Leveraging Deep Learning and Wearables for Automatically Identifying Gait Event: Effects of Age and Location of Sensors on the Assessment of Gait Events

The clinical gait analysis (CGA) is essential for understanding functional abilities and diagnosing neuromotor conditions but remains resource intensive, often requiring expensive, immobile equipment. Wearable inertial measurement units (IMU) offer a promising alternative for real-world functional assessment, yet a lack of standardisation in sensor placement presents a significant challenge. This study aimed to develop and evaluate deep learning (DL) models, specifically convolutional neural networks (CNN) and temporal convolutional networks (TCN), by comparing F1 scores and temporal error percentiles between predicted and target gait events using IMU placed at clinical locations [head (HD), pelvis (PV), wrist (WR), and foot (FT)] in young and older adults. Our findings demonstrated that TCN consistently scored higher F1 scores with lower standard error across both young and older adult groups. Specifically, the TCN with HD sensor achieved the highest F1 scores of 95.9% in younger individuals and 93.8% in older participants for heel strike, highlighting the robustness of TCN models in capturing temporal dependencies for accurate gait event detection. In contrast, CNN models exhibited a higher standard of error, with up to 20% differences between age groups showing age-dependent influence. Temporal error analysis showed comparable performance between CNN and TCN models across most placements. Integrating gyroscope data with accelerometer signals improved F1 score, particularly for HD and FT sensors. This article systematically evaluates optimal sensor placements and demonstrates the robust performance of TCN models, laying a foundation for future research. Expanded cohorts and validation in free-living environments are needed to enhance the generalisability of IMU-based gait analysis.