Deep Latent Variable Modeling of Physiological Signals

TL;DR

This work develops and evaluates deep latent variable approaches for cardiovascular and neurophysiological data. It combines probabilistic graphical models with deep learning to address high-dimensional, multimodal signals, presenting three core lines: (1) PPG-to-ECG translation with an attention-based deep state-space model enabling continuous AFib detection from wearable data, (2) EEG signal modeling that integrates GANs with graphical structures to yield interpretable latent representations and unsupervised epilepsy detection, and (3) joint neurocognitive modeling that links EEG and behavior through a Neurocognitive VAE capable of trial-level inference of cognitive parameters. Collectively, these methods advance interpretable, data-efficient, and deployable approaches for biomarker discovery, cross-species insights, and translational neurocognitive analysis, with implications for clinical diagnostics and consumer health monitoring. The work emphasizes probabilistic inference, structured representations, and task-relevant generative capabilities, aiming to translate complex physiological signals into actionable biomarkers and cognitive theories. Future directions include robustness to out-of-distribution data, scalable message-passing in DL contexts, and physics-informed priors to further enhance generalization and safety in clinical deployments.

Abstract

A deep latent variable model is a powerful method for capturing complex distributions. These models assume that underlying structures, but unobserved, are present within the data. In this dissertation, we explore high-dimensional problems related to physiological monitoring using latent variable models. First, we present a novel deep state-space model to generate electrical waveforms of the heart using optically obtained signals as inputs. This can bring about clinical diagnoses of heart disease via simple assessment through wearable devices. Second, we present a brain signal modeling scheme that combines the strengths of probabilistic graphical models and deep adversarial learning. The structured representations can provide interpretability and encode inductive biases to reduce the data complexity of neural oscillations. The efficacy of the learned representations is further studied in epilepsy seizure detection formulated as an unsupervised learning problem. Third, we propose a framework for the joint modeling of physiological measures and behavior. Existing methods to combine multiple sources of brain data provided are limited. Direct analysis of the relationship between different types of physiological measures usually does not involve behavioral data. Our method can identify the unique and shared contributions of brain regions to behavior and can be used to discover new functions of brain regions. The success of these innovative computational methods would allow the translation of biomarker findings across species and provide insight into neurocognitive analysis in numerous biological studies and clinical diagnoses, as well as emerging consumer applications.

Figures (20)

• Figure 1: Examples of different probabilistic graphical models.
• Figure 2: A PPG-ECG waveform pair. PPG signals can often become contaminated by noise.
• Figure 3: The graphical model for ECG translation from PPG. Shaded nodes represent observed variables. Clear nodes represent latent variables. Diamond nodes denote deterministic variables. Variables $\boldsymbol{x}_t, \boldsymbol{y}_t$, and $\boldsymbol{c}_t$ represent PP intervals, RR intervals, and context vectors, respectively. $\alpha_{t,i}$ are attention weights defines how well two intervals $\boldsymbol{x}_i$ and $\boldsymbol{y}_t$ are aligned. The attention mechanism is shown only at time step 2.
• Figure 4: The graphical model at latent state inference time. Variables $\boldsymbol{y}_t, \boldsymbol{h}_t, \boldsymbol{g}_t$, and $\boldsymbol{z}_t$ represent respectively RR intervals, backward, forward recurrent states, and latent states.
• Figure 5: Examples of the translated ECG signals. In each subfigure: the top panel shows the input PPG waveform and the bottom panel shows the reconstructed ECG waveform compared with the reference waveform. The average ECG waveform (dark blue) of all possible pulses overlaid on each individual pulse (light blue).