Non-invasive prediction of prognostic immune subtypes in lung adenocarcinoma using PET/CT-based radiomics

Background The treatment and prognosis of lung adenocarcinoma (LUAD) remain major clinical challenges. While transcriptomic profiling can define immune subtypes, its invasive nature limits clinical utility. Objectives We aimed to bridge this translational gap by developing and validating a non-invasive 18 F-FDG PET/CT radiomics-based biomarker to predict these subtypes and assess their therapeutic relevance. Methods We performed consensus clustering on transcriptomic data from 773 LUAD patients (TCGA/GEO) to identify novel immune subtypes, developed a radiomics-based machine learning classifier using their PET/CT data (n=115) and conducted a biologically and clinically supported external evaluation in an independent cohort from Tianjin Medical University Cancer Hospital (TMUCIH, n=249). In this cohort, we compared immune patterns with transcriptomic subtypes via immunohistochemical staining, and their prognostic significance was evaluated in patients receiving EGFR tyrosine kinase inhibitors (n=118) and immune checkpoint inhibitors (n=75). Results We identified three robust immune subtypes (log-rank p=0.0017): Inflammatory (ClusterA), Activated (ClusterB), and Evasive (ClusterC). Our radiomics model predicted these subtypes with high accuracy (one-vs-rest AUC = 0.84). Immunohistochemical patterns were consistent with predicted subtypes and stratified overall survival in both the TKI-treated (log-rank p-value=0.041) and ICI-treated (log-rank p-value =0.024) cohorts. These results support a validated, non-invasive approach linking genomics and radiomics to enhance patient stratification and guide personalized treatment strategies in LUAD.