AI-Driven Multimodal Risk Assessment Combining CT Imaging Biomarkers and Frailty Scores for Enhanced Mortality Prediction in Surgery Patients

Background: Automated frailty assessment predicts postoperative mortality, but routinely acquired CT imaging may provide complementary physiologic information. We evaluated whether AI-derived CT biomarkers improve one-year mortality prediction beyond an ICD-based Risk Analysis Index (RAI-ICD). Study Design: This retrospective cohort linked abdominopelvic CT scans obtained from 2013–2018 at a tertiary academic center to non-emergent surgery within 6 months. RAI-ICD was integrated with automated CT-derived muscle, adiposity, bone, and aortic calcification biomarkers using an XGBoost binary classification model, which we deemed the Unified Multimodal Model (UMM). The primary outcome was one-year all-cause mortality; discrimination, calibration, and decision-curve net benefit were evaluated. Results: Among 7,672 patients, one-year mortality was 12.3%; 7,638 patients were included in primary mortality models. RAI-ICD alone achieved an AUROC of 0.75 and outperformed every individual imaging biomarker. Adding all imaging biomarkers increased AUROC to 0.79 (ΔAUROC +0.04; P<0.001). UMM calibration (slope 1.00; intercept 0.01; ICI 0.01) exceeded RAI-ICD calibration (slope 0.72; intercept 0.04; ICI 0.03), particularly at higher predicted risk. At the Youden-optimal threshold, UMM PPV was 0.24, NPV 0.96, and number needed to screen was 4.1. UMM provided significantly greater net benefit overall and among Robust, Frail, and Very Frail strata. Conclusions: Integrating automated CT biomarkers with RAI-ICD improved one-year postoperative mortality prediction, high-risk calibration, and clinical net benefit compared with frailty assessment alone. External validation is required before clinical implementation.