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

Cheng Bian, Xiaoyu Li, Qi Bi, Guangpu Zhu, Jiegeng Lyu, Weile Zhang, Yelei Li, Zijing Zeng

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.

TL;DR

Researchers from OPPO Health Lab have developed the Latent Space Constraint Transformer (LSCT), a new deep learning framework that converts non-invasive photoplethysmography (PPG) signals—the kind measured by your smartwatch—into high-fidelity arterial blood pressure (ABP) waveforms. By using a quantized codebook and graph-based spectrum enhancement, the model remains remarkably robust even when the input PPG data is noisy or partially missing ("anomalous").

The Challenge: Latent Space Shift

Arterial Blood Pressure (ABP) provides a wealth of cardiovascular data, but measuring it traditionally requires an invasive catheter. While PPG sensors (optical sensors found in wearables) are a convenient alternative, translating a PPG signal into an ABP waveform is notoriously difficult.

The primary challenge is latent space shift. Because PPG data varies wildly across different hardware and individuals, and is often corrupted by motion or poor sensor contact, standard AI models struggle to maintain a consistent internal representation. When the input is "anomalous" (noisy or incomplete), the resulting ABP waveform becomes distorted, losing the clinical details necessary for diagnosing hypertension or other cardiac issues.

Visualization of anomalous PPG signals and the resulting latent space shift. The bottom plots compare the stability of latent representations with and without masking.

The LSCT Architecture

The authors introduce the Latent Space Constraint Transformer (LSCT) to "anchor" the model’s internal representations. The architecture consists of three novel components:

  1. Vector Quantized Latent Space: Instead of allowing the model to generate any arbitrary internal representation, the authors use a codebook (a dictionary of discrete "bases"). This forces the model to map noisy PPG signals onto a set of stable, pre-defined templates, effectively filtering out anomalies.
  2. Correlation-boosted Attention Module (CAM): Standard quantization methods usually pick the single "closest" template from the codebook. CAM instead uses a cross-attention mechanism to query and combine multiple relevant bases from the codebook globally. This allows the model to "fill in the blanks" if part of the input signal is defective.
The Correlation-boosted Attention Module (CAM) uses a cross-attention mechanism to query multiple relevant bases from a codebook, enhancing signal reconstruction.
  1. Multi-Spectrum Enhancement Knowledge (MSEK): To ensure the reconstructed waveform is precise, MSEK treats different frequency channels of the latent code as nodes in a graph. By modeling the information flow between these "spectrum nodes," the model captures fine-grained local details that are often lost in simpler transformations.

Datasets and Evaluation

The model was trained and evaluated on three distinct datasets:

  • MIMIC-III: A massive public clinical database. The authors used a subset of 127,260 signal segments from 942 patients.
  • VitalDB: A public dataset containing high-fidelity surgical data from 144 patients.
  • OML (Private): A downstream task dataset consisting of 20 subjects wearing professional-grade wearables for 30 days, used to test how well the waveform transformation generalizes to actual blood pressure value regression.

Key Insights

  • Discretization as a Denoiser: By constraining the latent space to a discrete codebook, the model creates a "bottleneck" that prevents noise from propagating into the output. It acts as a physiological prior, ensuring the output looks like a real ABP wave.
  • Global Querying vs. Local Lookup: By using the CAM module to query the codebook globally, the model doesn't just look for a match; it reconstructs the "idea" of the signal. This is why the model performs so well even when 50% or 70% of the PPG signal is masked.
  • Spectrum Graphing: Treating channels as nodes in a graph (MSEK) is a clever way to handle the multi-scale nature of biological signals, where different frequencies carry different physiological meanings.

Results and Clinical Impact

The LSCT framework achieved state-of-the-art (SOTA) performance across the board:

  • Accuracy: On the MIMIC-III dataset, it achieved a Root Mean Squared Error (RMSE) of 0.097, outperforming previous SOTA models like PPG2ABP (0.101).
  • Robustness: In "stress tests" where the input PPG signal was randomly masked, the LSCT maintained high fidelity. Even with a 90% mask ratio, the model's performance degradation was significantly lower than baseline models.
  • Clinical Utility: On the private OML dataset, the model predicted Systolic Blood Pressure (SBP) with a Mean Absolute Error (MAE) of 8.04 mmHg, a notable improvement over standard architectures like Swin-Transformer (11.17 mmHg).

Limitations and Future Directions

While the results are impressive, the authors note that real-world "ambulatory" data (people walking, exercising) presents different types of noise than the simulated masking used in this study. Additionally, the use of graph-based modules and cross-attention adds computational overhead, which may require optimization for deployment on low-power wearables. Future work could expand this approach to other medical signal transformations, such as ECG-to-ABP.

Summary updated

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.
BibTeX
@inproceedings{bian2024constraint,
  author = {Cheng Bian and Xiaoyu Li and Qi Bi and Guangpu Zhu and Jiegeng Lyu and Weile Zhang and Yelei Li and Zijing Zeng},
  title = {{Constraint Latent Space Matters: An Anti-anomalous Waveform Transformation Solution from Photoplethysmography to Arterial Blood Pressure}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=yhLAte_KtJf},
}
Figure