Predicting cancer patient mortality within 30 days of radiotherapy consultation to inform palliative radiotherapy fractionation decisions

Abstract

Background Radiation and medical oncologists evaluate patients' risk of imminent mortality with scales like Karnofsky Performance Status (KPS) and predicate treatment decisions on these evaluations. However, we hypothesized that statistical models derived from structured electronic health record (EHR) data could predict patient deaths within 30 days of radiotherapy consultation better than models developed only with patient age and physician-reported KPS. Methods Clinical data from patients who consulted in a radiotherapy department from June 2018 - February 2024 were abstracted from EHR databases, including patient demographics, laboratory results, medications, comorbidities, KPS, cancer stages, oncologic treatment histories, oncologist notes, radiologist reports, and pathologist narratives. A subset of structured features known or believed to be associated with mortality were curated and used to train and test logistic regression, random forest, and gradient-boosted decision classifiers. Results Of 38,262 patients, 951 (2.5%) died within 30 days of radiotherapy consultation. From 34.5 gigabytes of tabular data, 2,977 clinical features were chosen or derived by a radiation oncologist, then reduced to 1,000 features using ANOVA F values. Using an event probability classification threshold of 0.2, optimized logistic regression, random forest, and gradient-boosted decision classifiers tested with high accuracy (0.97, 0.98, and 0.98, respectively) and F1 scores (0.50, 0.54, and 0.52). The areas under receiver operating and precision-recall curves for the random forest model were respectively 0.94 and 0.55, which outperformed a model trained only with patient age and KPS (0.61 and 0.06). Models prominently weighed features that were rationally associated with mortality. Conclusion Statistical models developed from a physician-curated feature space of structured EHR data predicted patient deaths within 30 days of radiotherapy consultation better than a model developed only with a patient's age and physician-assessed KPS. With clinically explicable feature weights, these models could influence treatment decisions such as the length of palliative radiotherapy courses.