Abstract P2-02-29: AI-based Clinical Decision Support System (CDSS) for predicting response to primary systemic therapy in early breast cancer: multimodal modeling and pathology contribution

Le résumé fourni par la source

Abstract Background: Accurate prediction of tumor response to primary systemic therapy (PST) in early breast cancer is crucial for optimizing treatment decisions. The Breacs consortium developed an AI-based CDSS that integrates digital pathology data and pre-operative clinical variables to predict pathological complete response (pCR) in early breast cancer patients. Deciphering the factors on which the model relies to make these predictions may provide valuable new insights. Material and methods: A multimodal predictive model was developed using data from the BreaCS consortium, encompassing a cohort of 550 patients who underwent PST. The dataset included digital pathology data from biopsies and more than 100 pre-operative clinical variables from the standardised EUSOMA database. The data was split into a training and a test sets (80-20 stratified split on patient level). Model evaluation metrics included the area under the operator curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Quantitative cell segmentation analysis was conducted on the test set to identify histopathological features influencing the model’s predictions. Cell type fractions were compared across 4 different groups (true positives, false positives, false negatives, true negatives) within the confusion matrix. Results: The unimodal model based on pathology data achieved an AUC of 0.70 (training) and 0.81 (test). The model using only EUSOMA clinical data scored an AUC of 0.88 (training) and 0.82 (test). The multimodal model, incorporating both pathology and clinical data, demonstrated superior performance with an AUC of 0.91 (training) and 0.84 (test). The multimodal model demonstrated a significant improvement in sensitivity, NPV and PPV compared to the unimodal models, with specificity remaining relatively unchanged.The test set confusion matrix showed 72 true negatives (TN), 11 false negatives (FN), 14 true positives (TP), and 9 false positives (FP). The cell segmentation in the group of the TP (predicted and obtained complete remission) shows an excess in the fraction of tumor cells compared to the TN. The TN (non-responders) on the other hand have an excess in the fraction of connective tissue compared to the TP. Conclusions: The AI-based CDSS demonstrates promising predictive capabilities for response to PST in early breast cancer, with the multimodal approach showing enhanced predictive accuracy. This study underscores the importance of integrating multimodal data in predictive models. The observed differences in cellular composition between TP and TN reveals distinct cellular compositions and may suggest potential refinements for improving model accuracy. Citation Format: Barbara Bussels, Sandra Steyaert, Francesca Dedeurwaerdere, Isabelle Kindts, Philip Poortmans, Adelheid Roelstraete, Frederik Deman, Caroline De Beukelaar, Mona Bové, Sander Goossens, Peter De Jaeger. AI-based Clinical Decision Support System (CDSS) for predicting response to primary systemic therapy in early breast cancer: multimodal modeling and pathology contribution [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 P2-02-29.

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