Abstract
Background: Diabetics constitute a significant percentage of hemodialysis (HD) patients with higher mortality, especially among male patients. A machine learning algorithm was used to optimize the prediction of time to death in male diabetic hemodialysis (MDHD) patients.
Methods: This multicenter retrospective study was conducted on adult MDHD patients (2011-2019) from 34 HD centers affiliated with Shiraz University of Medical Sciences. As a special type of machine learning approach, an elastic net penalized Cox proportional hazards (EN-Cox) regression was used to optimize a predictive regression model of time to death. To maximize the generalizability and simplicity of the final model, the backward elimination method was used to reduce the estimated predictive model to its core covariates.
Results: Out of 442 patients, 308 eligible cases were used in the final analysis. Their death proportion was estimated to be 28.2%. The estimated overall one-, two-, three-, and eight-year survival rates were 87.6%, 74.4%, 67.2%, and 53.9%, respectively. The EN-Cox regression model retained 14 (out of 35) candidate predictors of death. Five variables were excluded through backward elimination technique in the next step. Only 6 of the remaining 9 variables were statistically significant at the level of 5%. Body mass index (BMI)<25 kg/m2 (HR=2.75, P<0.001), vascular access type (HR=2.60, P<0.001), systolic blood pressure (1.02, P=0.003), hemoglobin (11≤Hb≤12.5 g/dL: HR=3.00, P=0.028 and Hb<11 g/dL: HR=2.95, P=0.021), dialysis duration in each session≥4hour (HR=2.95, P<0.001), and serum high-density lipoprotein cholesterol (HDL-C) (HR=1.02, P=0.022) had significant effects on the overall survival (OS) time.
Conclusion: Anemia, hypotension, hyperkalemia, having central venous catheter (CVC) as vascular access, a longer dialysis duration in each session, lower BMI and HDL-C were associated with lower mortality in MDHD patients.