PhD Scientific Days 2026

Neurosciences

Toward Personalized Stress Assessment: A Multimodal EEG–ECG Machine Learning Pipeline

Name of the presenter

Bod, Réka

Institute/workplace of the presenter

HUN-REN RCNS

Authors

Réka Bod¹, László Grand², Lucia Wittner³
1: HUN-REN RCNS, Semmelweis University, Szentágothai János Neuroscience Division
2: Pázmány Péter Catholic University, Faculty of Information Technology and Bionics
3: HUN-REN RCNS

Text of the abstract

Introduction
Reliable stress assessment is challenging because physiological responses vary widely across individuals and reflect both current state and stable traits such as anxiety. Integrating brain activity (EEG), cardiac signals (ECG), and psychometrics with machine learning may enable more robust and personalized stress detection.
Aims
This study aimed to: (1) develop a reproducible multimodal processing pipeline integrating EEG, ECG, and psychometric data; (2) identify physiological and behavioral markers associated with stress; and (3) evaluate machine learning models for generalizable classification of stress states across individuals.

Methods
29 participants completed control and stress sessions across three tasks (Stroop, Tetris, mental arithmetic). EEG and ECG were recorded alongside STAI and PANAS. EEG was preprocessed with artifact handling, features including bandpower, relative power, entropy, and Hjorth measures; ECG yielded heart rate and HRV metrics. Models were trained using subject-wise cross-validation and evaluated with balanced accuracy across EEG-only and combined feature sets.

Results
Psychometric measures revealed structured individual differences, with higher anxiety associated with elevated heart rate and broadly reduced HRV. EEG analyses identified consistent stress-related changes in brain activity, particularly in alpha and theta frequency bands. Machine learning models based on EEG features alone achieved moderate performance in distinguishing stress from control conditions (balanced accuracy ~0.74). Performance improved substantially when incorporating ECG and psychometric features, with multimodal models reaching balanced accuracies of approximately 0.84–0.86. These findings demonstrate that combining complementary physiological and behavioral information enhances stress classification and better captures individual variability.

Conclusion
Multimodal integration improves stress assessment and highlights the importance of individual differences. The proposed pipeline supports interpretable, personalized stress modeling and future real-time applications.

Funding
Semmelweis 250+ Scholarship for Excellence and supported by the 2024-2.-1.-2-EKÖP-KDP-2024-00002 University Research Scholarship Programme of the Ministry for Culture and Innovation from the source of the National Research, Development and Innovation Fund.