🤖 machine learning

Domain Knowledge Based Temporal-Spatial Graph Convolution Network for ECG Recognition

This paper proposes a domain knowledge-based temporal-spatial graph convolution network that incorporates key ECG landmarks (PRQST) into a double-stream directed graph to effectively model intra- and inter-cycle relationships, achieving state-of-the-art performance in multi-class ECG recognition, particularly for rare categories.

Original authors: Wenting Ma, Zhipeng Zhang, Xiaohang Yuan, Ningwei Xie, Yuxin Xie, Xiaolin Wang, Meng Guo, Xingang Chai, Zhenjie Yao

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Wenting Ma, Zhipeng Zhang, Xiaohang Yuan, Ningwei Xie, Yuxin Xie, Xiaolin Wang, Meng Guo, Xingang Chai, Zhenjie Yao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: The "Black Box" Doctor

Imagine you have a very smart robot doctor that can look at your heart rhythm (an ECG) and tell you if you are sick. This robot is great at spotting patterns, but it's a bit of a "black box." You ask it, "Why do you think I have this condition?" and it just says, "I saw a pattern in the data." It doesn't explain which part of the heartbeat was weird or why it matters.

In the real world, doctors don't just look at the squiggly line; they look for specific landmarks, like the height of a wave or the time between two beats. They use a lifetime of medical training (domain knowledge) to make a diagnosis. The problem is that standard AI models often ignore this expert knowledge, especially when trying to spot rare diseases where there isn't much data to learn from.

The Solution: A "Map" Instead of a "Photo"

The authors of this paper decided to stop treating the ECG signal like a simple photo that a computer scans from left to right. Instead, they treated it like a road map.

Think of an ECG heartbeat as a journey. It has specific "checkpoints" (called P, Q, R, S, and T waves) that happen in a specific order.

The Landmarks: Just like a map has cities and towns, the ECG has these key points.
The Roads: The lines connecting them represent the time and distance between these points.

The researchers built a new type of AI called a Domain Knowledge-Based Graph Convolution Network. In simple terms, they taught the AI to stop looking at the whole messy line and start looking at the connections between the checkpoints, just like a human doctor does.

How It Works: The Double-Stream Detective

The AI uses a "Double-Stream" approach, which is like having two detectives working on the same case at the same time:

Detective A (The Spatial Stream): This detective looks at one single heartbeat. They check the map of that one trip. Is the distance between the "P" town and the "R" town too long? Is the "T" town too high? This helps spot problems happening inside a single beat.
Detective B (The Temporal Stream): This detective looks at a series of heartbeats. They compare the map of today's trip to yesterday's trip. Did the "R" town move closer to the "P" town compared to the last beat? Is the rhythm speeding up or slowing down? This helps spot problems that happen over time.

By combining the notes from both detectives, the AI gets a complete picture of what's wrong, not just a snapshot.

The Results: Better at Finding the "Needles in the Haystack"

The team tested this new AI on a massive dataset of heart recordings from a competition in China. They asked the AI to sort heartbeats into nine different categories, including some very common ones and some very rare ones.

The Score: The new AI achieved an average score of 88.1%, beating the previous best models.
The Rare Cases: This is where the "Domain Knowledge" really shined. When dealing with rare diseases (where there are very few examples to learn from), standard AI usually struggles. But because this AI was taught the "rules of the road" (the medical landmarks), it was much better at spotting these rare issues. It improved the detection of rare categories by a significant margin compared to models that didn't use this expert knowledge.

Why It Matters (According to the Paper)

The paper claims that by forcing the AI to look at the specific "landmarks" (P, Q, R, S, T) and the relationships between them, the model becomes:

More Accurate: It gets the diagnosis right more often.
More Understandable: Because it's looking at specific points doctors care about, it's easier to trust its decisions.
Better with Little Data: It doesn't need millions of examples to learn a rare disease because it already knows the "rules" of how a heart should look.

In a nutshell: Instead of letting an AI guess based on a blurry picture, the researchers gave it a detailed map with specific landmarks and two experts to check the map from different angles. This made the AI a much better heart doctor, especially for the tricky, rare cases.

Technical Summary: Domain Knowledge Based Temporal-spatial Graph Convolution Network for ECG Recognition

Problem Statement

The application of Artificial Intelligence (AI) in healthcare, specifically for Electrocardiogram (ECG) recognition, faces significant challenges regarding model interpretability and performance on rare categories. While Deep Learning (DL) models, particularly end-to-end Convolutional Neural Networks (CNNs), have achieved high performance, they often rely on massive, well-annotated datasets (e.g., ImageNet) which are scarce in medical domains. In healthcare, rare cardiac abnormalities often lack sufficient data, leading to poor model generalization. Furthermore, standard deep learning models function as "black boxes," lacking the integration of expert domain knowledge—such as the specific anatomical landmarks (P, Q, R, S, T waves) and their physiological relationships—that cardiologists use for diagnosis. The core problem addressed is how to effectively integrate this specific domain knowledge into a deep learning architecture to improve diagnostic accuracy, particularly for rare abnormalities, while enhancing model interpretability.

