PITN: Physics-Informed Temporal Networks for Cuffless Blood Pressure Estimation

TL;DR

The paper tackles cuffless blood pressure estimation under limited ground-truth data by introducing Physics-Informed Temporal Networks (PITN) that fuse physics-based Taylor residuals with a novel temporal block to capture personalized cardiovascular dynamics. It further augments data via adversarial training and enforces representation discipline through contrastive learning, integrating these components into an end-to-end training regime with a combined loss $\mathcal{L}_{\mathrm{total}} = \mathcal{L}_{\mathrm{clean}} + \mathcal{L}_{\mathrm{adv}} + \mathcal{L}_{\mathrm{con}} + \gamma \mathcal{L}_{\mathrm{physics}}$ where $\gamma=1$ and adversarial perturbations are constrained by $\epsilon=0.2$. The approach is validated on Graphene-HGCPT, Ring-CPT, and Blumio datasets across bioimpedance, mmWave, and PPG modalities, showing superior correlation and RMSE over state-of-the-art baselines and demonstrating robustness with minimal subject-specific labels. The work highlights a practical pathway to reliable cuffless BP monitoring in wearables, with potential to generalize to other medical time-series tasks by leveraging physics priors and data-efficient augmentation strategies.

Abstract

Monitoring blood pressure with non-invasive sensors has gained popularity for providing comfortable user experiences, one of which is a significant function of smart wearables. Although providing a comfortable user experience, such methods are suffering from the demand for a significant amount of realistic data to train an individual model for each subject, especially considering the invasive or obtrusive BP ground-truth measurements. To tackle this challenge, we introduce a novel physics-informed temporal network~(PITN) with adversarial contrastive learning to enable precise BP estimation with very limited data. Specifically, we first enhance the physics-informed neural network~(PINN) with the temporal block for investigating BP dynamics' multi-periodicity for personal cardiovascular cycle modeling and temporal variation. We then employ adversarial training to generate extra physiological time series data, improving PITN's robustness in the face of sparse subject-specific training data. Furthermore, we utilize contrastive learning to capture the discriminative variations of cardiovascular physiologic phenomena. This approach aggregates physiological signals with similar blood pressure values in latent space while separating clusters of samples with dissimilar blood pressure values. Experiments on three widely-adopted datasets with different modailties (\emph{i.e.,} bioimpedance, PPG, millimeter-wave) demonstrate the superiority and effectiveness of the proposed methods over previous state-of-the-art approaches. The code is available at~\url{https://github.com/Zest86/ACL-PITN}.

Figures (6)

• Figure 1: Illustration of the cuffless blood pressure estimation task by inputting bioimpedance signals in graphene-HGCPT dataset (left top) kireev2022continuous, PPG and millimeter-wave signal in blumio dataset (left bottom) blumio.
• Figure 2: Overall framework of physics-informed temporal network with adversarial contrastive learning for different modal cuffless blood pressure estimation. Our framework mainly contains two data flows: the upper part is about clean signals, which generate predicted values and differentiation for Taylor's approximation. The bottom part is adversarial samples generated by PGD accordingly. Notably, we introduce the adversarial training method in the PGD and construct the contrastive module to enable the model's ability in limited psychological data training. In the inference phase, the auxiliary LNs are not activated for different data distribution between clean and adversarial samples.
• Figure 3: Illustration of the proposed temporal block, which is designed to capture personalized cardiovascular cycles from the 1D signal by transforming it into a 2D tensor, followed by feature extraction using an inception block. The temporal block is stacked in a residual manner.
• Figure 4: Beat-to-beat SBP and DBP estimation based on Ours-Full model (shown in green) and PINN (shown in orange) model trained with the same number of instances and corresponding true BP (shown in dashed black). The figures in each column correspond to the same signal, with the subject listed at the top of each figure. Ours-Full model shows a more precise fit to the reference BP.
• Figure 5: Pearson’s correlation analysis with 2910 blood pressure values using subject #5 in the Graphene-HGCPT dataset. Predictions of the Ours-Full model are in green, while predictions of the PINN model are in orange.