B.A.R.T.: Biomarker Analysis for Respiratory Testing in Early Detection of Orodental Disease

Children in rural regions, low-income families, and communities without regular dental services often remain undiagnosed until damage becomes irreversible, ranking number #1 prevalent condition in low income areas.

Cavities & Oral Disease

Untreated cavities are about three times more common in children aged 2–5 from low-income households (18%), however is still prominent worldwide.

Cavities begin with subtle biochemical changes long before any structural damage OR pain occur. This phase is called biofilm acidification, effectively 95–99% are never noticed. However, biofilm acidification are caused by certain oral bacteria that produce measurable breath biomarkers.

Detection flaws in other diseases;

Conventional diagnostic pathways across major disease groups depend heavily on blood draws, tissue biopsies, and imaging, all of which are invasive, intermittent, and tuned to detect disease only after measurable structural or biochemical change has already occurred. These limitations (and more) exist in every major disease group. 80% of early molecular changes happen before biopsies. Diabetes requires frequent invasive blood tests, and early metabolic imbalances are often missed, because glucose only change after ketones begin rising.

According to World Health Organization, a primary contributing factor is late detection and low awareness, effectively causing preventive issues to build up into chronic illness. Breath Biopsy is the process of collecting and analyzing the VOC chemicals in exhaled breath to detect biological changes inside the body. This method of detection has been studied for preventive measure in a multitude of illness, but lacks research in oro-dental disease.

I started by selecting the biomarkers for sensors: acetone (a ketone body from fatty-acid oxidation, linked to diabetic ketoacidosis), ammonia (cleared by the liver and kidneys), hydrogen sulfide (from oral microbiota, tied to periodontal disease), TVOCs (a marker of oxidative stress and inflammation), and molecular hydrogen (H₂) (a gut-fermentation by-product). Breath enters through a disposable mouthpiece with a one-way valve, keeping clean sampling without back-inhalation. I then chose the sensor types; metal-oxide semiconductor (MOS) sensors, which work using electrochemical cells generating a current from a reaction at a working electrode. From there, I designed, wired, and soldered the full hardware (refer to figure 9); Most hardware was to ensure power regulation, as the sensors give readings in voltages. The enclosure was 3D printed in PLA;

Air Chambers / Plenum: (145.7 mL) Splits one breath into five even, smoothed flows. Acts as a pressure-equalizing manifold; the expanded volume drops flow velocity and dampens turbulence so each sensor receives a representative aliquot of the same exhalation.

Sensor Chambers: (One chamber = 41.6 mL) Five identical compartments, one sensor each at the chamber floor. A central aperture directs flow onto the sensing element, maximizing gas-surface contact. The base plate is detachable for inspection and recalibration.

Purging & Fan Chamber: (25 seconds for 99.97% cleared) A PWM-controlled fan drives a full purge cycle between samples to eliminate sample carry-over, a known confounding variable in serial breath analysis.

Firmware on an ESP32-S3 reads both ADCs, applies a form of filtering to reject transient noise, converts voltage readings to ppm/ppb concentrations through per-sensor calibration curves, and sends the results over BLE to a mobile app or Raspberry Pi. A finite state machine sequences warm-up, baseline, sampling, and purge phases, while readings stream to the onboard LCD and log to flash for later analysis.

Software, GUI, and Front-end

BART walks the user through a five-minute screening with prompts on screen at every stage. To begin, the user closes their lips around a straw-shaped intake tube and exhales while live voltage traces from each sensor stream onto the LCD. The app sequences four phases in order: a 30-second baseline that lets the chamber settle to room air, a 5-second ready cue, a 10-second "BLOW NOW" prompt, and finally a 30-second purge. Once breath capture finishes, the screen transitions into the endoscopy phase, where the user is prompted to take intraoral photos across ten mapped dental zones, giving the AI report visual context to pair with the gas readings. The user is then given their readings, alongside what they may mean, warning, and the chance to ask questions.

Back-End Software & Calibration Math

Internally, the ESP32-S3 samples both ADS1115 ADCs at 5.12 Hz, applies median filtering to reject transient noise, and converts raw voltages into ppm/ppb concentrations. For the four MEMS sensors (H₂S, NH₃, VOC, H₂), the firmware first calculates sensor resistance using

Rs = R_L × (V_cc − V_out) / V_out with a 4.7 kΩ load resistor.

After a 60-second warm-up in clean air, R₀ is captured as a baseline anchor; thereafter, concentration is estimated through a log-linear power law: ppm = a × (Rs/R₀)^b, with sensor-specific constants pulled from each datasheet's sensitivity curve. Acetone uses a different math path because the TGS1820 is a Wheatstone-bridge sensor, so it follows a linear ΔV approximation against its baseline. These are all done on the ESP.

