Interpretable Concept-Guided Polynomial Tabular Kolmogorov-Arnold Network for EEG-Based Mild Cognitive Impairment Detection

This study introduces CPTabKAN, a novel machine learning framework designed to address the limitations of existing EEG-based mild cognitive impairment (MCI) screening tools. Traditional models often treat heterogeneous EEG features as an undifferentiated mass, losing neurophysiological context, or rely on opaque deep learning architectures that lack clinical interpretability. CPTabKAN overcomes these issues by partitioning 1,379 EEG-derived features into ten physiologically motivated concept groups (e.g., power spectral density, sleep spindles, and slow oscillations). These groups are processed through a concept bottleneck, expanded via degree-2 polynomial transformations to capture pairwise interactions, and classified using a Fourier-parameterized Kolmogorov-Arnold Network (TabKAN).

Evaluated on the Study of Osteoporotic Fractures (SOF) cohort of 372 subjects, CPTabKAN achieved a weighted F1-score of 0.9038, significantly outperforming standard tabular classifiers like GradientBoosting. The ablation analysis confirmed that both the concept-structured bottleneck and the explicit second-order interaction modeling contributed independently to the model's predictive accuracy. Furthermore, the framework provided high-level interpretability: importance analysis revealed that while first-order features like power spectral density and entropy were predictive, cross-concept interactions—specifically those involving Lempel-Ziv-Welch complexity and slow oscillations—were even more critical for identifying MCI, aligning with established neurophysiological theories of cognitive decline.

By bridging the gap between high-performance deep learning and clinical transparency, CPTabKAN offers a more trustworthy tool for early MCI detection. Its ability to surface specific physiological interactions as drivers of its predictions allows clinicians to understand the reasoning behind a classification, potentially facilitating earlier intervention and better-informed clinical decision-making in community health settings.