NephroSense: Developing a Chronic Kidney Disease Screening Model for AI Breath Biomarker Testing

Biomarkers and Screening

Biomarkers are bodily chemicals that fluctuate in concentration with metabolic diseases. Common methods for proactive screening, which utilize said biomarkers as measurable signs of numerous conditions, are blood and urine sampling. These tests are crucial for early disease detection, allowing for increased treatment efficacy, reduced long-term expenses, and overall health maintenance (see Figure 1).

Problems

Regardless of their benefits, current biomarker techniques are often invasive. They require the extraction of bodily substances, which can induce patient anxiety; for instance, over 3/4 of American adults fear blood tests (see Figure 2). Additionally, the tests may take weeks for any results due to laboratory analyses. These delays further compound time beyond the screening process, as testing and treatment typically ensue. Other than patient convenience, common biomarker tests also have devastating environmental impacts from their biohazardous approaches (e.g., they contribute to over 15% of all healthcare-related waste, as seen in Figure 3).

Engineering Objectives

Despite being limited in application, breath analyses mitigate the flaws of current alternatives (see Figure 4). NephroSense utilizes them in a proof-of-concept prototype, testing for CKD through exhaled ammonia (see Figure 5). In doing so, the following objectives were defined:

Proactively address possible issues (e.g., insufficient diffusion and potential false positives from external variables).

Reliably calibrate sensor readings to patient ammonia concentrations.

Comply with controlled sampling standards in the physical design.

Accurately assess CKD probabilities and stages.

Corroborate the viability of proposed adaptations like a gas chromatography-mass spectrometry (GC-MS) device for generalized screening.

Physical Prototype

The physical device was built for the controlled collection of the breath samples needed for the machine learning evaluations. To construct it, the sensor and electrical components were connected to enable real-time voltage readings, which were translated into ammonia concentrations using a two-point calibration formula derived from the manufacturer-provided reference points of 1.4 V and 2 ppm (see Figure 6).

To maintain gas-sampling standards, the sensor compartment was lined with 5 layers of aluminum. Similarly, polytetrafluoroethylene (PTFE) was used for the collection tube, which was connected to the device through a check valve to inhibit backflow contamination of the user and unintended diffusion.

Timing

Voltages recorded before the initial stabilization time or after the plateau limit were not considered when computing the baseline and user voltages (Vout0 and Voutx, respectively) for the calibration formula. Meanwhile, the minimum purging time was used to ensure baseline re-establishment between usage diffusions (see Figure 7).

To vary tested ammonia readings while replicating the concentrations of those with and without CKD (see Figure 8), 20-20-20 fertilizer was utilized due to its nitrogen-heavy composition of compounds that produce ammonia during hydrolysis [10].

Machine Learning

The dataset linking exhaled ammonia concentrations with CKD was curated (see Figure 8) before being divided into 70% training, 15% testing, and 15% validation data. It was used to train a logistic regression model (BLRM) and multilayer perceptron (MLP) neural network. The BLRM was trained on the original dataset to predict CKD probabilities and label those over its determined threshold as being high-risk users (see Figure 9). The ammonia levels of those considered likely to have CKD were then normalized and assessed by the MLP model, which was trained over three iterations to classify people’s particular CKD stages through a modified version of the dataset (see Figure 10).

Relative Readings and Timing Results

Although advanced sensing equipment to cross-reference calibrated ammonia concentrations was not available for this project’s tests, relative comparisons of detection abilities revealed a strong positive relationship between 20-20-20 fertilizer concentrations and sensor readings.

The overlapping fluctuation patterns of the graphed sensor reactions were analyzed for the optimal initial stabilization time of 7 seconds (see Figure 11). Likewise, the determined plateau limit was 20 seconds, as the graphed sensor readings consistently had no significant changes past this time range once in a stable concentration of ammonia.

From finding the mean of the diffusion trials, a minimum required purging time of 140 seconds was chosen (see Figure 11).

