PPGFlowECG: Latent Rectified Flow with Cross-Modal Encoding for PPG-Guided ECG Generation and Cardiovascular Disease Detection

Xiaocheng Fang; Jiarui Jin; Haoyu Wang; Che Liu; Jieyi Cai; Yujie Xiao; Guangkun Nie; Bo Liu

cs.LGSep 24, 2025

PPGFlowECG: Latent Rectified Flow with Cross-Modal Encoding for PPG-Guided ECG Generation and Cardiovascular Disease Detection

Authors: Xiaocheng Fang, Jiarui Jin, Haoyu Wang, Che Liu, Jieyi Cai, Yujie Xiao, Guangkun Nie, Bo Liu,

Organizations: National Institute of Health Data Science, Peking University, China · School of Intelligence Science and Technology, Peking University, China · University of Chinese Academy of Sciences, China · Data Science Institute, Imperial College London, UK

Abstract

Electrocardiography (ECG) is the clinical gold standard for cardiovascular disease (CVD) assessment, yet continuous monitoring is constrained by the need for dedicated hardware and trained personnel. Photoplethysmography (PPG) is ubiquitous in wearable devices and readily scalable, but it lacks electrophysiological specificity, limiting diagnostic reliability. While generative methods aim to translate PPG into clinically useful ECG signals, existing approaches are limited by the misalignment of physiological semantics in generative models and the complexity of modeling in high-dimensional signals. To address these limitations, we propose PPGFlowECG, a two-stage framework that aligns PPG and ECG in a shared latent space using the CardioAlign Encoder and then synthesizes ECGs with latent rectified flow. We further provide a formal analysis of this coupling, showing that the CardioAlign Encoder is necessary to guarantee stable and semantically consistent ECG synthesis under our formulation. Extensive experiments on four datasets demonstrate improved synthesis fidelity and downstream diagnostic utility. These results indicate that PPGFlowECG supports scalable, wearable-first CVD screening when standard ECG acquisition is unavailable.

Figures & tables

Fig. 1 : Motivation of PPG-to-ECG translation for wearable cardiovascular assessment. Translated ECGs provide diagnostic cues from accessible PPG signals, supporting atrial fibrillation and CVD screening.

Fig. 2 : Overview of PPGFlowECG for PPG-guided ECG generation. Stage 1 learns a shared PPG–ECG latent space with global distribution alignment, local instance discrimination, and semantic decodability constraints; Stage 2 synthesizes ECG latents from PPG-conditioned latent rectified flow and decodes them into ECG waveforms.

Method	Ref.	MCMED			MIMIC-AFib			VitalDB			BIDMC
Method	Ref.	MAE	RMSE	$\boldsymbol{\mathrm{MAE}_{\mathrm{HR}}}$	MAE	RMSE	$\boldsymbol{\mathrm{MAE}_{\mathrm{HR}}}$	MAE	RMSE	$\boldsymbol{\mathrm{MAE}_{\mathrm{HR}}}$	MAE	RMSE	$\boldsymbol{\mathrm{MAE}_{\mathrm{HR}}}$
DDPM [ 13 ]	NeurIPS’20	0.94	1.36	7.00	0.73	1.22	9.55	0.83	1.31	11.57	0.79	1.24	5.82
CardioGAN [ 35 ]	AAAI’21	0.98	1.40	2.72	0.73	1.22	3.42	0.82	1.32	3.50	0.82	1.31	1.38
Rectified Flow [ 23 ]	ICLR’23	0.97	1.36	26.40	0.83	1.36	26.01	0.90	1.37	36.36	0.82	1.33	9.14
RDDM [ 38 ]	AAAI’24	0.99	1.41	7.45	0.82	1.37	15.35	0.87	1.40	30.27	0.83	1.34	10.23
CardioFlow [ 24 ]	ICASSP’25	0.95	1.35	15.37	0.76	1.33	17.92	0.85	1.34	28.87	0.79	1.31	7.37

TABLE I : Comparison with state-of-the-art approaches on four datasets. Lower values are better for all metrics. Best and second-best results are highlighted in bold and underline, respectively.

Dataset	Patients/Visits	Label	10-s Segments
			Train	Val.	Test
MCMED [ 17 ]	118,385	CVD	10,356,840	1,299,869	37,052
MIMIC-AFib [ 2 ]	35	AF	2,044	–	577
VitalDB [ 20 ]	6,388	–	1,007,165	–	10,720
BIDMC [ 32 ]	53	–	1,458	–	326

TABLE II : Dataset statistics after preprocessing. The last three columns report the numbers of high-quality 10-second PPG–ECG segment pairs.

Panel of Figure 5 — Fig. 3 : (a) Heart-rate estimation consistency. PPGFlowECG-generated ECGs yield HR estimates closer to real ECGs than raw PPG. (b) Downstream CVD categories evaluated on MCMED, spanning rhythm, conduction, vascular, and structural disorders.

Method	Ref.	I48	I71	I70	I44	I25	I50	Macro-AUROC
PPG-based
PaPaGei [ 31 ]	ICLR’25	61.0	57.3	55.1	58.9	55.7	55.4	57.2
Gen.ECG-based
DDPM [ 13 ]	NeurIPS’20	66.6	60.5	57.6	62.5	58.4	59.1	60.8
CardioGAN [ 35 ]	AAAI’21	61.2	65.7	56.7	57.0	57.4	53.9	58.7
Rectified Flow [ 23 ]	ICLR’23	49.9	50.2	50.4	49.6	49.7	49.7	49.9

TABLE III: Downstream multi-label CVD classification on MCMED. AUROC (%) is reported for each ICD category and Macro-AUROC; best and second-best results are bolded and underlined.

