Prediction Chronic Kidney Disease Progression in Diabetic Patients Using Machine Learning Models

Chronic kidney disease (CKD) significantly impacts healthcare system worldwide. Prediction suggesting intervention to possibly delay the CKD progression is a leading public health challenge especially in diabetes (DM) patients whose risk for CKD progression is relatively high. In this paper, a model for prognostication of kidney disease progression from stage 1 to stage 5 was determined by comparing the predictive accuracy among machine learning models including Cox regression (CoxPH), parametric survival models and random survival forest (RSF). The models were developed using a dataset of DM type 2 patients from a triary hospital in southern Thailand. The predictive accuracy parameters including time dependent concordance index (C-index) and area under the curve (AUC) were calculated and compared. The associated factors that influence renal disease progression were also assessed. From the data of 8037 DM patients, the cox regression and weibull was the best model for predicting CKD progression, with AUC of 0.934-0.965 and a C-index of 0.907 at time 6 to 60 months after DM diagnosis. The gompertz had relatively lower predictive accuracy with the C-index of 0.568 and AUC of 0.610-0.719. Based on the CoxPH model, the associated factors for CKD progression were demographic factors, comorbidities, laboratory parameters and medication received. Conclusion: CoxPH exhibit the best predictive accuracy which if implemented at the time of DM diagnosis, the results can provide insights for physician on the CKD prognosis leading to prescription adjustment and personalized recommendations to delay CKD progression, especially in DM patients with high CKD risk.