Heartbeat fusion algorithm to reduce false alarms for arrhythmias

There is a need for patient monitoring algorithms to reduce alarm fatigue by rejecting clinically irrelevant alarms. We developed an algorithm using multimodal physiological waveforms (electrocardiogram, blood pressure, photoplethysmogram) and noise classifiers to improve arrhythmia detection by reducing the incidence of false alarms while maintaining a high true alarm rate as part of the Physionet Challenge 2015. Combining information from multiple physiological signals our algorithm was able to discard 362 of 456 false alarms (true negative rate [TNR] of 79%), while correctly classifying 268 of the 294 true alarms (true positive rate [TPR] of 91%) on the training set, a score of 73.8. When applied to the test set which had 343 false alarms and 157 true alarms, we achieved a TNR of 81%, TPR of 86% and score of 70.2. Our results support the concept that false alarms can be reduced in the intensive care unit by removing noise segments in signals and combining information from multiple physiological signals.