부산대학교 도서관

Background: The prediction of individual patient risk has important implications in the process of informed consent. As there are no predictive indices that objectively quantify operative risk at the patient level, we compared the performance of artificial neural networks (ANN) with hierarchical logistic regression (HLR) in predicting outcomes in oesophagogastric surgery. Methods: POSSUM variables were collected from 505 patients undergoing elective (n = 339, 67.1 per cent) or emergency (n = 43, 32.9 per cent) oesophagogastric surgery in five UK hospitals between 1995 and 1999. Backpropagation ANN models were trained using genetic algorithms (18:4:6:1 network topology; input-hidden χ2 output nodes) and a 60:15:25 per cent split sample-validation-test technique adopted for model evaluation and comparison with the HLR equivalent. Bayesian analysis was used to assess model validity. Results: The ANN completed training following 18 generations with final mean squared errors of 0.056 and 0.065 in the development and test sets. HLR models performed better in group statistics. The ANN model offered better discrimination between high-risk and low-risk patients as compared to the HLR model (area under the ROC curve = 0.854 versus 0.838 development set; 0.937 versus 0.866 test set). Based on 9.3 per cent overall mortality (prior odds = 0.101) the likelihood ratios for a positive test at 10, 20, 30, 40 and 50 per cent probability of death were 1.1, 2.5, 4.3 and 10.2 for the ANN model. Conclusion: The ANN were significantly better in individualizing outcomes whereas statistical techniques were best suited for evaluating group outcomes. With additional training, ANN may become attractive adjuncts to clinical management and decision-making process. [ABSTRACT FROM AUTHOR]