Wearable Step Counting: A Comparative Analysis of Deep Learning and Traditional Methods Highlighting Data Imbalance Challenges

Abstract: Accurate and resource-efficient step counting from wearable devices in free-living conditions is crucial for health monitoring, yet it presents challenges related to sensor placement, data sampling rates, and individual demographics. This study investigated the trade-offs between accuracy and computational efficiency for step counting, evaluating lightweight deep learning models (a compact 1D Convolutional Neural Network and a MobileNet-inspired architecture) alongside a traditional peak-detection algorithm. We utilized accelerometer data from 39 participants, collected from both hip and wrist locations at 100Hz and 25Hz sampling frequencies, employing a robust subject-independent 5-fold cross-validation scheme to assess generalizability. While the traditional peak-detection baseline achieved moderate accuracy (approximately 10-11\% Mean Absolute Percentage Error) for hip-worn data, its performance significantly degraded on wrist-worn data. Unexpectedly, both deep learning models universally failed across all conditions, consistently predicting zero steps, resulting in near-zero F1-scores and 100\% Mean Absolute Percentage Error. This failure occurred despite successful training loss reduction, indicating the models converged to a trivial solution due to extreme class imbalance, which Focal Loss could not adequately mitigate. Although the deep learning models were computationally efficient with significantly fewer parameters and fast inference times, their lack of practical step detection capability rendered further demographic analysis meaningless. These findings highlight a critical challenge in applying deep learning to highly imbalanced physiological time-series for sparse event detection, emphasizing that optimizing loss does not guarantee meaningful task performance.

Subjects:	eess.SP; cs.LG
Cite as:	PX:2508.00027