Hemoglobin A1c control is an independent predictor of circulating troponin concentrations using machine learning

Abstract Background Diabetes mellitus (DM) is associated with multiple comorbidities that may precipitate cardiac damage and elevated cardiac troponin (cTn) concentrations including renal failure, hyperlipidemia (HLD), hypertension (HTN), coronary artery disease (CAD), and congestive heart failure (CHF). Diabetics >70 years old have a higher likelihood of adverse outcomes including major cardiovascular events and cardiac mortality. However, the risk to younger diabetic patients and the precipitating factor(s) that facilitate cardiac injury in those with DM are unknown. The objective of this study was to investigate factors predictive of cTn concentrations in young, relatively healthy diabetic patients. Methods We collected 1533 remnant plasma samples from outpatients between 06/22–09/22 with physician ordered hemoglobin A1c testing. cTn was measured using the Abbott ARCHITECT High Sensitivity Troponin-I assay (limit of detection = 1.7 ng/L, imprecision = 4.76% at 50 ng/L). Demographic information (sex, race, BMI) and pertinent medical history (diabetes, HTN, HLD, CAD, CHF) were collected from the electronic medical record, along with estimated glomerular filtration rate (eGFR) and hemoglobin A1c. Exclusion criteria included: patients with missing laboratory data, those undergoing cancer treatment, or a history of myocardial infarction/cardiomyopathy/cardiac surgery. Troponin results were classified as normal- or high-risk using a cut-off of 10 ng/L for females and 12 ng/L for males, thresholds previously shown to correlate with incidence of cardiovascular disease risk within 15 years. Univariate statistics were calculated using Wilcoxon rank sum tests. An XGBoost model was trained to predict high-risk troponinemia, and summary statistics were calculated on a held-out test set. Results Of the 1135 patients that met inclusion criteria, 621 (54.7%) were female. The median age was 60 years (IQR: 49–69 years) and the median A1c was 6.2% (IQR: 5.8%–7.0%). A total of 746 patients (65.7%) had a prior diabetes diagnosis, with 156 patients (13.7%) having prediabetes, 42 patients (3.7%) having Type 1, and 548 patients (48.3%) having Type 2. Median troponin concentration for patients with an A1c <5.7% was 1.6 ng/L (IQR: 0.8–3.4), 5.7%–6.4% was 2.2 ng/L (IQR: 1.3–4.8), and ≥6.5% was 2.9 ng/L (IQR: 1.6–6.7). Univariate analysis demonstrated significant differences in troponinemia by age (CI of difference: 4.4–8.9 years), A1c (0.2%–0.7%), and eGFR (28–38 mL/min/1.73 m2). The machine learning model demonstrated strong predictive capacity (sensitivity: 0.74, specificity: 0.86, PPV: 0.5, NPV: 0.92, area under ROC curve: 0.91). The features with the greatest impact on area under the ROC curve when removed were eGFR, Age, A1c, and CHF. Conclusion A combination of clinical and laboratory variables can be used to predict circulating troponin concentrations in outpatients. A1c control was an independent contributor to cTn concentrations, both in univariate and multivariate analyses. This study provides evidence that glucose control may be associated with cardiac damage and future cardiovascular events, warranting longitudinal outcome studies.