Artificial intelligence improves detection and classification of pulmonary venous hypertension related to left ventricular diastolic dysfunction by chest radiography.
Authors
Affiliations (7)
Affiliations (7)
- Division of Augmented Intelligence in Imaging, Mayo Clinic Florida, Jacksonville, FL, USA. [email protected].
- Division of Cardiothoracic Imaging, Department of Radiology, Mayo Clinic Florida, Jacksonville, FL, USA. [email protected].
- Department of Radiology, Mayo Clinic Florida, 4456 San Pablo Road, Jacksonville, FL, 32224, USA. [email protected].
- Division of Augmented Intelligence in Imaging, Mayo Clinic Florida, Jacksonville, FL, USA.
- Division of Cardiothoracic Imaging, Department of Radiology, Mayo Clinic Florida, Jacksonville, FL, USA.
- Department of Cardiovascular Medicine, Mayo Clinic Arizona, Phoenix, AZ, USA.
- Division of Pulmonary Medicine, Department of Medicine, Mayo Clinic Florida, Jacksonville, FL, USA.
Abstract
Isolated-Left Ventricular Diastolic Dysfunction [LVDD] ranges (and may progress) from preclinical asymptomatic, symptomatic-LVDD, to LVDD-predominate Heart Failure [HF] presentations; if recognized early, LVDD progression might be preventable. Current early-HF screening remains limited, providing opportunities for insights from a standard Chest X-Ray [CXR]. While CXR assessment for "pulmonary congestion" supports suspected-HF evaluation in evidence-based guidelines, the potential for systematic Pulmonary Venous Hypertension [PVH]-Staging to contribute to initial detection and scaling of LVDD is unclear. This study compared CXR-based PVH-Staging to Doppler Echocardiography [DEcho]-based LVDD-Grading in the absence of systolic dysfunction. Questions included: (1) With PVH-Staging performed by cardiothoracic radiologists, what intra-/inter-reader variabilities remain? (2) Does PVH-Staging track LVDD-Grading? and (3) Can AI-assisted PVH prediction of LVDD-Grade match human performance? CXR examinations of 1,682 (including 750 asymptomatic/healthy) subjects, without: (1) Anatomical/physiological confounders of DEcho or CXR examinations (≤ 24 h apart), and (2) AI model-training confounders, were independently assigned 1 of 11 (9 PVH-related) Pulmonary Vasculature Patterns [PVPs] by 4 cardiothoracic radiologists and repeated for reliability evaluation. Expert-consensus Human Ground Truth [HGT] PVH PVPs were correlated with LVDD Grades (0 to 3-4), as were PVH-Rank predictions by a transformer-based AI model ["PVPI"]. Despite experience-dependent intra-/inter-reader reliability in PVP assignment, there was significant (p < 0.001) overall consistency. With increasing HGT PVH Stage, a significant (p < 0.001) trend towards increasing LVDD Grade was found; while PVH-Staging achieved confidence backing Grade 0/No LVDD, confident LVDD Grade recognition was not achieved until Grades 3-4/Restrictive Filling. However, a significantly (p < 0.001) stronger incrementally positive trend in PVPI PVH-Ranking with LVDD-Grading was demonstrated. Although validated, PVH-Staging for LVDD-Grading is limited by reader variabilities. AI-assisted PVH-Ranking may facilitate earlier and widespread objective CXR screening for LVDD which is ubiquitous in HF.