Predicting mortality in patients with cirrhosis of liver with application of neural network technology

Prediction of mortality from cirrhosis is important in planning optimal timing of liver transplantation and other interventions. We evaluated the role of the Artificial Neural Network (ANN), which uses non-linear statistics for pattern recognition in predicting one-year liver disease-related mortality using information available during initial clinical evaluation.

The ANN was constructed using software with data from a training set (n = 46) selected at random from a cohort of adult cirrhotics (n = 92). After training, validation was performed in the remaining patients (n = 46) whose outcome in terms of one-year mortality was unknown to the network. The performance of ANN was compared to those of a logistic regression model (LRM) and Child-Pugh's score (CPS). Death (related to cirrhosis/its complications) within one year of inclusion was the outcome variable. The ANN was also tested in an external validation sample (EVS, n = 62) from another hospital.

Patients in the EVS were younger (mean age, 41 vs 45 years), infrequently of alcoholic etiology (5% vs 49%), had less severe disease (mean CPS 6.6 vs 10.8), and had lower one-year mortality (13 vs 46%). In the internal validation sample, ANN's accuracy was 91%, sensitivity 90% and specificity 92% in prediction of one-year mortality; area under the receiver-operating characteristic (ROC) curve was 0.94. The performance of the LRM (accuracy 74%) and the CPS (accuracy 55%) was significantly worse than ANN (P < 0.05, McNemar's test). Despite differences in the characteristics of the two groups, the ANN performed fairly well in the EVS (accuracy of 90%, area under curve 0.85).

ANN can accurately predict one-year mortality in cirrhosis and is superior to CPS and LRM.