Machine learning model for predicting DIBH non-eligibility in left-sided breast cancer radiotherapy: Development, validation and clinical impact analysis.

Objective: Multi-day assessments accurately identify patients with left-sided breast cancer who are ineligible for irradiation in Deep Inspiration Breath Hold (DIBH) and minimise on-couch treatment time in those who are eligible. The challenge of implementing multi-day assessments in resource-constrained settings motivated the development of a machine learning (ML) model using data only from the 1st day of assessment to predict DIBH ineligibility.

Methods: This prospective cohort study used data from 202 patients collected between January and December 2023 for model development. Patient-related and DIBH assessment-related variables (upper, lower, and average breath-hold amplitude; average breath-hold duration; breath-hold consistency) were included. Nine ML algorithms (and three modelling strategies) were evaluated, and a decision curve analysis was used to select the best model. The best model was temporally validated on a prospective dataset of 47 patients (January to March 2024). Further, a clinical impact study on another prospective cohort of 64 patients (April to August 2024) was performed, to assess its practical utility by comparing its predictions with the clinical team's decision to treat a patient in DIBH or not.

Results: The uncalibrated gradient-boosting ensemble model demonstrated the highest performance [AUC (95 % CI) = 0.803 (0.686-0.941); Recall = 0.526] and net benefit in decision curve analysis. Key predictors included average breath-hold duration and lower breath-hold amplitude levels. The clinical impact study suggests that the model reduces the need for additional DIBH assessments by up to 20 % without misclassifying eligible patients.

Conclusion: The developed ML model accurately predicts DIBH ineligibility using only first-day DIBH assessment data and could be a decision aid for patient selection in resource-constrained or busy departments. External validation is necessary to confirm its generalizability.