Reliability-aware biosignal ranking and selection framework using Trimodal Correlation (TRI-CORR) for adaptive mental fatigue assessment

Fatigue is a multidimensional construct associated with reduced productivity and adverse health outcomes. Existing studies fail to capture fatigue comprehensively and reliably, as they focus on isolated aspects of fatigue and rely on fixed biosignal configurations. This study conceptualizes fatigue into three aspects, namely temporal, spatial, and systemic aspects. As biosignals exhibit varying resilience to noise, a reliability-aware biosignals ranking and selection framework, termed Trimodal Correlation (TRI-CORR) was designed to assess mental fatigue adaptively. The framework begins with the acquisition and preprocessing of electroencephalography (EEG), electrocardiography (ECG), galvanic skin response (GSR), and photoplethysmography (PPG) data. Three types of correlations that quantify three aspects of fatigue across multimodal biosignals were captured and fed into adaptive classification module containing Light Gradient Boosting Machine (LightGBM). The framework adaptively weights and ranks the contribution of biosignals based on correlational data for reliable fatigue classification. Performance evaluation on custom dataset demonstrates its effectiveness, achieving 10-fold cross-validation accuracy of $$0.882\pm 0.021$$ , test accuracy of 0.757, test F1 of 0.777, and test ROC-AUC of 0.905. EEG and skin PPG contribute most significantly to the classification, while tree-based LightGBM demonstrates strong capability in learning correlation-derived features. Increasing the time window size, number of fatigue aspects and biosignals improve classification performance. In the cross-dataset validation on the publicly available Mental Fatigue Assessment through Neurophysiological and Biosignal Data (MEFAR) dataset against benchmarking approaches, the proposed framework demonstrates competitive performance, achieving 10-fold cross-validation accuracy of 0.9810, test accuracy of 0.9653, test F1 of 0.9650, and test ROC-AUC of 0.9959.