Clinical relevance of AI-based PD-L1 scoring in non-small cell lung cancer

Artificial intelligence (AI)-based algorithms are increasingly implemented in histopathological cancer diagnostics to enhance reproducibility and efficiency in biomarker assessment. Although AI-driven image analysis shows promise in standardizing immunohistochemical evaluation and reducing inter-observer variability, its clinical reliability as a substitute for expert assessment requires rigorous validation. Minor discrepancies in Programmed Death Ligand 1 (PD-L1) interpretation can alter patient stratification and treatment outcomes, especially in borderline expression ranges. This study evaluated the concordance between PD-L1 expression scores assessed by an expert pathologist and the uPath VENTANA PD-L1 (SP263) Assay Algorithm. The cohort included 333 non-small cell lung cancer (NSCLC) cases stained with the anti-PD-L1 (SP263) antibody. Digital slides were independently assessed by a board-certified pathologist and the uPath algorithm. Analysis used the Tumor Proportion Score (TPS), stratified into three categories: <1%, 1-49%, and ≥50%. The overall concordance of PD-L1 status (cutoff ≥1%) between the uPath algorithm and the pathologist was 95.2% (κ = 0.89), while the concordance for the three-tier categorization was 94.5% (weighted κ = 0.93). The median PD-L1 expression score from uPath was 1.2 percentage points (pp) higher than the pathologist's assessment. Discrepancies varied by TPS group, with mean differences of 0.2 pp, 4.4 pp, and 1.95 pp for TPS <1%, 1-49%, and ≥50%, respectively. Category changes occurred in 6% (20/333) of cases, potentially altering immunotherapy eligibility. AI-assisted assessment demonstrates high concordance with the expert pathologist, particularly in binary classification. However, due to discrepancies in the intermediate range, the algorithm is currently best positioned as a supportive tool for verifying borderline cases.