Use of Deep Belief Network Model to Discriminate Mild Cognitive Impairment and Normal Controls Based on EEG, Eye Movement Signals and Neuropsychological Tests

Background: Early detection of mild cognitive impairment (MCI) patients can be of great clinical significance in the primary care settings. Therefore, we developed an automatic, non-invasive MCI detection approach with multimodal measurements and Deep Belief Network (DBN) framework to perform first-line cognitive assessment for clinical practices. Methods: We recruited 152 patients with a clinical diagnosis of MCI and 184 Normal Controls (NC), who underwent electroencephalograph (EEG) and Eye movement signal recording under visual stimuli task, as well as neuropsychological assessments. We implemented linear and nonlinear analyses to extract 40 features from physiological signals (EEG and Eye movement) and 32 features from neuropsychological assessments. Also, we applied a DBN framework to construct three models for MCI detection: (i) Feature model with physiological features, (ii) Clinical cognitive model with neuropsychological features and 4 clinical variables, and (iii) Combined model with clinical variables, neuropsychological and physiological features. Results: The comparison of all three models showed the combined model had an excellent identification ability and was the best model for MCI detection (accuracy: Feature model: 84.21%; clinical cognitive model: 73.52%; combined model: 89.87%). Conclusions: Our proposed detection method may be a powerful tool to discriminate MCI patients from normal controls. Combining clinical cognitive assessments and multimodal measurements yielded a higher accuracy than single modal and bimodal detection tools in screening MCI patients in the large-scale population under primary care condition.