Personalizing TIPS Care: Machine Learning-Driven CT Body Composition for Predicting Post-Procedural Hepatic Encephalopathy

Monday, March 31, 2025

3:54 PM – 4:03 PM CT

Location: Music City Center, 205 A

Presenting Author(s)

TE

Tarig Elhakim, MD

Medicine Instructor & Research Associate
Penn (Medicine Instructor) & MGH (Research Associate), United States

Author/Co-author(s)

Allison Brea

Medical Student
Tufts University School of Medicine, United States
CB

Christopher P. Bridge, DPhil (he/him/his)

Instructor in Radiology
Martinos Center for Biomedical Imaging, Harvard Medical School, United States
DD

Dania Daye, MD, PhD

Associate Professor
Massachusetts General Hospital/Harvard Medical School, United States

Purpose: To assess the association between machine learning-based CT body composition metrics and the development of Hepatic Encephalopathy (HE) after Trans-jugular Intrahepatic Portosystemic Shunt (TIPS) placement.

Materials and Methods: Patients were included in this retrospective multicenter study if they underwent a TIPS procedure between 1995-2018 and had a CT scan of the abdomen and pelvis within one year preceding their TIPS placement. The patient charts were reviewed to collect various clinical variables, including the development of post-TIPS HE within 90 days following TIPS. A previously validated machine-learning algorithm, utilizing dual convolution neural network models, was used to extract the visceral fat area (VFA), subcutaneous fat area (SFA), and skeletal muscle area (SMA) from the L3 vertebral level. Independent t-tests and multivariate logistical regression models were used to evaluate the association between CT body composition metrics and the development of HE within 90 days after TIPS.

Results: 158 (57.8 years ± 11.5) patients met the inclusion criteria. Patients who developed HE (N=82) were older (60.1 years vs. 55.7 years, p=0.02) and had a higher frequency of hypertension (HTN) (69.5% vs. 52.6%, p=0.03). When comparing the HE group to the unaffected group, the frequency of coronary artery disease (CAD) (24.4% vs. 15.8%, p=0.18), diabetes (DM) (40.2% vs. 31.6%, p=0.26), and chronic kidney disease (CKD) (35.4% vs. 25.0%, p=0.16) were not significant. The HE group had more prescribed Rifaximin (81.7% vs. 40.8%, p< 0.01) and Lactulose after TIPS (90.2% vs. 68.4%, p< 0.01) and a higher frequency of noncompliance with these medications (18.3% vs. 5.26%, p< 0.01). The mean ammonia level of the HE group was higher than the unaffected group (144.5 μg/dL vs. 85.4 μg/dL, p< 0.02). The mean CT body composition metrics in the HE group compared to the unaffected were not statistically significant (SMA: 136.5 cm² vs. 144.4 cm², p=0.16 | SFA: 159.7 cm² vs. 183.9cm², p=0.16 | VFA: 143.7 cm² vs. 156.1 cm², p=0.46). After adjusting for age, comorbidities, and medication noncompliance, SMA (OR 0.98, p=0.03), SFA (OR 0.99, p=0.03), and VFA (OR 0.99, p=0.03) were all independent predictors of HE within 90 days after TIPS. Additionally, SMA (OR 0.98, p=0.03), SFA (OR 0.99, p=0.02), and VFA (OR 0.99, p< 0.01) were all independent predictors for Rifaximin use after TIPS.

Conclusion: Fully automated machine learning-based CT body composition metrics are independent predictors of post-TIPS HE development within 90 days. The metrics were also independently associated with Rifaximin use after TIPS.