Common-Mode and Differential-Mode Bilateral Imaging of Shock Microcirculation

When a critically ill patient's blood pressure and cardiac output have been restored to target but the smallest blood vessels in the tissues are still not delivering oxygen, the patient can keep deteriorating while looking stable on the monitor. Clinicians call this macro-microcirculatory dissociation. Today the main bedside window onto it is skin mottling, a patchy purple discoloration around the knee that is graded by eye. That grade predicts death and cardiovascular collapse better than blood pressure does. It is also subjective, and it has been unusable in patients with dark skin. Both landmark mottling studies excluded those patients outright. We propose to replace the eyeball grade with paired camera measurements of both legs. The left and right legs are read the way an electrical engineer reads a two-wire signal. The part the two legs share (the common mode) carries systemic shock, because shock clamps down peripheral vessels on both sides at once. The part on which they differ (the differential mode) carries focal problems such as an arterial injury or a compartment syndrome in one limb. Skin pigment is roughly constant along a leg, so a reference patch on the upper thigh of the same limb removes much of what pigment and individual anatomy add to the common mode.

The idea fixes a known failure of symmetry-based medical imaging. Every existing bilateral method uses the healthy side as a control and subtracts it, which silently cancels disease that is present on both sides. That is exactly what happens in shock. The setting is also a natural fit for the other two constraints. Mottling has reticulated, low-contrast borders that no annotator can trace reliably, and critical-care cohorts with imaging number in the tens to low hundreds. We therefore do not segment the mottled pixels. We estimate a single quantity per limb: how far the hypoperfused region extends from the knee toward the groin. That is what the clinical 0–5 score already encodes, and it can be learned from a few hundred patients.

Core Claim

Bilateral hyperspectral measurements of the legs can be decomposed into three parts: a common-mode perfusion gradient, normalized within each limb, that tracks systemic microcirculatory failure; a differential-mode residual that tracks focal limb ischemia once it is calibrated against healthy left–right asymmetry; and a per-limb "hypoperfusion frontier" that replaces pixel segmentation. Together these give a mottling-equivalent index that matches or beats the visual score at predicting adverse outcomes in light-skinned patients, and keeps that performance in dark-skinned patients, where the visual score cannot be obtained. The claim is about the measurement design first. A learned model is included only if it beats a hand-crafted version of the same decomposition.

Concretely, let $z_{k,s}$ be a vector of hyperspectral perfusion indices at limb $k \in \{L, R\}$ and site $s$ along the leg. The indices are tissue oxygen saturation $StO_2$, the near-infrared perfusion index, the tissue haemoglobin index and the tissue water index, all computed per pixel by the commercial camera. Let $s=0$ be a proximal thigh reference and $s=1$ the knee. The two modes are

$$ c = \tfrac{1}{2}\big[(z_{L,1}-z_{L,0}) + (z_{R,1}-z_{R,0})\big], \qquad d = \tfrac{1}{2}\big[(z_{L,1}-z_{L,0}) - (z_{R,1}-z_{R,0})\big]. $$

Here $c$ is the shared distal-versus-proximal deficit and $d$ is the side-to-side difference in that deficit. Under circulatory triage (the body sacrificing skin and muscle perfusion to protect heart and brain), $c$ should become more negative while $d$ stays near zero. A unilateral vascular injury should push $|d|$ outside the range of normal asymmetry. Pigment adds a nearly site-invariant offset to each limb, which the within-limb subtraction removes from both terms to first order.

