Photo Credit: Mohammed Haneefa Nizamudeen

The following is a summary of “Machine learning model for predicting DIBH non-eligibility in left-sided breast cancer radiotherapy: Development, validation and clinical impact analysis,” published in the April 2025 issue of Radiotherapy and Oncology by Chufal et al. 

Deep Inspiration Breath Hold (DIBH) is an essential technique for minimizing radiation-induced cardiac exposure in patients undergoing irradiation for left-sided breast cancer. Multi-day assessments have proven effective in accurately identifying patients ineligible for DIBH while optimizing on-couch treatment time for eligible individuals. However, implementing these multi-day assessments can be challenging in resource-limited clinical settings, necessitating the development of predictive models for early patient stratification. 

This study aimed to develop and validate a machine learning model that utilizes only first-day DIBH assessment data to predict patient ineligibility, thereby reducing the need for extended assessments. A prospective cohort of 202 patients undergoing DIBH evaluation between January and December 2023 was used to train and develop the model. Key patient-related and assessment-related variables, including upper, lower, and average breath-hold amplitude, breath-hold duration, and breath-hold consistency, were analyzed. Nine ML algorithms employing three different modeling strategies were assessed for predictive performance, with decision curve analysis guiding model selection. 

The gradient-boosting ensemble model exhibited the highest predictive performance, achieving an area under the curve of 0.803 (95% CI: 0.686–0.941) and a recall of 0.526, with net benefit demonstrated in decision curve analysis. The most influential predictive factors were average breath-hold duration and lower breath-hold amplitude levels. Temporal validation on a separate prospective dataset of 47 patients (January–March 2024) confirmed the model’s robustness. Additionally, a clinical impact study conducted on another cohort of 64 patients (April–August 2024) demonstrated that the model could reduce the requirement for additional DIBH assessments by up to 20% without misclassifying eligible patients. 

These findings suggest that the ML model can serve as a valuable decision-support tool in busy or resource-constrained oncology departments, facilitating efficient patient selection for DIBH while maintaining accuracy. External validation is warranted to confirm the model’s generalizability across diverse clinical settings.

Source: thegreenjournal.com/article/S0167-8140(25)00059-3/abstract