254P Development of machine learning models for the prediction of early progression (EP) to first-line (1L) cyclin-dependent kinase 4/6 inhibitors (CDK4/6i) therapy in oestrogen receptor (ER)-positive/human epidermal growth factor receptor 2 (HER2)-negative metastatic breast cancer (MBC)

The combination of hormonal therapy (HT) plus CDK 4/6i represents the standard treatment for 1L ER-positive/HER2-negative MBC. Nevertheless, a small proportion of patients (pts) experiences EP within 6 months. Machine Learning (ML) techniques have been adopted to investigate predictors of prognosis and progression. Feature selection methods (recursive feature elimination, stepwise regression and Least Absolute Shrinkage and Selection Operator LASSO) were employed to identified baseline clinical/pathological variables potentially influencing EP (progression free survival PFS <6 months) in a retrospective ER-positive/HER2-negative MBC patients' series. Supervised ML models (Random Forest, Neural Network, Support Vector Machine and Gradient Boosting Machine GBM) were trained to predict EP. The performance was evaluated by analysing the area under the receiver operating curve (AUC), sensitivity and specificity. Models were evaluated using stratified 10-fold cross-validation. Progression events occurred in 129 out of 267 eligible pts (48.3%) with median PFS of 31.4 months (95% CI 26.6-45.8). 40 pts (14.9%) experienced EP. GBM prediction model using LASSO features selection showed the best performance: AUC 0.71 (95% CI 0.62-0.80), sensitivity 0.40 (95% CI 0.26-0.55) and specificity 0.85 (95% CI 0.80-0.89). Variables importance analysis hierarchically indicated premenopausal status, followed by liver metastasis, presence of brain metastasis, HT type, primary metastatic disease, peritoneal metastasis, lobular histotype, and ER-expression as the most significant predictors. We developed and cross-validated a model to predict EP<6 months to 1L CDK 4/6i. The GBM utilizing LASSO-selected features exhibited the highest performance. These results highlight the potential of these models to guide treatment decisions and personalized therapeutic approaches.