1355862 - Prediction of late gadolinium enhancement (LGE) using non-contrast CMR biomarkers

Background: Although late gadolinium enhancement (LGE) cardiovascular magnetic resonance imaging (CMR) offers important diagnostic information to assess irreversible myocardial tissue damage, it requires administering a contrast agent, adding time and complexity to the CMR exam. The purpose of this study was to identify and rank non-contrast biomarkers to train a machine learning (ML) algorithm that predicts LGE in American Heart Association (AHA) segments.

Methods:

Patients referred for a CMR exam with ischemic cardiomyopathy (ICMP) were recruited. Exams were conducted on one of three scanners: 3T Magnetom Skyra™ (26%) (Siemens Healthineers, Erlangen, Germany), 3T SIGNA Premier (62%), or 1.5T Artist (12%) (GE Healthcare, Milwaukee, USA). Short- and long-axis cine images were acquired using a balanced steady-state free precession (bSSFP) sequence. Modified look-locker inversion recovery (MOLLI) and T2 SSFP sequences were used to acquire T1 and T2 maps, respectively. All data were analyzed using certified software (cvi42, Circle Cardiovascular Imaging Inc., Calgary, AB, Canada). Left ventricular (LV) longitudinal, circumferential, and radial strain values and myocardial wall thickness were computed from cine images. T1 and T2 relaxation rates were computed from MOLLI and T2-SSFP images, respectively. The CMR and individual biomarkers (body-to-mass-index (BMI), height, weight, blood pressure, and heart rate) were inputted into the Cardiom AI system (Area19 Medical Inc.) to train an ML algorithm to classify AHA segments as having positive or negative LGE. This dataset was split by 80% for training and 20% for validation. Cross-validation was conducted using a stratified 10-fold method. Finally, a Shapley analysis was used to identify the salient biomarkers which led to LGE prediction.  



Results:

Seventy-eight participants were enrolled in this study (mean age = 61y., 66 male). The ML algorithm was able to correctly classify LGE AHA segments with a sensitivity of 0.81, specificity of 0.70 and area-under-the-curve (AUC) of 0.84 (Fig.1A). The Shapley analysis ranked the 14 biomarkers with the greatest saliency to this prediction (Fig.2B). Circumferential strain was found to be the biomarker which contributed the most to LGE prediction, with decreasing strain values implying a greater likelihood of positive LGE (Fig.2A). T1 and T2 relaxation rates, as well as T1 standard deviation, were among the top 14 contributing biomarkers (Fig.2B, 2C, 2D). T1 values greater than 1250 ms, T2 values greater than 53 ms, and T1 standard deviations greater than 60 ms predicted positive LGE. 



Conclusion:

This study demonstrates a strong potential of non-contrast CMR biomarkers to predict irreversible myocardial tissue damage on a segmental level. The most salient biomarkers needed for this prediction were also identified and included circumferential strain, longitudinal strain, and T1 and T2 relaxation rates. This suggests that a multi-parametric, contrast-free CMR approach has the diagnostic potential to replace LGE.