Case-Anchored Digital Twins for Apneic VV-ECMO Airway Surgery

In some airway operations, such as cutting out and resewing a section of trachea invaded by tumor, surgeons now remove the breathing tube entirely. Veno-venous extracorporeal membrane oxygenation (VV-ECMO) keeps the patient alive during this "tubeless" phase. VV-ECMO is a machine that drains venous blood, adds oxygen and removes carbon dioxide outside the body, and returns the blood to the veins. During that window the lungs do nothing, and the perfusionist sets pump flow, gas flow, and oxygen fraction by hand from intermittent blood gases. Almost everything published about this technique is a single case report. The case reports make two limitations plain. Each one shows only that the technique can work in one patient. None of them records what the circuit settings would have needed to be for a patient with higher oxygen demand, lower cardiac output, or more blood recirculating between the cannulas.

We propose a small, physiologically explicit digital twin of apneic VV-ECMO gas exchange. A digital twin here means a simulation that is continuously re-fitted to a live patient's data and whose predictions are fed back as recommended settings. We use it for two linked jobs. The first job treats a single well-documented case as a constraint on the twin's parameters. The twin then maps the region of patient physiology where the reported protocol stays safe, which turns "it worked once" into "it would work for this fraction of plausible patients, and it fails first along these axes." The second job keeps the twin synchronized with streaming data during a case, so it can forecast desaturation or rising carbon dioxide minutes before a blood gas would reveal it. The point that makes this more than a sensitivity analysis is an empirical test. We ask whether a twin anchored on one case predicts the behavior of other, held-out cases better than a population-average model does, and whether live synchronization adds anything beyond a static twin.

Core Claim

A single case report can support a quantitative, falsifiable statement about generalization if its reported trajectory is used to calibrate a mechanistic gas-exchange model and the resulting feasible region is projected onto a realistic patient population. We claim that this case-anchored estimate predicts held-out cases better than both a population-prior model and a model-free predictor. We further claim that intraoperative synchronization, meaning sequential re-fitting of the drifting parameters, measurably improves short-horizon forecasts of oxygenation and carbon dioxide at the moments that matter: tube withdrawal, bleeding, and cannula repositioning. We deliberately do not claim that the twin should drive ECMO hardware autonomously. The feedback direction is advisory, and it is closed-loop only in simulation.

Where the Gap Is

The clinical motivation comes from the tubeless case literature. The trachea-invading thyroid carcinoma case of He et al. withdrew the tube for 50 minutes under femoral-to-jugular VV-ECMO and held oxygen saturation at 98–100% with manual management. The authors state that a single case cannot establish safety across patients with different lung, vascular, or tumor characteristics. Song et al. describe similar tube withdrawal during tracheal anastomosis. The review of Su et al. collects ECMO bridging for tumor airway obstruction without a way to pool the cases. The broader case-report literature supplies the epistemic framing. Hideshima et al. and Zhou et al. note that a single success cannot separate a generalizable technique from an exceptional patient. That is exactly the confound a mechanistic twin can make explicit, by asking whether the case's success depended on unusually favorable parameters.

On the twin side, the closest surgical systems are kinematic or static. Twin-S synchronizes an optically tracked drill with a simulated skull phantom at about 28 frames per second, with 35.7 ms latency and 1.39 mm error, but it models no physiology. The differentiable airway-planning system in Virtual reality-augmented differentiable simulations optimizes paths on static CT anatomy and receives no intraoperative telemetry. Asciak et al. set out digital-twin-assisted surgery as a concept and name biological variability and compute cost as the main barriers. All three motivate the work, and none is a method we can reuse for gas exchange.

The closest methodological precedent is Fox et al.. They trained a recurrent Soft Actor-Critic glucose controller across 30 virtual patients in the FDA-accepted UVA/Padova simulator and adapted it to new patients. Their work shows that a validated virtual cohort makes cross-patient claims testable. It also shows the key dependency: they had a simulator, and ECMO airway surgery has none. We therefore have to build and validate the virtual cohort before any controller claim is meaningful. We reuse two methods. DiffTune-MPC tunes model-predictive-control weights by differentiating through the solver, and we use it to tune the in-simulation advisory controller. MIMIC-IV preprocessing and MIMIC-Extract provide the pipelines that build population distributions of the physiology the twin needs.

