A deep learning AI model can predict long-term mortality in patients by analyzing aortic valve fibrotic and calcific tissue from CT angiography scans, completing analysis in 1.35 minutes with accuracy matching expert cardiologists and providing a novel biomarker for improved risk stratification in aortic stenosis.
- Deep learning model analyzed 1,674 patients across nine international medical centers to predict mortality from CTA scans
- AI completed tissue analysis in 1.35 minutes versus 7.2 minutes for expert readers, with 99.9% scan-rescan reproducibility
- Patients with highest fibrotic tissue volumes faced 2-fold higher death risk, while dense calcific tissue increased risk by 1.5-fold
- Over 4.1 years, 40% of patients died, with fibrotic tissue volume independently predicting mortality beyond traditional calcification measures
- First study to demonstrate deep learning enables rapid, accurate quantification of aortic valve tissue with excellent agreement to histology and expert assessment
A deep-learning AI model can predict long-term mortality in patients by quantifying aortic valve fibrotic and calcific tissue from CT angiography (CTA) scans, according to a study published November 5 in JACC: Cardiovascular Imaging.
The finding is from an international, multicenter study that included 1,674 patients who underwent CTA at nine sites for routine transcatheter aortic valve replacement (TAVR) planning, noted lead author Kajetan Grodecki, MD, PhD, of Cedars-Sinai Medical Center in Los Angeles, and colleagues.
"Computed tomography-based quantification of aortic valve fibrotic tissue provides novel prognostic insights, independently predicting long-term mortality after TAVR and offering a novel biomarker to improve risk stratification and therapeutic evaluation in aortic stenosis beyond traditional calcific measures," the group wrote.
Echocardiography is the standard modality for assessing aortic stenosis, but leads to discordant severity grading in roughly one-third of patients, while CTA-based calcium quantification has gained traction as a complementary tool, the authors explained. Yet manual CTA tissue composition analysis is labor-intensive and demands substantial reader expertise, limiting its clinical uptake, they added.
To determine whether AI can help close the gap, the researchers trained a deep learning system using the nnU-Net architecture on 468 patients who underwent preintervention CTA before TAVR, and then applied it to an external test set of 1,206 patients. Participating centers were located in Poland, Germany, the United Kingdom, and the United States. Performance was assessed against expert interpretations, histological valve analysis (n = 29), and scan-rescan reproducibility (n = 12). The primary endpoint was time to all-cause death after TAVR in 1,165 patients with clinical follow-up.
Comparison of deep learning segmentation (A) with quantitative histology (B) using Movat's Pentachrome staining (red = fibrotic tissue; yellow = calcific tissue; blue = aortic lumen), demonstrating excellent agreement between methods (C) for quantification of aortic valve composition. Fibrotic tissue is expressed as a percentage and was calculated as [(tissue component volume/total tissue volume) × 100].JACC: Cardiovascular Imaging
Over a mean 4.1 years, 469 patients (40%) died. Greater fibrotic tissue volume (adjusted hazard ratio [aHR] 1.42), denser calcific tissue attenuation (aHR 1.02), and reduced ejection fraction independently predicted mortality. Patients with the highest fibrotic tissue volumes (> 0.940 cm³) faced a 2-fold higher death risk, while those with dense calcific attenuation (> 885 Hounsfield units) faced a 1.5-fold higher risk, the researchers reported.
"To our knowledge, this is the first study to demonstrate that deep learning enables rapid quantification of aortic valve fibrotic and calcific tissue, with excellent agreement to expert assessments, strong concordance with histology, and excellent scan–rescan reproducibility," the group wrote.
Ultimately, the authors described the findings as hypothesis-generating and called for further study to clarify the mechanistic pathways linking valve tissue composition to adverse outcomes. Incorporation of fibrotic tissue assessment into clinical practice may improve risk stratification, identify patients at higher risk before irreversible myocardial consequences develop, and serve as a more sensitive endpoint than calcification alone in clinical trials evaluating disease-modifying pharmacotherapy, they suggested.
The full study is available here.
