Real-Time Analysis of Intracranial Pressure Waveform Morphology (Bioengineering in Neurological Disorders) Part 2

Detection of decreased cerebral blood flow (CBF)

The measure of cerebral blood flow (CBF) is an indicator of perfusion and is therefore very useful in neurocritical care. While imaging techniques provide a snapshot in time and invasive monitors offer continuous readings of the CBF, they usually carry additional risks and require additional equipments and increased cost. Furthermore, non-invasive techniques for CBF monitoring only provide intermittent measures of CBF. There is therefore a need to create a continuous, low cost technique that would not increase risk and could be easily integrated to bedside monitors.

Drawing from the fact that a physiological relation exists between ICP and CBF, we recently investigated if the ICP signal holds predictive information about CBF. Such a complex relationship has only been partially explored such that mean ICP (mICP) is used in the following equation to derive the driving pressure of blood flow through the cerebral vasculature:

where CPP stands for cerebral perfusion pressure, and ABP for systemic arterial blood pressure (ABP). To date, the influence of cerebral vascular changes on both ICP and CBF remain poorly understood. Subtle changes in the morphology of ICP pulses may reflect cerebral vascular changes. Because an ICP waveform can be thought as arising from an incidental arterial pressure pulse influenced by different intracranial compartments, we hypothesize that the ICP waveform carries information composed of changes in cerebral vasculature and hence CBF. In this section, we report our study [11] investigating the ability of ICP morphoology metrics to detect low CBF. A multi-modal dataset originating from brain injured patients with ICP monitoring, global average CBF, and Transcranial Doppler (TCD) assessment was analyzed. Detection of low CBF was posed as a classification problem and implemented using a regularized linear discriminant analysis (LDA). To further improve the performance of the framework, an optimization algorithm was used to find the subset of morphological metrics that maximizes a measure based on the combination of positive predictivity and sensitivity.

Data

The dataset used in our study originates from 63 patients among which 31 were admitted for SAH from aneurysm rupture and 26 had a TBI. The remaining patients were admitted either with arteriovenous malformation, brain tumor, and Intraparenchyma hemorrhage.

The mean global CBF was measured using the intravenous 133 Xenon clearance technique [27] for 11 minutes. TCD [28] was used to insonate the extracranial internal carotid artery (ICA) and the basilar artery (BA). Blood samples were taken immediately before or after CBF measurement. In addition to ECG, ICP was monitored using ventriculostomy and waveforms were recorded from bedside monitors at a sampling rate of 240 Hz. The ICP signal (selected as a one hour segment closest to the CBF measurement) was processed by MOCAIP to extract 24 morphological metrics for each three minute segment of data. A total number of 199 CBF-TCD-ICP segments were extracted from the 63 patients. In addition to the 24 morphological metrics extracted from ICP, the eight following variables weren also extracted: 1) average flow velocities of the right and the left internal carotid artery (ICA); 2) average diastolic flow velocities of the right and the left ICA; 3) average pulsatility indices (PI) of the right and the left ICA; 4) partial carbon dioxide pressure (CO2); 5) total amount of hemoglobin (Hgb); 6) fraction of the blood composed of red blood cells (Hct); 7) mean arterial blood pressure (MAP); 8) amount of CSF drainage in the hour of CBF measurement.

Experiments

A classification experiment was designed to evaluate the power of ICP morphology to discriminate between low and normal CBF value (averaged globally, threshold of 20 ml/min/100g). Four combinations of all 32 available metrics were considered, within which an optimal subset was obtained using the classifier training algorithm described below; with (a) all 32 metrics, (b) only includes MOCAIP, CSF drainage and TCD metrics, (c) only morphological ICP and CSF drainage metrics, and (d) only seven TCD and blood analysis metrics. Using those features, a regularized version [29] of the Gaussian quadratic classifier (QDC) was chosen, and differential evolution (DE) [30] used to optimize the model using feature selection. The objective function for the optimization algorithm is the average of the sensitivity and the positive predictivity (PPV). Each evaluation of the objective function involves a leave-one-patient-out cross-validation.

