Gait assessment and prognosis in knee osteoarthritis

Executive assessment

Gait assessment in knee osteoarthritis (OA) is most defensible when it answers a specified question: present walking capacity, reproducible change, movement strategy, real-world walking, or a future outcome. These questions need different evidence. A timed walk can document walking capacity without diagnosing the cause of a slow gait. A camera can reproduce a joint angle without establishing that the angle predicts cartilage loss. A biomechanical association with subsequent knee replacement is not a validated decision rule for surgery.

For routine clinical use, standardized short-distance walking and, where relevant, sustained walking form a useful foundation. Comfortable walking, fast walking, the 40-m fast-paced walk, and the 6-minute walk are related but not interchangeable tasks. Turning, acceleration, instructions, pain, aids and corridor length must remain visible. The OARSI recommendation supports a common measurement framework; it does not validate every test version or supply universal individual change thresholds. [1–3]

Original measurement studies support useful reproducibility but also expose important limitations. In a knee-specific community sample, the 40-m fast-paced walk had a one-week MDC90 of approximately 0.20 m/s. In an end-stage preoperative sample, its same-day SDC95 was 0.27 m/s. These are different confidence levels, intervals and populations. Several frequently repeated numbers are more problematic: one 20-m study calculated its "smallest detectable difference" from the confidence interval of a bootstrapped median difference, and a TUG paper's displayed SEM and MDC do not reconcile arithmetically. Neither should be copied uncritically into an individual responder rule. [3–7]

The commonly cited knee-OA gait-speed important-difference range of 4.1–6.9 m/min, approximately 0.068–0.115 m/s, is not a validated longitudinal patient-important change threshold. Its original OAI analysis used distribution-based and cross-sectional anchor estimates; longitudinal anchor correlations were too weak for the authors to use. This distinction is especially consequential for an app that labels treatment response. [8]

Wearable and video methods are increasingly feasible, but evidence is output-specific. Same-video rescoring, same-day retesting, comparison with another markerless system, and agreement against marker-based motion capture answer different questions. An at-home wearable study found weaker agreement between home and laboratory measurements than within-home reproducibility. A large multicentre markerless study demonstrated practical implementation and recognizable OA-associated patterns, but compared different participants across centres rather than retesting the same people at multiple sites. New 2026 studies show that good agreement for speed or sagittal motion can coexist with poor agreement for timing variables or frontal-plane range of motion. [9–16]

Prospective biomechanical evidence is strongest as evidence of association with particular structural outcomes, not as a calibrated personal risk calculator. Knee adduction moment peak and impulse are not equivalent, and medial tibiofemoral, lateral compartment and patellofemoral outcomes should not be combined. Associations also differ between non-traumatic and post-traumatic OA. A small gait-pattern study forecast later arthroplasty, but selection, model development and the meaning of the surgical endpoint restrict application. Cohort studies of gait-speed trajectories identify people more likely to experience poor mobility, while some ostensibly relevant prediction models were developed entirely in people without baseline radiographic OA. [17–25]

For rehabtools, the immediate opportunity is transparent, protocol-specific measurement with quality control and contextual change interpretation. Forecasting progression, falls or surgery requires a separate, externally validated prediction pipeline. Substitution of a smartphone estimate into a published laboratory model needs validation of both the substituted measurement and the final clinical prediction.

Scope and search approach

This report concerns adults with clinically diagnosed, symptomatic or radiographic knee OA before joint replacement. It covers clinical walking tests, instrumented gait, wearable and video measurement, reproducibility, absolute error, meaningful change, and genuinely later functional or structural outcomes. Patellofemoral and tibiofemoral disease are distinguished where the studies permit. Hip-OA mixtures, at-risk cohorts, arthroplasty recipients and healthy validation samples are included only where their restricted relevance is explicit. Treatment studies inform interpretation of measurements; comparative treatment effectiveness is not the principal question.

This critical narrative synthesis is not a registered systematic review, exhaustive screening exercise, meta-analysis or formal certainty grading. Searches completed on 2 October 2026 combined knee OA and gait or walking terms with measurement and prognostic terms. The dated PubMed query yielded 1,051 records, verified against the official NCBI records; a focused Scopus query yielded 343 unique records. A broader knee-OA review landscape, targeted searches and bounded citation chasing supplemented these sets. Counts describe retrieved records, not eligible or fully appraised studies.

Define the population and the construct first

OA diagnosis symptom state and stage

Knee OA is not usefully divided into "acute" and "chronic" in the same way as an acute neurological event. Record duration of disease or symptoms and distinguish an acute pain flare from longstanding structural disease. A person with a recent increase in pain, effusion, sleep disruption or analgesic change may walk differently despite no measurable structural progression. Conversely, substantial radiographic change can coexist with relatively stable symptoms or capacity.

Clinical ACR criteria, Kellgren–Lawrence grade at least 2, grade 1 or doubtful disease, MRI-defined cartilage lesions, self-reported physician diagnosis and a cohort "at risk" of OA are different entry definitions. Some early-OA studies include symptomatic knees with KL 0–2; they should not be represented as confirmed radiographic OA throughout. At the other extreme, patients selected for arthroplasty are a particular severe subgroup, often with restrictions imposed by trial recruitment. Findings from that subgroup may not describe community-managed mild disease. [3–6, 13, 26]

Record side, bilateral symptoms, compartment, previous ligament injury or reconstruction, alignment, current pain and body size. Calling the less painful limb "unaffected" is unsafe when OA is bilateral. The walking task is a whole-person activity, whereas a kinetic measure may represent only the selected knee. Studies selecting the worse knee and studies analysing both knees have different statistical units. Analyses of both knees should account for within-person dependence, as Chang and colleagues did with generalized estimating equations. [17, 18]

Age, obesity, cardiovascular capacity, other painful joints, neurological disease and use of a walking aid affect both performance and transportability. Several technically strong studies exclude aids, other lower-limb disease, severe obesity or substantial mobility limitation. Aily's original supplementary eligibility criteria excluded BMI at least 30 kg/m², despite obesity being common in knee OA. Such eligibility details are not minor background: they define the evidence boundary. [5, 6, 9, 10, 14]

Current Aily access update: gait-B020-S03 — The original explicitly excludes BMI ≥30 kg/m². The wider contextual claim about obesity being common in knee OA is not established by this eligibility table and has not been re-estimated here.

