Abstract 14440: Machine Learning-Based Approaches to Identify Diabetic Cardiomyopathy

Abstract

Introduction: Machine learning methodology can provide means for the identification of diabetic cardiomyopathy (DbCM), a severe and evolving complication of diabetes that leads to high morbidity and mortality. Phenotypic characterization of patient subgroups may support clinically relevant risk stratification in the population with DbCM. Methods: Among individuals with diabetes from the ARIC study cohort (training, n=953), unsupervised hierarchical clustering was performed with 24 candidate variables incorporating echocardiographic parameters, NT-proBNP, and hs-cTnT. The cluster with highest risk of HF was identified as DbCM. A deep learning (DL) classifier was developed to predict DbCM in the ARIC training cohort and validated in a pooled community-based cohort (ARIC testing plus CHS; n=1,050) and an electronic health record (EHR) cohort (n=3,139). Results: Clustering identified 3 phenogroups. Participants in group 3 (vs. 1 and 2) were more commonly men, had higher levels of creatinine, hs-cTnT, and NT-proBNP, higher LA size and LVMi, and increased prevalence of diastolic dysfunction and hypertension. The 5-year risk of HF was significantly higher in phenogroup 3 and thus identified DbCM (17.8% vs. 2.0% [phenogroup 2] vs. 3.5% [phenogroup 1]) ( Figure 1A ). The key predictors of DbCM were NTproBNP, LVMi, LA size, and diastolic dysfunction parameters ( Figure 1B ). The DL classifier demonstrated high model performance in identifying DbCM (AUROC = 0.96, accuracy = 0.93, and precision = 0.75). In the validation cohort (community-based), the DL classifier identified 16% of participants with DbCM with a two-fold higher risk of HF (HR [95% CI], 1.99 [1.47-2.67]; ref = no DbCM). A similar pattern of findings was observed in the EHR cohort (37% with DbCM; DbCM vs. no DbCM: HR [95% CI], 1.58 [1.17-2.12]). Conclusion: Machine learning-based techniques can be used to define and identify DbCM which is associated with higher risk of overt HF.