Fig. 3.

ROC and AUC analysis of incident dementia prediction model development and predictive value comparison. An elastic net regression model based on lasso penalty was used for dementia prediction. After 10-fold cross-validation, 24 of 249 metabolites were selected for the dementia prediction model. Xb1 curve used conventional risk factors as input signals, while the Xb2 curve was for 24 selected metabolites and Xb3 was for conventional risk factors and 24 selected metabolites. There was no clinically significant difference (P = 0.042) found between the AUC of Xb1 and Xb3. Conventional risk factors included age, gender, education level, systolic pressure, anti-hypertension treatment, diabetes mellitus, smoking status, history of stroke, history of coronary heart disease, and APOE ε4 allele. ROC, receiver operating characteristic; AUC, area under curve