Intelligent prediction of dynamic characteristics during exercise in patients with stroke.

Background: Accurate estimation of lower limb kinetic characteristics is critical for evaluating gait impairments and guiding rehabilitation in patients with stroke. Traditional three-dimensional (3D) optical motion capture systems provide high-precision measurements but are costly, require a laboratory environment, and are sensitive to marker placement errors. Inertial measurement unit (IMU) sensors, combined with machine learning models, may offer a portable and clinically feasible alternative.

Methods: Thirty patients with stroke performed level walking and stair negotiation tasks while wearing IMU sensors. Joint kinematic data derived from the IMUs were processed using principal component analysis (PCA) for dimensionality reduction, and lower limb joint torques were predicted using a backpropagation (BP) neural network. The proposed Principal Component Analysis - Back Propagation (PCA-BP) model was evaluated using normalized root mean square error (NRMSE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²).

Results: The PCA-BP model achieved high prediction accuracy for hip, knee, and ankle joint torques across sagittal, coronal, and transverse planes during both walking and stair tasks. Performance metrics indicated good agreement between predicted values and those obtained from OpenSim simulations based on IMU-derived kinematics.

Conclusion: IMU-based gait analysis in patients with stroke demonstrated the potential to serve as an alternative to traditional 3D optical motion capture systems, particularly in non-laboratory or resource-limited settings. This approach offers portability and practicality for sports scientists and clinicians, supporting its potential integration into routine clinical rehabilitation assessments.