PPGFlowECG: Cross-Modal ECG Translation

Updated 12 July 2026

PPGFlowECG is a cross-modal translation paradigm that converts accessible PPG signals into diagnostic ECG waveforms using cycle-wise reconstruction and latent alignment.
It leverages a range of methods—including DCT-based regression, attention-driven state-space models, GANs, and diffusion techniques—to preserve waveform fidelity and clinical features.
The approach emphasizes precise temporal matching, subject-independent learning, and physiology-guided refinements to support continuous cardiac monitoring and accurate disease detection.

PPGFlowECG designates PPG-to-ECG translation systems that infer or generate electrocardiographic waveforms from photoplethysmographic input. In the literature represented here, the term spans a cycle-wise inverse reconstruction pipeline based on discrete cosine transform (DCT) coefficients and ridge regression (Zhu et al., 2019), and, in later work, a two-stage framework that aligns PPG and ECG in a shared latent space through a CardioAlign Encoder and then generates ECG with latent rectified flow for downstream cardiovascular disease (CVD) detection (Fang et al., 24 Sep 2025). Across adjacent work, PPGFlowECG is situated within a broader technical program that includes attention-based deep state-space modeling, GANs with frequency-domain constraints, conditional diffusion, physiology-guided latent rectified flow, and flow-matching state-space models (Vo et al., 2023, Kong et al., 2024, Belhasin et al., 2024, Wang et al., 9 May 2026, Suzuki et al., 23 Jan 2026).

1. Scope and historical development

The motivating premise is consistent across the corpus: ECG is the gold standard for cardiac monitoring and provides electrophysiological information used for diagnosis, whereas PPG is accessible in wearables and suitable for continuous monitoring but does not, by itself, provide equivalent diagnostic morphology (Fang et al., 24 Sep 2025). Earlier work framed the problem as an inverse mapping between simultaneously recorded peripheral pulse and electrical activity; later work reformulated it as conditional generative modeling in aligned latent spaces, often with explicit downstream screening objectives (Zhu et al., 2019, Fang et al., 24 Sep 2025).

Representative work	Core mechanism	Reported outcome
"ECG Reconstruction via PPG: A Pilot Study" (Zhu et al., 2019)	DCT-cycle mapping with ridge regression	mean $\rho \approx 0.985$ under R2R segmentation
"Inferring ECG from PPG for Continuous Cardiac Monitoring Using Lightweight Neural Network" (Li et al., 2020)	Lightweight encoder-decoder with diagnosis-oriented training	MIMIC-III: $\rho = 0.94$ , rRMSE = 0.27
"PPG-to-ECG Signal Translation for Continuous Atrial Fibrillation Detection via Attention-based Deep State-Space Modeling" (Vo et al., 2023)	Subject-independent attention-based deep state-space model	PR-AUC = 0.986 using translated ECG in AFib detection
"PPGFlowECG: Latent Rectified Flow with Cross-Modal Encoding for PPG-Guided ECG Generation and Cardiovascular Disease Detection" (Fang et al., 24 Sep 2025)	CardioAlign Encoder + latent rectified flow	MAE = 0.73 mV; Macro-AUROC = 0.631 on MCMED
"PG-LRF: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation" (Wang et al., 9 May 2026)	Physiology-aware autoencoding + simulator-guided latent rectified flow	MAE 0.87→0.71; macro-AUROC 0.545→0.633

This progression suggests a shift in emphasis from waveform reconstruction alone toward semantically aligned generation, physiological plausibility, uncertainty handling, and diagnostic utility. A plausible implication is that PPGFlowECG is best understood not as a single algorithmic template, but as a family of PPG-conditioned ECG synthesis paradigms with increasingly structured inductive biases.

2. Cycle-wise inverse reconstruction and temporal correspondence

The 2019 pilot study formalized ECG reconstruction via PPG as a cycle-wise inverse problem. After alignment and detrending, raw PPG $x[n]\in\mathbb{R}^T$ and ECG $y[n]\in\mathbb{R}^T$ are segmented into $N$ cardiac cycles, each resampled by linear interpolation to a fixed length $L$ . Type-II DCT is applied to each cycle, and only the first $L_x$ PPG coefficients and first $L_y$ ECG coefficients are retained because most energy is in low frequencies. The retained ECG coefficients are predicted from retained PPG coefficients using ridge regression,

$f^* = (X_{\text{train}}^T X_{\text{train}} + \gamma I)^{-1} X_{\text{train}}^T Y_{\text{train}},$

and inverse DCT yields the reconstructed ECG waveform (Zhu et al., 2019).

