Supervised machine learning model to predict oncotype DX risk category in patients over age 50.

Abstract

Routine use of the oncotype DX recurrence score (RS) in patients with early-stage, estrogen receptor-positive, HER2-negative (ER+/HER2-) breast cancer is limited internationally by cost and availability. We created a supervised machine learning model using clinicopathologic variables to predict RS risk category in patients aged over 50years. From January 2012 to December 2018, we identified patients aged over 50years with T1-2, ER+/HER2-, node-negative tumors. Clinicopathologic data and RS results were randomly split into training and validation cohorts. A random forest model with 500 trees was developed on the training cohort, using age, pathologic tumor size, histology, progesterone receptor (PR) expression, lymphovascular invasion (LVI), and grade as predictors. We predicted risk category (low: RS ≤ 25, high: RS > 25) using the validation cohort. Of the 3880 tumors identified, 1293 tumors comprised the validation cohort in patients of median (IQR) age 62years (56-68) with median (IQR) tumor size 1.2cm (0.8-1.7). Most tumors were invasive ductal (80.3%) of low-intermediate grade (80.5%) without LVI (80.9%). PR expression was ≤ 20% in 27.3% of tumors. Specificity for identifying RS ≤ 25 was 96.3% (95% CI 95.0-97.4) and the negative predictive value was 92.9% (95% CI 91.2-94.4). Sensitivity and positive predictive value for predicting RS > 25 was lower (48.3 and 65.1%, respectively). Our model was highly specific for identifying eligible patients aged over 50years for whom chemotherapy can be omitted. Following external validation, it may be used to triage patients for RS testing, if predicted to be high risk, in resource-limited settings.