Low-Cost AI-Based Wearable Rehabilitation Monitoring System for Upper-Limb Stroke Patients

Abstract

The low cost wearble monitoring framework for the automatic evaluation of exercises for stroke recovery of the upper limb with special attention to discrimination between the intended and compensatory motion of the elbow joint is designed. The monitoring framework utilizes a 9 degrees-of-freedom IMU combined with the ESP32-based data acquisition and transmission device to obtain the motion of the forearm at 100 Hz. The signals are processed using the fourth order Butterworth filter, orientation estimation by Madgwick algorithm, quaternion representation, fixed window segmentation, and features extraction. A stacked 128- and 64-unit LSTM network is used to identify the characteristics of movements for the four classes of correct elbow flexion, correct elbow extension, compensatory flexion, and compensatory extension. The proposed system obtains an accuracy of 93.25%, precision of 93.3%, recall of 93.25%, F1-Score of 93.2%, and an AUC-ROC macro average of 0.822. Processing time and communication time are estimated at 312 ms, and the estimated cost for wearable hardware is USD 27.40. The system combines movement classification and correction feedback and offers an efficient system for inexpensive rehabilitation monitoring.