Explainable Machine Learning for Atrial Fibrillation in the General Population Using a Generalized Additive Model　- A...

Area of Science:

• Cardiology
• Medical Informatics
• Biostatistics

Background:

• Atrial fibrillation (AF) is a prevalent arrhythmia linked to stroke and heart failure.
• Existing machine learning models for AF risk lack interpretability for clinicians and patients.

Purpose of the Study:

• To develop an explainable and accurate model for predicting atrial fibrillation risk.
• To account for non-linear relationships between clinical factors and AF incidence.

Main Methods:

• Utilized data from 5,378 AF patients and 167,950 controls from health checkups (2009-2018).
• Employed a generalized additive model incorporating 47 clinical parameters.
• Validated model performance using area under the curve, sensitivity, and specificity.

Main Results:

• Achieved high model performance with an AUC of 0.964, sensitivity of 0.879, and specificity of 0.920.
• Identified key predictors: arrhythmia examination, coronary artery disease history, age, hematocrit, GGT, creatinine, hemoglobin, systolic blood pressure, and HbA1c.
• Visualized non-linear effects of clinical variables on AF probability.

Conclusions:

• Established a novel, interpretable, and accurate AF risk prediction model using general health checkup data.
• The model enhances AF risk assessment and understanding of contributing clinical factors.