Signal Processing and Machine Learning Approaches for frontal EEG-Based Mental Workload Assessment: From Protocol Design to Classification

Mental workload has become an important research topic, as increasing demands on multitasking and decision-making can impair performance and compromise safety in domains such as aviation, surgery, and driving. Understanding when and how workload arises enables the design of adaptive systems that monitor cognitive state and automate specific actions, improving safety and reducing fatigue. To investigate cognitive variations, psychophysiological measures such as ECG, respiration, electrodermal activity, and eye tracking are often used. In this work, frontal EEG was selected as the main modality, since brain electrical activity directly reflects cognitive processing. Eight frontal electrodes were used to minimize intrusiveness and simplify sensor placement. This thesis aimed to design an experimental protocol capable of eliciting distinct workload levels and to extract four feature groups (temporal, spectral, coherence, and entropy/non-linear) from the processed EEG, which were then used to train machine learning classifiers for automatic workload recognition. Nine task difficulty levels were defined within the Multi-Attribute Task Battery II (MATB-II) by manipulating event frequency. Eighteen participants completed a 5-minute rest phase followed by four MATB-II sessions combined with a secondary arithmetic task. Each session included five 2-minute task windows and a 20-second self-evaluation using the Bedford workload scale. EEG was recorded from eight frontal electrodes (F10, AF8, AF4, FP2, FP1, AF3, AF7, F9) with a g.HIAMP amplifier (g.tec). Two subjects were excluded due to corrupted data, yielding a final dataset of 16 participants. EEG pre-processing included resampling from 1200 Hz to 512 Hz, a 0.5–80 Hz band-pass filter and 50 Hz notch filter. Artifacts such as blinks and abrupt movements were removed using envelope subtraction and amplitude thresholding: samples exceeding three standard deviations, or 0.5 second windows around artifacts, were discarded based on signal quality. Each cleaned signal was segmented into 2, 3, and 4 second windows with 50% overlap. From each segment, 4 temporal, 15 spectral, 7 entropy/non-linear, and 168 coherence features were extracted. Four statistical indices (mean, variance, skewness, kurtosis) were computed for each feature, except coherence, where only mean and variance were used. Bedford ratings, collected on a 10-point subjective workload scale, were grouped into three levels (low, medium, high) and used as class labels for the machine learning models. A paired t-test (&#945;<0.05) confirmed significant differences in subjective workload between difficulty levels (easy–medium, medium–hard, easy–hard), validating the effectiveness of the experimental manipulation. A classification analysis was performed on normalized features to identify the optimal combination of feature type and window size. Each classifier was optimized using five-fold cross-validation, with 80% of the data used for training and 20% reserved for testing. In binary classification, both coherence features and entropy/non-linear features yielded the highest performance, with an F1 score of 92.46% and an accuracy of 98.51% on the test set. These results, consistent across multiple classifier–feature selector combinations, highlight signal complexity and inter-channel connectivity as key workload indicators. No consistent trend emerged for window length. In multiclass classification, performance predictably decreased due to class imbalance and subjective ratings used as class labels.