Abstract
Background
Late Gadolinium Enhancement (LGE) imaging is a reference standard technique for the differentiation of ischemic cardiomyopathy (ICM) from non-ischemic dilated cardiomyopathy (NIDCM) in patients with heart failure and reduced ejection fraction (HFrEF). 3D myocardial deformation analysis (3D-MDA) offers highly reproducible phenotypic assessments of regional architecture and function that may provide value for artificial-intelligence-assisted cardiomyopathy diagnosis without need for LGE imaging.
Purpose
In this study, we trained and validated a machine-learning-based model to enable automated diagnosis of ischemic versus non-ischemic dilated cardiomyopathy exclusively using regional patterns of deformation among patients otherwise matched by age, sex and global contractile dysfunction.
Methods
100 ICM and 100 NIDCM patients matched for age, sex, and LVEF underwent standard cine SSFP and LGE imaging. Patient diagnoses were established using a combination of clinical and LGE-based criteria. 3D-MDA was performed using validated software (GIUSEPPE) to compute regional 3D strain measures at each cardiac phase in both conventional and principal strain directions. Principal Component Analysis (PCA) was performed on the composite 3D-MDA dataset. The first 20 components were chosen, accounting for approximately 65% of the population variance. Subsequently, a support-vector-machine-based algorithm was used with 10-fold cross-validation to discriminate ICM from NIDCM.
Results
Patients were 63±10 years (ICM: 63±10 years, NIDCM: 63±10 years, p=0.955), 74% male (ICM: 74%, NIDCM: 74%, p=1.000), and had a mean LVEF of 27±8% (ICM: 27±7%, NIDCM: 28±7%, p=0.688). Global time to peak strain was significantly shorter in ICM patients relative to NIDCM patients across all surfaces and in all directions (p<0.05). The highest single-variable Area Under the Curve (AUC) achieved for the classification of ICM versus NIDCM from global data was for minimum principal strain (ICM: 43.7±7.8, NIDCM: 48.3±7.5, p<0.001, AUC: 0.682) (Figure 1). However, a multi-feature machine-learning-based model exposed to all available regional 3D deformation data achieved an AUC of 0.903 (sensitivity 87.7%, specificity 75.5%).
Conclusions
Machine learning-based analyses of3D regionaldeformation patterns allows for robust discrimination of ICM versus NIDCM. Further expansion of the presented findings is planned on a wider, multi-centre cohort.
Acknowledgement/Funding
Dr. White was supported by an award from Heart and Stroke Foundation of Alberta. This study was funded in part by Calgary Health Trust.
Automated diagnosis of heart valve degradation using novelty detection algorithms and machine learning
