Distributed Fiber Optic Sensing Platform

A pair of woven polymer-optical-fiber mats, joined at the head, that unfold across a sleep surface like a catamaran. Distributed acoustic and strain sensing along the fibers captures respiration, heart mechanical signal, limb movement, pressure distribution, and — preliminary — gross EEG-equivalent through skull-surface vibration.

The interrogator is a rackmount box today; the research question is whether it can be reduced to a shoebox that ships in the duffle with the mats. The cargo is one-night polysomnography-grade recordings from a user's own bed, at a price and a privacy envelope that no current sleep-clinic instrument can hit.

• Distributed acoustic sensing is cheap enough for non-infrastructure use. DAS on telecom fiber is now being used to detect earthquakes, whales, and intrusions. The cost curve per channel dropped an order of magnitude between 2018 and 2024. It is finally in the range where a home-deployable device is imaginable.
• Polymer optical fiber handles mechanical deformation. Glass-fiber DAS is rigid and fragile; POF is cheap, woven-fabric-compatible, and survives being slept on, washed, and folded. The engineering trade is sensitivity (POF is worse per channel) against deployment envelope (POF is usable in a bed).
• The recording density replaces the sensor stack. A full PSG exam today is fifteen electrodes, belts, thermistors, and oximeters. A fiber mat sees everything those sensors capture (plus some they don't) from a single strain-and-acoustic substrate. Fewer consumables. No adhesives.
• Home-only deployment solves a real problem. First-night-effect in a sleep lab is a well-documented confound. The recordings produced by this system are by construction in-bed, in-home, multi-night, normalized against the subject's own baseline.

The mat produces a 2,400-channel × 1 kHz acoustic + strain tensor, continuously, for the duration of a recording. This is ~10 GB per night uncompressed. The novel and difficult piece is that the physiological signals of interest are all embedded in this tensor at different spatial and temporal scales — respiration is a slow, large-amplitude pressure wave; heart mechanical signal is a faster, localized thoracic strain; limb movement is sharp and spatially discrete; gross EEG-equivalent lives at the boundary of what the fiber can resolve at skull contact.

Clinical workstations are not built for tensor data. The display model is as much of a research question as the hardware is. Early work with spectrograms per body region shows promise; a trained clinician can read the four-zone (head, thorax, pelvis, legs) spectrogram stack with ~20 min of training and extract the PSG staging from it.

Eleven in total; these are the four category labels, each with one representative.