including main tables; key table visually checked. Separate supplemental files were not supplied. Main-text/table inclusion conflicts remain unresolved; reported pooled estimates are not a reconstructed meta-analysis or a universal cutoff.

Source note: SRC-790963a08038 Sota Kobayashi 2026

15. Dennis R Louie, Janice J Eng. Berg Balance Scale score at admission can predict walking suitable for community ambulation at discharge from inpatient stroke rehabilitation. Journal of rehabilitation medicine. 2018. DOI 10.2340/16501977-2280 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-88dd903f1a1c Dennis R Louie 2018

16. Lisa Blum, Nicol Korner-Bitensky. Usefulness of the Berg Balance Scale in stroke rehabilitation: a systematic review. Physical therapy. 2008. DOI 10.2522/ptj.20070205 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-31b9ee12a9ba Lisa Blum 2008

17. Elizabeth Clark, Laura Podschun, Kelsie Church, Aaron Fleagle, Paige Hull, Samantha Ohree, et al. Use of accelerometers in determining risk of falls in individuals post-stroke: A systematic review. Clinical rehabilitation. 2023. DOI 10.1177/02692155231168303 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-3d681ff8144e Elizabeth Clark 2023

18. Yen-Chang Huang, Wei-Te Wang, Tsan-Hon Liou, Chun-De Liao, Li-Fong Lin, Shih-Wei Huang. Postural Assessment Scale for Stroke Patients Scores as a predictor of stroke patient ambulation at discharge from the rehabilitation ward. Journal of rehabilitation medicine. 2016. DOI 10.2340/16501977-2046 Source examined: Full official publisher PDF read, including methods/results/tables. Narrative/abstract and tables conflict over ambulatory outcome counts; rolling OR CI/p also inconsistent. Do not deploy PPV/prevalence or threshold without clarification.

Source note: SRC-6ba7285cb66f Yen-Chang Huang 2016

19. T J Stevenson. Detecting change in patients with stroke using the Berg Balance Scale. The Australian journal of physiotherapy. 2001. DOI 10.1016/s0004-9514(14)60296-8 Source examined: Full publisher article body and tables examined.

Source note: SRC-60fde2eaa9bd T J Stevenson 2001

20. Shuntaro Tamura, Kazuhiro Miyata, Satoshi Hasegawa, Sota Kobayashi, Kosuke Shioura, Shigeru Usuda. Pooled Minimal Clinically Important Differences of the Mini-Balance Evaluation Systems Test in Patients With Early Subacute Stroke: A Multicenter Prospective Observational Study. Physical therapy. 2024. DOI 10.1093/ptj/pzae017 Source examined: Complete licensed publisher HTML examined, including main-text tables and references; publisher PDF not retrieved. Three MCID estimates actually pooled; fourth method inapplicable. Therapist-only anchor; weaker assisted-walker correlation. No study-specific SEM/MDC: cited 3.4-point MDC is borrowed from Winairuk 2019.

Source note: SRC-39a35657d814 Shuntaro Tamura 2024

21. Butsara Chinsongkram, Nithinun Chaikeeree, Vitoon Saengsirisuwan, Nitaya Viriyatharakij, Fay B Horak, Rumpa Boonsinsukh. Reliability and validity of the Balance Evaluation Systems Test (BESTest) in people with subacute stroke. Physical therapy. 2014. DOI 10.2522/ptj.20130558 Source examined: Complete licensed original article PDF examined, including main tables; key table visually checked. Average-rating video rescoring, not patient test–retest; no original MDC. Current Fugl-Meyer motor-group classification, not prognosis. Printed Mini/BBS negative likelihood ratios conflict with corresponding sensitivity/specificity.

Source note: SRC-7b2639fb3de4 Butsara Chinsongkram 2014

22. Ulla-Britt Flansbjer, Johanna Blom, Christina Brogårdh. The reproducibility of Berg Balance Scale and the Single-leg Stance in chronic stroke and the relationship between the two tests. PM & R : the journal of injury, function, and rehabilitation. 2012. DOI 10.1016/j.pmrj.2011.11.004 Source examined: Full accepted manuscript read from official Lund University repository. Same 50 participants as 2005 gait-performance and knee-strength reliability papers; confirmed in Methods.

