[Paper Review] A lightweight hybrid CNN-LSTM model for ECG-based arrhythmia detection

Negin Alamatsaz, Leyla s Tabatabaei+4 more|arXiv (Cornell University)|Aug 29, 2022

ECG Monitoring and Analysis5 citations

TL;DR

This paper proposes a lightweight 11-layer hybrid CNN-LSTM model for automated ECG-based arrhythmia detection, achieving 98.24% mean diagnostic accuracy across 9 classes (8 arrhythmias and normal rhythm). The model processes 500-sample ECG segments end-to-end using resampled, baseline-wander-removed signals, leveraging CNNs for spatial feature extraction and LSTM for temporal sequence modeling, with inference time of just 5.127 ms on a Raspberry Pi, enabling real-time deployment in Holter monitors.

ABSTRACT

Electrocardiogram (ECG) is the most frequent and routine diagnostic tool used for monitoring heart electrical signals and evaluating its functionality. The human heart can suffer from a variety of diseases, including cardiac arrhythmias. Arrhythmia is an irregular heart rhythm that in severe cases can lead to heart stroke and can be diagnosed via ECG recordings. Since early detection of cardiac arrhythmias is of great importance, computerized and automated classification and identification of these abnormal heart signals have received much attention for the past decades. Methods: This paper introduces a light deep learning approach for high accuracy detection of 8 different cardiac arrhythmias and normal rhythm. To leverage deep learning method, resampling and baseline wander removal techniques are applied to ECG signals. In this study, 500 sample ECG segments were used as model inputs. The rhythm classification was done by an 11-layer network in an end-to-end manner without the need for hand-crafted manual feature extraction. Results: In order to evaluate the proposed technique, ECG signals are chosen from the two physionet databases, the MIT-BIH arrhythmia database and the long-term AF database. The proposed deep learning framework based on the combination of Convolutional Neural Network(CNN) and Long Short Term Memory (LSTM) showed promising results than most of the state-of-the-art methods. The proposed method reaches the mean diagnostic accuracy of 98.24%. Conclusion: A trained model for arrhythmia classification using diverse ECG signals were successfully developed and tested. Significance: Since the present work uses a light classification technique with high diagnostic accuracy compared to other notable methods, it could successfully be implemented in holter monitor devices for arrhythmia detection.

Motivation & Objective

• To develop a lightweight deep learning model for high-accuracy, real-time detection of cardiac arrhythmias from ECG signals.
• To overcome the limitations of heavy models that hinder real-time deployment in portable devices like Holter monitors.
• To classify 9 distinct ECG rhythm types, including 8 arrhythmias and normal sinus rhythm, using a single end-to-end framework.
• To improve generalization by training and testing on a combination of two diverse ECG databases with different sampling frequencies.
• To enable efficient deployment on low-power edge devices by minimizing model size and inference time.

Proposed method

• The model uses a 1D-CNN followed by an LSTM layer to extract hierarchical spatial and temporal features from 500-sample ECG segments.
• ECG signals were preprocessed via resampling and baseline wander removal to enhance signal quality and model robustness.
• The network is trained in an end-to-end manner without manual feature engineering, directly mapping raw ECG segments to rhythm class labels.
• The architecture is designed with only 11 layers to ensure lightweight inference, suitable for resource-constrained devices.
• Model performance was evaluated on two public databases: MIT-BIH Arrhythmia and Long-term AF, with cross-database testing to assess generalization.
• Inference efficiency was validated on a Raspberry Pi, measuring model size (0.16 MB) and inference time (5.127 ms per rhythm).

Experimental results

Research questions

• RQ1 Can a lightweight hybrid CNN-LSTM model achieve high accuracy in classifying multiple ECG arrhythmia types while remaining suitable for real-time deployment?
• RQ2 How does the proposed model perform across diverse ECG databases with different sampling rates and signal characteristics?
• RQ3 To what extent does combining CNN and LSTM improve classification accuracy compared to standalone models?
• RQ4 Can the model be efficiently deployed on low-power edge devices like Raspberry Pi, given its size and inference speed?
• RQ5 How does the model’s performance compare to state-of-the-art methods in terms of accuracy, parameter count, and inference time?

Key findings

• The proposed model achieved a mean diagnostic accuracy of 98.24% across 9 ECG rhythm classes, including 8 arrhythmias and normal sinus rhythm.
• The model demonstrated excellent generalization by being trained and tested on two distinct ECG databases with different sampling frequencies.
• The inference time on a Raspberry Pi was only 5.127 ms per ECG segment, indicating strong suitability for real-time applications.
• The model size was minimized to 0.16 MB, making it highly efficient for deployment on low-power wearable devices such as Holter monitors.
• The model achieved near-perfect AUC scores across all 9 rhythm classes, as shown in the ROC curves, indicating strong discriminative capability.
• The model outperformed several state-of-the-art methods in terms of accuracy and efficiency, particularly in balancing performance and computational cost.