Multi-Level EEG Signal Analysis for Mental Workload Classification using Hybrid CNN-Transformer-BiLSTM Model

For adaptive human-machine interaction systems and real-time cognitive monitoring, accurate mental workload (MWL) classification is necessary. Reliable workload discrimination is significantly complicated by the fact that electroencephalogram (EEG) data are inherently non-stationary and susceptible to different types of noise and artifacts, despite being highly informative for underlying cognitive processes. In order to overcome these difficulties, this paper suggests a Multi-Level EEG-based MWL classification framework that uses a hybrid deep learning architecture that combines Convolutional Neural Network (CNN), a Transformer Encoder, and Bidirectional Long Short-Term Memory (BiLSTM) networks to differentiate between easy, medium, and high workload states. At various stages of the framework, Multi-Level workload-related features were extracted after EEG preprocessing steps such as bandpass filtering, windowing, and normalization. CNN layers were used to learn spatial characteristics, the Transformer Encoder network was used to capture global contextual interactions, and the BiLSTM network was used to model temporal dependencies. Class separation and overall classification performance were considerably enhanced by integrating these complimentary feature representations. The proposed model was implemented in MATLAB R2025b and the suggested model outperformed current baseline methods with an accuracy of 99.27%, specificity 99.64%, precision 99.27%, Matthew's correlation coefficient (MCC) 98.71%, Cohen's kappa 98.71%, and F1-score 99.27%.