Application of AI in CT scan for Predicting Response to Immunotherapy and Immunochemotherapy in Non-Small Cell Lung Cancer: A Systematic Review and Meta-analysis

Background: Immune checkpoint inhibitors have transformed advanced non-small cell lung cancer (NSCLC) management, yet only few patients achieve sustained benefit. Traditional predictive biomarkers exhibit low accuracy and require invasive tissue sampling. Artificial intelligence (AI)-based computed tomography (CT) analysis offers a non-invasive alternative, but its clinical utility is unclear.

Purpose: To systematically evaluate the diagnostic accuracy of AI-based CT models for predicting immunotherapy or immunochemotherapy response in NSCLC patients.

Methods: This systematic review and meta-analysis followed PRISMA guidelines and was registered in PROSPERO (CRD420251055338). PubMed, Scopus, Web of Science, and Cochrane Library were searched from inception to May 2025. Studies evaluating AI-based CT models for predicting immunotherapy response in adult NSCLC patients were included. Quality assessment used QUADAS-2. Diagnostic accuracy meta-analysis employed a bivariate random-effects model within a hierarchical summary receiver operating characteristic framework. Subgroup analyses explored sources of heterogeneity.

Results: Fifteen studies were included. Pooled sensitivity was 0.80 (95% CI: 0.73–0.85) and specificity 0.76 (95% CI: 0.68–0.82), with HSROC area under the curve of 0.85. Positive likelihood ratio was 3.32 (95% CI: 2.52–4.37), negative likelihood ratio 0.21 (95% CI: 0.14–0.31), and diagnostic odds ratio 19.76 (95% CI: 11.33–34.46). Subgroup analysis revealed superior performance in test sets (AUC 0.90) versus external validation (AUC 0.73). Contrast-enhanced CT demonstrated better accuracy than non-contrast approaches. Supervised machine learning and deep learning showed comparable performance.

Conclusion: AI-based CT models demonstrate moderate-to-good diagnostic accuracy for predicting immunotherapy and immunochemotherapy response in NSCLC and show promise as non-invasive decision-support tools. However, lower external validation performance and methodological heterogeneity limit immediate clinical implementation. Prospective multicenter validation with standardized protocols is needed.