Real-Time Analysis of Intracranial Pressure Waveform Morphology

The cranial vault is composed of four fundamental components: arterial blood, venous blood, cerebrospinal fluid (CSF), and brain parenchyma. Intracranial pressure (ICP) represents the pressure within the brain parenchyma and cerebrospinal fluid (CSF). The environment within the cranial vault is unique compared to other organ systems; it is enclosed within a rigid skull and thus small volumetric changes in any of the four elements lead to significant changes in ICP. One example of this, is the periodic influx of arterial blood over the cardiac cycle; this change causes the ICP pulse pressure waveform. The pulse pressure waveform has three characteristic peaks hypothesized to correspond to different physiologic components. Early work [1] demonstrated the relationship between pulsations in the choroid plexus and the pulse pressure waveform. Moreover, other studies [2] compared the right atrium (venous) and the aortic (arterial) pressures to the intracranial waveform in the cistern magna and showed that although the cranial pulse pressure is related to arterial pulsations, there is also a venous component. The results of these studies support the current theories for the etiology of the characteristic peaks in the ICP pulse pressure waveform. The majority of the literature indicates that P1, the percussion wave, corresponds with the pulsation of the choroid plexus and/or large intracranial conductive vessels [3-5]. The rebound of the percussion wave is thought to contribute to P2, which has also been related to cerebral compliance [5]. Finally, the dicrotic wave, or P3, is thought to be venous in origin [2, 4, 6].


Introduction
The cranial vault is composed of four fundamental components: arterial blood, venous blood, cerebrospinal fluid (CSF), and brain parenchyma.Intracranial pressure (ICP) represents the pressure within the brain parenchyma and cerebrospinal fluid (CSF).The environment within the cranial vault is unique compared to other organ systems; it is enclosed within a rigid skull and thus small volumetric changes in any of the four elements lead to significant changes in ICP.One example of this, is the periodic influx of arterial blood over the cardiac cycle; this change causes the ICP pulse pressure waveform.The pulse pressure waveform has three characteristic peaks hypothesized to correspond to different physiologic components.Early work [1] demonstrated the relationship between pulsations in the choroid plexus and the pulse pressure waveform.Moreover, other studies [2] compared the right atrium (venous) and the aortic (arterial) pressures to the intracranial waveform in the cistern magna and showed that although the cranial pulse pressure is related to arterial pulsations, there is also a venous component.The results of these studies support the current theories for the etiology of the characteristic peaks in the ICP pulse pressure waveform.The majority of the literature indicates that P1, the percussion wave, corresponds with the pulsation of the choroid plexus and/or large intracranial conductive vessels [3][4][5].The rebound of the percussion wave is thought to contribute to P2, which has also been related to cerebral compliance [5].Finally, the dicrotic wave, or P3, is thought to be venous in origin [2,4,6].
ICP monitoring is fundamental to the management of numerous intracranial pathologies, including the management of traumatic brain injury (TBI), subarachnoid hemorrhage (SAH), hydrocephalous, and any other conditions where pathological changes in intracranial volume (ICV) may occur.Maintaining an ICP within normal limits is vital for adequate perfusion of the brain.An increase in ICP (intracranial hypertension) reduces arterial blood flow into the cranial vault, while decreased ICP (intracranial hypotension) results in severe headaches.Although the importance of monitoring ICP is well established, only recently has ICP pulse pressure waveform morphology yielded promising results due to advancing technology.However, the importance of ICP pulse pressure waveform morphology has been known for years; special interest should be focused not only on the height of mean ICP, but also on the pressure pulse curve, because the configuration and pulse amplitude of the CSF pulsations can be regarded to a certain extent as an index of the state of intracranial elastance or cerebral bulk compliance [6].

