Abstract
Adaptive deep brain stimulation has enabled machine learning models to track motor states from neural signals with improved accuracy, aiming to provide electrical stimulation accordingly. Such data-driven techniques necessitate extensive user-specific data collection involving repetitive tasks and additional sensors to quantify continuous movements, due to variations in neural signals between individuals. In this study, we introduce Neural-to-Gait Calibrator, a cross-subject deep learning framework that leverages collective neural data to track gait performance of users with Parkinson's disease. Our frame-work utilizes domain adversarial learning to calibrate target user's neural signals using data from other individuals, removing the need for synchronous gait recording systems thereby enabling personalized model calibration outside equipped clinical settings. The framework's effectiveness was demonstrated through a significant reduction in error rates compared to models trained with data from other individuals without calibration, achieving performance comparable to that of models trained directly with labeled target data.