CIGNN: A Framework for Cuffless Continuous Blood Pressure Estimation Based on Causality and Graph Neural Networks

Background Introduction

According to data from the World Health Organization (WHO), approximately 1.13 billion people globally are affected by hypertension, and this number is expected to increase to 1.5 billion by 2025. Hypertension is a significant risk factor for cardiovascular diseases, including heart disease and stroke, which are leading causes of death worldwide. The prevalence of hypertension further increases the burden of dementia and disability. Therefore, preventing and managing hypertension is crucial for improving global health outcomes.

Continuous blood pressure (BP) measurement can provide rich information for diagnosing and preventing hypertension. By continuously monitoring blood pressure, we can have a more comprehensive understanding of a patient’s blood pressure patterns and trends, which can indicate whether treatment or adjustment to the current treatment is necessary. Additionally, continuous BP monitoring has more advantages over traditional intermittent BP measurement since BP can be influenced by factors such as stress, physical activity, and medication compliance. Cuffless continuous BP measurement takes advantage of wearable physiological sensors, enabling non-invasive, convenient, and continuous monitoring. Therefore, cuffless BP estimation models have attracted widespread attention.

Existing cuffless BP estimation models are mainly divided into knowledge-driven and data-driven methods. Knowledge-driven models rely on expert knowledge in the cardiovascular system field, such as physiological models based on Pulse Transit Time (PTT) and Moens-Korteweg (M-K) equation. However, these models typically stand on certain assumptions that might not hold in reality. Another class, data-driven models, maps the relationship between BP-related information and BP through learning from data, but these models often depend on large amounts of high-quality data and might be influenced by motion artifacts and environmental noise.

Most existing studies focusing on either knowledge-driven or data-driven methods overlook the causal relationships between wearable features and BP changes. Causal relationships help improve the interpretability, robustness, and generalizability of cuffless BP estimation methods and can help identify potential physiological mechanisms.

Research Source

This paper was written by Lei Liu, Huiqi Lu, Maxine Whelan, Yifan Chen, and Xiaorong Ding. The research was partly supported by the Sichuan Science and Technology Program (2021YFH0179) and the National Natural Science Foundation of China (82102178). Huiqi Lu also received funding from the Daphne Jackson Trust Fellowship Grant of the Royal Academy of Engineering, the EPSRC Healthcare Technologies Challenge Award, and the Wellcome Trust. The paper was published in the IEEE Journal of Biomedical and Health Informatics in May 2024.

Research Process Introduction

This paper proposes a two-stage framework, CIGNN, that seamlessly combines causality with Graph Neural Networks (GNN) for cuffless continuous BP estimation.

Stage One: Causal Graph Generation

In the first stage, researchers use a causal inference perspective to generate a causal graph containing BP and wearable features to identify features causally related to BP changes. The innovation of this stage lies in surpassing the traditional Pulse Transit Time (PTT), identifying other causal features more closely associated with BP changes.

Specifically, the paper employs the Fast Causal Inference (FCI) algorithm to generate the initial causal graph and the Causal Generative Neural Networks (CGNN) algorithm to direct and refine the initial causal graph. Through these causal inference algorithms, the proposed causal graph can reveal the causal relationships between BP and the wearable features.

Stage Two: Spatio-Temporal Graph Neural Network

In the second stage, researchers utilize the causal graph obtained from the first stage, using a Spatio-Temporal Graph Neural Network (STGNN) model to learn spatial information in the causal graph and temporal information through heart signals for precise cuffless continuous BP estimation.

The researchers adopted three datasets, encompassing 305 subjects (including 102 hypertensive patients), with participants of different age groups and BP levels. The study results indicate that the mean absolute difference (MAD) for Systolic Blood Pressure (SBP) and Diastolic Blood Pressure (DBP) estimates using the CIGNN method is 3.77 mmHg and 2.52 mmHg, respectively, outperforming other comparative methods.

Research Results

1. Causal Graph Generation

   • The initial causal graph is generated through the FCI algorithm, showing some edges with undetermined directions.
   • The CGNN algorithm further refines the initial causal graph, clarifies the directions of all edges, and eliminates some incorrect edges.

2. Causal Feature Analysis

   • In the causal graph, AA and PTT features exhibited stronger influence than mere PTT, validating the rationality of the causal graph.

3. Blood Pressure Estimation Performance Evaluation

   • The CIGNN method performs superiorly on multiple evaluation metrics, with significantly lower estimation errors than other benchmark methods.
   • The estimation for different age groups and hypertensive patients also demonstrated high robustness and accuracy.

Conclusion and Value

This study significantly improves the accuracy of cuffless continuous BP estimation by introducing causal relationships. Specific contributions include:

1. Scientific Value

   • By revealing the causal relationships between BP and wearable features, the study enhances the interpretability of cuffless BP estimation methods.

2. Application Value

   • The study provides a cuffless BP estimation method that exhibits superior estimation performance under various conditions, assisting in the early diagnosis and prevention of hypertension.

3. Research Highlights

   • The study introduces a novel method combining causal inference with GNN, offering a new perspective on cuffless BP estimation.

Future research will further integrate physiological knowledge to reveal more potential causal relationships and be applicable to more clinical scenarios. This study not only advances the technology of cuffless continuous BP monitoring but also provides new tools and methods for the prevention and treatment of hypertension.