Accurate deep-learning model to differentiate dementia severity and diagnosis using a portable electroencephalography device

Mild cognitive impairment (MCI) and dementia present critical health challenges in aging populations, highlighting the need for prompt and accurate diagnostic methods. Current diagnostic approaches for dementia are constrained by limited access to specialists and the high cost or invasiveness of the diagnostic methods, necessitating the development of widely accessible, cost-effective, and noninvasive alternatives. Electroencephalography (EEG) is a promising, accessible, cost-effective, noninvasive biomarker of dementia. However, traditional EEG systems are limited by their lack of portability and need for skilled technicians. This study proposes a deep-learning-based model to differentiate healthy volunteers (HVs) from patients with dementia-related conditions using data from portable EEG devices. A dataset of 233 participants, including 119 HVs and 114 patients categorized by severity and clinical diagnosis, was analyzed. We developed a customized transformer-based model to analyze the EEG data transformed into frequency-domain features using a short-time Fourier transform. The model was evaluated using 10-fold cross-validation and a hold-out dataset. In cross-validation, the model achieved an area under the curve (AUC) of 0.872 and a balanced accuracy (bACC) of 80.8% in distinguishing HVs from patients. Subgroup analyses by severity and clinical diagnosis yielded AUCs of 0.812 to 0.898 and bACCs of 74.9 to 86.4%, respectively. Comparable performance was observed in the holdout dataset, except for small sample sizes. These findings highlight the potential of portable EEG devices and deep-learning models as practical tools for dementia screening.