Reducing false arrhythmia alarms using robust interval estimation and machine learning

Abstract

Reducing false arrhythmia alarms in the intensive care unit is the objective of the PhysioNet/Computing in Cardiology Challenge 2015. In this paper, an approach is presented that analyzes multimodal cardiac signals in terms of their beat-to-beat intervals as well as their average rhythmicity. Based on this analysis, several features in time and frequency domain are extracted and used for subsequent machine learning.Results show that alarm-specific strategies proved optimal for different types of arrhythmia and that obtained scores varied: While the score for reducing false ventricular tachycardia alarms was 68.91, false extreme tachycardia alarms could be suppressed with perfect accuracy. Overall, a top score of 75.55 / 75.18 could be achieved for real-time / retrospective false alarm reduction.

False arrhythmia alarms reduction in the intensive care unit: a multimodal approach

“…The classification model was selected as the Random Forest for all the arrhythmias. In the solution of Hoog Antink and Leonhardt (2015) different machine learning techniques were used depending on the arrhythmia. In their approach, very different strategies depending on arrhythmia provided an optimal solution.…”

“…Single-tree approaches have been presented previously with good results for integrating multiple signals for artifact detection in neonatal ICU (Tsien et al 2000) and patient specific alarming models (Zhang and Szolovits 2008). A binary classification tree was used also in the Challenge entry of Hoog Antink and Leonhardt (2015).…”

“…In addition, machine learning approaches were presented by Hoog Antink and Leonhardt (2015) and Kalidas and Tamil (2015).…”

Deep Neuroevolution: Training Deep Neural Networks for False Alarm Detection in Intensive Care Units

“…The participants used elementary algebra, descriptive statistics, [18]. Binary Classification Trees (BCTs) [19], [20], Discriminant Analysis Classifiers (DACs) [19], Random Forest [21], Support Vector Machine (SVM) [19], [22], and Fuzzy Logic [10] are used in different submissions as well. The highest score (81.39) in the event one is achieved [10] by applying Fuzzy Logic along with Elementary Algebra and Descriptive Statistics.…”

“…The highest score (81.39) in the event one is achieved [10] by applying Fuzzy Logic along with Elementary Algebra and Descriptive Statistics. [19] applied SVM, DACs, and BCTs and ranked fifth in this challenge with a score of 75.55.…”