Prediction of Life-Threatening Heart Arrhythmias Using Obstructive Sleep Apnoea Characteristics

Abstract

False alarms ratios of up to 86% in Intensive Care Units (ICU) decrease quality of care, impacting both clinical staff and patients through slowing off response time and noise tribulation. We present a novel algorithm to predict heart arrhythmias in ICUs. We focus on five life-threatening arrhythmias: Asystole, Extreme Bradycardia, Extreme Tachycardia, Ventricular Tachycardia, and Ventricular Fibrillation. The algorithm is based on novel features using only 12 seconds of one ECG channel to predict the arrhythmias. Our new feature sets include different SQI and physiological features and the features used in obstructive sleep apnoea detection. We also proposed a new morphological characteristic to count the abnormal patterns which are common in Ventricular Fibrillation arrhythmia. We evaluate our proposed features by ranking them using 19 different feature selector algorithms. Finally, NN classifiers were trained separately for every type of arrhythmia. Applying the algorithm on 750 data of bedside monitors, we achieved the score of 80.2% & 80.6% respectively in the real-time analysis for 12 and 16 seconds of one ECG channel.