Development and Multi-center Validation of a Breathomics-Based Triage Tool for Lung Cancer: A Prospective Study of 5,214 Participants.

🎯 Background & Rationale

The high false-positive rate of low-dose computed tomography (LDCT) for lung cancer screening necessitates efficient, non-invasive triage tools to differentiate malignant pulmonary lesions from benign mimics. While exhaled breath analysis of volatile organic compounds (VOCs) holds promise, clinical translation has been hindered by small-scale designs, lack of molecular resolution, and an absence of external validation in real-world symptomatic cohorts presenting with indeterminate pulmonary abnormalities.

👥 Study Design & Population

• Study Type: Large-scale, prospective, multicenter diagnostic accuracy study.
• Participants: A total of 5,214 symptomatic adult patients with radiologically detected pulmonary abnormalities requiring diagnostic workup across two campuses of a tertiary center.
• Cohorts: Partitioned into a discovery cohort (n = 4,669) for model training and internal tuning, and a geographically independent external validation cohort (n = 545).

📈 Methodology & Rigor

• Technology: Exhaled volatile organic compounds (VOCs) were analyzed using high-throughput proton-transfer-reaction time-of-flight mass spectrometry (PTR-TOF-MS).
• Modeling & Validation: An eXtreme Gradient Boosting (XGBoost) machine learning model integrated 40 prioritized VOC features with patient age and smoking history. A strict double-blind protocol was implemented, and the prediction model was locked prior to evaluation on the independent external validation cohort.
• Statistical Rigor: Sample size adequacy was confirmed via event-driven metrics (EPC ≈ 50) and the pmsampsize algorithm. Model performance was evaluated via AUC, sensitivity, specificity, NPV, PPV, calibration curves, and Decision Curve Analysis (DCA).

🔬 Key Findings [or Planned Endpoints]

• Internal Testing Set: The integrated model achieved an AUC of 0.891 (95% CI: 0.864–0.915), outperforming both clinical-only and VOC-only models.
• External Validation Cohort: The model maintained robust performance with an AUC of 0.850 (95% CI: 0.805–0.890). At a pre-specified "rule-out" decision threshold targeting high sensitivity, the model yielded a sensitivity of 93.1%, specificity of 55.9%, and NPV of 71.0% (driven by the high overall cohort malignancy prevalence of 76.7%).
• Intended-Use Subgroup (Pulmonary Medicine): In the targeted outpatient/respiratory setting with a prevalence of 44.5%, the model achieved an NPV of 89.0% and a specificity of 55.8% (sensitivity 91.4%).
• Subgroup Performance: Robust discriminative capability was maintained in early-stage disease (Stage 0/I AUC 0.849) and when distinguishing lung cancer from active pneumonia (AUC 0.849).

⚖️ Critical Appraisal

• Strengths: Exceptional sample size (over 5,000 patients), prospective multicenter design, independent geographical external validation, and strong biological grounding via mass-spectrometry-resolved metabolite and pathway attribution (e.g., pyruvate and microbial-host co-metabolism).
• Limitations: The high baseline malignancy prevalence in the validation cohort required Bayesian projection to estimate performance in lower-prevalence screening or outpatient settings. Retrospective exclusion of approximately 11.1% of patients lacking definitive pathological confirmation introduces potential selection bias. Furthermore, generalizability across different healthcare infrastructures and regional populations requires broader multi-center evaluation.

💡 The Clinical Bottom Line

This study provides robust prospective validation for a mass-spectrometry-based, molecularly resolved breathomics triage tool. In an outpatient pulmonary medicine setting, the model demonstrates a high negative predictive value (89.0%), establishing its potential utility as a non-invasive "gatekeeper" to safely rule out malignancy, streamline diagnostic pathways, and reduce unnecessary invasive procedures for indeterminate pulmonary abnormalities.