SOFEx: A Self-Organized Feature Extraction Approach for Detecting Mild Cognitive Impairment and Alzheimer's Disease using EEG Datasets.

Early differentiation of mild cognitive impairment (MCI) from Alzheimer's disease (AD) remains difficult because diagnosis integrates cognitive, imaging, laboratory, and neurological evidence. EEG-based models may provide complementary information, but fixed descriptors can miss signal-specific structure. SOFEx was developed as an adaptive and interpretable EEG feature-engineering framework. The analysis used 1,905 EEG segments from 451 recordings of 168 participants (59 AD, seven MCI, and 102 controls). SOFEx combined self-organized rose pattern extraction, fold-wise feature selection, channel-wise classification, and iterative majority voting. Validation used leave-one-subject-out (LOSO), leave-one-record-out (LORO), and participant-grouped 10-fold cross-validation (CV). Five classifiers were compared under LOSO, and a support vector machine (SVM) was retained. LOSO, used as the primary evaluation of generalization to unseen participants, achieved 84.09% accuracy (95% Wilson CI, 82.38-85.67%), with class-wise recall of 94.47% for AD, 56.59% for controls, and 21.18% for MCI (UAR, 57.41%; macro F1-score, 62.95%). LORO and participant-grouped 10-fold CV achieved 93.65% and 96.38% accuracy, respectively, indicating substantial performance variation across validation schemes. Frontal and parietal electrodes contributed repeatedly to the selected configurations, providing sensor-space information on the signals supporting classification. SOFEx provides an adaptive and interpretable feature-engineering framework for EEG-based classification, combining input-specific pattern selection with sensor-level traceability. Participant-independent evaluation showed high sensitivity for AD, while discrimination of MCI remained more challenging. These findings support the potential of the proposed framework and motivate further validation on larger and independent cohorts to assess its generalizability across clinical settings.