Results

A sensitivity of 81.8 ± 0.9% and specificity of 50.1 ± 0.2% were obtained using the optimal combination (d) of conventional TCD and blood analysis metrics as input. Using the optimal combination of the morphological metrics alone (c) was able to achieve a sensitivity of 92.5 ± 0.7% and specificity of 84.8 ± 0.8%. Searching for the optimal combination of all available metrics (a) achieved the best result that was marginally better than those from using morphological metrics alone (c). To visually assess how ICP pulse morphology is associated with different perfusion states, we present one typical case in Fig. 4 from a traumatic brain injury patient who had ICP recordings both in the normal and in the low CBF states. In each plot, we overlap the average ICP pulses extracted from every three minutes of data. In addition, we display the CBF value associated with the ICP recording. We observe that the elevation of the third peak within the pulse is associated with low CBF value. This pattern of elevated third peak was observed in six out of the eight patients with positive cases.

Fig. 4. Illustration of average ICP pulse with low (left) and normal (right) CBF.

Table 1 lists the mean and standard deviation of sensitivity, specificity, and positive predictivity value of the three after the leave-one-out (LOO) cross-validation and the bootstrapping (BS) cross-validation. Based on the bootstrapping results, it is observed that combining morphological ICP metrics, TCD, and blood analysis achieves the best performance.

The biggest gain of the performance is caused by the incorporation of the morphological ICP metrics. The number of times each metrics was selected over the experiment was accumulated and analyzed. The following metrics were always selected; dP 13, dP 3, diasP, mICP, L t, L 3, and ICAEd. There are 10 more metrics, including ICAPI and Hct, were selected for majority of the runs. Also, there are 10 metrics that were never selected as part of classifier features including PCO 2 , Hgb, and ICAMean. The complete list of the metrics can be found in the original paper.

Exp.	Val.	SE			SPE			PPV
MOCAIP+TCD+BA	LOO	0.933	±	0.000	0.862	±	0.003	0.356 ± 0.005
	BS	0.94!	±	0.014	0.852	±	0.013	0.341 ±0.024
MOCAIP + TCD	LOO	0.933	±	0.000	0.853	±	0.005	0.342 ± 0.008
	BS	0.920	±	0.014	0.846	±	0.011	0.316 ±0.018
MOCAIP	LOO	0.933	±	0.000	0.851	±	0.003	0.339 ± 0.005
	BS	0.925	±	0.007	0.848	±	0.008	0.320 ± 0.020
BA	LOO	0.867	±	0.000	0.516	±	0.000	0.127 ±0.000
	BS	0.818	±	0.009	0.501	±	0.002	0115 ±0.001

Se: sensitivity; Spe: specificity; PPV: positive predictivity value; LOO: leave-one out; BS: bootstrapping. Table 1. Illustration of the CBF classification results.

Besides the metrics that reflect P 3 elevation were selected as classifier features, the metrics including L t , L 1 , L 2 , and L x were also frequently selected. The engagement of L t in the classification process can be probably explained by the fact that it measures the timing difference between ECG QRS peak and the onset of ICP pulse, which is significantly influenced by systemic arterial blood pressure. Therefore, L t is a relevant measure as it contains information about the driving pressure of the cerebral blood flow.

Discussion

We tested the hypothesis that low global CBF may be detected using morphological metrics extracted from the ICP waveforms through a trained classifier. The main finding was that the incorporation of morphological metrics of ICP was able to significantly improve the performance as compared to only using conventional TCD and blood analysis measurements. Although the study was retrospective and data-driven, we believe that it should motivate further studies to investigate the implications and the underlying mechanisms of the association between ICP pulse morphology and cerebral blood perfusion.

One of the findings from the classification experiment is that the elevation of the third peak of an ICP pulse may indicate low global cerebral perfusion. Some questions can be raised regarding whether controlling ICP can also lead to the control of cerebral venous pressure. This is important because the true perfusion pressure is actually determined by the difference between arterial and venous pressure. Even when the mean ICP is well within the prescribed limit, the true perfusion pressure may be still low in situation of cerebral venous hypertension.

Predicting intracranial hypertension

Intracranial hypertension (IH) poses a constant threat to head injured patients because it may lead to secondary injuries due to decreased cerebral perfusion pressure and cerebral ischemia. Because bedside monitors are usually designed to report only a short-term history of the ICP, large scale patterns and trends on average ICP that might help to prevent IH are not available to the bedside clinician. Therefore, the constant attention of the nursing staff and their prompt reaction following detecting of an IH episode are critical aspects during the management of patients with IH issues. There is a clear need for a computerized monitoring support that would be accurate in predicting ICP hypertension several minutes ahead, offering enough time to attract the full attention of the bedside clinician.

