Radiomics risk score (RRS) on CT to predict survival and response to CDK 4/6 inhibitors in hormone receptor (HR) positive metastatic breast cancer (MBC).

Abstract

e13041 Background: The addition of CDK 4/6 inhibitors (CDK) to endocrine therapy (ET) is the standard of care for HR+ MBC. Currently, no robust predictive markers for CDK inhibitors exist. We hypothesized that baseline imaging features on pre-treatment CT scans using machine learning could correctly identify responders to CDK. Methods: From a retrospective pt registry, 46 pts with HR+ MBC who received palbociclib (palbo) plus ET and had liver metastases were identified, and divided at random into equally sized training and testing cohorts. Patients with objective response/stable disease per RECIST v1.1 were defined as ‘responders’ (65%), and those with progressive disease within 6 months were ‘non-responders’ (35%). The median age at diagnosis was 62. All pts were ER+ and 4% were HER2+. 65% pts received palbo as 1st or 2nd line tx. 503 radiomic texture features measuring subtle differences in lesion heterogeneity on a pixel level were extracted on pre-treatment CT within the first lesion measured for RECIST assessment. An elastic net Cox proportional hazards model was constructed within the training set to derive a Radiomics Risk Score (RRS), which was evaluated for association with prognosis and response within the testing set. A risk score threshold was chosen in the training set for stratifying patients into high and low risk groups. Results: The RRS consisted of 4 radiomic features indicating intratumoral heterogeneity to be associated with poor outcome. In the independent testing set (n = 23), RRS was continuously associated with survival (p = .0027) on multivariate analysis, along with age and RECIST response (Table). Categorical risk groups were similarly associated with survival in the testing set (p = .00491, HR = 4.03, CI = 0.6772). Median survival time in low and high risk groups were 2.33 and 0.673 years, respectively. When compared against RECIST response in the testing set, RRS was found to also be predictive of response with an ROC curve of 0.675. Conclusions: Radiomics analysis was able to predict response and survival in HR+ MBC pts prior to initiation of treatment with CDK. [Table: see text]