The Novel Indices of Short-Time Heart Rate Variability for Prediction of Cardiovascular and Cerebrovascular Events

Hai Zhao, Shiliang Shao, Ting Wang, Chunhe Song, Yun Su, Xingchi Chen, Hai Zhao 0002J. Medical Imaging Health Informatics 2020 · DOI

TL;DR

This paper introduces eight novel indices derived from short-time heart rate variability (HRV) signals, leveraging Bubble Entropy (BE) and Singular Value Decomposition (SVD), to predict cardiovascular and cerebrovascular events. These new indices demonstrate improved discrimination between patients with and without events, with the SVM classifier achieving 88.31% accuracy and 90.19% specificity.

What Problem Does This Paper Address?

Cardiovascular and cerebrovascular events pose significant health risks, making their early prediction crucial for timely intervention and improved patient outcomes. While Heart Rate Variability (HRV) analysis is a well-established, non-invasive method for assessing autonomic nervous system activity and has been used for disease diagnosis, its application in predicting these critical events remains an area with room for improvement. Traditional HRV analysis methods often rely on time-domain and frequency-domain indices, which primarily capture linear characteristics of the HRV signal. However, the underlying physiological processes are complex and exhibit nonlinear behavior. Existing nonlinear methods, while valuable, often focus on diagnosis rather than prediction or utilize a combination of traditional indices. This paper aims to bridge this gap by developing novel, more sensitive HRV-based indices specifically for the prediction of cardiovascular and cerebrovascular events.

Methods

The core of this research lies in the development of eight novel instantaneous indices for HRV analysis. The methodology involves several key steps:

  1. Signal Preprocessing: Raw ECG signals are first processed to extract HRV signals, which represent the fluctuations in successive RR intervals. This involves QRS complex detection and median filtering to remove artifacts.
  2. Hilbert Transform: To capture the instantaneous characteristics of the HRV signal, the Hilbert transform is applied. This allows for the estimation of:
  3. Instantaneous Amplitude (IA): The magnitude of the signal at any given time.
  4. Instantaneous Phase (IP): The phase of the signal.
  5. Instantaneous Frequency (IF): The rate of change of the phase.
  6. Novel Index Calculation: The eight novel indices are derived by combining the original HRV signal with its instantaneous amplitude, phase, and frequency using two advanced signal processing techniques:
  7. Bubble Entropy (BE): This method, an improvement over sample and approximate entropy, measures the complexity and disorder of a time series. The indices derived are BEHRV, BEIA, BEIF, and BEIP.
  8. Singular Value Decomposition (SVD): This technique decomposes a matrix into its constituent singular values, which represent the principal components of the signal. The indices derived are SVHRV, SVIA, SVIF, and SVIP, using the maximum singular value.
  9. Statistical Analysis: A t-test is employed to assess the statistical significance of the differences between the proposed indices in the "event" (E) group and the "no-event" (N) group.
  10. Classification: To evaluate the predictive performance of the novel indices, three common machine learning classifiers are used: K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Decision Tree (DT). 10-fold cross-validation is applied to ensure robust and reliable results.

Datasets

The study utilizes data from the PhysioNet database, specifically the Smart Health for Assessing the Risk of Events via ECG (SHAREE) dataset. This dataset comprises 24-hour ECG recordings from 139 hypertensive patients who were followed up for twelve months. During this period, 17 patients experienced major cardiovascular or cerebrovascular events. To address the class imbalance (17 event patients vs. 112 non-event patients), the data from the event patients was replicated six times, creating a balanced dataset for analysis.

Key Insights and Ideas

  • Beyond Linear Analysis: The paper strongly advocates for moving beyond traditional linear HRV analysis (time and frequency domains) to capture the complex, nonlinear dynamics of cardiac activity, which are crucial for predicting critical events.
  • Synergy of BE and SVD: The combination of Bubble Entropy and Singular Value Decomposition offers a powerful approach to extract rich information from HRV signals, including their instantaneous characteristics. BE captures complexity and disorder, while SVD reveals underlying signal components.
  • Instantaneous Features Matter: The inclusion of instantaneous amplitude, phase, and frequency, alongside the raw HRV signal, significantly enhances the predictive power of the derived indices. These features provide a more dynamic and responsive view of cardiac regulation.
  • SVM as a Strong Predictor: The SVM classifier consistently demonstrates superior performance in terms of accuracy and specificity when using the proposed novel indices, highlighting its effectiveness in distinguishing between patients at risk and those not.

Key Results

  • Statistical Significance: Seven out of the eight proposed novel indices (BEHRV, BEIA, BEIF, BEIP, SVIA, SVIF, and SVIP) showed statistically significant differences between the event (E) and no-event (N) groups, indicating their discriminative power.
  • Classification Performance:
    • The SVM classifier achieved the highest accuracy (Acc) of 88.31% and specificity (Spe) of 90.19% for predicting cardiovascular and cerebrovascular events.
  • The Decision Tree (DT) classifier demonstrated the highest sensitivity (Sen) of 88.64%.
  • Comparison with Existing Methods: The proposed novel indices, particularly when used with SVM, outperformed a previous method (Ref. [20]) that analyzed 5-minute HRV signals, achieving higher accuracy (88.31% vs. 85.7%) and sensitivity (86.36% vs. 71.4%).

Limitations and Future Work

  • Class Imbalance Handling: While the paper addresses class imbalance through data replication, it's worth noting that more advanced techniques like SMOTE or ensemble methods could be explored for future work.
  • Dataset Size: The study uses a relatively small dataset (139 patients, with 17 events). Future research could benefit from larger, more diverse datasets to further validate the findings and generalize the model.
  • Generalizability: The study focuses on hypertensive patients. Investigating the applicability of these novel indices to other patient populations or different types of cardiovascular/cerebrovascular events would be a valuable next step.
  • Real-time Prediction: While the indices are derived from short-time segments, exploring their suitability for real-time prediction in clinical settings could be a future direction.

Summary updated

BibTeX
@article{zhao2020novel,
  author = {Hai Zhao and Shiliang Shao and Ting Wang and Chunhe Song and Yun Su and Xingchi Chen and Hai Zhao},
  title = {{The Novel Indices of Short-Time Heart Rate Variability for Prediction of Cardiovascular and Cerebrovascular Events}},
  journal = {J. Medical Imaging Health Informatics},
  year = {2020},
  doi = {10.1166/jmihi.2020.2931},
}
Figure