# An open-source, externally validated neural network algorithm to recognize daily-life gait of older adults based on a lower-back sensor

## TL;DR

This paper presents an open-source neural network that accurately recognizes gait in older adults using lower-back sensor data, with strong performance in real-life settings.

## Contribution

An open-source, externally validated gait recognition model for older adults using lower-back inertial sensors and the impact of data augmentation is introduced.

## Key findings

- The best 6-channel model achieved 91.4% accuracy and 99.5% sensitivity in testing.
- The best 3-channel model achieved 96.5% accuracy and 98.9% sensitivity in testing.
- Models with data augmentation showed improved performance, especially for acceleration-only data.

## Abstract

Gait recognition is critical for daily-life fall risk assessment and rehabilitation monitoring, but existing models face many challenges that limit their use in daily life for older adults. We aimed to develop an open-source gait recognition for older adults using sensor data and explore the effect of data augmentation on model training. A convolutional neural network was trained using lower-back inertial sensor data from 20 participants (mean age 76.4) and externally validated on 47 participants (mean age 72.3). The model was trained using 6-channel data (accelerations and angular velocities) and 3-channel data (accelerations only), with and without data augmentation. On the testing dataset, the best 6-channel model achieved accuracy of 91.4%, precision of 59.7%, sensitivity of 99.5%, F1-score of 74.7%, and specificity of 90.3%, and the best 3-channel model achieved accuracy of 96.5%, precision of 78.7%, sensitivity of 98.9%, F1-score of 87.6%, and specificity of 96.1%. On the external validation dataset, the best models with both channels show near-perfect scores. This study demonstrates that the convolutional neural network algorithm based on lower-back inertial sensor data can accurately recognize daily-life gait of older adults, and data augmentation was especially beneficial for models using acceleration data only.

The online version contains supplementary material available at 10.1007/s11517-025-03466-z.