101
ICP Segmentation: The continuous raw ICP signal is segmented into a series of individual ICP pulses using a pulse extraction technique [21] combined with the ECG QRS detection [17,22] that locates each ECG beat.Because ICP recordings are subject to various noise and artifacts during the acquisition process, an average pulse is extracted from a series of consecutive ICP pulses using hierarchical clustering [23].
Peak Candidates: Peak candidates, each being a potential match to one of the three peaks, are detected at curve inflections of the average ICP pulse using the second derivative of the signal.

Second-order ICP Features:
We have shown in a recent study [20] that the first derivative of the ICP signal is very useful to discriminate between ICP peaks and therefore preoviding improve recognition.In our previous work, the first x L demonstrated the best improvements in comparison with the second derivative and the curvature within MOCAIP   framework.
The derivative x L is computed according to the smoothed version L of the ICP, =(,) ( 1 ,) . x where the ICP signal () Ix is first convolved with a Gaussian smoothing filter (; ) Gx with the standard deviation  to generate (, ) Lx , (, ) = (; ) *() .


(2) Peak Recognition: The peak recognition tasks consists in recognizing the three peaks 123 (, ,) p pp among the set of candidate peaks detected within the pulse.Several techniques have been developed and can be used such as independent Gaussian models [18], Gaussian Mixture Models (GMM), and spectral regression (SR) analysis [19].Depending on the technique, it can exploit the latency of the peak candidates, the raw ICP pulse, or different features extracted from the pulse.
Morphological metrics: Once the peaks have been detected in an ICP pulse, a set of metrics (Fig. 1) is used to describe the shape morphology in a parametric way.This allows to obtain a better understanding about the type of variations that take place.

Bayesian tracking of ICP morphology
In this section, we present a probabilistic framework [24] to track ICP peaks in real time.The tracking is posed as inference in a graphical model that associates a continuous random variable to the position of each of the three peaks, in terms of their latency within the pulse and pressure level.The model (Section 2.3) represents the dependencies between the peaks using a Kernel Density Estimation (KDE) from evidence collected from manually annotated pulses, while Nonparametric Belief Propagation (NBP) [25] is used during the detection process (Section 2.4).
We assume that the tracking framework is presented with a series of raw pulses extracted from the ICP signal.x is associated an observation , ti y directly extracted from the position of the peak within the current pulse.

Probabilistic tracking framework
The graphical model used in our tracking framework defines relations between pairs of nodes.


. Equation (3) formalizes the integration of the observation using a Gaussian model, where  is a smoothing parameter, and i  is a constant factor that accounts for missing peaks by the detector.

Compatibility and Temporal potentials
Temporal potentials define the relationship between two successive states of a peak.They are defined as a Gaussian difference between their arguments, where the standard deviation t  of the model was previously estimated using maximum likelihood (ML) on training data.Compatibility potentials ,,, (, ) i j ti t j xx  , however, are not expected to follow a Gaussian distribution.Each potential is represented by a KDE [26] ,,, (, ) = (; )


 that is constructed by collecting co-occurring ICP peak positions across the training set.

Tracking ICP peaks using nonparametric bayesian inference
Detecting peaks in an ICP pulse at time t amounts to estimating accumulated so far.Thus, peak detection is achieved through inference in our graphical model.One way to do this efficiently is to use Nonparametric Belief Propagation [25].It is a message passing algorithm for graphical models that generalizes particle filtering and Belief Propagation (BP).Messages are repeatedly exchanged between nodes to perform inference.Following the notation of BP, a message ij m sent from node i to j is written An example of inference is provided in Fig. 3 where the latency and the elevation of the three peaks is tracked simultaneously and in real-time on a pulse-by-pusle fashion.

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 (CO 2 ); 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.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 www.intechopen.com Real-Time Analysis of Intracranial Pressure Waveform Morphology 107 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 ,d P 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.
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 [33][34][35], 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][41][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.

Data source and pre-processing
The 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  is 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

Number of morphological ICP segments
This experiment aims at evaluating if the use of additional morphological vectors extracted up to ten minutes ( =10

Results
The best results of each method are reported in Fig. 5    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 (reproduced with permission from [65]).

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.

