Signed-Change Monitoring of Free Flaps After Surgery

When surgeons move a piece of living tissue, called a flap, with its own artery and vein to cover a wound that would otherwise not heal, the tissue lives only as long as those reconnected vessels stay open. The most dangerous period is the first one to three days after surgery. A clot in the vein or artery during that window kills the flap within hours unless someone notices and reoperates. Today "noticing" usually means a nurse checking color, warmth and capillary refill every hour or two. A camera-based method called remote photoplethysmography reads the faint pulse-driven color changes in skin from ordinary video. It has already been shown to tell a perfused flap from an ischemic one on the operating table. So far it has only been used for minutes, and only to answer the snapshot question "is there flow now?"

We propose to treat postoperative flap monitoring as a longitudinal reasoning problem. The model should answer a different question: over the last few hours, relative to this patient's own post-surgery baseline and to an untouched patch of their skin, is perfusion getting better or worse, and how fast? The work from the longitudinal medical imaging literature that motivates this is a specific, documented failure. Current models can often tell where something changed, but they are poor at the direction and size of the change. Direction and size are exactly the quantities that matter when a flap is slowly failing. We will build a monitor whose output is a signed rate of change and whose architecture guarantees that swapping the order of two observations flips the sign. We will test whether it gives earlier and more reliable warning than simple thresholding of the same camera signal.

Core Claim

The claim has two parts, and either can fail on its own. The first is clinical. Slow or partial flap compromise, especially venous congestion, shows up as a gradual downward trend in pulse signal quality relative to reference skin before it crosses any absolute threshold. That trend is only visible if you compare the flap against its own history. The second is methodological. A learned change model that is forced to be antisymmetric in time (score(a, b) = −score(b, a)) and is anchored to an internal reference region estimates the sign of change more reliably than models that learn temporal order only from data augmentation. It should also estimate sign more reliably than hand-set thresholds on the raw signal. If the second part fails, meaning a simple statistical trend detector does as well, the first part can still stand, and the project reduces to a data and monitoring protocol. We state that outcome in advance as acceptable.

Where This Sits in Prior Work

The optical starting point is Continuous intraoperative perfusion monitoring of free microvascular anastomosed fasciocutaneous flaps using remote photoplethysmography. That study recorded 15 head-and-neck free flaps with a surgical microscope at 60 frames per second and extracted the pulse signal with a standard skin-color projection method. It defined three quantities: a perfusion index, a correlation ρ between the flap pulse and a reference zone on the face, and a spectral magnitude M. Reperfusion was detected about 18 seconds after unclamping, ρ rose from near zero to about 0.56, and the one failed flap stayed flat (M ≈ 1234 against an ischemic baseline of ≈ 1246). We reuse its signal extraction and its central idea of comparing the flap against reference skin. We differ in two ways. We move from a minutes-long intraoperative window to days of postoperative recording, and we replace fixed deterministic readouts with a model of within-patient change. Its single gross failure was easy to see from a flatline. It gives no evidence about gradual or venous failure, which is the case we care about.

The motivation for focusing on the sign of change comes from How Good are Foundation Models in Longitudinal MRI Disease Progression Reasoning?. That benchmark found that vision-language models localize interval change well (up to 97.3% in an agentic setup) but reach only 0.19–0.63 Trend-F1 and under 16% on quantification. The same pattern holds in chest radiography. Learning to Exploit Temporal Structure for Biomedical Vision-Language Processing (BioViL-T) separates static from change features in a pair of X-rays. CoCa-CXR restricts attention to matching regions across the two images and trains on order-swapped pairs generated with an LLM. DeltaNet feeds an explicit feature difference into report generation. All of these work on two-timepoint 2D chest images. CoCa-CXR's order swap is the closest idea to ours, but it teaches sign symmetry only through training data. We impose it by construction and test whether that matters.

