Remote Assessment of Parkinson Disease Using Deep Learning on Structured Mouse-Trace Data From Suspected Cases: Machine-Learning Pilot Feasibility Study

Background: Parkinson disease (PD) is a pervasive neurodegenerative disorder globally, largely characterized by motor symptoms. Most existing artificial intelligence models for PD detection are trained on participants in well-resourced settings with confirmed clinical diagnoses. However, specialist-confirmed labels are often infeasible in low-resource settings. Objective: We developed a web platform for structured mouse data collection through pattern tracing tests. We sought to assess the feasibility of leveraging data from a community-recruited sample of participants with suspected but undiagnosed PD to train artificial intelligence models that achieve respectable performance in predicting diagnosed PD. We tested whether using weaker diagnostic labels that may be more feasible to collect in community or global health settings, where access to professional neurologists is sparse or nonexistent, can lead to models that learn predictive signals that are diagnostically useful. Methods: 261 participants (73 self-reported PD, 155 non-PD, and 33 suspected PD) were recruited from community organizations in Hawaii and completed 3 pattern tracing tasks on our custom web assessment: straight line, sine wave, and spiral wave. During each task, cursor positions, screen dimensions, and an in-target boolean flag were recorded. From these data, we engineered features and generated mouse trace images. We built 3 categories of classifiers: (1) a feed-forward neural network using engineered features, (2) fine-tuned computer vision deep learning models, and (3) multimodal models concatenating a feed-forward neural network with computer vision models. Performance was evaluated using 1 primary experiment and 2 secondary analyses. The primary experiment involved training on suspected PD versus non-PD and testing on self-reported PD versus non-PD. A secondary analysis evaluated the reverse direction by training on participants with self-reported PD and without PD and then testing on participants with suspected PD versus participants without PD. Additionally, a cross-validation analysis was conducted using participants with self-reported PD versus those without PD with 5-fold cross-validation to establish baseline performance under well-defined diagnostic labels. Results: The best-performing models included a multimodal Vision Transformer in the primary experiment (F1: mean 0.7619, SD 0.0535), a multimodal ResNet-50 in the secondary analysis (F1: mean 0.9353, SD 0.0334), and an image-based DenseNet-201 in the cross-validation analysis (F1: mean 0.9027, SD 0.0332). Training on patients with suspected PD yielded meaningful performance in predicting self-reported PD, supporting the feasibility of using lower-specificity labels for model development. Conclusions: This pilot feasibility study suggests that remotely collected mouse-tracing data can support PD screening models under data labeling conditions of low diagnostic specificity: models trained on suspected PD from a community sample may learn signals that can transfer to predicting actual PD. Future work may consider pretraining using weaker labels and then fine-tuning on stronger clinical labels.