Ordinary Video Aims to Measure Apraxia Beyond Observer Scores

A Cureus paper published 18 August 2026 lays out a full pipeline that would turn ordinary RGB video of standardized gestures into continuous spatial and temporal scores for ideomotor limb apraxia. Author Alireza Minagar frames the system as a measurement adjunct, not an autonomous diagnosis, and supplies both the clinical rationale and a validation design built around the one comparison that can kill the idea.

No patients were enrolled and no results are reported. The second-order stakes are larger than a cleaner bedside number: continuous, shareable features that could travel across raters, sites and telemedicine links, and that could finally give apraxia trials something better than ordinal observer scores.

What the Proposal Claims

Ideomotor limb apraxia is defined here as impaired production of learned or novel gestures under command or imitation, marked mainly by spatial and temporal execution errors that elementary weakness, sensory loss, ataxia or poor comprehension cannot fully explain. It is common after left-hemisphere stroke and appears in several neurodegenerative diseases.

Quantitative kinematic work has existed for decades. Three-dimensional computer-graphic analysis of apraxic gestures appeared as early as 1990. Later studies documented characteristic trajectory and timing abnormalities. Standardized batteries such as the TULIA upper limb apraxia test validation and its 12-item bedside screen improved reliability of human scoring. What stayed missing was a bridge that used only commodity video, required no lab motion-capture suite, and mapped recovered features onto the same error dimensions clinicians already use.

Minagar’s hypothesis is deliberately narrow. Markerless computer vision plus machine learning trained against blinded expert consensus can quantify selected spatial and temporal features, profile error dimensions at the trial level, and support detection and grading. Two independent claims sit inside it: a measurement claim that continuous features distinguish apraxic from non-apraxic production and track severity, and a criterion-validation claim that a model can approximate expert judgments of the same performances.

The claim is not that apraxia can be diagnosed automatically or that expert assessment can be replaced. The claim is narrower and correspondingly more defensible: that a markerless, video-based system can act as an objective measurement adjunct.

That sentence, from the paper itself, sets the tone for every later design choice.

Why Ordinary Video Suddenly Looks Plausible

Two technical shifts reopen the problem. Open-source pose estimators such as OpenPose and Google’s MediaPipe now recover body and hand landmarks from a single RGB camera in real time. Computer vision has already produced clinically usable upper-limb kinematic metrics with modest cameras. Machine learning routinely classifies movement patterns in rehabilitation and skill assessment.

Most recent markerless stroke work has stayed on elementary impairment: paresis, range of motion, compensation, or automated Fugl-Meyer scores. Praxis, the higher-order planning deficit, has been left largely untouched by the new tools. The paper’s novelty claim is therefore the intersection: video-only input, features mapped specifically onto an established apraxia error taxonomy, and validation against blinded consensus for bedside or remote use. That claim is presented as a probable gap that a formal literature search should confirm before any stronger assertion.

Adjacent proof-of-concept work already shows computer vision kinematic metrics on the drinking task can be extracted with acceptable accuracy in neurotypical participants. The same pipelines have begun tracking upper-limb reaching after stroke with MediaPipe. Extending them to the spatial and temporal signatures of apraxia is the untested step.

The Control Group That Decides Everything

If apraxic patients were compared only with healthy controls, a model could succeed by detecting nonspecific slowing or reduced range of motion. That would amount to recognizing left-hemisphere stroke with a weak arm rather than apraxia. The critical comparison is therefore Group A (apraxia present) versus Group B (neurologically affected patients matched for weakness, lesion burden, age and general cognitive impairment but without apraxia). Healthy controls may be added as a reference arm, never as the primary one.

Group B members may still show spasticity, intention tremor or cerebellar ataxia. Those conditions also produce movement fragmentation, extra velocity peaks and path curvature. The paper therefore states an explicit discriminating principle: content dependence. Cerebellar or pyramidal degradation should scale with mechanical demand and stay largely indifferent to symbolic content. Apraxic degradation should dissociate across elicitation conditions in the classical pattern: pantomime worse than actual tool use, command worse than imitation.

The primary signal is therefore a within-participant contrast across matched actions under different elicitation conditions, not a single kinematic number on one trial. Three secondary covariates are proposed for testing: spectral peaks that mark intention tremor, dissociation of peak velocity from path shape in spasticity, and the timing of deviation within the normalized trajectory for dysmetria versus apraxic plane or orientation errors. If none of these separate the groups once weakness is controlled, the stronger form of the hypothesis falls.

Features Mapped Onto the Error Taxonomy

The framework splits analysis into two levels. Level one addresses movement quality (how the gesture was executed). Level two addresses action content (whether the observed gesture matched the target). Kinematics support level one well and level two only partially; content errors are treated as an exploratory secondary endpoint.

Candidate features for level one are drawn from laboratory variables already shown to differ between apraxic patients and controls.

Body-part-as-object errors are deliberately kept out of the initial level-one feature vector. They depend on precisely the fine hand configuration and depth orientation that single-camera pose estimation recovers least reliably. Including them early would inject noise into the strongest recoverable signal. They re-enter only after multi-view or hand-mesh methods demonstrate acceptable stability against manual annotation.

The emphasis stays on hand-crafted, clinically motivated features rather than end-to-end deep nets. Apraxia cohorts are small; training large networks from scratch on a few tens of patients invites overfitting.

