Assessing the prediction potential of an in silico computer model of intracranial pressure dynamics.

OBJECTIVE Traumatic brain injury (TBI) frequently results in poor outcome, suggesting that new approaches are needed. We hypothesized that a patient-specific in silico computer model of intracranial pressure (ICP) dynamics may predict the ICP response to therapy. DESIGN In silico model analysis of prospectively collected data. SETTING Twenty-three and 16-bed pediatric intensive care units in two tertiary care academic hospitals. PATIENTS Nine subjects with severe TBI undergoing ICP monitoring (7 M/2 F, age range 3-17 years). INTERVENTIONS Random changes in head-of-bed (HOB) (0 degrees , 10 degrees , 20 degrees , 30 degrees , 40 degrees ) elevation and respiratory rate (to achieve a DeltaETco2 = +/-3-4 mm Hg) were performed daily according to a study protocol as long as an intracerebral monitoring device was in place. METHODS AND MAIN OUTCOME MEASURES A six-compartment dynamic ICP model was developed based on published equations and parametric data (baseline model parameter values). For each of 24 physiologic challenge sessions, patient-specific model parameter values were estimated that minimized the model fitness error, the difference between model-calculated ICP and observed ICP, both for baseline parameters and patient-specific parameter. Next, model prediction error was measured using two analyses. First, a "within" session analysis estimated parameter values using data from an initial Segment A, and then used those parameter values to predict the ICP during a later Segment B. The predicted ICP for B was compared with the observed ICP for B. Second, a "between" session analysis was performed. This analysis used parameter values estimated from earlier sessions to predict the ICP in later sessions. Fitness and prediction errors were measured in terms of mean absolute error (MAE). To normalize the errors, MAE was divided by the mean absolute deviation (MAD) for the associated segment or session, yielding a measure for both model fitness error and model prediction error that is favorable when <1. RESULTS For baseline parameter values, MAE/MAD was <1 in 2 of 24 (8%) sessions. For session-specific parameter values, MAE/MAD was <1 in 21 of 24 (88%) sessions and <0.5 in 9 of 24 (38%) sessions. Sessions with low (<12 mm Hg) (n = 8; 33%) or high (>18 mm Hg) (n = 6; 25%) ICP had lower error than moderate ICP (12-18 mm Hg) (n = 10; 42%). MAE/MAD was <1 for 6 of 22 (27%) for within-session predictions and 3 of 31 (10%) for between-session predictions. CONCLUSIONS The protocol for collecting physiologic data in subjects with severe TBI was feasible. The in silico ICP model with session-specific parameters accurately reproduced observed ICP response to changes in head-of-bed and respiration rate. We demonstrated modest success at predicting future ICP within a session and to a lesser extent between sessions.