Physics-Informed Learning Framework for Lower Limb Kinematic Prediction With Sparse Sensors and Its Application in Chronic Stroke.

Kinematic evaluation of gait is critical for biomechanical analysis and disease diagnosis. Stroke survivors, due to unilateral motor impairments, often exhibit asymmetric gait patterns, significantly altering lower limb joint kinematics. Accurate measurement of lower limb joint angles enables therapists to assess the functional status of stroke patients, thereby promoting the rehabilitation process. Although optical motion capture systems provide precise measurements, their use is constrained to laboratory and clinical settings. Inertial measurement units (IMUs) offer a promising wearable alternative for monitoring gait during daily life. However, comprehensive segment motion tracking typically requires multiple IMUs, leading to inconvenience and interference with daily activities. This study proposes a physics-informed learning framework utilizing a temporal convolutional network (TCN) for lower-limb kinematics prediction, significantly reducing sensor count to only two IMUs. Geometric physical constraints derived from IMU measurements and human gait modeling are integrated into the neural network during training. Validation on six healthy subjects and seventeen chronic stroke patients indicates the effectiveness of this proposed framework, achieving comparable accuracy with only two IMUs as with four or more IMUs, highlighting its potential for practical rehabilitation applications.