FB-Mamba, an Efficient Bidirectional State Space Model for Wearable Sensor-Based Gait Analysis

Wearable sensor-based gait analysis encompasses diverse tasks including human activity recognition, neurodegenerative disease diagnosis, and clinical event detection. Existing deep learning approaches are typically designed for specific tasks, lacking a unified framework that generalizes across different scenarios. We present the first systematic investigation of state space models for wearable sensor-based gait analysis, proposing FB-Mamba, a Forward-Backward Mamba architecture that achieves bidirectional temporal modeling with linear complexity through an adaptive fusion mechanism. We establish FB-Mamba as a general-purpose framework spanning three major task categories, including human activity recognition, disease detection, and event-level detection. Comprehensive experiments on five datasets with heterogeneous sensor modalities and sampling rates demonstrate that FB-Mamba achieves consistently competitive performance across all task categories while maintaining computational efficiency and favorable long-range dependency modeling capability.