Optimize Patient Care with Machine Learning-Based Biomarkers

• Activity Number: 315 - Statistical Innovation in Biomarker-Guided Clinical Development
• Type: Invited
• Date/Time: Tuesday, August 9, 2022 : 2:00 PM to 3:50 PM
• Sponsor: General Methodology
• Author(s): Jie Ding and Shuyang Wu and Jianghan Qu and Marla Johnson and Lori Lofaro and William Bulman and Sangeeta Bhorade and Sean Walsh and Jonathan Wilde and Giulia Kennedy and Jing Huang*
• Companies: Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc. and Veracyte Inc.
• Keywords: machine learning; cancer diagnosis; RNA sequencing; lung cancer; genomic

Abstract

Accurate cancer risk assessment is critical in the care of pulmonary nodule to minimize procedures for benign disease and for timely diagnosis and treatment for lung cancer. We demonstrate a machine-learning based clinical-genomic classifier can accurately assess such risk. It was developed in a training set of >1100 patients. Genomic indices were developed to better capture the genomic impact on cancer from smoking. A layered structure was utilized to handle cohort diversity. Different decision boundaries were chosen to maximize sensitivity and specificity for low and high risk nodules, respectively, yielding low, intermediate, and high risk categories. The performance was evaluated in an independent validation set of 249 patients. It identifies > 40% (41.7% Specificity) of benign patients as low risk with 96.3% sensitivity, resulting in 97.1% NPV in a population with 25% cancer prevalence; it identifies ~60% (58.2% sensitivity) of malignant patients as high risk with 90.4% specificity, resulting in 67.0% PPV in a population with 25% cancer prevalence. The performance is independent of nodule size or cancer stage and superior to commonly used clinical risk models (VA, Mayo, Brock)