Abstract
BACKGROUND: Recognizing knee hyperextension during gait in stroke patients is clinically challenging and involves cumbersome, costly procedures. This study proposes a simplified strategy based on a transformer-long-short-term memory (transformer-LSTM) approach for knee hyperextension recognition via surface electromyography (sEMG) data. AIM: This study proposes an algorithmic model to streamline knee hyperextension screening in stroke patients and reduce associated screening costs. METHODS: Using an open-source gait database of 50 stroke patients, co-contraction index (CCI) values-derived from Biceps Femoris (BF)/Rectus Femoris (RF), Tibialis Anterior (TA)/Gastrocnemius (GAS), and their combinations-were input into the model. Performance was evaluated using accuracy, F1 scores, Receiver Operating Characteristic (ROC) curves, and loss function convergence. RESULTS: The transformer-LSTM achieved a recognition accuracy of 83.38% (F1: 0.8335), outperforming linear regression (36.78%, F1: 0.3405), support vector machines (38.65%, F1: 0.3185), convolutional neural networks (72.56%, F1: 0.7259), gated recurrent units (70.94%, F1: 0.7103), long-short-term memory (76.86%, F1: 0.7687), and transformer (75.76%, F1: 0.7678). The transformer-LSTM demonstrated the fastest loss function convergence and an ROC Area Under Curve of 0.99. CONCLUSIONS: The transformer-LSTM provides superior recognition rates (83.38%) and offers a robust solution for screening knee hyperextension in stroke patients, highlighting its potential for clinical application.