AI reads routine chest CT scans to flag death risk in pulmonary hypertension

Artificial intelligence can pull prognostic signals out of ordinary chest CT scans that were never intended to assess the heart. In a retrospective study of 620 patients evaluated for suspected pulmonary hypertension at a specialist referral centre in Sheffield, England, two validated deep learning models measured the volumes of cardiac chambers, great vessels and coronary artery calcifications on routine non-gated, non-contrast CT. The automated measurements turned out to predict who would die during follow-up, and they did so even after accounting for age, sex, the subtype of pulmonary hypertension and an established clinical risk score. The work, published in European Radiology Cardiovascular and Thoracic, suggests that scans already sitting in hospital archives could yield additional prognostic biomarkers without exposing patients to any extra radiation.

The patients came from the ASPIRE registry, which enrols people who underwent specialist assessment for suspected pulmonary hypertension at the Sheffield Pulmonary Vascular Disease Unit and had not yet received any prior treatment. To enter the study, patients had to be at least 18 years old and to have undergone both right heart catheterisation and CT imaging within one month of assessment. From an initial pool of 884 patients, the investigators excluded 189 whose CT scan was performed more than a month from catheterisation, 9 with missing survival status and 66 who were unclassifiable because haemodynamic data or pulmonary hypertension subtypes were missing. The final cohort of 620 patients was enrolled between 2006 and 2020, with a mean age of 64 years, 62% female participants and 80% self-reporting as white British. Scans came from GE Healthcare systems in 78% of cases, Canon in 18% and Siemens in 4%.

Most of the cohort, 76%, had pulmonary hypertension. Within that group, 202 patients had pulmonary arterial hypertension, 84 had pulmonary hypertension due to left heart disease, 112 had pulmonary hypertension associated with lung disease, 65 had chronic thromboembolic pulmonary hypertension and 9 had unclear disease, while 148 patients did not have pulmonary hypertension. During follow-up, 388 patients died, 63% of the cohort, with survival durations ranging from 2 to 5,187 days and a mean of 1,659 plus or minus 1,207 days. Coronary calcifications were common: 45% of patients had calcification in the right coronary artery, 67% in the mainstem and left anterior descending artery, and 35% in the left circumflex artery, a pattern the authors read as a substantial burden of cardiovascular comorbidity.

The imaging analysis rested on two deep learning models developed and validated in earlier work. The cardiothoracic segmentation model, built on the nnU-Net framework, was trained on manual annotations from 50 non-contrast CT scans and evaluated on an external cohort of 55 scans, achieving Dice similarity coefficients of at least 0.87 and Hausdorff distances of at most 1.7 for the right and left ventricles, right atrium, left atrium, pulmonary artery, ascending aorta and descending aorta. Tested across 368 scans, it reached areas under the curve of 0.84 to 0.88, with sensitivities of 70 to 94% and specificities of 63 to 100%. The coronary calcification model, developed on 69 scans, was visually assessed against 100 scans by two experienced observers and tested on 473 scans, showing a sensitivity of 95%, specificity of 98%, positive predictive value of 99% and negative predictive value of 88% compared with radiologists’ reports.

Applied to the study cohort, the models quantified chamber and vessel volumes and detected calcifications in each coronary territory. Patients with pulmonary hypertension had significantly larger right atrial, ventricular and pulmonary artery volumes than those without the disease, with mean differences of 44.3 mL, 51.4 mL and 36.6 mL respectively, all p < 0.001, while left atrial, ascending aortic and descending aortic volumes did not differ significantly. The subgroup comparisons followed the same pattern with variations: pulmonary arterial hypertension patients showed larger right atrial, ventricular and pulmonary artery volumes; those with left heart disease showed enlargement across right atrium, left atrium, ventricles and all three vessels; lung disease-associated disease produced larger right atrial, ventricular and pulmonary artery volumes; and chronic thromboembolic disease showed larger right atrial and pulmonary artery volumes only.

The mortality analysis proceeded in stages. Univariable Cox regression, with volumetric hazard ratios expressed per 10-mL increase, linked larger left atrial, right atrial, ascending aortic, pulmonary artery and descending aortic volumes to higher mortality, as did calcification in each coronary territory. Ventricular volume and a pulmonary artery to ascending aorta ratio above 1 were not significant in these single-variable models. After adjustment for age, sex and pulmonary hypertension subgroup, right atrial volume remained independently associated with death, with a hazard ratio of 1.024 per 10 mL (95% confidence interval 1.009 to 1.039, p = 0.001). Calcification stayed significant in every segmented artery: the right coronary artery carried a hazard ratio of 1.378 (95% CI 1.113 to 1.705, p = 0.003), the mainstem and left anterior descending artery 1.868 (95% CI 1.441 to 2.421, p < 0.001) and the left circumflex artery 1.342 (95% CI 1.084 to 1.660, p = 0.007). The ratio of pulmonary artery to ascending aorta above 1 also emerged with a hazard ratio of 1.361 (95% CI 1.094 to 1.694, p = 0.006), whereas pulmonary artery, left atrial, aortic and ventricular volumes lost their independent significance.

