Use of the median filter in haemodynamic monitoring: Dissertation

Vigilant observation of the patient forms the basis of intensive care. Automatic monitoring equipment with limit alarms makes a vital contribution to this, but the specificity of these alarms is poor, as about 90 % of alarms are false. If the temporal characteristics of the changes causing true and false alarms differ, it is possible to separate them using median filtering, and to there by increase the reliability of limit alarms and the quality of haemodynamic data for computer processing. We have studied the properties of haemodynamic data in post operative open heart surgery patients. The studied variables were the heart rate, the systolic, mean and diastolic values of the systemic arterial pressure and the pulmonary arterial pressure, and the mean value of the central venous pressure. The incidence of outliers and the distribution of this data was analyzed and the duration of deviations of this data outside commonly applied alarm limits was studied. Instituting a delay before an alarm turned out to be the main method for reducing the alarm rate in current monitors. Changes of alarm limit values within a clinically rational value range did not produce significant changes in the alarm rate. The standard median filter and the vector median filter were applied to the preprocessing of haemodynamic data. Median filtering increased the specificity of alarms and decreased the alarm rate. Multivariate vector median filtering was able to decrease the alarm rate more effectively than one dimensional median filtering. Vectors derived from the systolic, mean and diastolic values of blood pressure have been used. A novel dual limit alarm for haemodynamic monitoring employed two standard median filters, followed by alarm limits. Its design ensured an alarm response to major deviations with a very short delay. The proportion of true alarms increased from 12 % to 49 %, while none were missed in the clinical evaluation. The median filter based preprocessing methods have been applied to providing a robust fusion of data for highly specific knowledge based alarm systems. During the clinical evaluation, a correct symbolic interpretation was provided for 99 % of the values and 93 % of the trends. Compared to other proposed methods, median filter based methods can be computationally significantly less demanding. The results support the hypothesis that the temporal characteristics related to true and false alarms are different.