Prediction of mortality and in-hospital complications for acute myocardial infarction patients using artificial neural networks
Mohammad Belall Ismael, W. Ed Hammond · 1999
Background. An acute myocardial infarction (MI) is one of the more common diagnoses for hospitalized patients in America. In treating this disease, physicians commonly have several different treatment regimens available. A decision support system that could identify patients at high risk for developing complications would be a great aid to physicians. The large volume of available patient data and the relative lack of theory relating patient characteristics to complications make this a problem well suited for machine learning technologies such as artificial neural networks. The purpose of this research was to determine whether or not it was possible to reliably predict in-hospital complications and 30-day mortality for acute MI patients using neural network technology. Methods. A database of 20,873 American patients who suffered an acute MI between 1990 and 1993 was used in this analysis. Twenty-two history and physical variables were analyzed as predictors for 16 distinct complications. First an extensive univariate and bivariate analysis was performed on the database to determine the feasibility of this study. Then sixteen separate feedforward, backpropagation neural networks were trained and analyzed, one for each complication. A final application incorporating the 16 trained neural networks was created that calculated a score for the patients representing their likelihood of having any complication. The index of discrimination for each of the neural networks, as well as for the scoring system, was defined to be the area underneath their respective ROC curves. Results and conclusions. The neural networks proved to be successful in predicting 4 of the specific complications: death within 30 days, shock, asystole, and congestive heart failure or pulmonary edema. The ROC areas for these complications ranged from 0.73–0.82. Nine other complications had ROC areas ranging from 0.61–0.68, indicating they were of marginal utility. Three complications had ROC areas under 0.57, suggesting that these complications were rather random events in relation to the history and physical variables. The scoring system had a rather low discriminatory ability with an ROC curve area of 0.59.