A Unified Contrastive Learning Framework for Neurological Disease Diagnosis from VGRF and IMU Gait Data

Accurate early diagnosis of neurological diseases (NDs) can effectively facilitate intervention and help with neural healthcare. While wearable informatics like Vertical Ground Reaction Force (VGRF) and Inertial Measurement Unit(IMU) offer complementary gait insights, a key challenge is fusing these heterogeneous modalities, especially when the data are unaligned and scarce. In this study, our primary contribution lies in proposing a novel contrastive learning framework designed specifically to unify VGRF and IMU gait data for ND classification. This framework learns a shared representation space during training, enabling the final trained model to perform accurate diagnosis from data streams of either a single VGRF or a single IMU modality. We evaluated this approach on the classification of four NDs-Parkinson's disease, Huntington's disease, amyotrophic lateral sclerosis, and stroke-against healthy controls using two public VGRF datasets and one real-world IMU dataset. As a key result, our model achieved 98.21% accuracy in the binary task (disease vs. healthy) and 96.99% accuracy in the multi-class classification. The high performance demonstrates the potential of our method to advance neural health diagnostics, providing a robust approach for gait-based neurological assessment using heterogeneous wearable sensors.