Accurate, simple and non-invasive tools are needed for efficient screening of abnormal glu-cose tolerance (AGT) and educating the general public.
To develop a neural network-based initial screening and educational model for AGT.
230 subjects with AGT and 3,243 subjects with normal glucose tolerance (NGT) were allocated into training, validation and test sets using stratified randomization. The ratios of AGT versus NGT in three groups were 150:50, 30:570 and 50:950, respectively. A feed-forward neural network (FFNN) was trained to predict 2-hour plasma glucose of 75 g Oral Glucose Tolerance Test (OGTT) using age, family history of diabetes, weight, height, waist and hip circumference. The screening performance was evaluated by the area under the receiver operating characteristic (ROC) curve (AUC) and the partial AUC (in the range of false positive rates between 35 and 65%) and compared to those from logistic regression, linear regression and ADA Risk Test.
Sensitivity, specificity, accuracy and percentage that needed further testing at 7.2 mmol/L in test group were 90.0%(95%CI: 78.6 to 95.7%), 47.7% (95%CI: 44.5 to 50.9%), 49.8% (95%CI: 46.7 to 52.9%) and 54.2% (95%CI: 51.1 to 57.3%) respectively. The entire and partial AUCs were 0.70 (95%CI: 0.62 to 0.78) and 0.26 (95%CI: 0.22 to 0.30). The partial AUC of the NN was higher than those of logistic regression (p = 0.06), linear regression (p = 0.06) and ADA Risk Test (P = 0.006).
NN can be used as a high-sensitive and non-invasive initial screening and educational tool for AGT.
