A PCA based feature reduction in intracranial hypertension analysis

Abstract

Traumatic brain injury (TBI) and its complications, including intracranial hypertension, are one of the leading causes of mortality. Many proposed algorithms have attempted to overcome the invasiveness of intracranial pressure monitoring with limited clinical applications. In medical practices, changes of intracranial hypertension are perceived manually, by clinical experts, via surgical placement of intraventricular catheters in the patient's skull. However, the method is invasive, time consuming and has poor reproducibility. In this paper, a principal component analysis has been conducted to perform non-correlated feature selection. An analysis has been performed on 31 TBI patients. The result of the analysis illustrates that two components can be extracted from routinely collected physiological signals of TBI patients. The finding is evaluated using Kaiser's Criterion, Scree test and parallel analysis. The predictivity of the components has been tested and the results show a promising accuracy with a mean absolute error of 0.025.