Harnessing exhaled breath for lung cancer early detection, results from the ExPeL study

The ExPeL study demonstrates that the Inflammacheck point-of-care device, which analyzes exhaled breath condensate using machine learning and metabolomics, effectively distinguishes early-stage lung cancer from controls with high accuracy and zero false positives, offering a promising non-invasive tool for primary care screening.

The Gist

Imagine your lungs are like a busy factory. When everything is running smoothly, the factory produces a steady, predictable stream of exhaust. But when a problem starts—like a small fire (cancer) beginning to smolder in the corner—the exhaust changes. It might smell different, get hotter, or carry tiny, unusual particles that weren't there before.

This paper is about a new, high-tech way to "sniff out" that smoke before the fire gets too big to see.

The Problem: Finding the Needle in the Haystack

Lung cancer is deadly, but it's much easier to treat if caught early. Currently, doctors use CT scans to look for it, but these scans are expensive, involve radiation, and can't be done on everyone. So, they try to guess who is at high risk based on age and smoking history. Unfortunately, this "guessing game" misses a lot of people who actually have cancer, while scanning many people who are perfectly fine.

The researchers wanted a better way to decide who needs a CT scan. They wanted a simple, non-invasive test that could act like a "smoke detector" for the lungs.

The Solution: The "Breathalyzer" for Lungs

The team developed a study called ExPeL. They used a handheld device called Inflammacheck®. Think of this device as a sophisticated breathalyzer, but instead of measuring alcohol, it measures the chemistry of your breath.

Here is how it works:

1. The Collection: You simply breathe normally (tidal breathing) into a mouthpiece for a few minutes. The device cools your breath down, turning the invisible water vapor and chemicals into a tiny drop of liquid called Exhaled Breath Condensate (EBC).
2. The Sensors: The device measures five things:
   • Hydrogen Peroxide: A marker of inflammation (like a "heat" sensor).
   • CO2: How well your lungs are exchanging gas.
   • Humidity & Temperature: The physical state of your breath.
   • Flow Rate: How hard you are pushing air out.
3. The Brain (AI): This is the magic part. The data from these sensors is fed into a computer brain (Machine Learning). The computer doesn't just look at one number; it looks at the pattern of all five numbers together, like a detective looking at the whole crime scene rather than just one clue.

The Experiment

The researchers tested this on 34 people from a UK lung screening program:

• Group A: People who had been found to have early-stage lung cancer.
• Group B: People who were smokers but had been cleared by a CT scan (the "controls").

The Results:
The computer brain was incredibly good at telling the two groups apart.

• Accuracy: It got it right about 86% of the time.
• The Superpower: Most importantly, it had zero false alarms. It didn't tell any healthy person they had cancer. In a screening program, this is huge because it means we won't scare people or waste money on unnecessary scans for healthy folks.
• The "Chaos" Factor: The researchers noticed something interesting. The healthy people's breath data was very consistent and clustered together. The cancer patients' data was all over the place (dispersed). It's like comparing a choir singing in perfect harmony (healthy) to a group of people shouting different notes (cancer). The cancer breath was "messier" and more chaotic, which the AI could detect.

The Secret Ingredients (Metabolomics)

To understand why the breath was different, the team also analyzed the liquid drop from the breath using a super-powerful microscope (Mass Spectrometry). They found over 2,000 tiny chemical molecules.

• They narrowed it down to just four specific molecules that acted like a "signature" for lung cancer.
• If you combined these four molecules, the test became even more accurate (97% accuracy).
• One of these molecules is a chemical often found in cleaning products. The researchers think the cancer might be messing up the body's ability to break it down, causing it to build up in the breath.

Why This Matters

Imagine a world where, instead of sending everyone to a CT scanner, a doctor in a regular clinic could just ask you to breathe into a small device.

• If the device says "All Clear": You go home, and you avoid radiation and anxiety.
• If the device says "Check this out": You get sent for a CT scan immediately.

This study shows that early-stage lung cancer leaves a unique chemical fingerprint on your breath. By using AI to read that fingerprint, we might be able to catch lung cancer much earlier, save lives, and stop wasting resources on people who don't need expensive scans.

The Bottom Line

This isn't a cure yet, but it's a very promising early warning system. It turns the simple act of breathing into a powerful diagnostic tool, using the chaos of cancer to distinguish it from the calm of health.

Technical Summary

1. Problem Statement

Lung cancer screening programs, such as the UK's Targeted Lung Health Check (TLHC), rely heavily on Low-Dose Computed Tomography (LDCT). However, current pre-screening clinical scores are insufficient, leading to low yield (only ~40% of lung cancer patients would have been eligible for CT screening prior to diagnosis) and high rates of false positives. There is an urgent need for scalable, non-invasive, point-of-care tools to triage patients effectively, distinguishing early-stage malignancy from benign respiratory conditions and low-risk controls in primary care settings.

2. Methodology

The ExPeL study was a feasibility pilot conducted within the Greater Manchester TLHC program, utilizing a multi-modal approach combining physiological breath analysis and untargeted metabolomics.

