Modelling Toxicity Risk and Uncertainty in Breast Radiotherapy with Bayesian Networks
P Ciunkiewicz1*, M Roumeliotis2, K Stenhouse2, P McGeachy2, P Grendarova3, S Quirk2, S Yanushkevich1, (1) University of Calgary, Calgary, Alberta, CA, (2) Tom Baker Cancer Centre, Calgary, Alberta, CA, (3) Alberta Health Services, Calgary, Alberta, CA
Presentations
WE-IePD-TRACK 5-1 (Wednesday, 7/28/2021) 12:30 PM - 1:00 PM [Eastern Time (GMT-4)]
Purpose: To predict post-treatment normal tissue toxicity in patients undergoing accelerated partial breast irradiation (APBI) therapy and to quantitatively identify which diagnostic, anatomical, and dosimetric features are contributing to these outcomes.
Methods: A retrospective study of APBI treatments was performed using 32 diagnostic, anatomical, and dosimetric features pertaining to various phases of the patient’s treatment journey. These features were used to inform and construct a Bayesian network (BN) based on both statistical analysis of feature distributions and relative clinical importance. The target feature for prediction was defined as a measurable worsening of telangiectasia, or subcutaneous tissue induration, or fibrosis when compared against the observed baseline. Parameter learning for the network was performed using data from the 299 patients included in the ACCEL trial. Predictive performance of the BN was measured and compared against a variety of conventional machine learning (ML) approaches.
Results: Cross validated performance of the BN for predicting toxicity was consistently higher across all binary classification metrics when compared against ML techniques including decision trees, naive Bayes, ensemble trees, and support vector machines. The measured BN receiver operating characteristic area under the curve (ROC-AUC) was 0.938±0.028 against the best ML result of 0.910±0.032 using 5-fold cross validation across 100 trials. The volume of the clinical target volume and baseline fibrosis measurements were found to have the highest mutual dependence with normal tissue toxicity in the network, representing the strongest contribution to patient outcomes.
Conclusion: The BN outperformed conventional ML techniques in predicting tissue toxicity outcomes and provided deeper insight into which features are contributing to these outcomes. With additional data and further development of the network topology, stronger predictive performance is possible. BNs are a powerful tool for many predictive and diagnostic clinical tasks, particularly for applications with limited data availability or completeness.
Funding Support, Disclosures, and Conflict of Interest: The authors acknowledge funding and regulatory support from the Institutional Clinical Trials Unit and an Investigator Initiated Trial grant supported by the Alberta Cancer Foundation that facilitated the conduct of the ACCEL trial.
ePosters
Keywords
Bayesian Statistics, Breast, Risk
Taxonomy
TH- Dataset Analysis/Biomathematics: Machine learning techniques
Contact Email