Does Sticking Harder Now Mean Failing Later? A Long-Horizon Cohort Test of Wet-Tissue Adhesive Chemistry

Surgical glues that bond to wet, bleeding tissue are improving quickly in the lab. New designs absorb blood or water at the wound surface, or pull it into themselves, and then form chemical bonds with tissue proteins. They seal arteries and bowel faster and more strongly than sutures or older sponge-type products. Every one of these studies, however, stops watching the animals after two weeks to three months. The authors themselves list late failures they could not see: false aneurysms forming at the repair, narrowing of the vessel, chronic foreign-body inflammation, and scar adhesions to nearby organs.

Human data that could answer this question already exist in hospital records. Approved products that use the same chemical routes to wet adhesion have been in surgical use for years. We propose a retrospective cohort study, meaning an analysis of routinely collected patient records followed forward in time. It would compare these products against one another, grouped by how they bond, and would treat the length of follow-up as the quantity to estimate rather than as a limitation to apologize for. The deliverable is a "crossover time" for each bonding mechanism: the point after surgery at which an early advantage in sealing turns into a late excess of complications, if such a point exists. This number tells biomaterial developers how long their animal studies must run before acute strength can be trusted.

Core Claim

The materials papers expose a real tension. Covalently reactive adhesives get their strength from dense chemical crosslinks with tissue. MonoSeal uses aldehyde groups that bind tissue amines and reaches 49.18 kPa wet lap-shear strength, against 8.16 kPa for cyanoacrylate, as reported in A Single‐Component Adhesive Sponge Based on Blood‐Triggered and Autopenetrative Adhesion for Robust Vascular Closure. The blood-imbibing microparticles form clots about 100 times stiffer than natural clots and cut rebleeding in rabbits to 4%, against 32–51% for commercial hemostats, as reported in Covalently reactive microparticles imbibe blood to form fortified clots for rapid hemostasis and prevention of rebleeding. Stiff, permanently bonded material placed on arteries that pulse or bowel that contracts may mismatch the tissue's mechanics over months.

The non-covalent alternative has the opposite problem. The hydrogen-bonded hydrogel in A Hydrogen Bonds-Crosslinked Hydrogels With Self-Healing and Adhesive Properties for Hemostatic self-heals and stays compliant, but its authors flag poor stability under full immersion.

Our claim is testable. Among patients who all received some topical sealant or hemostat on a vascular suture line, covalently tissue-reactive products will show lower early reoperation for bleeding but a hazard of late repair-site complications that rises over time. In most designs this rise would be averaged away by a single hazard ratio, or would never be seen at all under in-hospital or short follow-up.

None of the four wet-adhesion papers in the file, and none of the retrospective cohorts, connects these threads. The file's own survey finds no study in the corpus spanning wet adhesion and retrospective cohorts. The closest work is the adhesive papers' own calls for longer follow-up. This study answers those calls with human data and a mechanism-level hypothesis, not with a new material.

Research Program

The exposure is the bonding mechanism, not the brand. Each approved topical product recorded in the operative record would be assigned to one of three classes. The first is covalently tissue-reactive products, such as aldehyde-crosslinked protein glues and cyanoacrylates. The second is passive fluid-absorbing matrices, such as gelatin sponges, oxidized cellulose and microfibrillar collagen. These are the products the particle paper used as controls. The third is biologic or non-covalent sealants. The classes deliberately mirror the three strategies in the lab literature: active fluid evacuation, co-opting blood, and reversible networks. That mapping is what lets human findings feed back into material design.

Two engineers blinded to outcomes would do the class assignment from product chemistry before any outcome data are opened.

The primary clinical setting is open vascular and cardiac surgery. Adhesives are applied there to arterial suture lines, which is the setting MonoSeal targets. Late failure modes there are clinically defined and coded: false aneurysm, narrowing of the repaired vessel, and reintervention at the repair site. The secondary setting is gastrointestinal anastomosis, the surgical joining of two bowel ends. It corresponds to the cryogel particles of Adhesive cryogel particles for bridging confined and irregular tissue defects, with late narrowing of the join and adhesive bowel obstruction as outcomes.

