Wearable IMU-based gait-cycle signal processing with multi-scale attention for preliminary auxiliary identification of osteonecrosis of the femoral head

Abstract

Osteonecrosis of the femoral head (ONFH) often presents with nonspecific early symptoms, which may lead to missed or delayed diagnosis. Gait abnormality is an important functional manifestation of ONFH, and wearable inertial measurement unit (IMU)-based gait analysis may provide a portable and non-invasive approach for auxiliary assessment. In this study, we developed a wearable IMU-based gait-cycle signal processing and modeling framework for preliminary auxiliary identification of ONFH. Raw multichannel gait signals were segmented into gait cycles, followed by cycle-level quality control and temporal normalization to obtain fixed-length multichannel gait-cycle samples. A multi-scale gated attention network (MSGA-Net) was further proposed to capture multi-scale temporal dynamics within gait cycles while controlling model complexity. The framework was evaluated at the subject level using leave-one-subject-out cross-validation and compared with advanced time-series models and conventional machine-learning baselines based on handcrafted gait features. Complementary feature-based and hierarchical occlusion analyses were conducted to examine model interpretability. MSGA-Net achieved a subject-level accuracy of 0.941, sensitivity of 0.810, specificity of 0.996, and AUC of 0.982 at the default threshold. The interpretability analyses suggested that the predictions relied on multiple gait characteristics rather than a single feature. A post hoc threshold-dependent analysis showed a trade-off between sensitivity and specificity in the present dataset. These findings suggest that wearable IMU-derived gait-cycle dynamics may contain discriminative information for distinguishing ONFH patients from healthy controls. The proposed framework provides preliminary feasibility evidence for IMU-based auxiliary identification of ONFH and warrants validation in larger independent cohorts before practical clinical use.