1074 Background: CDK4/6 inhibitors (CDK4/6i) in combination with endocrine therapy (ET) are the current standard of care for pts with newly diagnosed HR+/HER2- metastatic breast cancer (MBC). Although multiple pivotal trials showed a robust benefit of CDK4/6i over ET alone in all the subgroups, there is a subset of pts with a suboptimal response that is difficult to characterize. To predict the benefit on first line CDK4/6i in pts with HR+/HER2- MBC, we created a machine-learning (ML) model based on clinico-pathological features (CPFs) at baseline. Methods: We identified 854 consecutive pts with HR+/HER2- MBC receiving CDK4/6i as 1st line therapy between 12/2013 and 10/2023 from our MSK translational database. OncoCast-MPM ML framework was used to stratify pts based on progression-free survival (PFS) on 1st line CDK4/6i. CPFs of interest were selected as input for the models. Pts with missing data in any of the key inputs were excluded and used as hold out set. We tested 4 algorithms (GBM, ENET, LASSO and SVM), and selected the only model that returned predictions when tested on the hold out dataset containing missing data. The Kaplan-Meier estimator and the log-rank test were used to assess differences in PFS, and Cox regression was used to measure hazard ratios (HR) between groups. Results: Overall, 651 out of 854 pts and 15 CPFs were used in models' creation. The median PFS of the cohort was 15.95 months (95%CI: 14.04-17.95). Of the 4 tested architectures, GBM was the selected architecture based on its ability to be tested on the holdout cohort containing missing data. GBM identified 3 risk groups: good-response risk group (N=279), with a median PFS of 23.8 months (95%CI: 19.6-29.9), intermediate-response risk group (N=231), with a median PFS of 14.7 (I95%CI: 12.2-16.9), and poor-response risk group (N=141), with a median PFS 7.4 (95%CI: 5.3,9.1). The HR between good-response and poor-response RGs was 2.55 (95%CI: 2.0,3.2) with a of p = 2.54e-14. The first 5 CPFs by importance, sorted by Garson algorithm score are: disease-free interval, liver involvement, age at metastatic disease, adjuvant treatment-free interval, and progesterone receptor status. No statistically significant differences were observed between the 3 risk groups identified in the train-validation and tested in the holdout cohort, in terms of median, 1yr, 3yr survival times, as well as in HR. Conclusions: ML-based models have the potential to identify subsets of pts likely to poorly respond to first line CDK4/6i leveraging clinico-pathological features. Risk stratification can help clinicians to select pts who may benefit from short-term interval imaging and potential cfDNA monitoring, as well as to identify the right population to design clinical trials with escalating treatment approaches.