LAUSR

271 Background: Unplanned hospitalizations may diminish quality of care among cancer patients receiving radiotherapy (RT). In patients undergoing RT for gastrointestinal (GI) cancers, we hypothesized that a machine learning approach would enable prediction of unplanned hospitalizations within 30 days of RT. Methods: We analyzed 836 abdominal (gastric, pancreatic, biliary, hepatic) and 514 pelvic (rectal, anal) courses of RT for GI cancers treated at our institution (3/2016—1/2019). Over 700 clinical/treatment variables and unplanned hospitalizations during or within 30 days after RT were mined from institutional databases. Using machine learning, we developed random forest (RF), gradient boosted decision trees (XGB), and logistic models for unplanned hospitalizations. Models were trained on 670 abdominal and 423 pelvic cases. Five-fold cross-validation (CV) was used to select model type and hyperparameters, using area under the ROC curve (AUC) to measure performance. The best model was validated on the subsequent 166 abdominal and 91 pelvic cases. AUC>0.70 was deemed clinically valid. Results: Among 1,350 cases, incidence of 30-day unplanned hospitalization was 12.3% (13.3% abdominal cohort; 10.7% pelvic cohort). Model CV AUCs are shown in table. The best models were XGB and RF for the abdominal and pelvic cohorts, respectively. Their validation testing AUCs are shown in table. For all models tested, lab values (e.g. potassium, lipoproteins, hemoglobin) prior to RT were significant predictors of unplanned hospitalizations. In the abdominal cohort, pancreatic primary and total RT dose were important. For the pelvic cohort, body mass index was important. Median healthcare costs from RT start - 30 days post-RT were $69,108 in non-hospitalized patients and $119,844 in hospitalized patients. Conclusions: In GI cancer patients undergoing RT, a machine learning model identified patients at risk of 30-day unplanned hospitalization. Predictive analytics may be a key tool to help providers identify high-risk patients and optimize interventions, while improving quality and value of care. [Table: see text]