Researchers are testing whether a simple breath test can help detect different types of cancer early, at a time when early detection still determines longevity.
A South Korean study shows that exhaled breath can be classified into distinct health signals, suggesting potential for future screening tools while also highlighting how much evidence is still needed.
Creating a cancer breath test
In the study, scientists analyzed breath samples from 206 people and used the results to distinguish between healthy participants and those with lung or stomach cancer.
Breath screening tends to fail when it cannot accommodate the noise of a real clinic, so engineers are focusing on more rigorous measurements and analysis.
The Electronics and Telecommunications Research Institute (ETRI) has combined sensors with software designed to make decisions quickly and reproducibly.
The project is led by ETRI’s Dr. Dae-Sik Lee, whose lab is building medical sensing tools for clinics.
This long-term effort currently aims to combine breath sampling and computer classification to screen for multiple cancers in one workflow.
Chemical signals in exhaled breath
Exhaled breath contains traces of chemicals from blood and respiratory fluids, some traces of which change depending on the disease.
Cancer can change cell metabolism, which can change volatile organic compounds (gases that easily evaporate from tissues) before symptoms appear.
Many factors such as smoking, infections, and diet can also cause these levels to fluctuate. Therefore, useful patterns typically include multiple compounds together.
Because no single breath chemical uniquely marks cancer, researchers often treat breath profiles as a matter of patterns.
Sensors catch various gases
Rather than tracking one compound at a time, this platform uses a multimodal sensor array consisting of many sensors with different chemical strengths.
The signal from the semiconductor metal oxide sensor, a solid-state chip whose resistance changes due to surface reactions, is highly sensitive to many exhaled gases.
Adding an electrochemical sensor, a cell that generates an electrical current during a target reaction, provides a clearer reading of the mixture.
The third channel comes from a photoionization detector, a sensor that ionizes the vapor with ultraviolet light, helping to catch chemicals that others miss.
Visualization of respiratory data
The respiratory sensor outputs wavy lines over time, so the team needed a way to convert those lines into a stable input.
The platform records time-resolved signals from all sensors as the sample passes, or readings that track changes from second to second.
The software reshapes these curves into a response map, an image-like grid stacked with sensors and time, allowing the model to learn the patterns.
This format is advantageous for deep learning tools built for images, but also requires careful tuning to keep day-to-day variations small.
A two-step decision for cancer
To distinguish between healthy breath and cancerous breath, the team trained a computer model that worked in two linked stages.
They used a model called a Hierarchical Deep Convolutional Neural Network (HD-CNN), which learns patterns through layered image filters, to make the call.
The first step of HD-CNN labels the sample as healthy or cancerous, and the second step classifies the cancer as lung or stomach.
By isolating the easy splits early, the model can spend more effort on the more difficult cancer-versus-cancer decisions.
How to check performance
To see how well the system worked, the team repeatedly tested it on different parts of the same respiratory dataset to avoid lucky results.
Across these tests, the model correctly identified most healthy people and correctly distinguished between lung and stomach cancer in the majority of cases.
We also demonstrated a strong balance between capturing true cancer cases and limiting false alarms, a key requirement for any screening tool.
When compared to simpler computer models, this approach separated groups more clearly and produced more consistent decision-making.
Challenges of breath analysis
Breath characteristics vary from person to person, so even a powerful model can stumble if the sample set remains narrow.
Food, alcohol, smoking, and drugs can change the chemicals exhaled within hours, potentially blurring the line between health and cancer.
Training requires careful labeling and matching, as gastric tumors and lung infections can generate overlapping chemical stress signals.
This reality makes external validation and repeated sampling essential before a platform can guide scans and endoscopy referrals.
What does that change?
Transitioning the platform into daily practice means adapting it to clinical hours, staff training, and clear decisions about next steps.
In 2019, an early ETRI prototype recorded 200 breath analyzes during clinical sampling and reached approximately 75 percent accuracy in the clinic.
“If this technology is put into practical use, it should strengthen the market competitiveness of lung cancer diagnostic equipment and help reduce government health insurance expenditures,” Lee said.
This real-world pressure explains why the platform is intended to act as a first filter rather than a final diagnosis.
The future of cancer breath testing
Scaling the platform from two cancers to many cancers will require more respiratory data while keeping the platform low cost and simple.
Hierarchical design already supports that idea, as new detailed classifiers can be added without retraining the entire frontend.
Professor Sanghoon Jeong, a thoracic surgeon at a university hospital in Bundang, South Korea, said, “Through our collaboration with ETRI, we have confirmed the possibility of testing for lung cancer in a low-cost and convenient manner.”
Without this scaling, even strong results for two cancers can mislead clinicians if the patient has other cancers or inflammatory diseases.
Together, the sensor array, response map format, and HD-CNN hierarchy demonstrate how exhaled breath chemistry can be read at the point of care.
Future work should test the platform in more hospitals, control sampling conditions, and prove that early flags lead to better outcomes.
This research npj digital medicine.
—–
Like what you read? Subscribe to our newsletter for fascinating articles, exclusive content and the latest updates.
Check us out on EarthSnap, the free app from Eric Ralls and Earth.com.
—–
