Eur Heart J Digit Health. 2026;7(3):ztag028.

Deep learning-based multi-view echocardiographic framework for comprehensive diagnosis of pericardial disease.

, , , , , , , , ,

제1저자 정시현, 문인태 · 교신저자 윤연이

PMID 41799660 PMC12966503 DOI 10.1093/ehjdh/ztag028 IF 5.3 순환기내과

소속 (6)

Department of Internal Medicine, Graduate School of Medical Science, Brain Korea 21 Project, Yonsei University College of Medicine, Seoul, Republic of Korea.
Division of Cardiology, Department of Internal Medicine, Uijeongbu Eulji Medical Center, Eulji University College of Medicine, Uijeongbu, Republic of Korea.
Ontact Health Inc., Seoul, Republic of Korea.
Division of Cardiology, Severance Cardiovascular Hospital, Yonsei University College of Medicine, Yonsei University Health System, Seoul, Republic of Korea.
Cardiovascular Center and Division of Cardiology, Department of Internal Medicine, Seoul National University Bundang Hospital, Seongnam, Gyeonggi, Republic of Korea.
Department of Internal Medicine, Seoul National University College of Medicine, Seoul, Republic of Korea.

Abstract

AIMS Pericardial disease spans a wide spectrum from small effusions to life-threatening tamponade or constriction. Transthoracic echocardiography (TTE) is the main diagnostic tool, but its interpretation is limited by operator dependence and incomplete functional assessment. Existing deep learning (DL) models focus mainly on effusion detection, lacking broader evaluation.

METHODS AND RESULTS We developed a DL-based framework that performs sequential assessment of pericardial disease: (i) morphological features, including effusion amount (normal/small/moderate/large) and pericardial thickening/adhesion (yes/no), from five B-mode views, and (ii) haemodynamic significance (yes/no), incorporating Doppler and inferior vena cava measurements. The developmental dataset comprises 2253 TTEs from multiple Korean institutions (225 for internal testing), and the independent external test set consists of 274 TTEs. In the internal test set, diagnostic accuracy was 81.8-97.3% for effusion, 91.6% for thickening/adhesion, and 86.2% for haemodynamic significance. External test set accuracy was 80.3-94.2%, 94.5%, and 85.5%, respectively. Area under the receiver operating curves for the three tasks were 0.92-0.99, 0.90, and 0.79 internally, and 0.95-0.98, 0.85, and 0.76 externally. Sensitivity for thickening/adhesion and haemodynamic significance improved from 66.7% to 77.3%, and 68.8% to 80.8%, respectively, when poor image quality were excluded. Similar performance gains were observed in subgroups with complete target views and a higher number of available video clips.

CONCLUSION This study presents the first DL-based TTE model for broader pericardial disease evaluation, integrating morphological with supportive functional assessments. The proposed framework demonstrated strong generalizability and aligned with the real-world diagnostic workflow. However, caution is warranted when interpreting results under suboptimal imaging conditions.

Keywords

Artificial IntelligenceDeep learningEchocardiographyHaemodynamic assessmentPericardial disease

Figures

**Figure 1** Two-stage deep learning framework for pericardial disease assessment. The framework sequentially evaluates morphological and haemodynamic features, reflecting the clinical diagnostic workflow. (A) Stage 1: Morphologic Feature Classifier—Multi-view B-mode echocardiographic videos (PLAX, PSAX, A4C, Modified A4C, and S4C) are processed by a shared R(2 + 1)D encoder. Per-view spatiotemporal features are fused and used for multi-task classification of pericardial effusion and thickening/adhesion. An uncertainty-weighted loss accounts for task imbalance. (B) Stage 2: Haemodynamic Significance Classifier—Functional features from Doppler and IVC analysis are encoded and combined with morphological embeddings to predict haemodynamic significance.

**Figure 2** Confusion matrices for internal and external test sets across three diagnostic tasks: (A) pericardial effusion classification, (B) pericardial thickening or adhesion detection, and (C) haemodynamic significance assessment.

**Figure 3** Receiver operating characteristics curves for internal and external test sets across three diagnostic tasks: (A) pericardial effusion classification, (B) pericardial thickening or adhesion detection, and (C) haemodynamic significance assessment.

**Figure 4** Diagnostic performance of the DL-based framework in the good/fair IQ subgroups of the internal test set. Confusion matrices (left) and Receiver Operating Characteristics Curves (right) are presented for the three diagnostic tasks: (A) Pericardial Effusion Classification, (B) Pericardial Thickening or Adhesion Detection, and (C) Haemodynamic Significance Assessment.

**Figure 5** GRAD-CAM visualization of selected cases from internal and external test sets across two diagnostic stages: (A) stage 1: pericardial effusion and thickening or adhesion, and (B) stage 2: haemodynamic significance.

그림 출처: PubMed Central 공개 논문 (PMC12966503)

Conflict of interest statement

Conflict of interest: J.J., J.K., S.A.L., Y.J., and Y.E.Y. are currently affiliated with Ontact Health, Inc. Y.E.Y., and H.J.C. hold stock in Ontact Health, Inc. The other authors have no conflicts of interest to declare.

그 밖의 정보