Detecting cardiovascular diseases using unsupervised machine learning clustering based on electronic medical records

Area of Science:

• Machine Learning
• Medical Informatics
• Cardiovascular Disease Research

Background:

• Electronic medical records (EMR) offer rich data for machine learning (ML) models in cardiovascular disease (CVD) risk prediction.
• Unsupervised ML approaches can potentially develop novel models for detecting prevalent CVDs in clinical practice.

Purpose of the Study:

• To test the hypothesis that unsupervised ML utilizing EMR data can develop a new model for detecting prevalent CVD.
• To evaluate the effectiveness of clustering models in identifying CVD cases within a large patient cohort.

Main Methods:

• Included 155,894 patients (≥18 years) from January 2014 to July 2022.
• Employed K-means clustering (k=2, 4, 8) and Bayesian theorem for predictive accuracy estimation.
• Utilized Principal Component Analysis (PCA) for dimensionality reduction.

Main Results:

• Clustering models achieved high predictive accuracy (0.85-0.87) on both training and testing sets.
• 2-, 4-, and 8-classification models demonstrated comparable predictive performance.
• PCA revealed significant separation between CVD and non-CVD cases after dimension reduction.

Conclusions:

• EMR data combined with unsupervised ML can create robust models for CVD detection.
• Further longitudinal studies are necessary to refine the model for clinical application.