Using machine learning algorithms to predict MACE in peritoneal dialysis patients

The study aimed to predict the risks of Major adverse cardiac events (MACE) in patients undergoing peritoneal dialysis (PD) with machine learning (ML) algorithm. In addition, we added the time factor and predicted the risk factors for MACE during the 1-year and 5-year follow-up. This retrospective study included 1006 PD patients from January 2010 to December 2016. XGBoost, Random Forest (RF) and Adaboost were used to train models for assessing risk of 1-year and 5-year MACE. The optimal ML algorithm was used to construct the models to predict the risk of the MACE end point. 409 patients developed MACE during the follow-up. The RF model (AUC = 0.80) was optimal for overall MACE prediction. The three most influential variables, ranked in descending order of importance were Parathyroid hormone, Congestive heart failure and Age.114 patients developed MACE during the first-year follow-up. The XGBoost model (AUC = 0.86) performed best for 1-year MACE. The three most influential variables, ranked in descending order of importance were High-Density Lipoprotein Cholesterol (HDL-C), Age and Calcium. 331 patients developed MACE during the 5-year follow-up. The RF model (AUC = 0.75) was the best predicting model for 5-year MACE. The three most influential variables, ranked in descending order of importance were Age, Creatinine and estimated Glomerular Filtration Rate. We developed and validated a novel algorithm to predict the risk factors of MACE in PD patients.