Selection of optimal features for classification of electrocardiograms

Area of Science:

• Biomedical Engineering
• Cardiology
• Machine Learning

Background:

• Accurate classification of electrocardiograms (ECGs) is crucial for diagnosing conditions like myocardial infarction (MI).
• Feature selection is a key step in developing robust ECG classification models.
• Evaluating different feature selection algorithms is essential for optimizing diagnostic performance.

Purpose of the Study:

• To compare the performance of forward sequential selection, backward sequential rejection, and branch and bound algorithms for ECG feature selection.
• To determine the most effective algorithm for classifying patients with and without old myocardial infarction based on ECG data.
• To assess the trade-offs between computational effort and classification accuracy for each algorithm.

Main Methods:

• Evaluated three feature selection algorithms: forward sequential selection, backward sequential rejection, and branch and bound.
• Applied algorithms to a dataset of ECGs from patients with old myocardial infarction and healthy subjects.
• Utilized metrics such as Mahalanobis distance, association index, sensitivity, and specificity for evaluation.

Main Results:

• Branch and bound algorithm is suitable for small feature sets but computationally prohibitive for large sets.
• Sequential selection algorithms achieved satisfactory and consistent classification accuracy by maximizing Mahalanobis distance.
• Maximizing the association index yielded better results but required more computation.
• Feature selection based on maximizing sensitivity at a fixed specificity level was less effective for high specificity requirements.

Conclusions:

• Sequential selection algorithms provide a practical and effective approach for ECG feature selection in myocardial infarction classification.
• The choice of feature selection algorithm should consider the size of the feature set and the required computational resources.
• Maximizing Mahalanobis distance offers a good balance between accuracy and computational efficiency for continuous ECG features.