Machine learning model for predicting DIBH non-eligibility in left-sided breast cancer radiotherapy: Development,...

Area of Science:

• Oncology
• Medical Physics
• Machine Learning

Background:

• Multi-day assessments for Deep Inspiration Breath Hold (DIBH) irradiation in left-sided breast cancer are effective but challenging in resource-limited settings.
• Developing a machine learning (ML) model using initial assessment data can predict DIBH ineligibility, optimizing patient selection.

Purpose of the Study:

• To develop and validate a machine learning model to predict DIBH ineligibility using only first-day DIBH assessment data.
• To assess the model's performance and clinical utility in a prospective cohort.

Main Methods:

• A prospective cohort study involving 202 patients for model development and validation on 47 patients.
• Evaluation of nine ML algorithms, with selection based on decision curve analysis.
• Clinical impact assessment on 64 patients comparing model predictions with clinical decisions.

Main Results:

• An uncalibrated gradient-boosting ensemble model achieved an AUC of 0.803, with average breath-hold duration and lower breath-hold amplitude as key predictors.
• The model demonstrated potential to reduce additional DIBH assessments by up to 20% without misclassifying eligible patients.
• Decision curve analysis confirmed the model's net benefit.

Conclusions:

• The developed ML model accurately predicts DIBH ineligibility using single-day assessment data.
• This model can serve as a decision aid for patient selection in DIBH radiotherapy, particularly in resource-constrained environments.
• External validation is recommended to ensure generalizability.