以類神經網路模式輔助外科重症病人預測死亡機率

Background: The aim of this study was to investigate predictors of mortality inpatients admitted to a surgical Intensive Care Unit (SICU). Methods: An artificial neural network model was constructed on 588 consecutive patients treated in an ICU of a medical center during Jan to May, 2012. The prognostic factors, retrieved from IntelliVue Clinical Information Portfolio (ICIP), included days of ICU stay, causes of transfer to ICU, APACHE II score at 24hrs stay and 48hrs stay in ICU, and the day of discharge from ICU. The data was randomly divided into a 441 patient training dataset and a 147 patient testing dataset. The best predicting model was established using MLP (MLP 59-19-2); the output variable was death (1, yes; 2, no). Results: The accuracy of mortality prediction was 96.9%; the area under the ROC curve was 0.975; sensitivity was 78.8%, and specificity was 98.7%. Conclusion: Our results show disease severity is highly correlated with mortality. The accuracy, sensitivity, and specificity are all also very high. In addition, this study demonstrates patients with nonintracranial hemorrhage disclosed less unpredictable mortality, but doctors can easier predict poor outcome in patients with intracranial hemorrhage and thus recommend hospice care to the family.