Grounding in Prior Work

Three lines of evidence motivate the clinical target. The mottling score predicts 14-day mortality in sepsis independently of vasopressor dose, with an odds ratio of 2.26 per point and a C-statistic of 0.90 in development and 0.76 on external validation (Mottling score is a strong predictor of 14-day mortality…). In 120 intubated ICU patients, mottling predicted post-intubation hemodynamic instability with AUC 0.72, against 0.58 for mean arterial pressure and 0.54 for cardiac output (Peripheral tissue hypoperfusion predicts post intubation hemodynamic instability). That study excluded 11 dark-pigmented patients, and the sepsis study likewise relied on visual contrast. Contact-free hyperspectral imaging with the TIVITA system detected lower palmar $StO_2$ (55.9% vs 65.8% in volunteers) in trauma patients, including those with low injury severity, and added no significant time to the resuscitation workflow (Implementation of hyperspectral imaging in a trauma resuscitation room). That trial averaged one hand-drawn palm region, compared neither sides nor sites, and reported group differences rather than per-patient prediction. Our proposal keeps its hardware and adds the paired, two-axis design. Neuromonitoring work points the same way: focal tissue-oxygen reactivity carries outcome information that global pressure reactivity misses, with an odds ratio of 11.3 for 6-month mortality (Revisiting the oxygen reactivity index in traumatic brain injury). This supports measuring local oxygenation against a within-patient reference rather than against population thresholds.

Bilateral imaging methods supply the technique and also the constraint. Siamese networks with shared weights compare mirrored halves for knee osteoarthritis (Automatic Knee Osteoarthritis Diagnosis from Plain Radiographs). Contralateral subtraction after spatial-transformer alignment improves chest X-ray detection (Contralaterally Enhanced Networks for Thoracic Disease Detection). Cross-hemispheric attention segments faint infarcts on CT (Symmetry-Enhanced Attention Network for Acute Ischemic Infarct Segmentation). A difference loss penalizes symmetric embeddings for femoral-head necrosis (AsT). All four treat asymmetry as the signal, and the AsT authors themselves note that symmetric disease defeats this design. The binocular diabetic retinopathy network (Two Eyes Are Better Than One) is the one method that exploits shared bilateral signal, through an inter-eye correlation of about 0.85. It does so implicitly, inside a contrastive loss, for a disease that is always systemic. We make the split between shared and differing signal explicit, so one model serves both systemic and focal states. We add a within-limb proximal reference that none of these methods has. And we confront a direct challenge to the symmetry assumption: paired organs have real physiological asymmetry (Fineide et al.). We do not assume $d=0$ in health. We estimate the distribution of $d$ in healthy volunteers and score focal disease only as departures from that distribution.

Boundary-ambiguity work constrains the output design. Polyp segmentation networks with reverse and boundary attention (MISNet, Parallel Res2Net with Reverse Attention) depend on hundreds of clean, compact masks, which reticulated mottling cannot provide. Bounding Maps (Bounding Maps for Universal Lesion Detection) showed that replacing hard binary targets with linearly decaying soft targets improves supervision near fuzzy edges. We borrow that idea in one dimension: the target is a soft distribution over frontier position along the knee-to-groin axis. On the small-sample side, the echocardiography pipeline of Dong et al. succeeded on 45 volumes by confining a network inside a strong anatomical prior. We follow its lesson by feeding the model physiological indices computed by the camera rather than raw 100-band spectra, and by tying both legs to one shared encoder.

Research Program

We will run a prospective two-site ICU observational study, targeting roughly 150 to 200 patients with sepsis, septic shock or undergoing intubation. Enrolment will be stratified so that at least a third have Fitzpatrick skin types V–VI; this stratification is the primary equity endpoint. Each patient is imaged with the TIVITA camera at both knees and both proximal thighs at enrolment, at 6 hours and around intubation. Standard macro-hemodynamics, lactate, capillary refill time and, where obtainable, the visual mottling score are recorded alongside. A separate group of healthy volunteers, again stratified by skin type, is imaged with the same protocol to estimate the normal spread of $c$ and $d$. A small trauma subgroup with documented unilateral limb vascular compromise supplies positives for the differential mode.

We build two versions of the index. The hand-crafted version computes $c$ and $d$ from fixed rectangular regions and fits a regularized logistic model. The learned version uses one encoder with shared weights for both legs and all sites, operating on index maps. Its features are combined only through the sum-and-difference structure above, and it predicts a soft frontier position per limb. Frontier supervision comes from visual scores in light-skinned patients. Outcome heads are supervised by 14-day mortality and post-intubation instability, which are available for every patient. That means dark-skinned patients contribute outcome labels without needing visual labels, which is what makes the equity test meaningful. All evaluation uses patient-level nested cross-validation with the outer folds stratified by skin type.

