Evaluation of decision-tree models of machine learning for the prediction of acute liver failure after resuscitation

Background Patients after cardiac arrest developing acute liver failure (ALF) show higher fatality rates and worse outcomes. As machine learning is able to support physicians in their decision-making with the help of big data in health records, the aim of this study is to evaluate decision-tree models for the prediction of ALF after resuscitation. Patients and methods The study is a secondary analysis of 347 Belgian patients after successful resuscitation from 2007 to 2015, whose anonymised dataset was published under the Creative Common Attribution Licence. For machine learning, a J48 decision-tree model as well as its AdaBoost variation and a random forest model were created. Evaluation was performed by sensitivity, specificity, positive / nega-tive predictive value (PPV / NPV) and by the area under the receiver operator curve (AUC-ROC). Results 184 patients (53 %) developed ALF after resuscitation. The main split points in J48 were the attributes international normalised ratio ad admission (INR, threshold 1.48) and total bilirubin ad ad-mission (threshold 0.99 mg/dl). Further splitting was performed in chronic renal and heart failure, administered amount of epinephrine during resuscitation, and the use of beta-lactam antibiotics. J48 yielded a sensitivity of 0.8 and a speci-ficity of 0.72 (0.73 PPV / 0.79 NPV, AUC-ROC 0.83). AdaBoost yielded a compa-rable sensitivity (0.79) and a higher PPV (0.75) and specificity (0.74, NPV 0.77, AUC-ROC 0.85). Random forest reached a sensitivity of 0.86 (PPV 0.74) and a specificity of 0.74 (NPV 0.84, AUC-ROC 0.87). The random forest achieved a significantly higher test quality in AUC-ROC than the other models. Conclusion Machine learning with decision-tree models seems to be suitable for the prediction of ALF after resuscitation. The clinical risk of developing ALF could be increased in conditions with elevated INR and total bilirubin, chronic renal and heart failure as well as prolonged resuscitation. Based on the applied at-tributes and developed models, further studies in large cohorts could lead to an advancement of the prediction models and therefore support physicians in diagnosis and decision-making.