Hyperparameter tuning of light gradient boosting machine for electromyography signal classification

Abstract Real-time myoelectric interfaces utilizing surface Electromyography (sEMG) data can be harnessed to create myoelectric prosthetic hands based on the captured muscular activities. The myoelectric signals must be classified accurately to effectively govern the operation of these devices. This paper proposes a classification framework for improving the classification accuracy of EMG signals by tuning the hyperparameters of the Light Gradient Boosting Machine (LightGBM) classifier. Features like Hudgins Time Domain (HTD) feature set, fused Time Domain Descriptors (fTDD), Auto Regressive (AR) - RMS and TDAR6 features were extracted from the EMG signals and given to the LightGBM classifier. The hyperparameters of LightGBM are tuned with Bayesian Optimization (BO) using the Tree Parzen Estimate (TPE) algorithm for robust and accurate classification. The accuracies of LightGBM were compared with traditional classifiers, attaining superior classification accuracies of 97.6% for Ninapro DB1 and 98.8% for DB2. The findings demonstrates significant performance compared to the state-of-the-art techniques.