AbstractHospital readmissions lead to greater demand for healthcare resources, financial costs, and poorer patient outcomes. They are also often preventable with improved care and management. This has led to their use as a quality-of-care indicator and the development of healthcare policy linking readmission outcomes to hospital funding in the USA, England, Germany, and most recently in Australia. The negative consequences of avoidable readmissions for funding and patient welfare have also spurred development of computational models predicting readmission risk to enable hospitals to identify high-risk patients for interventions.
These prediction models have been overwhelmingly characterised by classification approaches focusing on predicting readmission status at a single fixed time after discharge, most commonly 30 days. Research developing and validating these models is ongoing, driven by the poor performance of models currently used and the need for customisation to specific regions, populations, and conditions. To improve the performance of these models, research has increasingly considered machine learning techniques and leveraged novel data sources. Despite the additional information provided by survival models compared with classification models, survival approaches have received little attention with respect to available machine learning techniques, practical applications, and appropriate performance measures.
This research identified available and relevant machine learning survival techniques, including decision trees, ensembles, and artificial neural networks. The value of previously unconsidered machine learning survival techniques was investigated for predicting 30-day unplanned readmissions. This investigation considered adult patients admitted to hospital through the emergency department of Gold Coast University Hospital (n = 46,659) and Robina Hospital (n = 23,976) in Queensland, Australia. The value of both statistical and machine learning survival models for novel applications supporting managerial decision-making were also investigated. The proposed applications leverage survival predictions to dynamically rank patients by risk, account for patient-specific risk profiles, and forecast future readmissions. The important aspects of model performance for such applications were determined to be discrimination and calibration of predictions over time. Time-dependent concordance and D-calibration were identified as appropriate metrics capturing these aspects of model performance.
For the more complex population (Robina Hospital), machine learning survival models improved on statistical survival models for both 30-day readmission and risk over time prediction. Even compared to the benchmark classification model, select machine learning models exhibited competitive discrimination and better calibration in predicting 30-day readmissions. These models should be considered when developing tools supporting readmission management under classification approaches as well as survival approaches. The proposed model applications under survival approaches were demonstrated to be feasible, with varying levels of discrimination but consistent calibration across both machine learning and survival models. These models should benefit hospitals managing readmissions through better intervention targeting, follow-up care customisation, and demand forecasting. This in turn should lead to reduced costs and better outcomes for patients. The area is also advanced more generally through the highlighting of available machine learning techniques, applications, and performance measures under survival approaches.
|Date of Award||9 Feb 2022|
|Supervisor||Bruce Vanstone (Supervisor), Adrian Gepp (Supervisor) & Steven Stern (Supervisor)|