Explainable EEG-based machine learning for early diagnosis of Alzheimer's disease and frontotemporal dementia

The quick and accurate diagnosis of Alzheimer's disease (AD) and Frontotemporal Dementia (FTD) is a significant and unresolved challenge in clinical neurology, with early identification being crucial for prompt intervention and disease management. This study presents AutoSSM-ICA-EEG, an automated and interpretable framework that incorporates Fast Independent Component Analysis (FastICA) for artifact removal and relevant feature extraction, few-shot AutoML for efficient hyperparameter optimization and dynamic architecture adaptation, and State-Space Modeling (SSM) for temporal EEG dynamics, facilitating concurrent classification of dementia subtypes and regression of clinical severity. Three distinct datasets were utilized: (1) OpenNeuro ds006036 (open-eyes/photic stimulation EEG; AD, FTD, and Healthy Controls; n = 88), functioning as the primary tri-class benchmark; (2) OpenNeuro ds004504 (closed-eyes resting-state EEG; AD and HC; n = 65), employed for cross-condition generalization; and (3) ADFSU (resting-state EEG; AD and HC; external institution, UNESP Brazil), utilized for cross-institution transfer assessment. Rigorous subject-level leave-one-subject-out (LOSO) cross-validation was uniformly implemented throughout all four evaluation protocols, ensuring that all preprocessing, ICA decomposition, and model tuning were performed solely within each training fold to avert data leakage. In the primary tri-class benchmark, the model attained an accuracy of 87.5%, macro sensitivity of 87.5%, macro specificity of 95.4%, macro F1-score of 88.7%, and macro AUC of 0.99. The regression analysis of MMSE severity produced a R² of 0.81 ± 0.04, with a mean absolute error (MAE) of 1.64 ± 0.31 and a root mean square error (RMSE) of 1.71 ± 0.21. Cross-condition generalization (Closed→Open and Open→Closed) and cross-institution transfer (ADFSU↔OpenNeuro; AD sensitivity = 83.9%) further validate the framework's resilience beyond a singular dataset or recording situation. ICA-refined SHAP channel-level explainability identifies Ch11 and Ch13 as the principal neurophysiological biomarkers for subtype classification and MMSE severity regression, respectively, with Ch17 serving as the most sensitive predictor of disease duration - corroborating established frontotemporal EEG signatures of neurodegeneration. These findings establish AutoSSM-ICA-EEG as a potential, neurophysiologically interpretable paradigm for EEG-based dementia screening, subject to prospective multi-site clinical validation.