Source-linked AI summary
NeuroStrata: An Electroencephalographic Connectivity-Aware Deep Representation Learning Framework for Dynamic Brain Network Analysis of Mental Stress
Sayantan Acharya, Hamzeh Asgharnezhad, Abbas Khosravi, Douglas Creighton, Roohallah Alizadehsani, U Rajendra Acharya
TL;DR
Mental stress assessment remains limited by subjective questionnaires and the distributed, dynamic nature of the brain’s stress response. NeuroStrata models time-varying directed EEG connectivity with deep representations, achieving 97.3% accuracy for beta-PDC using LAION-CLIP-ViT-L14 and SVM.
Problem
Subjective questionnaires provide insufficiently comprehensive stress assessment because stress involves dynamic, distributed brain processes and individual perception varies substantially.
Method
NeuroStrata combines TV-PDC estimation of dynamic directed EEG connectivity with pretrained deep embeddings and lightweight machine-learning classifiers.
Results
97.3% accuracy was achieved by beta-PDC using the LAION-CLIP-ViT-L14 feature extractor with an SVM, while alpha-PDC remained stable across configurations.
Takeaways & Limitations
Beta connectivity formed distributed frontal–central–parietal networks, while classification performance improved and stabilized in central temporal windows as task engagement increased.
Takeaways & Limitations
TV-PDC estimates sensor-space electrode influences rather than source-localized causation, and the SAM 40 study population and offline setting limit broader real-time applicability.
Abstract
from arXiv · showhide
This study introduces NeuroStrata, a connectivity-aware deep representation learning framework for EEG-based mental stress analysis using Time-Varying Partial Directed Coherence (TV-PDC). Unlike conventional EEG classification approaches based on static features, NeuroStrata models the temporal evolution of frequency-specific directed connectivity across distributed brain regions. EEG signals from the 32-channel SAM 40 dataset recorded during mental arithmetic tasks were used to generate TV-PDC connectivity maps. These maps were processed using pretrained Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to extract deep connectivity embeddings, which were subsequently classified using lightweight machine learning models. Experimental results demonstrate that beta-band connectivity provides the highest discriminative capability, achieving a peak accuracy of 97.3% using the LAION-CLIP-ViT-L14 backbone with a Support Vector Machine classifier, while alpha-band connectivity exhibits consistently stable performance across model configurations. Connectivity analysis revealed prominent frontal-driven alpha influences and centrally integrated beta connectivity patterns associated with stress-related neural dynamics. Temporal evaluation further indicated that classification performance stabilizes in mid-to-late temporal windows, suggesting progressive consolidation of stress-related connectivity signatures. The proposed framework integrates time-varying effective connectivity modelling with deep representation learning to provide an interpretable and automated approach for EEG-based mental stress analysis.
1. Introduction
The introduction frames mental stress as a dynamic, multi-system neural response that subjective questionnaires cannot comprehensively or consistently measure. It motivates NeuroStrata, a connectivity-aware framework combining time-varying directed EEG connectivity, deep representations, lightweight classifiers, interpretability, and temporal analysis.
- Motivation: Mental stress involves progressive cortical, limbic, autonomic, and psycho-physiological processes during cognitively demanding circumstances.The response is described as a dynamic fight-or-flight process involving multiple interacting neural and physiological systems.
- Problem: Questionnaire-based assessment is subjective and variable across individuals, limiting accurate and comprehensive stress-level classification.Individual perception, emotional awareness, and situational context can produce inaccurate or inconsistent measurements.
- Motivation: Objective physiological markers, including EEG, are therefore integrated with subjective assessments to capture neural dynamics and rapid stress-related physiological changes.The introduction specifically identifies EEG, heart rate variability, skin conductance, and salivary measures as objective markers.
