Transcutaneous CO2 monitoring is not a new idea. Clinicians have long measured the partial pressure of carbon dioxide that diffuses out through the skin to track breathing adequacy in patients on ventilators, people with sleep apnea, and infants in intensive care. But conventional transcutaneous sensors are bulky, heated, and tethered to bedside equipment, because they must warm the skin to several degrees above body temperature to drive enough gas through the surface. Recent work has chased a lighter goal: a sensor small and low-power enough to be worn on the wrist during daily life, sleep, or exercise. The catch is that skin-emitted gas is extraordinarily hard to measure cleanly. Sweat, humidity, condensation, and the tiny shifts of the sensor against the skin all conspire to distort the infrared signal, sometimes producing readings that drift uselessly for many minutes before stabilizing.

The core detection technology in the new prototype is NDIR sensing, which exploits a simple physical fact: carbon dioxide molecules absorb infrared light strongly at a characteristic wavelength near 4.26 micrometers. The sensor shines broadband infrared light across a small optical cavity and measures how much of that wavelength arrives at a detector. The more CO2 in the cavity, the dimmer the detected light. In this prototype, the miniature NDIR module sits inside a 3D-printed microcavity mounted on a wristband, so the gas diffusing out of the skin accumulates in a tiny enclosed volume directly above the sensor’s optical path. Because the cavity is small, the gas concentration rises quickly and steeply, which is exactly what a sensor trying to pull a fast reading needs. The team, working with clinician Yunying Cai of the First People’s Hospital of Yunnan Province and Heng Su of Fuwai Yunnan Hospital, tested the device on the wrists of healthy adult volunteers in a low-risk, non-invasive observational study.

Even inside a microcavity, however, the raw signal remained hostage to the environment above the skin. Water vapor accumulates faster than CO2 in the confined space, condensation can fog optical surfaces, and micro-movements of the wristband change the diffusion geometry from second to second. The researchers’ first fix was a physical one: a membrane of polydimethylsiloxane, or PDMS, placed between the skin and the cavity. PDMS is a flexible silicone elastomer — commercially famous as the base of Silly Putty-like materials and ubiquitous in microfluidics — with a remarkable property for this application: although it is nominally hydrophobic, it is extraordinarily permeable to small gas molecules, including CO2, while suppressing the passage of liquid water and dampening the turbulence caused by movement. A 100-micrometer PDMS film is thin enough that CO2 diffuses through in seconds, yet robust enough to act as a mechanical and humidity buffer between sweaty skin and the optical cavity.

The second fix was algorithmic. Rather than waiting for the signal to plateau — which can take many minutes — the team asked a different question: at what moment does the sensor trace show a clean, physically meaningful upward sweep that can be trusted? To answer it, they first preprocessed the raw trace with a median filter and local polynomial smoothing, two standard techniques that suppress spike artifacts and slow wander without distorting genuine trends. Then they applied a sliding-window linear fit: across each successive short stretch of the smoothed signal, they fitted a straight line and computed three quantities — the coefficient of determination R-squared, which measures how well a straight line explains the data; the drop ratio, which flags intervals where the signal is falling or chaotic rather than rising; and the slope, which measures how fast the signal is climbing. Only windows that passed all three constraints were declared valid rate windows, and the window with the maximum R-squared defined the earliest trustworthy reading.

The head-to-head results were striking. In recordings from eight human wrist sessions without any membrane, the algorithm could identify a valid rate window in only four of eight cases — the other half of signals were too unstable or chaotic to interpret. With the PDMS membrane in place, all four PDMS-condition recordings yielded valid windows. The PDMS traces also climbed more consistently: the median rise rate was 0.954 raw units per second with a tight interquartile range of just 0.105, whereas the unmembraned traces, though faster on paper at a median 1.159 per second, showed a five-fold wider spread of 0.581 — a sign that some of those steep climbs were being driven by humidity and motion artifacts rather than by CO2. The median time at the center of the valid window also shifted from 360 seconds without the membrane to 280.5 seconds with it, meaning the silicone layer both stabilized the signal and moved the trustworthy reading earlier.

