A narrative review of the prediction of immunotherapy efficacy for treating NSCLC: An artificial intelligence perspective

This is the authors' abstract. We don't add key points for this paper.

Immunotherapy efficacy in non-small-cell lung cancer (NSCLC) remains variable, with traditional biomarkers (PD-L1 and TMB) limited by heterogeneity and sensitivity/specificity constraints. This review evaluates the role of artificial intelligence (AI) in advancing predictive biomarkers through the integration of radiomics, pathomics, and multi-omics. We conducted a narrative literature review (PubMed/Web of Science, 2012–2025) using keywords spanning NSCLC, immunotherapy, and AI methodologies (radiomics, pathomics, and multi-omics). The inclusion criterion was studies that developed predictive models for immune checkpoint inhibitor (ICI) outcomes; the exclusion criteria were non-original research and preclinical studies. Emerging evidence highlights AI’s transformative role in the prediction of immunotherapy efficacy for treating NSCLC: radiomics quantifies tumor heterogeneity through CT-based PD-L1 and TMB ≥ 10 mut/Mb biomarkers, while pathomics refines PD-L1 scoring and immune microenvironment classification. Multi-omics integration synergizes imaging, histopathology, and liquid biopsy data with dynamic models by resolving spatial discordance and temporal clonal evolution. However, clinical adoption requires the standardization of peri-tumoral radiomics, validation of delta-feature protocols, and prospective trials to address overfitting risks in small cohorts. AI-driven multi-omics redefines immunotherapy prognostication, but faces interpretability and ethical challenges. Future validation must prioritize prospective multicenter trials, causal biological linking of radiomic features, and federated learning frameworks to ensure equitable clinical implementation.