Generative AI for PPG Signal Synthesis: GANs, Diffusion Models, and Data Augmentation

Generative AI models create realistic synthetic PPG waveforms that are physiologically plausible, condition-specific, and statistically indistinguishable from real recordings. These synthetic signals address the fundamental data scarcity problem in clinical PPG research, enable privacy-preserving dataset sharing, and support robust model evaluation by generating edge-case conditions that are rare in real-world recordings.

Synthetic PPG generation has moved from simple template-based morphological models to neural generative models that capture the full complexity of inter-individual variability, motion artifacts, breathing modulation, and condition-specific waveform deformations.

Why Synthetic PPG Data Matters

Clinical PPG datasets face a fundamental collection barrier. Every labeled PPG recording requires simultaneous reference measurements (ECG for beat-level labels, PSG for sleep staging, intra-arterial catheter for BP ground truth) and often clinical supervision. The result: even the largest public PPG datasets have fewer than 200 subjects.

Synthetic data augmentation can expand effective dataset size by 5-20x. The critical requirement is fidelity: synthetic signals must preserve the statistical properties that discriminate between physiological states, not merely visual resemblance.

Three specific use cases drive most research:

1. Class imbalance correction: Rare arrhythmias (PVCs, 2nd-degree AV block, ventricular tachycardia) appear at low frequency even in arrhythmia-enriched datasets. Generating synthetic examples of rare conditions balances training distributions.

2. Privacy-preserving sharing: Real PPG waveforms from clinical trials may identify patients (particularly via cardiac biometric fingerprint). Synthetic PPG with similar statistical properties enables data sharing without privacy risk.

3. Out-of-distribution robustness testing: Systematically generating edge-case PPG conditions (extreme motion artifacts, unusual waveform morphologies, sensor failures) that may not appear in real validation sets reveals model failure modes before clinical deployment.

Generative Adversarial Networks for PPG

WaveGAN (Donahue et al., 2019) adapted DCGAN to raw audio waveforms, and the same approach applies to PPG. A 1D generator produces 250-sample (2-second) PPG segments from a noise vector; a 1D discriminator distinguishes real from synthetic segments.

The adversarial training objective produces PPG with realistic spectral characteristics: correct heart rate distribution, physiologically plausible systolic peak sharpness and diastolic notch presence, realistic amplitude variation. However, vanilla GAN training is unstable and may not generate condition-specific signals without conditioning.

Conditional GAN for PPG (CGAN-PPG): Condition the generator and discriminator on clinical labels (heart rate target, SpO2 level, rhythm type). A CGAN trained on MIMIC-IV PPG learns to generate distinct waveforms for sinus tachycardia at 110 BPM, sinus bradycardia at 48 BPM, and AF with irregular rhythm, rather than a blend of all observed patterns.

Golany et al. (2019, Improving ECG Classification Using Generative Adversarial Networks, AAAI workshop) demonstrated that GAN-augmented training improves rare arrhythmia classification by 12-18% F1. The PPG analog shows similar improvements for PVC detection when synthetic PVC segments are added to a predominantly sinus dataset.

Challenges with GAN-based PPG synthesis:

• Mode collapse: Generator converges to producing a narrow range of waveforms despite diverse conditioning inputs
• Training instability: GANs for 1D signals still require careful hyperparameter tuning; Wasserstein GAN with gradient penalty (WGAN-GP) provides more stable training for PPG
• Physiological validity: GAN outputs may pass visual inspection but fail quantitative physiological plausibility tests (e.g., diastolic notch at an impossible time relative to systolic peak)

Variational Autoencoders for Controlled PPG Generation

VAEs (Kingma & Welling, 2013) learn a continuous latent space where semantically similar PPG signals occupy nearby regions. Unlike GANs, VAEs provide stable training and enable smooth interpolation between physiological states.

A PPG-VAE with a 32-64 dimensional latent space can:

• Interpolate between a healthy sinus rhythm waveform and an AF waveform, generating intermediate states that represent borderline rhythm disorders
• Disentangle waveform attributes: separate latent dimensions for heart rate, waveform sharpness, dicrotic notch depth, and motion artifact level
• Condition on clinical variables: using a conditional VAE (CVAE) framework, generate PPG signals specifically for a 35-year-old female with SpO2=94% and mild tachycardia

The limitation of VAEs is blurry generation. The reconstruction objective encourages averaging over possible outputs, producing PPG signals with lower high-frequency detail than real recordings. This can be partially addressed with a perceptual loss that penalizes spectral differences rather than per-sample MSE.

Diffusion Models for High-Fidelity PPG Synthesis

Diffusion models (Ho et al., 2020, DDPM) iteratively denoise a Gaussian noise signal into a realistic sample. They have achieved unprecedented quality for image generation and the architecture translates well to 1D biosignals.

The denoising diffusion probabilistic model for PPG:

• Forward process: gradually add Gaussian noise to a real PPG waveform over T=1000 steps
• Reverse process: train a 1D U-Net to predict and remove the noise at each step
• Generation: start from pure Gaussian noise and run the reverse process T steps

Diffusion models produce higher-fidelity PPG than GANs in terms of:

• Spectral fidelity: Power spectral density of generated signals matches the real data distribution more closely
• Morphological diversity: Generated signals cover the full range of physiological variation including rare morphologies
• Sample quality consistency: No mode collapse; every sample is high quality, not just a fraction

Alcaraz et al. (2023, Diffusion-Based Conditional ECG Generation with Structured State Space Models, Computers in Biology and Medicine) applied structured state space diffusion models (S4-based UNet) to cardiac signal generation and demonstrated physiological plausibility scores comparable to real data across 8 morphological metrics. The approach transfers directly to PPG generation.

