Why standard knee assessment leaves gaps

A patient presents with moderate knee osteoarthritis. Their X-ray shows Kellgren–Lawrence grade II–III changes — cartilage loss, some osteophyte formation, a narrowing joint space. Their PROM score suggests meaningful pain and functional limitation. On paper, the picture looks clear enough. In the consulting room, it rarely is.

The difficulty is that two patients with identical radiographic grades can move in entirely different ways. One climbs stairs without hesitation; the other cannot complete a sit-to-stand without compensating at the hip. The X-ray cannot tell you which is which, because it was never designed to. It captures structural damage at a single moment in time — a snapshot of anatomy, not a record of how that anatomy performs under load, across a morning, or week by week as a treatment takes effect or fails to.

Patient-reported outcome measures address a different dimension: how the patient experiences their knee. That information matters, but it is subject to confounders that clinicians learn to read carefully — mood, expectations, previous consultations, the memory of a bad day rather than a representative week. PROMs tell you what the patient believes is happening; they are not the same as what is biomechanically happening.

The gap between these two sources of information — structural imaging on one side, subjective report on the other — is precisely where clinical judgement has historically done its most important and least visible work. Experienced surgeons develop an instinct for functional trajectory: whether a patient is quietly deteriorating, holding a plateau, or responding to treatment. That instinct is grounded in observation and pattern recognition accumulated over years of practice. It is not, however, easily audited, shared, or scaled.

For joint-preservation decisions — who is suitable for a biological intervention, what the appropriate timing is, and what realistic functional improvement looks like — a structural snapshot and a symptom questionnaire leave questions unanswered that matter clinically. Objective movement data has the potential to fill that space.

How markerless pose estimation works in practice

The basic components are straightforward. A standard RGB camera — a smartphone, a clinic-mounted camera, or an off-the-shelf depth sensor — records the patient performing a short sequence of functional tasks. No markers are attached to the skin. No reflective suit is required. The patient simply moves.

Behind the camera, a deep neural network analyses each video frame and estimates where the body's joints are positioned in space. Models such as VIBE (Video Inference for human Body pose and shape Estimation) reconstruct a full 3D skeletal representation from 2D footage, using spatial and temporal self-attention to track how limb segments relate to one another across time. The Microsoft Azure Kinect takes a complementary approach: its depth sensor achieves real-time skeleton tracking at 24 frames per second, trading some of VIBE's geometric richness for immediate throughput. VIBE, by contrast, processes roughly 40 seconds of computation for every 10 seconds of video — a trade-off that reflects where the technology currently sits rather than where it is heading.

The clinical tasks used to elicit meaningful data are intentionally simple: a squat and a sit-to-stand. Both are familiar movements that load the knee through its functional range and generate joint angle sequences for flexion, extension, and adjacent segments. From those sequences, principal component analysis condenses hundreds of data points per task into interpretable summary metrics — particularly movement smoothness (how consistently the joint moves across repetitions) and cumulative impulse (the overall volume of motion).

The clinical appeal is real but should be stated carefully. Replacing a £200,000 instrumented gait laboratory with a clinic camera is a meaningful reduction in barrier to access — though deployable does not mean trivially simple. Lighting conditions, camera positioning, and clothing all affect output quality, and the outputs require clinical interpretation rather than autonomous diagnostic use.

Digital biomarkers: what they measure and why they differ from PROMs

Patient-reported outcome measures and digital biomarkers are answering different questions. A PROM captures a patient's interpretation of their pain and function — it is clinically useful, but it passes through memory, expectation, and mood before it reaches the score. A digital biomarker, by contrast, is a quantitative metric derived directly from sensor or camera data: a number that describes what the joint actually did, not what the patient recalls it doing.

For knee pain, two specific metrics have emerged as sensitive in published work. Armstrong et al., reporting in Frontiers in Digital Health (2024), used a pre- and post-injection study design — giving patients a local anaesthetic to the knee and measuring movement before and after — as a methodologically sound way to isolate pain's effect on movement quality. The results showed statistically significant differences (p<0.05) in two distinct metrics: smoothness of maximum knee flexion curves during the squat task, and cumulative acceleration of elbow flexion and extension during the sit-to-stand. Smoothness reflects how consistently and fluidly the joint moves across the arc of motion; cumulative acceleration captures total compensatory load transfer through adjacent segments.

A practically relevant finding from the same work: the sit-to-stand task yielded more significant biomarkers than the squat. For clinics designing their assessment setup, that distinction matters — not all functional tasks are equally informative, and the simpler movement proved more discriminating.

The clinical value of this objectivity lies in reproducibility. The same patient, assessed on two separate dates using the same task protocol, produces comparable numerical outputs regardless of which clinician is present or how the consultation is framed. That is not a property PROMs can reliably offer. These metrics should be understood as sensitive monitoring tools — a complementary layer of evidence — rather than as standalone diagnostic instruments ready for immediate clinical-grade deployment.

What the accuracy evidence actually shows

Accuracy figures for markerless pose estimation systems vary enough that cherry-picking any single number would misrepresent the field. The honest picture is mixed — encouraging in some contexts, clearly insufficient in others.

For knee flexion range of motion, a study in healthy adults measuring computer vision-based markerless HPE against a reference goniometer found an ICC of 0.74 and a correlation of r=0.90. Those figures are clinically acceptable for longitudinal monitoring — tracking whether a patient's movement is improving over a rehabilitation course, for instance — but they fall below the ICC >0.90 threshold that is generally required before a measurement can stand alone as a diagnostic instrument. The distinction matters: monitoring and diagnosis are not the same clinical task.

