Automated opportunistic CT body composition markers and overall survival in patients with digestive system cancers: a retrospective cohort study

Routine abdominal CT contains quantitative muscle and skeletal information. We evaluated associations of automatically extracted CT body composition markers with overall survival (OS) and their incremental predictive value beyond clinical variables in digestive system cancers. This single-center retrospective study considered 405 patients with pretreatment CT and survival follow-up. TotalSegmentator masks of the L3 vertebra and bilateral iliopsoas muscles underwent manual quality control. Marker-OS associations were evaluated with univariable and adjusted Cox models. Complete cases were randomly divided into training and internal-validation sets. Clinical, CT body composition, and fusion ridge Cox models were compared using C-indices, time-dependent AUCs, and IPCW Brier scores. After imaging quality control, 392 patients had usable L3 imaging; 377 complete cases formed the common primary-analysis cohort (300 training; 77 internal validation). Higher muscle-quantity markers were associated with lower hazard in univariable analyses. After adjustment, L3 vertebral center-slice attenuation and whole-iliopsoas volume index were associated with OS (HR per 1-SD increase, 1.155 and 0.843, respectively). Internal-validation C-indices were 0.772 (95% CI, 0.699–0.841) for fusion, 0.762 (0.693–0.834) for clinical, and 0.571 (0.476–0.662) for CT body composition. The fusion-minus-clinical difference was + 0.009 (95% CI, -0.008 to 0.027; compareC, P = 0.28). All 1-, 2-, and 3-year AUC-difference CIs included zero. The fusion model had lower Brier scores (2 years, 0.190 vs. 0.248). Selected automatically extracted CT body composition markers were associated with OS, but adding them to clinical variables provided little incremental discrimination in internal validation (C-index difference, + 0.009; 95% CI, -0.008 to 0.027; compareC, P = 0.28). Continuous net reclassification improvement and integrated discrimination improvement were not statistically significant. Lower Brier scores were a secondary finding. Clinical utility is not established, and independent external validation is needed.