BUILDING CLINICAL CLASSIFIERS USING INCOMPLETE OBSERVATIONS - A NEURAL-NETWORK ENSEMBLE FOR HEPATOMA DETECTION IN PATIENTS WITH CIRRHOSIS

One objective of liver transplant evaluation is to identify patients t hat harbor a hepatoma, but standard screening techniques are not sensi tive enough. We trained neural network ensembles to predict the presen ce of hepatoma in patients with cirrhosis, based on information collec ted at the time of transplant evaluation. Network architecture and tra ining were modified to handle missing observations. Three ensembles we re trained: ensemble A using the subset with no missing observations ( 528 patients); ensemble B using the complete set, which included missi ng observations (853 patients); and ensemble C using the smaller subse t, originally with complete data, but after a fixed number of observat ions were deleted (i.e., made ''missing''). Ensemble performance on te sting sets was very good. The areas under the ROC curves were 0.91, 0. 89, and 0.90, for ensembles A, B, and C, respectively. Neural networks can successfully perform this classification task, and strategies can be developed that allow use of incomplete observations.