Interaction-informed EV FTIR spectroscopy reveals a covariate-modulated molecular ageing fingerprint of coronary atherosclerosis

Abstract Background Extracellular vesicles (EVs) are attractive candidates for label-free cardiovascular phenotyping, yet most spectroscopic approaches assume that disease exerts a uniform effect on vesicle composition. In reality, clinical covariates such as age and hypertension may remodel EV biochemistry and obscure disease-related signals. Purpose To determine whether explicitly modelling clinical covariate–spectrum interactions can uncover a latent EV FTIR fingerprint of coronary atherosclerosis. Methods Liquid-phase FTIR spectra (800–1,800 and 2,800–3,200 cm−¹, 1,022 wavenumbers) were acquired from serum-derived EVs in adults undergoing coronary CT imaging. After robust pre-processing and outlier removal, we first benchmarked elastic-net logistic regression and random forests for the detection of any coronary disease. We then quantified, at each wavenumber, the variance explained (R²) by age, sex, and risk factors, and identified wavenumbers with significant covariate×disease interactions. These terms were incorporated into a series of "interaction-informed" random forests. Discrimination was assessed by a 200-fold bootstrap AUC. Results Spectra-only models performed at chance level (AUC 0.53, 95% CI 0.47–0.60). Covariate scanning revealed a diffuse but structured footprint: sex influenced 167 bands, hypertension 112 and age 50, with 110 sex×disease and 25 age×disease interactions. Adding selected interaction terms increased AUC stepwise to 0.77 (0.72–0.82) in the differential-interaction model (ΔAUC +0.24 vs spectra-only, p0.001). The most stable predictors were age-dependent amide-II/III and carbohydrate bands (around 1,547 and 1,020 cm−¹) and a hypertension-linked CH2 deformation at 1,477 cm−¹, indicating accelerated, risk-factor-modulated "molecular ageing" of EV cargo in disease. Conclusions Circulating EV spectra alone carry little diagnostic information but become highly informative once clinical covariate–spectrum interactions are modelled. This interaction-informed FTIR framework uncovers a coherent molecular ageing fingerprint of coronary atherosclerosis. It provides a blueprint for integrating vesicle spectroscopy with risk-factor profiling in future multi-omic biomarker pipelines.Wavenumbers showing significant CADxageFor image description, please refer to the figure legend and surrounding text. Performance of random forest classifiersFor image description, please refer to the figure legend and surrounding text.