next up previous contents
Next: 2.3.5 Signal Quality Up: 2.3.4 Alarms and Inops Previous: 2.3.4 Alarms and Inops   Contents


2.3.4.1 Annotated alarms

Since no large annotated dataset of alarms is publicly available, a set of gold standard alarms to support the development and testing of a false alarm suppression algorithm was generated from the above alarms. Initially we have concentrated on life-threatening arrhythmia alarms, namely; Asystole, Extreme Bradycardia, Extreme Tachycardia, Ventricular Tachycardia and Ventricular Fibrillation. In order to assemble such a database we first searched for patient records in the MIMIC II database that met the following two criteria:

  1. Record contains at least one of the 5 above listed critical alarm categories.
  2. At least one of the alarms is associated with simultaneous ABP and ECG waveforms.

Our initial search yielded 496 patient records with a total of 45,370 hours of simultaneous ECG & ABP waveforms containing 8,636 alarms. Each alarm was manually reviewed by two independent experts, and discrepancies were adjudicated by a third expert.

Alarm repetitions referring to the same event, were removed. Furthermore, all 48 patients that possessed active intra-aortic balloon pumps (IABP) were removed, since their ABP waveforms do not appear as ``physiologically normal''. The final set comprises 448 patients with 5,386 alarms with simultaneous ABP & ECG waveforms. These annotations have been posted on PhysioNet with the file extension .alM.

Full details of how these alarms were annotated is available in Aboukhalil et al [11], together with an evaluation of their statistics.


next up previous contents
Next: 2.3.5 Signal Quality Up: 2.3.4 Alarms and Inops Previous: 2.3.4 Alarms and Inops   Contents
djscott 2010-08-24