Automated quality control tool quantifies how breathing irregularities skew lung washout test results

Every few minutes of quiet breathing can reveal how unevenly air moves through a patient’s lungs. The multiple-breath washout test, in which a person breathes pure oxygen to flush nitrogen out of the lungs, has become one of the most sensitive tools for detecting early lung disease, particularly in children with cystic fibrosis. Yet the test has a stubborn weakness: it depends on the patient breathing regularly for several minutes at a time, something that young children and people with diseased lungs often cannot manage. Coughs, sighs, leaks around the mouthpiece, and drifts in lung volume can all distort the final numbers. A new study published in Frontiers in Medical Engineering describes an automated, model-based quality control system that does something no previous tool could do: it estimates precisely how much each irregular behavior actually changes the test’s primary outcomes.

The research, led by Florian Wyler and colleagues at Bern University Hospital and the University of Bern in Switzerland, tackles a problem that has long frustrated clinicians and researchers alike. The main outputs of a washout test are the lung clearance index, or LCI, which measures how many lung turnovers are needed to clear the resident nitrogen, and the functional residual capacity, or FRC, the volume of air remaining in the lungs at the end of a normal breath out. Irregular breathing behaviors can bias both numbers, but the current gold standard for judging whether a measurement is trustworthy is expert quality control: a trained physician, nurse, or scientist visually inspects the signals and decides whether to accept or reject the test. That judgment is slow, requires rare expertise, and, crucially, is subjective.

The subjectivity is not a minor nuisance. Washout measurements typically last several minutes and are routinely performed in awake children from preschool age onward, so measurements full of minor, ambiguous irregularities are the norm rather than the exception. Two experienced raters can disagree about whether a given test should be kept, especially when the irregularities fall into a grey zone between clearly acceptable and clearly unusable. Existing guidelines also rely on fixed empirical thresholds, such as allowable variation in end-expiratory lung volume, without any direct way of estimating what those deviations actually do to the LCI or FRC. The Bern team set out to replace that guesswork with a quantitative, reproducible estimate of impact.

The core of the new approach is a mathematical model of how nitrogen is diluted during the washout. For any given breath, the model relates an input of breathing behavior to an expected output of nitrogen dilution, taking into account the current end-expiratory lung volume, physiological and machine dead spaces, and the fact that washout efficiency declines as the test proceeds, because well-ventilated lung regions empty first while poorly ventilated regions hold on to their nitrogen longer. The model’s parameters are estimated individually from each measurement, using the carbon dioxide and flow signals to locate dead space and fitting a linear efficiency curve over the course of the washout.

With this model in place, the algorithm runs paired simulations for each measurement. One simulation uses the observed breathing behavior, including whatever irregularities occurred; the other uses an idealized version in which the behavior in question, say a drifting end-expiratory lung volume, is replaced by a stable target. If the two simulations produce nearly identical LCI and FRC values, the irregularity was harmless. If they diverge, the size of the divergence, expressed as a percentage, becomes a behavior-specific quality control outcome. The team built seven such tests, covering changes in end-expiratory lung volume, tidal volume variability and size, leaks or trapped gas, the method used to determine end-tidal nitrogen concentration, signal synchronization, and breath detection. A measurement was deemed acceptable if no single test exceeded an estimated impact of 15 percent.

To validate the tool, the researchers re-analyzed an extraordinary dataset: 2,471 nitrogen washout measurements from 84 children with cystic fibrosis, collected between 2012 and 2019 across 754 visits, each visit including at least three completed measurements. The algorithm ran successfully on every single measurement without any manual intervention, taking roughly four seconds per test on an ordinary office computer. That speed matters, because expert quality control is one of the practical barriers to wider adoption of washout testing in clinical routine. The overall acceptance rate was 76 percent of visits, and the mean per-visit variance of LCI dropped from 0.72 turnovers when all measurements were included to 0.39 turnovers when only algorithm-accepted measurements were counted.

