Modeling mortality in the intensive care unit: comparing the performance of a back-propagation, associative-learning...

The objective of this study was to compare and contrast two techniques of modeling mortality in a 30 bed multi-disciplinary ICU; neural networks and logistic regression. Fifteen physiological variables were recorded on day 3 for 422 consecutive patients whose duration of stay was over 72 hours. Two separate models were built using each technique. First, logistic and neural network models were constructed on the complete 422 patient dataset and discrimination was compared. Second, the database was randomly divided into a 284 patient developmental dataset and a 138 patient validation dataset. The developmental dataset was used to construct logistic and neural net models and the predictive power of these models was verified on the validation dataset. On the complete dataset, the neural network clearly outperformed the logistic model (sensitivity and specificity of 1 and .997 vs. .525 and .966, area under ROC curve .9993 vs. .9259), while both performed equally well on the validation dataset (area under ROC of .82). The excellent performance of the neural net on the complete dataset reveals that the problem is classifiable. Since our dataset only contained 40 mortality events, it is highly likely that the validation dataset was not representative of the developmental dataset, which led to a decreased predictive performance by both the neural net and the logistic regression models. Theoretically, given an extensive dataset, the neural network should be able to perform mortality prediction with a sensitivity and a specificity approaching 95%. Clinically, this would be an extremely important achievement.(ABSTRACT TRUNCATED AT 250 WORDS)