Multiscale-Multistage Temporal Convolutional Network for EEG-Based Mild Cognitive Impairment Detection

Mild cognitive impairment (MCI) is an intermediate stage between normal ageing and dementia, with affected individuals at a higher risk of progressing to Alzheimers disease (AD). While several EEG-based MCI methods have used deep learning models, most models work at a single temporal scale, without stage-wise supervision. This limits their ability to capture the multiscale temporal structure present in EEG signals. To address this, we propose a multiscale-multistage temporal convolutional network (MS-MSTCN) for discriminating MCI from healthy controls (HC) using resting-state electroencephalography (EEG) signals. The model first processes each EEG segment through parallel dilated temporal convolution branches to capture short- and long-range patterns at multiple time scales. Three sequential stages with deep supervision then iteratively refine the class predictions at each stage, softmax probabilities are embedded and fed back into the next stage as class-conditional temporal features. We evaluated MS-MSTCN on two publicly available EEG datasets using 10-fold cross validation and compared it with two baseline models (Multiscale TCN and Multistage TCN) as well as existing EEG-based MCI methods. Across both datasets, MS-MSTCN consistently achieves superior performance and it reduces missed MCI cases and false positives (FP) in HC, which is significant for reliable early screening and clinical decision-making.