Wearable IMU analysis and semi-supervised recognition of hemiparetic gait states and phases

Wearable inertial measurement units (IMUs) can extend gait assessment beyond motion-analysis labora-tories, but rehabilitation applications require interpretable measures and reliable event timing. This sin-gle-center proof-of-concept pilot study developed a wearable framework for post-stroke gait assessment, gait-event detection and phase recognition. Eight participants with post-stroke hemiparesis and eight healthy controls completed walking trials instrumented with seven IMUs, Vicon motion capture, and em-bedded AMTI force platforms. Vicon-derived kinematics, spatiotemporal measures, and bilateral sym-metry characterized the deficits represented in the wearable signals. The affected side showed reduced knee flexion and ankle plantarflexion, shorter spatial progression, and greater reliance on double support. Heel- and toe-marker trajectories generated labels independently of the ankle-IMU inputs, while vertical ground-reaction-force thresholds verified heel-strike (HS) and toe-off (TO). Only bilateral ankle signals were supplied to the classifiers; the seven-sensor setup supported labeling and sensor-location comparison. A target-contralateral lightweight convolutional neural network (TC-LCNN) classified healthy, unaffect-ed-side, and affected-side gait. A semi-supervised gradient-reversal Transformer-BiLSTM (SSGR-TBiLSTM) recognized background, HS, full-foot contact (FF), heel-off (HO), and TO, using weak-to-strong consistency learning, confidence-filtered pseudo-labeling, and gradient reversal to exploit unannotated recordings and reduce participant-dependent variation. Under participant-independent eight-fold validation, TC-LCNN achieved 99.66% ± 0.36% accuracy and 99.71% ± 0.31% macro-F1 with 16.93 K parameters and 0.145 M MACs per window. Under eight-fold leave-one-subject-out validation, SSGR-TBiLSTM achieved left- and right-side macro-F1 scores of 93.50% ± 0.98% and 92.40% ± 0.59%, respectively. Event-level precision, recall, and F1 were 97.90%, 97.40%, and 97.65%; 95.40% and 99.00% of timing errors were within ±50 and ±100 ms. Mean absolute errors were 15-19 ms for HS, 17-23 ms for TO, and 33-41 ms for FF and HO. These preliminary within-cohort findings demonstrate label-efficient gait-state modeling and more precise HS/TO timing, while supporting multicenter and closed-loop valida-tion before phase-triggered clinical use.