Evaluation of Deep Learning-Based Event Detection for Parameter Estimation During Complex Walking in Parkinson's Disease.

Abstract

Despite recent advances in wearable technology and its use in quantifying movement, there is still a need for reliable methods of quantifying complex walking tasks beyond steady-state gait (SSG). The purpose of this study is to evaluate an inertial sensor-based processing pipeline that uses a deep learning method for event detection during stride segmentation and established methods for trajectory reconstruction and gait parameter calculation of simple and complex walking tasks. We propose a method that utilizes a Temporal Convolutional Network (TCN) during stride segmentation and pre-established methods for trajectory reconstruction and parameter extraction to accurately quantify spatiotemporal parameters of steady state gait (SSG), turn, gait initiation (GI), and termination (GT) strides. The results from this pipeline were evaluated against a pressure walkway as the reference system. Overall, our method was able to derive temporal and spatial parameters with small mean errors (≤ 1 ms and ≤ 2.3 cm, respectively) and strong correlation (r ≥ 0.96) with the pressure walkway for SSG strides. Turn, GI, and GT strides temporal and spatial parameters had similar performance (≤ 7 ms and ≤ 2.9 cm, respectively) and strong correlation (r ≥ 0.95) with the walkway. This study demonstrated that IMU derived gait metrics using TCN model event detection for stride segmentation and Gaitmap functions for stride reconstruction and parameter calculation can be used to quantify gait during both simple and complex walking tasks. The proposed method provides a reliable way to quantify complex walking tasks, allowing for a more complete understanding of mobility in home and community environments.