CRONOS acts as a constraint on our design rather than a method we reuse. It showed that standard sequence models often fail to beat simply copying the last observation, and that its continuous-time flow model beat that heuristic only modestly on glioma MRI (7.55 against 8.38 normalized error). Slowly changing medical signals punish models that predict blurry averages. We therefore adopt "the last state persists" as a mandatory reference point. We also predict a signed change rather than a future frame. The surgical literature defines where the monitor must work. Reconstruction of chronic radiation-induced ulcers in the chest wall using free and pedicle flaps reports 17/17 flap survival but delayed healing at flap margins in 2 patients, the kind of partial, edge-limited compromise a whole-flap threshold would miss. Simultaneous reconstruction of septic composite defects in lower extremities uses perforator flaps over infected beds, where the surrounding tissue is itself poorly perfused. That makes a reference region on the same limb unreliable, so the reference region must be chosen carefully.

Research Program

The system has three parts. The first is an acquisition protocol: a fixed bedside camera views the flap and one or more untouched skin regions. It records short clips every few minutes for 72 hours and logs heart rate from the standard monitor to help separate pulse from motion. The pulse signal extraction follows Schraven et al.

The second part is the change model. An encoder maps each clip to a per-region embedding of pulse quality: amplitude, synchrony with the reference and spectral shape. A comparison head then scores pairs of clips taken at times $t_i < t_j$:

$$s(x_i, x_j) = g\big(h(x_j) - h(x_i)\big) - g\big(h(x_i) - h(x_j)\big),$$

where $h$ is the encoder, $g$ is a small network, and $s$ is the signed perfusion change from $t_i$ to $t_j$. By construction $s(x_i,x_j) = -s(x_j,x_i)$ and $s(x,x)=0$. Before encoding, each flap region is normalized by the concurrent reference-region signal, so changes in the whole body, such as blood pressure or temperature, cancel out.

The third part is the trend estimator. The signed pairwise scores over a sliding window of irregularly spaced clips are combined, weighted by elapsed time, into a per-region rate of change. The alarm is raised on a sustained negative rate, not on an absolute level.

Labels come from three sources, in order of availability. The first is intraoperative transitions (baseline → ischemia → reperfusion), which give known-sign changes in every patient. The second is a healthy-volunteer study using a blood-pressure cuff on the forearm. Graded venous occlusion and graded arterial occlusion produce perfusion changes with known sign, magnitude and type, and they can be repeated many times. The third is the postoperative cohort, with outcomes defined as re-exploration, confirmed thrombosis, partial necrosis or uneventful survival.

The volunteer data is essential for a practical reason. Flap failure is rare, as the 1-in-15 in Schraven et al. illustrates, so no single-center cohort will contain enough events to train on. The model is therefore trained on volunteers and intraoperative transitions and evaluated on postoperative patients. To separate the time-structure question from flap-specific effects, we also run the same antisymmetric comparison head on MS-CXR-T and on the longitudinal MRI benchmark, where the sign failure was first documented.

Baselines

The naive reference point is "no change." It assumes every flap stays in its last observed state and raises no alarm until an absolute threshold on magnitude M or reference correlation ρ is crossed, with thresholds taken from the intraoperative ischemic and reperfused values. This is the direct analogue of CRONOS's last-observation baseline. A flatline flap like Schraven et al.'s failure case will be caught by it, so it sets the floor on gross failures.

The strongest competitor is a classical statistical trend detector on the same signal. We compute Schraven et al.'s ρ and M continuously on the same clips, with the same reference-region normalization, and run a cumulative-sum change-point test and a robust linear-slope test over the same sliding window. It is the strongest competitor because it uses exactly the physiological quantities already shown to separate perfusion states, and it has almost no free parameters to overfit on a small cohort. We tune its window length and alarm threshold on the same volunteer and intraoperative data the learned model sees, with a matched budget of threshold settings. For the learned model to beat it, it must achieve longer median warning time before clinical recognition of compromise at equal or lower false alarms per patient-day, or better sign accuracy on the graded venous-occlusion volunteer trials. For flaps that ultimately fail, we compare warning time using paired within-patient tests. We know of no prior warning-time figure: no reported result found on Lacuna for postoperative rPPG flap failure detection lead time. The only reported numbers are intraoperative: 18-second reperfusion detection and ρ ≈ 0.56 after reperfusion.

