Non-invasive detection of intracranial hypertension using random forests

In various pathologies, such as hemorrhagic stroke or traumatic brain injury, intracranial pressure (ICP) can rise to the point of causing neurological damage and should consequently be monitored. The invasiveness of current ICP monitoring procedures, however, limits the patient pool. Having a non-invasive screening tool for detecting intracranial hypertension would therefore be of high value. In this work, we developed a binary classifier to determine if ICP is above the clinical cutoff value of 20 mmHg. Arterial blood pressure (ABP) and cerebral blood flow velocity (CBFV) recordings serve as the only inputs to the classifier. We identified eighteen ABP and CBFV features reported to correlate with ICP. Given around 32 h of ABP, CBFV and invasive ICP recordings from 36 traumatic brain injury patients, we trained different binary classifiers via leave-one-patient-out cross validation. The random forest classifier resulted in the most stable and accurate prediction, yielding a sensitivity of 69.1% and specificity of 78.3%. These encouraging results and the ease by which features can be added to the framework, suggest possible extensions to include, for example, features derived from venous pressure or near-infrared spectroscopy recordings.