S274
Clinical - Breast
ESTRO 2026
predictive models are available. Using synthetic data generated from a real-world clinical database, we conducted a study to predict the occurrence of breast pain using several Bayesian networks. Material/Methods: A synthetic database of 6,000 patients with 70 variables was generated using a generative Bayesian network (GBN) modeling the data of patients followed at ICANS and collected via data farming in MOSAIQ© (UNITRAD dataset) (1-2). The objective was to predict maximal pain at three time points during RT (T1, T2, T3) and post-RT (Tpost). Pain scores, rated from 0 to 10, were grouped into three categories (0, 1-5, >5), which were highly imbalanced: 87%, 11%, and 1% at T1; 65%, 30%, and 4% at T2; 63%, 29%, and 7% at T3; and 73%, 22%, and 5% at Tpost.Three additional Bayesian networks were generated to identify variables associated with pain occurrence: a data- driven network (DBN), an expert-driven network (EBN), and a combined network based on both data and expert knowledge (DBEN). Subsequently, 1,000 synthetic patients were generated using the GBN, and predictions were made with the three models. Their predictive accuracies were compared. Results: The results are summarised in the table below.Time pointPain 0Pain 1-5Pain > 5OverallT1DBN / DBEN (99%)EBN (18%)DBN / DBEN (23%)DBN / DBEN (89%)T2DBN (96%)DBEN (40%)DBEN (33%)DBEN (71%)T3DBN (96%)DBN (52%)DBEN (26%)DBEN (79%)TpostDBN / DBEN (92%)EBN (47%)DBN / DBEN (6%)DBN (78%)The DBN performed best for predicting zero pain, regardless of treatment time, corresponding to the category with the largest available dataset. When data volume was lower, combining data and expert knowledge (DBEN) often yielded better predictions, although overall performance remained moderate. Conclusion: Bayesian networks can predict the occurrence of breast pain during breast cancer treatment. Combining data-driven and expert-driven approaches improves predictive performance. References: 1. Piot M, Bertrand F, Guihard S, Clavier JB, Maumy M. Bayesian Network structure learning algorithm for highly missing and non imputable data: Application to breast cancer radiotherapy data. Artif Intell Med. 2024 Jan;147:102743. doi: 10.1016/j.artmed.2023.102743. Epub 2023 Nov 30.2. Guihard S, Piot M, Issoufaly I, Giraud P, Bruand M, Faivre JC, Eugène R, Liem X, Pasquier D, Lamrani-Ghaouti A, Ghannam Y, Ruffier A, Guilbert P, Larnaudie A, Thariat J, Rivera S, Clavier JB. Données de vie réelle en radiothérapie : le data farming du groupe Unitrad [Real world data in radiotherapy: A data farming project by Unitrad]. Cancer Radiother. 2023 Sep;27(6-7):455-459.
response was associated with known risk factors for IMN involvement: 48.5% of patients with involved AXN showed a response in IMN, compared with only 21.8% of those without AXN metastases (p < .001) (figure B). 75.8% of patients with responding IMN had predominant internal mammary perforator vessel (IMPV) contact, while 6.5% had no contact (p < .001) (figure C). In patients with medial tumor location, an IMN response was observed in 52.3% compared to only 34.2% in patients with central or lateral tumors (p .037) (figure D).
Conclusion: We found an IMN response to NACT in a relevant proportion of patients with histopathological and clinical PT and AXN response without primarily suspected IMN metastases, assuming a higher incidence of IMN metastases than previously expected. This finding could improve the detection of IMN metastases using MRI and therefore has a high impact on further systemic therapy, radiotherapy and follow-up imaging strategies. Keywords: internal mammary lymph nodes, nodal response, NACT Comparison of three bayesian networks for predicting breast pain during breast cancer treatment using synthetic data Bastien Tarot 1 , Raphael Monnier 1 , Thuaut Sébastien 1 , Antoine Jean 2 , Clavier Jean-Baptiste 3 , Melanie Piot 1 , Sébastien GUIHARD 3 1 Laboratoire Learning, Data & Robotics, ESIEA, Ivry-sur- Seine, France. 2 Laboratoire, ESIEA, Ivry-sur-Seine, France. 3 radiotherapy, Centre Paul Strauss, Strasbourg, France Purpose/Objective: Breast pain is a common sequela of combined surgery and radiotherapy (RT) for breast cancer, yet few Digital Poster 2393
Made with FlippingBook - Share PDF online