LASSO-based machine learning algorithm to predict the incidence of diabetes in different stages

Background Formal risk assessment is crucial for diabetes prevention. We aimed to establish a practical nomogram for predicting the risk incidence of prediabetes and prediabetes conversion to diabetes. Methods A cohort of 1428 subjects was collected to develop prediction models. The LASSO was used to screen for important risk factors in prediabetes and diabetes and was compared with other algorithms (LR, RF, SVM, LDA, NB, and Treebag). Multivariate logistic regression analysis was used to construct the prediction model of prediabetes and diabetes, and drawn the predictive nomogram. The performance of the nomograms was evaluated by receiver-operating characteristic curve and calibration. Results These findings revealed that the other six algorithms were not as good as LASSO in terms of diabetes risk prediction. The nomogram for individualized prediction of prediabetes included Age, FH, Insulin_F, hypertension, Tgab, HDL-C, Proinsulin_F, and TG and the nomogram of prediabetes to diabetes included Age, FH, Proinsulin_E, and HDL-C. The results showed that the two models had certain discrimination, with the AUC of 0.78 and 0.70, respectively. The calibration curve of the two models also indicated good consistency. Conclusions We established early warning models for prediabetes and diabetes, which can help identify prediabetes and diabetes high-risk populations in advance.

Automated summary

This study aimed to establish a practical nomogram for predicting the risk incidence of prediabetes and prediabetes conversion to diabetes. The LASSO algorithm was found to be superior to other algorithms for diabetes risk prediction. The nomogram for prediabetes included Age, FH, Insulin_F, hypertension, Tgab, HDL-C, Proinsulin_F, and TG, and the nomogram for prediabetes to diabetes included Age, FH, Proinsulin_E, and HDL-C. Both models demonstrated good discrimination and calibration.