Machine Learning for Diabetes and Mortality Risk Prediction From Electronic Health Records

Abstract

Data science can provide invaluable tools to better exploit healthcare data to improve patient outcomes and increase cost-effectiveness. Today, electronic health records (EHR) systems provide a fascinating array of data that data science applications can use to revolutionise the healthcare industry. Utilising EHR data to improve the early diagnosis of a variety of medical conditions/events is a rapidly developing area that, if successful, can help to improve healthcare services across the board. Specifically, as Type-2 Diabetes Mellitus (T2DM) represents one of the most serious threats to health across the globe, analysing the huge volumes of data provided by EHR systems to investigate approaches for early accurately predicting the onset of T2DM, and medical events such as in-hospital mortality, are two of the most important challenges data science currently faces. The present thesis addresses these challenges by examining the research gaps in the existing literature, pinpointing the un-investigated areas, and proposing a novel machine learning modelling given the difficulties inherent in EHR data.

To achieve these aims, the present thesis firstly introduces a unique and large EHR dataset collected from Saudi Arabia. Then we investigate the use of a state-of-the-art machine learning predictive models that exploits this dataset for diabetes diagnosis and the early identification of patients with pre-diabetes by predicting the blood levels of one of the main indicators of diabetes and pre-diabetes: elevated Glycated Haemoglobin (HbA1c) levels. A novel collaborative denoising autoencoder (Col-DAE) framework is adopted to predict the diabetes (high) HbA1c levels. We also employ several machine learning approaches (random forest, logistic regression, support vector machine, and multilayer perceptron) for the identification of patients with pre-diabetes (elevated HbA1c levels). The models employed demonstrate that a patient's risk of diabetes/pre-diabetes can be reliably predicted from EHR records.


We then extend this work to include pioneering adoption of recent technologies to investigate the outcomes of the predictive models employed by using recent explainable methods. This work also investigates the effect of using longitudinal data and more of the features available in the EHR systems on the performance and features ranking of the employed machine learning models for predicting elevated HbA1c levels in non-diabetic patients. This work demonstrates that longitudinal data and available EHR features can improve the performance of the machine learning models and can affect the relative order of importance of the features.



Secondly, we develop a machine learning model for the early and accurate prediction all in-hospital mortality events for such patients utilising EHR data. This work investigates a novel application of the Stacked Denoising Autoencoder (SDA) to predict in-hospital patient mortality risk. In doing so, we demonstrate how our approach uniquely overcomes the issues associated with imbalanced datasets to which existing solutions are subject. The proposed model –– using clinical patient data on a variety of health conditions and without intensive feature engineering –– is demonstrated to achieve robust and promising results using EHR patient data recorded during the first 24 hours after admission.