Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-modal Heart Failure Risk Estimation based on Short ECG and Sampled Long-Term HRV

Sergio González, Abel Ko-Chun Yi, Wan-Ting Hsieh, Wei-Chao Chen, Chun-Li Wang, Victor Chien-Chia Wu, Shang-Hung Chang

TL;DR

This paper proposes several multi-modal approaches that combine 30-second ECG recordings and approximate long-term Heart Rate Variability data to estimate the risk of HF hospitalization and introduces two survival models: an XGBoost model with Accelerated Failure Time (AFT) incorporating comprehensive ECG features and a ResNet model that learns from the raw ECG.

Abstract

Cardiovascular diseases, including Heart Failure (HF), remain a leading global cause of mortality, often evading early detection. In this context, accessible and effective risk assessment is indispensable. Traditional approaches rely on resource-intensive diagnostic tests, typically administered after the onset of symptoms. The widespread availability of electrocardiogram (ECG) technology and the power of Machine Learning are emerging as viable alternatives within smart healthcare. In this paper, we propose several multi-modal approaches that combine 30-second ECG recordings and approximate long-term Heart Rate Variability (HRV) data to estimate the risk of HF hospitalization. We introduce two survival models: an XGBoost model with Accelerated Failure Time (AFT) incorporating comprehensive ECG features and a ResNet model that learns from the raw ECG. We extend these with our novel long-term HRVs extracted from the combination of ultra-short-term beat-to-beat measurements taken over the day. To capture their temporal dynamics, we propose a survival model comprising ResNet and Transformer architectures (TFM-ResNet). Our experiments demonstrate high model performance for HF risk assessment with a concordance index of 0.8537 compared to 14 survival models and competitive discrimination power on various external ECG datasets. After transferability tests with Apple Watch data, our approach implemented in the myHeartScore App offers cost-effective and highly accessible HF risk assessment, contributing to its prevention and management.

Multi-modal Heart Failure Risk Estimation based on Short ECG and Sampled Long-Term HRV

TL;DR

This paper proposes several multi-modal approaches that combine 30-second ECG recordings and approximate long-term Heart Rate Variability data to estimate the risk of HF hospitalization and introduces two survival models: an XGBoost model with Accelerated Failure Time (AFT) incorporating comprehensive ECG features and a ResNet model that learns from the raw ECG.

Abstract

Cardiovascular diseases, including Heart Failure (HF), remain a leading global cause of mortality, often evading early detection. In this context, accessible and effective risk assessment is indispensable. Traditional approaches rely on resource-intensive diagnostic tests, typically administered after the onset of symptoms. The widespread availability of electrocardiogram (ECG) technology and the power of Machine Learning are emerging as viable alternatives within smart healthcare. In this paper, we propose several multi-modal approaches that combine 30-second ECG recordings and approximate long-term Heart Rate Variability (HRV) data to estimate the risk of HF hospitalization. We introduce two survival models: an XGBoost model with Accelerated Failure Time (AFT) incorporating comprehensive ECG features and a ResNet model that learns from the raw ECG. We extend these with our novel long-term HRVs extracted from the combination of ultra-short-term beat-to-beat measurements taken over the day. To capture their temporal dynamics, we propose a survival model comprising ResNet and Transformer architectures (TFM-ResNet). Our experiments demonstrate high model performance for HF risk assessment with a concordance index of 0.8537 compared to 14 survival models and competitive discrimination power on various external ECG datasets. After transferability tests with Apple Watch data, our approach implemented in the myHeartScore App offers cost-effective and highly accessible HF risk assessment, contributing to its prevention and management.
Paper Structure (14 sections, 2 equations, 13 figures, 4 tables)

This paper contains 14 sections, 2 equations, 13 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Multi-modal HF risk estimation based on short ECG and sampled long-term HRV
  4. XGBoost AFT survival model based on comprehensive ECG features
  5. ResNet survival model based on raw ECG signal
  6. Approximate long-term HRVs from sampled beat-to-beat measurements
  7. Experimental Framework, Results and Analysis
  8. Experimental framework
  9. Model performance in survival analysis settings
  10. Approximating long-term HRV: Sample size and frequency
  11. Transferability to open ECG datasets
  12. Transferability to wearable devices: Apple Watch
  13. Use Case: myHeartScore App
  14. Discussion and Conclusions

Figures (13)

  • Figure 1: Preprocessing and features extraction of a 30s ECG.
  • Figure 2: Our features-based model relies on XGBoost AFT and the extraction of relevant features from the 30s ECG recording and the beat-to-beat measurements.
  • Figure 3: ResNet architecture with 4 residual blocks. In each convolutional layer, we indicate the kernel size, number of filters, and stride size, whose default value is 1. For example, 7 Conv. 64 /3 represents a 1D convolutional layer with a kernel size equal to 7, 64 filters, and a stride size of 3.
  • Figure 4: TFM-ResNet model includes a ResNet model to extract learned features from the 30s ECG signal, and a Transformer model encodes the HRVs time series. Then, an MLP combines these two sources of information and personal information to output the survival curve.
  • Figure 5: Flowchart of the implementation of the long-term HRVs from sampled beat-to-beat measurements.
  • ...and 8 more figures