Beyond MELD Score: Association of Machine Learning-derived CT Body Composition with 90-Day Mortality Post Transjugular Intrahepatic Portosystemic Shunt Placement

Abstract

Purpose: To determine the association of machine learning-derived CT body composition and 90-day mortality after transjugular intrahepatic portosystemic shunt (TIPS) and to assess its predictive performance as a complement to Model for End-Stage Liver Disease (MELD) score for mortality risk prediction.

Materials and methods: This retrospective multi-center cohort study included patients who underwent TIPS from 1995 to 2018 and had a contrast-enhanced CT abdomen within 9 months prior to TIPS and at least 90 days of post-procedural clinical follow-up. A machine learning algorithm extracted CT body composition metrics at L3 vertebral level including skeletal muscle area (SMA), skeletal muscle index (SMI), skeletal muscle density (SMD), subcutaneous fat area (SFA), subcutaneous fat index (SFI), visceral fat area (VFA), visceral fat index (VFI), and visceral-to-subcutaneous fat ratio (VSR). Independent t-tests, logistic regression models, and ROC curve analysis were utilized to assess the association of those metrics in predicting 90-day mortality.

Results: A total of 122 patients (58 ± 11.8, 68% male) were included. Patients who died within 90 days of TIPS had significantly higher MELD (18.9 vs. 11.9, p < 0.001) and lower SMA (123 vs. 144.5, p = 0.002), SMI (43.7 vs. 50.5, p = 0.03), SFA (122.4 vs. 190.8, p = 0.009), SFI (44.2 vs. 66.7, p = 0.04), VFA (105.5 vs. 171.2, p = 0.003), and VFI (35.7 vs. 57.5, p = 0.02) compared to those who survived past 90 days. There were no significant associations between 90-day mortality and BMI (26 vs. 27.1, p = 0.63), SMD (30.1 vs. 31.7, p = 0.44), or VSR (0.97 vs. 1.03, p = 0.66). Multivariable logistic regression showed that SMA (OR = 0.97, p < 0.01), SMI (OR = 0.94, p = 0.03), SFA (OR = 0.99, p = 0.01), and VFA (OR = 0.99, p = 0.02) remained significant predictors of 90-day mortality when adjusted for MELD score. ROC curve analysis demonstrated that including SMA, SFA, and VFA improves the predictive power of MELD score in predicting 90-day mortality after TIPS (AUC, 0.84; 95% CI: 0.77, 0.91; p = 0.02).

Conclusion: CT body composition is positively predictive of 90-day mortality after TIPS and improves the predictive performance of MELD score.

Level of evidence: Level 3, Retrospective multi-center cohort study.

Keywords: Artificial intelligence; CT body composition; Machine learning; TIPS prognostication.

Fig. 1

Flowchart of patient selection 122 patients met the inclusion criteria, and their CT studies were segmented and analyzed using machine learning algorithms to obtain CT body composition metrics

Fig. 2

Machine learning-based CT body composition metrics. In patients undergoing TIPS procedure, automated computation of body composition from pre-procedural abdominal CT scan was performed. Body composition metrics were extracted and used to predict post-procedural outcomes. DenseNet was used for automatic L3 slice selection, and U-Net architecture model was used for segmentation

Fig. 3

Receiver operating characteristic for prediction of 90-day mortality. Receiver operating characteristics for a model using MELD score alone (blue line) in predicting 90-day mortality with an AUC of 0.76 and a model using MELD score in addition to SMA, SFA, and VFA (red line) with AUC of 0.84 (p = 0.026). SMA skeletal muscle area; SFA subcutaneous fat area; and VFA visceral fat area