False arrhythmia alarm reduction in the intensive care unit using data fusion and machine learning

With aim of reducing the incidence of false critical arrhythmia alarms in intensive care units, a novel data fusion and machine learning algorithm is presented in this article. The 2015 PhysioNet/Computing in Cardiology Challenge database was used in this present algorithm, with each grouped as an asystole (AS), extreme bradycardia (EB), extreme tachycardia (ET), ventricular tachycardia (VT) or ventricular flutter/fibrillation (VF) arrhythmia alarm. A 10-second segment before the onset of the alarm was truncated from available signals, namely electrocardiogram (ECG), arterial blood pressure (ABP), and/or photoplethysmogram (PPG). By first assessing signal quality of available signals, a robust estimation of beat-to-beat intervals could then be derived. Features in heart rate variability (HRV) analysis and ECG parameters such as temporal statistical parameters, spectral analysis results, wavelet transformation coefficients, and complexity measurement etc were extracted and formed a vector. After feature selection through genetic algorithm (GA), a support vector machine (SVM) model was applied to conduct the classification of alarms for the specific arrhythmia type. The overall true positive rate (TPR) of classification algorithm is 93%, with the true negative rate (TNR) 94%. According to the method of performance evaluation in the 2015 Challenge, this algorithm achieved a gross score of 84.4.