#2654 USE OF SUPERVISED DEEP LEARNING ALGORITHM TO PREDICT RISK OF RENAL REPLACEMENT THERAPY(RRT) IN PATIENTS WITH CHRONIC KIDNEY DISEASE (CKD)

Abstract

Abstract Background and Aims Chronic kidney disease (CKD) is one of the most common causes of mortality, affecting around 10% adults worldwide [1]. Although statistical models for predicting risk of renal replacement therapy (RRT) have been developed for a decade [2], referral of patients with CKD to nephrology service is often based on clinical experience of primary care physicians. The use of deep learning algorithms (DLAs) allow clinicians to capture complex, multidimensional, non-linear relationships instead of relying on linear relationships between independent variables and outcome. Our study aims to develop a DLA that has at least non-inferior performance compared to the Kidney Failure Risk Equation (KFRE), which is a well-established and validated risk prediction tool [3]. Method This is a retrospective cohort study carried out in 3 major acute hospitals in Hong Kong providing a total of 3000 beds from Jan 1, 2009 to Mar 31, 2022. All patients aged > 18 with an estimated glomerular filtration rate (eGFR)<30ml/min/1.73 m2 according to the CKD-EPI formula, who attended follow-up for at least 3 months in nephrology clinic of the 3 hospitals, were recruited. Those who were on chronic RRT, received renal transplantation or had an eGFR<15 ml/min/1.73 m2 before referral were excluded. Supervised DLAs of different structures and their combinations were created and trained. Data in the test set were fed into models to predict the risk of requiring RRT in 2 years and 5 years. Predictive performance of these models were compared with that of KFRE [3]. Results 4992 patients were recruited in the study. Data of 499 patients were isolated as test set and were not used for model training. 1576 patients progressed to stage 5 CKD during their follow-up, and 989 patients required initiation of RRT during the study timeframe. When compared with KFRE (4-variable: ROC-AUC = 0.84, 95%CI:0.835-0.852; 8-variable: ROC-AUC = 0.84, 95%CI:0.830-0.847), almost all DLAs showed statistically significant superior robustness in predicting the risk of requiring RRT (Figure 1). No significant difference was found between performance of DLAs combining different neural network layers and those with single structures (CNN+LSTM+ANN layers ROC-AUC = 0.90, 95%CI:0.896-0.902; CNN ROC-AUC = 0.91, 95%CI:0.907-0.914) (Table 1). Conclusion The use of DLAs provided better prediction in the risk of requiring RRT when compared with KFRE.