Active learning and margin strategies for arrhythmia classification in implantable devices

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence in Healthcare
• Medical Informatics

Background:

• Massive datasets of cardiac arrhythmic episodes are available from Implantable Cardioverter Defibrillators (ICDs).
• Extracting electrophysiological knowledge requires accurate episode labeling, which is currently a manual, time-consuming, and expensive process performed by expert cardiologists.

Purpose of the Study:

• To propose and evaluate Active Learning (AL) strategies for streamlining the manual labeling of cardiac arrhythmic episodes.
• To develop classification models that reduce the human labeling burden in large-scale ICD datasets.

Main Methods:

• Adapted four large-margin-based Active Learning (AL) strategies to a pre-existing classification methodology.
• Benchmarked AL strategies on 3- and 8-class arrhythmia classification problems using a national ICD data repository of 9908 episodes.
• Evaluated the impact of episode-patient diversity on classification model performance.

Main Results:

• Achieved gold standard performance, comparable to using all available episodes, by utilizing only approximately 20% of episodes from 60% of patients for the 3-class model.
• For the 8-class model, expert-level performance was reached using approximately 50% of episodes from 85% of patients.
• Demonstrated the effectiveness of AL in reducing the data required for training accurate arrhythmia classifiers.

Conclusions:

• Active Learning (AL) techniques offer a significant advantage in designing classification models for cardiac arrhythmias.
• AL effectively streamlines the human labeling process for extensive Implantable Cardioverter Defibrillator (ICD) datasets.
• This approach promises to accelerate electrophysiological knowledge extraction from large clinical datasets.