Cognitive flexibility enables individuals to adapt to changing rules, goals, or uncertainty. This study evaluates the discriminative power of electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) signals from 42 healthy young adults during reversal learning tasks, using a comprehensive artificial intelligence (AI) framework. We compared classical machine learning models (e.g., Logistic Regression, Random Forest, XGBoost) and deep neural networks (MLP, GRU, BiLSTM), applied to EEG and fNIRS separately and in combination through multimodal fusion strategies. Interpretable machine learning played a significant role in our framework, with SHapley Additive exPlanations (SHAP) methodologies enabling transparent understanding of model predictions and feature importance across neurophysiological modalities. The interpretability fostered by SHAP enhances decision-making in cognitive assessments while cultivating deeper understanding of neurophysiological markers indicative of cognitive flexibility. Results show that EEG outperformed fNIRS in distinguishing "with objective (AO)" from "without objective (SO)" states. Logistic Regression achieved the best EEG performance (Accuracy = 0.810 ± 0.098, F1-score = 0.833 ± 0.089), while Random Forest was optimal for fNIRS (Accuracy = 0.647 ± 0.174, F1-score = 0.653 ± 0.228). Multimodal fusion strategies demonstrated varied effectiveness: early fusion with SVM achieved 74.17% ± 16.0% accuracy, intermediate fusion with XGBoost reached 75.3% ± 11.5%, while dynamic gating with instance-adaptive neural fusion provided dramatically superior performance at 95.00% ± 6.1% accuracy (95% Wilson CI: 0.8386, 0.9865)-representing a +14.00% improvement over best unimodal performance and +19.70% over static fusion methods. This exceptional accuracy was accompanied by 69.4% variance reduction compared to baseline approaches, demonstrating both superior classification capability and enhanced reliability. SHAP analysis revealed complementary modality roles: EEG demonstrated substantially higher gating importance (≈0.045) compared to fNIRS (≈0.028), with root mean square (RMS) features and spectral power emerging as most discriminative for EEG, while signal-to-noise ratio (SNR) features dominated fNIRS contributions. Train-test performance gap analysis (mean gap: 23.83%, range: 0.38%-40.91%) indicated generalization considerations requiring attention through expanded datasets and regularization strategies. These findings highlight the value of instance-adaptive multimodal AI integration for transparent identification of neurophysiological markers of cognitive flexibility. The dramatic performance advantage of dynamic gating over static fusion methods demonstrates that learned contextual modality weighting fundamentally surpasses fixed combination rules. However, the substantial train-test gaps underscore that these results establish methodological proof-of-concept in healthy young adults under controlled conditions, requiring extensive validation in heterogeneous populations, clinical cohorts, and naturalistic settings before clinical translation.
Mouazen et al. (Sun,) studied this question.