PENGUIN: General Vital Sign Reconstruction from PPG with Flow Matching State Space Models | ICASSP 2026

Transcript

1. PENGUIN: General Vital Sign Reconstruction from PPG with Flow Matching State Space Model Shuntaro Suzuki, Shuitsu Koyama, Shinnosuke Hirano, Shunya Nagashima Neurogica Inc., Japan @Neurogica Inc.
2. Background (1/2): Global burden of cardiovascular disease (CVD) n CVD is a leading cause of mortality worldwide n Hypertension affects >1.28 billion people n Many individuals remain undiagnosed @Neurogica Inc. Global Report on Hypertension [WHO, 25] Population [M] Continuous monitoring of relevant vital signs are essential for early CVD detection n Heart rate n Arterial blood pressure (ABP) n Respiratory rate etc.
3. Background (2/2): PPG as a scalable modality for CVD monitoring n Conventional modalities (ECG, sphygmomanometry) ✗ Bulky → not for continuous monitoring ✓ High SNR and clinically reliable @Neurogica Inc. n Photoplethysmography (PPG) ✓ Compact and energy-efficient → suitable for continuous monitoring ✗ Sensitive to noise and motion artifact Robust data-driven decoding of vital signs from PPG is essential [Wang+, IEEE TBMC16]
4. Previous PPG decoding methods @Neurogica Inc. n Multi-vital sign approaches n PaPaGei [Pillai+, ICLR25] ✓ Broad applicability ✗ Struggle to capture fine-grained waveform morphology n Vital sign-specific approaches n RDDM [Shome+, AAAI23], RespDiff [Miao+, ICASSP25] ✗ Rely on task-specific priors; limited generalizability ✓ High fidelity reconstruction of target vital sign RDDM [Shome+, AAAI23] ✗ PPG2ECG task-specific priors
5. Proposed method: PENGUIN Continuous reconstruction of multiple vital signs @Neurogica Inc. n Built upon flow matching framework for high-fidelity reconstruction n Dual-stream Flow-SSM block for long-range waveform modeling
6. Flow matching for PPG conditioned vital sign reconstruction @Neurogica Inc. n Building upon Optimal Transport Conditional Flow Matching [Lipman+, ICLR23], PENGUIN maps noise distribution to a vital sign distribution conditioned on PPG
7. Flow matching for PPG conditioned vital sign reconstruction @Neurogica Inc. n Building upon Optimal Transport Conditional Flow Matching [Lipman+, ICLR23], PENGUIN maps noise distribution to a vital sign distribution conditioned on PPG PENGUIN: 💡 The model learns time-dependent velocity field
8. Dual-stream Flow-SSM block for capturing long-range PPG / vital sign waveforms @Neurogica Inc. n Dual-stream DiT-based architecture for PPG / vital sign modeling n Extends DiT with state space models (SSMs) n Additive PPG conditioning for fine-grained per-timestep conditioning
9. n Dual-stream DiT-based architecture for modeling PPG / vital sign n Extends DiT with state space models (SSMs) n Additive PPG conditioning for fine-grained per-timestep conditioning Dual-stream Flow-SSM block for capturing long-range PPG / vital sign waveforms @Neurogica Inc. n SSMs (e.g., Mamba [Gu+, COLM24]) enable efficient long-range sequence modeling n We adopt S5 [Smith+, ICLR23], an SSM variant well-suited for continuous signals Hidden state Output signal Input signal