- Contribution: NeuroStrata integrates time-varying effective connectivity estimation, deep neural representations, and lightweight machine-learning classifiers for EEG-based stress analysis.Its connectivity-aware pipeline uses TV-PDC connectivity estimation and extracts informative embeddings from directed EEG connectivity patterns.
- Contribution: The framework uses pretrained CNNs and ViTs on TV-PDC maps, analyzes influential directed connections, and evaluates temporal segmentation across windows T1-T7.These analyses target discriminative representations, stress-modulated information-flow pathways, and temporal stability of connectivity-based classification.
2. Materials and methods
The study used SAM 40 multichannel EEG recordings to examine time-varying brain connectivity under stress. MAT trials were organized by stress level and analyzed using participant responses and dataset-defined EEG bands.
- Dataset and recordings: SAM 40 provided multichannel EEG recordings acquired during task execution, with corresponding participant responses and three stress levels: Relaxed, Low Stress, and High Stress.The EEG bands used were consistent with those defined in the SAM 40 dataset.
- Dataset and recordings: The dataset included 40 participants—26 males and 14 females—with a mean age of 21.5 years, and EEG sampled at 128 Hz using 32 channels.Participants performed Symmetric Mirror Image recognition, the Stroop Color–Word Task, and the Mental Arithmetic Task.
- Mental Arithmetic Task: Time-varying connectivity analysis used MAT-based EEG recordings because mental arithmetic reliably elicits stress-related neural changes in executive, attentional, and cognitive-control regions.During trials, participants mentally solved arithmetic problems and indicated whether displayed solutions were correct using a thumbs-up response.
- Stress labeling: MAT trials were labeled Low Stress or High Stress using self-reported stress ratings on a 10-point Likert scale, selecting each participant’s lowest- and highest-rated trials.Each participant contributed three MAT trials; separate Relaxed-state EEG recordings were acquired before each MAT session.
3. NeuroStrata
NeuroStrata is a multi-stage framework that models time-varying, frequency-specific directed connectivity from multichannel EEG during mental-stress processing. It combines TV-PDC representations with pretrained CNN and ViT feature extraction and PCA-based dimensionality reduction.
- Connectivity construction: NeuroStrata segments multichannel EEG into short-time windows and computes frequency-specific TV-PDC maps that encode directed causal information flow across channels.The implementation uses 5-second windows, 2-second steps, and a fixed 128 Hz sampling frequency.
- Deep representation learning: Pretrained CNNs and ViTs extract complementary deep features from TV-PDC maps, capturing localized connectivity patterns and global brain-network dependencies.VGG16 provides a stable spatial feature extractor, EfficientNet-V2 supports multiscale extraction, and LAION-CLIP-ViT-L14 supplies high-capacity global embeddings.
- Connectivity construction: PDC quantifies the magnitude and direction of frequency-specific causal interactions using an MVAR model of multichannel EEG signals.Sensor-level PDC captures directed information flow between EEG electrodes, while model order 5 was selected using multiple statistical criteria.
- Connectivity representation: The resulting TV-PDC matrices form temporal sequences of effective brain-network representations associated with mental-stress processing across successive windows.These representations encode pairwise causal interactions and evolving large-scale network dynamics.
- Feature reduction: PCA retains components explaining 95% of cumulative variance to compact deep features, reduce redundancy, and improve classifier stability.To prevent leakage, PCA is fitted within training folds and then applied to the corresponding test data.
4. Classification results
Classification performance varied substantially by frequency band and model combination: beta-PDC was most discriminative, while alpha-PDC provided stable high performance. Beta-PDC with LAION-CLIP-ViT-L14 and SVM reached 97.3% accuracy, with near-perfect class separability in the confusion matrix.
- Alpha-PDC: Alpha-PDC delivered consistently strong performance, including 79.2% accuracy with CLIP-ViT-B/16 and SVM and 81.0% accuracy with CLIP-ViT-B/32 and SVM.For CLIP-ViT-B/16, SVM also achieved 80.6% precision, 79.2% recall, and 79.2% F1 score.