The main hypothesis is that precursor features can be detected in the ICP signal prior to the elevation. Several studies have verified this hypothesis and offer various insights into which form the predictive features might take. Amplitude of ICP [31, 32], variance of changes [3335], and rounding of pulse waveform [36] have been shown to correlate with changes of the mean ICP. Decreases in ABP were observed at the beginning of plateau waves [37], and A waves [38]. Moreover, system analysis [39] suggested that a change in the transfer function that relates ABP to ICP may precede elevations. Several other investigators have also attempted to make predictions using wavelet decomposition of the ICP signal [40-42]. More recently, two studies [12, 43] demonstrated that morphological features extracted from the ICP waveform at various times before the elevation onset contains predictive information for IH. Despite more than 30 years [44, 45] of investigation, the automatic, real-time prediction of ICP hypertension is still beyond current methods. Drawing from the studies [12, 36, 43] indicating that ICP morphology contains relevant predictors of IH, we present in this section a framework [65] to predict IH based on morphological features of ICP. A key contribution of this study is to test the effectiveness of ensemble classifiers (AdaBoost, Extremely Randomized Decision Trees) to make temporal prediction. The proposed framework is evaluated on a representative database of 30 neurosurgical patients admitted for various intracranial pressure related conditions.

Methods

Data source and pre-processing

The dataset of ICP signals originates from the University of California, Los Angeles (UCLA) Medical Center. The ICP and ECG signals were acquired continuously at a rate of 240 Hz or 400 Hz using intraparenchymal sensors from a total of 30 patients treated for various intracranial pressure related conditions. These patients were monitored because of headache symptoms (idiopathic intracranial hypertension, Chiari syndrome, and slit ventricle patients with clamped shunts) with known risks of ICP elevation.

Intracranial hypertension episodes, defined as an elevated ICP greater than 20 mmHg for a period longer than five minutes, were manually delineated by retrospective analysis. The elevation onset was marked at the beginning of the plateau. From this analysis, 13 patients were identified with at least one IH episode, leading to a total of 70 episodes. Based on the expert review of the ICP signal and the manual annotation of the elevation onset, ICP and ECG segments were extracted to cover the period from 20 minutes before to one minute after the onset.

Control segments were constructed by randomly extracting ten-minute ICP segments from the 17 control patients who did not present a single episode of ICP elevation, and from the IH patients no less than an hour before or after an ICP elevation episode. There were a total of 70 control segments which are evenly distributed among all patients. The 140 IH and NON-IH ICP segments were then processed by MOCAIP so that morphological waveform features were extracted to describe each one minute segments of ICP.

Experimental setup

The prediction of ICP elevation is posed as a classification problem where each input exampleis a set ofMOCAIP vectorscalculated over a series of successive one-minute ICP segments. The corresponding outputis a binary variable which equals 1 if the ICP segment led to a treated IH episode. The performance in terms of Area Under the Curve (AUC) is reported for input segmentsextracted at various time-to-onset for the predictive model. The impact of the numberof successive ICP segments used to construct the input vectorsis evaluated for different prediction techniques. The main purpose of the experiment is to test the hypothesis that the use of ensemble classifiers improves the prediction of IH because it can exploit more efficiently the morphological information contained in longer ICP segments located prior to the elevation onset.

Performance Evaluation by Time-To-Onset Variation

For evaluation, a ten-fold cross-validation at the patient level is performed and three different prediction techniques are compared; Multiple Linear Regression (MLR) [46], Adaptive Boosting (AdaBoost) [47], and Extremely Randomized Decision Trees (Extra Trees) [48]. The models are trained on each fold such that each positive example

corresponds to thevectors of morphological metrics

whereis the number of MOCAIP metrics used in this study. The negative examples are randomly sampled from the pulses of the control set such that the training dataset is balanced and contains an equal number of positive and negative examples. For testing, the models are evaluated on the excluded fold by varying the time-to-onsetat which the ICP segmentsare extracted.

