Background and aim: Artificial intelligence (AI)-based models have been increasingly investigated for prognostic risk stratification in non-small cell lung cancer (NSCLC). This systematic review and meta-analysis evaluated the association between AI-predicted risk groups and survival outcomes.
Material and methods: PubMed/MEDLINE, Embase, Scopus, Web of Science, and CENTRAL were searched through November 10, 2026. Studies were eligible if they evaluated AI-based models in patients with NSCLC and reported hazard ratios (HRs) comparing high-risk with low-risk groups for overall survival (OS), progression-free survival (PFS), disease-free survival (DFS), or recurrence-free survival (RFS). Random-effects meta-analyses were performed using restricted maximum likelihood estimation. Methodological quality was assessed using the Newcastle–Ottawa Scale.
Results: Thirty-one studies involving 53,727 patients were included. High-risk classification was associated with poorer OS (21 studies; 50,105 patients; HR, 2.45; 95% CI, 2.09–2.89), PFS (15 studies; 42,998 patients; HR, 2.71; 95% CI, 2.20–3.35), DFS (5 studies; 2,431 patients; HR, 1.76; 95% CI, 1.30–2.39), and RFS (2 studies; 970 patients; HR, 3.94; 95% CI, 1.76–8.80). Heterogeneity was moderate to substantial across outcomes (I², 65.0%–85.0%). In subgroup analyses, immunotherapy status did not significantly modify the association with OS (P for subgroup difference=0.20); the difference for PFS was borderline (P=0.050). Twenty-six studies (83.9%) were rated as high quality. Funnel-plot asymmetry and statistical tests suggested potential small-study effects for OS and PFS.
Conclusions: AI-based high-risk classification was consistently associated with poorer survival across cohorts and all evaluated outcomes in NSCLC. Heterogeneity and potential small-study effects, together with limited evidence for DFS and RFS, warrant cautious interpretation and further prospective external validation.