1 Divisions of Anaesthesia, Addenbrooke's Hospital, University of Cambridge , Cambridge, United Kingdom .
2 Department of Surgery, Rady Faculty of Health Sciences, University of Manitoba , Winnipeg, Manitoba, Canada .
J Neurotrauma. 2019 Mar 1;36(5):713-720. doi: 10.1089/neu.2018.5987. Epub 2018 Sep 27.
The goal was to predict pressure reactivity index (PRx) using non-invasive transcranial Doppler (TCD) based indices of cerebrovascular reactivity, systolic flow index (Sx_a), and mean flow index (Mx_a). Continuous extended duration time series recordings of middle cerebral artery cerebral blood flow velocity (CBFV) were obtained using robotic TCD in parallel with direct intracranial pressure (ICP). PRx, Sx_a, and Mx_a were derived from high frequency archived signals. Using time-series techniques, autoregressive integrative moving average (ARIMA) structure of PRx was determined and embedded in the following linear mixed effects (LME) models of PRx: PRx ∼ Sx_a and PRx ∼ Sx_a + Mx_a. Using 80% of the recorded patient data, the LME models were created and trained. Model superiority was assessed via Akaike information criterion (AIC), Bayesian information criterion (BIC), and log-likelihood (LL). The superior two models were then used to predict PRx using the remaining 20% of the signal data. Predicted and observed PRx were compared via Pearson correlation, linear models, and Bland-Altman (BA) analysis. Ten patients had 3-4 h of continuous uninterrupted ICP and TCD data and were used for this pilot analysis. Optimal ARIMA structure for PRx was determined to be (2,0,2), and this was embedded in all LME models. The top two LME models of PRx were determined to be: PRx ∼ Sx_a and PRx ∼ Sx_a + Mx_a. Estimated and observed PRx values from both models were strongly correlated (r > 0.9; p < 0.0001 for both), with acceptable agreement on BA analysis. Predicted PRx using these two models was also moderately correlated with observed PRx, with acceptable agreement (r = 0.797, p = 0.006; r = 0.763, p = 0.011; respectively). With application of ARIMA and LME modeling, it is possible to predict PRx using non-invasive TCD measures. These are the first and as well as being preliminary attempts at doing so. Much further work is required.
目的是使用基于经颅多普勒（TCD）的脑血管反应性指数，即收缩期血流指数（Sx_a）和平均血流指数（Mx_a），来预测压力反应指数（PRx）。使用机器人 TCD 平行于直接颅内压（ICP）连续获得大脑中动脉脑血流速度（CBFV）的扩展持续时间时间序列记录。从高频存档信号中得出 PRx、Sx_a 和 Mx_a。使用时间序列技术，确定了 PRx 的自回归积分移动平均（ARIMA）结构，并将其嵌入 PRx 的以下线性混合效应（LME）模型中：PRx∼Sx_a 和 PRx∼Sx_a+Mx_a。使用记录的患者数据的 80%创建和训练 LME 模型。通过赤池信息量准则（AIC）、贝叶斯信息量准则（BIC）和对数似然（LL）评估模型的优越性。然后使用剩余 20%的信号数据来预测剩余 LME 模型中的 PRx。通过 Pearson 相关性、线性模型和 Bland-Altman（BA）分析比较预测和观察到的 PRx。10 名患者有 3-4 小时的连续不间断 ICP 和 TCD 数据，用于此初步分析。确定 PRx 的最佳 ARIMA 结构为（2,0,2），并将其嵌入所有 LME 模型中。确定 PRx 的两个最佳 LME 模型为：PRx∼Sx_a 和 PRx∼Sx_a+Mx_a。两个模型的估计和观察到的 PRx 值高度相关（r>0.9；p<0.0001），BA 分析的一致性也可以接受。使用这两个模型预测的 PRx 与观察到的 PRx 也有中度相关性，一致性也可以接受（r=0.797，p=0.006；r=0.763，p=0.011；分别）。通过应用 ARIMA 和 LME 建模，可以使用非侵入性 TCD 测量来预测 PRx。这是首次尝试，也是初步尝试。还需要做更多的工作。
