Detection of early-stage Parkinson's disease using wearable sensors at multiple body locations and convolutional neural networks.

Early-stage Parkinson's disease (PD) presents with subtle motor symptoms that complicate timely diagnosis. We developed a non-invasive detection framework using wearable sensors and a convolutional neural network (CNN) during a 6-min walk test. Time-series data were collected from 78 patients with early-stage PD and 50 controls across six body locations. Straight walking segments were analyzed, and 34 non-linear gait features were additionally extracted to complement deep learning with interpretable machine-learning models. The CNN achieved 95.6% accuracy using left-arm gyroscope data during the first minute of straight walking. Temporal analyses suggested that classification performance remained relatively stable across 1-2-min measurement windows, indicating the potential utility of short-duration gait assessments for early-stage PD detection. Feature-based machine-learning models using a reduced set of selected non-linear features demonstrated performance comparable to models using the full feature set, with peak discrimination observed during early and late test intervals. Although the highest accuracy was obtained from first-minute CNN analysis, multi-segment evaluation revealed complementary time-dependent motor signatures captured by interpretable features. These findings suggest that short-duration straight-walking data enable accurate and efficient early PD screening, while multi-segment analysis provides a more comprehensive and physiologically meaningful characterization of early motor dysfunction.