Can 3D movement analysis automate Parkinson's motor assessment and estimate disease duration?

A six-camera system turns routine exam tasks into 3D movement features

The paper asks whether video-based movement measurement can do more than reproduce a clinician's score: can it quantify motor impairment, distinguish disease stages, and estimate how long a person has had Parkinson's disease? The authors built a markerless capture setup in a Duke movement-disorders clinic: six hardware-synchronized cameras recording at 80 frames per second in a room of roughly 6 m by 3.3 m. After installation it needed only a brief daily check rather than per-session recalibration, and recordings were made during scheduled clinic visits.

Participants performed three tasks already used in the MDS-UPDRS Part III motor exam: repeated fist making (T-hand), pronation–supination of the hands (T-prosup), and gait (T-gait). The authors chose these to sample different motor subsystems — fine distal hand control, wrist and forearm rotation with rhythmicity, and whole-body locomotion and posture. 3D body and hand poses were triangulated from multiple camera views, then automatically segmented into repeated cycles such as individual fist closures, rotations, or strides.

From each cycle the team computed an interpretable feature bank: global kinematics such as range of motion, peak and mean velocity, temporal regularity, amplitude consistency, and left–right asymmetry; Parkinson's-oriented constructs such as bradykinesia-like slowing, amplitude decrement over repetitions, and side-dominant asymmetry; task-specific measures; and sequence-level descriptors of how performance changed across repetitions. The raw bank was large — 1719 features for T-hand, 1713 for T-prosup, and 1735 for T-gait — and was pruned for collinearity and low variance down to 14–21 features per task and side.

The cohort comprised 45 patients with Parkinson's disease (22 female, 23 male) diagnosed by a movement-disorders specialist, plus 6 healthy controls (5 female, 1 male). Patients were assessed in the practically defined ON state on their usual medication. People who could not safely walk were excluded, including Hoehn & Yahr stage 5 or an MDS-UPDRS-III gait subscore of 4, so advanced disease is outside the reported results. The small, acquisition-matched control group is a key constraint on how precisely diagnostic performance can be read.

Selected features behave the way clinicians would expect

Before any classifier, the authors checked whether the retained features showed clinically plausible distributions. In T-hand, patients showed slower hand-closing speed on the tested side and greater amplitude on the non-tested side, which the authors interpret as dyskinesia. Hand-closing speed declined stepwise from healthy to early Parkinson's to post-early Parkinson's and was negatively correlated with increasing MDS-UPDRS score, while hand amplitude showed wider spread and more outliers.

In T-prosup, patients showed weaker forearm-rotation speed on the tested side, reduced palm-orientation variability, and increased leg movement during what is an upper-limb task — described as spillover dyskinesia visible in knee-swing amplitude. Rotation speed and palm-orientation variability were negatively correlated with time since diagnosis and with MDS-UPDRS scores. In T-gait, patients showed slower calf-swing velocity and shifts in mean step width.

These are distributional and correlational findings, not causal claims about what produces the clinical score. Their value is that the features are named physical quantities a clinician can inspect, which is the paper's stated alternative to black-box prediction. The authors also note that in later phases and at higher MDS-UPDRS scores the feature distributions became broader and more skewed, capturing heterogeneity among patients.

Diagnosis is nearly saturated; staging is where models differ

The paper reports two classification targets that should not be conflated. Binary diagnosis separates patients from healthy controls. Duration-defined severity sorts participants into healthy, early Parkinson's (within 24 months of diagnosis), and post-early Parkinson's (beyond 24 months). The 24-month cut is described as a pragmatic modeling choice tied to a common clinical follow-up window, not a biological boundary, which is why the authors also model time since diagnosis continuously.

Demographics alone were weak: age and sex reached 50.0% balanced accuracy for diagnosis and 33.3% for phase, with 0% F1 for early Parkinson's. Clinician MDS-UPDRS task scores were strong for diagnosis — T-gait scores reached 100% balanced accuracy, and right-side T-prosup scores 98.0% accuracy and 98.9% balanced accuracy — but weaker for phase, with balanced accuracy of 61–67% and most configurations failing to identify early Parkinson's at all (early-phase F1 = 0%).

The 3D feature models were competitive on diagnosis; right-side T-prosup features reached 98.0% accuracy and 91.6% balanced accuracy. Their distinctive contribution was phase stratification. The multi-action 3D model reached 68.4% balanced accuracy for phase and, unlike the clinician-score models, produced non-zero early-phase recovery at 16.7% F1. Fusing 3D features with demographics and clinician scores gave the best phase result: 81.3% accuracy, 70.2% balanced accuracy, and 40.0% F1 for early Parkinson's.

