A separate set of data sets GSE128381 was selected, containing 183 Placental tissue samples, and the model was applied to these pregnant Placental tissue samples to analyze the accuracy of the model. Specifically, we selected a set from the Hasselt University Centre for The Environmental Sciences data set, GEO number is GSE128381, the expression matrix of 10 hub genes were extracted, our model was used to predict the sample and compared it with the clinical diagnosis. Among them, 178 of the 183 patients diagnosed as normal samples were predicted as normal samples, and 5 of the 6 patients diagnosed as GDM were predicted as GDM patients, with an accuracy rate of 97.3% (Fig.
6a), the area under the ROC curve was 0.773 (Fig.
6b), and the overall prediction performance was good, and a good predictive performance across data platforms. Furthermore, 88 (50%) samples were randomly selected from 177 known normal women using our model for prediction, and the number of normal samples was statistically predicted. In order, 1000 times were randomly selected, among which 400 (40%) times were correctly classified 100%, 5 (5.6%) were the biggest prediction errors, and the frequency was 133 (13.3%) times (Fig.
6c). This indicated that the model has good stability. To analyze the relationship between the model and the maternal history, the 183 cases from Hasselt University Centre for Environmental Sciences were predicted to be GDM group and normal group. The characteristics of the two groups of pregnant women were analyzed, and we found the age of the pregnant women predicted to be GDM were significantly higher than that the predicted normal sample (Fig.
6d). The pre-pregnancy BMI comparison also showed that the GDM sample was significantly higher than normal (Fig.
6e). It is well known that age and BMI are risk factors for GDM in pregnant women, and the model is consistent with maternal age and BMI. To run the double-blind trial, we used the expression profiles of HUVEC cells from umbilical cords in six pregnant women tested by Ambra R et al. [
30], our model was used to predict and identify three GDM and three normal samples. The oral glucose tolerance test (OGTT) was further performed between the 24th and 34th gestational weeks, and the three GDMs reported by the GTT were completely consistent with the model predictions. Furthermore, the expression profiles of Placental tissue samples from 183 pregnant women tested by Cox B et al. [
29] were predicted by our model to identify 11 GDM samples and 172 healthy group samples, However, according to clinical diagnosis of Cox B et al., 5 of the 11 predicted GDM samples were diagnosed as GDM, and 172 predicted healthy samples were all diagnosed as normal samples (Fig.
6f). This suggests that the model is suitable for different data platforms and is highly consistent with current clinical diagnostic methods.