Acute hypercapnic cerebral vasodilatation
The influence of changes within the cerebral vasculature and there impact on ICP remain poorly understood in humans.Studies [49] have shown that in head injured patients, the cerebral perfusion pressure is inversely proportional to the amplitude of pulsatile inflow and, consequently, the exponential shape of the pressure-volume relationship is not the only factor influencing the magnitude of ICP pulse wave [50].This section presents a recent study [14] conducted by our group to test the hypothesis that acute hypercapnic cerebral vasodilatation induces consistent changes in ICP waveform morphology.This hypothesis is tested on a dataset of ICP signals of uninjured patients undergoing a CO 2 inhalation challenge in which hypercapnia induced acute cerebral vasodilatation.For each morphological metrics extracted from the ICP waveforms using MOCAIP, the consistency and rate of change were analyzed.

Materials and methods
The hypercapnic dataset consists of the ICP and ECG recordings of four patients, who were admitted at UCLA medical center for the evaluation of their chronic headaches.During their hospitalization, the patients received continuous ICP monitoring using intraparenchymal microsensors situated in the right frontal lobe.They also underwent a CO 2 challenge test by inhaling a 5% CO 2 mixture for less than three minutes.During the test, ICP and ECG signals were recorded at a sampling rate of 400 Hz at the bedside with a dedicated acquisition system.
To quantify the rate of change of each metric over a specific time segment, a line was first fitted to the segment of interest and then the slope of this line was used to calculate the rate of the metric change.The sign of the obtained hourly rate of change (negative vs. positive) was used to determine the trend of change (decreasing vs. increasing).

Experimental protocol
The average duration of the selected data segment was ( 5.1 0.7  minutes) which included (1.5 0.5  minutes) of baseline, ( 2.5 0.5  minutes) of CO 2 challenge test and (1.1 0.2  minutes) of post-test data.The slope of the lines fitted to each of the extracted metrics over the rising edge of ICP signal during CO 2 challenge test and the falling edge of ICP signal during the post-test normal breathing, were used to define the hourly rate of change during the test and post-test normal breathing, respectively.For the purpose of comparing the hourly rate of change between different metrics, each metric rate was normalized by the average value of the corresponding metric over either the last ten beats of the baseline or the first ten beats of the stabilized part of the post-test data.
To evaluate the results from baseline, test, and post-test, we report 1) the consistency of changes of individual metrics; 2) differences in rate of metric changes; and 3) which of the peak regions (P1, P2 and P3) has a more dramatic change during the cerebral vasodilation; the specifics of how the region assignments were defined are in the publication [14].In addition, we calculated the region-weighted relative hourly rate of change averaged over the subset of consistent metrics.Fig. 6.Illustration of induced hypercapnia for a headache patient.There are three segments (Baseline, CO 2 Inhalation, and Post-test) with the linear fits shown by dotted lines (reproduced with permission from [14]).

Results
Fig. 6 depicts the mean ICP value for one of the headache patients during the baseline, the CO 2 challenge test, and the post-test normal breathing.When the patient inhales the 5% mixture of CO 2 , mean ICP increases over time, reaches a saturation level, and then stabilizes.When the patient returns to breathing normal concentrations of CO 2 , the mean ICP returned to baseline in less than one minute.
Investigating the hourly rate of change for all 128 ICP metrics during the hypercapnic and normal breathing post-test data, reveal that; out of 128 ICP metrics, 72 metrics had consistent changes in association with CO 2 changes for all four subjects.We observe that no metrics had the same trend during both the hypercapnic and normal breathing post-test data.This observation is consistent with our expectation that, if a variable has a specific trend of change in one condition, the change would be in the opposite direction as the condition is reversed.We also observe that for all subjects, 50 metrics consistently increased ("+" metrics) during hypercapnia and decreased when patients switched back to room air and 22 metrics consistently decreased ("-" metrics) during CO 2 inhalation phase and increased during post-test normal breathing.

