AI is learning to smell disease in human breath

A doctor does not need a machine to know that illness changes smell. Sweet breath in uncontrolled diabetes, ammonia-like breath in kidney failure, the stale odor of advanced liver disease, the sour edge of infection — medicine has lived with these clues for centuries. What has changed is the scale and resolution. Modern sensors can capture volatile organic compounds at concentrations too small for any clinician to notice. AI can then sort those signals into patterns that line up with disease, inflammation, microbial activity, oxidative stress, and drug response. That is why breath analysis has moved from bedside intuition to a serious branch of diagnostics research.

The bold claim that artificial intelligence can smell illness in human breath is only half right. Machines do not smell the way humans smell. They do not experience odor. They read chemistry. In one setup, a gas sensor array behaves like a rough synthetic nose and produces an electrical pattern when exposed to breath. In another, gas chromatography and mass spectrometry break breath into molecular components first, and machine learning looks for combinations that matter. The shift is subtle but important. AI is not replacing biology with magic. It is turning faint chemical noise into classification.

That distinction matters because the field is caught between real promise and repeated overstatement. Study after study shows that breath profiles can separate sick people from healthy controls under controlled conditions. Lung cancer, viral infection, ventilator-associated pneumonia, chronic airway disease, tuberculosis, kidney disease, and diabetes all have published signals in breath. Yet routine clinical use remains rare. The gap is not proof that the idea is flawed. It shows that finding a signal is easier than proving a test is stable, generalizable, clinically useful, and worth paying for.

A clinical hunch that never went away

Illness changes metabolism, and metabolism leaks. Some of those leaks leave the body as volatile organic compounds, or VOCs, in exhaled breath. They may come from the host, from microbes living in or on the body, from inflammation in the airways, from oxidative stress, or from shifts in liver and kidney handling of metabolites. Breath therefore carries a moving record of physiology. It is not a perfect record, and it is certainly not a clean one, but it is biologically rich. That richness is the reason breath analysis has kept returning, decade after decade, even after earlier waves of enthusiasm faded.

Breath is also appealing for a practical reason. It is painless, repeatable, fast, and potentially cheap. Blood draws are invasive. Imaging is expensive. Tissue biopsy is slower, riskier, and often unsuitable for frequent monitoring. Breath sampling looks almost trivial by comparison. A patient exhales into a bag, a sorbent tube, a mouthpiece, or a mask-based collector. That simplicity explains why breath analysis attracts clinicians, engineers, startups, and regulators alike. A technology that can screen, triage, or monitor disease from a few breaths is hard to ignore.

Still, human breath is not just a mirror of internal disease. It is also contaminated by diet, smoking, alcohol, medication, ambient air, oral bacteria, cleaning products, perfume, sampling materials, and the patient’s last meal or last walk outside. A breath test that ignores those variables will look smarter than it really is. That is one reason the strongest researchers in this area spend so much time on sample collection, background air correction, fasting rules, device calibration, and external validation. The hardest part is rarely the algorithm alone. It is the entire chain of measurement.

The European Respiratory Society technical standard on exhaled biomarkers made this point years ago. Breath tests are not one thing. They include exhaled nitric oxide, VOCs, exhaled breath condensate, and particulate measurements, each with its own sampling and analysis problems. That standardization push remains one of the field’s most important milestones because it forced breath researchers to treat collection and preprocessing as clinical issues, not mere lab housekeeping.

Breath is chemistry in motion

A useful way to think about breath analysis is to separate molecules, patterns, and models. Molecules are the underlying chemistry: alkanes, aldehydes, ketones, sulfur compounds, ammonia, and many others. Patterns are the collective signature those compounds create in a device readout. Models are the statistical or machine-learning systems that decide whether a given pattern resembles cancer, infection, asthma, COPD, kidney dysfunction, or something else. Each layer can fail on its own. A weak molecular signal produces a weak pattern. A good pattern paired with a biased model still collapses in real use.

