Patient classification with ensemble treebased modelling for decision support in acute clinical care settings

Decision Support is the collection of approaches that aim to support decision-making activities. It is used in clinical and health systems in order to ensure the provision of right care to right patients at the right time. Model-driven decision support relies on computational models and provides clinical users with analysis capabilities that can be used during the process of clinical decision-making. An important and widely used class of clinical decision support models is aimed at supporting diagnostic and prognostic decisions in acute patient care. This class of models is based on the use of statistical and machine learning classification techniques, and should assist clinicians with acute clinical decision making through classifying target patients or predicting treatment outcomes in an accurate and interpretable way.

Tree-based classification models are known to provide intuitively interpretable clinical insights, but their accuracy can be negatively affected by multiple intrinsic factors of underlying clinical data, such as sample size, number of input features used for classification task, the prognostic quality of these features, and the prevalence of the clinical outcome of interest. To overcome this limitation presented in Chapter 1, ensemble modelling can be used to improve classification performance by combining a set of individual classification models into a single ensemble classification model, for example, as is done in Random Forest (a widely used tree-based ensemble modelling technique as reviewed in Chapter 2). For existing ensemble modelling techniques, the improvement in classification performance comes at the expense of increased complexity and reduced model interpretability.

Overall, this thesis identifies the following gaps in the existing knowledge: 1. There is a lack of knowledge of whether and how acute clinical context and clinical features influence the design and development of the ensemble tree-based classification model for decision support in acute clinical care settings. 2. There is a lack of knowledge of potential benefits from the determination of the class membership in the ensembling level instead of in the base level as done with Random Forest. 3. There is a lack of knowledge of how to combine the classification results from various binary class classifications to provide systematic multi-class classification results to support acute clinical decision making. 4. There is a lack of knowledge of how the performance of ensemble tree-based models can be evaluated in imbalanced datasets.

In response to these gaps, in this thesis, a novel ensemble tree-based modelling framework is proposed in Chapter 3 to provide accurate and interpretable classification results to help clinicians with timely decision making in acute clinical settings. The proposed approach differs from the Random Forest approach in several ways: how the input features for classification or prediction tasks are selected, how the individual tree-based classification or prediction models are trained, and what information is contributed by individual models to the overall ensemble model.

The performance of the proposed novel ensemble tree-based modelling framework is investigated and compared to that of Random Forest models in Chapter 4 through extensive computational experimentation, where intrinsic characteristics of the simulated datasets such as sample size, number of input features used for classification or the prediction task, the prognostic quality of these features, and the prevalence of the clinical outcome of interest are systematically varied to affect the accuracy of the classification model within the settings of a full-factorial experiment. The original contribution to knowledge, based on these computational experiments, are that the classification performance of tree-based models is highly correlated with the above intrinsic factors of training datasets.

Subsequently, the proposed novel ensemble tree-based modelling framework is applied to provide decision support in three different real-world acute care case studies.

The first case study presented in Chapter 5 is to assist with classification of paediatric stroke. Differentiation of stroke from stroke mimics in a paediatric population is very challenging in acute clinical settings as many acute brain attack conditions can mimic stroke and their differences have not been explored by previous clinical studies. Accurate classification of stroke subtypes and stroke mimics can improve prioritising neuroimaging access and reduce the time from stroke symptom onset to the initiation of hyper-acute treatment. The ensemble tree-based models developed through the proposed framework display excellent performance on classifying stroke subtypes and stroke mimics.

The second case study presented in Chapter 6 is to assist with identification of patients at high risk of deterioration post-surgery using Medical Emergency Team (MET)-Call system. MET is a rapid response team to provide appropriate treatment to patients experiencing sudden adverse event in hospital. However, there is no recognized trigger that can be used to identify patients who will require MET assistance. The proposed ensemble tree-based modelling framework has been applied in this acute care clinical setting to develop a decision support model that is capable of identifying patients at high risk.

The third case study presented in Chapter 7 is to assist with prediction of neurological and functional recovery following acute endovascular treatment in adult stroke patients. Compared with Random Forest models, the original ensemble tree-based models can provide more accurate results on classifying patients with different status of functional and neurological recovery.

Through these case studies, as summarised in Chapter 8, this thesis provides an illustration of how the proposed ensemble tree-based modelling framework can be used to develop classification models across multiple acute clinical settings, each with their own system requirement. This thesis demonstrates how the ensemble tree-based modelling framework can be integrated into the clinical system to assist clinicians with acute clinical decision making. This thesis also establishes the position of the proposed novel ensemble tree-based modelling framework as a viable alternative to Random Forest models for providing accurate and interpretable classification for acute care decision support.