Discussion
Acute vasodilatation caused consistent changes in a total of 72 ICP pulse morphological metrics.In addition, it appears that the P 2 sub-region responded to cerebral vascular changes in the most consistent way with the greatest changes as compared to P 1 and P 3 sub-regions.Information with regard to how ICP pulse morphology responds to vasodilatation and vasoconstriction may allow surrogate, continuous monitoring of the cerebral vasculature.
In summary, the present work provides positive preliminary results related to the hypothesis that the dilation/constriction of the cerebral vasculature results in detectable consistent changes in ICP morphological metrics.Acute vasodilatation caused consistent changes in a total of 72 ICP pulse morphological metrics.In addition, it appears that the P 2 sub-region responded to cerebral vascular changes in the most consistent way with the greatest changes as compared to P 1 and P 3 sub-regions.

Morphological ICP waveform characteristics during slow waves
The diagnosis and management of NPH remain challenging mostly due to a lack of reliable methods of selecting candidates for shunt implantation and third ventriculostomy.There exists positive [51,52] and negative [53][54][55] evidence that frequent presence of ICP slow waves predicts a positive outcome after shunt implant.ICP slow waves, also known as Lundberg's B-waves, are defined as oscillations with a frequency of 2-0.5/minute and large amplitude [56].There exists indirect evidence that certain characteristics of ICP slow waves may contain useful information for correctly diagnosing NPH and predicting shunt response.It was recently found that increased ICP pulse pressure amplitude has a predictive value for shunt response [9,57].In this section, we describe our recent attempt [13] to detect and separate periods of ICP slow waves (BW, "B-wave") from those of flat or nearly flat ICP (NW, "no wave") in an overnight ICP recording.We hypothesized that mean values and variations of ICP pulse morphological metrics extracted by the MOCAIP algorithm can be effectively used as input feature vectors to a classification algorithm to distinguish between periods of flat ICP (NW) and those with slow waves (BW).Such a classifier can then be used to construct an automated ICP slow wave recognition algorithm.

Materials and methods
Pre operative (shunt) overnight ICP recordings performed in 44 patients hospitalized at the UCLA Adult Hydrocephalus Center.Hydrocephalus was diagnosed for all patients.An intraparenchymal ICP sensor was inserted in the right frontal lobe and simultaneous recordings of ICP and ECG signals were performed at a sampling rate of either 400 Hz or 240 Hz.The signal recordings were visually screened by three independent experts to select both BW and NW patterns.Fig. 7 presents an example of ICP overnight monitoring where initially fairly stable ICP recording transforms into clearly distinguishable slow waves with relatively high amplitude and asymmetrical shape.In our study, small ICP slow waves with amplitude less than six mmHg were classified as NW pattern.As both patterns of ICP (BW and NW) might occur multiple times during overnight monitoring in the same patient, we included several selections from the same study.The total number of selected patterns was 276 (NW-131 and BW-145).Fig. 7. Illustration of B wave segment (reproduced with permission from [13]).

Feature selection and classification
A total of 48 metrics (comprising 24 morphological metrics plus their standard deviation) were used to described the morphology of each ICP pulse.
To optimize the classification performance, three feature selection techniques (differential evolution (DE), discriminant analysis (DA) and analysis of variance based on Anova (V)) were applied to find an optimal set of MOCAIP metrics under different criteria.In addition, we selected three sets of metrics common to those found by combination of two selection methods, to be used as classification features (differential evolution and analysis of variance, discriminant analysis and analysis of variance, and combination of differential evolution and discriminant analysis).
A regularized linear quadratic classifier was used discriminate between BW from NW based on morphological ICP features.We repeated classification experiment seven times: first time for the set of optimal ICP metrics chosen by DE algorithm, second time for the set of ICP metrics selected based on analysis of variance (V), the third time for these metrics picked by step-wise DA approach, the next three runs were repeated for the metrics overlapped two sets: DE+V, DA+V and DE+DA, and finally we used only one parameter --S.D. of mean ICP to confirm possible advantage of using combination of metrics over single metric (SM).To compare the performance of classification for different sets of metrics, we ran the bootstrapping procedure 25 times to calculate average and S.D. of performance metrics including Se, Spe, PPV and Ac.

