Classification of Tetanus Severity in Intensive-Care Settings for Low-Income Countries Using Wearable Sensing

Ping Lu, Shadi Ghiasi, Jannis Hagenah, Ho Bich Hai, Nguyen Van Hao, Phan Nguyen Quoc Khanh, Le Dinh Van Khoa, Vital Consortium, Louise Thwaites, David A. Clifton, Tingting Zhu

TL;DR

This paper introduces a novel deep learning approach using wearable ECG sensors to classify tetanus severity in low-resource settings. By transforming ECG signals into spectrograms and employing a 2D CNN with a channel-wise attention mechanism, the system aims to automatically triage patients, reducing the burden on healthcare staff and enabling timely interventions.

What Problem Does This Paper Address?

Tetanus is a severe infectious disease that can lead to life-threatening autonomic nervous system (ANS) dysfunction, particularly in low- and middle-income countries where resources are often limited. Early and accurate assessment of tetanus severity is crucial for effective patient management, including deciding on intensive care unit (ICU) admission and mechanical ventilation. Traditional methods rely on clinical scores (like the Ablett score) and require expert assessment, which can be time-consuming and demanding for already stretched healthcare professionals. While ECG analysis, particularly heart rate variability (HRV), shows promise in detecting ANS dysfunction, it often requires specialized equipment and complex preprocessing. This paper addresses the need for a low-cost, automated, and accessible system to monitor tetanus severity, leveraging wearable sensors and advanced machine learning to provide timely alerts and support clinical decision-making in resource-constrained environments.

Methods

The core of the proposed method is a 2D Convolutional Neural Network (CNN) enhanced with a channel-wise attention mechanism. Here's a breakdown:

  • Wearable ECG Data Acquisition: The study utilizes low-cost, wearable ePatch monitors to collect single-lead ECG data.
  • Data Preprocessing: Raw ECG signals undergo noise removal using a Butterworth filter.
  • Spectrogram Transformation: To leverage image-based CNN architectures, the 1D ECG time series are converted into 2D log-spectrograms. This transformation captures time-frequency information, with the logarithmic scale emphasizing lower frequencies, which are important for detecting subtle physiological changes.
  • 2D CNN Feature Extraction: A series of convolutional blocks (each with 3x3 kernels, ELU activation, and batch normalization) are used to extract spatial features from the log-spectrograms. Max-pooling layers are incorporated in some blocks to reduce dimensionality.
  • Channel-Wise Attention Mechanism: This is the key novel component. At the end of each convolutional block, a channel-wise attention module is applied. This mechanism learns to assign different weights to feature maps across channels, allowing the network to focus on the most informative features and suppress less relevant ones. This is inspired by "squeeze-and-excitation" networks but applied to the spatial features of spectrograms.
  • Classification: The extracted features are fed through fully connected layers to perform a binary classification: mild tetanus (Ablett grades 1-2) vs. severe tetanus (Ablett grades 3-4).
  • Loss Function: Binary cross-entropy (BCE) is used for training.

The authors also explored and compared their method against several baseline architectures, including vanilla 2D-CNN, 2D-CNN with sequential models like LSTM and ConvLSTM, and a 2D-CNN with dual attention.

Datasets

The study utilized a dataset collected from 110 patients at the Hospital for Tropical Diseases in Ho Chi Minh City, Vietnam.

  • Data Source: ECG data was collected using the low-cost ePatch wearable monitor.
  • Data Collection: ECG waveforms were recorded at a sampling rate of 256 Hz. Data from the first 24 hours (day 1) and the second 24 hours (day 5) of hospitalization were collected.
  • Dataset Size: The study used a total of 178 ECG example files. For experiments, the data was split into training (141 files), validation (19 files), and testing (18 files) sets.
  • Data Preprocessing for Experiments: Trimmed ECG data was used. For spectrogram generation, 60-second windows were extracted, resulting in 30 samples per ECG example file. This yielded 4230 training, 540 validation, and 570 testing log-spectrograms.
  • Labeling: Patients were classified into "mild tetanus" (Ablett grades 1-2) and "severe tetanus" (Ablett grades 3-4) based on the Ablett score.

The paper does not introduce a new benchmark dataset but rather uses a previously collected and published dataset.