Capacity movement quality and everyday performance

A short clinic test estimates walking capacity under a stated instruction. A sustained walk adds endurance, repeated turning and symptom tolerance. Joint-angle and kinetic measurements describe how the person accomplished the task. A wearable recorded during daily life estimates performance within the recorded environment, conditional on wear time, algorithm validity and task detection. None directly measures participation, confidence or the importance of the achieved activity to the patient.

Self-report should therefore complement rather than be replaced by performance testing. Modest correlations with WOMAC or KOOS do not automatically invalidate a timed walk, because perceived difficulty and observed execution are not identical constructs. Equally, repeatedly failing prespecified construct or responsiveness hypotheses should not be dismissed merely because a test is popular. The appropriate response is to specify the intended construct more carefully and investigate the discrepancy. [1, 3, 4, 26, 27]

For technology development, distinguish measurement of gait speed from classification of OA, prediction of a laboratory variable, and prediction of a future clinical outcome. A regression that estimates today's 6-minute distance from today's 10-m speed is a concurrent estimation model. It is not a prognosis, even when "predicts" appears in the title. [28]

Table 1 Choose the task for the clinical question

Practical synthesis. These measures should not be collapsed into one unvalidated composite.

Clinical walking tests

Short walking tests

The OARSI minimum performance set includes a 40-m fast-paced walk, alongside a 30-second chair stand and a stair-climb task. The full recommended set also includes TUG and a 6-minute walk. The original consensus used feasibility, existing measurement evidence and scoring properties; it explicitly acknowledged gaps. The 40-m task uses four 10-m lengths, and the recommended version excludes turning time. A fast test is a capacity challenge, not an estimate of habitual community pace. [1]

A comfortable 10- or 20-m test may be preferable when the question concerns usual walking pace or comparison with a specific cohort. Preserve the exact timed distance and start rule. A moving-start middle 10 m within a 15-m course differs from a standing-start 20 m. Do not silently convert time to speed using the total course length when only the central segment was timed. For repeated testing, retain the same instructions, timing landmarks, footwear, aid, practice and aggregation rule.

The 30-second fast-paced walk offers an alternative in which distance is the outcome and test duration is fixed. Its original knee-OA study included 20 women with symptomatic physician-diagnosed OA and 20 matched controls. One practice walk and two scored trials were used, with the greater distance retained; retesting occurred after 2–7 days. Knee-OA within-rater ICC was 0.96 and MDC95 5.67 m. Simultaneous between-rater scoring was much more precise, MDC95 0.82 m, because it primarily samples scorer disagreement rather than variation between performances. The sample excluded walking aids and did not require radiographic confirmation. This is promising direct evidence for that protocol, not evidence that the 30-second and 40-m tests are interchangeable. [6]

The same study's strong convergent correlations were calculated across the combined OA and asymptomatic sample. Pooling groups with very different function can strengthen a correlation relative to the relationship within OA alone. Its favourable known-groups performance and practical fixed-time format should therefore be distinguished from proof of longitudinal responsiveness or patient importance; neither an MIC nor prospective prognosis was established. [6]

Sustained walking and composite mobility

The 6-minute walk measures distance accomplished over a fixed duration. It combines pace, endurance, turning, rests, motivation and symptoms. Corridor length changes the number of turns. A 15-m corridor, a 30-m corridor, and a 40-m corridor should not be treated as numerically identical without direct evidence in the intended patients. A 400-m corridor walk is another distinct task, with completion time and endurance demands rather than fixed-duration distance. [1, 8, 26, 28]

In a 58-person knee-OA study, a moving-start 10-m speed and 6-minute distance measured on a 15-m corridor were strongly related, r=0.913. The in-sample regression explained 83.4% of variance and had mean absolute error 18.17 m. Participants were relatively slow, mean 0.59 m/s, and nine used walking aids. This is useful evidence that the two capacity measures overlap in this setting. It does not show that a short test captures fatigue, rests or symptoms over six minutes, and the equation has not been independently validated here. Its reported coefficient requires speed in m/s, not the raw seconds needed to walk 10 m. [28]

TUG combines rising, a short walk, turning and sitting. Its time cannot isolate gait, standing balance or transfer capacity. Record the chair, arm use, usual versus fast instruction and aid. A TUG change may arise from a faster turn or rise despite unchanged straight-line speed. Its inclusion in a mobility battery is sensible when those transitions matter; using it as a universal knee-OA falls threshold is not supported by the measurement studies discussed here. [1, 7, 9]

Reliability and absolute error

Match the error estimate to the repeated measurement

An ICC quantifies relative reproducibility within a particular sample. A broad range of severity can yield a high ICC even when individual differences are substantial. SEM, limits of agreement and MDC/SDC express different aspects of error in original units. MDC90 and MDC95 have different confidence levels. An estimate from repeated scoring of one video is not the error of asking a patient to repeat the task on a later day.