- Beta-PDC: Beta-PDC showed the strongest discriminative behavior across evaluated bands, with LR reaching 91.2% accuracy for CLIP-ViT-B/16 and 89.3% accuracy for CLIP-ViT-B/32.For CLIP-ViT-B/16, LR achieved 91.3% precision, 91.2% recall, and 91.2% F1.
- Beta-PDC: 97.3% accuracy was achieved by beta-PDC with LAION-CLIP-ViT-L14 and SVM, alongside 98.3% precision, 97.0% recall, and 97.3% F1 score.This 97.3% accuracy was the mean across subject-wise 10-fold cross-validation folds.
- Beta-PDC: 100% accuracy was obtained for Relaxed and High Stress, while Low Stress reached 98.1% in the beta-PDC confusion matrix.The minimal confusion was primarily between Relaxed and Low Stress.
- Other frequency bands: Delta-, theta-, and gamma-PDC generally performed below alpha- and beta-PDC, although delta-PDC remained stable and theta-PDC outperformed delta-PDC across backbones.For gamma-PDC with CLIP-ViT-B/16, SVM achieved 74.2% accuracy, compared with 68.6% for RF and 68.8% for XGB.
5. Analysis of important TV-PDC features
XGB importance analysis highlighted alpha- and beta-PDC connections above a 0.9 threshold as interpretable stress-related features. Alpha connectivity was predominantly frontal-driven and lateralized, whereas beta connectivity involved fronto-central and central networks, with identified patterns corresponding to physiological pathways.
- Feature-importance selection: An importance score threshold of 0.9 selected the most relevant alpha-PDC and beta-PDC features for analyzing dynamic directed neural influences.The threshold focused the analysis on bands described as having the highest discriminative power for stress classification.
- Alpha-PDC organization: Alpha-PDC connections were dominated by frontal and fronto-central drivers, including Fp1, Fp2, F3, F4, F7, F8, Fz, FC1, and FC2.These frontal electrodes were described as major sources of directed information flow regulating posterior cortical dynamics.
- Alpha-PDC organization: Alpha-PDC exhibited lateralized projections, with left frontal electrodes influencing left posterior sites and right frontal electrodes influencing right parietal and occipital regions.Examples included Fp1 and F3 projecting toward P7, O1, and PO9, and Fp2, F4, and F8 influencing P8, O2, and PO10.
- Beta-PDC organization: Beta-PDC prominently involved fronto-central, central, and frontotemporal regions, implicating sensorimotor and executive control networks.The reported pattern contrasted with alpha-PDC and included electrodes such as Fz, FC1, FC2, C3, and C4.
- Physiological correspondence: 11 of 21 discriminative beta-PDC connections were retained among the top 10% strongest directed interactions.This overlap supported correspondence between XGB-identified beta connections and physiologically pronounced pathways; alpha analysis likewise linked selected connections to physiological connectivity, while other connections remained discriminative for class separation.
- Stress-related connectivity: Frontal-driven alpha and fronto-central or central-to-posterior beta interactions were consistently observed across classifiers and modeling strategies as robust stress-related connectivity markers.The study associated alpha patterns with cortical regulation and beta patterns with task engagement and sensorimotor stabilization under stress.
6. Impact of temporal windows
Classification performance generally rises across temporal windows, with consistent peaks in T4–T6 across diverse deep-learning and machine-learning pipelines. Early-window underperformance is attributed to transient adaptation, while later connectivity patterns become more pronounced and temporally coherent.
- Temporal performance trends: Accuracy increases from early segments toward central-to-late windows, with peak performance consistently occurring in T4 to T6 across ViT and CNN backbones.The analysis evaluated seven temporal segments, T1 to T7, across all DL-ML combinations.
- Temporal performance trends: Lower performance in T1 and T2 likely reflects transient neural adaptation and task familiarization before connectivity patterns fully differentiate across stress levels.
