Leveraging calcium score CT radiomics for heart failure risk prediction

Abstract

Studies have used extensive clinical information to predict time-to-heart failure (HF) in patients with and without diabetes mellitus (DM). We aimed to determine a screening method using only computed tomography calcium scoring (CTCS) to assess HF risk. We analyzed CTCS scans from 1,998 patients (336 with type 2 diabetes) from a no-charge coronary artery calcium score registry (CLARIFY Study, Clinicaltrials.gov NCT04075162). We used deep learning to segment epicardial adipose tissue (EAT) and engineered radiomic features of calcifications ("calcium-omics") and EAT ("fat-omics"). We developed models incorporating radiomics to predict risk of incident HF in patients with and without type 2 diabetes. At a median follow-up of 1.7 years, 5% had incident HF. In the overall cohort, fat-omics (C-index: 77.3) outperformed models using clinical factors, EAT volume, Agatston score, calcium-omics, and calcium-and-fat-omics to predict HF. For DM patients, the calcium-omics model (C-index: 81.8) outperformed other models. In conclusion, CTCS-based models combining calcium and fat-omics can predict incident HF, outperforming prediction scores based on clinical factors.Please check article title if captured correctly.YesPlease check and confirm that the authors and their respective affiliations have been correctly identified and amend if necessary.Yes.

Figure 1

Overview of project workflow. First, we segment epicardial adipose tissue (EAT) and coronary artery calcifications (CAC) from the CTCS image. We differentiate the EAT from paracardial adipose tissue (PAT). Next, radiomic features are extracted from EAT and CACs, from which we train time-to-event models and then analyze.

Figure 2

Heart failure free survival in full cohort, stratified by DM status. The term “at risk” in the given context indicates the number of patients who are susceptible to heart failure at the specified time points (0, 1, 2, 3, 4, 5 years).

Figure 3

DM Testing Set ROC Curves for HF prediction at 1 year post CTCS exam for DM-specific Age-sex-BMI, Agatston, EAT Volume, Calcium-omics, and Fat-omics Models. Statistical significance was determined using the DeLong test on AUC values. * indicates p-value < 0.05, ** indicates p-value < 0.005.

Figure 4

Demonstration of rule-based decision tree, to combine best models for DM and no-DM population. If the patient is diagnosed with diabetes, their heart failure (HF) risk is predicted using the calcium-omics model. Conversely, if the patient does not have diabetes, the HF risk is calculated using the fat-omics model.

Figure 5

Heart-failure-free survival in full cohort, stratified by mean EAT Thickness. The term “at risk” in the given context indicates the number of patients who are susceptible to heart failure at the specified time points (0, 1, 2, 3, 4, 5 years).