Abstract:

Predicting inpatient discharges has become an important operational problem as hospitals continue to experience frequent capacity crises caused by, among other factors, an imbalance of the timing of patient admissions and discharges. Often during capacity crises, what is needed beyond a list of discharge-ready patients is another list of potentially dischargeable patients. Having this short list can have a tremendous impact on the level of disruption to the hospital operations. We propose a prescriptive approach to identify such a list of dischargeable patients and prescribe associated interventions for their timely transition out of the hospital. The proposed approach starts with representing patients' clinical and administrative barriers to discharge in a clinically interpretable way, followed by building a neural network model to predict discharge likelihood of a given patient within 24 hours. Finally, using the trained neural network parameters, we employ a mixed integer programming model to identify the minimal subset of barriers that needs to be resolved, i.e., prescribed interventions, in order to lift a patient's discharge likelihood above a given threshold. We show the effectiveness of the proposed approach in a series of retrospective analyses conducted at the Massachusetts General Hospital. Specifically, the fraction of patients who are discharged within 48 hours (a measure of clinical readiness) were shown to be significantly higher among the patients identified by the model as actionable compared to that among non-actionable patients. Furthermore, the fraction of prescribed barriers resolved by the time of discharge (a measure of the importance of the barrier) was shown to be significantly higher compared to that of non-prescribed barriers.