Identificación de episodios de estrés con señales EEG y EDA mediante modelos de machine learning

Abstract

Stress is a critical emotional state with significant impacts on physical health and cognitive performance. Its traditional assessment relies on subjective methods such as questionnaires, which are prone to bias and lack temporal precision. While physiological signals like electroencephalography (EEG) and electrodermal activity (EDA) offer objective measurements, their individual use faces technical limitations and interindividual variability. The multimodal integration of these signals through machine learning (ML) models emerges as a promising alternative for the reliable detection of stress episodes. This study implemented a multimodal approach integrating EDA, EEG, and electrocardiography (ECG) signals from 10 participants undergoing cognitive tasks and sleep deprivation. After preprocessing, which included temporal segmentation and feature extraction (e.g., EDA phasic components, EEG spectral powers), binary labels (stress/no stress) were generated using the unsupervised k-means algorithm. Three supervised classification models (Decision Tree, SVM, KNN) were trained and optimized using only EDA features, with performance evaluated using metrics such as accuracy, F1-score, and AUC-ROC. In the task with balanced classes, the models showed acceptable performance, with an AUC-ROC of up to 0.8119 (KNN). In the task with class imbalance (85% stress), a drop in real discriminative capacity was observed (AUC-ROC of 0.4862 for KNN), despite deceptively high global metrics (e.g., 82.35% accuracy). Phasic EDA features (mean SCR amplitude and count) were identified as the most determinant. It was demonstrated that phasic EDA features contain sufficient information for the identification of acute stress, providing a basis for the development of low-cost wearable systems. The research underscores the criticality of class balance in data and the need to use robust metrics such as AUC-ROC for reliable evaluation. This work lays the foundation for future developments of stress monitoring systems applicable in clinical and everyday environments.