Results and discussion
Results indicate that the changes in six morphological metrics (S.D. of: dP 2 , dP 12 ,dP 13 , Curv T , dICP, mICP) were sufficient to distinguish between BW from NW with high specificity 96.2% and acceptable accuracy 88.9%.
Based on our experiments, using DE in conjunction with Anova to derive the final set of metrics leads to the best classification results.The combination of methods: DA+V as well as DE+DA did not improve the classification performance.Combining DE and V methods appeared to be complementary in our study.Since accuracy of classification experiment for Anova+DE is the highest (88.9%) and the number of input metrics is reasonable small (six metrics), we ranked this method as the best for separating BW from NW.
The final list of six metrics selected by DE+V (S.D. of: dP 2 , dP 12 , dP 13 , Curv T , dICP, mICP) reflects the most salient changes associated with ICP increase due to slow waves.We found that amplitude of P 2 component notably increases during ICP slow wave occurrence (hence S.D. of dP 2 ), which is also associated with increase in total curvature of the pulse (hence S.D. of Curv T ).Amplitude of P 1 , most likely related to systole of arterial pressure [58], shows a moderate increase during BW in comparison with both dP 2 and dP 3 (hence S.D. of dP 12 , S.D. of dP 13 ).

Predicting lumbar drain outcome from ICP morphology
Due to the complication rate and malfuction assoicated with existing shunt technologies, differentiating patient with NPH that will benefit from shunt implantation is significant.Extended lumbar drain (LD) for 72 hours has become a popular pre-shunt workup in many neurosurgical centers [59][60][61].Its popularity stems from the findings of an excellent study which demonstrated the high specificity and sensitivity of predicting shunt response based on LD test outcome [60,62].However, the mechanistic relationship between positive LD outcome and shunt response remains to be elucidated.Therefore, many centers also employ additional tests to assess patients for predicting their shunt response.Overnight ICP monitoring is one of such tests.It has been reported in many studies that overnight ICP monitoring is able to reveal many phenomena that a short-term ICP monitoring or lumbar puncture cannot reveal, e.g., large but slow oscillations of ICP that are termed B-waves.However, the capability of overnight ICP monitoring to predict shunt response remains controversial.[51,[53][54][55].
Given these existing contradicting results, elucidating the potential prognostic value of overnight ICP monitoring is significant.Overnight ICP monitoring could be an important economic alternative to LD because it requires a shorter hospital stay for patients if its prognostic value can be fully explored.Unfortunately, many existing analysis methods applied to overnight ICP recordings are very subjective and usually extract limited amounts of information from overnight ICP recordings.On the other hand, a few studies utilizing more advanced ICP signal analysis methods [7,10] have demonstrated that the amplitude of ICP pulse has excellent predictive value for shunt response.In addition to lack of complex analysis methods of overnight ICP, a potential limit to the existing studies that use shunt response as an end-point is that such outcome www.intechopen.comReal-Time Analysis of Intracranial Pressure Waveform Morphology 117 is also determined by many other post-implantation factors not necessarily associated with overnight ICP characteristics.
Therefore, this work aims to develop an automated method of feature extraction and decision rule construction from overnight ICP recording.This method is used to assess the predictive accuracy of LD outcome instead of shunt outcome.Specifically, we used the MOCAIP algorithm to analyze the overnight ICP recordings.MOCAIP was applied to consecutive short segments of an ICP recording resulting in 128 metrics per each segment.Then various feature functions were designed to summarize the distribution of the 128 metrics from all the segments per each ICP recording.The predictive value of each feature-metric pair was then assessed using the area under curve [60] of the receiver operator characteristic (ROC) curve.Based on the ROC curve, a simple decision rule involving one metric can be derived for predicting LD outcome.To further improve the performance, we proposed an automated way to combine two such rules.

