Automated error detection in physiotherapy training

Background: Manual skills teaching, such as physiotherapy education, requires immediate teacher feedback for the students during the learning process, which to date can only be performed by expert trainers. Objectives: A machinelearning system trained only on correct performances to classify and score performed movements, to identify sources of errors in the movement and give feedback to the learner. Methods: We acquire IMU and sEMG sensor data from a commercial-grade wearable device and construct an HMM-based model for gesture classification, scoring and feedback giving. We evaluate the model on publicly available and self-generated data of an exemplary movement pattern executions. Results: The model achieves an overall accuracy of 90.71% on the public dataset and 98.9% on our dataset. An AUC of 0.99 for the ROC of the scoring method could be achieved to discriminate between correct and untrained incorrect executions. Conclusion: The proposed system demonstrated its suitability for scoring and feedback in manual skills training.