Development and validation of a machine learning-based approach to identify high-risk diabetic cardiomyopathy...

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Diabetic cardiomyopathy (DbCM) is a subclinical stage of myocardial abnormalities preceding heart failure (HF) in individuals with diabetes.
• Comprehensive characterization of DbCM phenotypes is lacking, hindering targeted interventions.

Purpose of the Study:

• To develop and validate a machine learning (ML) clustering approach to identify a high-risk DbCM phenotype.
• To utilize echocardiographic and cardiac biomarker data for DbCM phenotyping.

Main Methods:

• Unsupervised hierarchical clustering was applied to echocardiographic and biomarker data from the ARIC cohort (n=1199).
• A deep neural network (DeepNN) classifier was developed and validated in external cohorts (CHS, n=802; UT Southwestern EHR, n=5071).
• Phenotypes were characterized by 5-year heart failure incidence.

Main Results:

• Clustering identified three phenogroups, with Phenogroup-3 (27%) exhibiting significantly higher 5-year HF incidence.
• Key predictors of high-risk DbCM included elevated NT-proBNP, increased left ventricular mass/atrial size, and impaired diastolic function.
• The DeepNN classifier identified 16-29% of individuals with DbCM in validation cohorts, who showed higher HF incidence.

Conclusions:

• ML techniques can identify a high-risk DbCM phenotype in 16-29% of individuals with diabetes.
• This phenotype is associated with increased risk of heart failure.
• Early identification may enable more aggressive implementation of HF preventive strategies.