Development and Validation of Feature-Based Machine Learning for ECG Artifact Detection and Classification

ECG signals are vital for diagnosing cardiovascular diseases, but artifacts like power line interference, baseline wander, and motion artifacts hinder accurate interpretation. This study aims to develop a robust machine learning model for reliable artifact classification. We extracted ECG features across Time, Frequency, Time-frequency, and Decomposition domains. After Recursive Feature Elimination (RFE), 45 features were selected. Light Gradient Boosting Machine (LightGBM) was used for classification, trained on the KURIAS ECG database (36,000 records). External testing included the PhysioNet CinC2011 (12,000 records) and MIT-BIH NST (800 records). The model achieved an F1 score of 89.88% for binary artifact detection. For multiclass classification across five artifact types, it obtained an average F1 score of 88.35%. External testing yielded average F1 scores of 94.35% on the CinC2011 database and 92.70% on the MIT-BIH NST database. Our model effectively classifies ECG artifacts, enhancing diagnostic reliability. Future work will validate the model in clinical environments and explore real-time implementation.