S1417
Interdisciplinary - Health economics & health services research
ESTRO 2026
Kanewoff 2 , Stefan Jovinge 5,6 , Per Munck Af Rosenschöld 7,2
intended radiotherapy prescription, and hospitalization within the last month.
1 DigIT/MT, Region Skåne, Lund, Sweden. 2 Hematologi Onkologi Strålningsfysik SUS, Region Skåne, Lund, Sweden. 3 Signal processing and Biomedical engineering, Chalmers, Gothenburg, Sweden. 4 AI Adoption, AI Sweden, Lund, Sweden. 5 Department of Clinical Sciences, Lund University, Lund, Sweden. 6 Skåne University Hospital, Region Skåne, Lund, Sweden. 7 Medical Radiation Physics, Lund University, Lund, Sweden Purpose/Objective: The aim of this study was to 1) minimize the risk for individual patients missing their radiotherapy appointment, and 2) ensure better utilization of radiotherapy capacity so that patients in queue do not miss a chance for earlier treatment. Therefore, a Machine Learning (ML) model was developed predicting the risk that a radiotherapy (RT) series start would be cancelled late (<7 days, including rescheduling and no-show). The performance of the model was evaluated and the most influential risk factors for late cancellation according to the model were determined. Material/Methods: Data from 28,000 RT series from 2019-2025 were used in ML model development. In total ~130 variables, from three data sources (oncology information system, patient administrative system, and the patient medical journal) were aggregated and presented to the model. Variables used in the model training were, among others, demographic variables, previous treatments, cancer type, other diagnoses, hospitalization, and prescribed radiotherapy. Predictions were made 7 days ahead of treatment start. A range of models (XGBoost, Random Forest) were trained, with varying data selections. Evaluation was made on 1,400 RT series starts made after those in the training data, no overlapping patients, with overall late-cancellation rate at 8%. The 50% with highest risk for late cancellation were grouped into medium-or-high risk, the remaining low risk. Metrics for evaluation were ROC-AUC and Positive Predictive Value (PPV, precision), recall and Negative Predictive Value (NPV). Results: The best performing model was an XGBoost model trained with data from 2022-2025 (15,000 RT series), with hyperparameters selected using Bayesian search. This model had an ROC-AUC at 0.71. In evaluation, the rate of late cancellations for medium-or-high risk was 12% (PPV, precision), and in the low-risk it was 3,5% (1- NPV). Of all late cancellations, 77% had medium-or- high risk (recall). The four most predictive features were: Days between booking and start, number of previously cancelled RT series starts, palliative-
Conclusion: Our model achieved moderate classification performance, despite having a large and well-curated data set. This indicates that predicting late cancellations for radiotherapy is a challenging problem, and a reason may be that a multitude of cancellation reasons apply: disease progression, treatment change, death and patient’s own request for cancellation to name a few. The late-cancellation risk predictions may help prioritization for interventions aimed at reducing late cancellations. Keywords: Radiotherapy, Machine Learning, Cancellation Risk
Proffered Paper 3278
Geographic accessibility of proton therapy services in Europe: current status and opportunities for enhancement
Dominik Wawrzuta, Katarzyna Ludwikowska, Katarzyna P ę dziwiatr, Marzanna Chojnacka
Made with FlippingBook - Share PDF online