좌각차단 환자군 특화 심전도 기반 좌심실수축기능 저하 예측 딥러닝 모델 = A Deep Learning Model for ECG-Based Prediction of Left Ventricular Systolic Dysfunction in Left Bundle Branch Block?Specific Patient Cohorts

다국어 초록 (Multilingual Abstract)

Biosignal-based screening using electrocardiograms (ECGs) has emerged as a practical approach for detecting latent cardiac dysfunction in real-world clinical settings. However, left bundle branch block (LBBB) introduces characteristic conduction-related morphological changes (e.g., widened QRS and altered repolarization patterns) that can substantially degrade the generalization performance of ECG-based left ventricular systolic dysfunction (LVSD) detection models trained on general populations. In this thesis, we propose an LBBB-specific deep learning framework for LVSD prediction using 12-lead ECGs, with an emphasis on robust performance under limited-data conditions and improved clinical interpretability.
First, we develop a residual-driven modeling pipeline that combines an autoencoder-based anomaly representation with a convolutional neural network (CNN) classifier. The autoencoder is trained to reconstruct non-LVSD ECG patterns, and the resulting reconstruction residuals are leveraged as informative cues to enhance separability between non-LVSD and LVSD in the LBBB cohort. Second, to fully exploit multi-lead information while maintaining model simplicity, we introduce a lead-wise ensemble strategy that aggregates predictions from single-lead CNN models, enabling stable decision-making without requiring a heavy multi-branch architecture. Finally, we apply Grad-CAM-based explainability to identify lead-dependent salient regions and support qualitative interpretation of model decisions.
Experimental evaluation on an LBBB-specific dataset demonstrates that the single-lead CNN baseline achieved an accuracy of 0.74 and an AUC of 0.69, whereas the proposed lead-wise ensemble improved performance to an accuracy of 0.81 and an AUC of 0.75. In addition, an external validation using a publicly available model trained on a general population showed marked performance degradation on the LBBB cohort (AUC 0.56), highlighting the necessity of LBBB-tailored modeling. A ResNet-based LBBB-oriented comparator exhibited high sensitivity but low specificity (AUC 0.86; specificity 0.40), suggesting that complex architectures may overfit or become sensitivity-biased under small-cohort constraints. Collectively, these results indicate that a residual-informed, lead-wise ensemble approach can provide a balanced and practically deployable solution for LVSD screening in the LBBB population.

목차 (Table of Contents)

• I. 서론
• 1. 연구 배경 및 필요성
• 2. 연구 내용 및 구성
• II. 관련 연구
• 1. 이론적 배경
• 1) 좌심실수축기능 저하
• 2) 심전도
• 3) 좌각차단
• 4) 좌심실수축기능 저하와 좌각차단 상호 연관성
• 2. 관련 연구 동향
• 1) 기존 접근
• 2) 심전도 기반 접근
• 3. 기존 연구 한계점
• III. 데이터 수집 및 전처리
• 1. 데이터셋 개요
• 2. 데이터 전처리
• 1) 대역 통과 필터링
• 2) 이상치 처리 및 정규화
• IV. 시스템 설계
• 1. 전체 시스템 구조
• 2. 오토인코더 기반 이상탐지 모델
• 3. 분류 모델
• 1) 합성곱 신경망 기반 분류 모델
• 2) Grad-CAM 분석
• 4. 리드 단위 앙상블 기법
• V. 실험 결과
• 1. 단일 리드: 합성곱 신경망 분류 모델
• 2. 멀티 리드: 리드 단위 앙상블
• 3. 성능 검증
• 1) 일반 대상군 모델 성능 검증
• 2) 좌각 차단 대상군 모델 성능 검증
• VI. 고찰
• VII. 결과
• 참고문헌