JoWUA

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative disorder that remains difficult to diagnose at an early stage, as most clinical symptoms appear only after substantial neuronal loss. This study proposes an integrated deep learning based auxiliary diagnosis system for PD using human gait data collected from wearable triaxial accelerometer sensors. The system employs stacked Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to capture complex temporal dependencies and motor anomalies associated with Parkinsonian gait. Experiments were conducted on the DeFOG dataset, which consists of multi-subject gait recordings annotated with freezing of gait (FoG) events. After rigorous preprocessing, including data cleaning, normalization, and sliding-window segmentation, the stacked GRU model achieved superior results, with 95.9% accuracy, 90.7% precision, 83.6% recall, 87.0% F1-score, and an AUC-ROC of 0.92. These results significantly outperformed traditional machine learning baselines such as Support Vector Machines (SVM) at 86.2%, Random Forests (RF) at 84.7%, and k-Nearest Neighbors (KNN) at 82.9%, as well as deep learning baselines, including CNN-only 90.1% and LSTM-only 92.5% models. Ablation studies confirmed the critical role of preprocessing and multi-layer temporal modeling in improving classification performance. With a low inference latency of 8 ms/sample, the system is well-suited for real-time deployment on wearable devices. While challenges remain in detecting brief FoG episodes and mitigating sensor variability, these results demonstrate that a stacked LSTM–GRU motion analysis system offers reliable, non-invasive support for early PD detection and continuous monitoring in real-world clinical and home settings.