The community study by Suwit and colleagues provides directly applicable knee-OA data for a one-week interval. Fifty-five participants completed the study; 60% had bilateral OA. Tests used an OARSI protocol, an outdoor 4×10-m fast walk with turning time excluded, familiarization and the same blinded test-specific assessor at retest. Walking aids were permitted. The 40-m test ICC was 0.85, SEM 0.084 m/s and MDC90 approximately 0.20 m/s. The 95% limits of agreement were approximately −0.248 to 0.243 m/s. The community, outdoor surface and protocol should accompany those figures; they are not a universal knee-OA error band. [3]

Dobson and colleagues studied a mixed hip/knee-OA population, excluding previous replacement. Fifty-one stable participants contributed the principal analyses after eight exclusions; most returned after seven days. Within-rater reproducibility was strongest for the 40-m walk and 6-minute walk. The study carefully separated within-session independent-rater performance from one-week same-rater performance, and used a global change screen. Its extracted body did not preserve all numeric tables for this appraisal; unsupported exact MDC values are therefore not reproduced. Mixed-joint estimates should not be presented as knee-only findings. [2]

Tolk's preoperative cohort contained 85 people, with a 30-person retest subgroup assessed after 30 minutes. For the 40-m walk, ICC was 0.93, SEM 0.10 m/s and SDC95 0.27 m/s. Subsequent responsiveness was assessed across TKA and 12-month postoperative follow-up: the walk met 75% of the authors' prespecified hypotheses, although baseline construct validity did not. These results do not supply nonoperative OA responsiveness. Their restrained lesson is that reproducibility, construct validity and responsiveness require separate judgements. [4]

Table 2 Selected error estimates and restrictions

Different confidence levels, intervals, scoring rules and populations are preserved. No pooled universal knee-OA threshold is recommended.

Learning and questionable threshold derivations

Motyl and colleagues tested 15 people with KL 2–3 OA and synovitis across eight 20-m walks over two days, 8–20 days apart. Analgesics were withheld for 48 hours under the study protocol, and no walking aids were used. First-session walks were slower; later trials stabilized. Practice and averaging two scored trials are therefore important operational lessons. However, the reported asymmetric "SDD" was derived from the 2.5th and 97.5th percentiles of 200 bootstrapped medians of paired differences. A confidence interval for a median change is not an individual 95% limit of agreement or conventional MDC. The reported −1.59 to +0.15-second range should not be treated as a proven boundary for an individual patient's real change. [5]

The widely cited Alghadir TUG study included 65 people labelled KL 1–3, with 38 in the doubtful KL-1 subgroup. A practice trial and average of two scored trials were used; same-rater retesting was after two days. Overall ICCs were 0.97 within rater and 0.96 between raters, but lower in the narrower KL-1 subgroup. More importantly, Table 3 reports SEMs of 0.16 and 0.17 s with MDCs of 1.10 and 1.14 s, despite specifying MDC=1.96×√2×SEM. Those displayed values do not reconcile. The Bland–Altman intervals are also materially wider than a simplistic "about one second" reading implies. Retain the reproducibility evidence, flag the original discrepancy, and do not silently choose or recalculate an authoritative threshold. [7]

Biomechanical studies have analogous problems of interpretation. Brisson and colleagues' reliability analysis followed knee moments, strength, power and daily steps at baseline, six months and 24 months. Such long-term repeatability includes biological change, activity variation and new comorbid conditions; several participants developed hip or ankle problems and were retained. The estimates are useful descriptions of longitudinal stability under that protocol, but should not be represented as pure short-term instrument noise. KAM impulse, KAM peak and KFM peak also had different reproducibility. [29]

Meaningful change and clinical validity

Important differences are not automatically within person MICs

Gilbert and colleagues examined 2,527 OAI participants with baseline radiographic knee OA. Their reported 20-m gait-speed important difference of 4.1–6.9 m/min corresponds to approximately 0.068–0.115 m/s; the 400-m range was 2.9–6.9 m/min. The estimators included SEM, half-standard-deviation approaches, and differences between groups defined by function anchors. All correlations between two-year changes in speed and the longitudinal anchors were below 0.2, so those longitudinal estimates were not considered usable. Thus the headline range cannot establish that an individual who improves by 0.08 m/s has experienced important benefit. [8]

An MIC should state the anchor, definition of improvement or deterioration, clinical context, follow-up interval, baseline function, discrimination and uncertainty. A distribution-based threshold does not become patient-important by being labelled clinically meaningful. A minimal detectable change says something about error, not whether the patient values the difference. For clinical communication, show observed change, a matching error estimate where defensible, and the patient's own walking goals as distinct information.

Mostafaee and colleagues' physiotherapy study directly addresses OARSI responsiveness and MIC, but its original full methods and results were not retrieved in this appraisal. It remains a high-priority gap rather than a source of unqualified software thresholds. Ramalho and colleagues' 107-person study is also important: its abstract reports favourable construct validity and six-month responsiveness for the 40-m walk, but not the chair stand. Because full text was unavailable, the exact hypotheses, missingness and analytic details have not been independently appraised here. [27, 30]

Table 3 Interpretability is not one statistical property

Gilbert 2021 [8] rejected all longitudinal anchors; its headline speed range cannot serve as a longitudinal knee-OA MIC.

Severity classification is a different use case

An early symptomatic KL 0–2 case-control study of 30 patients and 30 controls found worse walking and mobility performance in the symptomatic group. The 40-m walk had AUC 0.87 and a 26.8-s cut-point for distinguishing those particular groups; the 6-minute walk and TUG also discriminated. This is known-groups evidence in a small, predominantly female, unaided sample. It is not a diagnostic rule for OA in an unselected clinic, a disease-progression boundary, or a future-falls threshold. The controls were deliberately free of knee symptoms, which makes the comparison different from discriminating OA from other causes of knee pain or slow walking. [26]

Wearables video and markerless gait

Remote administration and same video scoring

Aily and colleagues compared video-guided and face-to-face OARSI testing in 32 people with KL 2–3 OA. Both conditions occurred in a prepared university environment, with equipment supplied, marked distances, local Wi-Fi and nearby assistance. Participants were already familiar with the tests. Interrater and six-week intrarater analyses rescored the recorded performances. Therefore, excellent ICCs mainly support observer scoring consistency, not six-week biological reproducibility or unsupervised home validity. [9]

Current Aily access update: gait-B066-S01 — The source has 32 participants: 22 with KL II and 10 with KL III. OARSI tasks were compared during separate video-guided and face-to-face administrations. gait-B066-S02 — Both modes took place at UFS Car. Equipment was supplied, distances were taped, and local Wi-Fi was used. The remote examiner was in a separate soundproof room in the same building and close enough to help. gait-B066-S03 — Previous test experience is explicit. Measurements took place at the second endpoint of the parent trial. gait-B066-S04 — Interrater assessment used stored audio/video. Investigator 1 reviewed the same recordings after six weeks. Both investigators were blinded to initial scores.

