Enhancing Brain–Machine Interface Applications for Disabled Individuals Using Visual Evoked Potentials and Explainable AI

Resumo

Individuals with severe physical disabilities often rely on Brain Machine Interfaces (BMI) to interact with their environment and regain some level of independence. One promising approach for BMI control in the use of Visual Evoked Potentials (VEP), which are electrical signals created in the brain in response to visual stimuli. However, the interpretation of VEP signals, is complex, presenting a barrier to the widespread adoption of VEP based BMIs. Therefore, in this study, a comprehensive model is proposed to boost the performance of VEP-based BMIs through the incorporation of advanced signal processing techniques and explainable Artificial Intelligence (AI) methods. The pre-processing of Electroencephalogram (EEG) signals plays a vital part in optimizing the performance of VEP based BMIs. To achieve this, an Orthogonal Wavelet Filter is employed, which offers extraordinary flexibility in managing the conservation of appropriate emotional data and the diminution of annoying interference. Next, to captivate the time-frequency representations of EEG signals, the Empirical Mode Decomposition (EMD) based Feature Extraction approach is used, which provides a data-driven approach to decay EEG signals into Intrinsic Mode Functions (IMFs), allowing for the extraction of relevant features related to brain activity. For prediction tasks, such as classifying intended movements or commands from EEG signals, a Modified Recurrent Neural Network with Long Short-Term Memory (RNN-LSTM) classifier is employed. The hyper parameters of the Modified RNN-LSTM model are tuned by employing Chaotic Bat Optimization (CBO), which effectively explores the hyperparameter space and optimizes model performance with rapid computational time. The approach is implemented through the python software and the performance metrics such as classification accuracy, recall, precision, and F1 score are compared with the traditional approaches to evaluate the effectiveness of the methodology. From the obtained outcomes, it is evident that the proposed Modified RNN-LSTM approach achieves high accuracy of (96%), thereby enhancing the daily functioning of disabled individuals.