Self-supervised Anomaly Detection Pretraining Enhances Long-tail ECG Diagnosis

TL;DR

Prospective validation in real-world clinical settings revealed that the AI-driven approach enhances diagnostic efficiency, precision, and completeness by 32%, 6.7%, and 11.8% respectively, when compared to standard practices, marking a pivotal step forward in the integration of AI within clinical cardiology.

Abstract

Current computer-aided ECG diagnostic systems struggle with the underdetection of rare but critical cardiac anomalies due to the imbalanced nature of ECG datasets. This study introduces a novel approach using self-supervised anomaly detection pretraining to address this limitation. The anomaly detection model is specifically designed to detect and localize subtle deviations from normal cardiac patterns, capturing the nuanced details essential for accurate ECG interpretation. Validated on an extensive dataset of over one million ECG records from clinical practice, characterized by a long-tail distribution across 116 distinct categories, the anomaly detection-pretrained ECG diagnostic model has demonstrated a significant improvement in overall accuracy. Notably, our approach yielded a 94.7% AUROC, 92.2% sensitivity, and 92.5\% specificity for rare ECG types, significantly outperforming traditional methods and narrowing the performance gap with common ECG types. The integration of anomaly detection pretraining into ECG analysis represents a substantial contribution to the field, addressing the long-standing challenge of long-tail data distributions in clinical diagnostics. Furthermore, prospective validation in real-world clinical settings revealed that our AI-driven approach enhances diagnostic efficiency, precision, and completeness by 32%, 6.7%, and 11.8% respectively, when compared to standard practices. This advancement marks a pivotal step forward in the integration of AI within clinical cardiology, with particularly profound implications for emergency care, where rapid and accurate ECG interpretation is crucial. The contributions of this study not only push the boundaries of current ECG diagnostic capabilities but also lay the groundwork for more reliable and accessible cardiovascular care.

Figures (7)

• Figure 1: Overview the proposed long-tail ECG diagnosis framework. (a) Long-tail distribution of cardiac types across the entire dataset. The dataset is characterized by a highly imbalanced distribution, with cardiac types divided into a Common Set, Uncommon Set, and Rare Set based on their frequency of occurrence. The red box highlights the expanded view of the Rare Set, where cardiac types occur fewer than 400 times. (b) The proposed two-stage ECG diagnosis framework. The framework consists of two main steps: Step 1: Self-supervised ECG anomaly detection pretraining, where an anomaly detection model is trained to identify abnormal ECG patterns using both global and local ECG information; and Step 2: Fine-tuning the classifier based on the pre-trained anomaly detection model to provide a detailed diagnosis. This approach improves the classification performance, especially for less common cardiac conditions, by leveraging anomaly detection as pretraining.
• Figure 1: Comprehensive analysis of the novel ECG-LT dataset.a. Hierarchical architecture of cardiac types. b. Comparison of the number of cardiac types in the ECG-LT dataset to those in existing ECG databases. c. Age distribution across the training, internal validation, and external validation sets. d. Gender distribution across the training, internal validation, and external validation sets.
• Figure 2: Performance of ECG diagnosis. (a) Diagnosis performance on the tail classes. A comparison of ECG diagnosis for each type in the Rare Set. The proposed method demonstrates superior performance, especially for anomalies with fewer samples. (b) Diagnosis fairness across sex. The model maintains consistent performance across male and female subjects, demonstrating balanced accuracy between sexes. (c) Diagnosis fairness across age. Diagnosis performance is presented for different age groups in ten-year intervals. The model exhibits stable and equitable performance across all age groups. (d) Visualization of anomaly localization. The proposed method's localization (Ours) is compared with a leading baseline (BeatGAN), with ground truth marked in pink boxes. Color-coded scores (ranging from 0 to 1) indicate the likelihood of an anomaly, with red marking the most likely location of anomaly.
• Figure 2: The details of multi-scale cross-restoration framework for ECG anomaly detection.
• Figure 3: Prospective validation results from the emergency department at Ruijin Hospital, comparing diagnostic timing, completeness, and accuracy between cardiologists and an AI-assisted system. The figure presents the diagnosis time and accuracy rates for Cardiologist A (under time limit), Cardiologist B (without AI assistance), AI alone, and Cardiologist C (with AI assistance). The metrics include the correct identification of comprehensive diagnoses (key diseases and detailed signal characteristics) and the correct identification of key diseases, along with the frequency of missed key diseases. The AI-assisted diagnosis demonstrated superior performance in both speed and accuracy, significantly reducing diagnostic time compared to cardiologists, especially under time constraints.