Using Machine Learning Methods to Improve Healthcare Delivery in Diabetes Management

Abstract

This dissertation includes three studies, all focusing on Analytics and Patients information for improving diabetes management, namely educating patients and early detection of comorbidities. In these studies, we develop topic modeling and artificial neural network to acquire, preprocess, model, and predict to minimize the burden on diabetic patients and healthcare providers.

The first essay explores the usage of Text Analytics, an unsupervised machine learning model, utilizing the vast data available on social media to improve diabetes education of the patients in managing the condition. Mainly we show the applicability of topic modeling to identify the gaps in diabetes education content and the information and knowledge needs of the patients. While traditional methods of the content decision were based on a group of experts' contributions, our proposed methodology considers the questions raised on social forums for support to extend the education content.

The second essay implements Deep Neural Networks on EHR data to assist the clinicians in rank ordering the potential comorbidities that the specific patient may develop in the future. This essay helps prioritize regular screening for comorbidities and rationalize the screening process to improve adherence and effectiveness. Our model prediction helps identify diabetic retinopathy and nephropathy patients with very high precision compared to other traditional methods. Essays 1 and 2 focus on Data Analytics as a research tool for managing a chronic disease in the healthcare environment.

The third essay goes through the challenges and best practices of data preprocessing for Analytics studies in healthcare. This study explores the standard preprocessing methodologies and their impact in the case of healthcare data analytics. Highlights the relevant modifications and adaptations to the standards CRISP_DM process. The suggestions are based on past research and the experience obtained in the projects discussed earlier in the thesis.

Overall, the dissertation highlights the importance of data analytics in healthcare for better managing and diagnosing chronic diseases. It unfolds the economic value of implementing state-of-the-art IT methods in healthcare, where EHR & IT are predominantly costly and difficult to implement. The dissertation covers ANN and text mining implementation for diabetes management.