Precise determination of biomarker status is necessary for clinical trial enrollment and endpoint analyses, as well as for optimal treatment determination in real-world practice. However, variabilities may be introduced into this process due to the processing of clinical specimens by different laboratories and assessment by distinct pathologists. Machine learning tools have the potential to minimize inconsistencies, although their use is not presently widespread. To assess the applicability of machine learning to the quality control process for biomarker scoring in oncology, we developed and validated an automated machine learning model to be applied as a quality control tool for monitoring the assessment of human epidermal growth factor-2 (HER2). The model was trained using whole slide images from multiple sources to quantify HER2 expression and measure immunohistochemistry stain intensity, tumor area, and the presence of artifacts or ductal carcinoma in situ across breast cancer phenotypes. The quality control tool was deployed in a real-world cohort of HER2-stained breast cancer sample images collected from routine diagnostic practice to evaluate trends in HER2 testing quality indicators and between pathology laboratories. Automated image analysis for HER2 scoring is consistent and reliable using this algorithm. Deployment of the HER2 quality control tool across 3 clinical laboratories revealed interlaboratory variability in HER2 scoring and inconsistencies in data reporting. These results support the future incorporation of quality control algorithms for real-time monitoring of clinical laboratories contributing to clinical trials in oncology and in the real-world setting of HER2 immunohistochemistry testing in local clinical laboratories and hospitals.
Read full abstract