Interpretable Hybrid Multichannel Deep Learning for 12-Lead ECG-based Heart Disease Classification
No Thumbnail Available
Date
2025-02
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
The electrocardiogram (ECG) is a noninvasive and affordable tool that offers valuable insights
into heart activity from multiple perspectives. However, medical practitioners often
face difficulties in diagnosing underlying heart conditions from ECG signals. To address
these challenges and improve diagnostic accuracy, researchers have investigated the potential
of deep learning (DL) techniques. Nevertheless, developing a robust and interpretable
deep learning model that performs well across diverse ECG datasets remains a key research
focus.
Thus, in this PhD research, an interpretable deep learning system is designed, incorporating
preprocessing of ECG signal and post-hoc interpretability. The designed model is a
multichannel hybrid deep learning architecture consisting of 12 blocks, each combining a
one-dimensional (1D) convolutional neural network (CNN) with bidirectional long shortterm
memory (BiLSTM) networks. After the 12 blocks, the feature maps are concatenated
and further processed by an attention mechanism and a two-dimensional (2D) CNN. All
components, including the 1D CNN, BiLSTM layers, attention mechanism, and 2D CNN,
are used as feature extraction backbones. Subsequently, fully connected (FC) layers are
incorporated for classification. The model was independently trained and tested on three
distinct 12-lead ECG datasets: (1) the PTB-XL dataset, using five super-diagnostic classes,
(2) the CODE-15% dataset, encompassing six heart disease classes, and (3) the Chapman
Arrhythmia datasets, which were analyzed using two configurations: seven reduced classes
(Chapman-Reduced) and four merged classes (Chapman-Merged). The model achieved average
test accuracy rates of 89.84%, 97.82%, 98.55%, and 98.80% for these datasets, respectively.
The result indicates the model’s effectiveness across different ECG datasets.
To understand how the model reached its classification result, we applied two post-hoc interpretability
techniques: Gradient-weighted Class Activation Mapping plus (Grad-CAM++)
and SHapley Additive exPlanations (SHAP). These techniques were used to visualize influential
segments of the ECG signal, both at the instance level for specific samples and at the
test set level for assessing the overall contributions of individual ECG leads. SHAP, with its
theoretical grounding, ensures consistent feature attribution by capturing causal relationships
within the ECG data. Meanwhile, Grad-CAM++, through causal localization, identifies regions
of the ECG signals that influenced the model’s decisions. The interpretability provided from both techniques were cross-checked against heart disease manifestations in ECG signals
using established cardiology literature, ensuring alignment with clinical patterns. The
model’s performance and the output interpretation techniques demonstrate that the proposed
approach is a practicable tool for ECG-based heart disease diagnosis
Description
Keywords
Heart disease, 12-leads ECG, CNN-BiLSTM, deep learning, interpretability, Grad-CAM++, SHAP.