Comparison of different calibration methods in a non-invasive ICP assessment model

Abstract

Objective: Previously we described the method of continuous intracranial pressure (ICP) estimation using arterial blood pressure (ABP) and cerebral blood flow velocity (CBFV). The model was constructed using reference patient data. Various individual calibration strategies were used in the current attempt to improve the accuracy of this non-invasive ICP (nICP) assessment tool. Materials and methods: Forty-one patients (mean, 52 years; range, 18–77 years) with severe brain injuries were studied. CBFV in the middle cerebral artery (MCA), ABP and invasively assessed ICP were simultaneously recorded for 1 h. Recording was repeated at days 2, 4 and 7. In the first recording, invasively assessed ICP was recorded to calibrate the nICP procedure by means of either a constant shift of nICP (snICP), a constant shift of nICP/ABP ratio (anICP) or by including this recording for a model reconstruction (cnICP). At follow-up days, the calibrated nICP procedures were applied and the results compared to the original nICP. Results: In 76 follow-up recordings, the mean differences (Bias), the SD and the mean absolute differences (ΔICP) between ICP and the nICP methods were (in mmHg): nICP, −5.6 ± 5.72, 6.5; snICP, +0.7 ± 6.98, 5.5, n.s.; anICP, +1.0 ± 7.22, 5.6, n.s.; cnICP, −3.4 ± 5.68, 5.4, p < 0.001. In patients with craniotomy (n = 19), the nICP was generally higher than ICP. This overestimation could be reduced by cnICP calibration, but not completely avoided. Discussion: Constant shift calibrations (snICP, anICP) decrease the Bias to ICP, but increase SD and, therefore, increase the 95% confidence interval (CI = 2 × SD). This calibration method cannot be recommended. Compared to nICP, the cnICP method reduced the Bias and slightly reduced SD, and showed significantly decreased ΔICP. Compared to snICP and anICP, the Bias was higher. This effect was probably caused by the patients with craniotomy. Conclusion: The cnICP calibration method using initial recordings for model reconstruction showed the best results.