Abstract 109: Reducing False Alarms in Intensive Care Units: A Generalizable Approach Based on Convolutional Neural Networks
Background: At intensive care units (ICUs), multiple physiological signals are continuously monitored so as to detect the onset of critical health events and generate alarms. In current systems, most such alarms happen to be false and correspond to innocuous events such as muscle movement or electrical disconnection. High rate of such false alarms (FAs) undesirably reduces staff responsiveness and annoys patients. Indeed, reducing the harm associated with clinical alarm systems is listed as a national patient safety goal of 2019.
Objective: We seek to reduce FA rate in ICUs, while preserving true alarms, by collating information from interdependent vital signals including electrocardiogram, photoplethysmogram and arterial blood pressure signals. In this regard, we propose an approach that generalizes to various arrhythmic conditions, and hence simplifies design.
Methods: We propose FARIC, a method of False Alarm Reduction in ICU based on Convolutional neural network (CNN). In particular, we use CNNs as building blocks (see Figure) to glean the heartbeat location and morphology information from multiple physiological signals by suitably modifying the training data. Here, as the features are algorithmically learnt from the example waveforms, we eliminate the need for handcrafted signal-specific and arrhythmia-specific features, thus making our method generalizable.
Results: We evaluated FARIC on the PhysioNet/CinC 2015 Challenge dataset, consisting of 1250 patient records collected at four hospitals in the USA and Europe. Our model was trained on the 750 public records, and tested on all records including 500 hidden ones. The Challenge at hand compares the efficacy of competing techniques in reducing ICU FA rate using a standardized score that penalizes false negatives heavily. Alongside an overall score, individual score was given for each of the following arrhythmias: Asystole, Extreme Bradycardia, Extreme Tachycardia, Ventricular Tachycardia and Ventricular Fibrillation/Flutter. Encouragingly, FARIC either matched or exceeded each of hitherto leading arrhythmia-specific scores and added more than 4 percent points to the leading overall score (see Table).
Conclusion: Our method promises to significantly reduce the FA rate in ICUs, and can be extended to additional arrhythmic conditions.