Constraint Latent Space Matters: An Anti-anomalous Waveform Transformation Solution from Photoplethysmography to Arterial Blood Pressure

TL;DR

The paper tackles the latent space shift that plagues PPG-to-ABP waveform transformation under device- and subject-related distribution changes. It introduces the Latent Space Constraint Transformer (LSCT), which discretizes latent representations via a codebook and refines them with Correlation-boosted Attention Module (CAM) and Multi-Spectrum Enhancement Knowledge (MSEK) to improve reconstruction fidelity. Across MIMIC-III, VitalDB, and a private OML dataset, LSCT achieves state-of-the-art performance on ABP waveform reconstruction (lower RMSE, PRD, and FD) and demonstrates robust generalization to downstream BP tasks. The approach offers a principled path toward robust cuffless ABP monitoring and shows potential for broader waveform transformation applications, supported by extensive ablations that validate the contributions of CAM and MSEK. Key mathematical constructs include the latent code $\mathbf{z_q}$, codebook $\mathbf{M}$, attended representation $\mathbf{z_v}$, graph-enhanced $\mathbf{z_g}$, and the decoding relation $\hat{u}_a = \mathcal{H}_\varphi(\text{MSEK}(\text{CAM}(\mathbf{z_q}))+\mathbf{z_q})$.

Abstract

Arterial blood pressure (ABP) holds substantial promise for proactive cardiovascular health management. Notwithstanding its potential, the invasive nature of ABP measurements confines their utility primarily to clinical environments, limiting their applicability for continuous monitoring beyond medical facilities. The conversion of photoplethysmography (PPG) signals into ABP equivalents has garnered significant attention due to its potential in revolutionizing cardiovascular disease management. Recent strides in PPG-to-ABP prediction encompass the integration of generative and discriminative models. Despite these advances, the efficacy of these models is curtailed by the latent space shift predicament, stemming from alterations in PPG data distribution across disparate hardware and individuals, potentially leading to distorted ABP waveforms. To tackle this problem, we present an innovative solution named the Latent Space Constraint Transformer (LSCT), leveraging a quantized codebook to yield robust latent spaces by employing multiple discretizing bases. To facilitate improved reconstruction, the Correlation-boosted Attention Module (CAM) is introduced to systematically query pertinent bases on a global scale. Furthermore, to enhance expressive capacity, we propose the Multi-Spectrum Enhancement Knowledge (MSEK), which fosters local information flow within the channels of latent code and provides additional embedding for reconstruction. Through comprehensive experimentation on both publicly available datasets and a private downstream task dataset, the proposed approach demonstrates noteworthy performance enhancements compared to existing methods. Extensive ablation studies further substantiate the effectiveness of each introduced module.

Figures (5)

• Figure 1: Top: Illustration of PPG, anomalous PPG and ABP signals. Bottom: Visualization of the latent space shift. Bottom Left: Visualization (t-SNE) results from naive discriminative model (a.k.a., Swin-Transformer swin_transformer). Bottom Right: Results from ours. Note that blue points denote the original features from PPG signals in the latent space, while orange points are the features from anomalous PPG which is random masked under the ratio of 10%. Corresponding signals (w or w/o masks) are connected by gray lines, where the length of each line represents the distance.
• Figure 2: Overview of our proposed Latent Space Constraint Transformer (LSCT) framework. Firstly, anomalous PPG waveforms are used as inputs fed to the encoder. Then, the proposed Correlation-boosted Attention Module (CAM) queries relative bases to form the attentional representation $\mathbf{z_v}$. Afterward, Multi-Spectrum Enhancement Knowledge (MSEK) constructs spectrum-wise graph flow from the comprehensive representation $\mathbf{z_g}$. Finally, the ABP waveforms are decoded by the summation of the latent code $\mathbf{z_q}$ and the comprehensive representation $\mathbf{z_g}$.
• Figure 3: Qualitative comparisons of ours with the state-of-the-art methods on the MIMIC-III dataset, using the mask ratio of 10%.
• Figure 4: RMSE comparison of typical state-of-the-art methods and ours given by different MRs.
• Figure 5: Performance with varied size of codebook and dimension of bases.