AI for tumour proportion scoring of programmed death‐ligand 1 immunohistochemistry in non‐small cell lung cancer: a review of commercial and non‐commercial tools

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Programmed death‐ligand 1 (PD‐L1) expression, commonly quantified as tumour proportion score (TPS), is a key biomarker guiding immunotherapy in non‐small cell lung cancer (NSCLC). Manual scoring of PD‐L1 immunohistochemistry (IHC) is time‐consuming and challenged by preanalytical variability and interpretive subjectivity, which can hinder consistent treatment selection. Artificial intelligence (AI) tools aim to support more standardized PD‐L1 scoring. We compare publicly available evidence on PD‐L1 AI tools and identify barriers to routine implementation. We mapped publicly available evidence for seven commercial and 13 non‐commercial or research‐stage PD‐L1 AI tools across six interconnected levels: data characteristics and validation setting, data preparation and reference standard, supervision strategy, performance evaluation and agreement, generalizability and clinical relevance, and deployment traceability. Commercial sources more often described workflow integration, intended use, reader assistance and certification or product status, whereas non‐commercial studies more often detailed model architecture, supervision strategy, code/data availability and experimental design. Heterogeneity in cohorts, assay–scanner settings, reference standards, agreement metrics and decision‐threshold reporting limited direct comparison. Key gaps included scarce prospective multi‐centre validation, limited reporting at decision‐relevant TPS thresholds, and incomplete separation between peer‐reviewed validation evidence, vendor claims and deployment traceability. The resulting evidence map supports conditional appraisal and local verification rather than a universal ranking. Publicly available evidence for PD‐L1 TPS AI tools remains heterogeneous, and reported performance is not automatically portable between laboratories because it depends on the combined case mix, assay–scanner workflow, reference standard and TPS‐threshold setting. Therefore, current evidence does not support a universal ranking for routine practice. The six‐level comparison structure and end‐user crosswalk support shortlisting intended‐use‐compatible tools and, where relevant, comparing them locally using the same representative cases, assay–scanner workflow, reference standard and prespecified TPS‐threshold endpoints. More transparent reporting, robust external and prospective multi‐centre validation, and ongoing quality monitoring are needed for safe integration into routine human–AI pathology workflows.