Adaptive Filtering Multi-Task Gait Network for Gait Phase Detection in Patients with Lower Limb Movement Disorders

This study presents a novel multi-task neural network model for continuous gait phase recognition and key gait event detection in patients with lower limb movement disorders. While traditional gait phase detection methods perform well in healthy individuals, they struggle to accurately capture the complex and irregular gait patterns typical in this patient population. To address this challenge, we introduce the Multi-Head Adaptive Filtering (MHAF) Block, which applies multi-scale adaptive low-pass filtering across the feature dimensions of gait data-effectively suppressing noise from abnormal gait features while retaining essential gait information. Built upon this block, we propose the Adaptive Filtering Multi-Task Gait Network (AFMG-Net), capable of concurrently classifying stance and swing phases and regressing continuous gait phases. Experimental validation using leave-one-group-out cross-validation (LOGO-CV) on gait data from 10 patients demonstrates that AFMG-Net significantly improves classification accuracy and reduces regression error in abnormal gait conditions. Compared with conventional neural networks, AFMG-Net shows strong potential for multi-task gait phase detection in pathological gait scenarios, offering a promising solution for real-world clinical applications.