Research Program

The twin is a lumped, low-dimensional model and not a multi-physics solver. This avoids the latency problem Asciak et al. raise. The model tracks body oxygen and carbon dioxide stores, oxygen consumption $\dot V_{O_2}$, carbon dioxide production, cardiac output $Q_C$, ECMO blood flow $Q_E$, the recirculation fraction $R$ (oxygenated blood pulled straight back into the drainage cannula), native lung shunt, hemoglobin, and blood volume. The central balance during the apneic phase, when the lungs contribute nothing, is

$$\dot V_{O_2}^{\text{del}} = Q_E\,(1-R)\,\big(C_{\text{post}} - C_{\text{pre}}\big),$$

where $C_{\text{post}}$ and $C_{\text{pre}}$ are the oxygen contents of blood leaving and entering the oxygenator. Arterial saturation falls whenever this delivered oxygen stays below $\dot V_{O_2}$ long enough to drain the body's oxygen stores. Carbon dioxide clearance follows an analogous term driven by sweep gas flow.

The surgical course is modeled as a sequence of discrete phases: ventilated, tube withdrawn, airway open, anastomosis, and reintubation. Bleeding and cannula events enter as parameter jumps. This reflects the non-smooth dynamics of Gap 3. We use gradients within a phase and handle the switches as known or detected events, rather than pretending the whole trajectory is differentiable.

Case anchoring works as follows. Let $\theta$ denote the patient parameters and $\pi$ the reported protocol, meaning the settings and timings. A case report supplies partial observations of the case, such as saturation bounds over the tubeless window, duration, and blood loss. We sample a posterior $p(\theta \mid \text{case})$ by simulation-based inference. We then define the protocol's feasible set $F_\pi$ as the parameters for which saturation and carbon dioxide stay within clinical thresholds for the planned apneic duration. The generalization estimate is

$$G_\pi = \Pr_{\theta \sim p_{\text{pop}}}\big[\theta \in F_\pi\big],$$

where $p_{\text{pop}}$ is a population distribution over the same parameters, estimated from ICU physiology through the MIMIC pipelines. The case posterior serves two purposes. It pins down the parameters the case actually informs, and it checks whether the case sat in an unusually favorable corner of $p_{\text{pop}}$, which is the "exceptional patient" confound. Along the boundary of $F_\pi$, the twin also solves for the smallest setting change that restores feasibility. That output tells a team which prospective patients would need a modified protocol.

Synchronization works as follows. During a case, an ensemble filter re-estimates the fast-drifting parameters ($R$, $\dot V_{O_2}$, $Q_C$, blood volume) from continuous oximetry, arterial pressure, circuit pressures, and pre- and post-oxygenator saturation. It incorporates intermittent blood gases when they arrive. The twin rolls forward over a 5 to 10 minute horizon. A controller tuned with DiffTune-MPC proposes flow and sweep changes. In the virtual cohort this loop is fully closed. Clinically it would only display a recommendation next to the forecast.

Baselines

The naive reference point is the manual protocol as reported, with fixed settings adjusted only when a simulated blood gas is "drawn" at clinically realistic intervals. The file's only relevant number for this is the He et al. result of 98–100% saturation over 50 minutes with manual care in one favorable patient. For any ECMO controller or forecaster on a cohort, we found no reported result on Lacuna. This baseline runs on exactly the same virtual patients and event scripts as every other arm.

The strongest existing closed-loop method is a reactive proportional-integral-derivative (PID) controller on saturation and carbon dioxide. This is the class of controller Fox et al. used as their clinical baseline, and they report that their predictive agent lowered median glycemic risk from 8.34 under PID to 4.24. That figure comes from a different organ system and serves only as evidence that predictive control can beat PID, not as a target here. We give PID gain tuning the same simulation budget as the MPC tuning. Beating it means fewer minutes below the saturation threshold and fewer hypercapnic excursions across the cohort, with no increase in large setting swings. A second strong competitor is a model-free recurrent Soft Actor-Critic agent in the style of Fox et al., trained on the same virtual cohort with the same number of simulated hours.