The complete original HTML and supplement also reveal numerical inconsistencies. For cross-method 40-m speed the text reports SEM 0.91 m/s and MDC 2.52 m/s alongside CV 4.70%; the figure's limits of agreement are approximately −0.40 to +0.28 m/s. For same-video intrarater scoring, Table 4 reports MDC 0.19 m/s while the prose later gives 0.97. These cannot be treated as interchangeable or automatically corrected decimal errors. The useful result is feasibility of a tightly supported video workflow, with technical issues in 11/32 sessions; the suspect numerical thresholds should be withheld from implementation pending correction or raw-data verification. [9]

Current Aily access update: gait-B067-S01 — Source inconsistencies are confirmed: gait intrarater MDC is 0.19 m/s in Table 4 versus 0.97 m/s in prose. Official eligibility and technical-issue supplements were recovered separately. Current recovery does not independently certify what the report author examined previously. gait-B067-S02 — The text prints SEM 0.91 m/s, CV 4.70% and MDC 2.52 m/s. Figure 2 D prints bias −0.06 m/s and limits −0.40 to +0.28 m/s. The report faithfully describes suspect source values. Do not silently repair them or use them as implementation thresholds. gait-B067-S03 — Table 4 gives intrarater MDC 0.19 m/s; later prose gives 0.97 m/s. SEM 0.07 × 1.96 × √2 ≈0.194 supports the table arithmetic, but does not establish the correct raw-data result. gait-B067-S05 — Technical issues in 11 of 32 sessions are confirmed; 21 had none. Feasibility of a supported university workflow is an appropriate description. Withholding inconsistent thresholds is an appraisal judgment.

At home inertial sensors

The 2023 wearable study provides a more informative test of transport to the home. Twenty people with physician-diagnosed knee OA received three Opal sensors, a supplied chair, a measured walking course and a tablet; researchers guided the home session by video. Participants removed and replaced sensors between two within-home blocks 15 minutes apart. Within-home walking-speed ICC was 0.85 and MDC approximately 0.15 m/s. Agreement between home and laboratory speed was weaker, ICC 0.63, with mean speed about 1.01 versus 1.06 m/s. Measurements differed between home and laboratory conditions despite similar equipment and walking instructions. This comparison does not isolate an environmental effect from sensor placement, visit timing or other procedural differences. [10]

Operational details matter: two participants had unusable sensor data; clock drift required manual timestamp correction in multiple recordings; two homes could not accommodate the intended 7-m course. The sample was predominantly female, White and highly educated, and required the ability to walk 20 minutes without assistance. This supports a supervised, equipment-supported service model. It does not establish fully autonomous deployment across arbitrary homes, aids and digital-literacy levels. The home and laboratory conditions should remain separate longitudinal series unless a validated adjustment is available. [10]

A 2026 pilot compared foot-mounted IMUs with optical motion capture in 10 knee-OA participants and 10 healthy controls. Pooled stride-level gait-speed agreement was strong, ICC 0.98, bias 0.01 m/s and limits approximately −0.14 to +0.16 m/s; mixed-effects analysis was described to account for clustered strides. Thousands of strides improve within-sample precision but do not create thousands of independent patients. The OA spectrum, aids, obesity and slower gait still require larger patient-level validation. Exploratory relations to self-reported activity are not longitudinal prediction of deterioration or treatment response. [16]

Avoid healthy sample extrapolation

The 2024 vision-based smartphone gait study is sometimes encountered in knee-OA searches because of its rationale, but its validation sample consisted of 20 healthy adults, mean age 35.5 years. It excluded recent lower-limb pain or injury. Speed was compared with timing sensors and loading-response knee flexion with Xsens. The initial speed bias was corrected using the same participants before a second validity test. Thus the result is developmental evidence, with neither an independent calibration cohort nor knee-OA validation. [11]

The published reliability table also contains discrepancies: the narrative and table give different left-knee ICCs, and displayed SEM-to-MDD relationships do not follow the stated formula. These are further reasons not to import its MDD as a knee-OA threshold. More generally, good performance for one sagittal event angle against an inertial reference does not establish accuracy for all gait phases, frontal alignment, joint moments or pathological compensations. [11]

Markerless motion capture promising but parameter specific

The primary abstract of the 2024 Outerleys repeatability study describes ten OA patients measured at three visits averaging eight days apart, using eight cameras and typical/fast overground walking. The original full-text methods and tables were not recovered. The subsequent 2025 multicentre study provides complete body evidence for clinical implementation. It analysed 351 people with end-stage knee OA and 135 asymptomatic participants across three hospital sites, using synchronized 8–10-camera Theia3D setups and a common processing pipeline. Recognizable group differences in speed, stride length, cadence and joint motion were observed. Mean intersite waveform differences were relatively small. [12, 13]

Those intersite comparisons involved different people, not the same participants retested at each site. They combine measurement and case-mix differences. Control ages differed substantially between sites, mobility-aid users were excluded, and long skirts or dresses were excluded after visual inspection. The finding supports feasibility and harmonization, but not a universal intersite MDC or automatic causal attribution of OA-control differences to the diseased knee. It also concerns a multicamera clinical system, not a single smartphone. [13]

The 2026 markerless-versus-marker-based study illustrates why parameter-level evidence is essential. Twenty-two knee-OA participants underwent concurrent optical and markerless capture with force plates. Sagittal knee waveform differences were generally around or below 5°, yet frontal knee range-of-motion agreement during walking was poor to moderate. KAM impulse agreement was strong, while second-half KFM peaks agreed less well. Importantly, kinetics in both pipelines used measured force-platform data. The study does not validate camera-only estimation of joint moments. The same trial registration as the at-home wearable paper signals a related research programme rather than an entirely independent replication. [10, 14]

A 2026 tablet-video comparison used a 3-D markerless reference during treadmill walking in 30 people. Eligibility included preoperative, conservative-care and postoperative unicompartmental-arthroplasty contexts, without a clean nonoperative-only analysis. Speed and step length agreed better than temporal outputs. Step and stance-time ICCs were near zero, and several agreement intervals were large relative to the measured quantities. The text's reassuring interpretation should not override those parameter-specific results. In addition, its description of initial-contact normalization is inconsistent across methods, tables and results. It is developmental evidence that requires clarification, not a blanket validation of all video gait outputs in knee OA. [15]

Table 4 Technology evidence applies to particular outputs

A high correlation or ICC for one output does not validate all device outputs, clinical interpretation or prognosis.

