Diabetes, which ranks 4th among the top ten causes of death in Taiwan, is
one of the most important medical issues in the 21st century. However, for most diabetic
patients, their blood sugar is not under good control, especially in ICUs. Good control of
blood sugar may reduce the risk of sepsis to 34% for patients in ICUs. The importance of
good blood sugar control for patients in ICUs is manifested by a signi cant reduction of
mortality and morbidity. This study designed a clinical decision support system (CDSS)
using a support vector machine (SVM) to predict if a critically ill patient can have good
glucose control after insulin administration. A lter method based on logistic regression
analysis (LRA) and a wrapper method based on recursive feature elimination (RFE) were
adopted to select salient features from 10 variables for CDSS design. Data on 231 patients
(2492 records) were collected covering four years from an ICU. Four signi cant variables
(p < 0:05) using LRA in contrast to ve ones using RFE algorithm, were selected. The
results show that the predictive accuracy under cross-validation was 93.50% for features
selected with LRA, and the accuracy, sensitivity and speci city with SVM-RFE were
95.75%, 92.71% and 99.81%, respectively. It could predict the outcome quite accurately
after having injected a certain dose of insulin. The proposed system may help doctors
effectively assess their patients in determining insulin dose for better glucose control in
an ICU setting.
關聯:
International Journal of Innovative Computing Information and Control 8(1(B)):779-790