BLRM Results

The coefficients for the linear predictor of the BLRM were determined, establishing the model’s equation as graphed on Figure 12. The receiver operating characteristic curve (see Figure 12) was analyzed to decide the model’s 0.2 threshold. Prioritization was given to the high recall rate of 99.88% to maintain clinical feasibility, which did not significantly compromise the false positive rate of only 0.53%. At this threshold, the model’s overall accuracy was 99.73%; any case with a CKD probability greater than or equal to 0.2 was labeled positive and further transferred to the MLP model.

MLP Results

The results of the final MLP model were ascertained from all previous iterations (see Figure 13). The third iteration was finalized due to its training and validation losses plateauing at 113 epochs (see Figure 14) through the provided patience. Moreover, the model’s overall loss was significantly lower than the other iterations at 0.18, with its final accuracy reaching 99.7% after also plateauing in the process (see Figure 14). Since the model’s losses and accuracies changed without divergence, it did not display signs of underfitting nor overfitting, signaling sufficient dropout regularization.

The finalized model’s precision in predicting all individual stages was always at least 99%. This did not come at the cost of sensitivity, as its lowest recall score was also 99% for all classes. Furthermore, the model’s confusion matrix reflects its ability to consistently differentiate between CKD stages (see Figure 14).

Device Usage

The following is the application process of NephroSense, combining physical, machine learning, and interface components (see Figure 15 for details).

User affirms adherence to controlled condition requirements, like alcohol avoidance for at least 12 hours [23].

User initiates baseline measurements to renew Vout0 (ensuring accuracy regardless of purging efficacy).

User breathes into the mouthpiece using a standardized procedure (one breath over 2 seconds into a controlled volume) to obtain Voutx.

Voltages are converted into an ammonia concentration using calibration.

The ammonia concentration is analyzed by the BLRM; if positive CKD is predicted, the concentration is then normalized and assessed by the MLP for CKD stage.

The GUI displays the user’s results.

Purging and disinfection/alteration of the mouthpiece are completed for subsequent users.

Overall, the device met the defined objectives (see Figure 15).

Central Device Outcomes

In meeting the defined objectives (see Figure 15), NephroSense was able to streamline biomarker screening processes. The physical components of the device were successfully combined to control the prototype’s airflow while maintaining its impermeability (see Figure 16). Meanwhile, the software aspects of the device can convey user results within minutes (see Figure 17). The device collects user breath samples through a non-invasive exhalation technique, allowing for user comfort while also producing no biohazardous waste when proper mouthpiece decontamination measures are ensured (see Figures 18). All of the prototype’s improvements from current methods do not compromise its performance metrics, as its ability to predict CKD presence and stages achieved overall accuracies of 99.73% and 99.7%, respectively. The entire device comes together via a custom, user-friendly interface (see Figure 19 for result page previews).

Significance of Device

The need for a novel alternative to current biomarker methods is constantly growing, particularly when considering the push for convenience, speed, and environmental sustainability in healthcare, among other fields. In achieving its objectives, Nephrosense corroborates the viability of breath-analysis testing, ultimately demonstrating its constructive potential in medical screening and beyond.

Key Limitations

Visualization of limitations can be found on Figure 20.

Lack of pressure control leads to the compression of gases in the sensor compartment.

Absence of clinical testing.

No consideration for the potential impacts of temperature and humidity on the sensor readings.

Limited accessibility features in the user interface.

Limited to illness-specific screening without overall biomarker evaluations.

Improvements

The lack of pressure control can be solved with a bag-based sensor compartment (see Figure 21). Meanwhile, the feasibility of the prototype can be improved through real-world testing, which can be completed by engaging a large number of subjects to further tune model parameters (see Figure 21). Future iterations will include temperature and humidity compensation as well as the integration of interface accessibility features (see Figure 22).

Adaptations

Constructing a GC-MS model could allow for generalized biomarker screening (see Figure 23), and altering the prototype to apply beyond biomarker screening (see Figure 24) can further assert its positive potential.