Some platforms aim for explicit identification of compounds. Gas chromatography–mass spectrometry, or GC-MS, is the classic example. It is analytically powerful and helps researchers identify candidate biomarkers, but it is slower, more complex, and less suited to true point-of-care use. Electronic noses take the opposite route. They usually do not identify every compound. They detect response patterns across a sensor array and let a model classify the resulting “breathprint.” That tradeoff is why eNoses are so attractive for screening and why they also draw skepticism. Speed and portability rise when molecular specificity falls.

Three ways machines read breath

GC-MS and related analytical platforms

Electronic nose sensor arrays

Breath condensate and wearable mask systems

This is why breath diagnostics rarely move in a straight line from chemistry paper to hospital product. The more precisely a platform measures molecules, the more burdensome the workflow tends to become. The more convenient the device becomes, the more the field leans on calibration, AI training, and rigorous controls to avoid false certainty.

AI enters at every level. It can classify raw sensor responses, reduce dimensionality in high-dimensional VOC data, detect non-linear relationships, flag outliers, and estimate the contribution of particular compounds or sensor channels. A newer line of work also tries to make these models more interpretable. A 2025 baseline study using GC-MS breath VOC data applied SHAP-based explainability to respiratory disease classification, which reflects a wider push to make models clinically legible rather than black-box curiosities. That shift is overdue. Clinicians tolerate high-performing models more easily when they can see why the model leans one way rather than another.

AI changes the unit of diagnosis

Classical diagnostics often ask a narrow question. Is this gene mutated? Is this pathogen present? Is this hormone high? Breathomics asks a different question: does the whole exhaled chemical profile look like a disease state? That is exactly the sort of task where machine learning shines. It can find structure in multivariate data without demanding that one molecule carry the entire burden of diagnosis. A breath test does not need a single miracle biomarker if a pattern of smaller changes is stable enough.

That shift also explains why AI breath diagnostics often look impressive in early work. A model trained on carefully selected cases and controls can achieve striking sensitivity or specificity because it is learning many weak signals at once. Lung cancer studies are full of that pattern. So are respiratory infection studies. The danger appears when a model leaves the original environment. Change the population, the device, the room air, the fasting protocol, the smoking status, or the circulating viral variant, and accuracy may fall quickly.

The COVID-19 literature made this painfully visible. In a 2023 JAMA Network Open diagnostic study, breath VOC biomarkers distinguished COVID-19 from non-COVID illness with high accuracy during one phase of the pandemic, but the rise of the Omicron variant altered the VOC profile enough that the earlier biomarker set lost performance and a new one had to be learned. That is not a failure of breath analysis. It is a lesson in dataset shift. A living pathogen changed the odor landscape faster than the model was built to adapt.

This is why breath diagnostics may end up looking less like one-off lab tests and more like adaptive systems. Models will need retraining, recalibration, site-specific monitoring, and strong quality control. The comparison is closer to modern radiology AI or clinical prediction models than to a static blood assay. The best breath tests of the future will not be clever gadgets alone. They will be measurement systems with ongoing learning and governance.

Cancer remains the field’s toughest and richest test case

Cancer is where breath analysis draws the most attention, and lung cancer sits at the center of that attention for obvious reasons. Tumors alter metabolism. Tumor-associated inflammation and oxidative stress alter metabolism too. Lung tumors also sit close to the organ that generates exhaled breath. That combination makes lung cancer an unusually plausible target for VOC profiling. The clinical need is also severe. Lung cancer kills because it is often found late. A fast, non-invasive breath test that flags suspicious cases before symptoms become unmistakable would have real value.

The evidence base is now too large to dismiss. A 2022 JAMA Network Open systematic review and meta-analysis reported that eNoses show high diagnostic accuracy for cancer detection using exhaled breath, while also stressing that most studies were small feasibility studies with weak standardization and a substantial risk of bias. That conclusion captures the field neatly: the signal is real enough to persist across studies, but the translational foundation is still uneven.

Individual lung-cancer studies tell the same story in more granular form. A 2017 study reported that electronic-nose analysis with support vector machine methods could discriminate patients with lung cancer from healthy and mixed-control groups. A 2024 study used a self-made electronic nose to identify lung cancer and even stage-related patterns in breath. A 2025 multicentre prospective external validation study then pushed the field a step closer to clinical seriousness by reporting accurate lung-cancer detection with external validation rather than single-center optimism. That progression matters more than any one headline accuracy figure. It shows the field inching from proof of concept toward reproducible deployment.

