A large validation study of AI-powered PD-L1 analyzer compared to pathologists’ assessment of PD-L1 expression in lung cancer.

Abstract

8535 Background: Programmed death ligand 1 (PD-L1) expression is a useful biomarker for immune checkpoint inhibitors in advanced lung cancer. However, pathologists' manual evaluation of PD-L1 expression has practical limitations, including observer bias. The development of artificial intelligence (AI)-powered PD-L1 evaluation models has recently progressed. We evaluated the concordance rate of PD-L1 expression as assessed by pathologists and an AI-powered PD-L1 analyzer in lung cancer patients. Methods: This multicenter prospective observational study included patients with stage II to IV or recurrent lung cancer (LC-SCRUM-IBIS). PD-L1 Tumor Proportion Score (TPS) was assessed in lung biopsy specimens, by using a 22C-3 Immunohistochemistry (IHC) assay and scanned at ×40 magnification using a whole-slide images scanner (Hamamatsu Photonics). The results of PD-L1 TPS were evaluated independently by three lung pathologists trained in IHC assessment of PD-L1 expression. We examined an AI-powered PD-L1 TPS analyzer, namely Lunit SCOPE PD-L1. Results: Between February 2017 and May 2018, 1,017 lung cancer patients were enrolled. Of these, adequate tumor samples allow for PD-L1 IHC assays; 847 non-small cell lung cancer (NSCLC) patients and 102 small cell lung cancer (SCLC) patients. Lunit SCOPE PD-L1 training included annotations of a total of 64,245,935 tumor cells. Regarding patients characteristics, the median age was 66, 31% were female, 75% were ever smokers, and the distribution of stages was as follows: stage II, III, IV, or recurrence in 37, 97, 632, and 183 patients, respectively. The histological subtypes included in NSCLC, non-squamous (666 patients), squamous (181 patients). Additionally, 85% were diagnosed by biopsy specimens. In comparing PD-L1 TPS assessed by AI and pathologists, the overall concordance rate was 70% with a kappa value of 0.56 (95% confidence interval [CI], 0.49–0.61). The concordance rate according to PD-L1 TPS ≥50%, 1-49%, and <1% was 84%, 94%, and 44%, respectively. Of the 416 patients whom pathologists determined to be TPS <1%, 231 (55%) were TPS 1-49%, and only one patient was determined to be TPS ≥50% by AI analyzer. In SCLC patients’ analysis, 84% of patients were determined to be PD-L1 <1% by pathologists, with a low concordance rate of 61% (k = 0.29) between pathologists and AI analyzer. Conclusions: PD-L1 TPS demonstrated a high concordance between pathologists and AI analyzers in lung cancer patients with TPS ≥50% and 1-49%. However, the concordance rate of TPS <1% was low regardless of histology. We will confirm if the AI analyzer accurately predicts treatment outcome, especially in TPS <1%. Clinical trial information: UMIN000026425 .