Two segmentation schemes were used. In SR, boundaries are set at $1/3$ cycle before each ECG R-peak. In R2R, boundaries are successive ECG R-peaks. On CapnoBase TBME-RR, with 42 sessions from 29 pediatric and 13 adult surgical patients, sampling rate 300 Hz, and cross-validated hyperparameters $\rho = 0.94$ 0, $\rho = 0.94$ 1, $\rho = 0.94$ 2, $\rho = 0.94$ 3, and $\rho = 0.94$ 4, the subject-dependent evaluation reported mean rRmse $\rho = 0.94$ 5 and mean $\rho = 0.94$ 6 for R2R, versus mean rRmse $\rho = 0.94$ 7 and mean $\rho = 0.94$ 8 for SR. The best average cycle-by-cycle correlation was $\rho = 0.94$ 9 (Zhu et al., 2019).

The same work also enumerated several limitations: reliance on accurate cycle detection, subject dependence, evaluation only on low-noise clinical data, incomplete compensation for variation in pulse transit time, and possible loss of high-frequency ECG features such as sharp QRS complexes. The suggested remedies were nonlinear regression, subject-independent training with more subjects and data augmentation, additional PPG features such as pulse arrival time and derivative waveforms, robust denoising, and adaptive selection of DCT basis orders (Zhu et al., 2019).

A complementary line of work addressed temporal correspondence explicitly through runtime monitoring and regression analysis. Timed automata were used to monitor ECG and PPG events in parallel, including policies such as PR interval versus systole period and a cross-signal temporal pattern ensuring that every R-peak is followed 600–700 ms later by a PPG systolic peak. On the BIDMC PPG & Respiration Database, the reported correlations were $x[n]\in\mathbb{R}^T$ 0 for RR versus peak-to-peak, $x[n]\in\mathbb{R}^T$ 1 for PR versus systole, and $x[n]\in\mathbb{R}^T$ 2 for RP versus diastole, all with statistically significant $x[n]\in\mathbb{R}^T$ 3-values. The fitted event-time regressions were $x[n]\in\mathbb{R}^T$ 4 and $x[n]\in\mathbb{R}^T$ 5, with composition verdicts remaining ok on more than 90% of cycles (Panda et al., 2022).

Taken together, these results establish a foundational PPGFlowECG principle: successful translation depends on precise temporal pairing, whether via cycle segmentation, event alignment, or latent conditioning mechanisms that compensate for the physiological delay between electrical depolarization and peripheral pulse arrival.

3. Neural translation before latent rectified flow

Subsequent work replaced explicit DCT-domain regression with neural sequence models that targeted morphology preservation, subject independence, robustness to noise, or data efficiency. A lightweight encoder-decoder model used multi-scale grouped convolutions, attention gates, channel shuffle, and a global residual skip. Its diagnosis-oriented training objective combined a weighted $x[n]\in\mathbb{R}^T$ 6 reconstruction term, a feature loss computed from a frozen ECG-based CVD classifier, and a contrastive loss over ECG features with different labels. The compressed student model fit under 40 K parameters, while the full teacher model used approximately 120 K parameters. Across MIMIC-III, BIDMC, and a motion dataset, the proposed full model reported average $x[n]\in\mathbb{R}^T$ 7 and rRMSE = 0.40, while the compressed model reported $x[n]\in\mathbb{R}^T$ 8 and rRMSE = 0.42; on MIMIC-III specifically, $x[n]\in\mathbb{R}^T$ 9 and rRMSE = 0.27 were obtained, and timing errors were PR MAE = 4 ms, QRS MAE = 5 ms, and QT MAE = 9 ms (Li et al., 2020).

The same study addressed motion artifacts by concatenating three-axis acceleration with PPG as a four-channel input. On the Motion dataset, this reduced rRMSE from 0.57 to 0.51 and increased $y[n]\in\mathbb{R}^T$ 0 from 0.81 to 0.84. It also reported that leave-one-out subject-independent training on BIDMC yielded $y[n]\in\mathbb{R}^T$ 1, whereas two epochs of subject-specific fine-tuning raised $y[n]\in\mathbb{R}^T$ 2 to 0.97, thereby quantifying the gap between generalization and personalization (Li et al., 2020).