What gait measures genuinely forecast

Structural progression

Laboratory gait kinetics measure net external moments under a defined model. KAM is a useful proxy related to medial-versus-lateral load distribution; it is not a direct measurement of cartilage stress or total joint contact force. Muscle co-contraction, alignment, body size, speed and the model affect the relationship. Peak, impulse and loading rate are different quantities. Normalization to mass, body weight×height, or an unnormalized value changes the estimand and must remain explicit.

Miyazaki's foundational six-year study related baseline KAM to radiographic progression in medial OA, and Bennell's 12-month study related medial loading to MRI change. Their original full texts were not obtained here, so they serve as historically important, abstract-verified signals rather than sources for deployable thresholds. A reported odds ratio per unit of normalized moment must not be converted into a person's probability without the original model and validation. [20, 21]

Chang and colleagues provide stronger directly examined evidence. The semiquantitative MRI analysis included 204 people and 391 knees; quantitative cartilage-thickness models used 203 people and 385 knees. Baseline moments were related to two-year outcomes, accounting for paired knees and adjusting for speed, age, sex, severity, pain and medication. KAM impulse was associated with several medial structural outcomes, including at least 5% cartilage-thickness loss; adjusted odds ratios were 2.39 (95% CI 1.28–4.48) for the medial tibial surface and 2.88 (1.66–5.00) for the central weightbearing femoral surface per 1 second × percent body weight × height of KAM impulse. Peak KAM did not perform identically, and peak KFM was unrelated to the measured medial progression outcomes. Some analysed knees had KL 0–1 despite person-level OA eligibility. Knees receiving TKA were excluded from the MRI analysis. These features qualify interpretation and demonstrate why "gait predicts progression" is too broad a summary. [17]

The small non-traumatic/post-traumatic study followed 35 participants for two years and found lower KAM measures associated with lateral-compartment cartilage loss, while medial associations were not substantial. Some EMG associations changed direction or strength by OA subtype. Multiple regional analyses, influential cases, attrition and a limited sample restrict confidence; numerical supplementary sensitivity analyses were not separately retrieved. Nevertheless, the study is an important counterweight to treating lower KAM as universally favourable across compartments and injury histories. Observational association does not establish that deliberately changing KAM will reproduce the estimated structural outcome. [18]

Patellofemoral outcomes require separate interpretation. Teng and colleagues analysed 61 people without tibiofemoral MRI OA, only 28 of whom had baseline patellofemoral OA. Ten progressed over one year. Higher second-half-stance peak KFM was associated with progression after adjustment; its impulse counterpart was not conventionally significant after adjustment. The small event count, inclusion of people without baseline PF OA and short interval preclude a general knee-OA prognostic threshold. This is evidence about a specific compartment and phase of stance, not a contradiction that can be resolved by averaging it with medial-compartment studies. [22]

Later arthroplasty is a distinct endpoint

Hatfield and colleagues measured gait in 80 people with moderate medial OA who were not surgical candidates initially. The final analysis included 54: 26 later underwent TKA and 28 remained without TKA and attended radiographic follow-up. A principal-component gait model combining knee and ankle moment features classified 74.1% correctly, with leave-one-out cross-validation giving a similar value. The model's optimized cut-point yielded sensitivity 84.6% and specificity 71.4%. [19]

This is temporally genuine prognosis, but not external validation. Screening many features, stepwise selection and cut-point optimization within a small retained cohort risk optimism. Principal components were derived using a larger gait dataset including asymptomatic people; implementation requires the exact feature basis and preprocessing. Those not receiving TKA but declining radiography were not in the final comparison. Follow-up differed between surgical and nonsurgical groups. Surgery additionally reflects symptoms, preferences, access and clinical decision-making, not just cartilage deterioration. The model should not determine referral, eligibility or a prediction that surgery is inevitable. [19]

Mobility trajectories and prediction models

The OAI gait-trajectory analysis found a small subgroup with relatively rapid four-year decline. Symptomatic radiographic OA was associated with membership in that group, with an adjusted odds ratio of 8.9 compared with neither radiographic OA nor pain. Radiographic OA without pain was not significantly associated in the same way. This is a useful warning against assuming that all people with structural OA follow one mobility course. It is a cohort-level association, not a validated individual probability from a single observed walk. [23]

The 2023 nomogram used 1,289 participants with symptomatic radiographic OA and an 80:20 random split. Reported AUCs were about 0.775 and 0.773 for identifying a slow gait-speed trajectory over eight years. The validation was internal, within the same OAI cohort. The trajectory classification incorporates baseline speed as well as subsequent measurements and may partly reflect persistently slower starting function rather than incident decline. Risk factors were dichotomized, missing predictors were filled with simple median/mode values, and calibration reporting relied partly on a nonsignificant goodness-of-fit test. These features limit transport to another clinical service. Inclusion of social variables also requires careful interpretation; they are contextual associations rather than immutable biological causes. [24]

A separate OAI-to-MOST risk-tree study is valuable because it actually used an external cohort, but it excluded baseline radiographic OA in both knees and baseline speed below 1 m/s. Its derivation and validation samples of 1,870 and 1,279 therefore concern people at risk, not a diagnosed knee-OA clinic population. The three-group tree's AUC was 0.70 in derivation and 0.67 in validation; larger trees performed somewhat better. Follow-up was ten versus seven years, and some high-risk ordering changed. It should not be described as an externally validated risk calculator for established knee OA. [25]

The Gilbert important-difference analysis, White gait trajectories, Liu nomogram and OAI derivation component of the Sharma risk trees draw from the same parent OAI cohort, albeit different subsets and outcomes. They are not four independent cohort replications. The companion strength report also examines OAI-derived RFD findings; cross-domain agreement must not be counted as an independent cohort merely because the assay changes. [8, 23–25]

Table 5 Keep prospective endpoints distinct

Structural progression, symptoms, mobility, falls and surgery require separate outcomes, horizons and validation.

