Predicting the Risk of Decompensation in Patients with Cirrhosis – a Validated Machine Learning Approach

Introduction: There is an unmet need for simple tools to predict development of hepatic decompensation among patients with compensated cirrhosis. We applied machine learning to several international datasets to develop and validate a straightforward predictive model of decompensation. Methods: We used routinely available clinical and laboratory data from 575 patients with compensated cirrhosis from the training cohorts in Nottingham (United Kingdom) and Modena (Italy), with a median follow-up of 4 years. Based on this data, we developed a predictive model using a random forest classifier and validated it across independent international populations involving over 2,100 patients from Dublin (Ireland), Menoufia (Egypt), Leeds (UK) and Ogaki (Japan). Results: In the training cohorts, 22% of patients developed liver decompensation. Using machine learning, we developed RODIC (Risk of Decompensation in Cirrhosis), a well-calibrated model based on albumin, bilirubin and the Fib-4 value (AUC-ROC = 0.86; weighted F1 score = 0.82) which predicts the risk of decompensation within 3 years (access free of charge at: https://antonkaly.pythonanywhere.com/predict). RODIC showed strong performance across all validation sets, with AUC-ROC scores ranging from 0.67 to 0.80 and weighted F1 scores from 0.70 to 0.81. Moreover, the model was effective regardless of cirrhosis aetiology. For HCV-positive patients, RODIC remained reliable irrespective of whether they achieved sustained virologic response. Discussion: Our validated machine learning model based on readily available clinical, and laboratory features accurately quantitates the risk of liver decompensation in patients with compensated hepatic cirrhosis.