A more explicitly probabilistic alternative was the subject-independent attention-based deep state-space model (ADSSM). ADSSM defines a nonlinear Gaussian state-space model with Gaussian emissions, Gaussian transition priors conditioned on an attention-weighted summary of the entire PPG history, and a structured variational posterior trained with an ELBO and KL warm-up. The model was described as robust to noise and data-efficient because it embeds probabilistic prior knowledge while using approximately 645K parameters. On 55 subjects from MIMIC-III, including noisy modifications designed to mimic real-world scenarios, translated ECG segments fed into the MINA AFib detector yielded PR-AUC = 0.986 and ROC-AUC = 0.99, compared with 0.987 and 0.995 on real ECG. Even on subjects with AFib, the model maintained translation quality of $y[n]\in\mathbb{R}^T$ 3 (Vo et al., 2023).

Adversarial and diffusion formulations then broadened the model class. The frequency-domain-constrained GAN used a conditional U-Net with attention gates as generator and a 1D PatchGAN-style discriminator. Its defining innovation was a frequency-domain loss

$y[n]\in\mathbb{R}^T$ 4

excluding the DC component and combined with the original GAN objective using $y[n]\in\mathbb{R}^T$ 5. On PPG-Dalia, using wrist PPG and chest ECG during daily activities, the method reduced validation-set heart-rate MAPE from approximately 14% to approximately 12% and reduced the standard deviation across 31 random seeds from approximately 12% to approximately 3%; on the test set, PPG-only peak detection gave 15%, the original GAN 14%, and f-GAN 12% (Kong et al., 2024).

Uncertainty-aware diffusion treated ECG synthesis as posterior inference rather than point estimation. A conditional 1D U-Net denoiser was trained with the DDPM simple loss on 1024-sample PPG and ECG segments, and inference used multiple posterior samples to estimate per-time-point means, variances, and 95% intervals. The work also introduced Expected Score Classification (ESC), which averages classifier outputs across multiple synthetic ECG samples. On MIMIC-III conversion quality, UA-P2E DDIM with $y[n]\in\mathbb{R}^T$ 6 reported RMSE $y[n]\in\mathbb{R}^T$ 7, 1-FD $y[n]\in\mathbb{R}^T$ 8, and 100-FD $y[n]\in\mathbb{R}^T$ 9; on CinC classification, synthetic ECG ESC reported macro-AUROC $N$ 0 and macro-AURC $N$ 1, improving markedly over single-sample baselines (Belhasin et al., 2024).

These pre-flow models collectively introduced four design motifs that recur in later PPGFlowECG systems: subject-independent parameter sharing, morphology-aware objectives, frequency-domain regularization, and explicit uncertainty quantification.

The 2025 system titled "PPGFlowECG: Latent Rectified Flow with Cross-Modal Encoding for PPG-Guided ECG Generation and Cardiovascular Disease Detection" formalized a two-stage “Align→Generate” architecture. Stage 1 uses a shared CardioAlign Encoder. For a 10 s PPG or ECG segment at 128 Hz, the encoder produces a latent posterior $N$ 2, from which modality-specific decoders reconstruct PPG and ECG. Alignment is enforced through distribution-level alignment,

$N$ 3

contrastive instance alignment via InfoNCE, and cross-modal reconstruction. The total Stage 1 objective combines within-modality reconstruction, KL regularization, $N$ 4, $N$ 5, and $N$ 6, with $N$ 7 and typical $N$ 8 (Fang et al., 24 Sep 2025).

Stage 2 applies latent rectified flow in the aligned latent space. Conditioning is $N$ 9, target is $L$ 0, and the straight-line interpolation is $L$ 1 for $L$ 2. A Transformer-based vector field with cross-attention at every layer learns $L$ 3 using the flow-matching objective

$L$ 4

At inference, the ODE

$L$ 5

is discretized with Euler steps, and the terminal latent is decoded by the frozen ECG decoder (Fang et al., 24 Sep 2025).

The model was evaluated on MCMED, described as a newly released clinical-grade dataset comprising over 10 million paired PPG-ECG samples from more than 118,000 emergency department visits with expert-labeled cardiovascular disease annotations. The detailed description specifies 118,385 adult ED visits, Lead II ECG and finger-PPG at 128 Hz, expert ICD-10 labels for six CVD categories, and official splits of 10.36 M train, 1.30 M validation, and 37 K test segments (Fang et al., 24 Sep 2025).

