Representation Learning Approaches to Detect False Arrhythmia Alarms from ECG Dynamics

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiovascular Signal Processing

Background:

• Intensive care units (ICUs) experience high rates of false arrhythmia alarms, leading to alarm fatigue and delayed critical care.
• Ventricular tachycardia (v-tach) alarms are frequent, and distinguishing true events from false positives is crucial for patient safety.

Purpose of the Study:

• To develop and evaluate machine learning models for accurate detection of false v-tach alarms using electrocardiogram (ECG) waveform data.
• To improve the reliability of arrhythmia detection systems in ICUs and mitigate the impact of alarm fatigue.

Main Methods:

• Utilized a Supervised Denoising Autoencoder (SDAE) model to learn a low-dimensional representation of ECG dynamics.
• Employed a combined reconstruction and classification loss function for training the SDAE.
• Applied Fast Fourier Transform (FFT) to ECG data for beat-by-beat analysis.
• Evaluated the SDAE model on the PhysioNet Challenge 2015 dataset, comprising over 500 ECG records with v-tach alarms.

Main Results:

• The SDAE model, when applied to FFT-transformed ECG data on a beat-by-beat basis, demonstrated superior performance in classifying false v-tach alarms compared to baseline methods.
• Exploiting physiological structure through beat-by-beat frequency distribution from multiple cardiac cycles significantly improved classification accuracy.
• The proposed method showed improvement over previous entries in the 2015 PhysioNet Challenge.

Conclusions:

• The SDAE model offers a promising approach for reducing false v-tach alarms in clinical settings.
• Beat-by-beat analysis of ECG frequency distributions is vital for accurate false alarm detection.
• Implementing advanced machine learning techniques can enhance the efficiency and effectiveness of ICU monitoring systems.