Challenges in Learning Universal Gait Fingerprints: Evaluating Adversarial Invariance and Demographic Bias for Wearable Step Counting

Abstract: Robust step counting from wearable accelerometers is crucial for digital health, yet current methods often lack generalizability across diverse sensor configurations and user populations. This paper investigated the feasibility of learning "universal gait fingerprints"—low-dimensional representations of purposeful steps inherently invariant to sensor location and sampling frequency, and adaptive to demographics. We proposed a deep learning framework featuring a 1D Convolutional Neural Network encoder and multi-task adversarial training with a Gradient Reversal Layer. This model was trained and rigorously evaluated on the OxWalk dataset, comprising triaxial accelerometer data collected from 39 participants using concurrent hip and wrist sensors at 25Hz and 100Hz. Our results demonstrate that while the adversarial approach largely succeeded in achieving invariance to sampling frequency, it critically failed to learn location-invariant representations, as evidenced by a 96.47\% accuracy in classifying sensor location from the learned embeddings and significant degradation in step-counting performance for wrist-worn data. Furthermore, the model exhibited substantial demographic bias, with Mean Absolute Percentage Error (MAPE) rising from 21.24\% for younger adults (19-30) to 75.04\% for older adults (45-81), and higher absolute errors for female participants. These findings suggest that the concept of a single, monolithic universal gait fingerprint is an oversimplification, underscoring the inherent challenges in developing truly generalizable step counting models without explicitly accounting for fundamental biomechanical and demographic variations.

Subjects:	eess.SP; cs.LG
Cite as:	PX:2508.00022