RadOnco-Priority: Machine Learning Decision Support for Radiotherapy Queue Prioritization Using Real-World Retrospective Radiotherapy Referral Data

Abstract

Radiotherapy queues are often managed by referral date and manual clinician judgment, although limited linear accelerator capacity requires prioritization that is clinically transparent, operationally auditable, and fair. This study evaluates RadOnco-Priority, a machine learning-enabled decision support framework for radiotherapy queue prioritization, using a de-identified real-world retrospective dataset of 240 radiotherapy referral records rather than simulated or synthetic patient records. The system combines a literature-informed rule-based urgency score with supervised machine learning models to identify patients requiring accelerated booking. Accelerated booking need was defined a priori as an operational triage label reflecting clinician-documented priority, urgent symptoms, time-sensitive tumor-site and treatment-intent combinations, accumulated referral delay, and planning complexity. Logistic regression, random forest, and gradient boosting were trained to predict accelerated booking need, while a capacity-aware scheduling simulation evaluated waiting-time redistribution. To address potential circularity, additional ablation analyses were performed with the aggregate urgency score removed from the predictors. In the held-out test set, logistic regression achieved the highest discrimination in the full-feature model (AUC 0.91), with sensitivity-oriented classification favoring reduced false negatives. Performance remained acceptable after removing the aggregate urgency score, indicating that the model did not rely solely on the pre-specified scoring logic. The scheduling simulation reduced median waiting time in the high-priority group and decreased the proportion of high-priority patients waiting more than 28 days. These findings support RadOnco-Priority as an interpretable, human-governed information-system framework for radiotherapy queue management. Prospective multicenter validation, fairness monitoring, and local ethics approval remain required before routine implementation.