Machine Learning Quantification of Pulmonary Regurgitation Fraction from Echocardiography
Jennifer Cohen1,2, Son Q Duong3,4,5,6, Naveen Arivazhagan7
1Department of Pediatrics, Icahn School of Medicine at Mount Sinai, New York, NY, USA.
Pediatric Cardiology
|May 10, 2024
Summary
Machine learning accurately predicts pulmonary regurgitation fraction using echocardiography, matching clinician accuracy. Branch pulmonary artery diastolic flow reversal is a strong indicator for mild to moderate regurgitation.
Area of Science:
- Cardiology
- Medical Imaging
- Artificial Intelligence
Background:
- Pulmonary regurgitation (PR) assessment is crucial for congenital heart disease management.
- Echocardiography has limitations in quantifying PR fraction (PRF), while cardiac MRI (cMRI) is the gold standard.
- Accurate PRF quantification is essential for guiding treatment decisions.
Purpose of the Study:
- To develop and validate a machine learning (ML) algorithm for predicting cMRI-quantified PRF from echocardiographic data.
- To compare the ML model's performance against clinician accuracy and previously reported algorithms.
- To evaluate the utility of branch pulmonary artery diastolic flow reversal (BPAFR) in PRF assessment.
Main Methods:
- A retrospective analysis of 243 patients with PR undergoing both echocardiography and cMRI within 3 months.
- Development of a gradient boosted trees ML algorithm using parameters like vena contracta ratio, PR index, PR pressure half-time, and BPAFR.
- Evaluation of the ML model's regression performance (MAE) and classification accuracy at clinical thresholds (AUROC), with external validation.
Main Results:
- The ML model achieved a Mean Absolute Error (MAE) of 7.0% for PRF regression.
- For predicting PRF > mild (≥20%), the Area Under the Receiver Operating Characteristic Curve (AUROC) was 0.96, though BPAFR alone showed superior sensitivity (94%) and specificity (97%).
- The ML model's AUROC for detecting severe PR (≥40%) was 0.86, but performance decreased to 0.73 with BPAFR; overall accuracy was comparable to clinicians (69% vs. 70%).
Conclusions:
- A novel ML model for echocardiographic PRF quantification shows promise, outperforming prior studies and matching clinician accuracy.
- Branch pulmonary artery diastolic flow reversal (BPAFR) is a highly effective marker for identifying mild to moderate PR but has moderate accuracy for severe PR.
- Challenges in external validation highlight the need for robust, reproducible algorithms in cardiovascular imaging research.
