Hospital Readmission Prediction with domain knowledge

Overview

The Australian Commission on Safety and Quality in Health Care (the Commission) has highlighted that reducing avoidable hospital readmissions supports better health outcomes, improves patient safety and leads to greater efficiency in the health system. Previous studies have reported that up to 11% of the Emergency (ED) population are "heavy users" with a higher prevalence of psychosocial problems and often co-existing chronic medical conditions.  All Australian governments have committed to reforms under the National Health Reform Agreement Addendum,1 and the ability to predict factors that influence readmission is of great importance as careful invention may reduce the occurrence of heavy use of services leading to greater efficiency and better health outcomes.

With increased adoption of Electronic Health Records (EHR) in hospitals across Australia, many analytic applications for predicting patient readmission have been proposed and suggested several influential predictors to readmissions. Existing models exhibit only moderate level of predictive performance and limit their approach to using aggregate features (e.g. the number of diseases), not considering interferences between diagnosis. It is common for patients to have one or more diagnoses, which is called comorbidity. Many studies have studied the importance of comorbidity and identified some associations between them to prevent or detect diseases. However, these existing approaches have not been applied to applications for hospital readmission prediction and therefore, this research will adopt frequent pattern and association mining to find patterns of patient characteristics that contribute to hospital readmission. In addition, as the Unified Medical Language System (UMLS) provides relationships between diagnoses, procedures and medications, this research will explore associations between them for improving prediction performance.

This research will investigate the characteristics associated with intensive hospital use by mining HER data.

Research engagement

• Comparison study on validating previous findings from clinical studies.
• Development of methods to identify latent relationships between features in EHR data.
• Design and implement a readmission prediction model based on the proposed algorithms.

Start date

18 November, 2024

End date

31 January, 2025

Contact

Contact the supervisor for more information.