Early adherence to biofeedback training predicts long-term improvement in stroke patients: A machine learning approach.

Biofeedback-based treadmill training generally involves 10 or more sessions to assess its effectiveness during stroke rehabilitation. Improvements are seen in some patients during the assessment, while others do not progress. Our aim in this study is to determine (i) if signs of progress are evident from the initial training session and (ii) whether quantitative measurements between consecutive training sessions can guide interventions for non-progressing patients. The study analyzes Minimum Foot Clearance (MFC) data from 15 stroke patients during their baseline and second training sessions to predict outcomes in the post-assessment phase. Based on early biofeedback training data, we propose a novel approach using cosine similarity (CS), correlation coefficient (CC) and cross-correlation distance (XCRD) measures to predict post-assessment improvements in stroke patients. We also introduce a new real-time adherence assessment metric (RAAM) metric to quantify improvements in adherence to feedback between consecutive training sessions, enabling more targeted interventions. The proposed approach using CS, CC and XCRD adherence indicators demonstrates 100% accuracy in predicting improvement during post-assessments. The results show that patients with MFC values dissimilar to their baseline while adhering to targeted feedback are more likely to improve. The work also indicates that patients who don't show significant overall improvement may benefit from extended training periods, suggesting the potential for personalized rehabilitation strategies.