Versatile and Risk-Sensitive Cardiac Diagnosis via Graph-Based ECG Signal Representation

TL;DR

VERSATILE AND RISK-SENSITIVE CARDIAC DIAGNOSIS via Graph-Based ECG Signal Representation introduces VARS, a graph-based framework that converts heterogeneous ECG signals into a unified graph representation, enabling versatile processing across lead counts, sampling rates, and durations. The approach combines a self-attention-based graph construction, a denoising reconstruction module, and a unified subgraph contrastive learning strategy to preserve essential ECG information while highlighting pathognomonic patterns. VARS demonstrates state-of-the-art performance across MITBIH, PTB-XL, ST-T, and Chapman–Shaoxing datasets, with substantial improvements in identifying risk signals and robust interpretability by localizing the exact waveforms driving predictions. The method emphasizes clinical practicality through a lean inference path and interpretable outputs, underscoring its potential for high-throughput cardiac screening and guiding future real-world validations and clinical trials.

Abstract

Despite the rapid advancements of electrocardiogram (ECG) signal diagnosis and analysis methods through deep learning, two major hurdles still limit their clinical adoption: the lack of versatility in processing ECG signals with diverse configurations, and the inadequate detection of risk signals due to sample imbalances. Addressing these challenges, we introduce VersAtile and Risk-Sensitive cardiac diagnosis (VARS), an innovative approach that employs a graph-based representation to uniformly model heterogeneous ECG signals. VARS stands out by transforming ECG signals into versatile graph structures that capture critical diagnostic features, irrespective of signal diversity in the lead count, sampling frequency, and duration. This graph-centric formulation also enhances diagnostic sensitivity, enabling precise localization and identification of abnormal ECG patterns that often elude standard analysis methods. To facilitate representation transformation, our approach integrates denoising reconstruction with contrastive learning to preserve raw ECG information while highlighting pathognomonic patterns. We rigorously evaluate the efficacy of VARS on three distinct ECG datasets, encompassing a range of structural variations. The results demonstrate that VARS not only consistently surpasses existing state-of-the-art models across all these datasets but also exhibits substantial improvement in identifying risk signals. Additionally, VARS offers interpretability by pinpointing the exact waveforms that lead to specific model outputs, thereby assisting clinicians in making informed decisions. These findings suggest that our VARS will likely emerge as an invaluable tool for comprehensive cardiac health assessment.

Figures (5)

• Figure 1: Illustrating our VARS approach. The approach involves three parts: unified structure construction, representation learning, and downstream tasks. a. Heterogeneous ECG datasets, including multi-lead and two-lead configurations with varied sampling rates and lengths, are transformed into a unified signal structure to ensure consistency. b. A contrastive learning method with a Graph Neural Network (GNN) encoder learns a unified ECG representation, with feature subgraphs and perturbed graphs, enhancing robustness. c. These graph structures support downstream tasks like classification and interpretation, categorizing signals into normal beats, supraventricular ectopic beats, ventricular ectopic beats, and other types.
• Figure 2: An overview of our VARS framework. VARS consists of three main parts. The graph construction process transforms ECG data into graph data, using attention mechanisms to capture semantic relationships between graph nodes. Feature subgraph contrastive learning is employed to generate a unified ECG representation. Downstream tasks utilize the trained GNN encoder for classification and interpretation of graph data.
• Figure 3: Visualization of ECG interpretability. The top portion is an interpretable demonstration at the heartbeat structure level, and the bottom portion is an interpretable demonstration at the heartbeat level.
• Figure 4: Match between clinical annotations and interpreter outputs. (a) Venn diagram on the 5,000-record set showing where the interpreter’s top-1 segment overlaps the clinically annotated interval. (b) Match-rate curve as the window tolerance increases, showing a smooth rise that indicates stable agreement between the interpreter’s top-1 segment and the clinical markings.
• Figure 5: Parameter sensitivity of VARS over five hyperparameters: $\Theta$, Top-$k$, masking rate $\rho$, $\gamma$, and $\tau$. Each curve uses a five-point grid centered at our default (the third tick), with two symmetric values on each side.