Abstract PS3-04-29: Machine Learning predicts early tumor progression in Hormone Receptor-positive, Human Epidermal growth factor Receptor 2-negative (HR+/HER2-) advanced breast cancer (aBC) patients treated with first-line endocrine therapy (ET) plus Cyclin Dependent Kinase 4/6 inhibitors (CDK4/6i)

Abstract Introduction: CDK4/6i combined with ET reshaped the first-line treatment paradigm for HR+/HER2-advanced Breast Cancer (aBC). However, early progressors, as defined as patients (pts) undergoing tumor progression within 6 months from treatment initiation, remain a major clinical challenge. Indeed, identifying early progressors may lead to the use of alternative, potentially effective, treatments. Machine learning (ML) has demonstrated significant capability in integrating data extracted from electronic health records to successfully predict clinical outcomes in oncology. Methods: PALMARES-2 (NCT06805812) is an observational, retrospective-prospective, multicenter Italian study that is investigating the effectiveness of first-line ET plus CDK4/6i in HR+/HER2- aBC pts. We used the log-rank test to evaluate the impact of early progression on real-world (rw) time to chemotherapy (rwTTC) and overall survival (rwOS). Four ML architectures, namely linear regression with elastic net penalisation (glmnet), random forest (RF), XGBoost (XGB) and neural network (NN), were trained to predict the risk of early progression (rw Progression-Free Survival 6 months). As model inputs, we used 73 variables, including clinical, pathological and laboratory ones. Missing data were imputed through the missForest method, while undersampling was applied to address class imbalance. The best hyperparameters were selected based on 10-fold cross-validation (CV) results. After model development in the training set, model performance was confirmed on a test set (20% of the overall population) and on an independent cohort of pts from Istituto Nazionale dei Tumori, Milan. Results: Among 4104 evaluable pts, 521 (12.7%) experienced early tumor progression. Early progressors had significantly worse rwTTC (median 5.3 vs 43.4 months, P 0.001) and rwOS (median 21.4 vs 73.0 months, P 0.001) when compared to pts who did not experience early progression. Early progressors were more likely to have lower tumor estrogen receptor and progesterone receptor expression, higher Ki67 expression, endocrine-resistant disease at diagnosis, worse ECOG PS and liver metastases (P 0.001). All models demonstrated consistent accuracy, with acceptable generalizability in the independent test cohort (Table). Conclusions: Early tumor progression during first-line ET+CDK4/6i is associated with remarkably worse outcomes in HR+/HER2- aBC pts. We developed the first ML model to predict the risk of early progression in this setting. If prospectively validated, this tool may assist clinicians in selecting pts who could benefit from alternative first-line treatment approaches, thus potentially improving long-term outcomes, such as TTC and OS. Citation Format: L. Provenzano, R. Caputo, M. Dieci, P. Vigneri, M. Giuliano, G. Curigliano, A. Toss, A. Botticelli, R. Pedersini, S. Cinieri, M. Lambertini, G. Rizzo, B. Tagliaferri, M. Sirico, M. Giordano, L. Gerratana, I. Meattini, M. Piras, A. Fabi, A. Zambelli, F. Pantano, A. Gennari, N. La Verde, D. Sartori, O. Garrone, D. Generali, F. Ligorio, G. Pruneri, F. De Braud, C. Vernieri, PALMARES-2 study group. Machine Learning predicts early tumor progression in Hormone Receptor-positive, Human Epidermal growth factor Receptor 2-negative (HR+/HER2-) advanced breast cancer (aBC) patients treated with first-line endocrine therapy (ET) plus Cyclin Dependent Kinase 4/6 inhibitors (CDK4/6i) abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-29.