Development of Computational Pipeline for Right Ventricular Hemodynamic Single-Beat Analysis
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Load-independent indices of right ventricular (RV) dysfunction aid in the prognosis of patients with pulmonary hypertension (PH), but their acquisition remains difficult. This study aimed to develop a novel computer vision artificial intelligence-based pipeline that can estimate load-independent RV functional indices using screenshots of the RV pressure-time waveform from a standard clinical right heart catheterization (RHC).
Methods
Prospectively collected clinical data and research-grade pressure-volume-time data were collected from 76 patients from three centers. Patients were referred for RHC for known or suspected PH. Thirty-nine patients from one center were used for internal development, with external validation performed on the remaining 37 patients from two independent centers. A MATLAB-based computational pipeline was developed to predict RV pressure-volume (P-V) loop and extract load-independent RV indices using image processing and single-beat analysis. Agreement with gold-standard single-beat analysis was assessed via Bland-Altman analysis, concordance correlation coefficient (CCC), and Pearson correlation coefficient. Kaplan-Meier survival analysis, Cox regression analysis, and K-means clustering were performed to evaluate prognostic value.
Results
The average age of the derivation cohort was 57±12 years. Strong concordance was observed between the novel and gold-standard methods for end-systolic elastance (Ees, R=0.96, CCC=0.58), effective arterial elastance (Ea, R=0.97, CCC=0.88), end-diastolic elastance (Eed, R=0.87, CCC=0.47), and Ees/Ea ratio (R=0.93, CCC=0.71). Agreement was validated with the external cohort. Prognostic analyses showed that pipeline-derived Ea (HR: 2.09 [1.04, 4.20]) and Ees/Ea (HR: 0.27 [0.08, 0.87]) were significant predictors of clinical outcomes. Cluster analysis identified two RV sub-phenotypes with distinct hemodynamic features, with the group exhibiting higher Ea (1.05±0.27 vs. 0.41±0.19 mmHg/mL, p<0.0001) and lower Ees/Ea (0.37±0.15 vs. 0.76±0.40 mmHg/mL, p=0.0002) demonstrating worse outcomes.
Conclusion
We have developed a novel computational pipeline tool that digitizes and generates single-beat estimates of RV-pulmonary arterial coupling from an image of the RV pressure waveform. Its output correlates with single-beat methods and predicts clinical outcomes.
