Research map / Probabilistic and scientific modeling / Molecules and proteins
Constraint Latent Space Matters: An Anti-anomalous Waveform Transformation Solution from Photoplethysmography to Arterial Blood Pressure
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LSCT, a VQ-VAE-style codebook transformer with cross-attention basis querying (CAM) and channel-graph enhancement (MSEK) to convert anomalous PPG into ABP waveforms robustly.
Problem
Converting PPG to arterial blood pressure waveforms is unstable under anomalous or masked PPG inputs, because the latent space shifts with changes in hardware and individuals, distorting the decoded ABP.
Approach
Quantize the latent space with a codebook, query all relevant bases globally via cross-attention (CAM), and add a channel-wise graph attention module (MSEK) over the latent channels to enrich the representation fed to the decoder.
Key results
- On MIMIC-III (10% mask), LSCT reaches RMSE 0.097, PRD 3.402, FD 3.107, versus 0.101 RMSE for PPG2ABP, the best baseline.
- On VitalDB, LSCT achieves RMSE 0.016, PRD 0.258, FD 0.566, improving on VQ-VAE (0.019 RMSE).
- On the private OML dataset, LSCT obtains SBP MAE 8.04 and DBP MAE 6.49, best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LSCT rather than the strongest listed baseline PPG2ABP | PPG-to-ABP waveform transformation · MIMIC III | RMSE: 0.101 -> 0.097 | Helped |
| Add LSCT rather than the strongest listed baseline Unet | PPG-to-ABP waveform transformation · MIMIC III | PRD: 3.618 -> 3.402 | Helped |
| Add LSCT rather than the strongest listed baseline PPG2ABP | PPG-to-ABP waveform transformation · MIMIC III | FD: 3.428 -> 3.107 | Helped |
| Add LSCT rather than the strongest listed baseline PPG2ABP | PPG-to-ABP waveform transformation · VitalDB | RMSE: 0.019 -> 0.016 | Helped |
| Add LSCT rather than the strongest listed baseline PPG2ABP | PPG-to-ABP waveform transformation · VitalDB | PRD: 0.281 -> 0.258 | Helped |
| Add LSCT rather than the strongest listed baseline VQ-VAE | PPG-to-ABP waveform transformation · VitalDB | FD: 0.641 -> 0.566 | Helped |
| Add LSCT rather than VQ-VAE | OML blood-pressure regression · OML | SBP MAE: 8.63 -> 8.04 | Helped |
| Add LSCT rather than VQ-VAE | OML blood-pressure regression · OML | DBP MAE: 6.73 -> 6.49 | Helped |
| Add CAM to the VQ-VAE K-nearest-neighbor baseline (ablation) | LSCT · VitalDB, MR 10% | RMSE: 0.019 -> 0.017 | Helped |
| Add MSEK to the VQ-VAE K-nearest-neighbor baseline (ablation) | LSCT · VitalDB, MR 10% | RMSE: 0.019 -> 0.018 | Helped |
| Add MSEK to CAM (ablation) | LSCT · VitalDB, MR 10% | RMSE: 0.017 -> 0.016 | Helped |
| Increase mask ratio from 10% to 90% (ablation) | LSCT · VitalDB | RMSE: 0.016 -> 0.047 | Hurt |
Surprising findings
- Codebook size (64 to 512) has a negligible effect on performance, according to the authors.
- Graph-based codebook variant (GBC) performs worse than plain self-attention graph flow, which the authors attribute to harming basis orthogonality.
Limitations
- Evaluated only on PPG-to-ABP tasks; generalization to broader waveform tasks is untested.
- Private OML dataset is not released and the downstream task uses SBP/DBP values rather than waveforms.
- Hardware and compute details are not reported.
Experimental setup
Datasets: MIMIC-III VitalDB OML
Benchmarks: MIMIC-III VitalDB OML
Metrics: RMSE PRD FD MAE ME SD
Baselines: LSTM U-Net PPG2ABP CycleGAN TTS-CGAN VQ-VAE Time-VQVAE InceptionTime ResNeXt-101 CoAtNet Swin Transformer
Builds on: VQ-VAE Swin Transformer Graph attention network PPG2ABP
Closest papers by meaning
- FactorizePhys: Matrix Factorization for Multidimensional Attention in Remote Physiological Sensing (NeurIPS 2024 · similarity 0.87)
- PaPaGei: Open Foundation Models for Optical Physiological Signals (ICLR 2025 · similarity 0.85)
- PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement (AAAI 2025 · similarity 0.84)
- Region-Disentangled Diffusion Model for High-Fidelity PPG-to-ECG Translation (AAAI 2024 · similarity 0.84)
- FacePhys: State of the Heart Learning (NeurIPS 2026 · similarity 0.84)
- Remote Photoplethysmography in Real-World and Extreme Lighting Scenarios (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.