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

Researchers have developed CPTabKAN, a machine learning model that detects mild cognitive impairment from EEG sleep data by organizing features into physiologically meaningful concept groups and modeling their interactions. The approach achieved 90.38% F1-score, outperforming gradient boosting while maintaining interpretability—a critical advantage for clinical deployment where understanding model reasoning builds physician trust.

This research addresses a significant gap in clinical AI: the tension between predictive performance and interpretability in medical diagnostics. Early detection of mild cognitive impairment remains clinically challenging, and existing EEG-based approaches either rely on manual feature engineering that loses neurophysiological context or employ black-box deep learning that clinicians cannot trust or validate. CPTabKAN bridges this divide by anchoring model architecture to domain knowledge—organizing 1,379 EEG-derived features into ten physiologically motivated concept groups before processing.

The model's innovation lies in explicit interaction modeling through degree-2 polynomial transformation, revealing that cross-concept relationships between complexity metrics, demographics, and oscillatory patterns matter more than individual features. This hierarchical reasoning mirrors clinical intuition, where diagnosis depends on feature combinations rather than isolated measurements. The Study of Osteoporotic Fractures dataset (372 subjects with overnight polysomnography) provides robust real-world validation, with 10-fold cross-validation and SMOTE-balancing confirming generalization beyond synthetic conditions.

The clinical significance cannot be overstated: a 5.65 percentage point improvement over gradient boosting translates to fewer false negatives—patients incorrectly cleared for cognitive decline—and fewer false positives that trigger unnecessary follow-up. More importantly, concept importance rankings provide actionable insights for clinicians, surfacing which EEG markers drive decisions and how they interact. This work demonstrates that interpretable-by-design architectures need not sacrifice performance, establishing a template for regulated AI applications beyond neurology. Future deployment hinges on prospective validation and regulatory framework alignment, particularly FDA clearance pathways for AI-assisted diagnostics.

• →CPTabKAN achieves 90.38% F1-score in MCI detection while remaining interpretable, addressing clinical barriers to AI adoption in neurology
• →Interaction modeling reveals that cross-concept feature relationships drive predictions more than individual metrics, informing clinical decision-making
• →The framework organizes raw EEG data into domain-informed concept groups, reducing feature engineering burden and improving generalization
• →Ablation analysis confirms each architectural component contributes independently, validating the design rationale for regulated medical environments
• →Concept importance analysis surfaces physiologically coherent reasoning patterns that support physician trust and clinical deployment readiness