A single smartphone walk-and-turn video measures Parkinsonian turning severity at clinician-level agreement (QWK 0.743) and can replace episodic in-clinic scoring for longitudinal tracking
Turning deficits and freezing of gait are primary drivers of falls in Parkinson's disease, yet the clinical gold standard, MDS-UPDRS Part III item 3.10, is an episodic 0-4 visual score assigned in clinic by a human observer. It cannot capture daily motor fluctuation, subtle turning hesitation, or the difference between a good week and a bad one, and optical motion capture precise enough to do better (over USD 100,000 per system) never reaches the district clinic where most patients are actually seen. The claim: a single smartphone video of a standardized walk-and-turn task is sufficient to produce a clinically interpretable, repeatable measure of turning severity, and the same measure can track motor decline and response to exercise in pre-frail older adults.
Evidence so far, from my own pipeline (HawkEye). A vision-language backbone with a bidirectional GRU temporal head first localizes the turning bout inside the clip (F1 at tIoU 0.5 = 0.86 on PD4T), so the segment that gets scored is the segment the UPDRS item actually asks about.
Severity is not regressed directly: the model routes through clinically inspectable primitives (turn duration, step count, left/right asymmetry, foot drag, arm-swing reduction) in a concept-bottleneck structure. On the official PD4T test split (n=8) that yields a quadratic weighted kappa of 0.743 (95% CI 0.54-0.87) for item 3.10; on the external TULIP set (n=12), with distribution alignment, QWK 0.70 and Spearman 0.66. Stated boundary: one annotator produced the labels and both test sets are small, so these results show that an interpretable structure works, not that the model is clinically validated.
The claim comes out of a decade of clinical practice in rehabilitation — musculoskeletal, neurological, geriatric and home-visit — where one question repeats at every visit: did this patient actually improve since last time? Gait speed and the timed up-and-go answer part of it, but turning, the component most closely tied to falls and to fear of falling, is still judged by eye and remembered from memory. For older adults with sarcopenia and elevated fall risk, the same phone-based task provides an objective baseline and a repeatable endpoint across a 12-week resistance and balance intervention, with the outcome defined before data collection rather than chosen after it.
Falsifiable predictions, and the experiment that would refute them. (1) In a prospective cohort of 50 people with Parkinson's disease, video-derived item 3.10 scores agree with two blinded MDS-UPDRS-certified raters at ICC at or above 0.75 and QWK at or above 0.70. (2) The same scores move in the direction and magnitude of the clinician score after a levodopa challenge. (3) In 40 pre-frail older adults, video-derived turn duration and step count improve over 12 weeks of supervised resistance and balance training while an age-matched control group does not. If agreement stays below ICC 0.60, or the model cannot separate medication-on from medication-off states better than chance, the claim that smartphone video substitutes for in-clinic turning assessment is refuted, and I will report that null result rather than a point estimate. The annotation protocol and evaluation harness will be released openly; no patient video or identifier leaves the contributing institution.