Deep Learning Prediction of Parkinson’s Disease using Remotely Collected Structured Mouse Trace Data

Parkinson’s Disease (PD) is the second most common neurodegenerative disorder globally, and current screening methods often rely on subjective evaluations. We developed deep learning-based classification models using structured mouse trace data collected via a web application. 261 participants (73 PD, 155 non-PD, 33 suspected PD) completed three hand movement tasks: tracing a straight line, spiral, and sinewave. We developed three types of models: (1) engineered features model, (2) computer vision models, and (3) multimodal models. The best-performing models were image-based DenseNet-201 model with an F1 score of 0.9027 ± 0.0332 (PD vs. non-PD), multimodal ResNet-50 with an F1 score of 0.9353 ± 0.0334 (suspected PD vs. non-PD), and multimodal ViT with an F1 score of 0.7619 ± 0.0535 (PD vs non-PD). Feature importance for the best-performing models was evaluated using Gradient Shapley Additive Explanations (GradShap). Image inputs consistently proved most predictive. The findings suggested that models trained on confirmed PD diagnoses hold promise for early-stage PD screening.