Abstract
Background: Computerised cotside monitoring has become routine in some neonatal intensive care units. Though a trend display of the vital signs of a neonate has practical benefits, the data is often significantly corrupted by artifact and is difficult to interpret by inexperienced staff.
Aims: To develop a system to reliably identify certain artifacts within monitoring data and to provide a measure of ‘abnormality' in the trended physiology of a neonate.
Methods: The study used 10 data channels sampled routinely in our intensive care unit at 1 second intervals, for 6 premature infants. The artifacts modelled were transcutaneous probe (TCP) recalibration, probe dropouts, and the blood pressure spike caused by taking a blood gas sample. A (factorial) hidden markov model (HMM) was implemented to analyse the data, a technique widely used in speech recognition and other machine learning problems. This builds up a probabilistic description of the way a baby's vital signs vary when it is stable or when certain defined events are happening. From this the probability of a particular event happening can be calculated from the observed data. In addition, the system was constructed to flag areas of the data where the signal varied in a way not due to any artifact but also not characteristic of the baby in its stable state.
Results: Probe dropouts are in general easy to identify. The study concentrated on transcutaneous probe recalibration and blood gas sample artifact, both of which frequently occur in monitoring data. Detection performance was evaluated with receiver operating characteristics (ROC) curves. TCP recalibration detection had an area under the ROC curve of 95.8% and an error rate of 3.77%. Blood gas sample detection had an area under the ROC curve of 96.0% and an error rate of 1.54%. Abnormal sections of data can be flagged once the system has been manually given a time where the baby is stable as a reference point.
Conclusion: Machine learning can be reliably used to identify artifacts in NICU monitoring data.
Log in or create a free account to read this content
Gain free access to this article, as well as selected content from this journal and more on nature.com
or
Author information
Authors and Affiliations
Rights and permissions
About this article
Cite this article
Quinn, J., Williams, C. & McIntosh, N. 219 Identification Of Artifact And Abnormal Variation In Nicu Monitoring Data. Pediatr Res 56, 501 (2004). https://doi.org/10.1203/00006450-200409000-00242
Issue date:
DOI: https://doi.org/10.1203/00006450-200409000-00242