Bump Testing & Sensor Validation

Sensors were validated using a 500 mL glass jar fed by a 1.5 L/min air pump, routed through BART's mixing plenum, air chamber, sensor stack, and purge cycle. Known volumes of 80% acetone solution (m, in mL) were injected, and the percent change from baseline was modelled as:

%Δ(m) = 100 · [ f( C_jar(m) · (1 − e^(−Qt/V)) ) − S₀ ] / S₀

where m = injected acetone volume, S(m) = predicted sensor value after exposure, and S₀ = clean-air baseline. Substituting Q = 25, V = 187.3, t = 35, and S₀ = 0.6381 gave:

%Δ(m) = 100 · [ f(0.99065 · C_jar(m)) − 0.6381 ] / 0.6381

Across the tested range, the response was approximately linear, simplifying to:

%Δ(m) ≈ 50m − 57.4

This confirmed the datasheet equations held in BART's actual chamber geometry and let the firmware predict sensor response directly from injected concentration (refer to figure leaveblank).

Clinical Validation

To strengthen the ppb thresholds and capture realistic breath signatures, I reached out to a clinic and collected multiple samples from patients with diagnosed oral disease. Working with real human breath, rather than synthetic gas mixtures, let me refine the calibration, study how disease-state patterns diverge from healthy baselines, and tighten the triage logic. After incorporating this data, the model scored 89% accuracy on single-test classification, averaging 93% across patients with multiple sessions; both figures approach the agreement levels seen in GC-MS, the clinical gold standard, at a fraction of the size and cost.

Cavities, periodontal disease, and a long list of systemic illnesses all begin at the molecular level, often months or years before any structural damage, pain, or visible sign appears, yet the tools we currently rely on for detection (blood draws, biopsies, dental X-rays) are invasive, infrequent, and tuned to catch disease only after it has progressed. For children in rural regions, low-income families, and communities without regular access to dental care, this delay is the difference between a preventable issue and a permanent one.

BART demonstrates that a small, wireless, multi-gas breath analyzer built from off-the-shelf MOS modules and a Raspberry Pi 5 and/or the mobile app hub is sufficient to detect breath-chemistry patterns that align quantitatively and qualitatively with the published clinical literature on halitosis, periodontitis, and breath acetone. Across 173 capture sessions on 131 subjects, the device flagged 24 sessions as periodontal-suspicious, 26 as caries-suspicious, and six as ketosis-suspicious, with three subjects positive on all three dental axes simultaneously and a sub-set of acetone-positive sessions that warrants metabolic follow-up. The single most consequential reliability mechanism added to the device, the runtime R₀-drift guard, has already prevented at least two false-positive sessions and is, in my view, a methodological prerequisite for any future MOS-array breath device.

The single most important outstanding piece of work is the construction of a clinician-labeled reference set so that descriptive results can be replaced by validated precision and recall.

The next phase is expanding BART's reach across both hardware and disease coverage. Migration to a custom PCB is already in progress, replacing the current layout with a compact, manufacturable board further reducing size and greatly reducing cost.

Beyond oral disease, I plan to train dedicated AI models for other condition groups; diabetes, gastrointestinal disorders, and respiratory illness, since the same five-biomarker panel already overlaps with their known signatures. A larger validation rig built around gas barrels would let me characterize sensor behavior at higher concentrations and extended exposure times, and broader testing will further strengthen the dataset.

First and foremost, I want to thank my parents for their consistent support throughout this project. There were many moments when BART felt overwhelmingly advanced. I am also grateful to my school for their support along the way and simply being there whenever I needed guidance.

I would also like to extend my sincere thanks to my mentors, Dr. Manuel Lagravere Vich and Dr. Hollis Lai from the University of Alberta, for offering insightful feedback on the the project and supporting its direction at important stages. I also thank Dr. Gaurav Sharma from Parhar Dental Group for his advice on the direction of the project. A heartfelt thank you to Enjoy Dental for taking the time to sample patients, as it directly contributed to BART's success!

