Dementia, characterized by memory and cognitive impairment, necessitates early detection for symptom mitigation and progression delay. However, existing diagnostic approaches are resource-intensive, demanding both high cost and expertise, which compli...
Dementia, characterized by memory and cognitive impairment, necessitates early detection for symptom mitigation and progression delay. However, existing diagnostic approaches are resource-intensive, demanding both high cost and expertise, which complicates early intervention. In this study, we propose an optimal model and hyperparameter settings suitable for early dementia diagnosis using machine learning and deep learning models trained and tested on 19-channel electroencephalogram (EEG) data (50 normal subjects, 175 mild cognitive impairment subjects) collected via a dry methodology. While previous studies utilized 64-channel high-resolution EEG data acquired through a wet methodology, this study is distinct in its use of 19-channel dry EEG data, which reduces the burden on researchers and patients. The experiment was conducted by training a binary classifier using an ensemble-based AutoML algorithm, Auto-sklearn, and three deep learning models (ShallowFBCSPNet, Deep4Net, EEGNet), followed by a comparison of their respective performances. EEGNet demonstrated the best results (accuracy = 68.9%). Next, to address the overfitting problem due to data imbalance, four data imbalance processing algorithms (class weight, Gaussian noise, frequency shift, and undersampling) were applied to EEGNet. The best performance (accuracy = 78.9%) was achieved with Gaussian noise applied to EEGNet. This study’s findings suggest the potential utility of combining 19-channel dry EEG data with the proposed method for early dementia diagnosis.