CT biomarkers improve postoperative mortality prediction

In a retrospective cohort of 7,672 patients undergoing non-emergent surgery at a tertiary academic center, researchers evaluated whether automated CT imaging biomarkers could improve postoperative mortality predictions. They linked abdominopelvic CT scans obtained within six months prior to surgery with an administrative frailty metric, the ICD-based Risk Analysis Index (RAI-ICD). Using an XGBoost algorithm, the team combined RAI-ICD with automated measurements of muscle, adiposity, bone, and aortic calcification into a Unified Multimodal Model.

Across the cohort, one-year mortality was 12.3%. In 7,638 patients included in primary mortality models, RAI-ICD alone yielded an AUROC of 0.75, outperforming every individual CT biomarker. Integrating all imaging biomarkers with RAI-ICD increased the AUROC to 0.79 (P < 0.001) and improved model calibration, shifting the slope from 0.72 to 1.00. The multimodal model also delivered greater clinical net benefit across robust, frail, and very frail patients.