Machine Learning Techniques for the Detection of Shockable Rhythms in Automated External Defibrillators

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Early recognition of ventricular fibrillation (VF) and prompt electrical therapy are critical for survival in out-of-hospital cardiac arrest (OHCA) patients.
• Automated external defibrillator (AED) algorithms are typically validated using public electrocardiogram (ECG) databases, which may not accurately represent ECGs during actual OHCA events.

Purpose of the Study:

• To evaluate the performance of VF-detection algorithms using both OHCA patient data and public Holter recordings.
• To compare the effectiveness of machine learning algorithms in identifying VF from different data sources and segment lengths.

Main Methods:

• Analyzed 4-s and 8-s ECG segments from OHCA patients and public Holter recordings.
• Computed 30 features per segment and utilized machine learning algorithms with built-in feature selection.
• Assessed algorithm performance using patient-wise bootstrap, measuring sensitivity (Se), specificity (Sp), and balanced error rate (BER).

Main Results:

• VF detection performance was significantly better for public data (Se: 96.6%, Sp: 98.8%, BER: 2.2%) compared to OHCA data (Se: 94.7%, Sp: 96.5%, BER: 4.4%).
• OHCA data required twice as many features (6 vs 3) for accurate detection compared to public databases.
• No significant performance differences were observed across different ECG segment lengths (4-s vs 8-s).

Conclusions:

• VF detection is more challenging using OHCA data than public Holter data.
• Accurate VF detection is feasible even with short ECG segments (as brief as 4 seconds).
• Algorithm validation should consider data representative of actual OHCA events for improved real-world performance.