Hybrid Network with Attention Mechanism for Detection and Location of Myocardial Infarction Based on 12-Lead Electrocardiogram Signals

Hongying Liu, Lidan Fu, Binchun Lu, Bo Nie, Zhiyun Peng, Hongying Liu 0001, Xitian PiSensors 2020 · DOI

TL;DR

This paper introduces a novel hybrid deep learning framework, MLA-CNN-BiGRU, that leverages a multi-lead attention mechanism to improve the accuracy of detecting and locating myocardial infarction (MI) from 12-lead ECG signals. By adaptively weighting the importance of each ECG lead and combining spatial and temporal feature extraction, the model achieves state-of-the-art performance, offering a more robust and efficient approach to MI diagnosis.

What Problem Does This Paper Address?

Myocardial infarction (MI), or heart attack, is a leading cause of death worldwide. Early and accurate diagnosis is crucial for patient survival and effective treatment. Electrocardiograms (ECGs) are a primary diagnostic tool due to their non-invasiveness and low cost. However, traditional ECG analysis relies heavily on expert interpretation and handcrafted features, which can be time-consuming and prone to human error. While deep learning has shown promise, existing methods often neglect the complex inter-lead relationships and temporal dynamics present in 12-lead ECG data. This paper aims to bridge this gap by developing an end-to-end deep learning framework that can automatically learn discriminative features from all 12 leads, considering both their spatial arrangement and temporal evolution, for improved MI detection and localization.

Methods

The core of this work is the MLA-CNN-BiGRU framework, a hybrid deep learning architecture designed to process 12-lead ECG signals. It comprises three main components:

  • Multi-Lead Attention (MLA) Module: This novel module is designed to adaptively learn the importance of each of the 12 ECG leads. It uses a custom activation function, StepReLU, to assign weights, effectively highlighting crucial leads and downplaying redundant ones. This addresses the issue that not all leads contribute equally to MI diagnosis.
  • CNN with Attention Mechanism (for Spatial Features): A 2D Convolutional Neural Network (CNN) is employed to extract spatial features by considering the inter-lead relationships. The paper innovatively uses three different kernel sizes (3x3, 5x1, and 7x1) to capture correlations across different leads. An attention layer is added after the convolutional and pooling layers to further refine these spatial features by weighting their importance.
  • BiGRU with Attention Mechanism (for Temporal Features): A Bidirectional Gated Recurrent Unit (BiGRU) is used to capture the temporal dynamics within each individual lead. BiGRU processes information in both forward and backward directions, allowing it to learn from past and future contexts. Similar to the CNN module, an attention layer is integrated to select the most relevant temporal features.

These spatial and temporal features are then concatenated and fed into a fully connected layer for final classification (MI detection or location). The entire framework is trained end-to-end.

Datasets

The study utilizes the Physikalisch-Technische Bundesanstalt (PTB) dataset. This dataset contains 549 records from 290 subjects, including conventional 12-lead ECGs and 3 Frank signals, sampled at 1000 Hz. The dataset includes records from 148 MI patients (with ten different types of MI) and 52 healthy controls. For MI location, five specific MI categories were selected: Anterior MI (AMI), Antero-lateral MI (ALMI), Antero-septal MI (ASMI), Inferior MI (IMI), and Inferolateral MI (ILMI).

Key Insights and Ideas

  • Lead Importance Varies: The study confirms that not all 12 ECG leads are equally informative for MI detection and location. The MLA module effectively identifies and prioritizes leads with higher diagnostic value, such as I, II, V5, and V6 for detection, and II, III, V5, and V6 for location, while deeming others like aVF and V2 as less critical or redundant.
  • Spatial and Temporal Fusion is Key: Combining spatial features (inter-lead relationships) extracted by CNN with temporal features (intra-lead dynamics) extracted by BiGRU significantly boosts performance. This hybrid approach captures a more comprehensive picture of cardiac electrical activity.
  • Attention Enhances Feature Integration: The integration of attention mechanisms within both the CNN and BiGRU modules is crucial for refining extracted features. It helps to focus on the most discriminative spatial and temporal information, preventing information overload and improving classification accuracy.
  • End-to-End Learning Outperforms Traditional Methods: The proposed end-to-end framework, which integrates feature extraction and classification, demonstrates superior performance compared to traditional methods like PCA combined with MLP, highlighting the power of deep learning for complex signal analysis.

Key Results

The MLA-CNN-BiGRU framework achieved state-of-the-art performance on the PTB dataset:

  • MI Detection (Intra-patient): Achieved an average Accuracy (Acc) of 99.93%, Sensitivity (Sen) of 99.99%, and Specificity (Spe) of 99.63%, with a very low standard deviation.
  • MI Detection (Inter-patient): Achieved an average Acc of 96.50%, Sen of 97.10%, and Spe of 93.34%. This scheme is more challenging due to patient variability.
  • MI Location (Intra-patient): Achieved an average Acc of 99.11%, Sen of 99.02%, and Spe of 99.10%.
  • MI Location (Inter-patient): Achieved an average Acc of 62.94%, Sen of 63.97%, and Spe of 63.00%. While lower than intra-patient, this is still competitive given the difficulty.

The study also showed that the proposed framework significantly outperformed other ablation studies and comparative methods, including PCA-MLP.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Model Complexity: The proposed hybrid framework, while effective, has a complex structure and numerous parameters, which poses challenges for deployment on resource-constrained devices like mobile phones. Future work should focus on optimizing the architecture for efficiency without sacrificing accuracy.
  • Lead Selection Explanation: While the model identifies important leads, a more in-depth pathological analysis and comparative experiments could provide a clearer understanding of why certain leads are more critical for MI diagnosis.
  • Dataset Diversity: The study was conducted on the PTB dataset. To confirm the framework's robustness, it should be evaluated on more diverse datasets with varying patient populations and recording conditions.
  • MI Location Improvement: The performance for MI location, particularly in the inter-patient scheme, needs further improvement. The authors suggest this area requires more attention.

Summary updated

BibTeX
@article{liu2020hybrid,
  author = {Hongying Liu and Lidan Fu and Binchun Lu and Bo Nie and Zhiyun Peng and Hongying Liu and Xitian Pi},
  title = {{Hybrid Network with Attention Mechanism for Detection and Location of Myocardial Infarction Based on 12-Lead Electrocardiogram Signals}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20041020},
}
Figure