An Interpretable Hand-Crafted Feature-Based Model for Atrial Fibrillation Detection.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Atrial Fibrillation (AF) is a prevalent cardiac arrhythmia impacting patient prognosis.
• Early AF diagnosis is crucial for effective treatment and improved outcomes.
• Machine Learning (ML) enhances Computer-Aided Diagnosis (CADx) for AF detection.

Purpose of the Study:

• To apply and evaluate explanation techniques for ML models in heart rhythm classification.
• To assess the interpretability and diagnostic utility of ML models for cardiologists.
• To compare the performance of different explanation techniques on hand-crafted features for AF detection.

Main Methods:

• Utilized hand-crafted features from Electrocardiogram (ECG) signals.
• Applied various explanation techniques, including SHapley Additive exPlanations (SHAP).
• Employed Random Forest (RF) and Support Vector Machine (SVM) classifiers on the 2017 CinC/PhysioNet challenge dataset.

Main Results:

• The SHAP technique combined with RF achieved a mean F-score of 0.746 for AF detection.
• This performance surpassed the cascaded SVM approach (mean F-score of 0.706) using the same features.
• The interpretable model provided cardiologists with a compact feature set and actionable diagnostic information.

Conclusions:

• SHAP and RF offer an effective and efficient approach for interpretable AF detection from ECG signals.
• Interpretable ML models can simplify complex algorithms, providing tangible diagnostic insights for clinicians.
• This methodology aids in improving the accuracy and clinical utility of CADx systems for cardiac arrhythmias.