Enhancing Gait Analysis for Parkinson's Disease Detection and Severity Staging With a Parallel Conv1D-Efficient Transformer and Bidirectional GRU Hybrid Architecture

Abstract

Gait analysis is a critical tool for diagnosing and assessing the severity of Parkinson's disease (PD). This study introduces a novel parallel hybrid architecture combining Conv1D, Efficient Transformers, and Bidirectional GRU layers to analyze gait data for both PD detection and severity staging. Conv1D layers extract local spatial features, Efficient Transformers capture contextual dependencies, and Bidirectional GRUs model temporal patterns in VGRF signals. Designed to balance computational efficiency and scalability, the model demonstrates state-of-the-art performance, achieving 95.7% accuracy in PD detection and 87.3% accuracy in severity staging on the PhysioNet gait dataset. Additionally, the architecture is highly versatile, offering potential for application in other 1D signal analysis tasks.