Enhancing performance of EEG based machine learning algorithm via Feature Fusion and Dimensional Reduction techniques

Neuro-prosthetics technology has swiftly emerged as a go-to solution for improving human lives. Currently, there are some challenges found in conventional prosthetic devices such as cosmetic prosthetic hands and myoelectric prosthetic hands. In the drive to improve the highest level of a disabled individual’s life, the EEG signal is one of thealternative methods to restore lost motor functions. This has resulted in the development of a machine learning (ML) algorithm that can be used for reliable neuro-prosthetics applications. Neuro-prosthetics is a fast-growing technology which aims to build a direct channel between humans and external devices without using the muscles and the peripheral nerves in the limbs. However, motor imagery (MI)-based neuro-prostheticsare limited in the number of control commands. Numerous features and channels lead to neuro-prosthetics devices requiring more advanced specifications and time to process the large data dimension. The effectiveness of the EEG-based ML algorithm after reducing the data dimensionality is required for evaluation. Consequently, the goal of this researchis to investigate the EEG signals by extracting hidden information from the signals of different motor movements and optimizing the data dimensionality without affecting the performance of the ML algorithm. The EEG signals of different motor movements can be recorded using EEG devices. Important features can be extracted and the data dimensionality can be reduced using advanced signal processing techniques. In this study, the acquisition protocol was established and the numerical algorithm was developed to optimize all variables extracted for the neuro-prosthetics applications. Advanced signal processing such as time-frequency multi-resolution analysis and non-linear analysis using MATLAB software was conducted for the characterization of brain wave features. Statistical-based and entropy-based features, Hjorth parameters, and HE of PSD were extracted from five frequency bands. The dominance of features from particular frequency bands in total frequency bands was investigated and the enormous data size was successfully reduced by eliminating the irrelevant features and channels using box plots and GA. The fusion features from multiple frequency bands for brain wave characterization provided an accurate understanding of developing an EEG-based ML algorithm for the neuro-prosthetics application. CSP and CCA were proposed for multiple-frequency band feature fusion. ANOVA and classification such as KNN, SVM, tree classifiers, and ensemble classifiers were used for the testing and validation stage. The remaining features and channels were statistical-based features of PSD and Pz-CPz’ (C26). CCA by summation and subspace KNN were the best combinations due to the accuracy, recall, and precision being 100.00%. The research outcome dictates the possibility of brain waves as an ultimate tool to restore lost motor functions. This finding contributes to an early understanding of the ML algorithm and proposes the revision of the robot arm especially on the number of control commands and the optimization of EEG features and channels for neuro-prosthetics applications. It is expected that this proposed ML algorithm can be extended to subject-independent tests with the help of hardware and the sensory reception can be included in the future.