Real-time machine learning-based intensive care unit alarm classification without prior knowledge of the underlying...

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Cardiology

Background:

• Intensive care units (ICUs) require precise and instantaneous identification of life-threatening arrhythmias.
• Existing bedside monitors have limitations in accuracy and response time for arrhythmia detection.

Purpose of the Study:

• To develop a standalone heart rhythm alerting system for ICUs.
• To improve the precision and reduce the latency of arrhythmia detection compared to current systems.

Main Methods:

• Utilized the PhysioNet 2015 Challenge dataset with re-annotated cardiac rhythm records.
• Developed an improved R-wave detection algorithm considering all electrocardiographic (ECG) leads and incorporating low-pass filtering to mitigate pacing spike interference.
• Employed a random forest classifier with 10-time five-fold cross-validation.

Main Results:

• Achieved a macro-average sensitivity of 81.54% for arrhythmia detection.
• Extracted arrhythmia-specific and signal quality features for analysis.
• The improved R-wave detection enhanced ventricular fibrillation (VF) detection accuracy.

Conclusions:

• The developed system demonstrated higher positive predictive values for asystole, extreme bradycardia, ventricular tachycardia (VT), and VF compared to PhysioNet 2015 competition bedside monitors.
• The system provides instantaneous arrhythmia alerts, detecting events up to 4 seconds earlier.