Reliability of clinical alarm detection in intensive care units

In hospital environments advanced medical devices are vital for both monitoring and therapy. Many have alarms, especially in intensive care areas. To ensure that important and unwanted clinical events are not missed, there is a tendency for devices to react to noise and artefact in the physiological waveforms, with many resulting false alarms. PhysioNet along with Computing in Cardiology have made available clinical alarm data, to allow improved algorithms for alarm detection to be developed. We present our results. Our analysis steps included: high pass filtering to remove baseline instability, scaling to normalise waveform amplitudes, detection of noisy and flat waveforms, diferentiation to accentuate sharp waveform edges, beat detection, timing between beats preceding alarm onset, and detection of alarm conditions. When the waveforms were assessed as noisy they were labelled as false alarms. When noise-free and alarm conditions were met they were labelled as true alarms. The original PhysioNet analysis algorithm analysed arterial blood pressure (ABP) and photoplethysmograph (PPG) waveform data, resulting in true alarm detection sensitivity of 89% and 88%, and specificity of 38% and 38%, for the training and test data sets respectively, indicating a similar range of data in both sets. We investigated the use of ECG data alone with the training data, and this resulted in overall gross sensitivity and specificity for the first ECG channel of 89% and 68%, and for the second 87% and 68% respectively, indicating similarity in the two ECG channels. When BP and PPG were analysed following detection of noise in the ECG the results were 92% and 56%, and 90% and 54% respectively. We have shown that analysis of the ECG alone can obtain average sensitivity of 88%, with little diference in results between two simultaneous ECG channels. When the arterial blood pressure and peripheral pulse were also analysed this additional physiological data improved sensitivity by 3% points, but decreased specificity by 13% points in the training set, and 4% and 9% respectively in the test set.