Machine learning model PD-L1 22C3 scoring of a multi-scanner, real-world reference laboratory NSCLC dataset generates scores comparable with manual pathologist scoring

PD-L1-targeting immunotherapy is a first line treatment for patients with metastatic or stage IV non-small cell lung cancer (NSCLC). Patient eligibility for these potentially life-saving treatments currently requires pathologist assessment of biopsy samples, a process that may miss some eligible patients2. PathAI has developed machine learning (ML)-based models that can reproducibly identify and quantify PD-L1 cells within tumor tissue samples. The PathAI Legacy PD-L1 NSCLC model, trained to quantify diverse PD-L1-stained tissue samples on slides digitized using two scanners, generated scores that were highly correlated with the continuous manual scores (Ventana: Pearson 0.87 p<0.05, Philips: Pearson 0.86 p<0.05). Here, we report on an optimized, generalizable ML-based NSCLC PD-L1 model and test its robustness through application to a real-world NSCLC clinical dataset also scanned using 2 scanners. ML models that perform well under such circumstances could be further developed for use in a clinical setting.