Machine learning-based MPO and Lp-PLA2 profiling reveals asthma-predominant inflammatory signature in asthma-COPD overlap

Background: Asthma-COPD overlap (ACO) remains poorly characterized at the molecular level, leading to diagnostic uncertainty and suboptimal treatment. We hypothesized that machine learning analysis of myeloperoxidase (MPO) and lipoprotein-associated phospholipase A2 (Lp-PLA2) could help reveal ACO as a distinct inflammatory endotype. Methods: We prospectively enrolled 609 patients with obstructive airway diseases (asthma n = 255, COPD n = 100, AECOPD n = 182, ACO n = 72) at First Hospital of Guangzhou Medical University (National Center for Respiratory Medicine) between July 2023 and August 2025. Serum MPO and Lp-PLA2 were quantified using chemiluminescence immunoassay. Five machine learning algorithms (XGBoost, LightGBM, CatBoost, Neural Network, Random Forest) were employed for supervised classification, while unsupervised clustering (hierarchical, k-means, DBSCAN) defined molecular phenotypes. Model interpretability utilized SHAP analysis and dimensionality reduction techniques (PCA, UMAP, t-SNE). Results: = 0.031). Unsupervised clustering revealed that 72.2% of clinically diagnosed ACO patients clustered with asthma, challenging traditional disease conceptualization. Ensemble machine learning model achieved exceptional performance (AUC 0.954, 95% CI: 0.926-0.982), significantly outperforming individual biomarkers (MPO/Lp-PLA2: AUC 0.782/0.714) and traditional clinical criteria (AUC 0.694). Clinically applicable ACORN (ACO recognition by neutrophil-eosinophil) score demonstrated 91.2% positive predictive value (PPV). Conclusion: Machine learning reveals ACO exhibits predominantly asthma-like inflammatory characteristics. MPO/Lp-PLA2 ratio and ACORN score enable precision phenotyping with potential therapeutic implications, advocating for biomarker-guided treatment strategies in obstructive airway diseases.