Summary
|
Post-traumatic stress disorder (PTSD) lacks objective, physiology-based tools for assessing severity and treatment response, despite strong links between PTSD, disrupted sleep, and elevated cardiometabolic risk. Sleep is a uniquely informative window because it captures hyperarousal, nightmares, and fragmentation without conscious modulation. This project leverages overnight polysomnography (PSG)—multichannel EEG, ECG, respiration, SpO, and movement—to discover interpretable biomarkers of PTSD from nonlinear, cross-modal dynamics. We will compute multiscale recurrence-analysis (RA) features from more than 10,000 publicly available, fully annotated PSGs and train disentangling variational autoencoders to compress them into a low-rank latent space. The objective is a representation whose axes align with physiologic motifs (arousal load, autonomic tone, respiratory instability) and remain stable across cohorts and devices. Using the learned encoder, we will map latent coordinates to continuous PTSD severity with heteroscedastic Gaussian processes, yielding calibrated predictions and principled uncertainty at the subject level. Decoder-side sensitivity analyses and feature attributions will link each salient latent factor back to specific RA descriptors and source channels, producing clinician-readable “factor cards.” Deliverables include an interpretable encoder, a vetted biomarker library tied to concrete physiology, and versioned, containerized pipelines with fairness and calibration reports. The resulting framework offers a replicable path to trustworthy, objective PTSD assessment and a template for biomarker discovery in other complex physiological disorders. |
Team
|
Cheng-Bang Chen (Industrial and Systems Engineering), Diane Lim (Sleep Medicine)
|