• Study Population:

  • Cases: Individuals with radiologically suspected lung cancer referred for biopsy (83% were early-stage, Stage I–II).
  • Controls: Ever-smokers from the TLHC program who underwent low-dose CT but were deemed low-risk (screen-negative) or had benign nodules. This represents a real-world screening comparator rather than lifelong healthy controls.
  • Sample Size: After quality control (excluding data due to sensor artifacts and insufficient sample acquisition), 34 participants with valid data were analyzed.

• Data Acquisition (Inflammacheck® Device):

  • Exhaled Breath Condensate (EBC) was collected via tidal breathing.
  • Physiological Parameters Measured: Hydrogen peroxide (H₂O₂) concentration, peak end-tidal CO₂, peak humidity, breath temperature, and mean exhalation flow rate.
  • Metabolomics: A subset of participants provided larger volume EBC samples for untargeted Liquid Chromatography–Mass Spectrometry (LC–MS) analysis.

• Analytical Approach:

  • Multivariate Statistics: Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Mahalanobis distance calculations were used to assess intrinsic group separation without prior class labeling (unsupervised) and with class labeling (supervised).
  • Machine Learning (ML): To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training data. Two ensemble models were trained and evaluated on held-out test sets:
    1. Stacked Ensemble: Random Forest + XGBoost (base) + Logistic Regression (meta).
    2. Soft Voting Ensemble: Random Forest + XGBoost + Logistic Regression.
  • Metabolomics: Sparse Partial Least Squares Discriminant Analysis (sPLS-DA) and Variable Importance in Projection (VIP) scores were used to identify discriminatory molecular features.

3. Key Contributions

• Real-World Validation: Unlike previous studies that excluded patients with chronic respiratory disorders, this study tested the device in a heterogeneous, real-world screening population including smokers with comorbidities.
• Integration of Physiology and Metabolomics: The study uniquely combined physiological breath parameters (H₂O₂, flow, temperature) with untargeted metabolomic profiling to create a comprehensive biomarker signature.
• Focus on Early-Stage Disease: The cohort was heavily skewed toward early-stage (I–II) cancers, demonstrating the feasibility of detecting malignancy before it becomes advanced.
• Novel Biomarker Discovery: Identification of specific volatile organic compounds (VOCs) and metabolic pathways associated with early lung cancer that can be integrated into future iterations of the Inflammacheck® device.

4. Results

• Multivariate Separation:

  • PCA: Showed distinct geometric structuring between cancer and control groups, with cancer cases exhibiting greater dispersion (physiological heterogeneity).
  • LDA & Mahalanobis Distance: Confirmed a linear discriminant axis and covariance-adjusted proximity that clearly separated the two groups, validating intrinsic physiological differences.

• Machine Learning Performance (Voting Ensemble Model):

  • Accuracy: 85.7%
  • Sensitivity: 80%
  • Specificity: 100% (Crucially, zero false positives were identified in the test set).
  • Precision (PPV): 100%
  • ROC–AUC: 0.90
  • Matthews Correlation Coefficient (MCC): 0.73
  • Note: The Voting Ensemble outperformed the Stacking Ensemble, which had lower specificity (50%).

• Metabolomic Findings:

  • LC–MS identified 2,132 molecular features.
  • A panel of four key metabolites achieved an AUC of 0.969 for cancer discrimination.
  • Identified Metabolites:
    1. Tri(propylene glycol) butyl ether: An organic solvent likely linked to disrupted degradation pathways in cancer.
    2. C₇H₁₀O (Putatively Hept-4-yn-2-one): Previously linked to lung cancer in urine studies.
    3. C₂H₆NO: Structural annotation pending.
  • Pathway Analysis: Revealed enrichment in sphingolipid biosynthesis and fatty acid biosynthesis, consistent with known cancer metabolic alterations.

• Physiological Observations:

  • Cancer patients showed lower median H₂O₂ levels compared to controls (contrary to some advanced-stage studies), likely due to steroid use suppressing NF-κB and inflammation in symptomatic early-stage patients. However, the dispersion (variance) of H₂O₂ and other physiological variables was significantly higher in cancer cases, a pattern captured only by multivariate analysis.

5. Significance and Clinical Implications

• Triage Utility: The 100% specificity and precision of the model suggest that Inflammacheck® could serve as an effective "rule-out" or triage tool in primary care. A negative result could confidently avoid unnecessary CT referrals, reducing radiation exposure and healthcare costs.
• Early Detection: The ability to distinguish early-stage (Stage I–II) disease is critical, as this is the stage where curative treatment is most viable.
• Scalability: The device is handheld, point-of-care, and non-invasive, making it suitable for integration into the UK's National Health Service (NHS) screening pathways.
• Future Directions: The study provides a proof-of-concept for integrating VOC metabolomics into the device. Future work requires prospective, multi-center validation in larger populations to confirm these findings and refine the predictive algorithm.

Conclusion: The ExPeL study demonstrates that exhaled breath condensate analysis, enhanced by machine learning and metabolomics, can effectively distinguish early-stage lung cancer from low-risk controls with high specificity, offering a promising adjunct to current CT-based screening protocols.