Efficient EMG-Based Facial Expression Classification Using Minimal Time-Domain Features and the k-Nearest Neighbors Classifier

Abstract

This study investigates the classification of six facial expressions using surface electromyography (EMG) signals recorded from the faces of ten healthy participants (aged 21–22 years). The objective is to identify effective time-domain features for distinguishing muscle activity patterns and to evaluate the performance of various classifiers. Three widely used features, i.e., integrated EMG, root mean square (RMS), and mean absolute value, were extracted and standardized using z-score normalization. Five classifiers were compared, namely k-nearest neighbors (KNN), support vector machine (SVM), ensemble (EN), neural network, and naive Bayes. The results indicate that the KNN classifier, in combination with the three selected features, achieved the highest classification accuracy of 98.73%. However, the study is limited by a small and homogeneous sample size, which may affect the model’s generalizability. Additionally, the exclusive use of time-domain features may reduce robustness under conditions such as muscle fatigue. Future work may explore frequency-domain features, deep learning models, or the integration of EMG with other data sources, such as video, to enhance accuracy and applicability in real-world scenarios.