For the generalization claim, the strongest competitor is a population-prior twin with the same model and the same $p_{\text{pop}}$ but no case anchoring. The key ablations isolate our two ingredients. One removes case anchoring, keeping synchronization with a population prior. The other removes synchronization, leaving a static case-anchored twin whose parameters are fixed at the start of the case. If the static twin forecasts as well as the synchronized one, the bidirectional part is not earning its cost.

Why not just…?

Why not just run a Monte Carlo sensitivity analysis on a lumped ECMO model? This is the most serious threat. For the population-feasibility estimate $G_\pi$, a sensitivity sweep and our method share almost everything. Our method differs in two places. The case posterior replaces an arbitrary parameter prior on the dimensions the case informs, and the twin is judged by prediction on held-out cases rather than by plausibility. If leave-one-case-out prediction shows no gain from anchoring over the population prior, we will report that the case adds nothing beyond a sensitivity sweep. That is a useful negative result about what case reports can support.

Why not just let an experienced perfusionist titrate from oximetry and point-of-care gases? Manual control clearly worked in the He et al. patient. The question is how much margin it has in less favorable patients. The prophylactic VA-ECMO trial (Prophylactic VA-ECMO During Complex High-Risk PCI) found a 19.4% rate of life-threatening complications when support was held in reserve, versus 0% when it was started beforehand. That concerns a different procedure, but it is consistent with reactive human management leaving gaps. Our cohort experiment measures this directly. It compares time-to-detection and time-below-threshold for the manual-interval policy against the twin forecast, stratified by distance from the boundary of $F_\pi$.

Why not just use a PID loop, or a model-free RL agent, with no twin? PID is included as a baseline. If it matches MPC on the cohort, a simple controller suffices and we will say so. The model-free agent is a subtler threat in simulation, because it may match MPC on the very simulator it was trained on. We therefore also evaluate every controller on a perturbed simulator with altered recirculation physics and hemorrhage profiles, where a model that is re-fitted online should degrade less.

Why not just run a multi-center registry? A registry should be run, and the twin is not a substitute for one. Tracheal-invading tumors that need tubeless resection are rare and anatomically diverse. The 70-patient single-center trial above still lacked power for 12-month outcomes (5.9% vs 16.7%, P = 0.26). The twin tells a registry which physiological variables to record, and in which patients the protocol is predicted to need modification.

Proposed Experiments

The first experiment validates the model on real data. We extract VV-ECMO-supported ICU stays from MIMIC-IV using the cited pipelines. Before committing, we will verify how many stays have charted flow, sweep, and blood-gas time series, because the file gives no count. Each stay is treated as a "single case." We anchor the twin on the first hours of one stay, or on one stay, and predict arterial oxygen and carbon dioxide at later blood-gas times in the same stay and in other stays. Performance is scored by prediction error and interval calibration, with the population-prior twin and a recurrent sequence model as comparisons. ICU patients on ultra-protective ventilation, as in the fungal ARDS case of Jung et al., are the closest available proxy for apneic support. We acknowledge that true tubeless apnea is more extreme, and that this experiment tests the gas-exchange model, not the surgical phases.

The second experiment reconstructs the published cases. We anchor on He et al., using the 50-minute apneic duration, 98–100% saturation, and roughly 400 mL blood loss. We report $G_\pi$, where the case's posterior sits within $p_{\text{pop}}$, and the setting changes needed at the feasibility boundary. The same procedure applied to the Song et al. case tests whether two independently anchored posteriors agree on the shared physiology.

The third experiment runs the controllers in the virtual cohort. We build scripted surgical scenarios, including tube withdrawal, a hemorrhage of variable size, and a sudden jump in recirculation from cannula migration, across a few thousand patients drawn from $p_{\text{pop}}$. We compare the manual-interval policy, tuned PID, model-free Soft Actor-Critic, the static twin with MPC, and the synchronized twin with MPC on minutes below 90% saturation, peak carbon dioxide, and forecast lead time. We also compute the twin's per-step compute cost against a real-time budget; Twin-S's 35.7 ms latency is a reference for what surgical twins have achieved.

The decisive checks are two. If case anchoring does not beat the population prior on held-out cases, the N=1 generalization claim fails. If synchronization does not beat the static twin during the scripted parameter jumps, the bidirectional component fails and the work reduces to a well-calibrated offline risk map. Both outcomes are publishable.