Materials and methods
The present retrospective study involved 54 patients undergoing pre-shunt workup that included overnight ICP monitoring and extended lumbar drain (three-day) while hospitalized at the UCLA Adult Hydrocephalus Center.An intraparenchymal ICP microsensor (Codman and Schurtleff, Raynaud, MA) was inserted in the right frontal lobe and monitoring started at least one night before the placement of the LD.Continuous waveform data including ECG and ICP was captured using the BedMaster system with a sampling rate of 240 Hz.Patients were assessed both pre-LD and post-LD before discharge from the hospital by ten meter walking exam, and an NPH routine assessment which includes the Mini-mental state examination (MMSE).The LD outcome used in the present work was retrospectively collected as indicated from the clinical report of the follow-up visit after the LD procedure and predominantly focused on the improvement in gait.The average age and standard deviation of the 54 patients is 72.05 ± 9.63 years respectively, with a gender distribution of 35 males and 19 females.Among the patient cohort, 12 patients (seven males and five females) showed no improvement in gait following the procedure.

Feature functions
Once the metrics for each dominant pulse have been computed an additional processing step is used to summarize each of the 128 MOCAIP metrics for one overnight ICP recording via five feature functions.For example, if a patient has ten hours of continuous ICP data there will be approximately 1200 dominant pulses, for each dominant pulse there is 128 MOCAIP metrics, each of the MOCAIP metrics can be summarized with the five feature functions (below).There is no prior knowledge with regard to the best way of summarizing an overnight ICP recording.Therefore, it is necessary to allow for an automated process to identify the best candidates.We have evaluated the following feature functions: Percentage feature: This function calculates the percentage of time when a metric is greater than a threshold.This threshold is determined by pooling data from all patients and then determined as the average of the corresponding MOCAIP metric.
Percentage of standard deviation feature: This function calculates the percentage of time of the standard deviation of MOCAIP metrics greater than a threshold calculated from a fiveminute ICP.The threshold was determined as the standard deviation of pooled MOCAIP metrics from all patients.
Range feature: This function calculates the difference between the 95 percentile and the 5 percentile of each individual metric of an overnight ICP recording.

Optimal single-metric rule
After applying a feature function, an overnight ICP recording is reduced to a vector of 128 metric-feature pairs.We next seek for a single-metric rule to predict LD outcome in the following form: if a metric-feature of an overnight recording is greater (or smaller) than a threshold, then LD outcome will be positive.Note, rules published in several existing studies [10,63] can be considered as specific instances of the above form.For each metric-feature, we generate a sequence of threshold values by pooling all the data.This sequence of threshold values can be used to generate two ROC curves for each metric-feature, one of which corresponds to the greater than situation and another of which corresponds to the less than situation.Then we retain the ROC that has an AUC greater than 0.5 and discard the other for each metric.Next, for each feature the optimal metric is defined as the one with the greatest AUC.The impact of the optimal metric selection (as a function of AUC) is explained in the results section.Following the optimal metric determination for each feature, one natural step further is to determine if one can combine two such rules to obtain a better performance (a combination of two of the five features described above).Considering the combination of two features (A and B) with the corresponding set of false positive, true positive, false negative, and true negative cases represented (from rule A) as FPA, TPA, FNA, TNA, respectively, the notation is analogous for rule B. Furthermore, we shall consider two possible combination operations: AND and OR.If the OR operator is used for combination, then the resultant sets of false positive, true positive, false negative, and true negative are FPA ∪ FPB, TPA ∪ TPB, FNA ∩ FNB, and TNA ∩ TNB.On the other hand, if the AND operator is used for combination, then the corresponding sets are FPA ∩ FPB, TPA ∩ TPB, FNA ∪ FNB, and TNA ∪ TNB.Based on this analysis, the accuracy of the combined predictive rules can be calculated using their respective true positive and true negative values along with the total number of patients.Therefore, one can choose the combinations with the maximal accuracy.

