Global circumferential strain predicts death and heart failure using fully automatic analysis of tagging MRI

Background

Feature tracking (FT) global longitudinal strain (GLS) is a popular measure of LV function and can predict adverse outcomes in a general population [1]. However, FT strain values differ from the validated reference standard (Tagging MRI) and other global measures such as circumferential strain (GCS) show more limited utility. Previous Tagging studies were limited by time-consuming manual analysis, but we have developed novel deep learning methods for rapid and automatic strain analysis.

Purpose

We assessed the prognostic value of tagging-derived GCS for all-cause mortality and major adverse cardiovascular outcomes in a large population from the UK Biobank.

Method

Three short-axis tagging MRI slices were prospectively acquired in UK Biobank participants [2] and strain was analysed with a fully automatic deep learning pipeline developed in house [3]. FT strains were derived using CVI42 (Circle Cardiovascular imaging) [1].

Statistical analysis was performed in Python (v3.11.8) using Lifelines, Statsmodels and scikit-learn. Four endpoints were studied: all-cause death, heart failure (HF), myocardial infarction (MI) and stroke. Association between Tagging GCS and the endpoints was assessed with Kaplan Meier survival plots. Each endpoint was then modelled using Cox proportional hazards regression including 1) clinical features (age, sex, ethnicity, BMI, diabetes, hypertension, coronary artery disease, alcohol use, smoking) 2) clinical features + LV ejection fraction (LVEF) and end-diastolic volume 3) clinical features + imaging markers + strain. The significance threshold was adjusted using Bonferroni correction.

Results

Of 43,014 participants with tagging MRI, 38,342 had analysable GCS. After excluding cases with missing entries, 33,258 participants (aged 64±8 years, BMI 26±4, 53% female, LVEF 61±6%) were included in survival analysis. During follow-up, 566 deaths, 285 HF, 405 MI, and 294 stroke incidents occurred. FT GLS was -18±2%, FT GCS was -19±2% and Tagging GCS was -18±2%.

In univariate analysis (Figure 1), Tagging GCS was associated with all four endpoints. After adjusting for covariates (Table 1), Tagging GCS remained an independent predictor of all-cause death, HF and stroke, but not MI. FT GLS remained an independent predictor of all outcomes, while FT GCS remained an independent predictor of HF only. All significant predictors improved model fit compared to the baseline with covariates only.

Conclusions

We demonstrate that an automated tagging-derived CMR biomarker provides independent incremental prognostic value for all-cause death and HF over clinical features and imaging markers. In contrast, FT GCS which is more widely used, did not have this prognostic significance.