Falls and broader prognosis

Slow gait, altered timing and poor mobility can be clinically relevant during a falls assessment, but the sources appraised here do not establish a transportable knee-OA gait-only falls calculator. Concurrent classification, a high ICC, or correlation with a balance score does not show prospective falls discrimination, calibration or benefit of acting on a cut-point. Falls require an explicit definition, prospective ascertainment and attention to exposure, previous falls, medication, balance, vision and environmental hazards. The companion standing-balance report examines the relevant falls evidence separately.

Practical assessment and rehabtools implementation

A minimum defensible workflow

First document the question and person: diagnosis definition, symptom duration and current flare, compartment and side, bilateral involvement, previous injury or surgery, pain before and after testing, recent treatment or medication timing, usual aid, and relevant comorbid constraints. Do not require analgesic withdrawal simply because a research protocol did; reproduce clinically appropriate conditions and record them.

Second select the task. Use a standardized comfortable or fast short-distance test for pace, a sustained walk when endurance and symptom tolerance matter, and TUG when transitions are part of the question. Record noncompletion and the reason. A person unable to finish a set-distance test does not automatically have a valid zero-speed score. Add observational movement information where it helps explain performance.

Third keep repeated conditions stable or label changes. A new walking aid, different corridor, home instead of clinic, different camera arrangement, or algorithm update can change the measurement. It may be clinically appropriate to change the aid; that improvement should be visible, rather than attributed solely to recovery of the knee.

Fourth report raw values and uncertainty. Show distance, elapsed time, derived speed, trial count and aggregation; identify the side and valid strides for instrumented outputs. Attach an MDC only when the population, task, interval and confidence level are sufficiently matched and the source is numerically coherent. Keep patient-important change separate. Do not automatically declare an individual "not improved" because a change falls just below a published point estimate.

Additional requirements for video and sensors

Retain device and algorithm version, sampling/frame rate, placement, calibration, camera distance and view, clothing, footwear, speed instruction, gait-event method, number of valid strides, turns excluded, missing-data handling and quality-control failures. Use person-level validation splits so that repeated strides or videos from one person do not appear in both training and test sets. Analyse person-level bias and limits of agreement, including heteroscedasticity, rather than relying on pooled correlations.

Validation should deliberately include slow walkers, high BMI, bilateral disease, deformity, different skin tones and clothing, aids, occlusion and real clinic/home constraints. Report failures as outcomes. A product that silently discards difficult videos may look accurate only among easy cases. Establish repeatability under genuine re-setup and between-day conditions, and test whether a device substitution preserves the intended clinical interpretation.

Before presenting a prognostic claim, define the later outcome and horizon; freeze the complete pipeline; assess calibration as well as discrimination in an independent population; compare against a simple clinical baseline; and determine whether the output usefully changes decisions. Structural progression, worsening symptoms, slower mobility, falls and TKA need separate validation. Measurement validity alone cannot supply these steps.

Evidence gaps and conclusion

Priority gaps are direct knee-OA longitudinal anchor studies with adequate stability/change definitions; full verification of recent OARSI MIC and responsiveness papers; patient-level validation and failure analysis of remote and markerless measures; reproducibility across re-setup, sites and algorithm versions; and external evaluation of clinically useful gait prognostic models. Unavailable original articles are evidence-access limitations, not proof that no relevant research exists.

The practical conclusion is positive but bounded. Gait testing can make knee-OA assessment more objective, repeatable and clinically informative. It works best as a transparent measurement of a particular task, supplemented by symptoms, goals and movement strategy. The evidence does not justify turning a convenient walk, a high technical correlation or a statistically associated joint moment into a universal change threshold or an automated prognosis.

Primary study characteristics

Primary studies supporting the narrative are grouped by measurement or prognostic question. Any consensus recommendation is explicitly identified. Population, protocol, endpoint and source access constrain interpretation. The linked bibliography identifies source-access limitations. Related publications from one cohort are not independent replications.

Table 6 Clinical measurement and interpretation

Table 7 Video wearable and biomechanical measurement

Table 8 Prospective structural surgical and mobility outcomes

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. Dobson F, Hinman RS, Roos EM, Abbott JH, Stratford P, Davis AM, et al. OARSI recommended performance-based tests to assess physical function in people diagnosed with hip or knee osteoarthritis. Osteoarthritis and cartilage. 2013;21(8):1042-52. DOI 10.1016/j.joca.2013.05.002 Source examined: Article body complete; consensus methodology, not threshold validation.

Source note: SRC-d7dd25666a3f Dobson F 2013

2. Dobson F, Hinman RS, Hall M, Marshall CJ, Sayer T, Anderson C, et al. Reliability and measurement error of the Osteoarthritis Research Society International (OARSI) recommended performance-based tests of physical function in people with hip and knee osteoarthritis. Osteoarthritis and cartilage. 2017;25(11):1792-1796. DOI 10.1016/j.joca.2017.06.006 Source examined: Article body complete; original numeric tables not retrieved.

Source note: SRC-55a59aa57129 Dobson F 2017

3. Suwit A, Rungtiwa K, Nipaporn T. Reliability and Validity of the Osteoarthritis Research Society International Minimal Core Set of Recommended Performance-Based Tests of Physical Function in Knee Osteoarthritis in Community-Dwelling Adults. The Malaysian journal of medical sciences : MJMS. 2020;27(2):77-89. DOI 10.21315/mjms2020.27.2.9 Source examined: Article body including tables complete.

