Prognostic Machine Learning Models for First-Year Mortality in Incident Hemodialysis Patients: Development and...

Area of Science:

• Nephrology
• Data Science
• Biomedical Informatics

Background:

• First-year survival in hemodialysis patients is poor.
• Existing mortality risk scores lack robustness and applicability.
• Need for improved prediction of early mortality in hemodialysis patients.

Purpose of the Study:

• Develop and validate machine learning models to predict first-year mortality in hemodialysis patients.
• Utilize clinical factors for accurate risk stratification.
• Assist physicians in identifying high-risk patients early.

Main Methods:

• Trained extreme gradient boosting models on large hemodialysis cohorts (n=5351 and n=5828).
• Developed two models: one at dialysis initiation and another 0-3 months post-initiation.
• Validated models using 10-fold cross-validation and assessed performance with AUC, sensitivity, specificity, precision, balanced accuracy, and F1 score.

Main Results:

• 10.93% and 13.11% of patients died within the first year in training and testing cohorts, respectively.
• Identified 15 key predictive features from 42 candidate variables.
• Model 1 (dialysis initiation) achieved AUC 0.83; Model 2 (0-3 months) achieved AUC 0.85, indicating strong predictive performance.

Conclusions:

• Successfully developed and validated two machine learning models for predicting first-year mortality in hemodialysis patients.
• Models demonstrate utility in stratifying patients by mortality risk early in the dialysis process.
• These models can aid clinical decision-making for high-risk hemodialysis patients.