Data analysis protocol
In this section, we summarize the data analysis protocol used in the present work to determine the best rule combination and its accuracy.The following steps were taken to analyze the 54 overnight ICP recordings: 1.Each recording is analyzed using the MOCAIP algorithm on consecutive 30-second segments.2. Resultant MOCAIP metrics are then manually checked for any errors in recognizing legitimate dominant pulse and placement of the landmarks.This step is facilitated by using software developed in house.3. Generate the threshold sequence for each metric that will be used to generate ROC curve.The i-th value of the sequence of thresholds is calculated for each metric-feature by the following equations: (i-1) where L is the mean plus the ten standard deviation of the metric-feature divided by the number of steps, which is 100 in the present work.4. Each value of the threshold sequence is used to form two simple rules (greater than and less than), which are then used to assess the 54 cases.The corresponding false positive rate (FPR) and true positive rate (TPR) are obtained.After sweeping through the sequence, two ROC curves are obtained but only the one with AUC greater 0.5 is retained.5. Determine the optimal metric-feature.This is achieved by selecting one metric-feature that has the largest AUC per each feature function evaluated.In other words, for each feature defined above, the metric (out of the 128 calculated by MOCAIP) with the greatest AUC is defined as the optimal metric-feature (five total, one for each feature).6. Determine the rule.For the optimal metric-feature, the optimal point on the ROC curve is selected as the operating point, which gives a false positive rate less than 0.35.7. Then the accuracy of the ten pairs (two combination rules and five feature) are calculated.8.The best rule combination is then determined as the one with the greatest accuracy.

Results
Panel F of Fig. 8 displays ROC curves of each of the 128 metric-features (average feature function is used here as an example).The ROC curve shown in bold represents the metric with the greatest AUC; which was selected as the optimal metric-feature for rule construction.The ROC curves of optimal metric-pair corresponding to each of the five feature functions (A: Average feature function; B: Standard deviation feature function; C: Percentage of average feature function; D: Percentage of five-minute standard deviation feature function; E: Inter 5-95 percentile feature function) are shown in Panels A through E of Fig. 8 where the selected MOCAIP metric is also spelled out.On each curve, the operating point is circled, which corresponds to the particular threshold value selected for rule construction.
In summary, the individual metric-feature rules found in the present work are: 1.If standard deviation of RLv2p2Lp1p2 of overnight ICP recording is > 0.0547, then patient will respond to LD. 2.If average RCurvp3Curvv2 of overnight ICP recording is > 0.9808, then patient will respond to LD. 3.If percentage of RC1 of overnight ICP recording greater than 8.76 is < 41.48%, then patient will respond to LD. 4.If percentage of five minute segments of ICP having a standard deviation of RRC3k2 greater than 57.68 is > 0.49%, then patient will respond to LD drain.5.If the inter 95-5 quartile range of RCurvp2Curvp3 of overnight ICP recording is < 64.25, then patient will respond to LD.The accuracy for the corresponding features is: 70.4%, 72.2%, 74.1%, 72.2%, and 79.6% respectively.Finally, the OR combination of rules 4 and 5 achieves the best accuracy of 88.9%.