Source note: SRC-a7af55b8593b Ulla-Britt Flansbjer 2012

23. Perez-Cruzado, David, González-Sánchez, Manuel, Cuesta-Vargas, Antonio Ignacio. Parameterization and reliability of single-leg balance test assessed with inertial sensors in stroke survivors: a cross-sectional study. Biomedical engineering online. 2014. DOI 10.1186/1475-925x-13-127 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-fa761f6a9a65 Perez-Cruzado 2014

24. Shirley Handelzalts, Flavia Steinberg-Henn, Nachum Soroker, Michael Schwenk, Itshak Melzer. Inter-observer Reliability and Concurrent Validity of Reactive Balance Strategies after Stroke. The Israel Medical Association journal : IMAJ. 2019. PubMed. Source examined: Full publisher PDF read; no DOI identified.

Source note: SRC-d8db83d745ae Shirley Handelzalts 2019

25. Jeannine Bergmann, Carmen Krewer, Friedemann Müller, Klaus Jahn. A new cutoff score for the Burke Lateropulsion Scale improves validity in the classification of pusher behavior in subactue stroke patients. Gait & posture. 2019. DOI 10.1016/j.gaitpost.2018.12.034 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-a6954ca21ce9 Jeannine Bergmann 2019

26. Vimonwan Hiengkaew, Khanitha Jitaree, Pakaratee Chaiyawat. Minimal detectable changes of the Berg Balance Scale, Fugl-Meyer Assessment Scale, Timed "Up & Go" Test, gait speeds, and 2-minute walk test in individuals with chronic stroke with different degrees of ankle plantarflexor tone. Archives of physical medicine and rehabilitation. 2012. DOI 10.1016/j.apmr.2012.01.014 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-b9d4b4be3302 Vimonwan Hiengkaew 2012

27. Shota Hayashi, Kazuhiro Miyata, Ren Takeda, Takamitsu Iizuka, Tatsuya Igarashi, Shigeru Usuda. Minimal clinically important difference of the Berg Balance Scale and comfortable walking speed in patients with acute stroke: A multicenter, prospective, longitudinal study. Clinical rehabilitation. 2022. DOI 10.1177/02692155221108552 Source examined: Complete licensed publisher HTML examined, including all three tables and references; not a publisher PDF. Anchor-specific ROC estimates verified. Patient-GRC change-difference 8.2 in Table 3 is not reproduced by Table 2 mean changes 10.7 minus 9.5; do not implement without clarification.

Source note: SRC-ebd84c78fce9 Hayashi S 2022

28. Meizhen Huang, Marco Y C Pang. Psychometric properties of Brief-Balance Evaluation Systems Test (Brief-BESTest) in evaluating balance performance in individuals with chronic stroke. Brain and behavior. 2017. DOI 10.1002/brb3.649 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-406871daf65d Meizhen Huang 2017

29. Aryan, Raabeae, Inness, Elizabeth, Patterson, Kara K, Mochizuki, George, Mansfield, Avril. Reliability of force plate-based measures of standing balance in the sub-acute stage of post-stroke recovery. Heliyon. 2023. DOI 10.1016/j.heliyon.2023.e21046 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-753e693a72b9 Aryan 2023

30. Aryan, Raabeae, Patterson, Kara K, Inness, Elizabeth L, Mochizuki, George, Mansfield, Avril. Concurrent validity and discriminative ability of force plate measures of balance during the sub-acute stage of stroke recovery. Gait & posture. 2025. DOI 10.1016/j.gaitpost.2024.12.001 Source examined: Full publisher article body and tables examined.

Source note: SRC-a7e616c6b412 Aryan 2025

31. David Jagroop, Raabeae Aryan, Alison Schinkel-Ivy, Avril Mansfield. Reliability of unconventional centre of pressure measures of quiet standing balance in people with chronic stroke. Gait & posture. 2023. DOI 10.1016/j.gaitpost.2023.03.021 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-6bd59f378221 David Jagroop 2023

32. Iva Fiedorová, Eva Mrázková, Mariana Zádrapová, Hana Tomášková. Receiver Operating Characteristic Curve Analysis of the Somatosensory Organization Test, Berg Balance Scale, and Fall Efficacy Scale-International for Predicting Falls in Discharged Stroke Patients. International journal of environmental research and public health. 2022. DOI 10.3390/ijerph19159181 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-b5740af37b2d Iva Fiedorova 2022

