Predicting 30-Day Hospital Readmissions among Patients with Type-2 Diabetes: A Comparison Between Machine Learning Classification Methods

The rising prevalence of diabetes worldwide presents a challenge for healthcare providers, as do the rising costs related to the disease. Among hospitalised patients who are readmitted within 30 days of discharge, individuals with diabetes are overrepresented, with comorbidities further increasing the risk of readmission. Identifying factors influencing readmissions can improve patient management, reduce healthcare costs, and optimize resource allocation, highlighting the need for predictive models to accurately identify high-risk diabetic patients. The thesis investigates factors influencing 30-day hospital readmission rates among adult patients diagnosed with type 2 diabetes and comorbidities. Patient data is leveraged to compare the effectiveness of three classifiers – logistic regression, support vector machines and random forest – in predicting early hospital readmissions. It is concluded that a high number of medications, diagnoses and healthcare visits in the year before the hospital stay are associated with a higher probability of readmission, while a change in medication implies the opposite. No clear connection between comorbidities and hospital rates is found. Moreover, the results indicate a low predictive power of the models, with logistic regression slightly outperforming the other classifiers. Thus, the need for accurate predictive models identifying patients at risk of readmission remains.