On MCMED signal-level metrics, PPGFlowECG reported MAE = 0.73 mV, RMSE = 1.14 mV, Fréchet Distance = 43.99, FID = 12.84, and HR-MAE = 1.80 bpm. Against cited baselines, CardioGAN had MAE = 0.98 and RMSE = 1.40, DDPM had RMSE = 1.36, and RectifiedFlow had FD = 104.35. For multi-label CVD detection on MCMED using Net-1d, AUROC values were 0.708 for I48, 0.626 for I71, 0.622 for I70, 0.619 for I44, 0.608 for I25, and 0.604 for I50, yielding Macro-AUROC = 0.631, compared with 0.608 for DDPM and 0.499 for vanilla flow. On MIMIC-AFib, a VGG-13 classifier using synthetic ECG achieved Accuracy = 0.82, F1 = 0.87, and AUROC = 0.87; in external zero-shot validation on MIMIC-AFib, Macro-AUROC = 0.83 without fine-tuning, with specialist performance reported as 0.87 (Fang et al., 24 Sep 2025).

Clinical realism and diagnostic relevance were also evaluated directly. In a cardiologist Turing test with $L$ 6, using 25 real and 25 synthesized ECGs, average accuracy was 0.52 and specificity was 0.19, which the paper interprets as indicating high perceptual realism. For AF screening, the reported F1 scores were 0.77 for PPG alone, 0.93 for PPG plus real ECG, and 0.94 for PPG plus generated ECG. Explainability analysis with Grad-CAM on Net-1d showed near-identical diagnostic focus for real and generated ECGs, including focus on the QRS region for AF (Fang et al., 24 Sep 2025).

This formulation makes semantic alignment central. Earlier methods largely learned a direct waveform mapping or an implicit conditional distribution; PPGFlowECG instead inserts an explicitly shared latent manifold before generation, thereby separating cross-modal correspondence from transport.

5. Physiology-guided, flow-matching, and multi-lead extensions

A prominent extension is PG-LRF, which argues that existing methods rely mainly on statistical alignment and data-driven generation and do not explicitly structure the latent space around physiology-aware electro-hemodynamic factors. PG-LRF introduces an electro-hemodynamic simulator based on a two-dimensional limit-cycle oscillator with shared cardiac phase dynamics, an ECG readout with P/Q/R/S/T phase components, and a PPG readout with pulse arrival delay $L$ 7. A Physiology-Aware AutoEncoder learns a shared latent manifold using reconstruction, KL, global posterior alignment, instance-level InfoNCE, weak cross-decodability, and a simulator-informed phase-delay consistency term. A PPG-conditioned latent rectified flow is then trained with both the rectified-flow loss and simulator residual losses on ECG and induced PPG beats (Wang et al., 9 May 2026).

On MC-MED, PG-LRF reported improvements over PPGFlowECG in ECG waveform fidelity: MAE 0.87→0.71, RMSE 1.32→1.07, FID 44.17→37.79, and HR-MAE 7.08 bpm→3.94 bpm. It also reported morphology and interval improvements, including PR MAE 68.9→59.9 ms, QRS 48.1→38.6 ms, QT 32.5→29.3 ms, QTcF 35.9→32.2 ms, and ST-J60 0.16→0.14. For disease classification, macro-AUROC improved from 0.545 for a PPG-only baseline to 0.633 for PG-LRF. Ablations showed that removing simulator guidance or the Physiology-Aware AutoEncoder degraded MAE to 0.84 and 0.85, respectively (Wang et al., 9 May 2026).

Flow matching also appeared in a more general framework, PENGUIN, which extends deep state-space models to reconstruct multiple vital signs as continuous waveforms from PPG. For the ECG task, PENGUIN models $L$ 8 with an optimal-transport conditional flow-matching ODE and a dual-stream Flow-SSM backbone. At inference it solves the ODE with Heun’s method using 25 steps. On PPG-DaLiA and WildPPG, the ECG-specific instantiation reported HR Error 15.64 bpm and 12.97 bpm, respectively, outperforming CycleGAN, RDDM, RespDiff, and PaPaGei-S. An ablation removing per-timestep PPG conditioning increased HR Error from 15.64 to 24.40 on DaLiA and from 12.97 to 21.75 on WildPPG (Suzuki et al., 23 Jan 2026).

