Development, Validation, and Evaluation of a Simple Machine Learning Model to Predict Cirrhosis Mortality

Area of Science:

• Hepatology
• Medical Informatics
• Biostatistics

Background:

• Traditional prognostic models for cirrhosis often lack the predictive accuracy of machine-learning algorithms.
• Machine learning models, while accurate, can be complex and opaque for clinical application.
• There is a need for more transparent and clinically applicable prognostic tools in cirrhosis management.

Purpose of the Study:

• To compare the performance of different machine learning methods in predicting overall mortality in patients with cirrhosis.
• To utilize machine learning to identify easily scorable clinical variables for a novel cirrhosis prognostic model.

Main Methods:

• A retrospective cohort study of 107,939 adult patients with cirrhosis was conducted across 130 Veterans Affairs healthcare system hospitals.
• Three machine learning methods were evaluated: gradient descent boosting, logistic regression with LASSO regularization, and a partial pathway logistic model.
• Predictive performance was assessed using area under the receiver operating characteristics curve (AUC) and calibration, comparing the developed Cirrhosis Mortality Model (CiMM) against the MELD-Na score.

Main Results:

• Machine learning models demonstrated good discrimination for 1-year mortality, with AUCs ranging from 0.78 to 0.81.
• The final CiMM model, derived from machine learning-identified clinical variables, significantly outperformed the MELD-Na score (AUC 0.78 vs. 0.67).
• All evaluated models showed good calibration.

Conclusions:

• Simple machine learning techniques achieved performance comparable to more advanced methods like gradient boosting.
• The developed Cirrhosis Mortality Model (CiMM) is more transparent than complex machine learning models and more predictive than the MELD-Na score.
• This study highlights the utility of machine learning in identifying key clinical variables for improved cirrhosis prognostication.