An Uncertainty-Aware Sequential Approach for Predicting Response to Neoadjuvant Therapy in Breast Cancer

Neoadjuvant therapy (NAT) is considered the gold standard preoperative treatment for reducing tumor charge in breast cancer. However, the tumor’s pathological response highly depends on patient conditions and clinical factors. There is a dire need to develop modeling tools to predict a patient response to NAT and thus improve personalized medical care plans. Recent studies have shown promising results of machine learning (ML) methodologies in breast cancer prognosis through the combination of several modalities, including imaging and molecular features derived from biopsy analyses. We here present a ML model to predict response to NAT through two sequential prediction stages. First, a pre-treatment dynamic contrast-enhanced magnetic resonance imaging model is trained, followed by a second model with molecular biomarkers-enriched data. We propose the integration of the Conformal Prediction (CP) framework in the first non-invasive model to identify patients whose predicted responses show large uncertainty and refer them to the second model that includes data from invasive tests. The major advantage of this procedure is in the reduction of unnecessary biopsies. Different alternatives for the standard ML algorithms and the CP functions are explored on a publicly available clinical dataset. Results clearly show the potential of our uncertainty-aware clinical predictive tool in such real scenarios.