Surgeons increasingly judge whether tissue is receiving enough blood using cameras that record many colour bands at video rate. These multispectral or hyperspectral cameras read the chemistry of blood in tissue without an injected dye. Models trained on small groups of patients, typically ten, have not transferred to new patients. When SPECTRALPACA recorded ten volunteers during stepwise arm-cuff occlusion, the spectra grouped by who the person was more than by how much blood flow was cut off. The best working system, Maktabi et al., gave up on learning across patients. It fits a model to a few seconds of each patient's own healthy tissue and flags anything that deviates. That works when a clean healthy recording exists. It does not work in emergencies such as bowel ischemia or trauma, where the tissue is already compromised when the camera arrives.
This proposal asks a narrower question that can be tested. Does the "who the patient is" signal occupy different spectral directions from the "how ischemic the tissue is" signal? If it does, a new patient's offset can be estimated from any frames, including ischemic ones, and removed. A population model can then read the perfusion state with no healthy baseline. If it does not, the ten-subject regime is fundamentally baseline-dependent, and the field should know that. The sample-size question sits at the centre of the study: we measure how many training subjects are needed to learn the patient directions well enough to match the per-patient baseline method.
Core Claim
We hypothesize that the subject clustering in SPECTRALPACA reflects a large but low-dimensional between-subject offset, not subject-specific responses to ischemia. The offset would come from skin pigment, tissue thickness and resting blood volume. It would be larger than the within-subject change caused by occlusion, which is why a principal component analysis (PCA, a method that finds the directions of largest variance) is dominated by it. But it would lie in directions that are nearly separable from the direction in which spectra move as oxygenation falls.
This hypothesis also reconciles two findings that seem to conflict. For static organs, HeiPorSPECTRAL found that inter-subject variability was only 2.3% of spectral variance after ℓ1-normalization, which divides each spectrum by its total so only its shape remains. Organ-type differences are large, so the subject share looks small. Perfusion changes within one tissue are small, so the same subject offset dominates. The claim is falsifiable. If subject-specific response directions carry most of the state information, the approach fails, and we will report that.
Formally, we model an ℓ1-normalized spectrum from subject $i$ at time $t$ as
$$ s_{i,t} = m + U z_i + V\,h(p_{i,t}) + \epsilon_{i,t}, $$
where each term is defined as follows:
- $m$ is the population mean spectrum.
- $U$ is a low-rank basis of between-subject offsets, and $z_i$ is subject $i$'s coordinates in it.
- $V$ is a basis for perfusion-state change.
- $h(p)$ is a monotone function of occlusion pressure $p$.
- $\epsilon$ is residual noise.
The separability of $U$ and $V$ is measured by their principal angles, the angles between the two subspaces. If those angles are large, then $z_i$ and the state can be jointly estimated from frames of any state by least squares. This is the step that removes the need for a healthy baseline. The estimate is well-posed only when the stacked matrix $[U\ V]$ is well-conditioned, which is why the angles are the quantity to measure.
Research Program
The study has three parts.
Part 1: characterize the variance structure. On SPECTRALPACA (10 subjects, 16 bands, about 200,000 frames with graded pressure labels), we fit the model above with a mixed-effects factor analysis. We then report four things:
- the rank of $U$;
- the share of within-subject state variance that lies inside versus outside $\mathrm{span}(U)$;
- the principal angles between $U$ and $V$;
- whether a subject-by-state interaction term improves held-out fit. An interaction would allow $V$ to rotate per subject, for example because pigment scales haemoglobin contrast multiplicatively.
We run the same decomposition on HeiPorSPECTRAL as a static positive control. This tests the reconciliation argument directly.
Part 2: build the simplest estimator the structure allows. The linear version learns $U$ and $V$ on training subjects. At test time it solves for $z_i$ and the state from a short window of frames. This is essentially nuisance-subspace projection, and we treat it as the primary method, not as a strawman. A nonlinear extension is added only if Part 1 shows a significant interaction. In that extension, a small conditional network lets the state readout depend on the estimated $z_i$. It is regularized to collapse to the linear model.
Part 3: measure the sample-size curve. We train on $k = 2,\dots,9$ subjects and plot held-out error against $k$, with a subject-level bootstrap. Estimating a rank-$r$ subspace from $k$ subject means needs $k > r$, and subspace error typically shrinks roughly as $1/\sqrt{k}$ for a fixed spectrum of eigenvalues. Extrapolating this curve tells future trials how many subjects to recruit before a baseline-free model matches personalized calibration.
Throughout, band selection, rank and regularization are chosen inside nested leave-one-subject-out loops. With ten subjects, post-hoc selection would reproduce the inflated-effect problem (the "Winner's Curse") that the connectome literature documents (Mendes et al.).
Baselines
Pooled model, no subject handling (naive reference). An ℓ1-normalized pooled regressor from spectrum to pressure: ridge regression and a small multilayer perceptron, trained on nine subjects and tested on the tenth. This is the "one-size-fits-all" approach that SPECTRALPACA's PCA suggests will fail. Any method must beat it on held-out subjects. No reported result found on Lacuna for cross-subject pressure regression or ischemia classification on SPECTRALPACA; the paper reports only clustering and sensitivity to oxygenation dips "across all pressure levels above 20 mmHg."