The most striking finding is how much of the natural variability in washout results can be attributed to quality control factors. The combined impact estimate from all tests explained around 45 percent of the within-visit variability of LCI in the children with cystic fibrosis, and the median Pearson correlation between estimated impact and observed within-visit LCI variation was 0.83. The largest single effect on LCI came from leaks or trapped gas, which exceeded the acceptability threshold in 31 percent of measurements, followed by changes in end-expiratory lung volume at 23 percent and variations in breath size around the end of the test at 19 percent. Leaks and trapped gas are particularly common in cystic fibrosis because the disease produces extended washout tails, in which the end-tidal nitrogen concentration hovers near the end-of-test threshold and occasionally ticks upward as poorly ventilated pockets of gas are slowly released.

The team also compared the automated judgments with human experts using a second dataset of 100 measurements from six-year-old children, half with cystic fibrosis and half healthy, drawn from the Bern and Swiss infant lung development cohorts. Agreement between the algorithm and expert raters was moderate, with Cohen’s kappa ranging from 0.32 against the rating made on the day of the test to 0.46 against a retrospective rating. Notably, that level of agreement was comparable to the agreement between two experienced human raters themselves, which stood at 0.65. Some disagreements were instructive: current guidelines instruct raters to reject any measurement with a leak, no matter how small, whereas the automated system tolerates minor leaks as long as their estimated impact on outcomes remains negligible. Conversely, experts sometimes struggled to judge visually whether a transient shift in end-expiratory lung volume mattered, while the model could quantify it directly.

The implications extend beyond cystic fibrosis clinics. Because the algorithm judges every irregularity on the same scale, its estimated impact on the primary outcomes, it replaces a patchwork of device-specific and behavior-specific thresholds with a single interpretable metric. The threshold itself can be tuned to the purpose of the test: a strict 15 percent cutoff for research repeatability, or a looser criterion when a clinical LCI is already far above the upper limit of normal and a modest bias would not change the diagnosis. The tool can also be applied retrospectively when measurements are re-analyzed with updated algorithms, and it inherently distinguishes between signal anomalies caused by genuine leaks and those fully explained by the patient’s underlying breathing pattern, a distinction that is especially valuable in patients who cannot achieve high-quality tidal breathing.

Limitations remain. The algorithm was developed for nitrogen washout data from the Ecomedics Exhalyzer D and relies on carbon dioxide, tracer gas, and flow signals; some tests are specific to that device setup, though the authors note that only minor adaptations would be needed for sulfur hexafluoride measurements or other devices. The current implementation runs after the measurement is complete, so a live version that warns operators mid-test to restart a compromised measurement would require further development. Still, the study demonstrates a feasible, fast, and reproducible method for assigning impact estimates to breathing irregularities, and it explains a substantial share of the measurement noise that has long plagued washout testing. For a technique whose value depends on detecting subtle changes in ventilation inhomogeneity over years of follow-up, an objective quality control standard could make the difference between a noisy research tool and a dependable clinical instrument.

Subject of Research: Automated model-based quality control for multiple-breath washout lung function testing in children

Article Title: Quality control in multiple-breath washout measurements: an automated, model-based approach quantifies effects on primary outcomes

Article References: Wyler, F., Borer, S., Curdy, M., Bovermann, X., Frauchiger, B. S., & Latzin, P. (2026). Quality control in multiple-breath washout measurements: an automated, model-based approach quantifies effects on primary outcomes. Frontiers in Medical Engineering, 4, Article 1774073. https://doi.org/10.3389/fmede.2026.1774073

Image Credits: AI Generated

DOI: 10.3389/fmede.2026.1774073

Keywords: multiple-breath washout, lung clearance index, quality control, cystic fibrosis, functional residual capacity, pediatric lung function, nitrogen washout, ventilation inhomogeneity, automated algorithm, respiratory medicine, clinical engineering, Bern University Hospital