Automated assessment of right ventricular systolic function from coronary angiograms with video-based artificial intelligence algorithms: Development, Validation, Comparison Against Humans and Prospective Deployment

Abstract Aims Right ventricular systolic function (RVSF) is a critical determinant of cardiovascular outcomes, yet assessment during coronary angiography remains challenging without prior imaging. We developed and validated DeepRV, a deep learning model predicting RVSF from routine coronary angiograms. Methods and Results DeepRV, a video-based deep neural network, was developed using 8,053 coronary angiography studies from 6,923 patients at Montreal Heart Institute (2017-2023), with RVSF determined by echocardiography. The model was externally validated at University of California, San Francisco and prospectively deployed during primary PCI for STEMI. In the internal test set (n=1,586; 10.5% reduced RVSF), DeepRV achieved AUROC 0.80 (95% CI: 0.76-0.84), sensitivity 70.5%, specificity 78.5%, and negative predictive value 95.8%. External validation demonstrated AUROC 0.75 (95% CI: 0.72-0.77) on the UCSF dataset (n=2,247 studies; 30% reduced RVSF). Prospective deployment of DeepRV during STEMI cases at our institution (n=82) achieved AUROC 0.83 (95% CI: 0.71-0.93) using post-PCI angiogram with a median 5.1-second inference time. In a human performance evaluation (n=200), AI assistance improved accuracy of identifying RVSF for cardiologists (72.1% to 77.6%) and medical students (43.5% to 64.0%). AI alone achieved the highest accuracy (79.5%) and sensitivity (70.0%), while cardiologists with AI achieved the highest specificity (84.6%). Conclusion DeepRV enables automated RVSF assessment from routine coronary angiograms and enhances diagnostic accuracy across experience levels. Real-time inference and open-weight availability support its potential as a point-of-care tool for risk stratification during coronary angiography.