To estimate pleural pressure in ventilated patients, clinicians thread an air-filled balloon catheter into the esophagus and read its internal pressure. That reading is only trustworthy within a narrow band of filling volumes. Too little air and the balloon does not touch the esophageal wall, so pressure is underestimated. Too much and the stretched balloon pushes back on its own air, so pressure is overestimated. Bench work shows this band is specific to each balloon: 0.7–1.7 ml for one commercial catheter and 2.2–4.6 ml for another. Clinical work shows the esophageal wall adds its own stiffness, which has to be measured and subtracted. Every published calibration rule comes from one catheter brand at one center, often with about 40 patients. None has been checked on an independent cohort, a different catheter, or a different hospital.
This note proposes to separate the balloon's own mechanics from the patient's. The balloon part can be measured once per catheter model on a bench with no patients. The patient part, mainly esophageal wall stiffness, varies across people but follows a population distribution that a few dozen patients can estimate. Combining the two gives a calibration model that can be locked at one center and handed to a second center using a different catheter. That second center needs only a short bench characterization and a handful of local patients. We would test this with a pre-registered external validation, and use the model's uncertainty to shorten bedside titration from about a dozen inflation steps to two or three.
Core Claim
The claim is that balloon calibration transports across catheters and centers when it is split along physical lines. The catheter-specific part is a balloon recoil curve, which is cheap to measure on the bench at many surrounding pressures. The patient-specific part is wall elastance and contact volume, which can be learned from small cohorts with partial pooling, meaning each patient's estimate is shrunk toward the population. This split is what makes the small-sample problem tractable. A new catheter model then costs one bench session plus fewer than about 20 target patients. It does not require a new clinical derivation study of 40 or more patients.
Concretely, we model the balloon pressure $P_b$ at filling volume $V$ as
$$ P_b(V) = P_s + R_c(V, P_s) + E_w \, \max(0,\, V - V_0), $$
where $P_s$ is the true surrounding (pleural-like) pressure and $R_c(V, P_s)$ is the catheter's intrinsic recoil. $R_c$ depends on surrounding pressure, consistent with the bench finding that the working range shifts upward at higher baseline pressure. $E_w$ is esophageal wall elastance and $V_0$ is the volume at which the balloon contacts the wall. The optimal filling range is the set of volumes above $V_0$ where $|R_c(V, P_s)| \le 1\ \text{cmH}_2\text{O}$. This is the same tolerance on the difference between balloon and surrounding pressure that the bench literature uses. $R_c$ comes from the bench. $E_w$ and $V_0$ get a hierarchical prior learned at the source center and are updated per patient from very few inflation points.
Research Program
The work builds directly on two calibration studies.
Optimal esophageal balloon volume for accurate estimation of pleural pressure at end-expiration and end-inspiration fit sigmoid curves to bench pressure–volume data for the Cooper, SmartCath-G and Microtek catheters. It showed the working range is balloon-specific and pressure-dependent, but it had no patients. We reuse its bench protocol to measure $R_c$, sampled densely across surrounding pressures, for each catheter.
Use of esophageal balloon pressure-volume curve analysis to determine esophageal wall elastance showed that the slope of the middle linear section of the in-vivo curve gives wall elastance. It proposed a two-point shortcut at 0.6 and 1.4 ml and reported a median elastance of 3.3 cmH₂O/ml and a median best volume of 1.0 ml in 40 patients, all with one catheter. Our model generalizes that slope to a population distribution. Its fixed volumes become a choice the model makes per catheter and per patient.
The design philosophy comes from A Left Ventricular Segmentation Method on 3D Echocardiography Using Deep Learning and Snake. That work segmented the heart from only 45 sequences by having a constrained geometric model do most of the work, leaving little to the learned component. We apply the same idea to a mechanical curve rather than an image boundary.
Esophageal virtual disease landscape using mechanics-informed machine learning supplies motivation and a constraint. It shows that conservation-law preprocessing of balloon data yields physically meaningful features. It also needed 804 single-center subjects and more than 220,000 augmented samples, so it says nothing about the regime of tens of patients that we target.
The program has three parts. First, a bench library: $R_c$ for three catheter models across the chamber elastances and baseline pressures used in the cited bench studies. Second, a source-center derivation with a Cooper catheter and full 0.2 ml titration in about 40 ventilated patients, mirroring the Sun et al. cohort. This fits the hierarchical prior over $E_w$ and $V_0$, and the model is then frozen. Third, an external center using a different catheter, again with full titration as the reference standard, in 20–30 patients.
The frozen model is evaluated at the external center at target sample sizes of 0, 5, 10 and 20 patients. Target patients update only the population prior, never the bench curve. Locking the model before the external data arrive follows the derivation-then-blind-test discipline of Prediction of Outcome in Acute Lower Gastrointestinal Bleeding Using Gradient Boosting. That design trained at one hospital and tested at another.
Because Yet Another ICU Benchmark found that preprocessing choices move external performance more than model family does, both centers will share a published, fixed processing script. That script covers occlusion timing, cardiac-artifact filtering and end-expiratory sampling.
On top of the model we build a sequential titration rule. After each inflation, the posterior over the optimal range is updated, and the next volume is chosen where it most reduces uncertainty about the range bounds. Titration stops when the posterior range is narrower than 0.2 ml, matching the step size of the clinical protocol.
Baselines
The naive reference point is a fixed volume per catheter. For the external catheter, we take the midpoint of the bench working range reported by Yang et al. For the Cooper catheter, we take the 1.0 ml median best volume from Sun et al. This costs nothing at the bedside, so any proposed method must beat it on calibrated-pressure error to justify extra steps.
