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

Introduction

Arterial Blood Pressure (ABP) waveforms provide critical insights into cardiovascular health, but obtaining them usually requires invasive arterial lines in a clinical setting. While Photoplethysmography (PPG) sensors—the optical sensors found in smartwatches—can track blood volume changes non-invasively, converting these signals into accurate ABP waveforms is difficult due to the "latent space shift."

Anomalous PPG signals with missing segments compared to clean PPG and ground truth ABP.

In real-world scenarios, PPG data is often "anomalous" due to hardware variations, skin tones, or sensor movement. Standard discriminative models struggle with this noise, often leading to distorted or "hallucinated" ABP waveforms. Researchers from OPPO Health Lab have developed the Latent Space Constraint Transformer (LSCT) to address this, creating a transformation that is robust enough to handle missing or corrupted PPG data without losing clinical detail.

Methods: The LSCT Framework

The LSCT architecture relies on three core innovations to stabilize the transformation process and ensure fidelity even under adverse conditions.

The full LSCT architecture, featuring the Correlation-boosted Attention Module (CAM) and Multi-Spectrum Enhancement Knowledge (MSEK) for robust transformation.

1. Quantized Codebook

Instead of mapping PPG to arbitrary points in a continuous latent space, the model uses a discrete "codebook" of learned signal bases (a VQ-VAE approach). This forces the model to reconstruct the ABP waveform using a set of stable, "vetted" building blocks, which prevents the latent space from drifting when the input is noisy.

2. Correlation-boosted Attention Module (CAM)

Traditional Vector Quantization (VQ) often relies on a simple "nearest neighbor" lookup. CAM instead uses a cross-attention mechanism to query the entire codebook globally. This allows the model to find the optimal combination of signal bases to "fill in the blanks" even if the input PPG is partially corrupted or missing.

3. Multi-Spectrum Enhancement Knowledge (MSEK)

The model treats different channels of the latent code as nodes in a graph. By modeling the relationships between these "spectrums" using a graph neural network, the model better captures the complex physiological dependencies inherent in blood pressure signals.

Datasets and Robustness Testing

The model was evaluated on three large-scale datasets: MIMIC-III (clinical), VitalDB (surgical), and a private OML dataset (wearable hardware). To test real-world robustness, the authors employed a Masked Strategy, randomly zeroing out up to 90% of the PPG signal to simulate sensor disconnects or extreme motion artifacts.

Key Results and Insights

The LSCT model outperformed several state-of-the-art methods, including CycleGAN and standard Transformers, across all primary metrics:

  • Accuracy: On the MIMIC-III dataset, it achieved a Root Mean Squared Error (RMSE) of 0.097, a significant improvement over previous benchmarks.
  • Resilience to Noise: In stress tests with 10% signal masking, LSCT maintained high fidelity while standard models saw error rates spike. Remarkably, the model's performance degraded gracefully even at a 90% mask ratio.
  • Clinical Utility: When predicting actual Systolic (SBP) and Diastolic (DBP) values, the model achieved a Mean Absolute Error (MAE) of 8.04 mmHg for SBP, outperforming specialized regression models.

Conclusion

The LSCT framework demonstrates that by restricting a model's "vocabulary" through a quantized codebook and leveraging global attention, it is possible to create a highly robust waveform transformation. While the computational complexity of graph modeling remains a challenge for low-power wearable chips, this research paves the way for more reliable, non-invasive cardiovascular monitoring in everyday life.

Summary updated

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