Energy-Efficient Real-Time Heart Monitoring on Edge-Fog-Cloud Internet-of-Medical-Things

Berken Utku Demirel, Islam Abdelsalam Bayoumy, Mohammad Abdullah Al FaruqueIEEE Internet of Things Journal 2021 · OpenReview

TL;DR

This paper introduces a novel, energy-efficient, three-layer system for continuous, real-time ECG monitoring on wearable devices. By intelligently distributing processing across edge, fog, and cloud layers using a novel CNN architecture, the system significantly reduces energy consumption and memory requirements, achieving high accuracy in detecting heart abnormalities.

What Problem Does This Paper Address?

Continuous monitoring of electrocardiogram (ECG) signals is crucial for early detection of cardiovascular diseases. However, existing solutions face a significant challenge: low-power wearable devices, the ideal platform for such monitoring, have limited energy and memory. Traditional methods often offload processing to powerful cloud servers, leading to high latency and energy drain from constant data transmission. Conversely, edge-only solutions struggle with the computational demands of real-time analysis, quickly depleting battery life. This paper aims to bridge this gap by developing a system that balances real-time processing, accuracy, and resource efficiency for wearable ECG monitoring.

Methods

The proposed methodology employs a three-layer architecture:

  1. Noise/Artifact Detection Layer (Edge): This initial layer on the wearable device filters out noisy ECG signals and artifacts. It uses a lightweight, rule-based algorithm to assess signal quality, preventing unnecessary transmission of corrupted data and thus saving energy.
  2. Normal/Abnormal Beat Classification Layer (Edge): This layer classifies individual heartbeats as either normal or abnormal. If a beat is classified as normal, only the heart rate is transmitted, further reducing data volume. Abnormal beats are passed to the next layer. This layer also dynamically controls the ECG sampling rate based on signal quality and beat classification, optimizing energy usage.
  3. Abnormal Beat Classification Layer (Fog/Cloud): This layer, running on more powerful fog or cloud resources, performs a more detailed classification of abnormal beats to identify specific cardiovascular diseases. It utilizes a novel, distributed multi-output Convolutional Neural Network (CNN) architecture.

The core innovation lies in the distributed multi-output CNN. This architecture is designed to minimize energy consumption and latency by strategically distributing its layers between the edge and fog/cloud. The first convolution block and an output block are placed on the edge device for initial classification (normal/abnormal), significantly reducing the amount of data that needs to be transmitted. The remaining convolution blocks and output blocks reside in the fog/cloud for more complex disease classification. This distribution drastically reduces the memory footprint on the edge device.

Datasets

  • MIT-BIH Arrhythmia Dataset: Used for training and evaluating the noise/artifact detection and the overall CNN classifier.
  • MIT-BIH Supraventricular Arrhythmia Database: Combined with the MIT-BIH Arrhythmia Dataset to increase the number of abnormal beats for training.
  • PhysioNet/CinC (PICC) 2011 challenge: Used for validation and comparison with other works.

Key Insights and Ideas

  • Intelligent Data Reduction is Key: The most significant energy savings come from intelligently filtering and classifying data at the edge, transmitting only essential information (heart rate for normal beats, or detailed abnormal beat data for further analysis).
  • Edge-Fog-Cloud Synergy: Distributing the CNN architecture across different computational tiers is crucial for balancing performance and resource constraints. Placing initial classification on the edge dramatically reduces communication overhead.
  • Lightweight Edge Processing: The noise/artifact detection and normal/abnormal beat classification layers are designed to be computationally inexpensive, making them suitable for resource-constrained wearable devices.
  • Dynamic Sampling Rate Control: Adapting the ECG sampling rate based on signal quality and classification outcomes is an effective strategy for further energy conservation.

Key Results

  • High Accuracy: The methodology achieves an accuracy of 99.2% on the MIT-BIH Arrhythmia dataset.
  • Significant Energy Efficiency: The proposed methodology achieves approximately 7x more energy efficiency compared to state-of-the-art works.
  • Reduced Communication Power: Communication power can be decreased by up to 1000x compared to transmitting raw ECG data.
  • Low Memory Footprint: The distributed CNN architecture reduces the memory requirement of the edge node by 87%, making it compatible with devices having a minimum RAM of 32 KB.
  • Fast Processing: The overall execution time for a heartbeat on the edge device is 36 ms with a 55 mW power consumption.

Limitations and Future Work

  • R-peak Detection Dependency: The performance of the abnormal beat detection and classification layers is heavily reliant on the accuracy of the R-peak detection. Errors in R-peak detection can lead to incorrect segmentation and HRV calculations, severely impacting classification performance.
  • Limited Scope of Abnormalities: The current work focuses on identifying a limited number of cardiac abnormalities (VEB, SVEB). Future work could explore classifying a wider range of arrhythmias to better represent the complexity of ECG interpretation.
  • Federated Learning for Distributed Architectures: The authors suggest that further research could explore federated learning approaches to enhance the distributed neural architecture across network nodes.
  • Enhanced Transmission Decision Unit: The performance of the transmission decision unit (Normal/Abnormal beat classification) could be further studied using more comprehensive ECG datasets with diverse arrhythmias.

Summary updated

Abstract

The recent developments in wearable devices and the Internet of Medical Things (IoMT) allow real-time monitoring and recording of electrocardiogram (ECG) signals. However, continuous monitoring of ECG signals is challenging in low-power wearable devices due to energy and memory constraints. Therefore, in this paper, we present a novel and energy-efficient methodology for continuously monitoring the heart for low-power wearable devices. The proposed methodology is composed of three different layers: 1) a Noise/Artifact detection layer to grade the quality of the ECG signals; 2) a Normal/Abnormal beat classification layer to detect the anomalies in the ECG signals, and 3) an Abnormal beat classification layer to detect diseases from ECG signals. Moreover, a distributed multi-output Convolutional Neural Network (CNN) architecture is used to decrease the energy consumption and latency between the edge-fog/cloud. Our methodology reaches an accuracy of 99.2% on the well-known MIT-BIH Arrhythmia dataset. Evaluation on real hardware shows that our methodology is suitable for devices having a minimum RAM of 32KB. Moreover, the proposed methodology achieves $7\times$ more energy efficiency compared to state-of-the-art works.
BibTeX
@article{demirel2021energy,
  author = {Berken Utku Demirel and Islam Abdelsalam Bayoumy and Mohammad Abdullah Al Faruque},
  title = {{Energy-Efficient Real-Time Heart Monitoring on Edge-Fog-Cloud Internet-of-Medical-Things}},
  journal = {IEEE Internet of Things Journal},
  year = {2021},
  url = {https://openreview.net/forum?id=XanLSHjm7U},
}
Figure