Hypertension is one of the most common chronic health conditions of our time and a leading risk factor for cardiovascular disease and stroke. Since the condition often causes no noticeable symptoms, regular blood pressure monitoring remains the only reliable means of timely detection. Standard measurement methods are either invasive or unsuitable for long-term continuous monitoring, making research into non-invasive alternatives highly active. One promising direction is the estimation of blood pressure from photoplethysmographic (PPG) signals, already captured by sensors in smartwatches and wearable devices. Despite encouraging results from deep learning approaches, most models operate as black boxes, without clear insight into which morphological features of the signal are decisive.
In this work, we develop an autoencoder with a disentangled latent space for PPG signals, in which each latent dimension corresponds to a single physiologically interpretable morphological parameter of the signal. Based on selected morphological parameters, we generate synthetic signals using a two-Gaussian model and Latin hypercube sampling of the parameters. On the synthetic data, we compare three models: a vanilla autoencoder and a standard variational autoencoder (VAE) without prescribed structure, and a supervised VAE, which uses an additional prediction head to explicitly align individual latent dimensions with the values of the morphological parameters. The trained models are then transferred to real PPG signals without any additional retraining, where we evaluate reconstruction quality and latent space structure of the real PPG signals.
Results on synthetic signals show that the supervised variational autoencoder achieves the most structured and interpretable latent space, while maintaining reconstruction accuracy comparable to the vanilla autoencoder. Intervention analysis further confirms that individual latent dimensions allow for direct interpretation and controlled manipulation of the morphological features of generated signals. The standard variational autoencoder fails at both reconstruction and latent space organization due to posterior collapse. When transferred to real PPG signals, the performance of all models deteriorates, indicating a domain gap between synthetic and real data. Although the supervised variational autoencoder achieves partial overlap between the two domains in the latent space, this does not necessarily lead to better morphological agreement between real signals and their nearest synthetic representations. The results show that supervised guidance of the latent space enables an effective and interpretable representation of the morphological characteristics of synthetic PPG signals, while transferring these representations to real data without additional adaptation remains limited due to the domain gap between the synthetic and real domains.
|
