A Machine Learning Approach to Respiratory Biomechanics: Identifying a PCA-Derived Membrane-Volume Coupling Pattern in Post-COVID Pulmonary Function.

BACKGROUND: Persistent dyspnoea after COVID-19 may not be fully explained by conventional pulmonary function tests. This retrospective cross-sectional study used secondary data to examine relationships between lung-volume and diffusion-related variables within an exploratory respiratory-biomechanical framework. METHODS: Participant-level data from 572 post-COVID participants and 72 healthy controls were obtained from an open-access multicenter dataset. Four derived relational descriptors and conventional pulmonary variables were entered into principal component analysis (PCA). Covariate-adjusted regression examined group differences in principal respiratory component (PRC) scores. An artificial intelligence model based on a multilayer perceptron (MLP) and a logistic regression benchmark used PRC1-PRC4 with a common stratified 70:30 training-testing partition. RESULTS: Four PCA-derived patterns explained 88.13% of variance. PRC2, descriptively labelled the membrane-volume coupling pattern from positive DLNO and DLCO loadings and a negative MVD loading, was lower in post-COVID participants after adjustment for age, sex, and body mass index (beta = -0.802, 95% CI -1.044 to -0.559, p < 0.001). Internal testing AUC was 0.814 (95% CI 0.734-0.894) for MLP and 0.738 (95% CI 0.657-0.819) for logistic regression; the paired difference was not significant (0.076, 95% CI -0.005 to 0.157, p = 0.066). CONCLUSIONS: Post-COVID status was associated with a lower score on the sample-derived diffusion-volume component pattern. The MLP showed numerically higher internal discrimination than logistic regression, while the between-model comparison remained statistically inconclusive. These findings highlight PRC2 as a sample-derived diffusion-volume pattern associated with post-COVID status and support independent validation of the component structure and classification performance in external cohorts.