Every breath we exhale carries a complex chemical narrative written in thousands of volatile organic compounds (VOCs). These tiny molecules, produced by human metabolic processes and invading pathogens, hold immense diagnostic potential. Yet, translating this molecular cloud into a definitive medical diagnosis has long faced an analytical bottleneck. Now, an international team of physicists from Botswana and South Africa has harnessed machine learning to decode these complex signatures, offering a rapid, non-invasive pathway to detect tuberculosis (TB) directly from human breath.

Breaking the analytical bottleneck

Traditional gas chromatography-mass spectrometry (GC-MS) excels at separating complex molecular mixtures based on volatility and mass-to-charge ratios. However, a single breath sample generates thousands of overlapping spectral peaks. Manually sifting through this dense landscape is a painstaking, time-consuming process.

To find a faster, more systematic approach to molecule identification, researchers in the Nano –Breath taking group led by George Chimowa from Botswana International University of Science and Technology (BIUST) collected breath samples from three distinct cohorts: patients with active TB, individuals with multidrug-resistant TB (MDR-TB), and healthy control volunteers.

Standard analytical pipelines often rely on aggressive data pre-processing to smooth and simplify the GC-MS spectra. However, this filtering risks erasing subtle, low-concentration molecular features that might hold critical diagnostic clues. Bypassing this limitation, the BIUST group fed a high-dimensional dataset of 1867 spectral features (from 87 breath samples) directly into four competing supervised machine-learning algorithms: decision trees, random forest, k-nearest neighbours, and support vector machines (SVM). The researchers published their findings in Discover Artificial Intelligence*. *

Geometric optimization avoids the pitfalls of high-dimensional data

The team found that the linear SVM classifier led the algorithmic comparison, distinguishing among the three patient cohorts with 93% accuracy.

This performance stems from how the algorithm handles high-dimensional data spaces. While distance metric models often suffer from overfitting errors, where experimental noise is misidentified as a meaningful diagnostic signal, SVM relies on geometric boundary optimization to separate classes.

Much like clearing a wide highway down the centre of two opposing crowds, it identifies a hyperplane that maximizes the margin between groups. Instead of tracking every individual in the crowd, the algorithm relies solely on a sparse subset of border points known as support vectors. By focusing strictly on these boundary “guards” and ignoring the background movement, this approach completely isolates the classifier from the noise inherent to the surrounding high-dimensional feature space.

Beyond raw classification accuracy, the model yielded a crucial insight for future clinical applications. It pinpointed a specific window within the gas chromatogram, a retention time between 10 and 30 min, where the most diagnostic, high-variance VOCs elute. This 20-min window corresponds to medium- and large-molecule VOCs, including specific chemical biomarkers associated with TB, such as tridecane, decane and O-cymene.

By demonstrating that these vital diagnostic features are concentrated within this single slice of data, the physicists have shown that future diagnostic hardware can be specialized, eliminating the need to analyse the entire spectrum. Narrowing the focus to this specific elution window can drastically shrink data-processing times and maximize instrument throughput in clinical settings.

AI and the future of medical diagnostics

TB remains one of the world’s deadliest infectious diseases. Traditional diagnostics rely heavily on slow sputum cultures and backlogged laboratory queues, delaying critical treatment. A rapid, non-invasive breath test powered by an optimized physical classifier offers an ideal alternative for resource-limited settings where laboratory infrastructure may be scarce.

Nanoflake-based breath sensor delivers ultrasensitive lung cancer screening

While the researchers emphasize that larger validation trials are required, the work signals a profound shift in medical diagnostics. It mirrors advanced data-modelling techniques used in engineering, such as those predicting structural degradation in self-healing aerospace composites.

By fusing analytical physics with artificial intelligence, this research points to a future where immediate, automated bedside disease detection could be rapid, accessible and driven entirely by the mathematical profiling of a patient’s breath.