Machine learning models predicting decompensation in cirrhosis

Background & Aim Since decompensation of cirrhosis significantly increases patients mortality, the prevention and early treatment is paramount. We applied machine learning techniques to identify parameters predicting decompensation.

Methods Using Python, Keras, and Scikit-Learn, several machine learning techniques including Random Forests, Neural Networks and Support Vector Machines (SVM) were trained and tested with 85:15 split on the INCA trial database containing 1,415 patients with cirrhosis from three German university hospitals. In addition to laboratory values and anamnestic data, genetic data including NOD2 genotypes were analysed. Permutation features importance (PFI) as model inspecting technique evaluated the impact of features on the prediction of decompensation.

Results At the index date, 313 patients were always compensated, 354 patients were decompensated before, and 748 were currently decompensated. 825 patients (46.5% decompensated) attended follow up. SVM showed the best performance in predicting decompensation, achieving an accuracy of 84.1% for the training- and 77.7% for the test data set (retrospective assessment) and 78.4% respectively 73.8% (prospective assessment). PFI revealed baseline levels of albumin, bilirubin and minimum serum sodium concentration were highest ranked to assess former decompensation. Maximum level of bilirubin and baseline levels of sodium and albumin were highest ranked for prospective data. In addition to parameters of established scores including MELD and Child-Pugh, NOD2 genotype and parameters related to infections were highly ranked.

Conclusions Among various machine learning models, the highest accuracy to predict decompensation was found for SVM. In addition to classical laboratory parameters, genetic factors and infections were critical parameters for individual predictions.

Publication History

Article published online:
26 January 2022