Yet even here, caution is warranted. Lung cancer rarely arrives alone. Smokers may have COPD. People undergoing evaluation may have pneumonia, fibrosis, or benign nodules. Surgery, anesthesia, medication, age, and comorbidity all reshape the breath profile. A clinically useful test does not merely separate clean cancer cases from healthy volunteers. It must survive the mess of real respiratory medicine. That is why external validation, mixed-control cohorts, and comparisons against confusable diseases matter far more than a polished conference demo.

The next frontier in cancer breath testing looks broader than simple yes-or-no detection. Researchers are now training hierarchical and multimodal models that attempt to sort multiple cancers, not just one, from breath signals. A 2026 NPJ Digital Medicine paper described a deep-learning breath platform for dual-cancer classification using a multimodal gas sensor array. That work is still early, but it points to where the science is heading: not “does this patient have cancer” in the abstract, but which disease state best matches the chemical pattern in front of us.

Infection leaves a faster but less stable signature

Infectious disease is another natural fit for breath analysis because infection changes both host metabolism and microbial ecology. Viral replication, airway inflammation, oxidative stress, fever, and immune activation all change the chemistry of exhaled compounds. That creates a faster-moving signal than many chronic diseases, which is useful for triage but also harder to stabilize across pathogens and variants.

A 2021 American Journal of Respiratory and Critical Care Medicine study showed that virus-induced VOCs are detectable in exhaled breath during pulmonary infection and linked a decane-related signature to rhinovirus infection in airway epithelial cells, experimentally infected healthy volunteers, and COPD patients with viral exacerbations. That kind of translational chain — from cell model to human challenge to naturally occurring disease — is unusually persuasive. It suggests the breath signal is not pure statistical artifact. It has biology underneath it.

COVID-19 accelerated the field dramatically. A 2021 ERJ Open Research study evaluated breath VOC analysis for identifying suspected or confirmed COVID-19. Later work in JAMA Network Open showed that the biomarker panel had to change once Omicron became dominant. A 2024 meta-analysis of 29 papers found pooled sensitivity around 0.92 and specificity around 0.90 for VOC-based breath analysis in COVID-19 detection, while also showing that performance varied by variant and detection method. That combination of promise and instability says a lot. Breath testing may be strong enough for rapid screening, but infectious disease models must be built to live with viral evolution.

Hospital-acquired infection may prove even more practical than pandemic screening. Ventilator-associated pneumonia is a notoriously difficult diagnosis, and treatment delays matter. A 2020 Respiratory Research study used electronic-nose sensor array signals with machine learning to detect ventilator-associated pneumonia and reported good diagnostic accuracy, while also stressing that protocols for data processing and modeling had to be tightened to improve generalizability. That is breath diagnostics in miniature: useful signal, then a warning label about reproducibility.

Tuberculosis is another compelling target because sputum collection is imperfect, lab infrastructure varies, and rapid non-invasive screening is badly needed. Breath-based pulmonary tuberculosis studies and reviews suggest real discriminatory VOC signatures, though field conditions, coinfections, nutrition, and local epidemiology complicate deployment. In other words, the biology is there, but the implementation problem is bigger than the biomarker problem.

Chronic lung disease shows why classification is hard

Cancer and acute infection generate headlines, but chronic airway disease may be where breath analysis becomes most clinically useful first. Asthma, COPD, bronchiectasis, cystic fibrosis, and mixed obstructive patterns all produce ongoing inflammatory and metabolic changes that could be tracked repeatedly. Breath testing is well suited to that kind of longitudinal monitoring because it can be repeated without burdening the patient.

The difficulty is that chronic respiratory disease is messy. Asthma overlaps with COPD. Smoking leaves its own chemical signature. Inhaled steroids, bronchodilators, exacerbations, airway colonization, and recent infections all shift the profile. A model that separates textbook asthma from textbook COPD may perform badly in the clinic where overlap is the rule, not the exception. That is why newer studies matter when they include PRISm, mixed chronic respiratory disease cohorts, or externally validated disease subtypes rather than idealized cases.