Dichotomising the significant variables at Youden-derived thresholds sharpened the picture. Right atrial volumes above 117.95 mL were associated with a 61% higher risk of mortality (hazard ratio 1.611, 95% CI 1.298 to 1.999, p < 0.001), and calcification volumes above 0.01 mL carried hazard ratios of 1.533 for the right coronary artery, 2.114 for the mainstem and left anterior descending artery and 1.465 for the left circumflex artery, all p < 0.001. Kaplan-Meier curves for five-year all-cause mortality separated the low-volume and high-volume groups for each of these measures. In an exploratory analysis adjusted for the REVEAL 2.0 risk score, a tool validated primarily for pulmonary arterial hypertension that incorporates demographic, clinical, functional, exercise, laboratory, imaging and haemodynamic parameters, the associations persisted: each 10-mL increase in right atrial volume conferred a 2.6% higher risk (hazard ratio 1.026, 95% CI 1.012 to 1.041), left atrial volume a 2.4% higher risk (hazard ratio 1.024, 95% CI 1.001 to 1.047, p = 0.043), and calcification hazard ratios reached 1.604 for the right coronary artery, 2.339 for the mainstem and left anterior descending artery and 1.576 for the left circumflex artery, all p < 0.001. The pulmonary artery to aorta ratio, by contrast, was no longer significant after REVEAL adjustment, suggesting its prognostic contribution largely reflects disease severity already captured by the clinical score.

Within the pulmonary arterial hypertension subgroup alone, the pattern held. Adjusted for REVEAL 2.0, left atrial volume showed a hazard ratio of 1.046 (95% CI 1.008 to 1.086, p = 0.016) and right atrial volume 1.039 (95% CI 1.018 to 1.061, p < 0.001), while coronary calcification produced the strongest associations of all: 1.616 for the right coronary artery (p = 0.001), 2.829 for the mainstem and left anterior descending artery and 1.662 for the left circumflex artery, both p < 0.001. The authors note that most prior work linking right atrial dilatation to adverse outcomes relied on ultrasound or magnetic resonance imaging rather than non-contrast CT, so the automated CT approach extends that literature to a modality already in widespread use. They also point out that although non-dedicated scans are not performed for coronary assessment, radiologists should routinely evaluate coronary calcification given its established role as an independent predictor of all-cause and cardiovascular mortality.

The study carries clear limitations. It was retrospective and single-centre, the cohort was predominantly white British, most scans came from one vendor, the sample size was not determined by an a priori power calculation, and scans acquired between 2006 and 2020 across multiple platforms and reconstruction settings may have introduced variability in volumetric measurements despite predefined quality criteria. The impact of the pipeline on clinical practice was not assessed, and the authors call for validation in multi-centre, ethnically diverse populations with varied scanner systems. Even so, the findings indicate that AI-derived right atrial volume and coronary artery calcification measurements on routine non-gated, non-contrast CT independently predict mortality in pulmonary hypertension and retain prognostic value beyond established risk scores, offering a route to more comprehensive, imaging-based risk stratification from scans that hospitals already possess.

Subject of Research: AI-derived cardiothoracic volumes and coronary artery calcification on non-contrast CT as predictors of mortality in pulmonary hypertension

Article Title: AI-derived cardiothoracic structures and coronary artery calcification volumes on non-contrast CT as predictors of mortality in pulmonary hypertension

Article References: Alnasser, T. N., Hokmabadi, A., Sharkey, M., Kiely, D. G., Alabed, S., & Swift, A. J. (2026). AI-derived cardiothoracic structures and coronary artery calcification volumes on non-contrast CT as predictors of mortality in pulmonary hypertension. European Radiology Cardiovascular and Thoracic, 1(1), Article 14. https://doi.org/10.1007/s44449-026-00005-2

Image Credits: AI Generated

DOI: 10.1007/s44449-026-00005-2

Keywords: pulmonary hypertension, artificial intelligence, deep learning, coronary artery calcification, non-contrast CT, right atrial volume, risk stratification, REVEAL score, ASPIRE registry, all-cause mortality, cardiothoracic imaging, right heart catheterisation