The learned sign baselines come from imaging. On MS-CXR-T, CoCa-CXR reports 65.0% average progression classification accuracy across five conditions, 4.8 points above BioViL-T. On the longitudinal MRI benchmark, the reported Trend-F1 range is 0.19–0.63, with a best final accuracy of 35.15% for InternVL3.5-Inst. We will reproduce CoCa-CXR-style order-swap training with the same backbone and data. Beating it means higher sign accuracy on held-out pairs, and in particular a smaller gap between forward-ordered and reversed-ordered test pairs.

The key ablation removes the antisymmetry constraint and keeps order-swap augmentation, with everything else fixed. A second ablation removes reference-region normalization. If the constrained and augmented variants perform the same, the architectural claim is refuted.

Why not just…?

Why not just threshold ρ and M, as Schraven et al. already do? For gross arterial failure this probably works, and we expect the no-change baseline to catch flatline cases. The gap is gradual compromise. Venous congestion first leaves the pulse present and then degrades it slowly, so an absolute threshold fires late or, if set tighter, fires constantly. The graded venous-occlusion volunteer trials test this directly: they show whether signal-level thresholds can track the sign of small, known changes.

Why not just run a change-point test on the raw signal instead of learning anything? This is the real threat, and we treat it as the primary baseline rather than a straw man. With few events, a nearly parameter-free detector may match any learned model. If it matches the learned model on warning time and false alarms in the postoperative cohort, we will report that and recommend the simple detector. In that case the contribution is the protocol, the reference normalization and the labeled volunteer dataset. The learned model earns its place only if it handles clip-level confounds better than the fixed signal readouts. The main confounds are motion, dressings, lighting and flaps partly under bandage.

Why not just give serial flap photos to a general vision-language model? The longitudinal MRI benchmark shows these models get the direction of change wrong and score under 16% on quantification. Nothing in their training involves pulse signals invisible to the eye. We include a prompted multi-image model on still frames as a lower-bound comparison in the volunteer study, mainly to confirm that color appearance alone lags the pulse signal.

Why not just use an implanted Doppler probe? Doppler is invasive and prone to false alarms from dislodgement, as the tissue-shielding literature notes. It also only measures the vessel at one point, not the skin territory. In the subset of patients who already receive one, we will record it as a comparison signal rather than treat it as the ground truth.

Proposed Experiments and Decisive Checks

The first check is on imaging benchmarks, and it isolates the sign mechanism. On MS-CXR-T and the longitudinal MRI benchmark, we measure sign accuracy and Trend-F1 for the constrained comparison head against an order-swap-trained copy with the same backbone. We report accuracy on reversed-order test pairs separately. The prediction is that the constrained model shows no gap between forward and reversed pairs while the augmented model does. If both close the gap equally, the constraint adds nothing.

The second check is the volunteer cuff study, which contains the most informative labels. Across graded venous and arterial occlusion levels, we measure sign accuracy, the correlation between predicted rate and cuff pressure change, and how many seconds each method takes to detect a change. We compare the learned model, the change-point detector and the thresholds. We also test reference choices: face, contralateral limb and adjacent skin on the same limb. This tells us whether same-limb reference skin is trustworthy, which matters for infected or irradiated beds like those in Liu et al. and Zhou et al.

The third check is a prospective postoperative cohort across several centers, recorded passively without influencing care. The primary endpoint is warning time before clinical recognition in flaps that need re-exploration or develop partial necrosis. The secondary endpoints are false alarms per patient-day in uneventful flaps and region-level detection of edge-limited compromise. The decisive result is whether the learned model gives earlier warning than the change-point detector at matched false-alarm rates. Either answer is informative, and we commit to publishing whichever one we find.