Supporting Real World Decision Making in Coronary Diseases Using Machine Learning.

Area of Science:

• Cardiology and Medical Informatics
• Application of Artificial Intelligence in Healthcare

Background:

• Cardiovascular diseases represent a significant global health burden.
• Machine learning algorithms (MLA) show promise in medical decision-making.
• Previous studies often utilize cleansed datasets, lacking 'real-world noise'.

Purpose of the Study:

• To assess machine learning's utility in supporting coronary artery disease diagnosis.
• To evaluate MLA performance on routinely collected, 'noisy' clinical data.
• To identify optimal MLAs for real-world cardiovascular databases.

Main Methods:

• A two-phase study was conducted using the Anonymous Cardiovascular Database (ACD) and the UCI Heart Disease Dataset.
• Phase 1: Trained and validated MLAs on the UCI dataset to establish baseline accuracy (ACU) and select optimal models.
• Phase 2: Applied selected MLAs to the ACD, a hospital-derived, real-world dataset, to assess performance on noisy data.

Main Results:

• Seven MLAs were selected, achieving a standard ACU of 0.85 on the UCI dataset.
• The same MLAs achieved significantly higher ACUs, around 0.96, on the ACD.
• Decision trees and neural networks demonstrated superior performance on the ACD, while linear regression and AdaBoost performed best on the UCI dataset.

Conclusions:

• Machine learning algorithms perform differently on 'clean' versus 'noisy' clinical data.
• Decision trees and neural networks are better suited for handling real-world clinical data noise.
• These MLAs can effectively support clinical decision-making for coronary artery diseases using routinely collected data.