A Novel Approach to Cognitive Load Measurement in N-Back Tasks Using Wearable Sensors, Empirical Mode Decomposition With Machine Learning, and Explainable AI for Feature Importance

Abstract

This study introduces a novel approach for detecting mental workload and stress, utilizing ECG and Fingertip-PPG data from the MAUS dataset. The dataset includes physiological recordings such as ECG, Fingertip-PPG, Wrist-PPG, and GSR signals from 22 participants exposed to varying levels of mental workload and stress conditions, serving as a valuable resource for developing mental workload assessment systems. Focusing on ECG and Fingertip-PPG signals, we evaluated their effectiveness in detecting mental workload and stress. Early detection of mental workload and stress is vital for preventing adverse health outcomes, emphasizing the importance of timely intervention. In this study, by applying Empirical Mode Decomposition (EMD) for feature extraction and machine learning classifiers, we achieved high performance with accuracy rates of 88.64% for ECG and 80.68% for Fingertip-PPG Signal. Combining both data types (ECG and Fingertip-PPG data) further improved recall to 94.87% and boosted overall accuracy to 87.50%, marking an improvement of approximately 13%. SHAP analysis identified key features contributing to this performance, revealing key features such as the waveform length of the first Intrinsic Mode Function (IMF) for the ECG signal, the mean of the second IMF for fingertip-PPG, and the waveform length of the second IMF for combined ECG and fingertip-PPG signals. These results highlight the potential of combining physiological signals and EMD with machine learning for detecting mental workload and stress.

Keywords

• Physiological signals
• electrocardiogram (ECG) signal
• photoplethysmography (PPG) signal
• machine learning
• empirical mode decomposition
• mental workload