ECG analysis using multiple instance learning for myocardial infarction detection.

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Cardiology Diagnostics

Background:

• Automated detection of myocardial infarction from ECGs is challenging due to the large volume of heartbeats and the cost of manual labeling.
• Supervised learning methods have shown limited success in ECG classification because of these labeling constraints.

Purpose of the Study:

• To develop a fully automatic technique for detecting myocardial infarction using ECGs.
• To address the limitations of supervised learning in ECG classification by proposing a novel Multiple Instance Learning (MIL) strategy.

Main Methods:

• The study proposes a 'latent topic MIL' strategy, mapping ECGs into a topic space derived from unlabeled heartbeats.
• A Support Vector Machine (SVM) is then applied directly to these ECG-level topic vectors for classification.

Main Results:

• The proposed latent topic MIL algorithm successfully detected myocardial ischemia in real ECG datasets from the PTB database without requiring any heartbeat labeling.
• The algorithm demonstrated improved classification performance, enhancing both sensitivity and specificity compared to existing MIL and supervised learning algorithms.

Conclusions:

• The latent topic MIL approach offers an effective solution for automated ECG classification, overcoming the need for manual heartbeat labeling.
• This method significantly improves the accuracy of detecting myocardial infarction and ischemia from patient ECGs.