33. Tatsuya Igarashi, Shota Hayashi, Shingo Hirano, Kazusa Saisu, Hironobu Kakima, Yuta Tani, et al. Development and validity of composite scores for the instrumented-modified Clinical Test of Sensory Interaction in Balance in inpatients with subacute stroke. Gait & posture. 2026. DOI 10.1016/j.gaitpost.2025.110067 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-ea26e3b26f52 Tatsuya Igarashi 2026

34. Sheehy, Lisa, Gal-Dev, Emma, Sveistrup, Heidi, Bilodeau, Martin, Finestone, Hillel. Accuracy and Variability of a Commercial Markerless Motion Capture System Compared to a Pressure Mat for Weight Distribution in Standing: Cross-Sectional Observational Study. JMIR formative research. 2025. DOI 10.2196/73575 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-dda9a452317f Sheehy 2025

35. Schröder, Jonas, Hallemans, Ann, Saeys, Wim, Yperzeele, Laetitia, Kwakkel, Gert, Truijen, Steven. Is a portable pressure plate an alternative to force plates for measuring postural stability and interlimb coordination of quiet standing balance control? Journal of rehabilitation and assistive technologies engineering. 2024. DOI 10.1177/20556683241234858 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-aa06f7e8b10a Schroder 2024

36. Julie K Tilson, Samuel S Wu, Steven Y Cen, Qiushi Feng, Dorian R Rose, Andrea L Behrman, et al. Characterizing and identifying risk for falls in the LEAPS study: a randomized clinical trial of interventions to improve walking poststroke. Stroke. 2012. DOI 10.1161/strokeaha.111.636258 Source examined: Full article body and tables examined via PubMed Central.

Source note: SRC-665c4514879f Julie K Tilson 2012

37. Shylie F Mackintosh, Keith D Hill, Karen J Dodd, Patricia A Goldie, Elsie G Culham. Balance score and a history of falls in hospital predict recurrent falls in the 6 months following stroke rehabilitation. Archives of physical medicine and rehabilitation. 2006. DOI 10.1016/j.apmr.2006.09.004 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-42951cc0ea94 Shylie F Mackintosh 2006

38. A Ashburn, D Hyndman, R Pickering, L Yardley, S Harris. Predicting people with stroke at risk of falls. Age and ageing. 2008. DOI 10.1093/ageing/afn066 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-a30397b9c5e9 A Ashburn 2008

39. Carina M Samuelsson, Per-Olof Hansson, Carina U Persson. Early prediction of falls after stroke: a 12-month follow-up of 490 patients in The Fall Study of Gothenburg (FallsGOT). Clinical rehabilitation. 2019. DOI 10.1177/0269215518819701 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-1f2a2780152c Carina M Samuelsson 2019

40. Carina U Persson, Per-Olof Hansson, Katharina S Sunnerhagen. Clinical tests performed in acute stroke identify the risk of falling during the first year: postural stroke study in Gothenburg (POSTGOT). Journal of rehabilitation medicine. 2011. DOI 10.2340/16501977-0677 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-c729528a2eeb Carina U Persson 2011

41. Surekha Gangar, Shajicaa Sivakumaran, Ashley N Anderson, Kelsey R Shaw, Luke A Estrela, Heather Kwok, et al. Optimizing falls risk prediction for inpatient stroke rehabilitation: A secondary data analysis. Physiotherapy theory and practice. 2023. DOI 10.1080/09593985.2022.2043498 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-1f8aa4a60b51 Surekha Gangar 2023

42. Johanna Jonsdottir, Fabiola Giovanna Mestanza Mattos, Alessandro Torchio, Chiara Corrini, Davide Cattaneo. Fallers after stroke: a retrospective study to investigate the combination of postural sway measures and clinical information in faller's identification. Frontiers in neurology. 2023. DOI 10.3389/fneur.2023.1157453 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-3aab8ae4f3e3 Johanna Jonsdottir 2023

43. Lisa A Simpson, William C Miller, Janice J Eng. Effect of stroke on fall rate, location and predictors: a prospective comparison of older adults with and without stroke. PloS one. 2011. DOI 10.1371/journal.pone.0019431 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-a3af282a3237 Lisa A Simpson 2011

44. de Haart, Mirjam, Geurts, Alexander C, Huidekoper, Steven C, Fasotti, Luciano, van Limbeek, Jacques. Recovery of standing balance in postacute stroke patients: a rehabilitation cohort study. Archives of physical medicine and rehabilitation. 2004. DOI 10.1016/j.apmr.2003.05.012 Source examined: Abstract verified; full text not retrieved.

