Data-driven, two-stage machine learning algorithm-based prediction scheme for assessing 1-year and 3-year mortality...

Area of Science:

• Nephrology
• Medical Informatics
• Data Science

Background:

• Chronic hemodialysis (CHD) patients face significantly reduced life expectancy.
• Identifying mortality risk factors in CHD patients is crucial for improving outcomes.
• Serum laboratory data from regular dialysis offers a potential source for predictive modeling.

Purpose of the Study:

• To develop a machine learning (ML)-based mortality prediction model for CHD patients.
• To identify key risk factors contributing to mortality in this population.
• To evaluate the performance of different ML algorithms in predicting CHD mortality.

Main Methods:

• Retrospective observational cohort study of 800 CHD patients.
• Analysis of 44 laboratory indicators using five ML methods: logistic regression (LGR), decision tree (DT), random forest (RF), gradient boosting (GB), and eXtreme gradient boosting (XGB).
• Development of a two-stage ML algorithm-based prediction scheme, with a stepwise RF model incorporating important risk factors.

Main Results:

• The random forest (RF) model demonstrated superior accuracy and area-under-the-curve (AUC) compared to other ML methods for 1- and 3-year mortality prediction.
• The stepwise RF model, integrating key risk factors identified by multiple ML methods, outperformed LGR in predicting mortality.
• The developed ML model showed satisfactory performance in predicting mortality over 1- and 3-year periods.

Conclusions:

• A two-stage ML algorithm, specifically the stepwise RF, effectively predicts mortality in CHD patients.
• The study highlights the utility of routine laboratory data for risk stratification in CHD.
• Findings can aid nephrologists in patient-centered decision-making and early identification of high-risk individuals.