Source note: SRC-81b45b905c32 Suwit A 2020

4. Tolk JJ, Janssen RPA, Prinsen CAC, Latijnhouwers DAJM, van der Steen MC, Bierma-Zeinstra SMA, et al. The OARSI core set of performance-based measures for knee osteoarthritis is reliable but not valid and responsive. Knee surgery, sports traumatology, arthroscopy : official journal of the ESSKA. 2019;27(9):2898-2909. DOI 10.1007/s00167-017-4789-y Source examined: Original article including methods, results, flow diagram and Table 2.

Source note: SRC-a6727b5507a5 Tolk JJ 2019

5. Motyl JM, Driban JB, McAdams E, Price LL, McAlindon TE. Test-retest reliability and sensitivity of the 20-meter walk test among patients with knee osteoarthritis. BMC musculoskeletal disorders. 2013;14:166. DOI 10.1186/1471-2474-14-166 Source examined: Article body including tables complete.

Source note: SRC-460d3522ed70 Motyl JM 2013

6. Hoglund LT, Folkins E, Pontiggia L, Knapp MW. The Validity, Reliability, Measurement Error, and Minimum Detectable Change of the 30-Second Fast-Paced Walk Test in Persons with Knee Osteoarthritis: A Novel Test of Short-Distance Walking Ability. ACR open rheumatology. 2019;1(5):279-286. DOI 10.1002/acr2.1040 Source examined: Article body including tables complete.

Source note: SRC-25f80c58982a Hoglund LT 2019

7. Alghadir A, Anwer S, Brismée JM. The reliability and minimal detectable change of Timed Up and Go test in individuals with grade 1-3 knee osteoarthritis. BMC musculoskeletal disorders. 2015;16:174. DOI 10.1186/s12891-015-0637-8 Source examined: Article body including tables complete; numeric inconsistencies flagged.

Source note: SRC-5986d610ec92 Alghadir A 2015

8. Gilbert AL, Song J, Cella D, Chang RW, Dunlop DD. What Is an Important Difference in Gait Speed in Adults With Knee Osteoarthritis? Arthritis care & research. 2021;73(4):559-565. DOI 10.1002/acr.24159 Source examined: Article body including tables complete.

Source note: SRC-365721962bf4 Gilbert AL 2021

9. Aily JB, da Silva AC, de Noronha M, White DK, Mattiello SM. Concurrent Validity and Reliability of Video-Based Approach to Assess Physical Function in Adults With Knee Osteoarthritis. Physical therapy. 2024;104(6). DOI 10.1093/ptj/pzae039 Source examined: Complete original article, Figure 2 and official supplement.

Source note: SRC-4a456057208e Aily JB 2024

10. Rose MJ, Neogi T, Friscia B, Torabian KA, LaValley MP, Gheller M, et al. Reliability of Wearable Sensors for Assessing Gait and Chair Stand Function at Home in People With Knee Osteoarthritis. Arthritis care & research. 2023;75(9):1939-1948. DOI 10.1002/acr.25096 Source examined: Article body including tables complete; supplement not separately retrieved.

Source note: SRC-fa1d061d5068 Rose MJ 2023

11. Leung KL, Li Z, Huang C, Huang X, Fu SN. Validity and Reliability of Gait Speed and Knee Flexion Estimated by a Novel Vision-Based Smartphone Application. Sensors (Basel, Switzerland). 2024;24(23). DOI 10.3390/s24237625 Source examined: Article body including tables complete; healthy-adult extrapolation only.

Source note: SRC-fcb1cb76d063 Leung KL 2024

12. Outerleys J, Mihic A, Keller V, Laende E, Deluzio K. Markerless motion capture provides repeatable gait outcomes in patients with knee osteoarthritis. Journal of biomechanics. 2024;168:112115. DOI 10.1016/j.jbiomech.2024.112115 Source examined: Abstract only; original full text not retrieved.

Source note: SRC-ce6c19a86196 Outerleys J 2024

13. Outerleys J, Laende E, Malek M, Civiero S, Madden K, Ruder M, et al. Clinical integration of markerless motion capture: A multicentre study of gait in knee osteoarthritis. Journal of biomechanics. 2025;192:112952. DOI 10.1016/j.jbiomech.2025.112952 Source examined: Article body complete; supplementary protocol not separately retrieved.

Source note: SRC-270ddc3874de Outerleys J 2025

14. Torres RTG, Senderling B, Kim E, Rose M, Gheller M, Neogi T, et al. Agreement between markerless and marker-based motion capture for knee kinematics and kinetics during functional activities in knee osteoarthritis. Osteoarthritis and cartilage open. 2026;8(3):100853. DOI 10.1016/j.ocarto.2026.100853 Source examined: Article body including tables complete; supplemental model comparisons not separately retrieved.

Source note: SRC-b8110b4d465a Torres RTG 2026

15. Wegner M, Molt M, Hansen C, Gellhaus F, Simon MJK, Seekamp A, et al. Gait assessment using a 2D video-based pose estimation app in comparison to a markerless motion capture system in subjects with osteoarthritis of the knee - a pilot study. Frontiers in bioengineering and biotechnology. 2026;14:1869878. DOI 10.3389/fbioe.2026.1869878 Source examined: Article body including tables complete; internal reporting inconsistencies flagged.

Source note: SRC-54c9f53b59e4 Wegner M 2026

16. Odonkor CA, Bohacek S, Hirani S, Zhang W, Muaremi A, Leutheuser H, et al. Wearable Sensor-Derived Gait Parameters Across Self-Reported Physical Activity Levels in Individuals With Knee Osteoarthritis and Healthy Controls: Pilot Cross-Sectional Validation Study. JMIR formative research. 2026;10:e80728. DOI 10.2196/80728 Source examined: Article body including tables complete; supplementary intake form not separately retrieved.

