Abstract 11780: Machine Learning Models to Predict Development of CKD and/or HF in Early Stages of Type 2 Diabetes Patients

Chronic kidney disease and/or heart failure (CKD/HF) are the first and most frequent comorbidities that are associated with adverse prognosis in early stages of type 2 diabetes (T2DM) patients. However, efficient screening and risk assessment strategies for the development of CKD/HF remain to be established. We aimed to develop novel machine learning models that can predict the risk of CKD/HF manifestation as well as prognosis in early stages of T2DM patients without a history of CKD or cardiovascular diseases (CVDs). Prediction models were developed with extreme gradient boosting machine (XGB), neural network, logistic regression (LR), and Cox proportional hazards using real-world data of T2DM patients aged ≥18 years without a history of CKD or CVDs (n=217,054) extracted from a Japanese hospital-based administrative claim database. A separate dataset of 16,822 patients from another Japanese database was used for external validation. The outcomes used to construct the models were diagnosis of CKD/HF, hospitalization for CKD/HF and all-cause death during a 5 -year follow-up period. AUROC of XGB and LR for each outcome were as follows. Diagnosis of CKD/HF: 0.777 (XGB); 0.732 (LR), Hospitalization for CKD/HF: 0.785 (XGB); 0.767 (LR), All-cause death: 0.918(XGB) and 0.903(LR). The best-performing XGB models were externally validated; AUROC for diagnosis of CKD/HF , hospitalization for CKD/HF and all-cause death were 0.718 , 0.837 and 0.869 respectively. Furthermore, the Kaplan-Meier curves showed that for both outcomes, the hazard ratio of event-free survival in the group predicted as high-risk and low-risk in the XGB model was statistically significant (Figure). These results demonstrated successful construction of prediction model and using an administrative hospital database and suggest that our model may contribute to early diagnosis and intervention of CKD/HF among early stages of T2DM patients and ultimately improving prognosis of the patients.