Temporal characterization of patient state with applications to prediction of tachycardia in anesthesia via induction of inhaled desflurane

Abstract

It has always been assumed that using clinically measurable parameters is the most efficient way to characterize patient state. By adding additional sensors, monitors, and derived statistics (e.g. mean arterial blood pressure from diastolic and systolic), it was hoped that more information could be garnered about patient state. This thesis challenges the assumption that providing the physician with a full set of clinically measurable parameters is the most efficient way to characterize patient state. The thesis presents a novel way to consider patient state by utilizing reduced dimensionality and by estimating noise. It then explores an application, namely prediction of tachycardia, which often occurs at the onset of induction of inhaled desflurane. One unexpected initial finding was that all 46 patients exhibited tachycardia or hypertension within the first hour of the operation. Three models for predicting tachycardia episodes are proposed, including one model based on use of Blind Noise Adjusted Principal Component Analysis1 (using Iterative Order and Noise Estimate (ION)2 and Principal Component Analysis (PCA)3). Without ION, PCA-based methods alone yielded only 2 useful degrees of freedom, with the rest being relegated to noise. The ION PCA-based method allows one to capture with 5 principal components the information contained in 31 fundamental and derived patient variables, while at the same time reducing the effects of noise. Furthermore, the five discovered significant principal components representing patient state were characterized quantitatively and their physiologic correlates are hypothesized qualitatively. Examination of the 31 original patient parameters in the ION PCA model that predicts tachycardia revealed the relative importance of the original patient parameters to the tachycardia problem. The receiver operating characteristic (ROC) curve for the ION PCA-based predictor suggested a 70% detection rate with 3% false alarms when predicting tachycardia two minutes and twenty seconds into the future. While the patient state characterization method was used for tachycardia prediction, it is potentially useful in myriad medical domains involving multivariate analysis.