Deep learning-based CT-derived vertebral bone mineral density and metformin therapy: a longitudinal study in the Multi-Ethnic Study of Atherosclerosis.

Abstract

Evidence on metformin's skeletal effects remains conflicting. We emulated a target trial to evaluate associations between metformin therapy and deep learning-based CT-derived vertebral bone mineral density (vBMD). We also assessed variation across prespecified subgroups. Within the Multi-Ethnic Study of Atherosclerosis (MESA), incident metformin users (Exams 4 and 5) were compared with propensity score-matched controls. Noncontrast chest CT scans from Exams 5 and 6 were processed using a previously validated deep learning-based vertebral segmentation and calibration pipeline to quantify trabecular vBMD from T1 to T10. Median imaging follow-up was 6.4 years. Linear mixed-effects models estimated annualized Fracture Risk Assessment Tool (FRAX) absolute vBMD change, applying inverse probability of censoring weights. Prespecified subgroup analyses examined demographic, metabolic, and inflammatory modifiers. Among 238 trial entries (86 metformin, 152 controls), metformin was not associated with overall vBMD change (time × treatment interaction β, 0.27 mg/cm³ per year; 95%CI, -0.07 to 0.61; p = 0.12). Substudy-specific and per-protocol estimates were consistent. Favorable associations were observed in women, body mass index (BMI) < 30 kg/m², ASCVD risk < 0.2, hs-CRP < 2 mg/L, never-smokers, and nondrinkers. Significant effect modification was found for gender, hs-CRP, and smoking status, with borderline trends for BMI. Metformin use was not associated with overall CT-derived vBMD change. Subgroup analyses indicate heterogeneity of association across demographic, metabolic, and inflammatory profiles, with more favorable associations among women and participants with healthier risk profiles. Findings support metformin's skeletal safety and warrant future context-specific trials.