Projectboard References

[1] World Health Organization. (2022). Global oral health status report: Towards universal health coverage for oral health by 2030. https://www.who.int/team/noncommunicable-diseases/global-status-report-on-oral-health-2022

[2] Balaji, S. M. (2022). Global oral health report – WHO. Journal of Global Oral Health, 5(2), 61–62. https://doi.org/10.25259/JGOH_50_2022

[3] Sahni, V. (2022). WHO report. British Dental Journal, 233(11), 982. https://doi.org/10.1038/s41415-022-5364-6

[4] Jain, A., Bhaskar, D. J., Gupta, D., Yadav, P., Dalai, D. R., Jhingala, V., Bumb, S. S., & Karim, F. (2024). WHO's global oral health status report 2022: Actions, discussion and implementation. Oral Diseases, 30(2), 73–79. https://doi.org/10.1111/odi.14516

[5] Centers for Disease Control and Prevention. (2024). Health disparities in oral health. U.S. Department of Health and Human Services. https://www.cdc.gov/oral-health/health-equity/index.html

[6] Shastri, M. D., Allam, V. S. R. R., Shukla, S. D., Jha, N. K., Paudel, K. R., Peterson, G. M., Patel, R. P., Hansbro, P. M., Chellappan, D. K., & Dua, K. (2021). Streptococcus mutans and dental caries: A review of pathogenesis, biofilm formation, and management. Microbial Pathogenesis, 159, 105132.

[7] Kageyama, S., Furuta, M., Takeshita, T., Ma, J., Asakawa, M., & Yamashita, Y. (2019). High-level acidogenic microbiomes in Lactobacillus-rich saliva are associated with caries severity. Microbiology Spectrum, 7(3).

[8] Owlstone Medical. (2024). Non-invasive VOC biomarkers. https://www.owlstonemedical.com/science-technology/voc-biomarkers/

[9] Sociedade, F., Lima, A. R., Pinto, J., Carvalho-Maia, C., Jerónimo, C., Henrique, R., & Bastos, M. de L. (2023). Breath volatile organic compounds (VOCs) as biomarkers for the diagnosis of pathological conditions: A review. Biomedical Journal, 46(4), 100623. https://doi.org/10.1016/j.bj.2023.04.001

[10] Ferrari, A. M., & Pellegrino, M. (2014). Volatile organic compounds as exhaled biomarkers of inflammation and oxidative stress in respiratory diseases. In Studies on respiratory disorders: Oxidative stress in applied basic research and clinical practice (pp. 75–105). Springer. https://doi.org/10.1007/978-1-4939-0497-6_4

[11] Schulz, E., Woollam, M., Grocki, P., Davis, M. D., & Agarwal, M. (2023). Methods to detect volatile organic compounds for breath biopsy using solid-phase microextraction and gas chromatography–mass spectrometry. Molecules, 28(11), 4533. https://doi.org/10.3390/molecules28114533

[12] Chan, M.-J., Li, Y.-J., Wu, C.-C., Lee, Y.-C., Zan, H.-W., Meng, H.-F., Hsieh, M.-H., Lai, C.-S., & Tian, Y.-C. (2020). Breath ammonia is a useful biomarker predicting kidney function in chronic kidney disease patients. Biomedicines, 8(11), 468. https://doi.org/10.3390/biomedicines8110468

[13] Qiao, Y., Gao, Z., Liu, Y., Cheng, Y., Yu, M., Zhao, L., Duan, Y., & Liu, Y. (2014). Breath ketone testing: A new biomarker for diagnosis and therapeutic monitoring of diabetic ketosis. BioMed Research International, 2014, 869186. https://doi.org/10.1155/2014/869186

[14] Saxena, R., Yadav, P., Bhardwaj, A., & Kumar, A. (2018). Sensing technologies for detection of acetone in human breath for diabetes diagnosis and monitoring. Sensors, 18(8), 2581.

Image Sources

[I-1] World Health Organization. (2022). Global oral health status report: Towards universal health coverage for oral health by 2030. World Health Organization. https://www.who.int/team/noncommunicable-diseases/global-status-report-on-oral-health-2022

[I-2] U.S. Department of Health and Human Services. (2000). Oral health in America: A report of the Surgeon General. National Institute of Dental and Craniofacial Research. https://www.nidcr.nih.gov/sites/default/files/2017-10/[email protected]

[I-3] Biofilm acidification diagram. Original figure created by the author.

[I-4] National Center for Health Statistics. (2024). Oral and dental health statistics. Centers for Disease Control and Prevention. https://www.cdc.gov/nchs/fastats/dental.htm

[I-5] Owlstone Medical. (2024). Breath Biopsy: Non-invasive VOC biomarkers. https://www.owlstonemedical.com/science-technology/voc-biomarkers/

[I-6] Figaro USA, Inc. (2005). General information for TGS sensors (Rev. 03/05). Figaro Engineering. https://www.electronicaembajadores.com/datos/pdf2/ss/ssga/tgs.pdf