Authors:
Rui Wang、Mengshi Qi、Yingxia Shao、Anfu Zhou、Huadong Ma
Paper:
https://arxiv.org/abs/2408.08488
Introduction
Time series analysis is a critical task in data mining, especially in medical applications such as continuous health monitoring. Blood pressure (BP) is a vital indicator of cardiovascular health, with irregularities like hypertension and hypotension being potential markers for severe conditions such as stroke or chronic kidney disease. Traditional BP measurement methods using cuffs are uncomfortable, leading to the development of cuffless BP estimation techniques using wearable devices. These devices utilize multimodal signals like bioimpedance, millimeter-wave, and photoplethysmography (PPG) to estimate BP. However, these methods often require large amounts of training data and ground truth labels, which are challenging to obtain. This paper introduces a novel Physics-Informed Temporal Network (PITN) with adversarial contrastive learning to address these challenges and improve BP estimation accuracy with limited data.
Related Works
Blood Pressure Monitoring
Blood pressure monitoring is traditionally done using cuff-based devices, which are uncomfortable for continuous use. Recent advancements have explored cuffless BP measurement methods using wearable devices that capture pulse activities and infer BP using learning-based methods. These methods, however, require significant amounts of invasive or cuff-based BP measurements as ground truth labels.
Adversarial Training
Adversarial training has been shown to improve model robustness by generating adversarial examples that align network representations with salient data characteristics. This approach has been applied in various domains, including computer vision and natural language processing, but its application in medical time series data is limited.
Contrastive Learning
Contrastive learning has achieved success in generating semantically rich embeddings without labeled data. It works by pulling instances with the same labels together in the latent space while pushing different classes apart. However, its application in time series analysis, especially in medical data, is challenging due to the unique characteristics of the data.
Preliminaries
Problem Definition
The goal of cuffless blood pressure estimation is to predict BP values (systolic BP (SBP) or diastolic BP (DBP)) from input time series signals. The neural network can be formulated as ( \hat{y} = f(x, u, \theta) ), where ( x ) is the input signal, ( u ) represents physiological features, ( \theta ) are the model parameters, and ( \hat{y} ) is the predicted BP.
Physics-Informed Neural Networks
Physics-Informed Neural Networks (PINNs) incorporate laws of physics described by partial differential equations (PDEs) as network residuals. For BP estimation, the relationship between physiological features and BP can be approximated using Taylor’s approximation, leading to a physics-based loss function that enhances model training with reduced labels.
Proposed Approach
Overview
The proposed framework comprises three key components: Physics-Informed Temporal Network (PITN), adversarial training, and contrastive learning. The PITN uses temporal blocks to extract personal physiological features, while adversarial training generates additional samples for model training. Contrastive learning captures BP dynamics by comparing clean and adversarial samples based on their true BP values.
Physics-Informed Temporal Networks
The PITN enhances traditional PINNs by incorporating a novel temporal modeling block that captures complex temporal variations from transformed 2D tensors using a multi-scale Inception block. This approach allows for precise personal cardiovascular cycle modeling.
Adversarial Training
Adversarial training generates additional samples to augment the training data, improving model robustness. The perturbation is controlled to ensure it remains within the problem domain, and the Projected Gradient Descent (PGD) method is used for generating adversarial samples.
Contrastive Learning
Contrastive learning introduces additional constraints in the latent space to regularize embeddings with similar BP values. This approach helps the model capture cardiovascular dynamics more effectively, improving BP estimation accuracy.
Training and Inference
During training, both clean and adversarial samples are used to obtain BP predictions. The overall loss is computed by summing the regression loss, physical loss, and contrastive loss. During inference, the network uses primary layer normalizations to predict BP values from test data.
Experiments
Datasets
Experiments were conducted on three public benchmark datasets: Graphene-HGCPT, Ring-CPT, and Blumio. These datasets contain various wearable signals for cuffless BP estimation, including bioimpedance, PPG, and millimeter-wave signals.
Baselines
The proposed model was compared against several baseline models, including Hybrid-LSTM, PINN, ResNet1D, iTransformer, and TimesNet. These models represent state-of-the-art approaches in cuffless BP estimation and general time series modeling.
Evaluation Protocols
Common evaluation metrics, including root-mean-square error (RMSE) and Pearson’s correlation coefficient, were used to assess model performance. Additionally, mean error (ME) and standard deviation of the error (SDE) were computed following the AAMI standard.
Results and Analysis
Results on Graphene-HGCPT Dataset
The proposed model demonstrated superior performance compared to baseline methods, achieving significant improvements in correlation and RMSE for both SBP and DBP estimation. The results highlight the effectiveness of the temporal block, adversarial training, and contrastive learning in improving BP estimation accuracy.
Results on Ring-CPT Dataset
Similar to the Graphene-HGCPT dataset, the proposed model outperformed baseline methods on the Ring-CPT dataset, demonstrating its robustness and effectiveness across different datasets.
Results on Blumio Dataset
The proposed model also showed superior performance on the Blumio dataset, which includes different modalities of cuffless signals. This demonstrates the model’s versatility and applicability to various signal types.
Statistical Analysis
Pairwise t-tests confirmed the statistical significance of the proposed model’s superior performance compared to baseline methods, further validating its effectiveness.
Qualitative Results Analysis
Visual comparisons of beat-to-beat SBP and DBP estimations showed that the proposed model more accurately captured personalized characteristics, yielding higher correlations and lower absolute errors compared to the PINN model.
Ablation Study
Temporal Block
The temporal block significantly improved BP prediction accuracy, demonstrating its importance in capturing personalized cardiovascular cycles.
Adversarial Training
Adversarial training enhanced model performance by generating additional samples, improving robustness and accuracy.
Contrastive Learning
Contrastive learning further improved correlation by capturing BP dynamics more effectively, demonstrating its value in personal modeling.
Parameter Sensitivity Analysis
The model’s performance was stable across different values of minimal sensitivity and physics loss weight, indicating its robustness to parameter changes.
Discussion
The proposed model meets the AAMI standard for BP estimation in most cases, with some exceptions in hypertensive subpopulations. Future work could explore integrating multiple sensors to further improve BP estimation accuracy.
Conclusion
This paper presented an adversarial contrastive learning-based Physics-Informed Temporal Network for cuffless blood pressure estimation. The proposed framework demonstrated superior performance across multiple datasets and modalities, highlighting its effectiveness and robustness. Future work will explore applying this framework to other medical applications.