The comparison is between products (active comparator), and every patient enters the cohort on the day the product is used (new user). Comparing sealant against no sealant would be badly confounded, because surgeons reach for glue when bleeding is already hard to control. Gap 3 of the file makes this point.

Within the comparison between products, confounding by indication is handled in two independent ways. The first is propensity-score weighting on preoperative and intraoperative covariates. A propensity score is the estimated probability of receiving a given product given the patient's recorded characteristics. This method follows the matching used in Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach. The second is an instrumental-variable analysis. It uses the fact that hospitals and individual surgeons stock and prefer different product classes for reasons largely unrelated to the patient in front of them.

The prostate study is also a warning. It concedes that matching cannot remove unmeasured surgeon preference. Here we turn that preference into the instrument.

Clustering by hospital and surgeon would be modeled with multilevel terms. This follows how Risk factors associated with 31-day unplanned readmission in 50,912 discharged patients after stroke in China separated patient-level from hospital-level effects across 375 hospitals.

The time dimension is the methodological core. We let the hazard ratio vary over time, $HR(t) = h_{\text{cov}}(t)/h_{\text{ref}}(t)$. Here $h_{\text{cov}}(t)$ is the hazard of a repair-site complication at time $t$ after surgery for the covalent class, and $h_{\text{ref}}(t)$ is the same hazard for a reference class. The model is a flexible parametric survival model with time-by-exposure interaction splines. We then define the crossover time $t^*$ as the earliest time at which the difference in cumulative incidence of the composite outcome changes sign:

$$\Delta(t) = F_{\text{cov}}(t) - F_{\text{ref}}(t), \qquad t^* = \inf\{t : \Delta(t) > 0\},$$

where $F(t)$ is the cumulative incidence of the composite of reintervention, false aneurysm, narrowing and reoperation for bleeding by time $t$, estimated with death as a competing risk. A $t^*$ with a confidence interval is directly actionable. If $t^*$ falls at, say, eight months, then any animal study that stopped at 14 days or 3 months, as all four adhesive papers did, could not have detected the trade-off.

Data would come from multi-institution EHR extracts linked to claims or death registries so follow-up extends past discharge. Discharge-censored designs are exactly what Occurrence rates and risk factors of in-hospital venous thromboembolism, major bleeding, and death in patients receiving fondaparinux after orthopedic surgery or trauma surgery concedes leaves "long-term post-discharge risks … unquantified."

Cohort construction would reuse the code mapping and time-binning approach of An Extensive Data Processing Pipeline for MIMIC-IV. We are candid about a constraint: none of the datasets in the file records product-level topical sealant exposure. MIMIC-IV serves as a pipeline-development and outcome-definition testbed, with procedure codes and readmissions available. The primary analysis requires institutional data in which operative notes or supply-chain charges identify the product used. A natural-language extraction step from operative notes, checked against manual chart review on a random sample, is therefore part of the build.

Baselines

The naive reference point is an unadjusted comparison of patients who received any sealant against those who received none, with outcomes collected only until discharge. This is what a quick chart review would produce. We expect it to show sealant users doing worse, because sealants are used in harder cases. It defines the confounded floor and is reported only to show how far adjustment moves the estimate.

The strongest existing approach in the file is propensity-matched comparison of treatment arms with long follow-up. Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach used it to report 68% lower overall mortality for prostatectomy versus conservative care. On our question, sealant-class effects on late vascular complications, we found no reported result on Lacuna, so there is no published number to beat.

