MOS acetone sensors based on platinum doped tungsten oxide (SB-33) were purchased from FIS, Inc. Non-rebreathing T-pieces were purchased from Vacumetrics, Inc. Teflon membranes (VE8) were purchased from Membrane Solutions, Inc. The humidity sensor (HIH-5030-001) was purchased from Honeywell, Inc.

The assembly diagram of the diffusive gas separation device is shown in Figure 1A, a non-rebreathing T-piece with two check valves is used for breath sampling and gas separation. The MOS sensor is placed in the reservoir for acetone sensing. Direct airflow on the sensor surface will change the conductivity of the sensor by altering the heated layer temperature. Therefore, we placed a porous Teflon membrane between the reservoir and the sensor. The membrane allows diffusion of the gases, while reducing the influence of the air speed. In addition, a humidity sensor is integrated into the reservoir for humidity monitoring and correction.

When the commercial MOS sensor (SB-33) is exposed to acetone, acetone reacts with oxygen ions adsorbed on the surface of the platinum doped tungsten oxide, resulting in electron transfer and the impedance change of the sensor.

The MOS sensor resistance was detected using a STM32 MCU based module integrated with an AD7792 A/D converter. We monitored the sensor baseline (R_air) for 20 s before tests, and the response of the MOS sensor was defined as the ratio of resistance (R _air/R).

In addition, the concentrations of BrAce were validated using a selected ion flow tube-mass spectrometer (SIFT-MS) (Instrument Science, Profile Series, Crewe, United Kingdom) in multiple ion monitoring (MIM) modes. H3O+ (m/z: 19) was chosen as the precursor ion for reaction with breath samples. The diffusive gas separation device was directly connected to the SIFT-MS injector (Supplementary Figure S1), so that a breath sample can be tested by both the acetone sensor and the SIFT-MS.

To collect a breath sample, a human subject directly breathes into the T-piece for about 10 s to trap the end portion of the breath. During exhalation, the elastic membranes of the check valves are bent by the breath pressure, allowing the breath to pass through (Figure 1A). After exhalation stops, the two elastic membranes restore their original flat shape (Figure 1B), which forms a breath reservoir, traps the breath sample, and prevents reverse airflow caused by inhalation. Captured gas molecules can escape from the reservoir through the small gaps between the elastic membrane and its holder (Figure 1C), and different types of gas molecules escape at different speeds because of different diffusion coefficients. Thus, the ratio of acetone concentration to interference gases changes over time. By optimizing the diffusion and measurement time, acetone can be detected with improved accuracy.

Figure 2A shows the signal of one online breath test. The MOS sensor responds immediately after exhalation starts. After the exhalation stops, it takes some additional time for the gas molecules to diffuse to the MOS sensor before peak response is reached. Because the gas molecules keep diffusing into the ambient air, the signal decreases after the peak response. We further conducted three consecutive real breath tests in Supplementary Figure S2, which showed a good consistency. An intuitive way of signal analysis is simply using the peak value. To evaluate the performance, we tested the breath samples of one subject for 3 days. A selected ion flow tube mass spectrometer (SIFT-MS) was used as a reference instrument. As shown in Figure 2B, there is a clear offset between day 1 and day 3. Since no alcohol was consumed, and the ambient humidity was similar for the 3 days, other interference gases in exhaled breath can be the cause of the offset.

