Self-Supervised Feature Learning for Robust and Interpretable Step Event Detection in Multi-Fidelity Wearable Data

Abstract: Accurate step event detection from wearable accelerometer data is critical for health monitoring but faces challenges from limited annotated data and variability in sensor placement and sampling frequency. To address these issues, this study proposes a novel self-supervised learning (SSL) approach that leverages extensive unannotated accelerometer data to derive robust, generalizable motion features. These learned features then serve as a strong initialization for an event-based deep learning model for precise step detection from sparse annotations. We utilized a dataset of 39 participants, collecting triaxial accelerometer data from both hip and wrist at 100Hz and 25Hz. Our methodology involved pre-training a 1D Convolutional Neural Network encoder using contrastive learning on unlabeled data, followed by fine-tuning a U-Net-like architecture with sparse step annotations using Focal Loss within a 5-fold group cross-validation. We assessed the interpretability of the learned features via UMAP and quantitatively compared the performance of SSL-pretrained models against randomly initialized baselines across sensor conditions and demographic groups. Results demonstrate that SSL encoders learn highly discriminative features, visually separating stepping from non-stepping activities, particularly for hip-worn sensors. Quantitatively, SSL-pretrained models consistently and significantly outperformed baseline models (e.g., for Hip 100Hz, F1-score was 0.96 vs. 0.92, and Mean Absolute Percentage Error was 4.8\% vs. 8.2\%). Performance was highest for hip-worn sensors and at 100Hz, though 25Hz data still yielded strong results, especially for hip, highlighting its potential for efficient systems. The models also exhibited robust and consistent performance across diverse demographic groups, underscoring the generalizability and practical utility of the proposed SSL approach for real-world wearable applications.

Subjects:	eess.SP; cs.LG
Cite as:	PX:2508.00025