Developing risk of Mortality and early warning score models using routinely collected data
Student thesis: Doctoral Thesis
Aim. The aim of this study was to contribute to the building of effective and efficient methods to predict adverse clinical outcome. It has been done by developing risk of mortality and early warning score models using routinely collected data that are available from hospital computer systems. Methods. To predict risk of mortality, firstly we used logistic regression using (Biochemistry and Haematology Outcome Model - BHOM dataset) to generate a model, and the performance of each model was then compared using discrimination (AUROC or c-index) and calibration (the Hosmer- Lemeshow test). Secondly, we focused on decision trees (DT) to be compared with logistic regression (LR). In addition, we used cross validation to compare LR with other various machine learning methods. We developed early warning score algorithmically using decision trees (DTEWS) using vital sign dataset and compared the performance of DTEWS with other EWSs based on clinical expertise using c index, early warning score efficiency curve and distribution score. We also compared DTEWS with another EWS based on statistics and applied DTEWS to BHOM dataset. Results. In BHOM dataset, there were 9497 adult hospital discharges, and it was divided into four subsets. A model was built using one training set and then applied to three other testing data sets. The model in logistic regression satisfied both discrimination and calibration value when the c-index in the range 0.700-0.800 is reasonable discrimination and the p-value > 0.05 indicates there is no evidence of significant lack of fit. We also found that decision trees gave a satisfactory result followed by some other machine learning methods.Using a large vital signs dataset (n = 198,755 observation sets) from acute medical admissions, DTEWS can provide a discrimination (c-index) as good as other EWSs, has a better c-index, and also is better in other measurements including EWS efficiency curve, and distribution of score. We found DTEWS can also be applied to BHOM dataset with satisfactory results. Conclusion. The results of this study support the idea that decision trees can be applied to medical problems. When we produced a model for risk of mortality, we have shown that the decision trees model has reasonable discrimination and could be considered as an alternative technique to logistic regression. We have shown that a structured methodology using decision trees to develop early warning score has satisfactory result and contributes additional evidence that suggests an algorithmical method can be employed to quickly produce EWSs for employment in particular types of medical purpose.
3.26 MB, PDF document