Automatic error detection in the clinical measurement of gastric impedance spectra

• Published in: 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology Vol. 2010; pp. 1000 - 1003
• Main Authors: Godinez-Garcia, M M, Yanez-Suarez, O, Sacristan-Rock, E
• Format: Conference Proceeding Journal Article
• Language: English
• Published: United States IEEE 01.01.2010
• Subjects: Algorithms; Artificial Intelligence; Diagnosis, Computer-Assisted - methods; Dielectric Spectroscopy - methods; Humans; Impedance; Impedance measurement; Measurement uncertainty; Monitoring; Pattern Recognition, Automated - methods; Reproducibility of Results; Sensitivity and Specificity; Spectroscopy; Stomach - physiology; Support vector machines
• ISBN: 1424441234; 9781424441235
• ISSN: 1094-687X; 1557-170X
• DOI: 10.1109/IEMBS.2010.5627795

Gastric impedance spectroscopy has been proposed as a method of monitoring mucosal injury due to hypoperfusion and ischemia in the critically ill. During validation tests for this procedure, it was found that 60% of the measurements had errors by factors inherent to the clinical setting, indicating that some kind of automatic error detection should be incorporated to potentially avoid the loss of measurements. This paper presents an algorithm developed to detect errors due to bad connection, bad location or bad contact of the electrode probe. A labeled database with 20,908 sets of 92 spectral measurements each, obtained from critically ill patients was used as training/testing data. To reduce the dimensionality, the database was resized by dividing the spectral range into four bands, and then by computing mean and standard deviation in magnitude, phase, resistance and reactance for each band and measurement. Initial exploration into the data space was performed by k-means clustering, establishing the number of classes. Sequential Forward Selection was performed to determine best features from the reduced data set. Finally, Support Vector Machine classifiers were designed in a one-vs-rest hierarchical scheme to classify the quality of the spectra. Each classifier gave a hit rate greater than 95% and an area under the relative operating characteristic curve of 0.99. In a validation run with cardiac surgery and intensive care unit patient spectra, the error rates were 2.3% and 8.4% respectively.