부산대학교 도서관

This doctoral thesis aims to investigate the application of different machine learning techniques to improve outcome prediction in the trauma field. The thesis covers three topics to fulfil that aim. Firstly, an interpretable machine learning (IML) method based on vital sign variables for predicting trauma patients' outcomes is developed. It contains different methods, including a maximum likelihood evidential reasoning framework, belief rule-based inference methodology based on evidential reasoning, and a non-linear optimisation for parameter tuning. Common ML techniques have been applied to find the most accurate model for trauma outcome prediction. Furthermore, to enhance the prediction of trauma outcome, the prediction accuracy of multiple models based on vital sign features has been evaluated, as vital signs features are commonly collected in trauma centres and units. Secondly, the evidential reasoning (ER) rule is introduced for feature selection to highlight the key features impacting the outcomes. The ER rule finds the optimal weight for each feature that maximises the prediction accuracy during model training. Other feature selection methods have also been implemented, such as random forest and ReliefF. Thirdly, the ER rule has been applied for ensemble learning and has the advantage of adjusting the weight for each classifier in the ensemble learning process. In this thesis, two sets of trauma data are acquired to implement the proposed techniques. The results show that the IML method improves prediction accuracy over other common ML techniques. Similarly, the ER rule achieves good prediction accuracy after ensemble learning. The results highlight the role of the proposed feature selection techniques in finding the key predictors of patients' outcomes.