SIMULATED · NO HUMAN SUBJECTS

Non-Ambulatory Bathing System

Thermographic pressure-injury surveillance for prolonged casualty care

Empowered Disease Management Health Services, LLC UEI FSMJREGJGPJ3CAGE 91US7 Toledo, OhioTRL 3

The problem

Detection of pressure injury during prolonged holding is, today, a tired medic looking for redness under bad light on skin covered in dirt or blood.

USAF aeromedical evacuation, 2001–2014
137,433
patients, 210,863 flights
Immobile on a litter
~19%
transports of 8–16 hours
Prolonged casualty care
10 h
median to transport; range 4–120 h
Burn casualties expected in LSCO
5–25%
with evacuation likely delayed

Prolonged casualty care research has addressed the support surface — the Bridges randomized mattress trial is the most rigorous recent example, and it notes explicitly how little operational research exists. Detection has not been addressed at all. Erythema is also least reliable on darker skin, so the current method fails worst where the population is most diverse.

What we found

We built the simulator expecting the camera to be the constraint. It is not.

Across 633 simulated scan cycles, instrument noise contributes under 5% of the error variance. Over half is physiological reference variance — the fact that ΔT has no meaning until you name what you are subtracting, and every candidate reference carries biological spread plus a precondition that can be false.

error variance share 4.8% instrument NETD, FPN, FFC drift · σ 0.0605 °C 54.9% physiological reference σ 0.2050 °C — the binding constraint 40.3% evaporative artifact σ 0.1758 °C — engineerable

A better sensor buys almost nothing. Constraining the reference buys almost everything — and a fixed enclosure that positions the body identically every cycle is precisely a reference-constraining instrument. The hygiene function is not a feature bolted onto the diagnostic. It is what makes the diagnostic well-defined.

Referencing each site to itself
0.917
measured AUC
Same sensor, baseline ablated
0.581
measured AUC
Published dry-handheld anchor
0.885
Asif et al. 2022, n=197 burns

The anchor is the pooled infrared-thermography burn-depth meta-analysis (sensitivity 0.84, specificity 0.76 → d′ = 1.70). It is never adjusted to improve results. If a bench measurement cannot be reproduced by a simulator that still reproduces the anchor, the model is wrong — not the anchor.

How the scan is ordered

The 8 × 8 tile board over the focal plane is scheduled twice over. A Warnsdorff knight's tour orders the tiles; Warnock recursive subdivision decides how hard to look inside each one. Warnsdorff says where to look next; Warnock says how hard.

The reason for a tour rather than a raster is specific. Over a 93-second scan, FFC drift and enclosure warm-up add an offset that depends on acquisition index. Under raster, acquisition index is an affine function of row — so drift becomes a smooth function of the cephalocaudal axis, which is the same functional form as the perfusion gradient the detector references against. The two stop being separable.

Tile
1 / 64
Onward degree chosen
corr(index, row)
drift aliased onto anatomy

Dashed squares are the admissible successors, each labeled with its onward degree. The amber square is the arg-min. Tick raster order and watch the correlation pin to +0.99 — that single number is the whole argument for the ordering.

How hard to look

Warnock terminates when a quad's thermal variance falls below the sensor noise floor: there is nothing left to resolve that the instrument could distinguish. The homogeneity test runs over body pixels only — the skin-to-air edge is a nine-degree step, and testing it unmasked declares every margin tile heterogeneous, degenerating the subdivision into a uniform raster that spends the whole dwell budget on empty air.

Threshold Θ 0.060 °C
Leaf quads
Scanned / skipped
Dwell on zones

Untick body pixels only to see the degeneration. Production geometry is 160 × 120 with 20 × 15 tiles; this runs the same predicate on a 64 × 64 resampling with the size floor scaled proportionally, so counts differ from the 249-scanned figure.

What the device refuses to say

ΔT is undefined until you name the reference. Four schemes exist, each with a precondition evaluated at runtime — never from ground truth — and the cascade takes the first admissible one. When none holds, the zone is reported indeterminate rather than guessed.

R0  longitudinal   P0: prior baseline exists AND global thermal state unshifted
R1  contralateral  P1: unilateral involvement
R2  cephalocaudal  P2: intact sympathetic vasomotor tone
R3  surround       P3: lesion small relative to the surround annulus

Decidable(z)  ↔  ∃R . Admissible(z, R)
(S1 ∨ DT) → D     no staging call on an inadmissible reference
I → ¬N            an abstention is never recorded as "no injury"

Scenarios one and two are the same sacrum, the same measured ΔT of +0.61 °C, the same sensor — called stage 1 with a baseline on file and no deviation without one. Nothing about the patient changed; only what the device was permitted to subtract from.

What it looks like across a mission

Forty scan cycles through a full bath profile, with a febrile episode from cycle 21. Steam compresses contrast during the wash phase; fever shifts the patient's whole thermal state and withdraws the longitudinal reference. Both are visible, and the device reports them rather than hiding them.

Cycle 26

The raw global-shift trace is steam-biased — at cycle 18 it reads −0.88 °C with no fever at all, 70% of the way to the withdrawal limit on artifact alone, while simultaneously masking a real 1.85 °C arterial rise. Correcting for path transmittance before any cross-cycle comparison fixes both directions.

Where it stands

MeasureValueNote
Nominal cycleAUC 0.945 · d′ 2.26 · σ_eff 0.301 °Cτ 0.921, mild steam and residual water
Full scan cycle time93.5 s705 frames, Lepton 3.5 at 8.7 Hz
Edge inference latency62.2 ms8 zone crops, ResNet-50 INT8, Jetson Orin Nano
Operating envelope0.705 – 0.988378 cycles across steam × residual water
Worst wet condition → mitigated0.743 → 0.864path purge, dry-down gate, reference blackbody
Drift aliasing, raster → tour+0.9923 → +0.0820corr(acquisition index, tile row)
Mission accounting68 correct · 12 missed · 2 wrong-sign80 lesion-zone observations over 40 cycles

What we are not claiming

This is a long-wave infrared instrument. It maps surface and near-surface perfusion. It does not image the peritoneum, the alveoli or the carotid, and no amount of development will make it do so. The output is advisory against a documented protocol — the device proposes, the clinician disposes. No learned component writes to an actuator.

The skin-tone argument is stated carefully: long-wave emissivity of human skin is governed by tissue water rather than melanin, so this modality plausibly does not degrade where visual inspection of erythema degrades worst. That is a physical premise with a mechanism and an open, fundable question. It is not a demonstrated result.

The ask

TRL 3 today. The path to TRL 4 is instrumented hydrogel phantoms with software-controlled perfusion analogues — no human subjects are required to get there. What we want from DARPA is the phantom validation program and access to the prolonged casualty care community that would field this.