Real-Time Estimation of Anterior-Posterior Ground Reaction Forces in Individuals with Stroke Using Wearable Sensors and Deep Learning

Reduced limb propulsion is linked to impaired mobility and fall risk. Biofeedback retraining that targets anterior ground reaction forces (AGRFs) can enhance propulsion, but current approaches rely on costly lab-based equipment. Inertial measurement units (IMUs) offer a low-cost, portable alternative. We developed a hybrid convolutional–recurrent neural network to estimate AGRF time series during walking in 14 individuals with chronic hemiparesis post-stroke, using kinematic signals from wearable IMUs. The model was trained with leave-one-subject-out cross validation to emphasize generalizability and evaluated using mean absolute error (MAE), coefficients of determination (R²), normalized root mean squared error (NRMSE), and minimal detectable change (MDC). With IMUs across all lower-limb segments, the model achieved a MAE of 3.3 ± 1.2% BW; accuracy was preserved with only pelvis and thigh sensors (3.4 ± 1.5% BW), supporting clinical practicality with reduced sensor sets. Although group-level errors exceeded MDC thresholds, over half of trials fell within the between-session MDC, suggesting reliable within-subject tracking of AGRF. Outlier cases, primarily among participants with greater mediolateral sway, indicated gait stability as a key factor influencing accuracy. This highlights the potential need for personalized calibration or adaptive modeling in heterogeneous populations. Designed for streaming, frame-by-frame inference with low computational cost, the model supports real-time deployment on portable hardware. By enabling accurate AGRF estimation with a few wearable IMUs, this approach could facilitate in-clinic propulsion assessments and biofeedback-based gait training outside controlled labs, improving accessibility and ecological validity of rehabilitation.