Long operations cause more infections and complications. The evidence is consistent across procedures: an operative duration over three hours was the strongest predictor of meningitis after vestibular schwannoma surgery, and operation time was the most influential feature in a machine-learning model of cholangitis after biliary stenting. Yet "minutes on the clock" is only a stand-in for whatever actually damages the patient. Two mechanisms are usually proposed. Bacteria may settle on the open wound over time, or tissue may gradually lose blood supply and immune defense. The clock cannot tell these apart, and it cannot tell a long operation on well-perfused tissue from a shorter one on starving tissue.
We propose to replace the clock with a measured quantity, the integrated hypoperfusion burden. This is the accumulated time, weighted by severity, that the tissue in the surgical field spends below a perfusion threshold. It is computed continuously from the video the surgical microscope already records, using remote photoplethysmography (rPPG): recovering the tiny color changes each heartbeat causes in skin or tissue by filtering the video in the frequency band of the pulse. The contribution is methodological and scientific. We build an integral that stays honest when the view is blocked or corrupted, which snapshot measures never had to handle. We then test whether this burden predicts postoperative complications beyond operative duration, and whether it accounts for part of duration's effect.
Core Claim
Two things are true at once. Intraoperative rPPG can already tell perfused from ischemic tissue. Duration-based risk models already work reasonably well. What nobody has done is turn the perfusion signal into a cumulative exposure variable and ask whether it carries the information that duration only approximates.
Our claim has two parts. First, a reference-locked, quality-aware integral of local perfusion loss can be computed reliably through a whole operation. It is reliable in a specific sense: occlusions, smoke, lamp flicker, and ventilator motion add little spurious burden. Second, in free-flap reconstructive surgery this burden adds predictive value for flap complications and surgical-site infection beyond total operative time. We also expect it to absorb a measurable share of the association between duration and those outcomes.
If the second part fails, meaning the burden adds nothing once duration is known, then the clock really is sufficient. That would be a useful negative result for a field that increasingly proposes camera-based monitoring.
Relation to Prior Work
The motivating evidence comes from four studies that treat time as a coarse, after-the-fact variable:
- Risk factors for postoperative meningitis after microsurgery for vestibular schwannoma found an odds ratio of 7.927 for operations over three hours. All 27 meningitis cases were culture-negative, which leaves open whether the damage was bacterial or ischemic and inflammatory.
- Risk prediction of cholangitis after stent implantation based on machine learning ranked operation time as the top feature of a random forest that reached an AUROC of 0.87, or 0.89 after feature selection.
- The FLEXOR registry analysis found total surgical time independently predictive of complications, with an odds ratio of 1.12 per unit of time.
- Outbreak of Sepsis Following Surgery linked procedure duration to maximal organ-failure score (R = 0.933) and used 16S sequencing to show workspace contamination that culture missed.
All four use duration as an unexplained aggregate. We aim to measure one of the physical processes it stands for.
The method comes from intraoperative rPPG, which is also the closest existing work:
- Perfusion assessment via local remote photoplethysmography adapted the plane-orthogonal-to-skin (POS) algorithm to transplanted kidneys and flaps. It computed a perfusion index and a correlation with a reference pulse, and a flat signal in one case preceded necrosis by five weeks. - Continuous intraoperative perfusion monitoring of free microvascular anastomosed fasciocutaneous flaps recorded 15 flap surgeries at 60 fps. It showed a clear difference in pulse magnitude between clamped and reperfused states (p < 0.001), detected reperfusion about 18 seconds after clamp release, and saw flap-to-face correlation rise from near zero to 0.56.
Both papers evaluate discrete states, before clamping versus after release, using manually chosen regions. Neither accumulates anything over time, handles tool occlusion or smoke, or links the signal to later infection. The second paper explicitly notes that safety thresholds remain undetermined.
Our proposal changes the object of measurement from a state check to a time integral. That change creates a problem those papers never faced: every corrupted second now adds to the total.
