Reduction of false cardiac arrhythmia alarms through the use of machine learning techniques

Due to the so-called “crying wolf” effect, frequent false cardiac arrhythmia alarms have been shown to diminish staff attentiveness and thus reduce the quality of care patients receive in the ICU. The PhysioNet/Computing in Cardiology 2015 Challenge seeks to improve patient care by decreasing the number of these false cardiac arrhythmia alarms. Using a training set of 750 multi-parameter recordings organized by type of arrhythmia alarm, we developed a decision tree for each arrhythmia category. We derived the features utilized in the decision tree from the arterial blood pressure (ABP) waveform and the photoplethysmogram (PPG). For Phase 1 of the challenge, our score for the realtime test set = 57.64 and retrospective test set = 61.15, resulting in an overall score of 59.39. For Phase 11, our score for the real-time test set = 65.19 and retrospective test set = 72.19. In conclusion, decision trees have been shown to generate reasonable results in reducing false cardiac arrhythmia alarms; future work will involve more sophisticated machine learning algorithms to improve performance.