Model that blends machine learning, traditional risk score may improve cirrhosis prediction
Using clinical variables identified from simple machine learning in a cirrhosis mortality model produced a new score that was more transparent than a commonly used version of the Model for End-Stage Liver Disease, a study found.
Applying machine learning to a traditional risk score for cirrhosis may lead to a blended, simpler, clinically explainable risk score, a study found.
To compare machine learning methods in predicting overall mortality in cirrhosis, as well as how machine learning could choose easily scored clinical variables, researchers created a prognostic study cohort of 107,939 adult patients with cirrhosis or its complications from 130 hospitals and ambulatory clinics in the Veterans Affairs health care system from October 2011 to September 2015. Potential predictors of mortality included demographic characteristics; cause, severity, and complications of liver disease; use of health care resources; comorbid conditions; and laboratory and medication data. Two-thirds of patients were randomly selected for model development and one-third for validation.
The researchers evaluated three statistical and machine learning methods: gradient descent boosting, logistic regression with least absolute shrinkage and selection operator (LASSO) regularization, and logistic regression with LASSO constrained to select no more than 10 predictors (partial-pathway model). The predictors identified in the most parsimonious (partial-pathway) model were then refit using maximum-likelihood estimation to develop the Cirrhosis Mortality Model (CiMM). The researchers compared its predictive performance with that of the widely used Model for End-Stage Liver Disease with sodium (MELD-Na) score. Results were published Nov. 3 by JAMA Network Open.
The main outcome of annual mortality rate ranged from 8.8% to 15.3%. In the study, 32.7% of patients died within three years and 46.2% died within five years after the index date. The researchers noted that models predicting one-year mortality had good discrimination for the gradient descent boosting (area under the receiver-operating characteristics curve [AUC], 0.81; 95% CI, 0.80 to 0.82), logistic regression with LASSO regularization (AUC, 0.78; 95% CI, 0.77 to 0.79), and the partial-pathway model (AUC, 0.78; 95% CI, 0.76 to 0.78). All models showed good calibration.
The authors concluded that the CiMM with machine learning-derived clinical variables offered significantly better discrimination than the MELD-Na score. AUCs were 0.78 (95% CI, 0.77 to 0.79) versus 0.67 (95% CI, 0.66 to 0.68) for one-year mortality, respectively. The CiMM's accuracy was similar to that of less interpretable machine learning algorithms yet higher than that of the traditional MELD-Na severity score, the authors wrote. The CiMM was more transparent than the machine learning models that helped select the variables and could be more easily applied in a variety of point-of-care situations, they added.
“Although we tested this approach in cirrhosis, it holds promise for improving prognostication across other medical conditions,” the researchers wrote. “If confirmed in other conditions, this blended approach could improve data-driven risk prognostication through the development of new scores that are more transparent and more actionable than machine learning and more predictive than traditional risk scores.”