Predicting Unplanned Hospitalization Among Patients Receiving Definitive Radiation.
A small fraction of cancer patients receiving definitive radiation experience unplanned hospitalizations related to treatment, which potentially leads to substantial morbidity or mortality. Identifying patients at high risk of unplanned hospitalizations could help with risk stratification, and potentially help inform future interventions aimed at improving patient outcomes. This study evaluated the predictive capacity of a machine learning prediction algorithm to identify unplanned hospitalizations and intensive care unit (ICU) visits among a large cohort of patients undergoing definitive radiation therapy.
We identified cancer patients undergoing definitive radiation therapy from the Veterans Health Administration (VHA) between 2000 and 2017. We used an extreme gradient boosting model to separately predict the risk of unplanned hospitalization and ICU admission within 30 days of completion of radiation treatment. Model covariates included patient demographics, cancer stage and subtype, and treatment information including previous chemotherapy and surgery. The data were split 75%/25% into training/testing datasets. We constructed the model with training data and evaluated performance within the test data using area under the curve (AUC), with an AUC of 1.0 indicating perfect prediction.
We identified 60,804 cancer patients from 14 tumor subtypes receiving definitive radiation therapy, of which 936 (1.5%) experienced an unplanned hospitalization with 113 (0.19%) escalated to the ICU. Our final predictive model included 30 features, which contained 10 demographic variables and 20 tumor or treatment characteristics. The predictive models achieved an AUC of 0.82 and 0.79 for predicting unplanned hospitalization and ICU visit, respectively. Emergency department visits prior to radiation treatment was identified as the most important associated feature for predicting both unplanned hospitalizations and ICU visits. Other features including tumor subtype, Charlson comorbidity score, previous chemotherapy, and biologic effective dose were also important predictors in the final models.
This study demonstrates that a machine learning model can help provide a robust prediction of unplanned hospitalizations and ICU admission among cancer patients receiving definitive radiation. Machine learning prediction models show promise in helping to identify patients at risk of severe adverse outcomes, though additional validation is needed as well as research studying how to effectively implement these tools into practice.
Qiao, EM; Qian, AS; Nalawade, V; Kotha, NV; Voora, RS; Murphy, JD
Volume / Issue
Start / End Page
Electronic International Standard Serial Number (EISSN)
International Standard Serial Number (ISSN)
Digital Object Identifier (DOI)