A personalized prediction model for hospital readmission risk for cancer patients

7057 Background: Cancer patients (pts) are at high risk of unplanned hospital readmissions. Predicting which cancer patients are at higher risk of readmission would improve post-discharge follow-up/navigation, decrease cost, and improve pt outcomes. Methods: We conducted a retrospective cohort study of non-surgical cancer pts hospitalized at our center between 12/2014 to 7/2018. A machine learning algorithm was trained on 348 medical, sociodemographic and cancer-specific variables with a total of 1,801,944 data points. The cohort was randomly divided into training (80%) and validation (20%) subsets. Prediction performance was measured by area under the receiver operator characteristic curve (AUC). Results: A total of 5,178 hospitalizations were included, of which 45.1% were women, and 27.6% experienced an unplanned readmission within 30 days. The most frequently represented cancers were hematologic malignancies (30.5%), followed by GI (18.1%), lung (13.7%), and GU (10.9%). Significant variables that impacted the algorithm decision are ranked from the most to the least important, including: days from last admission; planned index chemotherapy admission; number of vascular access lines, drains, and airways in use; length of stay; cancer diagnosis; total ED visits in past 6 months; age; discharge lab values (sodium, albumin, alkaline phosphatase, bilirubin, platelets); number of prior admissions; and discharge disposition. The AUC for the validation subset was 0.80. To ease the translation of this model into the clinic, we developed a web application whereby users can supply the aforementioned variables to the model and receive a personalized prediction that highlights those variables most affecting a subject’s readmission risk status: www.Cancer-Readmission.com. Conclusions: A cancer-specific readmission risk model with high AUC for 30-days unplanned readmission has been developed. The model is embedded in a freely available web application that provides personalized, patient-specific predictions. Programs that integrate this model can identify cancer patients with a greater risk for unplanned hospital readmission, thus providing a personalized approach to prevent future unplanned readmissions.