Cognitive Load Estimation Using Brain Foundation Models and Interpretability for BCIs

Figure 1: Overall pipeline of cognitive load estimation with brain foundation model (BFM) in adaptive training systems.

We investigate the use of Brain Foundational Models (BFMs) for continuous cognitive-load monitoring and examine key challenges in scalability, generalization, interpretability. We apply BFMs to continuous cognitive-load estimation and analyze their behavior in a multi-day training setting. Our contributions are:

• A scalable and cross-participant pipeline for long-term cognitive load estimation using BFM-derived features.
• A flexible group-average channel alignment for heterogeneous layouts, improving cross-subject generalization
• An adaptation of Partition SHAP to interpret EEG feature and region importance, aligned with neuroscience [23].
• A longitudinal analysis across multiple days revealing learning progression w.r.t. cognitive load and other neural markers.

We find that cognitive load decreases over time while prefrontal neural relevance increases. Our results further show that BFMs, particularly LaBraM [18], improve estimation accuracy and consistently emphasize frontal regions linked to working memory and executive function, supporting their use in real-world cognitive monitoring.

Fig. 4: Topomap of SHAP feature relevance, averaged globally (left) and per day (right); red indicates high importance.
Daily averages of cognitive load (decreasing over time), focus stability (declining), blink duration (increasing over time) are also shown.

Fig. 4 (left) shows normalized global SHAP feature relevance, averaged across all days and folds, while Fig. 4 (right) presents daily
SHAP maps alongside behavioral metrics (cognitive load, focus stability, blink duration). LaBraM emphasizes frontal and prefrontal
regions linked to cognitive control and decision-making [28], as well as parieto-occipital areas linked to visual working memory [29, 30]. In contrast, CBraMod primarily highlights parieto-occipital regions, consistent with the visual nature of the task, but lacks prefrontal emphasis, which may explain its weaker performance. These patterns were consistent across all estimators (SVM, Linear, DNN), underscoring the robustness of the extracted features and SHAP’s model-agnostic design. The longitudinal design, with participants returning over multiple days, enables analysis of temporal trends in cognitive load and training progression. Fig. 4 (right) presents daily SHAP relevance maps alongside behavioral metrics of focus stability and blink duration extracted from Varjo VR3. LaBraM shows increasing prefrontal relevance over time, while cognitive load decreases from 0.90 to 0.64 and blink duration increases from 0.19 to 0.23. This suggests participants became more proficient and cognitively efficient over time, while (potentially) more relaxed. These trends align with behavioral indicators such as reduced focus stability and longer blink durations, supporting the neurophysiological validity of LaBraM’s feature representations [31]. In summary, prefrontal and parietal channels dominate SHAP relevance, with prefrontal emphasis differences explaining LaBraM and CBraMod performance gaps.

Abstract: Accurately monitoring cognitive load in real time is critical for Brain-Computer Interfaces (BCIs) that adapt to user engagement and support personalized learning. Electroencephalography (EEG) offers a non-invasive, cost-effective modality for capturing neural activity, though traditional methods often struggle with cross-subject variability and task-specific preprocessing. We propose leveraging Brain Foundation Models (BFMs), large pre-trained neural networks, to extract generalizable EEG features for cognitive load estimation. We adapt BFMs for long-term EEG monitoring and show that fine-tuning a small subset of layers yields improved accuracy over the state-of-the-art. Despite their scale, BFMs allow for real‑time inference with a longer context window. To address often-overlooked interpretability challenges, we apply Partition SHAP (SHapley Additive exPlanations) to quantify feature importance. Our findings reveal consistent emphasis on prefrontal regions linked to cognitive control, while longitudinal trends suggest learning progression. These results position BFMs as efficient and interpretable tools for continuous cognitive load monitoring in real-world BCIs.