Weak Supervision for Affordable Modeling of Electrocardiogram Data.

Area of Science:

• Computational Biology and Bioinformatics
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Electrocardiograms (ECGs) are crucial for diagnosing heart disease.
• Current machine learning models for ECG analysis require extensive manual annotation, which is costly and time-consuming.
• Automated detection of abnormal heartbeats is hindered by the lack of large, labeled datasets.

Purpose of the Study:

• To explore the use of multiple weak supervision sources for training diagnostic models of abnormal heartbeats.
• To develop a method that bypasses the need for manually annotated ground truth labels on individual data points.
• To investigate the application of weak supervision directly on time series data for ECG analysis.

Main Methods:

• Utilized human-designed heuristics as weak supervision sources.
• Defined weak supervision directly on time series ECG data.
• Inferred high-quality probabilistic label estimates for heartbeats using a limited number of heuristics.

Main Results:

• Successfully generated probabilistic labels for over 100,000 heartbeats with minimal human effort.
• Achieved competitive classifier performance using the inferred labels on held-out test data.
• Demonstrated the efficacy of using as few as six intuitive time series heuristics.

Conclusions:

• Weak supervision offers a viable and efficient alternative to manual annotation for ECG analysis.
• This approach significantly reduces the cost and effort associated with creating labeled datasets for heart disease diagnosis.
• The proposed method enables the training of effective machine learning models for abnormal heartbeat detection.