Exhaled Breath Electronic Nose for Multi-Disease Risk Prescreening (BREATH-PAN)

This is an investigator-initiated, prospective, multicenter, observational, staged platform cohort study involving hospitals and health examination centers. The study is designed to establish a standardized and traceable exhaled-breath electronic-nose platform and to evaluate its potential role as a first-line risk prescreening tool rather than as a stand-alone diagnostic test. Up to 20,000 participants will be prospectively enrolled.

The first stage is a methodological standardization and cross-center consistency cohort of approximately 500 clinically stable participants. Each participant will complete three predefined breathing maneuvers in randomized order: end-expiratory sampling without breath holding, sampling after a 5-second breath hold, and sampling after a 10-second breath hold. Each maneuver will be repeated twice. At least 250 participants will undergo repeat sampling at 7±3 days and 30±7 days, and selected participants will undergo cross-device or cross-center bridge sampling. Repeatability, within-participant variability, sample validity, tolerability, and device/center consistency will be evaluated to determine the locked sampling and quality-control procedure for the subsequent large cohort.

The second stage is a multi-disease risk prescreening cohort of approximately 19,500 participants. Prespecified major disease modules include lung cancer/clinically significant pulmonary nodules, chronic respiratory diseases, and metabolic/endocrine diseases. Liver, kidney, cardiovascular, and systemic inflammatory/infectious/immune conditions will be evaluated as secondary or exploratory modules. Because participants may have more than one condition, disease outcomes will be modeled using a multi-label rather than mutually exclusive classification framework.

Approximately 12,000 participants will be used for model development, approximately 3,000 for temporally independent evaluation, and approximately 4,500 from prespecified held-out centers for whole-center external evaluation. Preprocessing procedures, model architecture, features, module-specific screening thresholds, and the overall screening rule will be locked before evaluation data are analyzed. Temporal and whole-center external evaluation datasets will not be used for model tuning or threshold selection.

Reference status for each disease module will be determined from routine clinical or health-screening information, including laboratory testing, pulmonary function testing, imaging, pathology, specialist assessment, and routine follow-up as applicable. The study will not require additional CT scans, blood sampling, endoscopy, surgery, or other invasive procedures solely to establish the reference standard. Participants with insufficient reference information will not be included in the corresponding primary accuracy analysis.

Primary evaluation of the locked risk prescreening models will include sensitivity, specificity, false-negative rate, and 95% confidence intervals at fixed thresholds. Positive predictive value, negative predictive value, and screening-positive rate will also be reported in consecutively recruited intended-use populations without case enrichment. Secondary analyses will assess AUROC, AUPRC, likelihood ratios, Brier score, calibration, decision-curve net benefit, subgroup performance, and transportability across centers, devices, and acquisition periods.

A nested longitudinal/repeat-sampling cohort of approximately 600-1,000 participants, with an upper limit of 3,000, may undergo repeat breath sampling at routine preoperative/postoperative, treatment, follow-up, or health examination visits to explore within-person temporal changes in breath profiles. These analyses are exploratory and are not part of the primary risk prescreening conclusions.