PhD Scientific Days 2025

Neurosciences

Multimodal Stress Marker Recognition Using Artificial Neural Networks: Toward Real-Time, Personalized Stress Assessment

Name of the presenter

Bod Réka

Institute/workplace of the presenter

HUN-REN Research Centre for Natural Sciences

Authors

Réka Bod¹, László Grand², Lucia Wittner¹

1: HUN-REN Research Centre for Natural Sciences
2: Polaritás-GM, Pázmány Péter Catholic University, Faculty of Information Technology and Bionics

Text of the abstract

Introduction
Stress is a complex psychophysiological condition that affects well-being and long-term health. Objective stress detection through physiological signals holds great promise for clinical, occupational, and sports medicine applications, enabling adaptive, real-time mental state assessment.
Aims
This study aims to develop a multimodal artificial neural network (ANN) framework for accurate stress marker recognition using EEG, ECG, respiratory activity, and validated self-report data. The model is designed to integrate heterogeneous inputs and generalize under limited-sample conditions.
Method
Data were collected from 29 participants across two sessions: one control and one stress-inducing, randomly assigned. Each session involved Stroop, mental arithmetic, and Tetris tasks. The stress session introduced strict time limits, negative feedback, and randomly “sticky” arrow keys to increase cognitive-emotional load. Signals recorded included 32-channel EEG, ECG, respiratory effort and frequency, and self-reported STAI and PANAS scores. The ANN, implemented in PyTorch, includes dual EEG branches (a Temporal Convolutional Network and a Spatial CNN), GRU layers for ECG and respiration, and a multimodal fusion module with attention mechanisms to weigh modality importance dynamically.
Results
Preliminary results indicate distinct stress-related changes: involving EEG power spectra changes, reduced heart rate variability, elevated respiratory rate, and significant differences in questionnaire scores. These findings validate the effectiveness of the stress paradigm and the utility of combined physiological and psychological data.
Conclusion
The proposed ANN architecture shows promising prospects for real-time, automated stress classification across diverse settings. Final training and optimization of the model are ongoing and expected to be optimized by the beginning of the conference. Future directions include real-world deployment in wearable systems and integration into digital mental health tools.
Funding
This research is supported by the EKÖP-KDP scholarship program and the Semmelweis 250+ Excellence PhD Scholarship.