Determination of lung cancer exhaled breath biomarkers using machine learning - a new analysis framework

Sci Rep 2025 AI 5 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Overview: Exhaled Breath as a Window Into Lung Cancer

The Promise of Breath-Based Diagnosis Lung cancer kills approximately 1.8 million people annually, yet most cases are caught at advanced stages when survival is poor. This study explores a radically non-invasive alternative: analyzing the air patients breathe out to detect cancer-related chemical signatures called volatile organic compounds (VOCs).

What Are VOCs? Volatile organic compounds are small molecules released into exhaled breath as byproducts of cellular metabolism. Cancer cells have abnormal metabolic activity, so they produce different VOC patterns than healthy cells. GC-MS (gas chromatography-mass spectrometry) was used to precisely identify and measure these compounds in 52 participants.

The Study Population Participants included lung cancer patients, healthy controls, and patients with tuberculosis (TB) - a critical design choice that allowed the researchers to separate cancer-specific VOC signals from those caused by other lung diseases.

Key Goal The researchers developed a new machine learning pipeline to identify the best VOC signatures for lung cancer diagnosis, moving beyond simple statistical comparisons to find robust biomarker combinations that could withstand real-world confounders.

TL;DR: This study used machine learning and exhaled breath analysis to identify chemical biomarkers specific to lung cancer, testing 52 participants including controls and tuberculosis patients to filter out false positives.
Pages 2-3
Machine Learning Pipeline: Filtering Signal from Noise

GC-MS Breath Collection Exhaled breath samples were collected from participants and analyzed using gas chromatography-mass spectrometry, which separates and identifies individual chemical compounds by their molecular mass and retention time. This yielded a large dataset of VOC measurements per subject.

Confounder Elimination A critical methodological step was identifying VOCs that were elevated in both lung cancer and tuberculosis. These shared compounds - which could lead to false positive cancer diagnoses in TB patients - were systematically removed from the candidate biomarker pool before any classification models were built.

PLS-DA as the Core Model Partial Least Squares Discriminant Analysis (PLS-DA) was selected as the primary classification model. PLS-DA is a supervised method that finds combinations of variables (VOCs) that best separate defined groups (cancer vs. healthy), making it well-suited for high-dimensional chemical data with many correlated features.

Model Validation Performance was evaluated using sensitivity (how well the model catches true cancer cases), precision (how many flagged cases are truly cancer), accuracy (overall correct classification rate), F1 score (harmonic mean of sensitivity and precision), and AUC (area under the ROC curve).

TL;DR: The pipeline used GC-MS to measure exhaled VOCs, then applied confounder elimination and PLS-DA machine learning to build a lung cancer classifier robust to tuberculosis as a confounding disease.
Pages 3-5
Key Findings: 10 VOCs and Strong Classification Performance

PLS-DA Outperforms Other Models Among the machine learning models tested, PLS-DA achieved the best performance: sensitivity of 82%, precision of 90%, accuracy of 80%, F1 score of 86%, and AUC of 0.96. The AUC of 0.96 indicates near-excellent ability to distinguish lung cancer patients from healthy individuals.

10-VOC Biomarker Signature After the confounder elimination step, 10 VOCs were identified as specific lung cancer biomarkers. These compounds were elevated in lung cancer patients but not in tuberculosis patients or healthy controls, making them candidates for a disease-specific diagnostic test.

Tuberculosis as a Critical Confounder Without the confounder elimination step, several VOCs associated with lung cancer were also found in TB patients. This finding demonstrates why comparing only cancer patients against healthy controls - as many prior studies did - can produce misleading biomarker candidates that would fail in clinical practice.

Biological Interpretation The identified VOCs are consistent with disrupted metabolic pathways known to be active in lung tumors, including oxidative stress, lipid peroxidation, and altered amino acid metabolism - providing biological plausibility for their use as cancer markers.

TL;DR: PLS-DA achieved AUC 0.96 with 82% sensitivity and 90% precision, and 10 VOCs specific to lung cancer (not confounded by tuberculosis) were identified as candidate breath biomarkers.
Pages 5-6
Why Breath-Based Diagnosis Could Change Clinical Practice

Advantages Over Current Screening Low-dose CT (LDCT) scanning is the current gold standard for lung cancer screening, but it requires expensive equipment, radiation exposure, and radiologist expertise. A breath test could be performed anywhere, repeated as often as needed, and at a fraction of the cost.

The Confounder Problem in Clinical Settings Clinics serving high tuberculosis prevalence populations - common in many regions of Asia, Africa, and Eastern Europe - would encounter exactly the confounding problem this study addressed. A VOC test validated only against healthy controls would produce unacceptably high false positive rates in such settings.

Screening vs. Diagnosis Breath-based VOC analysis is most likely to serve as a triage tool - identifying patients who need further evaluation by CT or biopsy - rather than a definitive diagnostic test. The high AUC of 0.96 suggests it could be a very effective first-line screen.

Standardization Challenges GC-MS is a laboratory instrument requiring expertise to operate. For breath testing to reach primary care settings, the technology would need to be miniaturized and standardized - challenges that ongoing sensor technology development is beginning to address.

TL;DR: A VOC breath test could serve as a low-cost, non-invasive lung cancer screening tool, but validation against confounding diseases like tuberculosis is essential for real-world accuracy.
Pages 6-7
Limitations and Future Directions

Small Sample Size With only 52 participants, this study is exploratory rather than definitive. Larger cohorts are needed to confirm the 10-VOC signature and validate performance across diverse populations, smoking histories, and cancer histological subtypes.

Single Institution All participants were recruited from one center, which limits generalizability. Breath VOC profiles can be affected by diet, ambient air quality, medications, and other lung diseases that may vary by geography and population.

External Validation Needed The PLS-DA model was evaluated on data from the same study. Independent validation in a completely separate cohort is the critical next step before any clinical translation can be considered.

Future Work Future studies should expand the panel of confounding diseases tested (COPD, pneumonia, asthma), develop portable sensor alternatives to GC-MS, and investigate whether VOC profiles differ between lung cancer subtypes (adenocarcinoma vs. squamous cell carcinoma) or early versus late stage disease.

TL;DR: The 52-patient study is preliminary; larger multi-center validation studies testing a broader range of confounding conditions are needed before this breath test could be used clinically.
Citation: Open Access, 2025. Available at: PMC12274551.