The signal-processing literature constrains the design. Remote Heart Rate Measurement from Face Videos under Realistic Situations showed that a fixed 0.7–4 Hz bandpass alone performs poorly in realistic video (correlation 0.36). It needed illumination correction and rejection of the most volatile segments to reach 0.81. Segment rejection is exactly what an integral cannot absorb silently.
Contrast-Phys learns rPPG without contact labels by contrasting power spectra across face regions. It reaches 0.64 bpm error on UBFC-rPPG, but it relies on facial spatial structure that surgical fields lack.
Research Program
The measurement pipeline has three pieces.
Reference-locked filtering. The anesthesia pulse oximeter already reports the patient's heart rate. We also film a patch of uncovered skin as a reference, following the flap-monitoring study. Instead of passing everything between 0.7 and 4 Hz, we track the reference fundamental frequency $f_0(t)$ and read local tissue power in a narrow band around $f_0(t)$ and its first harmonic.
This band exploits the fact that other periodic sources rarely share the patient's heart rate, including ventilator motion at a fraction of a hertz, microscope or headlamp flicker, and repetitive instrument sweeps. Dividing local band power by reference band power also separates local perfusion loss from systemic changes such as hypotension or vasoconstriction. Systemic changes affect both regions; local ischemia affects only the flap.
Observation-quality model. For each tissue region and each one-second window, a quality score $w(t) \in [0,1]$ estimates whether the pulse is observable at all. It combines three inputs: instrument and blood-pool segmentation masks, a spectral signal-to-noise ratio, and a coherence check against the reference phase. Low-quality windows are treated as missing, not as zero perfusion. Missing time is filled by a simple state-space model, which carries forward the last well-observed perfusion level with growing uncertainty.
The burden. Let $r(t)$ be the normalized local pulse power in window $t$, $\tau$ a perfusion threshold, and $\hat r(t)$ the filled estimate. The burden over an operation of length $T$ is
$$B = \int_0^T \max\big(0,\; \tau - \hat r(t)\big)\, dt,$$
reported alongside the fraction of time imputed and an uncertainty interval on $B$ propagated from the state-space model. We learn $\tau$ from the planned ischemia intervals in the flap dataset rather than setting it by hand. Planned clamping supplies free, reliable labels for "no local perfusion."
Contamination arm (secondary, prospective). A wound-bed swab at closure is sequenced with the 16S amplicon pipeline from the sepsis-outbreak study. This gives a direct bacterial-burden measure next to $B$, so the two candidate mechanisms behind duration can be entered into the same outcome model.
Baselines
The naive reference point is total operative duration, used as a continuous variable and as the published three-hour cutoff. This is the variable every cited clinical study found predictive, so it is the bar any new measure must clear. The strongest duration-based result reported in the file is an AUROC of 0.87 for a random forest with operation time as its top feature, rising to 0.89 after selection (Risk prediction of cholangitis after stent implantation). That is a different procedure and outcome, so it is context, not a target.
The fair comparison is within our cohort. We fit logistic and random-forest models on identical clinical covariates (age, blood loss, comorbidity, duration) with the same cross-validation folds and tuning budget, with and without $B$. Beating this baseline means a confidence interval on the AUROC gain from adding $B$ that excludes zero, together with a net reclassification gain.
The strongest existing perfusion method is the stage-based rPPG readout of the flap-monitoring study: pulse magnitude and flap-to-reference correlation measured once after reperfusion. It reported a correlation of 0.56 after reperfusion and an 18-second detection latency. No reported result was found on Lacuna for outcome prediction from these metrics. We compute them with the published POS pipeline on the same videos. Beating it means that $B$ predicts outcomes better than the end-of-reperfusion snapshot, which isolates the value of accumulation itself.
Two ablations isolate the key ingredients:
- Integral without reference locking, using the standard 0.7–4 Hz band, tests whether the narrowband filter matters.
- Integral without the quality model, in which corrupted windows either count as hypoperfused or are dropped as in the face-video work, tests whether honest handling of missing time matters.
A learned extractor adapted from Contrast-Phys, with the face-region contrast replaced by contrast between tissue patches sharing the reference frequency, is run under the same data and compute. This checks whether a neural extractor beats the simpler filter.