That fusion gain was non-additive: adding MDS-UPDRS scores to the multi-action plus demographics model raised early-phase F1 from 16.7% to 40.0%. The authors read this as evidence of non-linear interactions between kinematic and clinical information, motivating multimodal rather than kinematics-only models. Because predictions were made per repetition but aggregated to one decision per subject, and evaluated with subject-level five-fold cross-validation, these numbers describe subject-level performance rather than per-movement accuracy.

Estimating time since diagnosis is the harder and more novel target

For continuous time since diagnosis, the paper reports mean R² across five folds, with Pearson correlations and p-values. Three participants lacked time-since-diagnosis data and were excluded, leaving n = 48 for regression and n = 51 for classification. Demographics explained little variance (age R² = 0.046, p = 0.124), and clinician MDS-UPDRS baselines offered minimal explanatory power: T-gait R² = 0.089, p = 0.015, while the composite MDS-UPDRS R² was −0.121.

The explainable multi-activity 3D model reached R² = 0.223, p = 0.0009, with Pearson r = 0.46 and a 95% confidence interval of [0.27, 0.62], outperforming both demographics and MDS-UPDRS. Single-activity variants were weaker — T-gait R² = 0.058, p = 0.060; left T-prosup R² = 0.043, p = 0.045 — so combining actions materially improved fit. The multimodal model, adding demographics and clinician scores, reached R² = 0.277, p = 0.0002, Pearson r = 0.51, 95% CI [0.30, 0.66].

An R² of 0.223 means the multi-activity kinematic model explains roughly a fifth of the variance in time since diagnosis, so it is a signal rather than a precise clock. The authors also report that when predicting duration, T-hand features, including non-tested knee-swing amplitude and tested palm pitch drop, were the top contributors. Stratified analyses found diagnosis sensitivity was equally high in early and post-early disease, while duration regression remained statistically significant within the post-early group alone.

A comorbidity check supports, but does not prove, disease specificity

A plausible alternative explanation is that the duration signal reflects other conditions rather than Parkinson's kinematics. The authors tested this by comparing model performance between subjects with and without chronic pain, a comorbidity more prevalent in early than in post-early Parkinson's in this cohort. Performance was comparable across model configurations in both subgroups, and other prevalent comorbidities such as osteoarthritis had similar prevalence between the two phase groups.

This is a targeted robustness check, not a full causal isolation. It reduces concern that chronic pain drives the duration estimates, but the study remains cross-sectional: each participant contributes movement recorded at one time point, and time since diagnosis is inferred from a model trained across participants rather than from repeated measurements of the same person. The paper's framing of the framework as potentially supporting future longitudinal monitoring is therefore a forward-looking possibility, not a demonstrated result.

The results fit ambulatory, medicated, early-to-middle-stage patients

Several scope conditions shape how far the numbers travel. All patients were ambulatory and assessed in the ON state on their typical medication regimen, reflecting routine clinic evaluation; time since last levodopa dose and formal motor-fluctuation status were not recorded. Advanced disease was excluded by design. Freezing of gait was present in 10 of 45 patients (22%), and these ambulatory subjects were retained.

The comparison group was only 6 healthy controls, recruited under the same protocol and equipment, which gives acquisition-matched comparisons but limited statistical precision for diagnostic metrics. The system also required a dedicated multi-camera room and on-site consent, which the authors note limits recruitment throughput for both patients and controls in an elderly population. Data are not publicly available because they contain sensitive health information.

The clinical baseline is also narrower than the full exam: comparisons use only the three corresponding MDS-UPDRS item scores — 3.5 hand movements, 3.6 pronation–supination, and 3.10 gait — not the complete Part III. The authors state explicitly that the results are not intended as a comparison with the full motor examination.

Fusion with clinical scores, not kinematics alone, carries the staging gain

The paper's most decision-relevant comparison is not 3D features versus clinician scores on diagnosis, where both perform well, but what each contributes to staging and duration. Clinician task scores were strong diagnostically yet largely failed to flag early Parkinson's, with 0% F1 in most configurations. Multi-action 3D features produced the first non-zero early-phase signal at 16.7% F1, and adding clinician scores and demographics raised it to 40.0%.

That ordering suggests the practical value of this framework lies in combining interpretable kinematic features with existing clinical inputs rather than replacing the bedside exam. The duration regression points the same way: R² rose from 0.223 for kinematics alone to 0.277 for the multimodal model, while clinician scores alone were near or below zero.

The priority for readers evaluating this work is therefore the multimodal phase and duration results, read against a small single-center cohort of ambulatory, medicated, early-to-middle-stage patients. Whether the framework supports longitudinal monitoring, as the authors propose, remains untested here; the evidence establishes cross-sectional separation and estimation, not tracked change within individuals over time.

Sources used in this answer