More encouraging results come from a specific application in knee spasticity assessment. Video pose estimation algorithms (AlphaPose and STCFormer) achieved ICCs of 0.931 and 0.911–0.94 respectively for the normalised relaxation index during a pendulum test, with strong negative correlation to the Modified Ashworth Scale (P1: ρ=−0.747 to −0.781, p<0.01). That is a meaningfully different performance profile — though also a more constrained, controlled movement than the variable gait patterns seen in routine OA assessment.

A 2025 cross-sectional clinical evaluation of a commercial markerless motion capture system against manual goniometry tells a more sobering story. Active range-of-motion agreement across joints reached only fair levels (kappa 0.32–0.37), ICCs ranged from 0.121 to 0.960 depending on the joint and task, and roughly 40% of absolute measurement differences fell within ±5 degrees. The one positive anchor: active knee extension was the only joint where the difference from goniometry did not reach statistical significance — a finding that offers a starting point for clinically reliable deployment rather than a general endorsement.

Two-dimensional systems show reduced precision for movements outside the sagittal plane — hip rotation and ankle mechanics are particularly vulnerable. Environmental variables compound this: poor lighting, loose or baggy clothing, and unfavourable camera angles all degrade output quality in ways that small validation studies under controlled conditions may not fully reflect.

Proof-of-concept studies and small observational cohorts have established that these tools can detect clinically meaningful differences in movement. That is not the same as demonstrating that they do so reliably across the heterogeneous patients, settings, and clinicians encountered in real-world practice. Large prospective studies validating pose-derived metrics as primary endpoints against gold-standard gait laboratory data are absent from the current literature — a gap that should temper the enthusiasm these early results reasonably generate.

From movement data to joint-preservation decisions

Joint preservation is not an ideology — it is a clinical calculation that depends on knowing where a patient currently sits on the trajectory of their disease. Is this knee stable, declining, or responding to treatment? Subjective pain scores shift with sleep, mood, and the framing of a consultation. They cannot reliably answer that question across months of follow-up.

Objective motion metrics can. Measured repeatedly on the same task protocol, they produce a longitudinal signal: the joint's movement quality at a given point in time, expressed as numbers that do not vary with how the patient felt about the morning. That is precisely the kind of data that allows a consultant to time an intervention with confidence — not acting too early, when a conservative approach may yet succeed, and not waiting too long, when the window for meaningful joint preservation has already closed.

For staging, Kaya et al. (2025) demonstrated that combining markerless gait kinematics with deep learning can classify knee osteoarthritis severity without imaging. If that finding holds in larger prospective cohorts, it offers a non-invasive grading pathway — one that could help select patients for the right intervention tier before committing to an MRI queue or an arthroscopic assessment.

Treatment response is the other dimension where this matters. If a patient receives an injection and their motion biomarkers show no improvement across subsequent assessments, that is a quantifiable non-response — a clearer basis for moving to the next intervention than 'the pain feels about the same.' The work published by Armstrong et al. in the Journal of Arthritis (2022) and subsequently supported by Innovate UK funding of £872,000 at MSK Doctors explored exactly this application: not replacing clinical judgement but giving it a more reliable informational foundation in knee OA detection and treatment monitoring.

The mechanism by which AI motion capture supports joint preservation is, therefore, indirect but consequential. Better data enables sharper selection, more confident timing, and earlier recognition of non-response — each of which reduces the likelihood that a patient reaches arthroplasty before every appropriate alternative has been properly tested.

Honest limits and the road to routine clinical use

Three gaps stand between the evidence reviewed here and routine adoption in consultant-led MSK care — and naming them precisely is more useful than a generic call for further research.

The first is metric standardisation. VIBE and Azure Kinect extract joint trajectories differently; a smoothness score derived from one platform is not directly comparable to the same label from another. Without an agreed minimum metric set validated across systems, multi-site studies and cross-institution audits cannot be meaningfully pooled. That is a solvable problem — but it requires the field to reach consensus rather than each research group publishing its own specification.

The second gap is regulatory-grade evidence. No large prospective trial has yet pre-specified a pose-derived digital biomarker as a primary endpoint. The injection-validated work published in Frontiers Digital Health (2024) is a necessary predecessor to that step, not a substitute for it. Commissioners and regulators — who have flagged standardisation and early regulatory engagement as preconditions since at least 2017 — will require pre-specified primary endpoint data before these metrics can influence formal treatment pathways or reimbursement decisions.

Third, clinical workflow integration remains unsolved. A busy orthopaedic clinic cannot absorb a video capture step without a clear operational model: who runs the system, what it adds to consultation time, and who interprets the output. Computational throughput compounds this — processing 10 seconds of video currently takes approximately 40 seconds on a high-specification GPU, which is incompatible with real-time clinical use on standard hardware.

None of this reflects a fundamental flaw in the underlying approach. The Innovate UK-funded programme at MSK Doctors — spanning motion AI and MRI intelligence grants — represents exactly the kind of structured investment aimed at these translational problems. The trajectory from proof-of-concept to clinical tool is credible; the honest position is simply that arrival has not yet happened.

1. [1] Reliability and validity of computer vision‐based markerless human pose estimation for measuring hip and knee range of motion. (2025). https://doi.org/10.1049/htl2.70002 https://doi.org/10.1049/htl2.70002
2. [2] A Feasible Method for Evaluating Post-Stroke Knee Spasticity: Pose-Estimation-Assisted Pendulum Test. (2025). https://doi.org/10.3390/life15111760 https://doi.org/10.3390/life15111760