Abstract P2-04-22: Predictive Modeling of Cancer Treatment-Related Cardiac Events in Breast Cancer Patients: Utilizing Dosiomic and Radiomic Features with Machine Learning

Le résumé fourni par la source

Abstract Purpose: Conventional clinical study consists of clinical, radiological and dosimetric data and follow-up for many years. In the last decade, personalized medicine has become the main subject of scientific research with genome maps and biomarker discoveries. This study focused on predicting treatment-associated cardiac side effects in breast cancer patients, before they occur, using cardiac related biomarkers (hs-TopT, radiomics, dosiomics). This is the first study of heart-segmented dosiomics in breast cancer patients.Methods: In this retrospective study, clinical and dosimetric data, along with radiomic and dosiomic features extracted from medical imaging, were analyzed for 42 women diagnosed with localized breast cancer (patients whose blood troponin levels were measured 2-3 weeks following the completion of radiotherapy). Patients with pre-existing cardiac comorbidities or previous troponin-T elevation were excluded from the study. The cutoff value of heart specific troponin T levels at 2-3 weeks following radiotherapy was determined 14 ng/L, and the patients were classified into two groups. Those above this limit value were considered to be associated with a cancer treatment-related cardiac event. For the extraction of radiomic and dosiomic features, an open-source Python package PyRadiomics was utilized. In this study, we employed the TPOT (Tree-based Pipeline Optimization Tool) to select and optimize machine learning models and hyperparameters. This process led us to identify the 'Gradient Boosted Classification Algorithm' as the algorithm that would yield the best performance. Gradient boosted recursive feature elimination, a hybrid method, was used to select important features for prediction. Gradient boosted classification algorithm, an embedded method, was used to create and test the model. 5-layer cross-validation and nonparametric permutation testing were performed to evaluate model generalizability and randomness. The area under the curve (AUC) method was used to evaluate model performance. Results: In total, 111 dosimic and 119 radiomic features were extracted for each patient. A total of 6 different models were created with different feature groups (clinical, dosimetric, radiomics, dosiomics). The highest prediction model was obtained with clinical + dosiomics + radiomics parameters (test set-AUC = 0.96). This value was much lower for the clinical + dosimetric model (test set-AUC = 0.67). These two models were found to be non- coincidence based on confirmation from nonparametric permutation tests (p<0,05). Other models were considered random based on permutation testing (p>0,05). Cross-validation analyses for the clinical+ dosiomic+ radiomics model showed that the generalizable performance of the model was relatively lower but still fair-to-good (mean AUC value 80.33 ± 21%). Discussion: This study can demonstrates that imaging biomarkers, specifically radiomics and dosimics, exhibit superior predictive capability for treatment-related cardiac events in breast cancer patients compared to traditional clinical and dosimetric parameters (96% vs. 67%). Incorporating radiomics and dosiomics parameters into future clinical practice can serve as critical indicators. Fundamentally altering treatment approaches and follow-up strategies by achieving nearly perfect predictive success (96%) in anticipating cardiac side effects before their occurrence. This may suggests that high statistical imaging biomarkers should be integrated into personalized clinical practice. Further investigation is needed. Citation Format: Sefika Dincer, Sefika Dincer, Muge Akmansu. Predictive Modeling of Cancer Treatment-Related Cardiac Events in Breast Cancer Patients: Utilizing Dosiomic and Radiomic Features with Machine Learning [abstract]. In: Proceedings of the San Antonio Breast Cancer Symposium 2024; 2024 Dec 10-13; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(12 Suppl):Abstract nr P2-04-22.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.