Prediction of outcome in critically ill patients using artificial neural network synthesised by genetic algorithm

Background Decisions about which patients to admit to intensive care and how long to keep them there are difficult. A flexible computer-based mathematical model which is sensitive to the complexity of intensive care medicine, and which accurately models prognosis, seems highly desirable. Methods We have created, optimised by genetic algorithms, trained, and evaluated the performance of an artificial neural network (ANN) in the clinical setting of systemic inflammatory response syndrome and haemodynamic shock. 258 patients were selected from an intensive care database of 4484 patients al a London teaching hospital and randomised to a network training set (168) and a test set (90). The outcome evaluated was death during that hospital admission and the performance of the neural net was compared (by receiver operating characteristic [ROC] curves and by Brier scores) with that of a logistic regression model. Findings Artificial neural network performance increased with successive generations; the best-performing ANN was created after 7 generations and predicted outcome more accurately than the logistic regression model (ROC curve area 0.863 vs 0.753). Interpretation In this study, ANNs have lent themselves particularly well to modelling a complex clinical situation; we suggest that this relates to their inherently flexible nature which accommodates interactions between the clinical input fields. In addition, we have demonstrated the value of a second computational technique (genetic algorithms) in ''tuning'' ANN performance. These techniques can potentially be implemented in individual intensive care units; the outcome models which they will generate will be sensitive to local practice. Analysis of such accurate clinical outcome models may empower clinicians with a hitherto unappreciated degree of insight into those elements of their clinical practice which are most relevant to their patients' outcome.