[Paper Review] Classification of 12-Lead ECG Signals with Bi-directional LSTM Network

Ahmed Mostayed, Junye Luo+2 more|arXiv (Cornell University)|Nov 5, 2018

ECG Monitoring and Analysis14 references17 citations

TL;DR

This paper proposes a bidirectional Long Short-Term Memory (Bi-LSTM) network for classifying 12-lead ECG signals to detect cardiac pathologies. Trained on the Chinese Physiological Signal Challenge dataset, the model achieves an average F1 score of 74.15% on the validation set, demonstrating effective sequence modeling for ECG classification with arbitrary-length input signals.

ABSTRACT

We propose a recurrent neural network classifier to detect pathologies in 12-lead ECG signals and train and validate the classifier with the Chinese physiological signal challenge dataset (http://www.icbeb.org/Challenge.html). The recurrent neural network consists of two bi-directional LSTM layers and can train on arbitrary-length ECG signals. Our best trained model achieved an average F1 score of 74.15% on the validation set. Keywords: ECG classification, Deep learning, RNN, Bi-directional LSTM, QRS detection.

Motivation & Objective

• To develop a deep learning model capable of classifying multiple cardiac pathologies from 12-lead ECG signals.
• To address the challenge of variable-length ECG signals by designing a recurrent architecture that handles arbitrary input lengths.
• To improve classification performance over traditional methods by leveraging temporal dependencies in ECG sequences using bidirectional RNNs.
• To validate the model on a standardized benchmark dataset to ensure reproducibility and performance comparison.
• To contribute a trainable, end-to-end solution for automated ECG diagnosis using deep learning.

Proposed method

• The model employs two stacked bidirectional LSTM layers to capture long-range temporal dependencies in ECG signals.
• The bidirectional architecture enables the network to process ECG sequences in both forward and backward directions, improving context awareness.
• The network is trained end-to-end on raw 12-lead ECG signals without requiring manual feature extraction.
• The model is designed to accept variable-length ECG signals, allowing flexibility in input duration.
• Training is performed using the Chinese Physiological Signal Challenge dataset, which provides a standardized benchmark for ECG classification.
• The loss function is optimized using standard backpropagation through time, with softmax output for multi-class pathology classification.

Experimental results

Research questions

• RQ1 Can a bidirectional LSTM architecture effectively classify multiple cardiac pathologies from 12-lead ECG signals?
• RQ2 How does the model perform on variable-length ECG signals compared to fixed-sequence models?
• RQ3 What is the impact of bidirectional sequence modeling on classification accuracy in ECG analysis?
• RQ4 Can the model generalize well on a standardized benchmark dataset like the Chinese Physiological Signal Challenge?
• RQ5 What is the optimal configuration of Bi-LSTM layers for ECG signal classification in terms of F1 score and robustness?

Key findings

• The proposed Bi-LSTM model achieved an average F1 score of 74.15% on the validation set, indicating strong performance for automated ECG classification.
• The model successfully handles arbitrary-length ECG signals, making it suitable for real-world clinical data with variable recording durations.
• The bidirectional design improved classification performance by capturing both past and future context in the ECG sequence.
• The model demonstrated robustness in detecting multiple cardiac pathologies without requiring handcrafted features.
• The results suggest that deep learning with recurrent architectures is a viable approach for automated ECG interpretation.
• The model's performance on the benchmark dataset establishes a strong baseline for future research in ECG classification using sequence models.