A data-driven approach to predicting diabetes and cardiovascular disease with machine learning

Area of Science:

• Data science
• Medical informatics
• Machine learning

Background:

• Diabetes and cardiovascular disease are leading causes of death in the US.
• Early identification and prediction of these diseases are critical for patient outcomes.
• This study focuses on leveraging patient data for disease risk detection.

Purpose of the Study:

• To evaluate machine learning models for detecting at-risk patients for cardiovascular disease, prediabetes, and diabetes.
• To identify key variables contributing to these diseases using survey and laboratory data.
• To develop an ensemble model for improved detection accuracy.

Main Methods:

• Utilized the National Health and Nutrition Examination Survey (NHANES) dataset.
• Applied supervised machine learning models including logistic regression, support vector machines, random forest, and gradient boosting.
• Developed a weighted ensemble model and used information gain for variable importance analysis.

Main Results:

• The ensemble model achieved high AU-ROC scores for cardiovascular disease (83.9%) and prediabetes (73.7%).
• XGBoost model demonstrated strong performance in diabetes detection (95.7% with lab data) and prediabetes (84.4% with lab data).
• Key predictors identified include waist size, age, self-reported weight, leg length, sodium intake, blood pressure, and chest pain.

Conclusions:

• Machine learning models analyzing survey data can automate the identification of patients at risk for diabetes and cardiovascular diseases.
• Identified key predictive variables offer insights for electronic health record (EHR) system improvements.
• This approach facilitates early intervention and disease management.