Method	Ref.	Acc. $\uparrow$	F1 $\uparrow$
PPG-based
Shen et al. [ 37 ]	KDD’19	0.54	0.51
PaPaGei [ 31 ]	ICLR’25	0.69	0.75
Gen.ECG-based
DDPM [ 13 ]	NeurIPS’20	0.70	0.80
CardioGAN [ 35 ]	AAAI’21	0.71	0.83

TABLE IV: AF detection performance on the MIMIC-AFib dataset. Higher values are better.

Variant	CAE	LRF	$\text{MAE}_{\text{HR}}$	Macro-AUROC
Baseline	–	–	3.28	53.1
w/ CAE	✓	–	1.93	61.5
w/ LRF	–	✓	2.12	58.9
Full model	✓	✓	1.80	63.1

TABLE V: Ablation study on the MCMED dataset.

Fig. 4 : Qualitative comparison of PPG-guided ECG synthesis across four datasets. Rows show the input PPG, reference ECG, and ECGs generated by different methods.

Fig. 5 : Ablation study of the effect of sampling steps.

Fig. 6 : Grad-CAM attribution maps for I70 diagnosis on real and PPGFlowECG-generated ECGs.

Fig. 7 : Performance of five cardiologists on the Turing Test.

Input Modality	Acc. $\uparrow$	Sens. $\uparrow$	Spec. $\uparrow$
PPG only	$0.81\pm 0.13$	$0.68\pm 0.15$	$0.85\pm 0.07$
PPG + Real ECG	$\mathbf{0.89\pm 0.07}$	$0.88\pm 0.09$	$\mathbf{0.93\pm 0.05}$
PPG + Synth. ECG (Ours)	$0.87\pm 0.09$	$\mathbf{0.92\pm 0.04}$	$0.88\pm 0.11$

TABLE VI : Cardiologist AF diagnosis under different input modalities. Values are reported as mean

\pm

standard deviation over five cardiologists. Best results are highlighted in bold.

Variant	GDA	LID	SDC	$\text{MAE}_{\text{HR}}$	Macro-AUROC
Baseline	–	–	–	3.28	53.1
+ GDA	✓	–	–	2.11	58.3
+ GDA + LID	✓	✓	–	1.97	60.7
CardioAlign Encoder	✓	✓	✓	1.93	61.5

TABLE VII: Progressive ablation of CardioAlign Encoder.

Electrocardiography

Electrocardiography studies the heart’s electrical signals to measure cardiac function and support diagnosis, monitoring, and research.

See all 108 papers

PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation
May 9, 2026Xiaoda Wang, Minxiao Wang, Kaiqiao Han +10Electrocardiography Physics-Informed ML
Department of Computer Science, Emory University · Department of Computer Science, University of California, Los Angeles · Nell Hodgson Woodruff School of Nursing, Emory University +2
BeatFlow-ECG: Rectified Flow for ECG Reconstruction from Indirect Wearable Signals
Oct 6, 2026Mohamed Kamel, Sahar Selim, Walaa Medhat +1Electrocardiography Wearable Health Monitoring
Department of Computer Science, Virginia Commonwealth University, Richmond, VA, USA · School of Information Technology and Computer Science, Nile University, Giza, Egypt
Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model
May 10, 2026Zhangdaihong Liu, Chang Liu, Fenglin Liu +4Cross-Modal Representation Learning Masked Autoencoders
Department of Engineering Science, University of Oxford, Oxford OX3 7DQ, UK · Oxford Suzhou Centre for Advanced Research, Suzhou, 215123, China · School of Public Health, Shanghai Jiao Tong University, Shanghai, China

May 9, 2026J/K move · Enter open · S save

PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation

Xiaoda Wang, Minxiao Wang, Kaiqiao Han, Defu Cao+9

Department of Computer Science, Emory University · Department of Computer Science, University of California, Los Angeles · Nell Hodgson Woodruff School of Nursing, Emory University+2

Electrocardiography Physics-Informed ML Photoplethysmography Conditional Generative Modeling Conditional Generation

Open paper AskPDF arXiv

Electrocardiography (ECG) is the clinical standard for cardiac assessment but requires dedicated hardware that does not scale to daily-life monitoring. Photoplethysmography (PPG) is ubiquitous in wearables but lacks ECG-specific diagnostic morphology and is corrupted by motion and sensor noise. PPG-to-ECG generation aims to bridge this gap by recovering electrical morphology and timing from peripheral pulse signals. However, existing methods largely rely on statistical alignment and data-driven generation. They fail to explicitly structure the latent space around physiology-aware electro-hemodynamic factors and lack constraints from forward physiological dynamics. To address these challenges, we propose PG-LRF, a physiology-guided latent rectified flow framework. PG-LRF introduces an electro-hemodynamic simulator that co-models ECG and PPG through shared cardiac phase dynamics. Guided by this simulator, a Physiology-Aware AutoEncoder learns a structured electro-hemodynamic latent space. Then we integrate this simulator guidance into a PPG-conditioned latent rectified flow, enforcing ECG-side morphology consistency and ECG-to-PPG forward hemodynamic consistency during generative transport. Experiments on the large-scale MC-MED dataset demonstrate that PG-LRF significantly improves PPG-to-ECG generation and downstream cardiovascular disease classification, proving its ability to generate ECGs that are both signal-faithful and physiologically plausible under the ECG-to-PPG hemodynamic pathway

Abstract

Electrocardiography

Explore similar work

PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation

BeatFlow-ECG: Rectified Flow for ECG Reconstruction from Indirect Wearable Signals

Biosignal Fingerprinting: A Cross-Modal PPG-ECG Foundation Model