Bidirectional Cross-Task Transfer Learning for Gait Phase Prediction in Frail Older Adults Using Wearable Sensors

Transfer learning has been widely applied to gait analysis using wearable sensors to address data scarcity and improve model generalization. However, existing approaches predominantly focus on intra-task transfer (classification-to-classification or regression-to-regression), while cross-task transfer between fundamentally different learning objectives remains largely unexplored. This study proposes a novel bidirectional cross-task transfer learning framework to jointly improve both classification and regression performance in wearable-sensor-based gait analysis. As a proof of concept, we establish knowledge transfer between continuous gait cycle percentage prediction (regression) and discrete gait phase classification using the GSTRIDE dataset of frail older adults. Bidirectional transfer is implemented through end-to-end fine-tuning of deep neural network and Transformer architectures augmented with attention mechanisms. Experimental results demonstrate that model-level transfer yields an F1-score of 0.9778 (a 4.13% improvement) for gait phase classification and a mean absolute error (MAE) of 0.0358 for gait cycle percentage prediction (a 10.9% improvement) compared with models trained without transfer, while feature-level transfer provides a computationally efficient alternative with comparable performance. The proposed bidirectional framework shows strong potential for practical deployment in wearable systems, enabling widespread applications in fall risk assessment, rehabilitation monitoring, and early detection of neurodegenerative diseases in aging populations.