Model-based noninvasive estimation of intracranial pressure from cerebral blood flow velocity and arterial pressure

Abstract

Intracranial pressure (ICP) is affected in many neurological conditions. Clinical measurement of pressure on the brain currently requires placing a probe in the cerebrospinal fluid compartment, the brain tissue, or other intracranial space. This invasiveness limits the measurement to critically ill patients. Because ICP is also clinically important in conditions ranging from brain tumors and hydrocephalus to concussions, noninvasive determination of ICP would be desirable. Our model-based approach to continuous estimation and tracking of ICP uses routinely obtainable time-synchronized, noninvasive (or minimally invasive) measurements of peripheral arterial blood pressure and blood flow velocity in the middle cerebral artery (MCA), both at intra-heartbeat resolution. A physiological model of cerebrovascular dynamics provides mathematical constraints that relate the measured waveforms to ICP. Our algorithm produces patient-specific ICP estimates with no calibration or training. Using 35 hours of data from 37 patients with traumatic brain injury, we generated ICP estimates on 2665 nonoverlapping 60-beat data windows. Referenced against concurrently recorded invasive parenchymal ICP that varied over 100 millimeters of mercury (mmHg) across all records, our estimates achieved a mean error (bias) of 1.6 mmHg and SD of error (SDE) of 7.6 mmHg. For the 1673 data windows over 22 hours in which blood flow velocity recordings were available from both the left and the right MCA, averaging the resulting bilateral ICP estimates reduced the bias to 1.5 mmHg and SDE to 5.9 mmHg. This accuracy is already comparable to that of some invasive ICP measurement methods in current clinical use.

Fig 1

Progressive abstraction of cerebrovascular physiology. (A) Relevant cerebrovascular anatomy: brain tissue (BT), cerebrospinal fluid (CSF), and cerebral arterial network (AN). (B) Schematic representation of the main cerebrovascular compartments and associated physiological variables: cerebral blood flow (CBF), arterial blood pressure (ABP), and intracranial pressure (ICP); the collapsed venous segment is also shown. (C) Lumped circuit-model representation of cerebrovascular physiology: cerebral blood flow q(t), cerebral arteriovenous flow q₁(t), and arterial blood pressure p_a(t). ICP denotes both extra-luminal pressure and the effective downstream pressure for cerebral perfusion.

Fig. 2

Schematic representation of data acquisition, showing representative intracranial pressure (ICP), arterial blood pressure (ABP), and cerebral blood flow velocity (CBFV) waveforms. (A) Possible direct, invasive recordings of ICP over time through a parenchymal probe (PP) or ventricular catheter (VC). (B) Invasive recording of ABP waveform through radial artery catheterization (RAC) and noninvasive recording of middle cerebral artery (MCA) blood flow velocity waveform by transcranial Doppler (TCD) ultrasonography, used together for noninvasive estimation of ICP.

Fig. 3

Comparison of measured and estimated ICP in four brain-injured patients. In panel A, ICP was estimated on a sliding 60-beat data window. In panels B–D, the estimates were obtained on 60-beat non-overlapping data windows. (A) Single plateau wave. (B) Severe progressive intracranial hypertension. (C) Two consecutive plateau waves. (D) Borderline normal ICP. All patient data are summarized in table S2.

Fig. 4

Bland-Altman plots of overall estimation performance. ICP is mean measured intracranial pressure and nICP is the noninvasive estimate, each computed on a 60-beat estimation window. (A) ICP and nICP on 2,665 non-overlapping windows from 45 patient records. (B) ICP and nICP on 1,673 non-overlapping windows from 30 records with bilateral CBFV recordings, where averaging of left and right estimates reduced the bias and SDE from (A). (C) ICP and nICP averaged across all windows in each of 45 patient records. For all three plots, the bias is shown as the dashed line and dash-dotted lines indicate the limits of agreement, computed as bias ± 2SDE.

Fig. 5

Receiver operating characteristics (ROC) for detection of intracranial hypertension, defined as ICP > 20 mmHg. One ROC was computed for all 2,665 nICP/ICP data pairs (red), and a second one for nICP/ICP data averaged across each of the 45 patient records (blue).