Towards More Nuanced Patient Management: Decomposing Readmission Risk with Survival Models

Unplanned hospital readmissions are costly and associated with poorer patient outcomes. Overall readmission rates have also come to be used as performance metrics in reimbursement in healthcare policy, further motivating hospitals to identify and manage high-risk patients. Many models predicting readmission risk have been developed to facilitate the equitable measurement of readmission rates and to support hospital decision-makers in prioritising patients for interventions. However, these models consider the overall risk of readmission and are often restricted to a single time point. This work aims to develop the use of survival models to better support hospital decision-makers in managing readmission risk. First, semi-parametric statistical and nonparametric machine learning models are applied to adult patients admitted via the emergency department at Gold Coast University Hospital (n = 46,659) and Robina Hospital (n = 23,976) in Queensland, Australia. Overall model performance is assessed based on discrimination and calibration, as measured by time-dependent concordance and D-calibration. Second, a framework based on iterative hypothesis development and model fitting is proposed for decomposing readmission risk into persistent, patient-specific baselines and transient, care-related components using a sum of exponential hazards structure. Third, criteria for patient prioritisation based on the duration and magnitude of care-related risk components are developed. The extensibility of the framework and subsequent prioritisation criteria are considered for alternative populations, such as outpatient admissions and specific diagnosis groups, and different modelling techniques. Time-to-event models have rarely been applied for readmission modelling but can provide a rich description of the evolution of readmission risk post-discharge and support more nuanced patient management decisions than simple classification models.