A data mining approach to reduce the false alarm rate of patient monitors

Patient monitors in intensive care units trigger alarms if the state of the patient deteriorates or if there is a technical problem, e.g. loose sensors. Monitoring systems have a high sensitivity in order to detect relevant changes in the patient state. However, multiple studies revealed a high rate of either false or clinically not relevant alarms. It was found that the high rate of false alarms has a negative impact on both patients and staff. In this study we apply data mining methods to reduce the false alarm rate of monitoring systems. We follow a multi-parameter approach where multiple signals of a monitoring system are used to classify given alarm situations. In particular we focus on five alarm types and let our system decide whether the triggered alarm is clinically relevant or can be considered as a false alarm. Several classification algorithms (Naive Bayes, Decision Trees, SVM, kNN and Multi-Layer Perceptron) were evaluated. For training and test sets a subset of the freely available MIMIC II database was used. Alarm-specific classification accuracy was between 78.56% and 98.84%. Suppression rates for false alarms were between 75.24% and 99.23%. Classification results strongly depend on available training data, which is still limited in the intensive care domain. However, this study shows that data mining methods are useful and applicable for alarm classification.