(PC098) Fully Automated Multimodal Risk Assessment Using Body Composition Biomarkers and Frailty Scores Enhances Mortality Prediction in Patients Undergoing Cardiovascular Surgery

Objectives: Opportunistic computed tomography (CT) analysis repurposes routine clinical CT scans for disease prognostication, often utilizing machine learning. This study harnesses artificial intelligence to develop a multimodal risk assessment model by extracting body composition features from abdominopelvic CT scans and integrating them with ICD code-based frailty assessments to predict post-operative mortality in cardiovascular surgery patients.

Methods: An initial cohort of 36,354 intravenous contrast-enhanced abdominopelvic CT scans performed 2013-2018 at our tertiary care emergency department were identified. From this cohort, 1,831 CT-patient pairs who underwent any cardiovascular procedures either within the subsequent 6 months or the prior 3 months were included. Clinical frailty scores were calculated using the validated Risk Analysis Index (RAI)-ICD frailty index, derived from ICD-9/10 codes. A fully automated deep-learning pipeline was utilized to extract skeletal muscle area and density; spinal bone mineral density; subcutaneous and visceral adipose tissue cross-sectional areas; and abdominal aortic calcification volume from the CT scans. Operative severity was classified using the Operative Stress Score (OSS; range: 1-5). Predictive models for post-operative all-cause mortality were developed using XGBoost, a gradient-boosting algorithm. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC).

Results: In this heterogenous cohort, RAI-ICD outperforms individual morphometric markers in predicting all-cause mortality (Table), with the highest AUROC (0.82) and AUPRC (0.76). Among individual markers, bone density demonstrates strong predictive value (AUROC: 0.80), but all markers alone underperform compared to RAI-ICD. The predictive performance for all-cause mortality improves, as additional imaging biomarkers are integrated with RAI (Figure), with the combination of RAI adipose, muscle, bone, and aortic calcification achieving the highest AUROC (0.86; P-value compared to RAI-ICD:0.008) and AUPRC (0.83; P-value compared to RAI-ICD < 0.001). The best combination model has comparable performance across all four frailty categories - with the “Robust” group achieving AUPRC: 0.74 and the “Very Frail” group AUPRC: 0.71. Similarly, the best combination model performed well across all OSS classifications, with the best performance observed in high-stress procedures OSS 3-5 (AUROC: 0.89, AUPRC: 0.85).

Conclusions: A fully automated, multimodal approach integrating morphometric biomarkers from CT scans and frailty score from electronic medical data improves mortality prediction in cardiovascular surgery patients. Although RAI-ICD is a robust individual predictor of all-cause mortality, its combination with additional imaging biomarkers enhances model performance. Future research is needed to externally validate these models and develop prospective risk-prediction models with clinical implementation in vascular-surgery-specific cohorts.