Deep learning algorithms for predicting renal replacement therapy initiation in CKD patients: a retrospective cohort study

Abstract

Background: Chronic kidney disease (CKD) requires accurate prediction of renal replacement therapy (RRT) initiation risk. This study developed deep learning algorithms (DLAs) to predict RRT risk in CKD patients by incorporating medical history and prescriptions in addition to biochemical investigations. Methods: A multi-centre retrospective cohort study was conducted in three major hospitals in Hong Kong. CKD patients with an eGFR < 30ml/min/1.73m2 were included. DLAs of various structures were created and trained using patient data. Using a test set, the DLAs' predictive performance was compared to Kidney Failure Risk Equation (KFRE). Results: DLAs outperformed KFRE in predicting RRT initiation risk (CNN + LSTM + ANN layers ROC-AUC = 0.90; CNN ROC-AUC = 0.91; 4-variable KFRE: ROC-AUC = 0.84; 8-variable KFRE: ROC-AUC = 0.84). DLAs accurately predicted uncoded renal transplants and patients requiring dialysis after 5 years, demonstrating their ability to capture non-linear relationships. Conclusions: DLAs provide accurate predictions of RRT risk in CKD patients, surpassing traditional methods like KFRE. Incorporating medical history and prescriptions improves prediction performance. While our findings suggest that DLAs hold promise for improving patient care and resource allocation in CKD management, further prospective observational studies and randomized controlled trials are necessary to fully understand their impact, particularly regarding DLA interpretability, bias minimization, and overfitting reduction. Overall, our research underscores the emerging role of DLAs as potentially valuable tools in advancing the management of CKD and predicting RRT initiation risk.