Hand Gesture Recognition: Multichannel sEMG-driven Classification for Finger-level Motor Rehabilitation

In recent years, the need to enhance interaction between therapists and patients during rehabilitation sessions has grown, leading to the increasing use of engineering systems that incorporate algorithms for movement recognition. This work focuses on developing a classifier for finger gestures, as hand rehabilitation is crucial for post-stroke patients with motor impairments that may prevent or hinder their movements. This thesis proposes the development of a classifier that uses an Artificial Neural Network (ANN) to recognize and differentiate hand movements. Specifically, the study focuses on finger movements, such as counting, which are essential for hand functionality and rehabilitation. A preliminary study was conducted using the publicly available GRABMyo and Hyser datasets to extract valuable information for constructing a custom dataset. The GRABMyo dataset provided insights into inter-subject variability. The Hyser dataset was used to generate power maps, highlighting major activation points for different movements, which were then used to develop a protocol for positioning the acquisition devices. The acquisition system consisted of six acquisition devices developed by the eLiONS research group. These devices simultaneously acquire surface ElectroMyoGraphy (sEMG) and Average Threshold Crossing (ATC) signals using dedicated Python software. A total of seven different movements (Hand Close, Thumb up, One, Two, Three, Four, Five), along with the idle state, were recorded. The recording process was supported by a user-friendly GUI for real-time monitoring and control, along with a dedicated interface to guide the subject and improve recording quality. After a pre-processing step, the features for the classification were extracted, and feature selection was performed using PCA, identifying 10 features sufficient to explain 95% of the variance in the sEMG signal. The final dataset includes 24 healthy subjects, 21 of which were used for training and 3 for testing. Two classifiers were built: one including all previously introduced movements and one excluding the two most difficult movements to classify. The classifier, considering all the movements, yielded an average per-class accuracy of 92.88% and an F1 score of 71.84%. A separate classifier was developed by excluding the two most challenging movements to discriminate, resulting in significant improvement across all metrics, among which an accuracy of 95.53% and F1 score of 86.86%. A final test was performed using only one subject for both training and testing, with acquisitions on different days to assess the classifier's performance with a single subject. For the classifier considering all movements, the accuracy was 96.82% and the F1 score was 87.30%. For the classifier excluding the two most challenging movements to recognize, the accuracy increased to 98.53%, and the F1 score reached 95.60%. The performance difference between the classifier trained on all subjects and the single-subject classifier is attributed to the well-documented inter-subject variability in the literature. Consequently, the relatively small number of subjects may affect generalization. In conclusion, the proposed method represents a robust approach for developing a hand gesture recognition classifier, without the need for a high-density acquisition device, achieving high performance across the evaluation metrics. The classifier effectively discriminates finger movements using an acquisition setup that relies on features extracted from sEMG signals.