2789 - Machine Learning-Based Prediction of Acute Care during Radiation Therapy Using the NCCN Distress Thermometer

J. Ejie^1,2, S. C. D. Hampson¹, M. Ho^2,3, M. Elia³, I. Friesner¹, J. J. Chen⁴, C. I. Nnadi⁵, J. Chew⁴, N. W. Cho⁴, H. Vasudevan⁴, L. Boreta⁴, S. Sinha⁴, S. E. Braunstein⁴, and J. C. Hong⁴; ¹University of California, San Francisco, Bakar Computational Health Sciences Institute, San Francisco, CA, ²University of California, San Francisco, School of Medicine, San Francisco, CA, ³Bakar Computational Heath Sciences Institute, University of California San Francisco, San Francisco, CA, ⁴Department of Radiation Oncology, University of California San Francisco, San Francisco, CA, ⁵University of California San Francisco, San Francisco, CA

Purpose/Objective(s): A cancer diagnosis, coupled with its associated treatments, including surgery, chemotherapy, and/or and radiation therapy (RT), can be a profound stressor that impacts mental and physical health. We previously demonstrated that an electronic health record (EHR)-based machine learning (ML) model can accurately triage patients and reduce acute care utilization. In this study, we hypothesized that the NCCN DT would identify patients at high risk for acute care and complement an EHR-based approach. Materials/

Methods: This single-institution, retrospective cohort study included patients who underwent RT from 2019 to 2021. NCCN DT scores were extracted through retrospective chart review. A ML approach to predict acute care (emergency visits and hospitalizations) based on RT treatment and high dimensional EHR data was generated. A logistic regression model based on NCCN DT scores was generated to predict occurrence of an acute care event during RT. We then trained two Least Absolute Shrinkage and Selection Operator (LASSO)-based logistic regression models using EHR data with and without NCCN DT. All three models were trained on a randomly selected training cohort (75%) and tested on a separately held-out test cohort (25%). Models were evaluated for discrimination using area under the receiver operator characteristic curve (AUC), which combines sensitivity and specificity. A calibration curve was visualized.
Results: 1147 patients completed the NCCN DT prior to or during the first week of RT, of whom 605 (45.9%) were male and the median age was 62 (IQR 56.3-73.4). The median NCCN score was 2 (IQR 0-5). In total, 39 patients (3.4%) experienced an acute event during their course. On logistic regression, NCCN DT was associated with increased risk of acute care utilization (OR 1.19 per point, 95% CI 1.05-1.34, p = 0.045). Model performance was good (AUC 0.693 [95% CI 0.455-0.931]). The EHR-based model demonstrated superior performance (0.873 [0.780-0.965]). In the combined LASSO model, variable selection selected NCCN DT as a model feature over other EHR variables, and NCCN DT modestly improved model performance (0.892 [0.818-0.967]).
Conclusion: The NCCN DT has a strong correlation with acute care events during RT, suggesting a relationship between patient distress and acute care requirements. However, it has limited discriminatory ability to identify high risk patients, when compared to an EHR-based approach, though provided additional information to improve predictions when combined with an EHR-based approach. Further investigation is warranted, particularly given the limited sample sizes and events in this study.