Number of morphological ICP segments

This experiment aims at evaluating if the use of additional morphological vectors extracted up to ten minutesprior to the tested time-to-onset can improve the prediction performance of the models. To do so, the classifiers are trained such that the inputof the positive examples are themorphological vectorsextracted from the segments prior to the onset plus the segment concurrent to the IH onset. Training of the models is performed for ten different lengthsfrom one to ten minutes.

Controls of corresponding lengths are randomly extracted from the current fold but remain the same across the different time-to-onset. Each model is evaluated using the average AUC results computed from a ten-fold cross-validation, and ten different time-to-onset

Results

The best results of each method are reported in Fig. 5 in terms of AUC. While the linear classifiers obtain an AUC of [0.87,0.78,0.71] for the time-to-onset corresponding to [1,3,6] minutes prior to the elevation respectively, the use of ensemble classifiers significantly improves the AUC. AdaBoost reaches an AUC [0.93,0.84,0.80], and Extra-Trees performs even better with an AUC of [0.96,0.91,0.87].

The Extra-Trees method shows significant improvement in terms of average AUC when the length of the input is increased. For [1,5,10] minute(s)-long segments, the AUC is respectively [0.77,0.88,0.9]. AdaBoost also shows significant increases in AUC, from [0.75,0.83,0.84] for segments of [1,5,10] minute(s) length, respectively. The linear models are unable to efficiency exploit the larger segments of morphological ICP data. AUC results are [0.76,0.65,0.73] for segments of [1,5,10] minute(s) length, respectively. The best average AUC for the linear model was observed using a single ICP segment. In contrast, the best performance for AdaBoost and Extra-Trees models is obtained withICP segments, corresponding to ten minutes of ICP data extracted prior to the tested time-to-onset.

Although the previous experiments demonstrate that the ensemble classifier models trained over ten minutes of ICP data perform better than the one trained over shorter segments, it is not clear if each additional one minute ICP segment contribute independently to the improvements, or if they are complementary; in which case the dynamic of change of ICP morphology over time would appear to be useful. To test these hypothesis, three different learning strategy of the Extra-Trees models are compared. The first model (a), which is used as baseline, is trained on a singleone-minute long segmentThe second strategy (b) consists in learning independently ten different models on each one-minute long segment and then fuse their output using an arithemic mean. Finally, (c) is to build a single model trained on a series of tenone-minute long segments concatenated to a single input vectorAlthough, the use of ten independent classifiers (b) improves the overall performance versus the use of only one classifier (a), there is an additional increase in AUC by using all ten vectors at the same time (c). This results indicates that the relative values of successive morphological vectors contain relevant precursors of IH. It is an important finding because it means that the dynamic of ICP changes holds critical predictive information.

Fig. 5. Illustration of an IH ICP episode (left). Elevated episodes are divided into 21 segments of one minute. For each segment, MOCAIP is applied to extract morphological vectors. On the right, the AUC is reported after a leave-one-out crossvalidation for each technique. Extra-trees ranks first and is followed by the AdaBoost and Multi-linear classifiers.

Discussion

Thanks to the use of a series of successive one minute ICP segments as input to classifier ensemble techniques, the proposed study has demonstrated that ensemble classifiers can exploit more efficiently the morphological information contained in the pulse. The performance improvement observed in our experiments can be attributed to the following reasons:

• First, as it has been shown in other applications, the two ensemble classifiers perform better than the multiple linear classifier because they can better capture the nonlinearity between the morphological vectors and the outcome.

• Second, the use of a larger segment of ICP segment prior to the onset improves the accuracy.

• Finally, the use of a full sequence of successive morphological vectors at once leads to better models than the one based on individual vectors which indicates that the relative values and the order between successive morphological vectors contain additional precursors.

Although the ICP of brain injured patients is continuously managed by the bedside clinicians, changes in ICP prior to elevation are reflected by complex variations in the morphology of the signal that are difficult to be recognized in real-time. Decision support tools that would alert the bedside clinicians of future ICP elevation would add a new proactive dimension to the current treatment of ICP elevations, which largely remains a reactive procedure. Further improvement of the technical methodology and a better understanding of the physiological meaning of these morphological variations should be possible. Ideally, we would like to translate the rules learned by ensemble classifiers into a physiological model in an attempt to represent ICP dynamics explicitly.