- Temporal performance trends: As the task progresses, fronto-parietal alpha and fronto-central beta interactions become more pronounced and temporally coherent, improving classification performance.Peak accuracies converge around T5 to T6 across diverse DL-ML pipelines.
- Convergence-group analysis: The top 20 high-performing DL-ML pipelines were selected from the global window-wise comparison and stratified into 5 convergence groups for detailed investigation.Selection was based on overall classification performance across temporal windows.
- Convergence-group analysis: Across the 5 hybrid DL–ML groups, accuracy generally improves from T1 to T3 toward T4 to T6, followed by mild stabilization or decline at T7.Most pipelines show uniform performance peaks within T4 to T6 despite variations in backbones and classifiers.
7. Discussion
The discussion positions NeuroStrata as a layered framework combining time-varying, frequency-specific directed connectivity with deep representations and lightweight classifiers for mental-stress analysis. Results emphasize beta-PDC discrimination, frontal-to-posterior alpha organization, temporal consolidation, and limitations related to sensor-space estimation and dataset scope.
- Framework contribution: NeuroStrata integrates dynamic effective-connectivity estimation, deep neural representations, and lightweight machine-learning classifiers into a hybrid DL–ML stress-analysis pipeline.The framework models time-varying PDC to represent evolving information flow rather than static connectivity snapshots.
- Neurophysiological interpretation: Alpha-PDC showed frontal-driven influences toward parietal and occipital regions, whereas beta-PDC formed an integrative frontal→central→parietal network with prominent central-hub involvement.The alpha organization suggests regulatory control of posterior sensory regions, while the beta structure reflects distributed cognitive engagement and vigilance.
- Temporal dynamics: Classification accuracy improved from early to central windows, peaked consistently in T4–T6, and gained no additional discriminative advantage from extended aggregation at T7.These findings indicate that connectivity embeddings stabilize after adequate temporal consolidation and that optimal temporal span matters across DL–ML configurations.
- Frequency-band findings: Beta-PDC exhibited the strongest discriminative improvement across CNN models, alpha-PDC remained relatively separable, and performance declined for delta-, gamma-, and especially theta-PDC.The discussion notes that prior stress studies have reported inconsistent frequency-band effects, motivating frequency-specific connectivity analysis.
- Limitations and future work: The framework’s limitations include sensor-space TV-PDC estimation, restriction to 35 young neurologically healthy SAM 40 participants, and evaluation in a controlled offline environment.Further investigation is required for real-time monitoring across diverse stress-inducing conditions.
8. Conclusion
NeuroStrata is an EEG connectivity-driven deep representation learning framework that uses dynamic, frequency-specific effective connectivity from TV-PDC for stress quantification. Directional temporal connectivity improves stress discrimination over conventional static connectivity approaches, with beta-PDC showing the strongest performance among EEG bands.
- Framework: NeuroStrata uses dynamic effective connectivity derived from TV-PDC for EEG-based stress quantification.The framework applies deep representation learning to connectivity-driven EEG analysis.
- Findings: Directional and frequency-specific temporal connectivity significantly improves stress discrimination compared with conventional static connectivity approaches.
- Findings: Beta-PDC exhibits the strongest performance among the evaluated EEG bands.
CRediT authorship contribution statement
The authors contributed across writing, conceptualization, visualization, software, methodology, supervision, analysis, data curation, resources, validation, and project administration.
- Sayantan Acharya contributed to writing, conceptualization, visualization, validation, software, resources, and methodology.
- Abbas Khosravi contributed to writing, conceptualization, visualization, supervision, formal analysis, and data curation.
- Hamzeh Asgharnezhad contributed to visualization, software, and methodology, while Douglas Creighton contributed to writing, project administration, and formal analysis.
- Roohallah Alizadehsani contributed to writing, supervision, resources, and conceptualization.
Funding
The study received no funding.
- No funding was received for this study.