Abstract Background Diabetes mellitus (DM) is linked to more diffuse and extensive coronary artery disease (CAD), yet binary DM status may miss risk granularity. Continuous glucose monitoring (CGM) data summarized with machine-learning could reveal glycemic phenotypes associated with coronary anatomy in acute coronary syndrome (ACS). Purpose To derive CGM-based patient phenotypes using unsupervised learning and test their association with CAD extent in ACS, beyond binary DM status. Methods ORACLE is a prospective, multicenter registry of high-risk ACS patients with continuous time-series data from multiple wearables. CGM (FreeStyle Libre 3 Plus, Abbott) was initiated during index hospitalization. From CGM time-series, 34 gold-standard features were computed and standardized; k-means clustering (k=3) identified CGM phenotypes. Coronary anatomy characteristics were compared across clusters using Kruskal–Wallis (continuous) and χ² tests (categorical). Results Among 251 ACS patients, three CGM phenotypes were identified. Cluster 0 (C0) had a higher prevalence of type 2 DM (72.1%) and higher HbA1c/discharge glucose; clusters 1 (C1) and 2 (C2) had similar, lower DM prevalence (36.0% and 36.4%). All other risk factors were broadly comparable among the three groups. Despite clustering being CGM-only based, CAD burden showed a stepwise gradient among clusters (C0C1C2): three-vessel disease 41.0% vs 27.1% vs 23.3% (p=0.05); number of lesions treated 2.2±1.0 vs 2.1±1.0 vs 1.9±1.2 (p=0.02); total stent length 59±41 mm vs 48±33 mm vs 47±36 mm (p=0.04); left main treatment 16.4% vs 12.8% vs 4.7% (p=0.03). Notably, C1 had worse anatomy than C2 despite similar DM prevalence and risk profiles. Conclusions Unsupervised clustering of CGM time-series identified three ACS phenotypes that stratified anatomic CAD burden in a graded fashion (C0 C1 C2). The C1–C2 separation, despite similar DM status and conventional risk factors, indicates that CGM dynamics capture risk granularity not reflected by binary DM classification. These findings support development of CGM-driven prediction tools to improve patient stratification and refine anatomic risk profiling after ACS.CGM-derived phenotypes in ACS: 3D UMAP vFor image description, please refer to the figure legend and surrounding text.
Costa et al. (Sun,) studied this question.