Extraction of Clinically Relevant Temporal Gait Parameters from IMU Sensors Mimicking the Use of Smartphones

TL;DR

This study shows that smartphone-like sensors can accurately track gait patterns in real-world settings, making remote health monitoring more accessible.

Contribution

A CNN-LSTM model achieves accurate gait parameter prediction from a single IMU in smartphone-like positions.

Findings

Stride time predicted with <5% error across different IMU placements.

Stance and swing times showed moderate errors, while double support time had >20% error.

Predictions correlated moderately strongly with lab data, preserving inter-subject gait patterns.

Abstract

What are the main findings?
Single IMU + CNN-LSTM predicts stride time with <5% errors across hand, pocket, and jacket placements.Stance/swing times show moderate errors; double support > 20%, yet all correlate moderately strongly with lab data.

Single IMU + CNN-LSTM predicts stride time with <5% errors across hand, pocket, and jacket placements.

Stance/swing times show moderate errors; double support > 20%, yet all correlate moderately strongly with lab data.

What is the implication of the main finding?
Smartphone-based IMU enables remote, real-world gait tracking.Robust predictions across positions and speeds support scalable monitoring of gait disorders.

Smartphone-based IMU enables remote, real-world gait tracking.

Robust predictions across positions and speeds support scalable monitoring of gait disorders.

As populations age and workforces decline, the need for accessible health assessment methods grows. The merging of accessible and affordable sensors such as inertial measurement units (IMUs) and advanced machine learning techniques now enables gait assessment beyond traditional laboratory settings. A total of 52 participants walked at three speeds while carrying a smartphone-sized IMU in natural positions (hand, trouser pocket, or jacket pocket). A previously trained Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM)-based machine learning model predicted gait events, which were then used to calculate stride time, stance time, swing time, and double support time. Stride time predictions were highly accurate (<5% error), while stance and swing times exhibited moderate variability and double support time showed the highest errors (>20%). Despite these variations, moderate-to-strong correlations between the predicted and experimental spatiotemporal gait parameters suggest the feasibility of IMU-based gait tracking in real-world settings. These associations preserved inter-subject patterns that are relevant for detecting gait disorders. Our study demonstrated the feasibility of extracting clinically relevant gait parameters using IMU data mimicking smartphone use, especially parameters with longer durations such as stride time. Robustness across sensor locations and walking speeds supports deep learning on single-IMU data as a viable tool for remote gait monitoring.

Figures