Prediction of Outcome in the Critically Ill Using an Artificial Neural Network Synthesised By a Genetic Algorithm

This paper demonstrates that neural nets have the capacity to 'mould' themselves to data sets which relate to critically ill patients. Furthermore, they outperformed the conventional approach of logistic regression analysis and were successfully 'bred' for improved performance using genetic algorithms. Neural net analysis of databases relating to ICUs has already been performed by others. Doig et al [29] used 15 physiological parameters measured on 422 ICU patients to train a neural net for mortality-rate prediction. Whilst their performance figures were an apparent improvement over those obtained from a logistic regression model, the low frequency of mortality in their dataset precluded a valid comparison between the two methods. Buchman et al [30] also compared the performance of a neural net with a logistic regression model with regard to predicting the length of stay in ICU. The network, constructed using daily records from from 491 patients, outperformed the conventional approach (true positive rate 0.97 versus true positive rate 0.72). As concerns mathematical modelling in ICU for predicting outcome, Atkinson et al [31] used trend analysis based on APACHE II scores. Their performance figures cannot, however, be compared with table 3 since they did not select their ICU population for SIRS and shock; furthermore, many ethical questions concerning the use of such data remain unanswered [32]. Although the optimization of neural networks by genetic algorithms has been used successfully elsewhere within medicine [33], the use of genetic algorithms to improve the performance of neural nets in such complex clinical scenarios has not been previously proposed. 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Symbol Model ROC-curve area Mean Brier Random probability generator Logis..