A 2025 cross-sectional study reported disease-specific VOC signatures that differentiated COPD, asthma, PRISm, and healthy controls. Another 2025 baseline study used interpretable machine learning on GC-MS-derived breath VOCs to distinguish asthma, bronchiectasis, and COPD. Neither paper solves the problem of routine classification. Both do show that the field is moving beyond simple disease-versus-healthy comparisons and toward the more clinically relevant question of which airway disease pattern is present, how severe it is, and whether it is shifting over time.

That longitudinal angle may become the real prize. Breath analysis might prove most valuable not as a once-only diagnostic oracle but as a repeatable monitoring layer. Exacerbation risk, treatment response, inflammatory phenotype, and early relapse are all more plausible medium-term targets than flawless first-pass diagnosis across every patient group. Some asthma and cystic-fibrosis studies already point in that direction.

Metabolic and systemic illness widen the map

Respiratory disease gets most of the attention because breath is produced in the lungs, but the chemistry of exhaled air reaches well beyond the chest. Breath carries systemic metabolism. That is why diabetes, kidney disease, and liver disease remain important parts of this story. The underlying idea is old. The modern twist is that AI lets researchers interpret multiple metabolites together instead of chasing a single famous compound such as acetone or ammonia.

Diabetes is the classic example. Breath acetone rises with ketosis and has long been studied as a non-invasive signal related to glucose metabolism. The new work is less about proving acetone exists and more about building usable sensor systems. A 2024 study described a TinyML-powered eNose for noninvasive diabetes detection from human breath, reflecting a broader attempt to shrink models and sensors into portable, real-time systems rather than bench instruments. That direction makes sense. Diabetes monitoring only matters clinically if the device is simple enough to use repeatedly.

Kidney disease offers another strong biochemical rationale because impaired nitrogen handling influences breath ammonia. A 2020 study found that breath ammonia predicted kidney function in chronic kidney disease patients, and newer reviews continue to frame ammonia sensing as one of the more plausible non-invasive pathways in renal diagnostics. That does not mean a breath ammonia reading will replace serum creatinine or estimated GFR any time soon. It does mean systemic disease leaves an airway-accessible trace that is measurable and clinically interpretable.

Liver disease may be even more intriguing because the liver sits at the center of metabolism and detoxification. A 2021 PLOS One pilot study used machine learning on breath volatolomic profiles to identify non-invasive biomarkers of liver disease. Other work has focused on cirrhosis, hepatocellular carcinoma, and disease-related compounds such as limonene and dimethyl sulfide. These are still early pathways, but they reinforce the central proposition of breathomics: illness does not need to sit in the lungs to alter what leaves the lungs.

This wider map matters strategically. If breath diagnostics were limited to respiratory conditions, the field would still be useful. If robust systemic signatures become reproducible, breath becomes something bigger: a front-door screening medium for multiple classes of disease. That is the ambitious vision driving many breath-AI companies and laboratories. The science does not justify universal screening yet, but the logic behind the ambition is not fanciful.

The ugly problem of confounding variables

Every exciting breath study sits beside an awkward question: what else could have produced that signal? Confounding is not a side issue here. It is the central engineering and clinical problem. Smoking changes VOC profiles. Diet changes them. Oral hygiene changes them. Environmental exposure changes them. Collection bags and tubing can add compounds of their own. A model that quietly learns the hospital ward, the local cleaning products, or the smoking status of the control group is not learning disease. It is learning context.

That is why standardization papers deserve more respect than flashy classifier papers. Sampling duration, expiratory fraction, fasting state, room-air correction, humidity control, storage conditions, thermal desorption methods, sensor drift, cross-validation strategy, and external validation all determine whether a result survives first contact with another site. A 2017 paper on optimizing sampling parameters for standardized breath sampling made the point plainly: poor repeatability often begins before the model ever sees the data.

AI adds a second layer of risk because high-dimensional models can overfit beautifully. Small datasets, synthetic balancing, unblinded preprocessing decisions, and weak separation of training and test workflows can all inflate performance. Breath research has not been uniquely sloppy in this respect. It has simply been vulnerable to the same traps seen across medical AI. What makes breath harder is that the raw input is already variable before modeling begins.