Discussion
We have demonstrated a systemic way of deriving two-rule combination of simple decision rules from overnight ICP recording to predict the corresponding outcome of three-day LD for patients with NPH clinical triad undergoing pre-shunt evaluation.Using a cohort of 54 patients, we are able to find optimal two-rule combination that reaches an accuracy of 88.9%.This rule can be explained in plain English such that it can be readily communicated to non-technical clinicians and patients.Although further validation with large patient cohort is needed, technically, the rules found in the present work in combination with an automated analysis of overnight ICP using the existing MOCAIP algorithm can readily be built into a decision support system of NPH diagnosis and evaluation that has the potential to reduce the time and cost associated with the existing three-day LD procedure.
A systemic way of discovering predictive rules is needed if a large number of ICP pulse morphological metrics are involved to construct rules.Compared to the existing studies that focus on either mICP or ICP pulse amplitude, the increased number of ICP metrics demands an objective way of discovering rules to guarantee optimality and avoid bias.Indeed, the five individual optimal metrics found in the present work do not include either mean ICP or ICP wave amplitude demonstrating the potential power of adopting a more comprehensive ICP pulse morphology characterization.
The proposed rule discovery framework can be easily extended in several aspects.We have proposed five simple feature functions to summarize an overnight ICP recording.More complex feature function can certainly be implemented in future work.In particular, we believe that feature functions characterizing the relationship among MOCAIP metrics could potentially offer better predictive power as compared to single-metric feature functions used www.intechopen.comReal-Time Analysis of Intracranial Pressure Waveform Morphology 121 here.Furthermore, we have only studied the combination of two rules, one could readily derive the equations to calculate the accuracy of combining multiple rules and check if the predictive accuracy can be further improved.
One fundamental limitation of the present work is the lack of cases to conduct independent evaluation of the discovered rules.It therefore remains to be demonstrated whether the same level of predictive accuracy can be retained when applying the rules discovered using our dataset to process data from other centers and how the data from multi-centers can be utilized to refine the rules.The present work has clearly demonstrated the technical feasibility of such future studies and hopefully more precise and robust predictive rules can be discovered through multi-center collaboration and the adoption of advanced ICP signal analysis and data mining methods.

Conclusions
Pulse pressure amplitude has been linked to several conditions with promising results in NPH for predicting shunt responsiveness [9,10].Work has also been done in chronic headaches; again, showing the usefulness of pulse pressure amplitude.Recent, detailed pulse pressure morphology analysis by our group has shown promise as a possible indicator of low global cerebral blood perfusion [11].The resistance from many clinicians and researchers is the unknown relationship between these pulse pressure metrics (MOCAIP) and their physiologic meaning.Therefore, the additional information provided by knowing the origins of these pulse pressure features would help bridge the gap between a strictly data mining approach and physiologic meaning.Finally, the importance of this work was confirmed in a recent review article by Wagshul et al.: "Given the success of invasive pulsatility measurements in clinical prognosis, studies which can provide a link between changes in pulse pressure and changes in non-invasive TCD-or MRI-based measures of pulsatility will be particularly valuable."[64] States , ti xx  and observations , ti y y  are represented in the graphical model and illustrated in Fig. 2 by white, and shaded nodes, respectively.Edges represent dependencies between states, and possibly observations, by two types of functions: observation potentials formulation and can be used by conditioning them in either directions during inference.By introducing compatibility potentials between states of the same peak at successive times name temporal potentials, the model becomes a dynamic Markov model.

Fig. 3 .
Fig. 3. Peak latency (left) and ICP elevation (right) estimated on ICP sequences by NBP tracking algorithm.The predictions of the tracking are obtained in real-time and are robust to transient perturbations that frequently occur during the ICP signal recording.
24 is 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-onset  at which the ICP segments the tested time-to-onset can improve the prediction performance of the models.To do so, the classifiers are trained such that the input i x of the positive examples are the s N segments prior to the onset plus the segment concurrent to the IH onset.Training of the models is performed for ten different lengths = {1, ,10} s N  , from 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 ={0, ,9}   .
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 ten ( =10 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 .
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 (reproduced with permission from[65]).
Average feature: This feature function simply calculates the average of each individual metric across one overnight ICP recording.Standard deviation feature: This function calculates the standard deviation of each individual metric across one overnight ICP recording.
The model consists of three distinct states 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 [1,3,6]s 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].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 single ( =1 [1,5,10]a-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]forsegmentsof[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]forsegmentsof[1,5,10]minute(s) length, respectively.The best average AUC for the linear model was observed using a single ICP segment =1 s N .In contrast, the best performance for AdaBoost and Extra-Trees models is obtained with =10 s N ICP segments, corresponding to ten minutes of ICP data extracted prior to the tested time-toonset.