To verify that the MOS sensor responds to other gas components in exhaled breath, a gas washing bottle filled with water was connected to the inlet to filter out acetone in breath. The removal of acetone by the bottle filter was verified by SIFT-MS (Supplementary Figure S3). Figure 3A shows the results of two consecutive breath tests, one with a filter and one without. It can be observed that the MOS sensor showed a significant response to the filtered breath. The peak response shows up earlier and the signal decreases faster than the unfiltered breath, which indicates that the interference gases in the filtered breath have large diffusion coefficients. Hydrogen is probably the major interference component, which is not absorbed by water and has a large diffusion coefficient compared to acetone. Supplementary Figure S4 shows the cross-sensitivity of the used MOS acetone sensor. Although it is less sensitive to hydrogen than acetone, the hydrogen concentration in breath can increase up to 50 ppm. Hydrogen interference is a major issue not only because of its high concentration in exhaled breath but also because its presence is hard to control by diet like breath alcohol. We solve the hydrogen interference issue by using diffusive gas separation (Figure 3B). 790 ppb acetone and 50 ppm hydrogen were prepared representing a normal BrAce level and a high breath hydrogen level. The two samples were pumped through the T-piece separately. Although the MOS sensor’s response to 50 ppm hydrogen was much higher than 790 ppb acetone, the response of the MOS sensor to hydrogen decreased much faster than acetone after sampling stopped. At 130 s, almost no hydrogen can be detected, but the response of the MOS sensor to acetone is still about 60% of the peak response. In addition, we further confirmed the diffusion coefficients of acetone-air (0.124 cm²s⁻¹) and hydrogen-air (0.627 cm²s⁻¹) under the circumstance (p = 101.325 kPa, T = 293.15 K) (Marrero and Mason, 1972).

Conventionally, MOS acetone sensors are calibrated and tested using synthetic gas samples, but this doesn’t translate into accurate acetone detection in real breath samples, considering the complex components in exhaled breath. To calibrate the device, 37 real breath samples were tested in the same way as real application scenarios, i.e., directly breath into the device for 10 s. Humidity interference is another common issue for gas sensing. Although the humidity level in the reservoir is always 100 %RH after breath sampling, it can be different before each test, so a humidity sensor was placed in the reservoir to record the initial humidity level. The acetone concentration of each breath sample was confirmed by SIFT-MS. Figure 4A shows the MOS sensor signals of all the tests at 100 s (0 s indicates the start of breath sampling). The color of each data point represents the initial humidity level in the reservoir, which clearly has an impact on the response of the MOS sensor. A multiple linear regression equation was used to predict acetone concentration: [ A c e t o n e ] = A 1 + A 2 × ( R a i r R 100 s ) + A 3 × H (1) Where [Acetone] is the acetone concentration, R _air is the resistance of the MOS sensor before breath sampling, R _100s is the resistance of the MOS sensor 100 s after breath sampling starts, and H is the initial humidity level. Figure 4B shows that the BrAce concentrations obtained from our device and the SIFT-MS have a strong correlation (R = 0.95, p < 0.001). The best fitting values of A1, A2 and A3 for Eq. 1 are given in Supplementary Table S1.

The limit of agreement between the device and the SIFT-MS was determined from a Bland-Altman plot as shown in Figure 4C. The difference between the readings from the MOS sensor and the corresponding reading from SIFT-MS was plotted against the average reading of SIFT-MS and the MOS sensor. The limit of agreement was ±1.96 standard deviations (SD) from the mean, which was 321 ppb. This means that within 95% confidence the reading from the MOS sensor is within 321 ppb difference from the reading obtained from SIFT-MS. Regression equations using MOS signals at different time points were compared to show the benefits of the used approach (Figure 4D). Simply using the peak response value leads to the largest variation between the MOS sensor and SIFT-MS. Including humidity signal can reduce the variation. The smallest variation is obtained using R _150s (SD = 158 ppb). This allows accurate detection of low concentration BrAce.

As a demo application, here we show that the device can be used to track diet (Figure 5). On day 1, the initial high BrAce level of a healthy male subject was reached via a high fat and low carbohydrate dinner the day before. A high carbohydrate breakfast about 4 h before the test was consumed to switch body energy source from fat to carbohydrate. It is noteworthy that eating high carb content will make bacterias in the gut produce methane from breaking down the carbs. However, the sensor (SB-33) employed in this study is much more sensitive to acetone than methane as shown in Supplementary Figure S4 in Supplementary material. As expected, a continuous decrease of BrAce was detected by both the device and SIFT-MS. A control experiment performed on day 2 shows the detection of basal BrAce under normal diet conditions. The good correlation and agreement between the device and SIFT-MS validate its reliability in the detection of real breath samples. Table 1 compares this work with recent studies of MOS acetone sensors, showing its unique feature of online breath sampling, yet with the capability of quantifying low-level BrAce.