C Cao1, N McIntosh, I S Kohane, K Wang. 1. Informatics Program, Children's Hospital, Harvard Medical School, 300 Longwood Avenue, Boston, MA 02115, USA.
Abstract
BACKGROUND: Artifacts in clinical intensive care monitoring lead to false alarms and complicate later data analysis. Artifacts must be identified and processed to obtain clear information. In this paper, we present a method for detecting artifacts in PCO2 and PO2 physiological monitoring data from preterm infants. PATIENTS AND DATA: Monitored PO2 and PCO2 data (1 value per minute) from 10 preterm infants requiring intensive care were used for these experiments. A domain expert was used to review and confirm the detected artifact. METHODS: Three different classes of artifact detectors (i.e., limit-based detectors, deviation-based detectors, and correlation-based detectors) were designed and used. Each identified artifacts from a different perspective. Integrating the individual detectors, we developed a parametric artifact detector, called ArtiDetect. By an exhaustive search in the space of ArtiDetect instances, we successfully discovered an optimal instance, denoted as ArtiDetector. RESULTS: The sensitivity and specificity of ArtiDetector for PO2 artifacts is 95.0% (SD = 4.5%) and 94.2% (SD = 4.5%), respectively. The sensitivity and specificity of ArtiDetector for PCO2 artifacts is 97.2% (SD = 3.6%) and 94.1% (SD = 4.2%), respectively. Moreover, 97.0% and 98.0% of the artifactual episodes in the PO2 and PCO2 channels respectively are confirmed by ArtiDetector. CONCLUSIONS: Based on the judgement of the expert, our detection method detects most PO2 and PCO2 artifacts and artifactual episodes in the 10 randomly selected preterm infants. The method makes little use of domain knowledge, and can be easily extended to detect artifacts in other monitoring channels.
BACKGROUND: Artifacts in clinical intensive care monitoring lead to false alarms and complicate later data analysis. Artifacts must be identified and processed to obtain clear information. In this paper, we present a method for detecting artifacts in PCO2 and PO2 physiological monitoring data from preterm infants. PATIENTS AND DATA: Monitored PO2 and PCO2 data (1 value per minute) from 10 preterm infants requiring intensive care were used for these experiments. A domain expert was used to review and confirm the detected artifact. METHODS: Three different classes of artifact detectors (i.e., limit-based detectors, deviation-based detectors, and correlation-based detectors) were designed and used. Each identified artifacts from a different perspective. Integrating the individual detectors, we developed a parametric artifact detector, called ArtiDetect. By an exhaustive search in the space of ArtiDetect instances, we successfully discovered an optimal instance, denoted as ArtiDetector. RESULTS: The sensitivity and specificity of ArtiDetector for PO2 artifacts is 95.0% (SD = 4.5%) and 94.2% (SD = 4.5%), respectively. The sensitivity and specificity of ArtiDetector for PCO2 artifacts is 97.2% (SD = 3.6%) and 94.1% (SD = 4.2%), respectively. Moreover, 97.0% and 98.0% of the artifactual episodes in the PO2 and PCO2 channels respectively are confirmed by ArtiDetector. CONCLUSIONS: Based on the judgement of the expert, our detection method detects most PO2 and PCO2 artifacts and artifactual episodes in the 10 randomly selected preterm infants. The method makes little use of domain knowledge, and can be easily extended to detect artifacts in other monitoring channels.
Authors: Marion Verduijn; Niels Peek; Nicolette F de Keizer; Erik-Jan van Lieshout; Anne-Cornelie J M de Pont; Marcus J Schultz; Evert de Jonge; Bas A J M de Mol Journal: J Am Med Inform Assoc Date: 2007-12-20 Impact factor: 4.497