AI modeling photoplethysmography to electrocardiography useful for predicting cardiovascular disease

Figures

• 

  Fig. 1. Overview of the study design. Stage I: Development of CardioPPG. Using PulseDB53 (include MIMIC-III16 and VitalDB17 dataset) and MESA datasets18, we trained two core components of CardioPPG: (1) a PPG model trained with PPG-ECG contrastive learning to capture relevant information from ECG, and (2) an ECG-generation model trained with an autoregressive model conditioned on PPG. Stage II: Downstream tasks. The dataset used is from the PulseDB-MIMIC dataset. All tasks used only PPG data to validate the model’s performance, including: a screening for cardiovascular diseases, b generation quality assessment of different generation methods.

• 

  Fig. 2. Overall assessment of cardiovascular diseases and label utilization efficiency. a The classification AUC metrics for mitral and aortic valve diseases (MAVD), old myocardial infarction (OMI), paroxysmal ventricular tachycardia (PVT), atrioventricular block (AVB), atrial fibrillation (AF), cardiomyopathy (CM), heart failure (HF), ischemic heart disease (IHD), paroxysmal supraventricular tachycardia (PSVT) and coronary atherosclerosis (CAS) in PulseDB-MIMIC dataset. CardioPPG is compared against masked self-supervised baselines—PPG-only (SSL-PPG) and ECG-only (SSL-ECG)—and exceeds SSL-PPG on all conditions, while matching or surpassing SSL-ECG (regarded as an approximate upper bound) on several. Statistical significance was assessed using the DeLong test (*p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001). b AUC on an external PPG–AF dataset comprising two channels measured on different fingers; we assessed channels separately, and CardioPPG outperformed SSL-PPG on each, demonstrating strong generalization. c Label efficiency was analyzed by measuring the model’s performance with varying fractions of training data, providing insight into the amount of data required to achieve…

• 

  Fig. 3. Multiclass evaluation of arrhythmia. We analyzed three arrhythmia—paroxysmal supraventricular tachycardia (PSVT), paroxysmal ventricular tachycardia (PVT), and atrial fibrillation (AF). Using the PulseDB-MIMIC dataset, we constructed a balanced cohort (1:1:1:1) of normal rhythm and the three arrhythmia subtypes, matched in size to the AF subset. Top panels: One-vs-rest (OvR) receiver operating characteristic (ROC) curves for normal rhythm and each arrhythmia subtype comparing three models: SSL-ECG (green), CardioPPG (orange), and SSL-PPG (blue). Shaded regions indicate 95% bootstrap confidence intervals (n = 1000). Bottom panels: Confusion matrices showing four-class predictions from the three models, illustrating classification performance across normal, PSVT, PVT, and AF categories.

• 

  Fig. 4. Comparison of ECG-generation results guided by PPG across three models: CardioPPG, WGAN, and RecQRS. a Evaluation metrics for generative performance, including Marginal Distribution Difference (MDD, lower is better), Autocorrelation Difference (ACD, lower is better), Skewness Difference (SD, lower is better), Kurtosis Difference (KD, lower is better), Mean Squared Error (MSE, lower is better), Fréchet Inception Distance (FID, lower is better), Mean Absolute Error (MAE, lower is better), Fréchet Distance (FD, lower is better), Structural Similarity Index (SSIM, higher is better), and Peak Signal-to-Noise Ratio (PSNR, higher is better). Error bars denote bootstrap 95% CIs computed with n = 1000 resamples; in a few cases, intervals are very small and may be indistinguishable at the figure scale. Full bootstrap CIs are provided in Supplementary Table 7. b Comparison between generated and real ECG signals using t-SNE, where the distributions of CardioPPG-generated ECGs almost completely overlap with those of real ECGs. c Three cases for qualitative assessment.

• 

  Fig. 5. Model diagram. a The self-supervised model for signal learning is trained using a masking-based self-supervised approach for both PPG and ECG, aiming to obtain higher-quality feature representations. b The PPG-ECG contrastive learning model initializes the PPG and ECG encoders with pretrained model parameters from subfigure a, enabling aligned representations of PPG and ECG in the feature space. c PPG-guided ECG-generation model. (i) An STFT-based dual-branch VQ-VAE encodes low- and high-frequency ECG latent features; (ii) two bidirectional Transformers are trained with masked-token prediction, with the high-frequency Transformer additionally taking the unmasked low-frequency tokens as auxiliary input; (iii) iterative decoding reconstructs low-frequency first, then high-frequency conditioned on it for high-fidelity ECG. d Fine-tuning the aligned PPG encoder to adapt to various downstream tasks and using a pretrained generation model in subgraph (c) generate the corresponding ECG signals.