Our project would not be the same without all of the amazing people that helped us in its creation.

We would like to first sincerely thank Dr. Ryan LaRue, who took time out of his busy schedule to review our project and provide us with guidance in preparation for the Canada-Wide Science Fair. We would also like to thank all of the BASEF organizers and chaperones, as they connected us to Dr. LaRue and gave us this wonderful opportunity.

We also want to appreciate our parents for their support in the completion of our work, both in encouragement and enthusiasm. They helped motivate us, and they provided all of the funding needed for our project.

Words cannot express how grateful we are for the help we have received in the time of this project. A great thank you to everyone mentioned!

References

[1] Mayo Clinic Labs. (2025, April 21). 4 statistics that showcase the broad impact of laboratory medicine. Mayo Clinic Laboratories. https://news.mayocliniclabs.com/2025/04/21/4-statistics-that-showcase-the-broad-impact-of-laboratory-medicine/#:~:text=Below%20are%20a%20few%20highlights,insights%20at%20an%20accelerated%20pace.

[2] Canadian Society for Medical Laboratory Science. (2023). Surgical Volume Trends 2009. In Canadian Institute for Health Information. https://labweek.csmls.org/wp-content/themes/nmlw2015mem/toolkit/stat_cards_EN.pdf

[3] Echelon Health. (2025, December 15). How Early Detection Can Improve Survival Rates for Major Diseases. https://www.echelon.health/how-early-detection-can-improve-survival-rates/#:~:text=Breast%20cancer:%20Over%2090%25%20of,bypass%20operations%20or%20stent%20placements.

[4] Pennic, F. (2024, October 18). 33% of patients delay diagnostic tests, 78% fear blood drawn, survey reveals. HIT Consultant Media. https://hitconsultant.net/2024/10/18/33-of-patients-delay-diagnostic-tests-78-fear-blood-drawn/

[5] Nall, R. (2019, August 1). How long does it take to get blood test results? Healthline. https://www.healthline.com/health/how-long-for-blood-test-results

[6] Duong, D. (2023). Improper disposal of medical waste costs health systems and the environment. Canadian Medical Association Journal, 195(14), 518–519. https://doi.org/10.1503/cmaj.1096046

[7] Bentall, A. (2025, December 9). Chronic kidney disease - Symptoms and causes. Mayo Clinic. https://www.mayoclinic.org/diseases-conditions/chronic-kidney-disease/symptoms-causes/syc-20354521

[8] Mark P, Stafford L, Grams M et al.

Global, regional, and national burden of chronic kidney disease in adults, 1990–2023, and its attributable risk factors: a systematic analysis for the Global Burden of Disease Study 2023

The Lancet, 2025; 406, 2461-2482

[9] Fresenius Kidney Care. (2024, January 1). Ammonia breath & Chronic Kidney Disease, CKD. Fresenius Kidney Care. https://www.freseniuskidneycare.com/thrive-central/ammonia-breath

[10] Grigar, J. (2024, September 11). How Nitrogen Fertilizers React with the Soil. No-Till Farmer. https://www.no-tillfarmer.com/articles/13590-how-nitrogen-fertilizers-react-with-the-soil#:~:text=Urea%20(46%25%20N)%20is,and%20water%20in%20various%20proportions.

[11] DFRobot. (2025, December 9). Gravity: Electrochemical Oxygen Sensor (0-25%Vol, I2C & UART) | DFRobot Wiki. Gravity: Electrochemical Oxygen Sensor (0-25%Vol, I2C & UART) | DFRobot Wiki. https://wiki.dfrobot.com/sen0465#Specification

[12] Hibbard, T., & Killard, A. J. (2011). Breath ammonia levels in a normal human population study as determined by photoacoustic laser spectroscopy. Journal of Breath Research, 5(3), 037101. https://iopscience.iop.org/article/10.1088/1752-7155/5/3/037101