Source note: SRC-5cc1ad5ee7a1 Odonkor CA 2026

17. Chang AH, Moisio KC, Chmiel JS, Eckstein F, Guermazi A, Prasad PV, et al. External knee adduction and flexion moments during gait and medial tibiofemoral disease progression in knee osteoarthritis. Osteoarthritis and cartilage. 2015;23(7):1099-106. DOI 10.1016/j.joca.2015.02.005 Source examined: Article body including tables complete.

Source note: SRC-6049fc26a333 Chang AH 2015

18. Robbins SM, Pelletier JP, Abram F, Boily M, Antoniou J, Martineau PA, et al. Gait risk factors for disease progression differ between non-traumatic and post-traumatic knee osteoarthritis. Osteoarthritis and cartilage. 2021;29(11):1487-1497. DOI 10.1016/j.joca.2021.07.014 Source examined: Article body complete; numerical supplements not separately retrieved.

Source note: SRC-64f1571b745d Robbins SM 2021

19. Hatfield GL, Stanish WD, Hubley-Kozey CL. Three-dimensional biomechanical gait characteristics at baseline are associated with progression to total knee arthroplasty. Arthritis care & research. 2015;67(7):1004-14. DOI 10.1002/acr.22564 Source examined: Article body including tables complete.

Source note: SRC-11a60da07d3f Hatfield GL 2015

20. Miyazaki T, Wada M, Kawahara H, Sato M, Baba H, Shimada S. Dynamic load at baseline can predict radiographic disease progression in medial compartment knee osteoarthritis. Annals of the rheumatic diseases. 2002;61(7):617-22. DOI 10.1136/ard.61.7.617 Source examined: Abstract only; original scanned article not retrieved.

Source note: SRC-02981b3cd145 Miyazaki T 2002

21. Bennell KL, Bowles KA, Wang Y, Cicuttini F, Davies-Tuck M, Hinman RS. Higher dynamic medial knee load predicts greater cartilage loss over 12 months in medial knee osteoarthritis. Annals of the rheumatic diseases. 2011;70(10):1770-4. DOI 10.1136/ard.2010.147082 Source examined: Abstract only; original full text not retrieved.

Source note: SRC-7179e3d175a9 Bennell KL 2011

22. Teng HL, MacLeod TD, Link TM, Majumdar S, Souza RB. Higher Knee Flexion Moment During the Second Half of the Stance Phase of Gait Is Associated With the Progression of Osteoarthritis of the Patellofemoral Joint on Magnetic Resonance Imaging. The Journal of orthopaedic and sports physical therapy. 2015;45(9):656-64. DOI 10.2519/jospt.2015.5859 Source examined: Article body including tables complete.

Source note: SRC-d0480efccd6a Teng HL 2015

23. White DK, Niu J, Zhang Y. Is symptomatic knee osteoarthritis a risk factor for a trajectory of fast decline in gait speed? Results from a longitudinal cohort study. Arthritis care & research. 2013;65(2):187-94. DOI 10.1002/acr.21816 Source examined: Article body including tables complete.

Source note: SRC-b2def3bd5d4a White DK 2013

24. Liu P, Wang C, Chen H, Shang S. Development of a nomogram prediction model for gait speed trajectories in persons with knee osteoarthritis. Scientific reports. 2023;13(1):11291. DOI 10.1038/s41598-023-37193-y Source examined: Article body including tables complete.

Source note: SRC-898ae5079518 Liu P 2023

25. Sharma L, Kwoh K, Lee JJ, Cauley J, Jackson R, Hochberg M, et al. Development and validation of risk stratification trees for incident slow gait speed in persons at high risk for knee osteoarthritis. Annals of the rheumatic diseases. 2019;78(10):1412-1419. DOI 10.1136/annrheumdis-2019-215353 Source examined: Article body including tables complete; supplements not separately retrieved; at-risk cohort without baseline radiographic OA.

Source note: SRC-a9cba6b60b61 Sharma L 2019

26. Lee SH, Kao CC, Liang HW, Wu HT. Validity of the Osteoarthritis Research Society International (OARSI) recommended performance-based tests of physical function in individuals with symptomatic Kellgren and Lawrence grade 0-2 knee osteoarthritis. BMC musculoskeletal disorders. 2022;23(1):1040. DOI 10.1186/s12891-022-06012-2 Source examined: Article body including tables complete.

Source note: SRC-87937e420ab7 Lee SH 2022

27. Ramalho RB, Casonato NA, Montilha VB, Chaves TC, Mattiello SM, Selistre LFA. Construct Validity and Responsiveness of Performance-based Tests in Individuals With Knee Osteoarthritis. Archives of physical medicine and rehabilitation. 2024;105(10):1862-1869. DOI 10.1016/j.apmr.2024.05.024 Source examined: Abstract only; original full text not retrieved.

Source note: SRC-e79f96a4423e Ramalho RB 2024

28. Lozano-Meca J, Montilla-Herrador J, Gacto-Sánchez M. Gait speed in knee osteoarthritis: A simple 10-meter walk test predicts the distance covered in the 6-minute walk test. Musculoskeletal science & practice. 2024;72:102983. DOI 10.1016/j.msksp.2024.102983 Source examined: Article body complete; unavailable table images not transcribed.

Source note: SRC-ed2678690b7e Lozano-Meca J 2024

29. Brisson NM, Stratford PW, Maly MR. Relative and absolute test-retest reliabilities of biomechanical risk factors for knee osteoarthritis progression: benchmarks for meaningful change. Osteoarthritis and cartilage. 2018;26(2):220-226. DOI 10.1016/j.joca.2017.11.003 Source examined: Article body complete; numeric tables not retrieved.

Source note: SRC-7b73c2e20195 Brisson NM 2018

30. Mostafaee N, Rashidi F, Negahban H, Ebrahimzadeh MH. Responsiveness and minimal important changes of the OARSI core set of performance-based measures in patients with knee osteoarthritis following physiotherapy intervention. Physiotherapy theory and practice. 2024;40(5):1028-1039. DOI 10.1080/09593985.2022.2143253 Source examined: Abstract only; original full text not retrieved.

Source note: SRC-0ae5c66bba6f Mostafaee N 2024