Systematic Literature Review of Machine Learning Methods for Emotion Recognition Using EEG and Physiological Signals in Healthcare

Summary

This review synthesizes 70 peer-reviewed studies (2014–2024) on emotion recognition from physiological signals — EEG, ECG, electrodermal activity (EDA), PPG/blood-volume pulse, and respiration — for healthcare uses like mental-health monitoring, stress assessment, and patient-adaptive support.

It organizes the field around four questions: (1) which physiological modalities and multimodal fusion strategies are used, (2) the most common machine learning and deep learning methods, (3) datasets, protocols, and metrics — carefully separating subject-dependent (SD) from subject-independent / leave-one-subject-out (SI/LOSO) evaluation, and (4) explainable AI (XAI) practices for clinical trust.

A consistent finding: SD results are higher than SI/LOSO, and domain adaptation plus multimodal fusion partially close the gap (e.g., ~3–5 points on DEAP for fusion over EEG alone, and larger gains on SEED-IV/SEED-V under LOSO). The review also maps barriers to clinical translation — artifacts and noise, inter-subject variability, limited cohort diversity, privacy constraints, and non-standardized interpretability — and outlines directions like protocol-aware benchmarking, cross-site validation, and clinically useful XAI.

Authors

Y. Abdollahinejad, S. M. Reza, B. Ashwini, A. Mousavi

Summary

Authors

Links