Neural networks as predictors of outcomes in alcoholic patients with severe liver disease

We developed and evaluated neural networks as predictors of outcomes in alcoholic patients with severe liver disease using commonly available clinical and laboratory values. Hospital charts of 144 patients were reviewed. Nine variables (five laboratory, four clinical) were recorded along with in-hospital death or survival. Data were organized into separate development and validation sets. Neural network predictions of survival were compared with those of the Maddrey discriminant function and logistic regression models developed on the same data. Model performance was evaluated by comparing areas under receiver-operating characteristic (ROC) curves and the distributions of model scores. Survivors had significantly different laboratory and clinical characteristics, the most important being a higher prothrombin time, lower bilirubin, and lower incidence of encephalopathy. Neural network performance was significantly better than that of the Maddrey score (ROC areas, 81.5% vs. 73.8%; P = .04). The ROC area for neural networks was similar to that of logistic regression (ROC area 78.2%; P = .3), but the neural networks were more successful in classifying patients into low- and high-risk groups (P < .001). A neural network score with laboratory data from hospital-day 7 improved prognostic accuracy further to 84.3%. After adjusting for baseline risk, the neural network change in illness severity was still a significant predictor of mortality (P = .001). Neural networks using clinical and laboratory data showed a high prognostic accuracy for predicting mortality in alcoholic patients with severe liver disease.