Metabolomic profiling refines cardiovascular risk stratification beyond SCORE2-diabetes in patients with concurrent type 2 diabetes and pulmonary dysfunction

Abstract

Background Patients presenting with both type 2 diabetes (T2D) and lung function impairment (LFI) are exposed to an amplified risk of major adverse cardiovascular events (MACE). Although the SCORE2-Diabetes algorithm is widely endorsed for clinical risk assessment, its prognostic accuracy within this specific multimorbid phenotype remains inadequately explored. This study sought to evaluate the baseline utility of SCORE2-Diabetes and to investigate whether integrating a novel, machine learning-derived metabolomic signature could optimize 10-year MACE prediction in this highly vulnerable population. Methods We conducted a prospective analysis of UK Biobank participants with type 2 diabetes (T2D) and no cardiovascular disease at baseline. A reference cohort of 12,130 participants with preserved lung function was used to assess the base clinical model, and the primary cohort comprised 3,706 participants with lung function impairment (LFI). We profiled 249 plasma metabolites using nuclear magnetic resonance (NMR) spectroscopy. A machine-learning framework combining least absolute shrinkage and selection operator (LASSO) Cox regression, random forest survival analysis, and extreme gradient boosting (XGBoost) was used to derive a consensus signature. Internal validation used 1,000 bootstrap resamples; the complete modeling pipeline was repeated within each resample. Incremental performance beyond SCORE2-Diabetes was assessed using Harrell’s concordance index (C-index), net reclassification improvement (NRI), calibration, and decision curve analysis (DCA). Results During a median follow-up of 12 years, SCORE2-Diabetes showed lower discrimination in the LFI cohort than in the reference cohort (C-index, 0.670 vs. 0.707). Among the 3,706 participants with LFI, 845 developed major adverse cardiovascular events (MACE). Adding the 12-metabolite signature increased the C-index from 0.670 to 0.722 (absolute difference, 0.052; P < 0.001), representing a statistically significant but moderate improvement in discrimination. The five-metabolite model had a C-index of 0.708, while the two-metabolite sensitivity model yielded a C-index of 0.699 (95% CI, 0.683-0.716). Using the 2023 European Society of Cardiology SCORE2-Diabetes categories, the categorical NRI was 14.8% (95% CI, 10.8%-18.8%). Internal calibration and DCA suggested improved agreement and net benefit, but these results require external validation. Conclusions In participants with T2D and LFI, adding a metabolomic signature to SCORE2-Diabetes produced a statistically significant but moderate improvement in risk discrimination. These internally validated findings support further evaluation in independent cohorts and prospective impact studies; they do not establish immediate clinical usefulness.