The Science Journal of the American Association for Respiratory Care

2009 OPEN FORUM Abstracts

ARTIFICIAL NEURAL NETWORK TO IMPROVE CAPNOMETRY DURING SEDATION

Joseph Orr, Ken B. Johnson, Lara Brewer; Anesthesiology, University of Utah, Salt Lake City, UT

Background: Capnometers are used during sedation in non-intubated patients to monitor respiratory rate (RR) and end-tidal CO2 (etCO2). Since capnometers detect RR based on changes in CO2 concentration rather than gas movement, it is common for spurious changes in CO2 to be mis-identified as breaths. This leads to false alarms and missed detection of periods of apnea. Cardiogenic oscillations, attempts by the patient to speak and unsuccessful breath attempts during obstruction are common causes of changes in CO2 that can be falsely detected as breaths. An artificial neural network (ANN) is a type of computer algorithm that can be trained to recognize patterns and events. We trained an artificial neural network to estimate the tidal volume (TV) of each breath based on the shape of its capnogram. If the trained ANN identified the capnogram as having a TV of at least 200 mL it was categorized as a valid breath; otherwise it was rejected as artifact. We evaluated the accuracy of this combination ANN algorithm and a simple threshold algorithm compared to a reference RR, as measured using a pneumotach. Methods: 24 volunteers were fitted with a tight-fitting, sealed mask connected to a combination flow and CO2 sensor. Combinations of propofol and remifentanil concentrations delivered as target controlled infusions were administered to each volunteer to simulate various sedation conditions. The threshold algorithm defined a breath as any deviation of the CO2 signal above 10 mm Hg and below 5 mm Hg. Using data from 12 volunteers, the ANN was trained to estimate the TV associated with each breath. The detected TA breath was considered valid only for breaths identified by the ANN as having a TV larger than the airway dead space (200 mL). Both algorithms were compared to the reference resp. rate using data from the remaining 12 volunteers. Results: Using the threshold method, the average detected RR was 3 breaths per minute (bpm) higher than the reference rate. The standard deviation of the difference (SD) was 4.7 bpm. When the ANN modification was applied, the average RR was 0.07 bpm lower than the reference with SD of 2.5 bpm. The plot shows RR data of both methods for a typical volunteer. Conclusions: An ANN can be applied as a intelligent filter to provide breath-by-breath analysis of the capnogram and improve alarm reliability during sedation. Sponsored Research - Philips/Respironics

You are here: RCJournal.com » Past OPEN FORUM Abstracts » 2009 Abstracts » ARTIFICIAL NEURAL NETWORK TO IMPROVE CAPNOMETRY DURING SEDATION