Table 3.
Evaluation results of prediction models according to type of data and classification model.
| Logistic regression (Ordinary least square) | Logistic regression + Elastic net | Random forest | Extreme gradient boosting | |||||
|---|---|---|---|---|---|---|---|---|
| AUROC | Deviance | AUROC | Deviance | AUROC | Deviance | AUROC | Deviance | |
| MMSE | 0.803 | 23.845 | - | - | - | - | - | - |
| DM + MMSE | 0.664 | 31.380 | 0.803 | 25.234 | 0.748 | 26.200 | 0.752 | 26.995 |
| RSEEG | 0.824 | 23.037 | 0.824 | 22.183 | 0.773 | 23.843 | 0.807 | 23.929 |
| sensERP | 0.697 | 30.332 | 0.647 | 28.832 | 0.605 | 28.465 | 0.500 | 28.979 |
| attERP | 0.891 | 20.363 | 0.882 | 21.608 | 0.857 | 20.679 | 0.882 | 24.134 |
| RSEEG + sensERP + attERP | 0.739 | 42.722 | 0.849 | 21.146 | 0.832 | 22.569 | 0.849 | 21.193 |
| DM + RSEEG + sensERP + attERP | 0.571 | 73.514 | 0.832 | 21.439 | 0.832 | 21.954 | 0.832 | 21.295 |
| MMSE + RSEEG | 0.798 | 25.300 | 0.849 | 22.141 | 0.807 | 21.581 | 0.849 | 22.123 |
| MMSE + sensERP | 0.807 | 25.795 | 0.798 | 23.338 | 0.790 | 22.598 | 0.756 | 29.949 |
| MMSE + attERP | 0.803 | 23.845 | 0.849 | 23.064 | 0.798 | 24.330 | 0.739 | 26.390 |
| MMSE + RSEEG + sensERP + attERP | 0.782 | 40.182 | 0.866 | 20.855 | 0.866 | 20.875 | 0.866 | 20.986 |
| DM + MMSE + RSEEG + sensERP + attERP | 0.605 | 86.032 | 0.849 | 22.048 | 0.866 | 21.140 | 0.874 | 21.150 |
| Significant-variables | 0.874 | 20.628 | 0.891 | 19.397 | 0.798 | 22.908 | 0.874 | 20.461 |
The MMSE score, demographic information (i.e., sex, age, education level, interaction between sexes, and other variables), resting-state EEG (eight variables), sensory ERP (five variables), and selective-attention ERP (four variables) were the data sources. For the models with demographic variables, the terms corresponding to the interaction between sex and other variables were included in each model. DM, Demographic information; RSEEG, resting state EEG; sensERP, sensory ERP; attERP, selective-attention ERP. For models with demographic variables, the terms corresponding to the interaction between sex and other variables are included in each model.“Significant-variables” models contain MEF, peak frequency, alpha/theta, frontal asymmetry from resting state EEG, response time from sensory ERP, and number of correct responses, response time and weighted error percentile from attention ERP. The bold fonts indicate outstanding prediction accuracies.