10. Dual-stream Flow-SSM block for capturing long-range PPG / vital sign waveforms @Neurogica Inc. n Dual-stream DiT-based architecture for modeling PPG / vital sign n Extends DiT with state space models (SSMs) n Additive PPG conditioning for fine-grained, per-timestep conditioning n Directly injects PPG morphology into the generative process (e.g., PPG systolic peak → ECG R-peak) n Unlike cross-attention or post- concatenation used in DiT models
11. Experiments @Neurogica Inc. Target PPG decoding tasks Task Dataset Window Metric ECG monitoring (PPG2ECG) PPG-DaLiA 8.0 sec. Heart rate (HR) error: MAE (bpm) of HR estimated by Hamilton method WildPPG Respiratory monitoring (PPG2Resp. rate) BIDMC 60.0 sec. Respiratory rate (RR) error: MAE (bpm) for dominant non-negative frequency WESAD ABP monitoring (PPG2ABP) UCI-BP 8.0 sec. Systolic / diastolic BP (SBP / DBP) error: MAE (mmHg) for max./ min. value of ABP MIMIC-BP Baseline methods n Specialist model: CycleGAN [Aqajari+, EMBC21], RDDM [Shome+, AAAI23], RespDiff [Miao+, ICASSP25] n Generalist model: PaPaGei [Pillai+, ICLR25]
12. Quantitative results: PENGUIN consistently outperformed baseline methods Dataset Metric Specialist model Generalist model CycleGAN RDDM RespDiff PaPaGei-S PENGUIN ECG reconstruction PPG-DaLiA HR error [bpm] ↓ 23.61 16.43 22.75 40.89 15.64 WildPPG 23.21 16.02 20.57 55.42 12.97 Respiratory monitoring BIDMC RR error [bpm] ↓ 9.78 13.88 3.71 4.48 2.98 WESAD 11.93 10.12 5.12 5.84 4.45 Arterial blood pressure monitoring UCI-BP SBP error [mmHg] ↓ 25.79 44.37 78.83 37.01 12.61 DBP error [mmHg] ↓ 12.76 16.57 26.13 13.34 7.14 MIMIC-BP SBP error [mmHg] ↓ 20.26 22.84 97.65 38.42 17.43 DBP error [mmHg] ↓ 10.49 11.83 19.75 11.52 11.34
13. Dataset Metric Specialist model Generalist model CycleGAN RDDM RespDiff PaPaGei-S PENGUIN ECG reconstruction PPG-DaLiA HR error [bpm] ↓ 23.61 16.43 22.75 40.89 15.64 WildPPG 23.21 16.02 20.57 55.42 12.97 Respiratory monitoring BIDMC RR error [bpm] ↓ 9.78 13.88 3.71 4.48 2.98 WESAD 11.93 10.12 5.12 5.84 4.45 Arterial blood pressure monitoring UCI-BP SBP error [mmHg] ↓ 25.79 44.37 78.83 37.01 12.61 DBP error [mmHg] ↓ 12.76 16.57 26.13 13.34 7.14 MIMIC-BP SBP error [mmHg] ↓ 20.26 22.84 97.65 38.42 17.43 DBP error [mmHg] ↓ 10.49 11.83 19.75 11.52 11.34 - 0.79 - 3.05 - 0.73 - 0.67 - 13.18 - 5.62 - 2.83 + 0.85 Quantitative results: PENGUIN consistently outperformed baseline methods
14. @Neurogica Inc. Ground truth Vital sign PENGUIN PPG PaPaGei-S ECG monitoring (WildPPG) Resp. monitoring (BIDMC) ABP monitoring (MIMIC-BP) Qualitative results: PENGUIN preserves morphological characteristics
15. Resp. monitoring (BIDMC) ABP monitoring (MIMIC-BP) PPG ECG monitoring (WildPPG) @Neurogica Inc. PENGUIN PaPaGei-S Ground truth Vital sign J Consistent morphology, and timing of ECG R-peaks and ABP systolic peaks. Qualitative results: PENGUIN preserves morphological characteristics
16. Recap @Neurogica Inc. n Background n Decoding vital signs from PPG is crucial for continuous CVD monitoring n Proposed method: PENGUIN n Enables continuous reconstruction of multiple vital signs from PPG n Results n Outperformed both specialist and generalist methods while preserving waveform morphology Our code is available here!