The team did not stop at the primary comparison. A series of supplementary experiments probed the physical design parameters. They tested different gas-gap heights between the sensor and the skin and found that a 0.3-centimeter gap gave the best balance between fast accumulation and mechanical tolerance. They varied PDMS film thickness and confirmed that the 100-micrometer membrane delivered repeatable performance, thick enough for robustness yet thin enough for rapid gas transfer. In repeatability trials, the same volunteer sessions yielded valid windows again and again rather than succeeding only by chance. And in a post-exercise experiment — arguably the most demanding scenario for any wearable, since elevated skin temperature, sweat, and blood flow all change dramatically after exertion — the early-window extraction still worked, forming clean rate windows in the critical minutes right after physical activity when metabolic CO2 output is highest.

The significance of the R-squared-maximum method goes beyond this one device. Continuous biomedical signals are full of intervals that look plausible to a naive algorithm but are actually contaminated by artifacts, and most wearable systems simply delay reporting until a signal stabilizes, wasting the most informative early minutes of a recording. By demanding statistical quality — linearity, monotonic rise, adequate slope — before accepting a reading, the method turns the sensor’s own rise transient into an asset rather than a nuisance. The moment when CO2 floods the microcavity is precisely when the signal carries the most information about the rate of gas emission from the skin, and the algorithm now knows how to harvest that moment safely. This echoes a classical lineage: the sliding-window linear fit refines ideas dating back to Savitzky–Golay smoothing, while the robust statistics of outliers, drawing on decades of work from Rousseeuw onward, guard against false positives.

The authors are careful about scope, and that caution matters. This is a prototype-level engineering validation on healthy volunteers, not a clinical accuracy trial. The study demonstrates that the raw signal is stable, that early windows can be extracted reproducibly, and that the design choices are sound — but it does not yet establish that the device’s numbers can stand in for arterial or clinical transcutaneous PCO2 measurements in patients. Calibration against gold-standard blood gas analysis, testing in people with respiratory disease, and validation during sleep and long-term wear remain future work. The prototype also deliberately skipped the skin-heating elements that conventional transcutaneous monitors use, which means its readings reflect gas diffusing at native skin temperature; whether this suffices for quantitative clinical use is precisely what the next phase must determine.

Still, the direction of travel is clear. A wearable that can hang off a wristband, run a low-power infrared sensor, shrug off sweat and motion through a silicone interface, and tell its user — or a physician — within about four to five minutes that its readings can be trusted, would transform how breathing adequacy is monitored outside the hospital. For the millions of people with COPD, sleep-disordered breathing, or neuromuscular conditions who need intermittent CO2 checks without needles or clinic visits, the combination of a humble silicone film and a clever statistics trick may prove to be the difference between a laboratory curiosity and a device worn every day. The Kunming team’s prototype does not yet diagnose anything — but it has demonstrated the two things every diagnostic wearable must have first: a signal you can believe, and a way of knowing when to believe it.

Subject of Research: A wearable microcavity NDIR transcutaneous CO2 sensing prototype using a PDMS interface layer and rate-window signal analysis

Article Title: Artifact-resilient design and early reading extraction in a wearable microcavity NDIR transcutaneous CO2 prototype: validation of a PDMS interface layer and the R2-max rate-window method

Article References: Xiong, X., Xu, H., Cai, Y., Mo, X., Su, H., & He, J. (2026). Artifact-resilient design and early reading extraction in a wearable microcavity NDIR transcutaneous CO2 prototype: validation of a PDMS interface layer and the R2-max rate-window method. Biomedical Engineering Letters. https://doi.org/10.1007/s13534-026-00611-x

Image Credits: AI Generated

DOI: 10.1007/s13534-026-00611-x

Keywords: wearable sensors, transcutaneous CO2, NDIR spectroscopy, PDMS membrane, artifact suppression, rate-window method, biomedical engineering, non-invasive monitoring, signal processing, microcavity sensor, respiratory monitoring, Kalman-free smoothing