Objective:

To describe a prototype AI algorithm trained to quantify movement from natural single-camera smartphone video into MDS-UPDRS motor scores.

Background:

The MDS-UPDRS motor subscale is a widely used tool in clinical and research assessments of Parkinsonian disorders. However, it requires specially trained raters and captures only predefined movements. A machine learning model could capture both standardized MDS-UPDRS part III assessments and currently unquantified metrics such as variability in joint movement and finger tap decrement rate.

Design/Methods:

Videos of individuals with Parkinson's disease (PD) performing MDS-UPDRS were ingested by the algorithm. Object detection models (YOLO real-time detection) classified gestures. Hand and body landmarks were detected by MediaPipe. These spatial features were used to calculate motor metrics for training the AI algorithm (AWS Sagemaker) using a Random Forest Classifier. An 80/20 training-validation split was used, and performance was compared with a 60/40 split. Items not filmed—rigidity and postural stability—were excluded. Initial analysis focused on hand movements and tremor.

Results:

From 71 videos (30 unique patients), 382 video segments were extracted. Hand movements were analyzed, and with an 80/20 training split, concordance with expert rater scores was 0.70 for finger tapping, 0.85 for hand opening/closing, and 0.75 for pronation/supination. In the 60/40 split, these scores were 0.50, 0.83, and 0.75, respectively. The AI’s performance met the endpoint of matching expert MDS-UPDRS scores within a ±1.0-point error margin.

Conclusions:

This prototype demonstrates the feasibility of automating MDS-UPDRS hand movement scoring using single-camera smartphone video. Sample size was the primary limitation, and additional data will enable modeling of more MDS-UPDRS items. This method could also reduce scoring subjectivity, e.g. rater tendency to preferentially score in the middle range (1-2), thereby increasing precision and objectivity. This approach could improve access to accurate movement assessments across clinical and research settings, enhancing both variability and resolution.