The strongest existing clinical method is the Sun et al. two-point calibration, transported unchanged. Its volumes are rescaled to the new catheter's bench range, since the original 0.6 and 1.4 ml are specific to the Cooper catheter. In its home setting it agreed with multi-point regression to a bias of −0.1 cmH₂O/ml, and its bench elastance estimate had a bias of 0.5 cmH₂O/ml. The proposal beats it if, at the external center, our two- or three-step method gives lower absolute error in calibrated end-expiratory pressure against the full-titration reference. That error must also stay within the ±1 cmH₂O tolerance in a larger fraction of patients, with the same number of inflations.
The ceiling is per-patient sigmoid fitting on the full 13-step titration (Yang et al.'s method applied in vivo). The goal is to match it with far fewer steps, not to beat it. On the bench, Yang et al. report all transmural pressures within ±1.0 cmH₂O at the fitted optimal volumes. No reported in-vivo result for sigmoid fitting was found on Lacuna.
A flat learned baseline is a gradient-boosting regressor that predicts the optimal range bounds from patient covariates, catheter identity and the first two inflation readings. It is the analogue of the boosting models that worked on small tabular cohorts in Ayaru et al. and of the balloon-volume model in Machine learning-based prediction of the risk of moderate-to-severe catheter-related bladder discomfort. That study found balloon volume to be the dominant predictor, with an internal validation AUC of 0.881, but at a single center.
The key ablation removes the bench term. It is the same hierarchical model with $R_c$ replaced by a generic sigmoid learned only from patient data. This isolates whether the bench-derived balloon curve is what makes transport work.
All learned variants get the same source data, the same external patients and the same hyperparameter budget, tuned by cross-validation on source data only. Everything is frozen before external data are unblinded.
Why not just…?
Why not just do the full titration on every patient? This is the real ceiling, and if bedside time and repeated occlusions were free it would win. The case for the model rests on two points. It needs fewer steps. It also gives an answer when the full curve is corrupted by cardiac oscillations or patient effort, a case the open-problems literature flags because current protocols were developed in paralyzed patients. The step-count experiment below settles whether the saving is large enough to matter. If three adaptive steps do not match thirteen fixed ones within 0.5 cmH₂O, the proposal reduces to an external-validation study of full titration. That would still be useful, but it is a smaller result.
Why not just carry over the two-point calibration? Yang et al. show that the linear section shifts with surrounding pressure and that working ranges differ severalfold between catheters. A fixed pair of volumes chosen for one 2.8 ml balloon has no principled transfer to a larger one. Whether rescaling by the bench range is enough is exactly what the baseline comparison tests. If rescaled two-point calibration performs as well externally, the conclusion is that bench ranges alone suffice, and we would report that.
Why not use a general domain-adaptation method? Instance re-weighting as in Boosting Transfer Learning with Survival Data from Heterogeneous Domains is formulated for censored survival outcomes and matches covariates without regard to physics. Input mapping into a frozen predictor, as in Adapt the Data, Not the Model, needs labeled target data and was built for ICU vital-sign series. Text summarization as in Can we generate portable representations for clinical time series data using LLMs? adapts from 16 samples but discards precisely the sub-cmH₂O differences that calibration depends on. A new catheter is a known, bench-measurable change, so it is cheaper to measure that shift directly than to learn it from scarce patients. The no-bench ablation is effectively this purely data-driven route, and it tests the claim directly.
Why not add a hierarchical statistical model without physics? That is the ablation. If it performs as well as the full model, the bench library is unnecessary and the simpler model should be recommended.
Proposed Experiments
The primary endpoint is absolute error in calibrated end-expiratory esophageal pressure against the full-titration reference at the external center, with a different catheter, at a target sample size of 10 patients. Secondary endpoints are the fraction of patients within ±1 cmH₂O, error in the predicted range bounds (ml), and the number of inflations needed.
A learning curve over target sample sizes of 0, 5, 10 and 20 shows how fast the prior adapts. The zero-patient point shows whether the bench curve plus the source prior already transports with no local patients.
The bench is the only setting with true surrounding pressure, so we also hold out one catheter model entirely on the bench and predict its in-chamber range from its free recoil curve alone. This is a direct test of whether $R_c$ captures what makes a balloon specific.
A drift check repeats the external evaluation on patients enrolled later in the study. All models are local argues that a single external test does not guarantee lasting transport, and this check probes that concern.
The decisive result for the idea is the zero- and five-patient external error compared with the rescaled two-point baseline and the no-bench ablation. If bench-informed transport does not beat both there, the decomposition is not earning its keep.
References
- Optimal esophageal balloon volume for accurate estimation of pleural pressure at end-expiration and end-inspiration: an in vitro bench experiment
- Use of esophageal balloon pressure-volume curve analysis to determine esophageal wall elastance and calibrate raw esophageal pressure: a bench experiment and clinical study
- A Left Ventricular Segmentation Method on 3D Echocardiography Using Deep Learning and Snake
- Esophageal Virtual Disease Landscape Using Mechanics-Informed Machine Learning
- Prediction of Outcome in Acute Lower Gastrointestinal Bleeding Using Gradient Boosting
- Yet Another ICU Benchmark: A Flexible Multi-Center Framework for Clinical ML
- Machine learning-based prediction of the risk of moderate-to-severe catheter-related bladder discomfort in general anaesthesia patients: a prospective cohort study
- Boosting Transfer Learning with Survival Data from Heterogeneous Domains
- Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors
- Can we generate portable representations for clinical time series data using LLMs?
- All models are local: time to replace external validation with recurrent local validation