Reduction of False Alarms in Intensive Care Unit using Multi-feature Fusion Method

Abstract

In thist study, we proposed a multi-feature fusion method for accurately classifying the true or false alarms for five life-threatening arrhythmias: asystole, extreme bradycardia (EB), extreme tachycardia (ET), ventricular flutter/fibrillation (VF) and ventricular tachycardia (VT). The proposed method consisted of four steps: 1) signal pre-processing, 2) detection validation and feature calculation, 3) real-time determining and 4) retrospectively determining. Up to four signal channels, that is, two ECGs, one arterial blood pressure (ABP) and/or one photoplethysmogram (PPG) signals were analyzed to obtain the classification features. Multi-features from those signals were merged to reduce the maximum number of false alarms, while avoiding the suppression of true alarms. Two events existed: Event 1 for “real-time” and Event 2 for “retrospectively”. The optimal results of true positive ratio (TPR) for the training set were: 100% for asystole, EB, ET and VF types and 94% for VT type. The corresponding results of true negative ratio (TNR) were 93%, 81%, 78%, 85% and 50% respectively, resulting in the corresponding scores of 96.50, 90.70, 88.89, 92.31 and 64.90, as well as with score 80.57 for Event 1 and 79.12 for Event 2. The results of the our open source entries for the Challenge obtained the optimal scores of 88.73 for asystole, 77.78 for EB, 89.92 for ET, 67.74 for VF and 61.04 for VT types, with the final scores 71.68 for Event 1 and 75.91 for Event 2.