Pose Estimators and the Occlusion Problem

Acquisition standards are strict: at least 300 lux frontal diffuse light, plain contrasting background, short sleeves, fixed tripod at shoulder height 1.5-2 m away, 1920×1080 at 60 frames per second preferred, armless chair, 10-second calibration segment, full logging of device and deviations. Frame rate is treated as a design parameter because several temporal features depend on it; 30 fps is marginal for closely spaced velocity peaks and unreliable for higher derivatives.

Self-occlusion is named as the principal technical vulnerability. Gestures such as combing hair, brushing teeth or saluting bring the hand across the body or in front of the face. Single-camera accuracy for fine finger configuration and hand orientation in depth is therefore a genuine constraint that must itself be validated. A two-camera default (frontal plus 45-degree oblique) is recommended for the validation study so that the performance drop of the single-camera clinical configuration can be measured rather than assumed.

Estimator family is a prespecified experimental factor, not a settled choice. At least one lightweight convolutional model and one transformer-based model should run over identical recordings.

• MediaPipe / BlazePose, real-time on ordinary hardware, 33 landmarks, browser-friendly, permissive license.
• OpenPose, multi-person part-affinity classic; non-commercial license limits product use.
• HRNet and similar high-resolution CNNs, preserve spatial precision via parallel multi-resolution streams.
• ViTPose / ViTPose++, plain vision-transformer backbones that capture long-range dependencies; higher compute cost.

A 2026 production survey of the four main pose-tracking model families in 2026 underscores the same trade-offs: MediaPipe for on-device speed, RTMPose-class models for CPU throughput, ViTPose when GPU accuracy is the priority, and the hard limit that monocular depth estimates are not clinical-grade three-dimensional measurements. Multi-view triangulation remains the route to reliable depth.

Optional inertial measurement units on upper arm, forearm and hand, fused via Kalman or complementary filter, form a research-grade comparison arm only. Sensors would forfeit the bedside and telemedicine advantage that motivates the work. Their role is to bound how much the video-only system gives up under occlusion and lighting change.

How the Model Would Be Trained and Judged

Ground truth comes from at least three trained raters scoring every performance with a prespecified manual and a validated instrument. Interrater reliability is reported before consensus adjudication. Soft-label training retains the distribution of ratings rather than forcing a hard majority; trials with disagreement stay in both training and evaluation. Model performance is interpreted against observed panel reliability, treated as the realistic ceiling.

Given typical cohort sizes under 100 patients, the primary classifier is an interpretable model (random forest or support-vector machine) over the engineered features. A transformer over the same sequences is a secondary arm only, admitted with pretraining or transfer, strong regularization and strictly participant-level evaluation. Leave-one-subject-out cross-validation is mandatory; all splitting, selection and tuning stay inside folds. Primary metrics are the intraclass correlation between predicted and consensus severity and, for detection, area under the ROC curve.

Generalization claims are ordered by difficulty: new trials of trained gestures, new patients, unseen gestures, different elicitation conditions, different camera or site. Multisite held-out evaluation would be required before any clinical claim. The full analysis plan, including primary versus secondary arms, should be pre-registered.

Earlier kinematic analysis of aiming movements in apraxia already showed spatial deficits specific to the disorder when left-hemisphere patients without apraxia served as controls. The present design deliberately inherits that logic.

What Would Count as Success or Failure

If the hypothesis holds, continuous features recovered from video would differ between apraxic patients and the matched neurological controls, severity estimates would correlate strongly with expert consensus, and the error-dimension profile would align with clinician-assigned types for the dimensions accessible to level one. Content-error performance is expected to be weaker; that boundary is anticipated rather than spun as a surprise.

Refutation criteria are stated in advance. Failure of features to separate groups once weakness is controlled, or failure of a model to generalize to new patients, raters or recording conditions, disconfirms the stronger claim. Transformer pose estimators are predicted to stabilize distal landmarks under occlusion more than lighter models, yet that advantage may not reach classification once features are trial-averaged. Video-inertial fusion is predicted to improve feature stability; the size of the gain bounds the video-only shortfall. At realistic sample sizes the simpler interpretable classifier may outperform the attention model; that result would be reported as data, not as framework failure.

Should even the bounded form succeed, continuous measures would make severity comparable across examiners and time, support remote structured assessment, give therapists measurable feedback, and supply finer dependent variables for mechanistic studies. Diagnostic responsibility stays with the clinician; automated output remains open to override.

Limitations are listed at length: the system is unbuilt; the reference standard is imperfect expert judgment; pose estimation error is known and worse under occlusion and atypical posture; content errors largely exceed single-gesture kinematics; confounding motor and cognitive deficits require the special control group; small datasets plus high-dimensional video create overfitting risk; ideomotor apraxia is heterogeneous; good pantomime scores do not automatically predict real-world tool use. Ethical questions around video consent in aphasic patients, disagreement between clinician and score, and responsibility for erroneous automated numbers are treated as design requirements rather than afterthoughts.

Researchers already run markerless pipelines on elementary reaching and ADLs. The higher-order layer of praxis has stayed mostly outside those instruments. This framework is an explicit bet that the same cameras can be pointed at the error taxonomy clinicians already use, provided the validation refuses the easy healthy-control shortcut. The next step is not another conceptual paper. It is the first multi-rater, multi-view recording set that can tell whether the continuous numbers separate apraxia from everything that merely looks like weak movement.