Decreasing the false alarm rate of arrhythmias in intensive care using a machine learning approach

We present a novel algorithm for classifying true and false alarms of five life-threatening arrhythmias in intensive care. This algorithm was entered in the PhysioNet/Computing in Cardiology Challenge 2015 Reducing False Arrhythmia Alarms in the ICU. The algorithm performs a binary classification of the alarms for a specified arrhythmia type by combining signal quality information and physiological features from multiple sources, such as electrocardiogram (ECG), photoplethysmogram (PPG), and arterial blood pressure (ABP). Signals were selected for feature computation by first assessing the quality for available signals. Random Forest classifiers were trained separately for every type of arrhythmia with arrhythmia-specific features. Hence, the complete algorithm leverages five different predictive models. Classification sensitivities of true alarms 75-99 % (average 93 %) on the training set with cross-validation and 22-100 %(average 90 %) on the unrevealed test set. Classification specificities on the training and test set were 76-94% (average 80%) and 75-100% (average 82%), respectively. The best performance was for extreme bradycardia, whereas the poorest results were for ventricular arrhythmias. The results are for the real-time category when only information prior to the alarm is considered. The final challenge score was 75.54.