Commentary based on Chao CJ, Jeong J, Arsanjani R, et al. Echocardiography-based deep learning model to differentiate constrictive pericarditis and restrictive cardiomyopathy. JACC Cardiovasc Imaging 2023;Oct 25:[ePub ahead of print].¹

Background

• Constrictive pericarditis (CP) remains a challenging diagnosis despite implementation of multimodality imaging over the previous decade, resulting in misdiagnosis and delays in diagnosis.² Moreover, distinguishing CP from restrictive cardiomyopathy (RCM) can be challenging.
• Established echocardiographic criteria to diagnose CP include characteristic ventricular septal motion, augmented septal mitral annular motion, and restrictive mitral diastolic filling with respiratory variation (Figure 1),³ but correctly interpreting these signs requires a high level of training and expertise.
• Deep learning algorithms that derive predictions on the basis of raw imaging features have been used for disease classification, including hypertrophic cardiomyopathy and left ventricular hypertrophy but not CP because of the rarity of CP cases to be used for model training and dearth of reference-standard cases of CP to be used as "ground truth."^4,5

Figure 1: Echocardiographic Criteria for Constrictive Pericarditis That Are Present on an Apical Four-Chamber View

Study Aims and Methods

• This study assessed the ability of deep learning of transthoracic echocardiograms (TTEs) to distinguish CP from RCM in a retrospective analysis.
• Enrolled patients had a diagnosis of CP confirmed via surgery at Mayo Clinic Rochester from 2003 to 2021. Control patients with RCM were derived from those with cardiac amyloidosis (CA) confirmed via endomyocardial biopsy or advanced imaging.
• Patients were excluded if they had inadequate image quality, moderate or greater aortic/mitral regurgitation or aortic/mitral stenosis, a larger than moderate-sized pericardial effusion, a prosthetic valve, mitral or tricuspid annuloplasty, conduction delay, an intracardiac device, only contrast-enhanced images, increased respiratory effort (chronic obstructive pulmonary disease or severe obesity), atrial fibrillation/flutter, or severe right ventricular dysfunction; patients were also excluded if they were missing hepatic vein, mitral inflow, or medial e′ measurements.
• A deep learning approach using a standard apical four-chamber (4C) TTE view was used for model training because of its ability to demonstrate ventricular septal motion, mitral annular motion, and transmitral filling patterns.
• Cases were split into training, validation, and hold-out testing sets in a 60:20:20 ratio. Models were trained for binary CP and CA classification using the ResNet50 model, which outperformed other models in initial testing.
• Results were externally validated using TTEs from Taipei Veterans General Hospital from 2010 to 2022.

Study Findings

• A total of 381 patients (mean age 68.7±11.4 years, 72.8% male) were included: 184 with CP and 197 with CA.
• The artificial intelligence (AI) algorithm demonstrated excellent discrimination to distinguish CP from CA with an area under the curve of 0.97 on internal validation and 0.84 on external validation.
• Adding motion information improved model performance, particularly model sensitivity, compared with that of single-frame views.
• The model predominantly focused on septal regions when making classifications, suggesting the importance of septal motion to CP pathophysiology.

Study Implications

• Using a simple 4C TTE view, a deep learning model distinguished patients with CP from those with RCM with high accuracy, suggesting a possible role for AI in improving workflow and reducing misdiagnosis of CP.
• Further prospective validation studies are needed to demonstrate whether use of this model alters clinical decision making, improves detection of CP, and improves distinction from RCM.

References

1. Chao CJ, Jeong J, Arsanjani R, et al. Echocardiography-based deep learning model to differentiate constrictive pericarditis and restrictive cardiomyopathy. JACC Cardiovasc Imaging 2023;Oct 25:[ePub ahead of print].
2. Welch TD, Ling LH, Espinosa RE, et al. Echocardiographic diagnosis of constrictive pericarditis: Mayo Clinic criteria. Circ Cardiovasc Imaging 2014;7:526-34.
3. Oh JK, Hatle LK, Seward JB, et al. Diagnostic role of Doppler echocardiography in constrictive pericarditis. J Am Coll Cardiol 1994;23:154-62.
4. Morita SX, Kusunose K, Haga A, et al. Deep learning analysis of echocardiographic images to predict positive genotype in patients with hypertrophic cardiomyopathy. Front Cardiovasc Med 2021;8:[ePub ahead of print].
5. Li J, Chao CJ, Jeong JJ, et al. Developing an echocardiography-based, automatic deep learning framework for the differentiation of increased left ventricular wall thickness etiologies. J Imaging 2023;9:48.