Abstract PD2-07: Prediction of CDK inhibitor efficacy in ER+/HER2- breast cancer using machine learning algorithms

• Published in: Cancer research (Chicago, Ill.) Ill.), 2020-02, Vol.80 (4_Supplement), p.PD2-07-PD2-07
• Main Authors: Mason, Jeremy, Gong, Yutao, Amiri-Kordestani, Laleh, Wedam, Suparna, Gao, Jennifer J, Singh, Harpreet, Pazdur, Richard, Kuhn, Peter, Blumenthal, Gideon, Beaver, Julia A
• Format: Article
• Language: English
• ISSN: 0008-5472; 1538-7445
• DOI: 10.1158/1538-7445.SABCS19-PD2-07

Background: Three cyclin dependent kinase 4/6 inhibitors (CDKIs) are approved by the Food and Drug Administration (FDA) for the treatment of patients with hormone receptor (HR)-positive, human epidermal growth factor receptor 2 (HER2)-negative advanced or metastatic breast cancer in combination with hormonal therapy. However, the choice of when to initiate CDKI therapy (in the first- or second-line setting, or with aromatase inhibitor or fulvestrant) can be clinically challenging as studies have not directly compared these two approaches, nor have studies determined the benefit of continuation of CDKI therapy with changes in hormonal therapy backbone. Methods: Patient-level data from eight randomized controlled pivotal trials submitted to the FDA for new or supplemental marketing applications were pooled. We performed exploratory analyses of the predictive power of the baseline patient and disease characteristic data to forecast outcomes based on a specific treatment pathway. Variables included were: age, race, ethnicity, country and continent of origin, height, weight, body mass index (BMI), menopause status, estrogen receptor (ER) status, progesterone receptor (PR) status, ECOG performance score, histological type, histological grade, and initial disease stage. Additionally, baseline levels at the time of enrollment of alanine aminotransferase (ALT), albumin, alkaline phosphatase, aspartate aminotransferase (AST), bilirubin, blood urea nitrogen (BUN), calcium, creatinine, hemoglobin, magnesium, potassium, sodium, lymphocytes, neutrophils, platelets, and white blood cells. Statistical and machine learning algorithms were used to build predictive models based on the clinical trial results. Results: The pooled studies included 4580 breast cancer patients. Baseline random survival forest models were constructed for patients that received a CDKI plus hormonal therapy and for patients that received hormonal therapy alone. Preliminary findings for these models produced prediction accuracies of 69.2% for the CDKI model and 70.6% for the hormonal therapy alone model. Median predicted survival times for patients were calculated in both treatment scenarios and used to identify those that were more likely to benefit from initial use of CDKIs and those less likely to benefit. Through further analysis of the patients in each group, we identified patient characteristics that are prognostic of survival. Conclusions: This exploratory analysis utilized clinical trial data