Spatio-temporal graph convolutional networks with transfer learning for continuous ground reaction force estimation in hemiparetic gait.

Three-dimensional ground reaction force (GRF) is an important biomechanical indicator of weight-bearing, propulsion, and bilateral asymmetry in hemiplegic gait. However, conventional GRF measurement relies on laboratory-based force plates, limiting its use in continuous rehabilitation assessment. A specific methodological challenge is to estimate continuous three-dimensional GRF in post-stroke hemiplegic gait without using force-plate signals as model input, while still preserving whole-body kinematic coordination and affected-unaffected side asymmetry. This study proposed a force-plate-independent, marker-based method for estimating continuous stance-phase three-dimensional GRF in patients with hemiplegia by combining a spatio-temporal graph convolutional network (ST-GCN) with two-stage transfer learning. Data were collected from 30 chronic stroke patients with hemiplegia and 60 healthy controls. The model used 39 raw Plug-in Gait markers as graph nodes, with 10-dimensional node features consisting of three-dimensional position, velocity, acceleration, and laterality encoding. The model was pretrained using healthy participant data and then fine-tuned and evaluated on hemiplegic gait data using leave-one-subject-out cross-validation. The main contribution of this work is the integration of marker-level body topology, explicit kinematic derivatives, pathological laterality encoding, and healthy-to-hemiplegic transfer learning within a unified ST-GCN framework. The proposed ST-GCN achieved Pearson's