Simple Machine Learning Method Predicts Cirrhosis Mortality Risk

A simple machine learning approach predicted cirrhosis mortality risk as accurately as less interpretable algorithms, and also outperformed traditional risk methods, a study published in JAMA Network Open revealed.

Cirrhosis is a high-risk condition with a progressive clinical course, researchers noted. While prognostic risk models exist for cirrhosis, the team said that these existing scores aren’t always the most comprehensive when assessing mortality risk – including the most widely used cirrhosis prognostic models, such as the Model for End Stage Liver Disease with sodium (MELD-Na).

“None of these scores account for a wide range of clinical and psychosocial factors that are likely to be associated with mortality in cirrhosis. Machine learning techniques have been used to help fill these gaps for cirrhosis but have not seen widespread use,” researchers stated.

Researchers developed three separate machine learning algorithms using data from a group of patients with cirrhosis seen at 130 VA hospitals from October 1, 2011 to September 30, 2015. The algorithms had varying levels of complexity and range of variables that predicted risk of mortality in cirrhosis.

To achieve a balance among accuracy, interpretability, and feasibility, the team then developed and validated a blended model, called the Cirrhosis Mortality Model (CiMM). This model used the variables selected from the machine learning algorithms and implemented them in an accessible platform.

The results showed that all three machine learning models achieved good discrimination when predicting one-year mortality, as well as good calibration. Overall, mortality predictions for year two onward showed similar trends, although overall discrimination fell as time elapsed.

Researchers then selected variables from the machine learning models to develop the CiMM, including older age, higher bilirubin level, hepatic encephalopathy, and hepatocellular carcinoma, all of which were associated with higher risk of mortality.

The team then compared the performance of the CiMM with MELD-Na, the traditional risk score method. The CiMM achieved an AUC of 0.78 for one-year mortality, 0.76 for two-year mortality, and 0.72 for three-year mortality. The corresponding AUCs for MELD-Na were 0.67 for one-year mortality, 0.65 for two-year mortality, and 0.61 for three-year mortality.

The findings show that simple machine learning techniques can perform as well as less interpretable, more advanced models, as well as outperform traditional methods.

“Better understanding of prognosis can frame patients’ preferences, help prioritize goals of care, and inform decision-making across many medical conditions. Machine learning models have greatly enhanced the accuracy of such predictions, but their black box analytics have limited their usefulness,” researchers stated.

“The most useful models would combine the high predictive accuracy with the transparency and easy measurability of more traditional risk scores.”

The team noted that although they used cirrhosis as a use case for their simpler approach, the techniques could be applied to other medical conditions as well. The model can also be easily implemented at the patient and population levels, researchers stated.

“The CiMM can be incorporated into electronic health records from EPIC, Cerner Corporation, and others to easily automate and display prediction scores at the individual patient level. The CiMM can also be incorporated into population dashboards as part of quality improvement strategies,” the group said.

“Within these population management systems, CiMM could identify high-risk patients for linkage to care, vigilant surveillance, and proactive care coordination. Matching the interventions with patients’ risk of mortality may allow more tailored approaches rather than one-size-fits-all strategies, enhancing the overall effectiveness of quality improvement initiatives.”

With this simple technique, researchers and provider organizations may be able to easily and accurately predict mortality risk across a wide range of diseases – while still being able to interpret results.

“The promise of machine learning–based medical prognostication has been limited by implementation and interpretation challenges. We found that machine learning can help select important variables for more transparent risk scores while maintaining high rates of accuracy,” researchers concluded.

“The resultant blended CiMM performed better than the widely used MELD-Na score. 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.”