Transformer‐Based Multi‐Channel K‐Complex Detection Algorithm Tailored for Elderly Patients With Amnestic Mild Cognitive Impairment

K-complexes (KCs) are hallmark waveforms of non-rapid eye movement stage 2 (NREM2) sleep, associated with sleep maintenance and memory consolidation. KC density and amplitude decline with ageing and are further altered in amnestic mild cognitive impairment (aMCI). Manual scoring, while considered the gold standard, is labour intensive and subjective. Existing automated KC detectors, often trained on small public datasets of young healthy subjects using single-channel electroencephalography (EEG), may underperform in elderly aMCI individuals whose KC morphology is more variable. Hence, the goal of this study is to develop and validate AdaPatchFormer, an automated multi-channel Transformer-based KC detection algorithm optimised for elderly individuals with aMCI. AdaPatchFormer integrates a period embedding module, which adaptively identifies physiologically relevant rhythms across multiple frequency bands, with a multi-granularity encoder that progressively fuses temporal features across channels. The model was trained on full-night polysomnography (PSG) recordings from 268 elderly aMCI patients and evaluated against expert-labelled gold standards on four independent test datasets: private aMCI and cognitively normal cohorts, plus two public elderly cohorts. Across all datasets, AdaPatchFormer outperformed the two open-access detectors by Chambon et al. and Lechat et al., achieving higher recall, precision, specificity, accuracy, F1 score, Matthews correlation coefficient (MCC) and a well-balanced recall-precision profile. Moreover, the KC density and amplitude detected by AdaPatchFormer closely mirrored expert annotations. These results suggest that AdaPatchFormer is a robust, accurate, and objective algorithm for KC detection in elderly individuals, with the potential for supporting early and cost-effective identification of aMCI in real-world settings.