Identification of patients at high risk for preventable emergency department visits and inpatient admissions after starting chemotherapy: Machine learning applied to comprehensive electronic health record data

1511 Background: Acute care use is one of the largest drivers of cancer care costs. OP-35: Admissions and Emergency Department Visits for Patients Receiving Outpatient Chemotherapy is a CMS quality measure that will affect reimbursement based on unplanned inpatient admissions (IP) and emergency department (ED) visits. Targeted measures can reduce preventable acute care use but identifying which patients might benefit remains challenging. Prior predictive models have made use of a limited subset of the data available in the Electronic Health Record (EHR). We hypothesized dense, structured EHR data could be used to train machine learning algorithms to predict risk of preventable ED and IP visits. Methods: Patients treated at Stanford Health Care and affiliated community care sites between 2013 and 2015 who met inclusion criteria for OP-35 were selected from our EHR. Preventable ED or IP visits were identified using OP-35 criteria. Demographic, diagnosis, procedure, medication, laboratory, vital sign, and healthcare utilization data generated prior to chemotherapy treatment were obtained. A random split of 80% of the cohort was used to train a logistic regression with least absolute shrinkage and selection operator regularization (LASSO) model to predict risk for acute care events within the first 180 days of chemotherapy. The remaining 20% were used to measure model performance by the Area Under the Receiver Operator Curve (AUROC). Results: 8,439 patients were included, of whom 35% had one or more preventable event within 180 days of starting chemotherapy. Our LASSO model classified patients at risk for preventable ED or IP visits with an AUROC of 0.783 (95% CI: 0.761-0.806). Model performance was better for identifying risk for IP visits than ED visits. LASSO selected 125 of 760 possible features to use when classifying patients. These included prior acute care visits, cancer stage, race, laboratory values, and a diagnosis of depression. Key features for the model are shown in the table. Conclusions: Machine learning models trained on a large number of routinely collected clinical variables can identify patients at risk for acute care events with promising accuracy. These models have the potential to improve cancer care outcomes, patient experience, and costs by allowing for targeted preventative interventions. Future work will include prospective and external validation in other healthcare systems.[Table: see text]