Cross-Configuration Transfer Learning Framework for Robust Step Counting in Free-Living Conditions

Abstract: Reliable step counting in free-living conditions is essential for health monitoring, but its accuracy is challenged by the diversity of wearable sensor configurations and user populations. This study addresses these challenges by developing a cross-configuration transfer learning framework to assess the generalizability of machine learning models for step counting. Using Leave-One-Subject-Out Cross-Validation, we trained a LightGBM model on high-fidelity hip-worn accelerometer data (100Hz) from 39 participants. We then rigorously evaluated its zero-shot transferability to data from different sensor locations (wrist) and reduced sampling frequencies (25Hz), aiming to identify generalizable motion patterns. While the source model demonstrated strong baseline performance (Mean Absolute Error: 387.54 steps, Mean Absolute Percentage Error: 12.88\%), direct transfer resulted in significant and statistically confirmed performance degradation across all target configurations. Errors escalated considerably for wrist-worn data and lower sampling rates, culminating in a Mean Absolute Error of 1978.11 steps and a Mean Absolute Percentage Error of 66.91\% for the Wrist 25Hz configuration. This degradation was characterized by systematic step underestimation and increased inter-individual variability. Interestingly, statistical analyses revealed no significant differences in transfer performance based on participant sex or age range, indicating that the challenges posed by cross-configuration transfer affect demographic subgroups equitably. These findings underscore the inherent difficulties of directly applying models across vastly different sensor configurations without adaptation, and suggest that demographic factors may not be the primary determinants of performance loss in zero-shot transfer scenarios for step counting.

Subjects:	eess.SP; cs.LG
Cite as:	PX:2508.00026