[13] Chan, M., Li, Y., Wu, C., Lee, Y., Zan, H., Meng, H., Hsieh, M., Lai, C., & Tian, Y. (2020). Breath ammonia is a useful biomarker predicting kidney function in chronic kidney disease patients. Biomedicines, 8(11), 468. https://www.mdpi.com/2227-9059/8/11/468

[14] Hayes, A. (2026, March 23). Empirical Rule: Definition, Formula, and Example. Investopedia. Retrieved April 18, 2026, from https://www.investopedia.com/terms/e/empirical-rule.asp

[15] Adrover, R., Cocozzella, D., Ridruejo, E., García, A. I., Romé, J., & Podestá, J. (2020). Breath-Ammonia Testing of Healthy Subjects and Patients with Cirrhosis. Digestive Diseases and Sciences, 57, 10.1007/s10620-011-1858–9. https://doi.org/10.1007/s10620-011-1858-9

[16] Brooks, S. M., Haight, R. R., & Gordon, R. L. (2006). Age does not affect airway pH and ammonia as determined by exhaled breath measurements. Lung, 184(4), 195–200. https://doi.org/10.1007/s00408-005-2580-1

[17] Numiqo. (2023, January 8). Logistic Regression [Simply explained] [Video]. YouTube. https://www.youtube.com/watch?v=C5268D9t9Ak

[18] GeeksforGeeks. (2026, April 6). Logistic regression in machine learning. GeeksforGeeks. https://www.geeksforgeeks.org/machine-learning/understanding-logistic-regression/

[19] GeeksforGeeks. (2025, November 29). Data normalization in data mining. GeeksforGeeks. https://www.geeksforgeeks.org/machine-learning/data-normalization-in-data-mining/

[20] GeeksforGeeks. (2025, July 12). Leaky ReLU activation function in deep learning. GeeksforGeeks. https://www.geeksforgeeks.org/machine-learning/leaky-relu-activation-function-in-deep-learning/

[21] GeeksforGeeks. (2025, November 17). Softmax activation function in neural networks. GeeksforGeeks. https://www.geeksforgeeks.org/deep-learning/the-role-of-softmax-in-neural-networks-detailed-explanation-and-applications/

[22] GeeksforGeeks. (2025, July 26). What is Sparse Categorical Crossentropy. GeeksforGeeks. https://www.geeksforgeeks.org/machine-learning/what-is-sparse-categorical-crossentropy/

[23] Sullivan Nicolaides Pathology. (2021, June 4). Urea breath test (UBT). Sullivan Nicolaides Pathology. https://www.snp.com.au/patients/patient-resources/pre-test-information/urea-breath-test-ubt/

Images

AlexRaths. (2011, November 26). hand in protective glove holding a blood sample. iStock. https://www.istockphoto.com/photo/blue-gloved-hand-holding-a-blood-sample-in-a-vile-gm135018352-18443766?searchscope=image%2Cfilm

Astiyah, R. (n.d.). Ocean Plastic Pollution Illustration. Icon Scout. https://iconscout.com/illustration/ocean-plastic-pollution-illustration_7220417

Bet_Noire. (2016, April 26). Blood test tubes. Blood samples isolated on white. 3D illustration. iStock. https://www.istockphoto.com/photo/test-tubes-blood-samples-isolated-on-white-gm519960294-90798031

Bkmam. (n.d.). Lab Use Gas Sampling Bag with Single Double Valves 0.1L 1L 2L 5L 15L [Video]. Made-in-China.com. https://bkmbio.en.made-in-china.com/product/RQjYzklUYgcN/China-Lab-Use-Gas-Sampling-Bag-with-Single-Double-Valves-0-1L-1L-2L-5L-15L.html

Bradford, A. (2018, February 2). Lungs: Facts, Function and Diseases. Live Science. https://www.livescience.com/52250-lung.html

Creative Fabrica. (2020, December 29). Temperature Icon. Creative Fabrica. https://www.creativefabrica.com/product/temperature-icon-4/