A separate extension moves from single-lead synthesis to multi-lead translation. The demographic-aware diffusion framework P2Es targets wearable PPG to 12-lead ECG reconstruction. Its forward process combines frequency-domain blurring and temporal noise injection, while the reverse process uses multi-scale temporal recovery followed by frequency deblurring. Demographic adaptation is handled by KNN clustering over chest-lead Pearson matrices and demographic features, with contrastive assignment of affinity matrices. On MIMIC-IV, P2Es reported overall MSE 0.0902 versus 0.1078, DTW 0.0583 versus 0.0792, variance 0.4647 versus 0.9583, KL 0.4920 versus 0.7664, Wasserstein 8.09 versus 9.24, and electrical-axis similarity $L$ 9 0.9542 versus 0.8718 relative to cited baselines; against an upper-bound Lead-I→11-leads model with best MSE 0.0814, P2Es achieved approximately 89% of that fidelity using only PPG (Ji et al., 29 Sep 2025).

These extensions indicate two active frontiers. One is stronger inductive structure, exemplified by electro-hemodynamic simulators and phase-delay constraints. The other is broader output scope, exemplified by general vital-sign reconstruction and 12-lead ECG generation.

6. Evaluation logic, misconceptions, and open problems

The literature evaluates PPGFlowECG with heterogeneous criteria. Early reconstruction work emphasized waveform-level correlation and relative RMSE; the 2019 DCT model used rRmse and Pearson correlation, while the 2020 lightweight network likewise reported $L_x$ 0 and rRMSE (Zhu et al., 2019, Li et al., 2020). Later generative work introduced heart-rate errors, Fréchet-type distances, classifier-centered AUROC and F1, selective-classification AURC, and cardiologist studies (Kong et al., 2024, Belhasin et al., 2024, Fang et al., 24 Sep 2025). This suggests that direct numerical comparison across papers must be made cautiously, because the target datasets, sampling rates, supervision regimes, and endpoints differ materially.

A recurrent misconception is that wearable PPG can be treated as diagnostically equivalent to ECG without translation. The later PPGFlowECG formulation states the opposite: PPG offers accessible, continuous monitoring but “lacks definitive electrophysiological information, preventing conclusive diagnosis,” which is precisely why PPG-guided ECG generation is proposed (Fang et al., 24 Sep 2025). Another misconception is that morphology preservation follows automatically from heart-rate agreement. Multiple papers argue otherwise by adding diagnosis-oriented feature losses, spectral constraints, interval metrics, or simulator consistency terms to preserve QRS, T-wave, ST-segment, and related semantics (Li et al., 2020, Kong et al., 2024, Wang et al., 9 May 2026).

Limitations are also recurrent and explicit. The 2019 pipeline requires accurate cycle detection and was only tested on clean high-quality PPG and ECG; performance may drop in subject-independent settings and under motion artifacts (Zhu et al., 2019). The runtime-monitoring framework assumes reliable event detection and therefore requires pre-filtering when signal quality is poor (Panda et al., 2022). The lightweight neural model documented a measurable subject-independence gap before fine-tuning, and the motion study showed that accelerometry remains relevant even in deep models (Li et al., 2020). ADSSM, diffusion, and flow-based systems address some of these issues through probabilistic priors, posterior sampling, or cross-modal latent alignment, but their reported results still depend on dataset-specific preprocessing and labeling conventions (Vo et al., 2023, Belhasin et al., 2024, Fang et al., 24 Sep 2025).

The future directions stated in the corpus are correspondingly concrete. They include nonlinear regression and more DCT basis functions for sharp QRS recovery, global subject-independent models with more subjects and data augmentation, robust denoising and artifact removal, additional PPG features such as pulse arrival time and derivative waveforms, true SDE noise to model aleatoric uncertainty, variable-length windows and real-time streaming inference, cross-population generalization, and multi-lead ECG synthesis (Zhu et al., 2019, Suzuki et al., 23 Jan 2026, Wang et al., 9 May 2026, Ji et al., 29 Sep 2025). A plausible implication is that the field is converging on hybrid systems in which waveform fidelity, latent alignment, physiological plausibility, and task-level diagnosis are optimized jointly rather than sequentially.