Key Insights and Ideas

  • Wearable ECGs for Low-Resource Tetanus Monitoring: The paper demonstrates the feasibility of using low-cost wearable ECG sensors for automated tetanus severity classification, a significant step towards accessible healthcare solutions in resource-limited settings.
  • Spectrograms as a Powerful Representation: Transforming 1D ECG signals into 2D log-spectrograms allows the application of powerful image-based CNN architectures, effectively capturing complex time-frequency dynamics relevant to disease severity.
  • Channel-Wise Attention Enhances Feature Discrimination: The integration of channel-wise attention significantly improves the model's ability to identify and weigh the most discriminative features within the spectrograms, leading to better classification performance compared to a vanilla CNN.
  • Noise Resilience of the Approach: Despite the inherent low signal-to-noise ratio of low-cost wearable ECGs, the proposed method, particularly with the attention mechanism, shows robustness in classifying tetanus severity.
  • Limitations of Sequential Models on Noisy Data: The study found that sequential models like LSTMs and ConvLSTMs performed poorly on this low-quality ECG data, suggesting that the noise significantly hinders their ability to capture temporal dependencies effectively.

Key Results

The proposed 2D-CNN + Channel-wise Attention method achieved promising results on the test dataset:

  • Overall Performance (60s window, no downsampling):
    • F1 Score: 0.79 ± 0.03
    • Precision: 0.78 ± 0.08
    • Recall: 0.82 ± 0.05
    • Specificity: 0.85 ± 0.08
    • Accuracy: 0.84 ± 0.04
    • AUC: 0.84 ± 0.03
  • Comparison to Baselines: The proposed method outperformed the vanilla 2D-CNN and other baseline models, including those with LSTM/ConvLSTM and dual attention, across most metrics.
  • Window Length Robustness: The model demonstrated good performance with window lengths as short as 20 seconds, with the 50-second window showing slightly higher performance metrics. However, 10-second and 5-second windows were found to be too short for effective deep learning analysis.
  • Impact of Channel-Wise Attention: Visualizations (Grad-CAM) showed that the attention mechanism highlighted more critical regions in the spectrograms, particularly influencing the final decision for mild tetanus classification.

Limitations and Future Work

  • Low Signal-to-Noise Ratio (SNR): The authors acknowledge that the low SNR of wearable ECG data presents a significant challenge for noise reduction and feature extraction.
  • Dataset Size: The dataset, while valuable, is relatively small. The authors plan to extend the dataset to enable prediction of day 5 severity using only day 1 data.
  • Optimal Window Length: While 50s showed good performance, further exploration into the optimal time window for spectrogram generation is suggested.
  • Future Directions:
    • Sequence Learning with Transformers: Investigating transformer networks, potentially in combination with CNNs, to capture both local and global dependencies.
  • Knowledge Distillation: Applying knowledge distillation techniques to train smaller, more efficient models from larger, complex teacher models, which could further improve accuracy and reduce computational costs.
  • Expanding to Other Infectious Diseases: The data-driven nature of the approach suggests its potential transferability to classifying the severity of other infectious diseases.

Summary updated

Abstract

Infectious diseases remain a common problem in low- and middle-income countries, including in Vietnam. Tetanus is a severe infectious disease characterized by muscle spasms and complicated by autonomic nervous system dysfunction in severe cases. Patients require careful monitoring using electrocardiograms (ECGs) to detect deterioration and the onset of autonomic nervous system dysfunction as early as possible. Machine learning analysis of ECG has been shown of extra value in predicting tetanus severity, however any additional ECG signal analysis places a high demand on time-limited hospital staff and requires specialist equipment. Therefore, we present a novel approach to tetanus monitoring from low-cost wearable sensors combined with a deep-learning-based automatic severity detection. This approach can automatically triage tetanus patients and reduce the burden on hospital staff. In this study, we propose a two-dimensional (2D) convolutional neural network with a channel-wise attention mechanism for the binary classification of ECG signals. According to the Ablett classification of tetanus severity, we define grades 1 and 2 as mild tetanus and grades 3 and 4 as severe tetanus. The one-dimensional ECG time series signals are transformed into 2D spectrograms. The 2D attention-based network is designed to extract the features from the input spectrograms. Experiments demonstrate a promising performance for the proposed method in tetanus classification with an F1 score of 0.79 ± 0.03, precision of 0.78 ± 0.08, recall of 0.82 ± 0.05, specificity of 0.85 ± 0.08, accuracy of 0.84 ± 0.04 and AUC of 0.84 ± 0.03.
BibTeX
@inproceedings{lu2022classification,
  author = {Ping Lu and Shadi Ghiasi and Jannis Hagenah and Ho Bich Hai and Nguyen Van Hao and Phan Nguyen Quoc Khanh and Le Dinh Van Khoa and VITAL Consortium and Louise Thwaites and David A. Clifton and Tingting Zhu},
  title = {{Classification of Tetanus Severity in Intensive-Care Settings for Low-Income Countries Using Wearable Sensing}},
  booktitle = {Sensors},
  year = {2022},
  url = {https://openreview.net/forum?id=sE09qule5L},
}
Figure