Personalized healthy-baseline anomaly detection (strongest existing method). The per-patient approach of Maktabi et al. uses an ensemble of invertible neural networks, generative models that give exact likelihoods. It fits them to each patient's first seconds of perfused tissue and scores deviation as an "ischemia index." On their 10-patient nephrectomy cohort it reached a median AU-ROC of 1.0. It is the strongest competitor because it sidesteps cross-subject variance entirely.
We reimplement it on SPECTRALPACA using each subject's 0 mmHg segment as the baseline. We give it the same frame budget that our method uses for offset estimation and tune it with the same nested budget.
It is expected to be near-ceiling when a clean baseline exists, so we compare against it in two ways:
- Matched-information comparison. With a clean baseline, beating it means matching its AU-ROC within the bootstrap interval while using no subject-specific training.
- Corrupted-baseline comparison. The "baseline" window is drawn from frames already at 60 mmHg or higher. Beating it here means a statistically significant AU-ROC advantage.
Oracle subject centering. Subtract each test subject's true 0 mmHg mean spectrum, then apply the population state readout. This upper bound shows how much of the gap our baseline-free offset estimate closes.
Ablation of the key ingredient. The same pipeline, with $U$ replaced by a random subspace of equal rank, and separately with $U$ omitted. This isolates whether the learned between-subject subspace does the work.
Why not just…?
Why not just ℓ1-normalize and train a pooled classifier? ℓ1-normalization already removes intensity effects. SPECTRALPACA reports that normalized data still cluster by subject, and HeiPorSPECTRAL's 2.3% figure comes from a static setting where state differences are large. The pooled baseline above tests this directly. If it matches our method, the separability story is unnecessary and we will say so.
Why not just use the per-patient baseline method, since it already hits AU-ROC 1.0? When a clean baseline exists it may well be as good or better, and we do not expect to beat it there. Its stated weaknesses are two. It needs pristine healthy tissue, and it raised false alarms in one of ten patients (Patient 7) under scene changes. The corrupted-baseline experiment is designed to show whether those weaknesses matter in practice. If its performance holds up with a contaminated baseline, our case for baseline-free estimation weakens substantially.
Why not just compute tissue oxygenation by physical unmixing, or reference a healthy region in the same frame? Both are serious threats.
Oxygenation estimated from haemoglobin absorption is designed to be subject-invariant. If subject offsets are mostly pigment and scattering, unmixing might already remove them. We therefore include linear-unmixing oxygenation as a baseline and test whether $\mathrm{span}(U)$ overlaps the unmixing null space. If it does, our learned subspace is rediscovering physics, which is still useful because it tells us what the physics misses.
A same-frame healthy reference, such as the unoccluded region used in ratio methods, requires that some visible tissue be healthy. In SPECTRALPACA the whole palm is occluded, so this baseline cannot be run there, and we state that limitation openly. It remains plausible in surgical scenes with partial ischemia.
Why not physics-informed deconvolution of dye curves instead? Differentiable deconvolution embeds tracer-kinetic equations in a network and generalized across 94 CT perfusion cases, with a Dice of 0.359 versus 0.353 for experts. It addresses dye-bolus timing, not dye-free spectral offsets. Applying it to indocyanine green (ICG) fluorescence video would bring back the roughly 30-minute dye washout problem. That problem is exactly why spectral monitoring is attractive.
Proposed Experiments
The primary benchmark is SPECTRALPACA with leave-one-subject-out evaluation. We report three metrics:
- Spearman correlation between predicted state and cuff pressure;
- AU-ROC for detecting occlusion at 60 mmHg or higher versus 0 mmHg;
- the lowest pressure step detectable per subject, compared against the paper's greater-than-20-mmHg sensitivity.
All intervals come from subject-level bootstrap, since ten subjects is the real sample size, not 200,000 frames.
The secondary checks are HeiPorSPECTRAL, for the static variance decomposition, and the Maktabi nephrectomy cohort if the authors can provide access, for a surgical replication of the corrupted-baseline test.
Decisive checks. Three results would settle the question:
- Separability. The principal angles between $U$ and $V$ in Part 1 must be clearly non-trivial, and the interaction term must add little held-out fit. If both fail, the project reports a negative result: baseline-free transfer is not supported at $N=10$.
- Offset estimation from ischemic frames. In the corrupted-baseline setting, our estimator must close most of the gap between the pooled baseline and oracle centering, and the random-subspace ablation must not.
- Sample size. The learning curve over $k$ must be monotone, and its extrapolation must predict how many subjects are needed to reach oracle-centering performance. That number is the concrete, reusable answer to the sample-size question this intersection raises.
References
- The SPECTRAL Perfusion Arm Clamping dAtaset (SPECTRALPACA) for video-rate functional imaging of the skin
- Spectral imaging enables contrast agent–free real-time ischemia monitoring in laparoscopic surgery
- HeiPorSPECTRAL - the Heidelberg Porcine HyperSPECTRAL Imaging Dataset of 20 Physiological Organs
- A functional connectome phenotyping dataset including cognitive state and personality measures
- Differentiable Deconvolution for Improved Stroke Perfusion Analysis