Why not just…?
Why not just use the clock? This is the real threat, and the design is built around it. In retrospective data, duration already carries large effects, such as an odds ratio of 7.927 in the meningitis cohort. The test is incremental validity: does $B$ improve prediction once duration is in the model, and does duration's coefficient shrink when $B$ is added? The flap study's failed case suggests it can. That flap flatlined, with reperfusion magnitude about 1234 against an ischemic baseline of about 1246, on a normal timeline. One case is not evidence of a population effect, though, and if the incremental gain is null we will report that the clock suffices.
Why not just take a single rPPG snapshot at reperfusion or closure? A snapshot catches complete failure but not partial, intermittent, or slowly worsening perfusion across a long operation. Whether that difference matters for outcomes is an empirical question, and the snapshot baseline is included precisely to settle it. If the snapshot matches $B$, the integral adds complexity for nothing, and we will say so.
Why not just use systemic monitors or contact probes? Cumulative hypotension from the anesthesia record is free and continuous, so we include it as a covariate. The claim is only that local tissue burden adds to it; the reference normalization is designed to remove the systemic component. Implantable Doppler probes are invasive and prone to false alarms, as the flap-monitoring authors note. Added hardware in the sterile field also introduces the kind of access points the sepsis-outbreak study implicated as contamination routes. A probe also gives one point rather than a map.
Proposed Experiments
Signal validation on existing video. The first experiment uses the FFCF ARRIscope dataset (15 patients, baseline, ischemia and reperfusion stages) and, where available, the SurgTissue recordings. Known clamp intervals serve as labels. We report per-second ischemia classification, reperfusion detection latency against the reported 18 seconds, and the between-state separation against the published p < 0.001. No numeric benchmark is reported for SurgTissue.
Stress test of the integral. This is the decisive check that is new to this proposal. We take reperfused segments and synthetically overlay real instrument masks, smoke-like haze, and 50/60 Hz-derived flicker, then measure how many spurious burden-seconds each method accumulates per hour. Ventilator motion is already present in real recordings. A method that accrues substantial false burden under realistic corruption is disqualified, whatever its classification accuracy. To confirm the reference-locked filter is sound in ordinary conditions, we also run it on UBFC-rPPG and MAHNOB-HCI faces. Contrast-Phys reports 0.64 bpm error on UBFC-rPPG, and the realistic-situations pipeline reports correlation 0.81 on MAHNOB-HCI.
Prospective outcome study. Fifteen patients cannot validate outcome prediction, so this step requires a prospective multi-surgeon cohort of free-flap reconstructions. It is sized for the number of complication events rather than patients, since both reference clinical cohorts were small and single-center. Primary outcomes are flap compromise and surgical-site infection at 30 days. The decisive analyses are three:
- the AUROC gain from adding $B$ to duration plus clinical covariates;
- the attenuation of the duration effect when $B$ is added, estimated with a mediation model that respects the strong collinearity between $B$ and time;
- a model with both $B$ and 16S wound bacterial burden, asking which mechanism carries duration's effect.
A positive result on the first analysis establishes a new intraoperative risk variable. A positive result on the third begins to answer the question the culture-negative meningitis cohort left open.
References
- Risk factors for postoperative meningitis after microsurgery for vestibular schwannoma
- Risk prediction of cholangitis after stent implantation based on machine learning
- Does age impact outcomes of retrograde intrarenal surgery in the elderly? Results from 366 patients from the FLEXible ureteroscopy outcomes registry (FLEXOR)
- Outbreak of Sepsis Following Surgery: Utilizing 16S RNA Sequencing To Detect the Source of Infection
- Perfusion assessment via local remote photoplethysmography (rPPG)
- Continuous intraoperative perfusion monitoring of free microvascular anastomosed fasciocutaneous flaps using remote photoplethysmography
- Remote Heart Rate Measurement from Face Videos under Realistic Situations
- Contrast-Phys: Unsupervised Video-Based Remote Physiological Measurement via Spatiotemporal Contrast