Baselines

The naive reference is mean arterial pressure and cardiac output alone. These are the macro-hemodynamic variables whose failure defines the problem. In the intubation cohort they reached AUC 0.58 and 0.54 for instability (Peripheral tissue hypoperfusion…). We refit them on our cohort, with the same folds, as a floor.

The strongest existing clinical comparator is the visual mottling score combined with lactate and capillary refill time. Mottling reached AUC 0.72 for post-intubation instability in that cohort, and the sepsis mortality model reached a C-statistic of 0.76 on external validation (Mottling score…). In light-skinned patients we refit this model on identical folds. Beating it means a higher AUC with overlapping-or-better calibration, tested by DeLong's comparison on the pooled out-of-fold predictions. In dark-skinned patients this comparator is unavailable by construction. There the bar is an AUC within 0.05 of the light-skinned AUC, with the gap confidence interval excluding a larger loss.

The strongest imaging comparator is the protocol of the trauma imaging trial (Implementation of hyperspectral imaging…): a single averaged hyperspectral region per patient, here the knee, entered as raw $StO_2$ and perfusion index. That trial reported group means, not a predictive AUC, so no reported result was found on Lacuna for hyperspectral outcome prediction. We will produce this number ourselves. It is the most important comparison, because it isolates whether the two-axis design adds anything over simply buying the camera.

Two ablations isolate the key ingredient. The first is the subtraction-only bilateral model in the style of CEN and AsT, which keeps $d$ and discards $c$. The second drops the proximal thigh reference, which keeps left–right pairing but loses pigment normalization. Every learned variant gets the same encoder size, the same 50-trial hyperparameter search inside the inner folds, and the same training budget.

Why not just…?

Why not just average the knee's hyperspectral oxygenation and feed it to a logistic regression? This is the real threat, and it may be enough in light skin. Our argument against it is specific to pigment and inter-patient variance. Raw $StO_2$ in the trial differed by about 10 points between groups, while melanin absorption and baseline anatomy shift individual readings by comparable amounts. Only a within-patient reference removes that shift. The single-region baseline above settles the question directly. If it matches the two-axis index in both skin strata, we will report that the camera alone suffices and drop the learned model.

Why not just train a standard single-image network to regress the mottling score from knee images? With a few hundred patients, an unpaired network can learn skin tone as a proxy for whether a visual label exists and for outcome. Chest X-ray work warns that small cohorts invite exactly such shortcut learning ([paper). The no-reference ablation tests this, as does a per-stratum calibration check. If the unpaired model's dark-skin AUC holds up, the concern is unfounded and we will say so.

Why not just use existing bilateral subtraction architectures? Because they discard the shared signal that defines systemic shock. We expect the subtraction-only ablation to keep its focal-limb sensitivity but lose most of its mortality and instability prediction in patients with $|d|$ in the normal range. If it does not, our core mechanism is wrong.

Proposed Experiments and Decisive Checks

The first check runs before any patient is enrolled, on public data. On EyePACS we compare a sum-and-difference head against the binocular network. That network raised average classification accuracy from 48.20% to 52.12% with a ResNet-50 backbone, and reached 55.13% with BiRA-Net (Two Eyes Are Better Than One). We use the ResNet-50 backbone and the reported split. If explicit common-mode pooling does not match the 52.12% figure, the decomposition is not carrying the information we claim, and we would rethink the design before data collection.

The clinical decisive checks are as follows. The common-mode index must beat mean arterial pressure and the single-region baseline for instability and mortality, specifically within the subgroup whose mean arterial pressure is at least 65 mmHg. That subgroup is the dissociated population and the reason the study exists. Performance must hold within 0.05 AUC in Fitzpatrick V–VI patients. The soft frontier estimate must correlate with the visual score in light skin. Its change over the first 6 hours should track survival, as visual mottling change does in the sepsis cohort. Finally, $|d|$ must separate the unilateral vascular-compromise patients from both healthy volunteers and systemic-shock patients. Systemic-shock patients with severe common-mode deficits should show differential values inside the volunteer range. That pattern, a large shared deficit with an unremarkable side-to-side difference, is the signature that every subtraction-based bilateral method would erase.