GaryAlvis. (2012, March 12). A red bio-hazard bag containing medical waste with a bio-hazard. iStock. https://www.istockphoto.com/photo/bio-hazard-bag-small-gm185318589-19563743?searchscope=image%2Cfilm

GaryAlvis. (2012, March 20). This image shows various forms of bio-waste including needle disposal. iStock. https://www.istockphoto.com/photo/bio-waste-with-clipping-path-gm155369782-19659208?searchscope=image%2Cfilm

Hamz. (n.d.). Water droplet icon 7126517 Vector Art at Vecteezy. Vecteezy. https://www.vecteezy.com/vector-art/7126517-water-droplet-icon

iStock. (2016, December 14). Disposal container; reducing medical waste disposal. Small Medical. iStock. https://www.istockphoto.com/photo/disposal-container-reducing-medical-waste-disposal-gm623595016-109417955?searchscope=image%2Cfilm

JHVEPhoto. (n.d.). Markham, On, Canada - October 23, 2020: Emergency Department sign is seen at Markham Stouffville Hospital in Markham, Ontario, Canada. Markham Stouffville Hospital is a community hospital. Adobe Stock. https://stock.adobe.com/ca/images/markham-on-canada-october-23-2020-emergency-department-sign-is-seen-at-markham-stouffville-hospital-in-markham-ontario-canada-markham-stouffville-hospital-is-a-community-hospital/388003741

Nikolayev. (2019, August 16). Scalpel with a removable blade. Manual surgical instrument. Medicine. iStock. https://www.istockphoto.com/vector/scalpel-with-a-removable-blade-manual-surgical-instrument-medicine-and-health-gm1167889129-322251126?searchscope=image%2Cfilm

Pidjoe. (2023, April 22). Pile of Microcentrifuge Tubes, Syringes, Gloves-Medical Plastic. iStock. https://www.istockphoto.com/photo/medical-hazardous-hospital-waste-close-up-gm1483773591-510285508?searchscope=image%2Cfilm

Pixelimage. (2020, September 21). Blood Sample on Blue Background. iStock. https://www.istockphoto.com/photo/blood-sample-on-blue-background-gm1273400357-375290370?searchscope=image%2Cfilm

Rahman Anis, H. (2025, August 4). Detailed 3D CT scan of the abdomen in arterial phase showing real kidneys, complete abdominal arteries, and inverted bony structures. Ideal for clinical reference and anatomical study. Shutterstock. https://www.shutterstock.com/image-photo/detailed-3d-ct-scan-abdomen-arterial-2661787345?trackingId=fd0e45d7-8dcc-47db-bdc7-14a4b794ff14&listId=searchResults

Requejo, O. (2025, December 10). A close-up of a person in a lab coat holding a urine specimen cup. iStock. https://www.istockphoto.com/photo/gloved-hand-with-sample-gm2249764868-663517502?searchscope=image%2Cfilm

RW Medical Canada. (n.d.). Isopropyl alcohol 70% 500mL. RW Medical Canada. https://rwmedical.ca/products/isopropyl-alcohol-70-500ml?srsltid=AfmBOoptsf4vzwH_qm2hp1uG9H-keFQa0ejKHV8n3aF6bC7Z8W3chd-p

SRAinstruments. (n.d.). Agilent 7000E Triple Quadrupole GC/MS - High Performance - SRA Instruments. SRAIntruments. https://www.srainstruments.com/en/product/gc-qqq-7000d-agilent-technologies.aspx

Suradech. (2019, February 27). White blank plastic spray bottle isolated on white background. iStock. https://www.istockphoto.com/photo/white-blank-plastic-spray-bottle-isolated-on-white-packaging-mockup-gm1132296573-300112516?searchscope=image%2Cfilm

Surfr. (2008, February 12). Syringe. iStock. https://www.istockphoto.com/photo/syringe-gm145118075-5307835?searchscope=image%2Cfilm