We would run this baseline on identical cohorts and covariates, with the same tuning budget for the propensity model. It would be run as a sealant-versus-no-sealant matched Cox model with one constant hazard ratio. The proposal beats it under two conditions. The first is that the active-comparator time-varying design yields estimates that pass the negative-control checks described below while this baseline fails them. The second is that the time-varying model shows a statistically supported departure from a constant hazard ratio, detected by a likelihood-ratio test of the interaction splines. This second condition would mean the single-number summary is misleading.

The short-horizon baseline is the same active-comparator, propensity-weighted model with follow-up cut off at discharge, then at 30 days, then at 3 months. These cut-offs mimic the in-hospital design of the fondaparinux cohort and the 3–6 month window of A predictive model using left atrial function and B‐type natriuretic peptide level in predicting the recurrence of early persistent atrial fibrillation after radiofrequency ablation. It isolates what the limited-follow-up lens adds. If the conclusion at 3 months has the opposite sign from the conclusion at 3 years, the follow-up window itself is driving inference.

Two ablations isolate the key ingredients. Pooling all products into one "sealant" exposure tests whether mechanism grouping matters at all. Replacing chemistry-based classes with a random grouping of the same product counts tests whether any grouping would show heterogeneity by chance.

Why not just…?

Why not just run longer animal studies? Longer studies in pigs and rabbits are needed and complementary, but they cannot supply what this study does. Animal models use healthy young vessels, small numbers, and a single lesion type. Late human failures arise in atherosclerotic, diabetic and reoperated tissue over years. The adhesive papers themselves frame their two-week and three-month windows as a translational gap (Adhesive cryogel particles for bridging confined and irregular tissue defects), not a protocol oversight. The cohort's $t^*$ estimate is what tells a future animal study how long it must run.

Why not just do a propensity-matched sealant-versus-no-sealant study? This is the obvious design and a real threat if sealant indication were well captured. It is not: intraoperative bleeding severity, tissue quality and surgeon judgment are mostly unrecorded. The prostate study names unmeasured surgeon preference as its residual weakness. We run this design as a baseline and subject it to the same negative-control outcomes. If it passes them as well as the active-comparator design does, we would report it as adequate and simplify.

Why not just wait for a randomized trial? A trial could randomize product class, but it would need years of follow-up and thousands of patients to detect late events. It also cannot ethically randomize uncontrolled hemorrhage to a weaker agent. Timing to intubation COVID-19 patients: can we put it off until tomorrow? makes the analogous argument for delayed intubation. Our estimates would size such a trial and indicate which comparison deserves it.

Why not just mine device adverse-event reports? Spontaneous reports have no denominator, no comparator, and heavy reporting bias toward dramatic early events. They cannot estimate a time-varying hazard, which is the quantity at stake.

Proposed Experiments and Decisive Checks

The primary experiment estimates $HR(t)$ and $t^*$ for covalent versus passive-matrix products on arterial suture lines over five years. Propensity weighting and the surgeon-preference instrument give two independent estimates. Agreement between them is required before any claim is made.

Three falsification checks decide credibility. First, a negative-control outcome unrelated to the repair site, such as fracture admissions, must show no class difference. Second, the time-varying effect must be absent in patients whose sealant was applied to a non-pulsatile site, which would be the case if mechanical mismatch is the mechanism. Third, the E-value for each late-window estimate must be reported. The E-value is the strength of unmeasured confounding needed to explain the effect away.

Outcome definitions would be validated by blinded chart review on a random sample before the main analysis, with positive predictive value reported per outcome code.

The secondary experiment repeats the design for bowel anastomosis, with late narrowing and adhesive obstruction as outcomes. It tests whether $t^*$ generalizes across tissue types. If it does not, the result points to tissue-specific timing for degradation, which is the open problem the cryogel paper raises.

A clean null would also be informative. That would be a constant hazard ratio near one across all windows with adequate power, computed in advance from expected late-event rates. It would tell developers that acute strength is a fair proxy for durability in this mechanism class, and that the two-week animal endpoint is defensible.

All analyses, class assignments and window definitions would be preregistered before outcome data are unlocked.