Deep Learning and Dynamical Modeling Frameworkfor EEG-Based Cognitive State Evolution inBrain-Computer Interfaces

Umair, Muhammad Khurram; Khawaja, Ayesha Arif; Abrar, Muhammad Faisal; Ali, Sikandar; Lee, It Ee; Jan, Salman · 2026 · Crossref

DOI: 10.21203/rs.3.rs-9208513/v1

archive: archived pipeline: cataloged

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical limitation of current deep learning models in Brain-Computer Interfaces (BCIs), which often lack temporal consistency and mechanistic interpretability, hindering their clinical applicability. To bridge the gap between data-driven pattern recognition and physiological understanding, the authors propose a unified framework integrating Long Short-Term Memory (LSTM) networks with a three-state compartmental Ordinary Differential Equation (ODE) model. The framework uses a probabilistic coupling mechanism where LSTM classification probabilities dynamically modulate ODE transition rates between Active, Passive, and Fatigued cognitive states, enabling both accurate detection and interpretable modeling of state evolution. The study evaluated this framework using the OpenNeuro ds004148 dataset, comprising 60 participants and 61 EEG channels sampled at 500 Hz. Data were preprocessed into 256-sample sequences and split using stratified temporal cross-validation to prevent leakage. The integrated LSTM-ODE model achieved 85.57% accuracy (95% CI: 85.19–85.93%) and an F1-score of 0.854, significantly outperforming traditional baselines such as Random Forest (82.29%) and XGBoost (80.92%). Statistical tests confirmed medium effect sizes for these improvements. Crucially, the ODE component did not compromise classification accuracy compared to a standalone LSTM-Attention model, while substantially enhancing temporal consistency. ODE-smoothed predictions reduced state oscillation frequency by 63.2% and increased mean state duration from 1.18 to 3.21 seconds, aligning better with physiological timescales. Explainability analyses using gradient-based attribution, permutation importance, and SHAP identified anterior frontal channels (AF7, AF3, Fp1) as primary contributors, consistent with known neurocognitive correlates of attention. The learned ODE parameters revealed physiologically meaningful dynamics: the Passive-to-Fatigued transition was the fastest (time constant 3.33 seconds), indicating rapid vigilance decrement, whereas full recovery from fatigue was much slower (time constant 100 seconds). A fatigue accumulation ratio of 0.91 highlighted strongly asymmetric dynamics where fatigue-directed transitions dominate over recovery. Cross-subject generalization tests showed graceful performance degradation, maintaining over 77% accuracy even when 30% of subjects were held out. This work establishes a principled approach for combining deep learning with mechanistic dynamical modeling in BCIs. By providing clinically meaningful transition rate interpretations, the framework enables proactive safety interventions based on forecasted state trajectories rather than reactive responses. The findings suggest that integrating interpretable ODE dynamics with deep learning can enhance the reliability and trustworthiness of EEG-based cognitive monitoring systems in clinical and assistive technology applications.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.8-27b-gittensor on 2026-08-23; verification: pending re-verification.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices
• drowsiness detection algorithms