Predicting 3-year all-cause mortality in patients undergoing hemodialysis using machine learning.

Abstract

Background: Survival rates for patients after starting hemodialysis (HD) remain low, and cardiovascular events remain the most common cause of death. However, few reports have investigated risk models that include electrocardiogram (ECG) findings. The present study aimed to develop a novel risk model including ECG findings for predicting all-cause death in patients undergoing HD.

Methods: We enrolled 454 patients undergoing HD at 4 facilities from April 2008 to March 2021. Multivariate Cox regression analysis was performed to identify predictive factors, which were used to create a nomogram. We calculated the area under the curve (AUC) and used calibration plots to evaluate the risk model. Bootstrapping was also performed to evaluate the relationship between predicted and observed probabilities.

Results: During the 3-year follow-up period, 98 (21.5%) patients died. Age (P < 0.001), serum albumin level (P = 0.03), history of stroke prior to HD initiation (P = 0.001), atrial fibrillation (P = 0.01), and corrected QT interval (P = 0.005) were identified as independent predictors of all-cause death. The predictive model was constructed using all these parameters with good discrimination of all-cause death, showing an AUC of 0.83 with 80.1% sensitivity and 75.6% specificity. The AUC based on the tenfold cross-validation was 0.82, with 78.2% sensitivity and 75.1% specificity, suggesting a good model.

Conclusion: This novel risk model can effectively stratify high-risk patients and predict 3-year all-cause mortality in patients undergoing HD. We anticipated that this risk model might contribute to identify high-risk cases earlier and provide safer prescriptions and treatments for individual patients.