Prediction of real-world progression free survival (rwPFS) using a multimodal machine learning (ML) model for patients with HR+ HER2- metastatic breast cancer (mBC) undergoing first line (1L) treatment with cyclin-dependent kinase 4 and 6 inhibitors

e13088 Background: CDK4/6i combined with ET is 1L standard of care treatment for HR+/HER2- for mBC patients, however duration of response varies with some patients experiencing disease progression within 12-months. Limited predictive factors related to ET+CDK4/6i treatment response exist. This pilot aims to assess the feasibility of developing a ML derived risk score for CDK4/6i+ET across 5 different data modalities to inform escalation and de-escalation therapeutic strategies. Methods: A pilot study on a retrospective cohort of 131 patients with mBC from the Memorial Sloan Kettering Cancer Center (MSK) treated with CDK4/6i and ET in the 1L setting was carried out to develop a ML based algorithm for predicting rwPFS at an individual patient level. Baseline multimodal data (including clinical, demographic, histopathological, genomic and radiomic) alongside clinical outcomes were collected. For radiomic analysis, up to 5 lesions were segmented in 3D on baseline PET/CT scans per patient using the SOPHiA DDMTM Radiomics platform. Radiomic features, extracted per IBSI standards, were combined with other data modalities. Baseline tumor genomic analysis was performed using the MSK-IMPACT assay, including genes with an alteration frequency > 5% in the cohort. A filter-based variable selection method was applied prior to training multiple ML algorithms, with the optimization criteria being the Brier score for rwPFS prediction. Due to the limited cohort size, a nested cross-validation approach was employed to ensure robust performance estimation. Results: A Random Forest survival model, combined with k-Nearest Neighbor (k-NN) imputation techniques for handling missing data, achieved an AUC of 0.787 (95% CI, 0.716–0.859) at 12 months for predicting rwPFS. Presence of liver metastases, SUVmax, total tumor volume, CEA, CA 15-3, TP53 mutation and PET uptake heterogeneity were among the top weighted features. These clinically plausible features, combined into a multimodal algorithm, allowed to stratify patients into two risk groups based on their predicted rwPFS with a median rwPFS of 45.2 (95% CI, 27.1-n/a) in low-risk vs 11.3m (95% CI, 9.1-16.1m) in high-risk group (hazard ratio: 4.10, 95% CI, 2.64–6.37). Conclusions: This pilot study demonstrates the feasibility of a ML, multimodal approach to predict rwPFS for patients with HR+/HER2- mBC treated with 1L CDK4/6i + ET. This approach could be implemented in the clinical setting to guide treatment choice and to inform clinical trial design by identifying high-risk individuals. Further training and validation of the model is planned in a larger, multicentric cohort.