The field knows this now. Reviews published in 2024, 2025, and 2026 are less interested in raw novelty and more interested in reproducibility, transparent modeling, multicenter validation, and clinical transition. That tonal shift is healthy. It suggests the discipline is growing up. The era of “our device reached 95% accuracy” is giving way to the harder question of whether the same device reaches the same result elsewhere.

Accuracy is not the same as usefulness

A breath model may classify disease accurately in a study and still fail as a medical product. Clinical usefulness depends on where the test sits in the pathway. Is it screening asymptomatic people? Triage in the emergency department? Monitoring during therapy? Replacing an existing test? Acting as an adjunct that reduces unnecessary imaging or biopsy? The same sensitivity and specificity will matter very differently in each setting.

The best reality check comes from regulation. The FDA’s classified breath VOC analysis device in this area is not a sweeping AI nose for cancer or infection. It is the Heartsbreath test, indicated only as an adjunct in diagnosing grade 3 heart transplant rejection in patients within the first post-transplant year, and not as a substitute for endomyocardial biopsy. That narrow indication says more than a dozen enthusiastic press releases. Regulators reward defined clinical claims, controlled workflows, and clear limitations.

That pattern is likely to repeat. Breath tests will reach practice first where the clinical question is narrow, the workflow is controlled, and the consequence of a wrong answer is buffered by confirmatory testing. ICU infection triage, post-transplant surveillance, lung-cancer workup support, and chronic disease monitoring all fit that logic better than grand claims about universal disease detection from a single exhale. The first durable wins will probably be modest and very practical.

Reimbursement and operations will matter too. A hospital will not buy a breath platform because a model is elegant. It will buy one if the device fits existing workflows, reduces downstream costs, earns clinician trust, avoids contamination headaches, and produces stable results across shifts and sites. Breath diagnostics is now old enough that these unglamorous questions are impossible to ignore. That is a good sign, not a depressing one. It means the field is finally talking like medicine instead of pure invention.

Wearables push breath analysis beyond the lab

One of the most interesting changes in the field is the move from one-time sampling toward continuous or semi-continuous monitoring. A 2024 Science paper described EBCare, a smart mask designed to harvest exhaled breath condensate and analyze biomarkers in real time using cooling, microfluidics, biosensors, and wireless electronics. NIH highlighted the same work as a step toward monitoring chemicals in exhaled breath outside traditional lab settings.

This matters because a single breath sample captures a moment. A wearable system captures a trajectory. Disease often reveals itself as change over time rather than one abnormal value. Airway inflammation rises and falls. Infection evolves. Recovery after COVID shifts gradually. A wearable breath platform could, in principle, detect trends that a static clinic test misses. That is a stronger use case for AI, since models are often better at reading sequences and deviations from personal baseline than at making absolute judgments from one isolated measurement.

There are tradeoffs. Wearables inherit all the confounders of ordinary breath analysis and add new ones: motion, environment, long-term sensor stability, user adherence, and device maintenance. They may end up being excellent research tools long before they become everyday clinical products. Still, the direction is revealing. The field no longer imagines breath analysis only as a fancy lab assay. It increasingly imagines breath as a live physiological stream.

The real future of smelling disease

The strongest version of this technology will not behave like science fiction. You will not breathe into a sleek gadget and receive a flawless list of hidden diseases. What is more plausible is both narrower and more powerful. A patient at risk of lung cancer may use breath testing as an early sorting tool before imaging. A ventilated ICU patient may get a breath-based infection alert hours before a conventional workup settles the question. A person with asthma or COPD may be monitored for inflammatory drift or impending exacerbation. A transplant recipient may use a breath test as one more layer of surveillance. Those are not cinematic outcomes. They are exactly the kind of outcomes medicine adopts.

AI is already able to detect disease-linked signatures in human breath. That statement is well supported. The harder statement — that AI can reliably diagnose illness from breath across hospitals, populations, seasons, variants, smoking histories, diets, and device generations — remains unfinished. The field is moving toward that goal, but it has not arrived. The honest view is not cynical and not breathless. It is simpler. The chemistry is real. The signal is often real. The clinic is harder.

