Plasma Lipidomic Signatures Across the Healthy-Pre-COPD-COPD Continuum Identified by Machine Learning.

Objective: Chronic obstructive pulmonary disease (COPD) imposes a substantial global burden, and Pre-COPD is regarded as an early, high-risk window that comprises distinct phenotypes, including small airway dysfunction (SAD), emphysema, and preserved ratio impaired spirometry (PRISm), whose early molecular heterogeneity is not fully captured by spirometry or imaging. This study aimed to characterize plasma lipidomic profiles across the healthy-Pre-COPD-COPD continuum, including these three phenotypes, and to apply regularized machine learning to identify lipid signatures shared across or specific to individual phenotypes. Methods: In this single-center, cross-sectional, observational study, 124 participants were enrolled, comprising 30 healthy controls, 63 individuals with Pre-COPD (SAD, n=23; emphysema, n=20; PRISm, n=20), and 31 patients with COPD. Untargeted plasma lipidomics was performed by UHPLC-high-resolution mass spectrometry. Differential lipids were identified by OPLS-DA (VIP > 1) and P values, and those that remained associated with disease group after adjustment for sex, age, BMI, and smoking (FDR < 0.05) were retained as candidate features. Elastic-net-regularized multinomial logistic regression was then applied for feature selection and to assess the discriminative performance of the selected lipids, evaluated by internal five-fold nested cross-validation with bootstrap stability selection. For each task, clinical-only, lipid-only, and combined models were compared, under a two-stage design comprising three phenotype-specific models (Healthy-SAD-COPD, Healthy-Emphysema-COPD, Healthy-PRISm-COPD) and a merged Healthy-Pre-COPD-COPD model. Results: A total of 171 differential lipids were identified. The merged Healthy-Pre-COPD-COPD model achieved a macro-averaged area under the curve (AUC) of 0.85 (95% CI 0.79-0.91), and the Healthy-SAD-COPD, Healthy-Emphysema-COPD, and Healthy-PRISm-COPD models achieved 0.90 (0.85-0.95), 0.92 (0.87-0.97), and 0.83 (0.74-0.90), respectively; across all tasks, models incorporating the selected lipids outperformed clinical-only models, whose macro-averaged AUCs ranged from 0.54 to 0.73. A compact set of recurrently selected, high-stability lipids, namely the ceramides Cer(d18:0/14:0), Cer(m18:0/18:0), and Cer(d18:1/22:0), PG(16:0/0:0), and LacCer(d16:0/16:0), was associated with disease stage and with pulmonary function. The phenotypes shared this signature but were further distinguished by phenotype-specific lipids, namely PI(18:0/20:4) and Cer(d18:1/22:0) in SAD, both associated only with small-airway indices and the latter being especially discriminative in this phenotype, PE(22:6/0:0) in emphysema, and GM3(d18:1/22:0) in PRISm. Conclusion: Under internal cross-validation, plasma lipidomic profiles discriminated healthy individuals, Pre-COPD phenotypes, and patients with COPD and were associated with pulmonary function, and a compact set of candidate lipids was identified. As this was an exploratory, cross-sectional analysis without longitudinal or external validation, these candidate markers require external and prospective validation before any clinical application.