Methodology

The authors propose a Domain Knowledge-based Graph Convolution Network that models ECG signals as a Double-Stream Directed Graph. The methodology proceeds through the following stages:

1. ECG Preprocessing and Key Point Extraction

Signal Processing: Raw ECG signals undergo analog band-pass filtering (50 Hz cutoff) and are sampled at 500 Hz. Digital filters (cascaded low-pass and high-pass) are applied to remove noise sources like muscle artifacts, baseline drift, and power-line interference. Nonlinear transformations (differentiation or square root) are used to enhance QRS complexes.
Landmark Detection: Critical domain knowledge is encoded by extracting key landmarks (P, Q, R, S, T points) using a dyadic wavelet transform. This identifies significant fluctuations by locating local maxima in the wavelet transform modulus across various scales.

2. Graph Construction

The extracted key points are structured into a Directed Acyclic Graph (DAG) representing the ECG signal:

Vertices ( $V$ ): Represent the key points (P, Q, R, S, T) within each cardiac cycle.
Edges ( $E$ ): Represent directed connections.
- Intra-cycle (Spatial): Connections follow the physiological sequence within a single cycle (e.g., $P_t \to Q_t \to R_t \to S_t \to T_t$ ).
- Inter-cycle (Temporal): Connections link corresponding points across adjacent cycles (e.g., $P_t \to P_{t+1}$ ), capturing rhythm and long-term dependencies.
Attributes: The graph includes vertex attributes (key point coordinates) and edge attributes (displacement vectors between adjacent points).

3. Double-Stream Directed Graph Neural Network (DGNN)

The model employs a dual-stream architecture to process the graph structure:

Spatial Stream: Utilizes Spatial Directed Graphs (SDGs) to model the morphological relationships and positional dependencies of key points within a single cycle.
Temporal Stream: Utilizes Temporal Directed Graphs (TDGs) to model the temporal dependencies and rhythm changes between adjacent cycles.
Architecture: Each stream consists of a Directed Graph Network (DGN) block and Residual GTN (Graph Transformer Network) blocks. The DGN updates vertex and edge attributes based on adjacent nodes, while the residual blocks facilitate deeper feature aggregation.
Fusion: The encoded spatial and temporal features (vertex and edge attributes) are concatenated along the channel dimension.
Classification: The fused features undergo average pooling and are passed through a Fully Connected (FC) layer and a Softmax function to predict one of nine ECG categories (including Normal and eight abnormal types).

Key Contributions

Domain Knowledge Integration: The paper constructs a specific graph structure for ECGs that explicitly incorporates expert domain knowledge by representing PRQST landmarks and their physiological connections, rather than relying solely on raw signal processing.
Double-Stream Directed Graph Model: A novel architecture is proposed that simultaneously models intra-cycle (morphological) and inter-cycle (rhythmic) information using Spatial Directed Graphs (SDGs) and Temporal Directed Graphs (TDGs).
Performance on Rare Categories: The method is validated on a real-world dataset, demonstrating particular efficacy in detecting rare abnormal categories where data is scarce, outperforming state-of-the-art models that do not integrate domain knowledge.

Experimental Results

The proposed model was evaluated on the First Chinese ECG Intelligent Competition dataset, comprising 6,500 ECG recordings (12-lead, 500 Hz) with 5-fold cross-validation. The dataset includes nine categories (Normal + 8 abnormalities).

Overall Performance: The proposed method achieved an average F1 score of 88.1%, outperforming three state-of-the-art baselines (Method1: 87.9%, Method2: 87.6%, Method3: 87.5%).
Rare Category Performance: The model showed significant improvement in detecting rare abnormalities. Specifically, for the combined categories of Left Anterior Fascicular Block (LAFB) and Early Repolarization (ER)—which have the smallest sample sizes—the model achieved an average F1 score of 76.3%. This is notably higher than the baselines (ranging from 72.2% to 74.2%), suggesting that domain knowledge mitigates the performance degradation typically caused by data scarcity.
Specific Class Results: The model achieved high F1 scores for Normal (0.921), Atrial Fibrillation (0.958), and Premature Ventricular Contractions (0.965).
Interpretability: The paper notes that the model's reliance on key points allows for more straightforward explanations of decisions. Misclassifications were primarily attributed to excessive noise leading to inaccurate key point detection, rather than a failure of the model architecture itself.

Significance and Claims

The paper claims that integrating domain knowledge into deep learning architectures is a viable strategy to overcome the limitations of purely data-driven models in specialized medical domains. By explicitly modeling the topological relationships of ECG landmarks, the proposed method enhances diagnostic performance, particularly for rare categories where training data is insufficient.

The authors assert that this approach not only improves accuracy but also increases model interpretability, as the decision-making process is grounded in clinically relevant features (waveform morphology and rhythm) rather than abstract latent features. While the primary validation is on ECG, the authors suggest the methodology has potential applicability to other cyclic signals in healthcare (e.g., PPG) and other fields (e.g., communications, meteorology), though they maintain a modest stance, noting that further investigation is required to construct more intricate graph connections for broader applications.