Classification of Hand Gestures with sEMG Signals

Placement of the EMG sensors at the targeted muscle locations on the forearm of the participant.

Objective

Hand function plays a crucial role in completing activities in daily living. Due to the importance of hand function and its direct physical interaction with the world, a significant part of the motor cortex is spent on controlling our hands. Individuals who suffer from mobility impairment, experience a significant drop in their quality of life when control over their hand function is impacted. Accurate functional assessment is a crucial step in the rehabilitation and therapy planning process. Current assessment procedures are limited in measuring hand function with specific objects. Furthermore, these clinical methods are manually assessed within a controlled environment and limited duration, which are unreliable and not necessarily an accurate representation of hand function in the real world.

Noninvasive surface electromyography (sEMG) is popular for measuring musculoskeletal activity due to its safety, reliability, and efficient capture of temporal signals. This has led to increased sEMG use for medical and rehabilitative purposes. As a result, many researchers have shown interest in developing mathematical models to derive meaningful insights from minimal sensors. In this project, we explore the potential of machine learning (ML) and deep learning (DL) methods to identify pre-defined hand gestures from sEMG data collected from the forearm.

Design

Seven sensors were placed at targeted muscle locations on the forearm of the subjects. Three participants were chosen for the study, and sensors were placed over the following muscle locations of the dominant forearm: Flexor Carpi Radialis (FCR), Palmaris Longus (PL), Flexor Carpi Ulnaris (FCU), Flexor Digitorum (Flx Digi), Extensor Carpi Radialis (ECR), Extensor Carpi Ulnaris (ECU), Extensor Digitorum (Ext Digi), and the back of the hand. 10 wrist and hand movements were chosen for the classification study: Wrist Flexion (WFLX), Wrist Extension (WEXT), Abduction (WBD), Adduction (WADD), Hand Open (REST), Hand Closed (HCLO), and individual finger flexions (THMB, RING, MIDL, INDX, PINK). Participants alternated between a specified movement and a resting state for 90 trials per movement, a total of 2700 active trials and 2700 resting trials. The preprocessing data pipeline is as follows: EMG signals were first passed through a Bandpass filter (4th order, Zero-lag Butterworth filter, 25-300 Hz frequency range) to remove background noise. Next, the data was rectified to baseline, and then a Low-pass filter was applied (4 Hz cut-off) to smoothen the signals. Finally, the smoothed data was normalized for each muscle by its maximum measured value across all movements.

sEMG data cleaning and preprocessing pipeline.

The following features along with their formulation were chosen for this study.

Application

Traditional ML algorithms, including logistic regression, decision trees, LDA, KNN, SVM, and DL models such as MLP and CNN, were employed using Tensorflow and Scikit-learn to investigate the effectiveness of each model for gesture classification. A set of 5 features was extracted from the data: [L2, Variance, RMS, Skewness, Kurtosis]. To evaluate each model's performance, we utilized accuracy, precision, recall, receiver operating characteristics curve, and the area under the curve metrics. These metrics are widely used for classification problems and provide a graphical representation of the varying true positive and false positive rates.

Metric-wise accuracy variation across ML & DL Models. Better-performing models (KNN and SVM) yielded higher accuracy with Metric 1, Metric 2, and Metric 3 compared to other models

Parameters and accuracy outcomes for each ML and DL model. The KNN-approach achieved the highest the testing accuracy with minimal hyperparameters.

Our results show that the KNN model achieved the highest accuracy of 81.42%, followed by SVM with 81.32% and the decision tree with 75.57% for a given combination of features. A three-layer MLP achieved an accuracy of 77%, and a CNN attained an accuracy of 99%. Our findings suggest that the ML models are highly accurate in predicting hand signs, but CNN requires higher data variability to avoid overfitting and address potential ethical considerations.

Page updated

Google Sites

Report abuse