67P Artificial Intelligence to predict early progressors among aBC patients treated with first-line CDK4/6i: A sub-analysis from PALMARES-2 study

CDK4/6 inhibitors (CDK4/6i) in combination with endocrine therapy (ET) are nowadays the standard of care, 1st line treatment for Hormone Receptor positive, Human Epidermal growth factor Receptor 2-negative (HR+/HER2-) advanced Breast Cancer (aBC). However, some patients (pts) experience low PFS during ET+CDK4/6i and could be candidate to treatment intensification. Artificial Intelligence (AI) methods, in particular Machine Learning (ML), are effective in integrating data to generate predictive models. We used ML to build a predictive model based on Real-world data (RWD) of HR+/HER2- aBC pts enrolled in the multicenter Italian study PALMARES-2. All pts received ET+CDK4/6i as 1st therapy. PFS status at 18 months was used as the clinical outcome for classification model. The entire dataset was split in training and test cohorts, and undersampling method was used to balance outcomes. Logistic regression (LR), random forest (RF), XGBoost and neural networks with model averaging (avNNet) were used as classifiers. Permutation-based variable importance was used for AI explainability (XAI). After excluding patients with less than 18 months of follow up and adjusting for outcomes unbalance, a total of 1000 pts from 18 Italian centers were included in the analysis. To fit the model, 52 clinical features were selected based on clinical relevance. All the 4 models consistently demonstrated similar prediction ability, with AUC ranging from 0.72 to 0.74 in training set and from 0.68 to 0.70 in test set. Performance values are reported in the table. XAI revealed endocrine resistance, liver metastases and performance status as the most informative features. ML demonstrated promising ability in predicting early progressors among HR+/HER2- aBC patients treated with 1st line ET+CDK4/6i using RWD. Integrating different sources of data, e.g. genomic, radiomics and digital pathology, could further improve model accuracy.Table: 67PLRRFXGBoostavNNetTrainingTestTrainingTestTrainingTestTrainingTestAUC0.740.700.730.680.740.680.720.70Sensitivity0.690.690.690.730.660.670.690.73Specificity0.700.700.660.620.680.680.660.66 Open table in a new tab