Development and validation of a screening model for lung cancer using machine learning: A large-scale, multi-center study of biomarkers in breath

Jing Li, Yuwei Zhang, Qing Chen, Zhenhua Pan, Jun Chen, Meixiu Sun, Junfeng Wang, Yingxin Li, Qing YeFrontiers in Oncology 2022 · DOI

TL;DR

Researchers have developed a non-invasive lung cancer screening model that identifies specific "breath-biomarkers" using high-resolution mass spectrometry and machine learning. By analyzing volatile organic compounds (VOCs) in exhaled air, the team identified a six-feature panel that can distinguish lung cancer patients from healthy controls with high accuracy, offering a potential radiation-free alternative to traditional CT scans.

What Problem Does This Paper Address?

Lung cancer (LC) remains the leading cause of cancer-related deaths globally. While early-stage detection can boost five-year survival rates to 80%, most cases are caught too late for curative treatment.

Currently, the gold standard for screening is Low-Dose Computed Tomography (LDCT). However, LDCT has significant drawbacks: it involves ionizing radiation, carries a high rate of false positives (leading to unnecessary invasive biopsies), and is expensive to implement at scale. Breath analysis—detecting the metabolic "waste" of cancer cells—has long been proposed as a solution, but previous studies were often too small, lacked external validation, or relied on "electronic noses" that couldn't provide precise chemical quantification.

Methods

The researchers employed a sophisticated analytical pipeline combining real-time chemistry with gradient-boosted machine learning:

  1. Sampling: They used Buffered End-Tidal (BET) online sampling. Unlike traditional methods where patients breathe into a bag (which can introduce contaminants), BET allows the subject to breathe directly into the instrument, capturing air specifically from the deep lungs (alveoli) where gas exchange with the blood occurs.
  2. Analysis: The air was analyzed using Proton Transfer Reaction Time-of-Flight Mass Spectrometry (PTR-TOF-MS). This technique allows for real-time, high-sensitivity detection of VOCs at parts-per-trillion levels without complex sample preparation.
  3. Modeling: They utilized eXtreme Gradient Boosting (XGBoost) to rank 318 extracted features by importance.
  4. Optimization: To prevent overfitting and ensure clinical utility, the authors used an iterative process to find the minimum number of features required to reach 99% of the maximum possible Area Under the Curve (AUC). This resulted in a streamlined six-biomarker panel.

Datasets

This study is notable for its scale and multi-center design:

  • Derivation Cohort: 2,043 participants (1,007 LC patients, 1,036 healthy controls) from the Cancer Institute and Hospital, Tianjin Medical University.
  • External Validation Cohort: 265 participants (158 LC patients, 107 healthy controls) from the General Hospital of Tianjin Medical University.
  • Significance: Most previous "pilot" studies in this field involved fewer than 200 participants. This larger dataset provides much-needed statistical power and helps account for individual physiological variability.

Key Insights and Ideas

  • The "Minimalist" Panel: The researchers found that while 12 features provided the absolute highest accuracy, a 6-feature panel captured nearly all the predictive power. This simplicity is vital for future development of portable, low-cost diagnostic hardware.
  • Biological Fingerprinting: The identified biomarkers—including acetone, methanol, and phenol—are not random. They correlate with known metabolic shifts in cancer cells, such as altered lipid metabolism and increased oxidative stress. For example, acetone is a byproduct of ketone metabolism, which is often upregulated in tumor environments.
  • Overcoming the "Small Study" Trap: By using a multi-center approach and an independent validation set from a different hospital, the authors addressed the "replicability crisis" that has plagued breath-analysis research for decades.

Key Results

The model demonstrated strong performance, though it highlighted the "real-world" challenge of external validation:

  • Training Performance: In the derivation cohort, the model achieved an AUC of 0.963, with a sensitivity of 87.1% and a specificity of 93.5%.
  • External Validation: When tested on the independent hospital dataset, the performance dropped to an AUC of 0.771, with 67.7% sensitivity and 73.0% specificity.
  • Consistency: The model remained effective across different histological subtypes (like adenocarcinoma and squamous cell carcinoma) and across different stages of the disease.

Limitations and Future Work

The authors are transparent about several hurdles:

  • The "Validation Gap": The significant drop in AUC from 0.96 to 0.77 suggests that environmental factors at different hospitals (e.g., ambient air quality) or differences in patient populations still influence the results.
  • Case-Control Bias: The study compared confirmed cancer patients with healthy controls. In a real-world screening scenario, the model would encounter "borderline" cases—people with benign lung nodules or other inflammatory lung diseases—which might lower the specificity.
  • Next Steps: The team emphasizes the need for prospective studies in primary care settings (where cancer prevalence is much lower) to truly validate the model as a frontline screening tool before patients ever reach a specialist hospital.

Summary updated

Abstract

Our results suggested that breath analysis may serve as a valid method in screening lung cancer in a borderline population prior to hospital visits. Although our breath-biomarker panel is noninvasive, quick, and simple to use, it will require further calibration and validation in a prospective study within a primary care setting.
BibTeX
@inproceedings{li2022development,
  author = {Jing Li and Yuwei Zhang and Qing Chen and Zhenhua Pan and Jun Chen and Meixiu Sun and Junfeng Wang and Yingxin Li and Qing Ye},
  title = {{Development and validation of a screening model for lung cancer using machine learning: A large-scale, multi-center study of biomarkers in breath}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.975563},
}
Figure