Source note: SRC-be39257f43f0 de Haart 2004

45. de Haart, Mirjam, Geurts, Alexander C, Dault, Mylène C, Nienhuis, Bart, Duysens, Jacques. Restoration of weight-shifting capacity in patients with postacute stroke: a rehabilitation cohort study. Archives of physical medicine and rehabilitation. 2005. DOI 10.1016/j.apmr.2004.10.010 Source examined: Abstract verified; full text not retrieved.

Source note: SRC-3c49022d19ae de Haart 2005

46. Dongni Buvarp, Lena Rafsten, Tamar Abzhandadze, Katharina S Sunnerhagen. A cohort study on longitudinal changes in postural balance during the first year after stroke. BMC neurology. 2022. DOI 10.1186/s12883-022-02851-7 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-7ac2e2f4102f Dongni Buvarp 2022

47. Joanna Jenkin, Stephanie Parkinson, Angela Jacques, Lay Kho, Kylie Hill. Berg Balance Scale Score as a Predictor of Independent Walking at Discharge among Adult Stroke Survivors. Physiotherapy Canada. Physiotherapie Canada. 2021. DOI 10.3138/ptc-2019-0090 Source examined: Abstract plus public PMC Methods passage read; full article body not recovered.

Source note: SRC-201528eabd42 Joanna Jenkin 2021

48. Francesc Medina-Mirapeix, M José Crisostomo, Rodrigo Martín San Agustín, M Piedad Sánchez-Martínez. Prognostic value of balance performance for improvements of community ambulation among stroke patients: a cohort study. European journal of physical and rehabilitation medicine. 2022. DOI 10.23736/s1973-9087.21.06996-3 Source examined: Abstract and original public indexed statistical-analysis/results/discussion sections read; full Methods not retrieved. Possible shared Jerez cohort with chair-rise reports; exact recruitment/person overlap unconfirmed.

Source note: SRC-56cea77586dc Medina-Mirapeix 2022

49. Sarah F Tyson, Marie Hanley, Jay Chillala, Andrea B Selley, Raymond C Tallis. The relationship between balance, disability, and recovery after stroke: predictive validity of the Brunel Balance Assessment. Neurorehabilitation and neural repair. 2007. DOI 10.1177/1545968306296966 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-0721fea48069 Sarah F Tyson 2007

50. Heegoo Kim, Chanmi Lee, Nayeong Kim, Eunhye Chung, HyeongMin Jeon, Seyoung Shin, et al. Early functional factors for predicting outcome of independence in daily living after stroke: a decision tree analysis. Journal of rehabilitation medicine. 2024. DOI 10.2340/jrm.v56.35095 Source examined: Full article body and tables examined via Europe PMC.

Source note: SRC-25e9b854eb9d Heegoo Kim 2024

51. Shuntaro Tamura, Hiroyuki Saito, Hiroki Iwamoto, Sota Kobayashi, Yoshito Takahashi, Satoshi Hasegawa, et al. Prediction model for timing of independent walking on varied surface conditions in patients with early subacute stroke: A multicenter prospective study. Physiotherapy theory and practice. 2026. DOI 10.1080/09593985.2026.2733860 Source examined: Abstract/metadata verified; full text not retrieved.

Source note: SRC-945eac2e8d81 Shuntaro Tamura 2026

52. van Nes, Ilse J W, van Kessel, Marlies E, Schils, Fanny, Fasotti, Luciano, Geurts, Alexander C H, Kwakkel, Gert. Is visuospatial hemineglect longitudinally associated with postural imbalance in the postacute phase of stroke? Neurorehabilitation and neural repair. 2009. DOI 10.1177/1545968309336148 Source examined: Abstract verified; full text not retrieved.

Source note: SRC-83b11d472ac5 van Nes 2009