That is why this area still deserves serious attention. Breath analysis sits at the meeting point of metabolomics, sensor engineering, clinical medicine, and AI. Very few diagnostic fields have that combination of biological plausibility and practical accessibility. Exhaled air is available every minute of the day. If researchers keep solving standardization, external validation, and workflow design, medicine may eventually treat breath the way it now treats blood pressure or pulse oximetry — not as a curiosity, but as a routine window into disease.

FAQ

Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

Research progress of electronic nose technology in exhaled breath disease analysis
A broad review of breath diagnostics, eNose design, and disease applications.

A European Respiratory Society technical standard on exhaled biomarkers in lung disease
A foundational standard on breath biomarker collection and analysis.

Fulfilling the promise of breathomics
A translational review focused on why breathomics struggles to move into clinical care.

Considerations for the discovery and validation of exhaled breath VOC biomarkers
A concise review on validation, biomarker discovery, and practical pitfalls.

Optimisation of sampling parameters for standardised exhaled breath sampling
A study centered on sampling repeatability and standardization.

Diagnostic performance of electronic noses in cancer diagnoses using exhaled breath
A systematic review and meta-analysis of eNose performance in cancer detection.

Detection of lung cancer in exhaled breath with an electronic nose using support vector machine analysis
An early clinical lung-cancer study using eNose classification.

Detection of lung cancer and stages via breath analysis using a self-made electronic nose device
A breathomics study examining lung-cancer detection and staging signals.

Lung cancer detection by electronic nose analysis of exhaled breath
A multicentre prospective external validation study for lung-cancer detection.

Virus-induced volatile organic compounds are detectable in exhaled breath during pulmonary infection
A translational study linking viral infection biology to breath VOC changes.

Diagnosis of COVID-19 by exhaled breath analysis using gas chromatography-ion mobility spectrometry
A clinical study on breath-based COVID-19 detection.

Portable breath-based volatile organic compound monitoring for the detection of COVID-19 during the circulation of the SARS-CoV-2 Delta variant and the transition to the SARS-CoV-2 Omicron variant
A diagnostic study showing how viral variants changed breath-model performance.

A comprehensive meta-analysis and systematic review of breath analysis in detection of COVID-19 through volatile organic compounds
A pooled analysis of VOC-based breath testing for COVID-19.

Diagnosis of ventilator-associated pneumonia using electronic nose sensor array signals
A study on ICU breath analysis and machine learning for pneumonia detection.

Integrated exhaled VOC and clinical biomarker profiling for asthma and COPD diagnostics and bronchodilator response prediction
A recent study combining breath VOCs and clinical markers in chronic airway disease.

A baseline study of interpretable machine learning using GC-MS-derived breath VOCs for respiratory disease classification
A paper on explainable machine learning for asthma, COPD, and bronchiectasis classification.

Breath ammonia is a useful biomarker predicting kidney function in chronic kidney disease patients
A clinical study supporting breath ammonia as a renal biomarker.

Machine learning analysis of volatolomic profiles in breath can identify non-invasive biomarkers of liver disease
A pilot study applying supervised learning to liver disease breath profiles.

A breathomics based pulmonary tuberculosis detection method
A study on breath-based tuberculosis detection.

Noninvasive diabetes detection through human breath using TinyML-powered e-nose
A portable breath-analysis system for diabetes screening research.

A smart mask for exhaled breath condensate harvesting and analysis
A landmark wearable-breath paper introducing the EBCare platform.

Smart mask for monitoring chemicals in exhaled breath
An NIH summary of the EBCare platform and its monitoring potential.

Product classification for volatile organic compounds breath analysis
FDA classification details for the Heartsbreath VOC breath analysis device.

Humanitarian Device Exemption approval for Heartsbreath
FDA approval information defining the device’s narrow adjunctive use.

The electronic nose emerging biomarkers in lung cancer diagnostics
A clinician-facing review on lung-cancer eNose research.

Exhaled volatile organic compounds and respiratory disease
A review examining the biological basis and current evidence for respiratory VOC biomarkers.