Artificial Neural Network Models for Prediction of Acute Coronary Syndromes Using Clinical Data From the Time of Presentation

Abstract

Clinical and ECG data from presentation are highly discriminatory for diagnosis of acute coronary syndromes, whereas definitive diagnosis from serial ECG and cardiac marker protein measurements is usually not available for several hours. Artificial neural networks are computer programs adept at pattern recognition tasks and have been used to analyze data from chest pain patients with a view to developing diagnostic algorithms that might improve triage practices in the emergency department. The aim of this study is to develop and optimize artificial neural network models for diagnosis of acute coronary syndrome, to test these models on data collected prospectively from different centers, and to establish whether the performance of these models was superior to that of models derived using a standard statistical technique, logistic regression. The study used data from 3,147 patients presenting to 3 hospitals with acute chest pain. Data from hospital 1 were used to train the models, which were then tested on independent data from the other 2 hospitals. From 40 potential factors, variables were selected according to the logarithm of their likelihood ratios to produce models using 8, 13, 20, and 40 factors. Identical data were used for logistic regression and artificial neural network models. Calibration and performance were assessed, the latter using receiver operating characteristic (ROC) curve analysis. Although the performance of artificial neural network models generally increased with increasing numbers of factors, this was insignificant. The 13-factor model was therefore used for the rest of the study owing to its marginally improved calibration over the smallest model. Area under the ROC curve (with standard error) was 0.97 (0.006). The overall sensitivity and specificity of this model for acute coronary syndrome diagnosis using the training data was 0.93. ROC curves for logistic regression and artificial neural network models applied to data from the 3 hospitals were identical. For the 13-factor artificial neural network model tested on data from hospitals 2 and 3, area under the ROC curves (standard error) were 0.93 (0.006) and 0.95 (0.009), respectively. Investigation of the performance of the artificial neural network models throughout the range of predicted probabilities showed that they were well calibrated. This study confirms that artificial neural networks can offer a useful approach for developing diagnostic algorithms for chest pain patients; however, the exceptional performance and simplicity of the logistic model militates in favor of logistic regression for the present task. Our artificial neural network models were well calibrated and performed well on unseen data from different centers. These issues have not been addressed in previous studies. However, and unlike in previous studies, we did not find the performance of artificial neural network models to be significantly different from that of suitably optimized logistic regression models.