Two-Stage Machine Learning Approach to Predict the Survival State of Liver Cirrhosis Patients with Correctness Estimates

Accurate prediction of survival state for liver cirrhosis patients is important for prioritizing liver transplantation. Using a Mayo Clinic study dataset on liver cirrhosis patients, this study aims to utilize clinical measurements to train machine learning (ML) classifiers to predict the patient survival state. Two derived features, logBXProthrombin and logBXCopper combining log(Bilirubin) with Prothrombin or Copper respectively are used for their informativeness for this task. After model evaluations using 20-fold cross validation, the Random Forest (RF) model is chosen for this binary classification task. This RF model achieves a ROC-AUC of 0.83 on a held-out testing dataset. The primary RF model's prediction metrics (deceased class probability, variance, and confidence interval lower bound of the probabilities) are then used to train a secondary ML classifier for estimating correctness of the primary classifier predictions. This secondary ML classifier achieves ROC-AUC = 0.78 and outperforms simple confidence heuristic (ROC-AUC of 0.66).