Diagnosis of Alzheimer’s Disease Via Machine Learning Approaches with Integrated Resting-State EEG and ERP Characteristics

Electroencephalogram (EEG) has proven to be a cost-effective and non-invasive tool for detecting electrophysiological changes during the preclinical stage of Alzheimer’s disease (AD). Validating the efficacy of EEG biomarkers is essential for early discrimination and prediction of cognitive decline. This study aimed to investigate whether the combined use of resting state EEG and event-related potentials (ERPs) could enhance diagnostic performance through multiple machine learning (ML) models. Resting-state EEG and auditory ERP data of 240 individuals, including 101 Alzheimer’s disease (AD) subjects, 82 mild cognitive impairment (MCI) subjects, and 57 subjects with normal cognition, were recorded. The extracted features after parameter screening were fed to multiple machine learning classifiers to perform binary and three-way classifications. The relative power and functional connectivity measures in the theta band, and Hjorth complexity showed significant differences between normal cognition and AD. People with AD exhibited decreased amplitudes of MMN compared to normal cognition, but MCI showed a compensatory increase. The amplitudes of P50 waveforms and the latency of N100 exhibited a significant increasing trend from NC to AD. Comparative results showed that when multimodal resting-state EEG/ERP features were used as input, an accuracy of 0.7519 with the Gradient Boosting classifier in three-way classification was achieved. Interestingly, classification performance was further improved when neuropsychological scales were added, with an accuracy of 0.8036 in Ensemble learning. The multimodal approach, integrating resting-state EEG and ERP features, significantly enhances early diagnostic accuracy for cognitive impairment. These findings underscore the potential of EEG-based multimodal frameworks for clinical screening for cognitive decline.