Stress is a major factor affecting mental health and cognitive performance, yet its reliable detection in real time remains challenging. This thesis investigates EEG-based stress monitoring with a focus on real-time suitability, interpretability, and practical integration into brain–computer interface (BCI) systems. Using the SAM40 dataset, EEG signals were preprocessed with a lightweight, causal pipeline and segmented into short overlapping epochs. Importantly, classifications were performed at the epoch level rather than on averaged trials, allowing a more fine-grained and dynamic assessment of stress compared to many previous works. Features spanning time, frequency, entropy, Hjorth parameters, ratios, and asymmetry indices were extracted, followed by hierarchical feature selection using Kruskal–Wallis (KW) and Minimum Redundancy Maximum Relevance (mRMR). Multiple classifiers were evaluated, with the Support Vector Machine (SVM) achieving the most robust and efficient performance. Feature analyses consistently highlighted frontal theta and beta activity, parietal alpha suppression, and selected ratio features as reliable neurophysiological markers of stress. Channel selection further reduced the electrode set from 32 to 14 without loss of performance, enhancing feasibility for wearable systems. A proof-of-concept demonstration in OpenViBE validated that stress-related features, such as the theta/beta ratio, can be computed and streamed in real time, confirming the practical readiness of the proposed framework for real-time integration into BCI environments and in the later steps, for closed-loop adaptive feedback systems. Overall, this work delivers a practical and interpretable solution for EEG-based stress detection, bridging the gap between offline analysis and real-world applications.
Lo stress è un fattore determinante che influisce sulla salute mentale e sulle prestazioni cognitive, ma la sua rilevazione affidabile in tempo reale rimane una sfida. Questa tesi indaga il monitoraggio dello stress basato su EEG con particolare attenzione all’idoneità per l’elaborazione in tempo reale, all’interpretabilità e all’integrazione pratica nei sistemi brain–computer interface (BCI). Utilizzando il dataset SAM40, i segnali EEG sono stati pre-processati con una pipeline leggera e causale e segmentati in brevi epoche sovrapposte. È importante sottolineare che le classificazioni sono state eseguite a livello di epoca e non come media delle epoche, consentendo una valutazione dello stress più fine e dinamica rispetto a molti studi precedenti. Sono state estratte caratteristiche nei domini temporale, frequenziale, entropico, parametri di Hjorth, rapporti e indici di asimmetria, seguite da una selezione gerarchica delle feature mediante Kruskal–Wallis (KW) e Minimum Redundancy Maximum Relevance (mRMR). Sono stati valutati diversi classificatori, con la Support Vector Machine (SVM) che ha ottenuto le prestazioni più robuste ed efficienti. Le analisi delle feature hanno costantemente evidenziato l’attività frontale nelle bande theta e beta, la soppressione alfa parietale e alcuni rapporti selezionati come marcatori neurofisiologici affidabili dello stress. La selezione dei canali ha inoltre ridotto l’insieme degli elettrodi da 32 a 14 senza perdita di prestazioni, migliorando la fattibilità per sistemi indossabili. Una dimostrazione proof-of-concept in OpenViBE ha validato che feature legate allo stress, come il rapporto theta/beta, possono essere calcolate e trasmesse in tempo reale, confermando la prontezza pratica del framework proposto per l’integrazione in tempo reale in ambienti BCI e, in fasi successive, per sistemi di feedback adattivo in closed-loop. Nel complesso, questo lavoro propone una soluzione pratica e interpretabile per la rilevazione dello stress basata su EEG, colmando il divario tra analisi offline e applicazioni nel mondo reale.
Design and evaluation of an explainable EEG-based stress monitoring framework: towards real-time integration
Ghaffari Elkhechi, Nastaran
2024/2025
Abstract
Stress is a major factor affecting mental health and cognitive performance, yet its reliable detection in real time remains challenging. This thesis investigates EEG-based stress monitoring with a focus on real-time suitability, interpretability, and practical integration into brain–computer interface (BCI) systems. Using the SAM40 dataset, EEG signals were preprocessed with a lightweight, causal pipeline and segmented into short overlapping epochs. Importantly, classifications were performed at the epoch level rather than on averaged trials, allowing a more fine-grained and dynamic assessment of stress compared to many previous works. Features spanning time, frequency, entropy, Hjorth parameters, ratios, and asymmetry indices were extracted, followed by hierarchical feature selection using Kruskal–Wallis (KW) and Minimum Redundancy Maximum Relevance (mRMR). Multiple classifiers were evaluated, with the Support Vector Machine (SVM) achieving the most robust and efficient performance. Feature analyses consistently highlighted frontal theta and beta activity, parietal alpha suppression, and selected ratio features as reliable neurophysiological markers of stress. Channel selection further reduced the electrode set from 32 to 14 without loss of performance, enhancing feasibility for wearable systems. A proof-of-concept demonstration in OpenViBE validated that stress-related features, such as the theta/beta ratio, can be computed and streamed in real time, confirming the practical readiness of the proposed framework for real-time integration into BCI environments and in the later steps, for closed-loop adaptive feedback systems. Overall, this work delivers a practical and interpretable solution for EEG-based stress detection, bridging the gap between offline analysis and real-world applications.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_10_GhaffariElkhechi_Thesis_01.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Thesis
Dimensione
19.28 MB
Formato
Adobe PDF
|
19.28 MB | Adobe PDF | Visualizza/Apri |
|
2025_10_GhaffariElkhechi_Executive_Summary_02.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary
Dimensione
1.68 MB
Formato
Adobe PDF
|
1.68 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/243415