Noninvasive intracranial hypertension detection using machine-learning of cerebral blood flow velocity waveforms

The monitoring of intracranial pressure (ICP) is crucial in the clinical management of various cerebral diseases and injuries, including head trauma, hydrocephalus, intracranial tumors, and cerebral edema. It could also play a broader role, for example, in the clinical management of stroke. The objective of monitoring is to prevent intracranial hypertension (IH) from causing further brain damage, which can be irreversible. However, current technology for ICP monitoring is invasive and typically requires pressure probes to be inserted through the skull, which is associated with potential complications. This procedure is reserved for the most severe clinical conditions, potentially overlooking IH injuries in other cases. To address this issue, we propose a non-invasive framework for IH detection that analyzes the morphology of cerebral blood flow velocity (CBFV) waveforms using non-invasive transcranial Doppler (TCD) ultrasound, thereby identifying and preventing IH injuries without the need for invasive procedures. Such a non-invasive framework could help detect IH injuries outside neurointensive care units and help provide timely treatment. The proposed methodology was evaluated on a cohort of 89 patients treated for various ICP-related conditions. Compared to previous frameworks based on semi-supervised learning of specific CBFV metrics, we found that using the raw waveform as input to a machine learning model improves the area under the ROC curve (AUC) to 96%. This is a significant improvement, as, by comparison, the pulsatility index (PI) achieved much lower accuracy in detecting IH at 59%.