eCAPRI: a novel tool combining clinical and imaging data for post-TAVI mortality prediction

Abstract

Introduction

TAVI has become a validated alternative to surgical valve replacement for severe aortic stenosis, yet up to a quarter of high-risk patients still have poor outcomes at one year, largely because of comorbidities. The CAPRI score of Lantelme et al. predicts one-year mortality after TAVI by adding thoracic aortic calcium (TAC) volume measured on CT to clinical, biological and demographic factors (AUC 0.68). Its main barrier to adoption is the manual annotation of calcium, which is slow.

Imaging biomarkers measured automatically on routine CT have also shown value for long-term outcome prediction. eCAPRI automates TAC measurement with deep learning and adds further CT biomarkers to improve the CAPRI score.

Method

2.1The CAPRI score

Variables are selected with univariate Cox regressions and a likelihood ratio test, then checked for mutual correlation. The score of a patient sums, over the selected variables, the Cox coefficient β_i times the deviation of the variable from its training-set mean.

2.2Automated aortic calcium segmentation

A 3D U-Net (TensorFlow 2.10) is trained with a masked binary cross-entropy: heart, great-vessel and spine segmentations provide exclusion masks for the spine and the aortic and mitral valves and a dilated aorta mask, so the network only learns what calcium looks like, not where it is. This allows a small network at full image resolution. Images are cropped around the aorta, windowed to [−200, 1000] HU and normalized to [−1, 1]; training uses batches of 5 random 100³ patches, Adam with a learning rate starting at 10⁻³ and decayed by 0.88 per epoch.

2.3Imaging biomarkers

Pre-trained, locked models segment the heart chambers, great vessels, lungs, spine and abdominal muscles at L3. From them are derived the cardiac chamber and myocardium volumes, the maximal aorta and pulmonary-artery diameters (topological thinning), emphysema and healthy lung volumes and their ratio (voxels below −950 HU), and the abdominal muscle surface at L3, a marker of sarcopenia. Every biomarker except the emphysema ratio is also indexed to body surface area (BSAi). Fat and density measures were excluded because contrast agent biases them.

2.4Biomarker selection and eCAPRI

1. Univariate filter. Keep biomarkers whose likelihood-ratio p-value is below 0.2, a loose threshold that allows for confounding interactions.
2. Redundancy. Drop a biomarker correlated above 0.85 with a CAPRI variable or with a biomarker that has a better p-value.
3. Combination search. Evaluate every combination of the remaining biomarkers added to the CAPRI variables with 4-fold cross-validation; the best mean AUC defines eCAPRI, which is then refit on the whole training set.

eCAPRI is compared with CAPRI (reweighted on the training set) and the EuroSCORE Logistic with ROC curves, a 10⁴-sample permutation test between CAPRI and eCAPRI, and calibration curves with Brier scores.

Experiments

3.1Data

The TAVI-Lyon cohort (479 men, 478 women, mean age 82.35 ± 7.53) was split 80/20 with an equal proportion of deaths. Comorbidities were frequent: renal impairment of stage III or worse in over half of patients, NYHA class III/IV heart failure in 49% and atrial fibrillation in 31%.

3.2Calcium segmentation

Key findings

• Accurate on a large cohort: a mean Dice of 0.777 on 1,111 contrast scans, from only 66 training patients.
• Low Dice comes from tiny volumes: 60 scans fall below 0.6; the 29 below 0.5 have a mean calcium volume of 0.4 mL, versus 4.2 mL overall.
• Slight overestimation: predicted volumes run a little above the ground truth; heavily injected vena cava, artefacts and tracheal calcifications cause most errors.

3.3Biomarker selection

Key findings

• Lungs carry little signal: pulmonary biomarkers were largely uncorrelated with one-year death.
• Three biomarkers retained: BSA-indexed right-ventricle volume, BSA-indexed pulmonary-artery maximal diameter and L3 muscle surface — clinically recognized markers of right heart dysfunction, pulmonary hypertension and frailty.

3.4Mortality prediction

Key findings

• Better discrimination: eCAPRI reaches AUC 0.731, significantly above CAPRI (0.669) and the EuroSCORE Logistic (0.588).
• Better calibration: the lowest Brier score (0.079) and the calibration curve closest to ideal; the EuroSCORE underestimates risk in the highest deciles.
• Automation preserves CAPRI: with automated calcium, CAPRI reaches AUC 0.669, close to the original publication (0.68).

Discussion & limitations

Because every TAVI candidate already has a pre-procedural contrast CT, all eCAPRI inputs are available in routine care. With automated segmentation and biomarker extraction, it could run in the background on radiology workstations or PACS and give clinicians a standardized risk estimate without extra workload. The segmentation still struggles with hardening artefacts near the aorta and tracheal calcifications, and slightly overestimates calcium volume. The −950 HU emphysema threshold may not suit contrast scans, which could explain why emphysema was not retained. Further validation on broader populations remains important.

Related links

1. eCAPRI: a novel tool combining clinical and imaging data for post-TAVI mortality prediction, European Radiology 36:3542–3555 (2026). Open access under a CC BY 4.0 licence; the figures on this page are reproduced from the article, and the tables are adapted from it.
2. P. Lantelme et al., the original CAPRI score for one-year mortality after TAVI.

BibTeX

@article{djahnine2025ecapri,
  title = {eCAPRI: a novel tool combining clinical and imaging data for post-TAVI mortality prediction},
  author = {PJ. Lartaud and B. Harbaoui and A. Della Corte and A. Djahnine and O. Nempont and JM. Rouet and A. Cecilia Vlachomitrou and B. Riche and M. Rabilloud and S. Si-Mohamed and P. Douek and F. Derimay and G. Rioufol and P. Lantelme and L. Boussel},
  journal = {European Radiology},
  year = {2026},
  doi = {10.1007/s00330-025-12184-x}
}