Abstract
Real-time monitoring of intracranial pressure (ICP) is a routine part of neurocritical care in the management of brain injury. While mainly used to detect episodes of intracranial hypertension, the ICP signal is also indicative of the volume-pressure relationship within the cerebrospinal system, often referred to as intracranial compliance (ICC). Several ICP signal descriptors have been proposed in the literature as surrogates of ICC, but the possibilities of combining these are still unexplored. In the present study, a rapid ICC assessment consisting of a 30-degree postural shift was performed on a cohort of 54 brain-injured patients. 73 ICP signal features were calculated over the 20 minutes prior to the ICC test. After a selection step, different combinations of these features were provided as inputs to classification models. The goal was to predict the level of induced ICP elevation, which was categorized into three classes: less than 7 mmHg ("good ICC"), between 7 and 10 mmHg ("medium ICC"), and more than 10 mmHg ("poor ICC"). A logistic regression model fed with a combination of 5 ICP signal features discriminated the "poor ICC" class with an area under the receiving operator curve (AUROC) of 0.80 (95%-CI: [0.73-0.87]). The overall one-versus-one classification task was achieved with an averaged AUROC of 0.72 (95%-CI: [0.61-0.83]). Adding more features to the input set and/or using nonlinear machine learning algorithms did not significantly improve classification performance. This study highlights the potential value of analyzing the ICP signal independently to extract information about ICC status. At the patient's bedside, such univariate signal analysis could be implemented without dependence on a specific setup.