PET/CT-derived whole-body composition and survival in resectable NSCLC: double machine learning-based adjusted association analysis of intermuscular adiposity burden and sex-specific metabolic phenotypes

To evaluate whether whole-body PET/CT-derived body composition features are associated with survival in patients with resectable non-small cell lung cancer (NSCLC), using double machine learning to quantify adjusted associations with restricted mean survival time. This retrospective multicenter study included 769 patients with stage ≤ IIIA NSCLC who underwent preoperative 18F-fluorodeoxyglucose positron emission tomography/computed tomography ( 18 F-FDG PET/CT) and curative resection. Center 1 was used for model development ( n = 555) and Center 2 for external validation ( n = 214). Automated whole-body segmentation was used to extract volumetric, attenuation-based and metabolic features of skeletal muscle and adipose tissue. Restricted mean survival time-based double machine learning with generalized propensity score weighting estimated DML-adjusted differences in restricted mean survival time for overall survival (OS) and progression-free survival (PFS). Ridge-penalized Cox models were further used to evaluate prognostic performance and incremental value. Higher intermuscular adipose tissue (IMAT) volume index was associated with shorter survival (DML-adjusted RMST difference, -4.29 months for OS and − 2.74 months for PFS per 1-SD higher IMAT volume index). Higher TAT SUR (Mean) showed an exploratory favorable adjusted association with longer OS (+ 4.42 months). Sex-stratified analyses suggested stronger adverse adipose-volume associations in male patients and stronger favorable adipose-metabolic associations in female patients. Model analyses showed moderate external discrimination, whereas incremental-value metrics were modest and endpoint-dependent. Whole-body PET/CT body-composition phenotyping may provide prognostic information in resectable NSCLC. Higher IMAT burden was associated with shorter survival, whereas TAT SUR (Mean) showed an exploratory favorable adjusted association with OS. A local software framework may support reproducible feature extraction and research-oriented risk stratification.