Deep learning-based automation of myocardial work and pressure-strain loop morphology analysis for differentiation of healthy and diseased in a population cohort
I Yttervoll, M A Azad, P Aadahl, I Kirkeby-Garstad, H Dalen, B Grenne, A OstvikAbstract
Background
Non-invasive pressure-strain loop (PSL) analysis enables estimation of myocardial work by integrating left ventricular (LV) longitudinal strain with afterload. However, current semi-automated approaches are time-consuming and PSL morphology is typically assessed qualitatively, limiting reproducibility and broad implementation.
Purpose
To develop a fully automated pipeline for myocardial work estimation and quantitative PSL morphology analysis, evaluate agreement with expert measurements, and assess its ability to differentiate healthy individuals from patients with structural and functional cardiac disease in a population-based cohort (Fig. 1).
Methods
We developed a deep learning-based pipeline for automated myocardial motion tracking and valve event detection.1,2 Longitudinal strain was derived and estimated valve timings were combined with systolic blood pressure to estimate a continuous LV pressure waveform. The estimated waveform was integrated with strain to derive PSLs and global work index (GWI). The pipeline was then validated in an independent cohort. Agreement between the automated pipeline and expert analysis was assessed using Bland-Altman analysis in a healthy subset (n=103) from a population study matched to the distribution of age, sex, and blood pressure of the healthy population. For morphological analysis, PSL shapes were compared between healthy participants and patients with coronary artery disease (n=85), hypertension (n=71), LV hypertrophy (n=20), heart failure (n=20), and dilated or non-compaction cardiomyopathy (n=14) in non-healthy participants from the same population-based study using dynamic time warping (DTW) and Procrustes alignment.3
Results
Mean GWI was 1822 ± 350 mmHg% for Expert 1, 1798 ± 335 mmHg% for Expert 2, and 1635 ± 294 mmHg% for the automated pipeline. Interobserver comparison between experts showed a mean bias in GWI of -36 mmHg% (limits of agreement: -327 to 255). The automated pipeline demonstrated mean differences of -197 mmHg% (-500 to -106) and –154 mmHg% (-459 to 151) relative to Expert 1 and Expert 2, respectively. Automated GWI differed significantly between healthy participants and all disease groups (p < 0.001), except LV hypertrophy (p=0.204). Quantitative PSL morphology analysis demonstrated significantly greater DTW distance and Procrustes dissimilarity between healthy participants and all disease groups (p < 0.001) (Fig. 2).
Conclusion
Fully automated deep learning-based analysis of PSL morphology enabled differentiation between healthy individuals and patients with cardiac disease, including LV hypertrophy where automated GWI did not differ significantly. Automated GWI showed good agreement matched to expert interobserver variability. Thus, fully automated PSL analysis may facilitate scalable, load-integrated assessment of myocardial function with less reliance on manual measurements.Figure 1Figure 2