Predicting Edema in Liver Cirrhosis Patients: Development and Validation of Multivariable Machine Learning Models

Purpose: Edema is a medical condition characterized by the abnormal accumulation of fluid within body tissues, commonly affecting the extremities, such as legs and arms. It is a common complication of liver cirrhosis, a disease characterized by fibrosis and impaired liver function. Although edema may not pose an immediate threat to life, managing it is essential for reducing patient discomfort, keeping track of cirrhosis progression, and avoiding further complications. This study aims to construct machine learning models to predict the likelihood of edema in liver cirrhosis patients. Methods: This research uses data from a Mayo Clinic trial on primary biliary cirrhosis (1974-1984) with 424 patients. Eight classification algorithms—logistic regression, decision trees, support vector machine, random forests, neural networks, XGBoost, AdaBoost, and k-nearest neighbors—were evaluated for predicting the probability of edema. Correlations between the decisive attributes were also analyzed. Risk stratification was performed based on predicted probabilities from the top performing model. Results: Random Forest with 10-fold cross validation, exhibited better performance(accuracy > 90%) in identifying potential edema (precision > 80%), whereas neural network and logistic regression showcased better sensitivity (> 81%). Conclusion: Random forest model is the most suitable model to identify potential edema in cirrhosis patients. Such a system can help physicians to mitigate the risk of fluid imbalance early on.