LONDON — Support vector machines had the highest accuracy in predicting liver decompensation among a subset of patients in Germany compared with other machine learning models, according to data presented at the International Liver Congress.

“Since decompensation of cirrhosis significantly increases mortality, early detection of patients at risk of worsening liver function is of paramount importance,” Sophie Elisabeth Müller, of Saarland University Medical Center in Homburg, Germany, and colleagues wrote.

Seeking to identify predictors of hepatic decompensation, Müller and colleagues applied machine learning techniques, including decision trees, random forests, neural networks and support vector machines (SVMs), to trial data from 1,415 patients with cirrhosis from three German university hospitals. Researchers analyzed laboratory values, medical history and genetic data and trained and tested models with an 85:15 split. Permutation feature importance (PFI) was used to evaluate the impact of features on the prediction of decompensation.

At the index date, researchers reported that 313 patients were permanently compensated, 354 patients had been decompensated before and 748 patients were currently decompensated.

Of 825 patients that could be followed-up (median duration, 12 months), 46.5% decompensated, with SVM demonstrating superior performance in predicting decompensation with an accuracy of 84.1% for training and 77.7% for test data set in retrospective assessment, and 78.4% and 73.8% for prospective analysis.

According to PFI, baseline albumin, bilirubin and minimum serum sodium concentration levels associated with former decompensation. Maximum level of bilirubin and baseline sodium and albumin levels were the most accurate variables for prospective data.

“Among tested machine learning models, SVM seems to have the highest accuracy in predicting liver decompensation,” Müller and colleagues concluded. “In addition to classical laboratory parameters, genetic factors and infections were critical parameters for individual predictions by SVM.”