Prognostic model of complications in type 2 diabetes : traditional statistical model versus machine learning

Abstract

Diabetes and its complications have become the major public health concerns globally. Many prognostic models of diabetic complications have been developed, but their performances are still varied. Only limited studies derived the models by using the appropriate methods and were satisfactorily reported. The accurate prognostic model related risk factors are essentially needed for early detection and prevention strategies notably for Type 2 diabetic (T2D) patients. This study aimed to validate the existing models that have been published in the literature and develop a new model based on the traditional statistical model and machine learnings (MLs). This study was a retrospective cross-sectional design. We obtained the nationwide survey data and linked it with the standard hospital databases. We synthesized the existing model by the recent systematic review. Traditional metrics (i.e., discrimination, calibration, and reclassification) were compared to evaluate their predictive performances. We updated the equations and developed the new models using the substantial prognostic factors for adjusting the model performance in our settings. A total of 2, 3, 3, and 4 prognostic studies for developing DR, CKD, ESRD and CVDs were included into external validations. We updated the equations by including the routine clinical factors which significantly improved C-statistics. After the recalibration and update, the O/E ratios were getting closer whereas the C-statistics also significantly increased. In validation, most CKD and ESRD models had moderate performance whereas DR and CVDs still varied from low to poor performances. We derived a successful Thai DR risk score with good discrimination and calibration. The traditional statistical models by logistic regression and seven MLs (i.e., Neural Network, Support Vector Machine, Decision Tree, Extreme Gradient Boosting, Random Forest, Naïve Bayesian, and LASSO Logistic Regression) were performed well and compared for their evaluation metrics. The Thai DR risk score is easy to use in identifying the T2D patient who have the high risk to develop DR in the routine clinical practice. Further external validations are needed to clarify the performance with the parallel development of the software applications