OBJECTIVES: (1) To predict at the time of diagnosis of sepsis the subsequent occurrence of multiple organ failure and patient death; and (2) to compare the prediction accuracies of standard multiple logistic regression (MLR) and neural network (NN) models. METHODS: The data were collected during a 5-year period for all patients (n=173) who met prospectively determined criteria for sepsis and had positive blood culture results while admitted in the surgical intensive care unit at the University Hospital of Geneva, Switzerland. These data formed the basis for a retrospective cohort study described elsewhere. The MLR model was adapted from existing data. An NN model of the feed-forward, back-propagation type was constructed for predicting the outcome of sepsis with bloodstream infection. Both models were constructed from randomly chosen subsets of patients and subsequently were evaluated on the remaining (independent) patients. RESULTS: Survival after sepsis was predicted with an accuracy of 80% by the NN model, which used only information collected at the time of the diagnosis of sepsis. The development of multiple organ failure after the diagnosis of sepsis was predicted accurately (81.5%) with either the MLR or the NN model. Both the MLR and the NN methods depended on the interpretation of a likelihood quantity, requiring the choice of a threshold to make a survival prediction. The accuracy of the MLR models was very sensitive to the threshold value. The accuracy of the NN models was not sensitive to the choice of threshold, because they generated likelihood predictions that were distributed far from the middle range where the threshold was placed. CONCLUSION: Compared with MLR models, the NN models were slightly more accurate and much less sensitive to the arbitrary threshold parameter.
OBJECTIVES: (1) To predict at the time of diagnosis of sepsis the subsequent occurrence of multiple organ failure and patient death; and (2) to compare the prediction accuracies of standard multiple logistic regression (MLR) and neural network (NN) models. METHODS: The data were collected during a 5-year period for all patients (n=173) who met prospectively determined criteria for sepsis and had positive blood culture results while admitted in the surgical intensive care unit at the University Hospital of Geneva, Switzerland. These data formed the basis for a retrospective cohort study described elsewhere. The MLR model was adapted from existing data. An NN model of the feed-forward, back-propagation type was constructed for predicting the outcome of sepsis with bloodstream infection. Both models were constructed from randomly chosen subsets of patients and subsequently were evaluated on the remaining (independent) patients. RESULTS: Survival after sepsis was predicted with an accuracy of 80% by the NN model, which used only information collected at the time of the diagnosis of sepsis. The development of multiple organ failure after the diagnosis of sepsis was predicted accurately (81.5%) with either the MLR or the NN model. Both the MLR and the NN methods depended on the interpretation of a likelihood quantity, requiring the choice of a threshold to make a survival prediction. The accuracy of the MLR models was very sensitive to the threshold value. The accuracy of the NN models was not sensitive to the choice of threshold, because they generated likelihood predictions that were distributed far from the middle range where the threshold was placed. CONCLUSION: Compared with MLR models, the NN models were slightly more accurate and much less sensitive to the arbitrary threshold parameter.