Latent diffusion for PPG: Instead of diffusing in the raw 500-sample signal space, encode PPG into a lower-dimensional latent space with a VAE, then run diffusion in latent space. Denoising in 32-64 dimensional latent space is 10-100x faster than denoising in raw signal space while maintaining high output quality.

Classifier-free guidance enables conditioning on clinical attributes without training a separate classifier. By randomly dropping condition labels during training, the model learns both unconditional and conditional generation. At test time, combining conditional and unconditional predictions at scale factor w > 1 produces sharper, more condition-specific outputs. For PPG, guidance on heart rate and rhythm type with w=3-5 generates highly specific waveforms without mode collapse.

Physiological Constraints in Generative PPG Models

Unconstrained neural generators may produce physiologically impossible PPG signals that pass visual inspection but fail consistency checks. Incorporating physiological constraints improves both generation quality and downstream model training value:

Morphological constraints: The diastolic notch must occur after the systolic peak and before the next systolic peak; its timing relative to systolic peak (diastolic notch fraction) should be in [0.3, 0.7] for healthy adults. Waveform valley-to-peak amplitude ratio should be > 0.3 for adequate signal quality. These constraints can be enforced via differentiable validity penalties added to the generator loss.

Inter-beat consistency: Consecutive beat morphologies should be similar for sinus rhythm and more variable for AF. A recurrent structure in the generator (LSTM conditioning on previous beat) enforces temporal coherence.

Spectral plausibility: The fundamental frequency (heart rate) must appear as the dominant spectral peak. A spectral loss term comparing generated and real spectrograms in the 0.5-4 Hz band ensures frequency-domain realism.

Evaluation of Synthetic PPG Quality

Evaluating synthetic PPG requires domain-specific metrics beyond general generative model benchmarks:

Fréchet Inception Distance (FID) analog: Use a PPG classifier (pre-trained on real data) as the "inception" model. Compare distribution statistics of real and synthetic signals in the classifier's penultimate layer. Lower FID indicates better statistical similarity.

Downstream utility: Train a clinical PPG model on synthetic data (or synthetic + real). Evaluate on a real held-out test set. If synthetic data adds value, downstream task performance should improve with more synthetic data. This is the gold-standard evaluation.

Physiological plausibility score: Automated checks for morphological validity (notch timing, peak sharpness, AC/DC ratio), spectral consistency (HR fundamental frequency presence), and condition specificity (does AF-conditioned generation actually show irregular IBI?).

Expert review: Cardiologists and biomedical engineers are asked to distinguish real from synthetic segments in a forced-choice paradigm. High confusion rates (near 50%) indicate generation fidelity.

For context on PPG signal properties these generators must reproduce, see PPG morphology features, PPG waveform decomposition, and PPG noise types and classification. For the augmentation context, see deep learning for PPG heart rate estimation.

Key Papers

• Ho, J. et al. (2020). Denoising diffusion probabilistic models. NeurIPS. https://doi.org/10.48550/arXiv.2006.11239
• Golany, T. et al. (2019). Improving ECG classification using generative adversarial networks. AAAI workshop on Health Intelligence. https://doi.org/10.48550/arXiv.1903.09949
• Alcaraz, J.M. et al. (2023). Diffusion-based conditional ECG generation with structured state space models. Computers in Biology and Medicine, 163, 107115. https://doi.org/10.1016/j.compbiomed.2023.107115
• Donahue, C. et al. (2019). Adversarial audio synthesis. ICLR. https://doi.org/10.48550/arXiv.1802.04208

FAQ

Are synthetic PPG signals good enough to replace real clinical data entirely? Not yet, particularly for rare and high-stakes conditions. Synthetic data is most valuable as a supplement to real data, not a replacement. For common conditions with abundant real training data, synthetic augmentation provides diminishing returns. For rare arrhythmias or minority demographic groups, it can provide transformative benefits. The downstream utility evaluation (train on synthetic, test on real) is the most honest measure.

Can generative models learn from small datasets (fewer than 50 real PPG subjects)? GANs and VAEs typically need 100+ subjects for stable training and physiologically diverse generation. Diffusion models are more data-efficient; latent diffusion with a VAE encoder can produce plausible PPG from 30-50 subjects. Transfer learning (pre-train on ECG or other biosignals, fine-tune on limited PPG) further reduces the real-data requirement.

Do regulatory bodies allow training on synthetic data? The FDA's stance on synthetic data for AI/ML medical devices is evolving. The 2023 FDA guidance on AI in medical devices acknowledges synthetic data as a valid training source when accompanied by rigorous real-data validation showing comparable downstream performance. Synthetic data alone is not sufficient for device clearance; it must be accompanied by real-world performance evidence.

How are diffusion models for PPG different from simple signal augmentation? Standard augmentation (amplitude scaling, time stretching, noise injection) transforms existing signals without generating new physiological patterns. Diffusion models learn the distribution of real PPG signals and generate new samples from that distribution, including combinations of features (heart rate + morphology + rhythm type) not present in the training data. This generative diversity is the key advantage over augmentation.

Can synthetic PPG be used to train on conditions not present in the training dataset at all? With careful conditioning and physiological constraint enforcement, synthetic generation for conditions not seen in training is possible but risky. The generator may extrapolate outside its training distribution in non-physiological ways. Cross-validation with domain experts is essential. Simulation-based generation (using cardiac and vascular physiological models to produce forward-model PPG for a given condition) is more reliable for truly novel conditions.