Study On The Prediction Of Intracranial Hypertension Based On Waveform Feature Extraction And Support Vector Machines Classification

In neurological settings, intracranial hypertension (ICH) is a commonly happenedcritical condition. ICH can cause the decrease of cerebral perfusion pressure (CPP) andreduce the cerebral blood flow (CBF), which leads to brain hypoxia-ischemia and evenbrain herniation. Normally when the ICH is observed, the nursing procedure can bestarted. Due to that the ICH may not be found timely, it usually progresses rapidly anddrugs take effect with a long time, the prompt and effective treatment that can stopprogress or shorten the duration and prevent complications from an already existingdisease process may be delayed.In current clinical settings, there is short of effective alarm device for predictingthe occurring of ICH. This kind of device can warn the nursing staff the happening ofICH and let them be well prepared for preventing the increase of intracranial pressure(ICP) and secondary brain injury, improving the clinical outcome of the treatment,preventing the happening of brain herniation and lowering the mortality. Thus, anautomatic ICH alert algorithm is of great value for the ICP management and its nursingcare.With the development of the signal processing technology, many ICH forecastingmethods have been proposed to alarm the ICH in advance. The conventional methodsinclude the setting up of a threshold and the independence hypothesis of the ICP signal.From90s, the autocorrelation of the ICP signal is taken into consideration and timeseries regression model is employed to deal with this problem. With the success of theapplication and development of machine learning technology in many fields, severalresearchers adopt the machine learning technology in the prediction of ICH in recentyears, e.g., the artificial neural network. And researchers combine several advancedsignal processing methods to attack this problem, such as the wavelet analysis, kalmanfiltering and approximate entropy. In these methods, the ICP signal is averaged and themean value is considered as the predictor. It ignores the dynamic characteristic of thesignal during the cardiac cycle. In addition, the ICP signal is a nonlinear andnonstationary process and the fundamental knowledge of the mechanism ofcerebrovascular regulation is insufficiency, which makes the above models and methodsnot good enough and not adopted in clinical practice. However, recent research findings have shown that the configuration of thecharacteristic peaks and other waveform features of the pulse waveform have strongpositive correlation with the neurological conditions as well as the evolvement of theICP. They also can disclose important information of intracranial environment, such asintracranial compliance (IC) and cerebral autoregulation (CA).In view of the high relation between the waveform feature and the brainpathophysiology, this paper proposes a novel scheme to judge if the ICH will occurafter5minutes. The procedure of forecasting the ICH of the proposed method is asfollows.①The continuous ICP signal is segmented to individual pulse by a new pulseonset detection algorithm.②The three characteristic peaks are identified by a peakrecognition algorithm and the morphological feature metrics are extracted.③Aclassification system is constructed based on the support vector machines (SVM). Thefeature metrics is the input of the classification system and the classification resultscorrespond to ICH/non-ICH. Specifically, the system first employs a global searchalgorithm, deferential evolution (DE) algorithm, to select the optimal featurecombinations and Wrapper scheme to perform the optimal feature selection. Then theselected optimal feature combinations are classified by the classifier SVM. The outputof the classifier corresponds to the ICH or non-ICH group.The main research findings of this paper are as follows.First, a novel algorithm is proposed to detect the pulse onset of continuous ICPsignal. There are on other methods existing to detect the pulse onset of continuous ICPsignal. Following the idea of a well known algorithm in image processing field, shapecontext, a descriptor, waveform descriptor, to describe the points on one-dimensionalphysiological signal is constructed. Then the waveform descriptor is employed toextract the feature of given points and the features are compared with a customizedtemplate and finally the onset whose feature is most similar with the template isidentified. After the onset is identified, the continuous ICP signal is segmented intoindividual pulse.Second, a novel algorithm to recognize the characteristic peaks of ICP pulses isproposed. There are on other methods existing to recognize the characteristic peaks ofeach individual ICP pulse. The algorithm first employs the waveform descriptor toextract the features of the points on the ICP pulse, and then utilizes a classifier, SVM, toclassify the features and the three characteristic peaks of ICP pulses are identified. Finally, the morphological features such as latency, slope and peak-to-peak amplitudeare extracted.Three, an ICH forecasting algorithm is proposed based on optimal feature selectionand SVM classification. For a given signal, the algorithm aims to discriminate betweenthe pre-ICH segment and stable segment5min before the occurring of the ICH. Thealgorithm is able to judge if the signal is ICH before5minutes to happen. It firstemploys the DE algorithm to select the optimal feature sub-set, Wrapper method (DEalgorithm for feature selection, SVM for classification and the mean value of thesensitivity and positive predictive value as the objective function) to evaluate theperformance of the selected sub-set and determine the optimal feature set and then usesSVM to classify the optimal feature to different classes. The output of the classifiercorresponds to the pre-ICH segment (ICH) or stable segment (non-ICH).Four, one important feature of the proposed ICH forecasting system is theclassification way is adopted, rather than the prediction of the future value of the ICPsignal. For a given signal, the algorithm can discriminate between the pre-ICH segmentand stable segment. Once it is regarded as stable segment, it means that the ICH will nothappen in5minutes. Otherwise, once it is regarded as pre-ICH segment, it means thatthe ICH will happen in5minutes. The classification system sets up some directconnection between the morphological features and ICH, forecasts the ICH by detectingthe change of the morphological features. It provides a new window to explore therelation between the morphological features and neurological conditions.Finally, the proposed method is validated. First, the onset and three characteristicpeaks are annotated manually and the onset detection and peak recognition algorithmare performed, respectively. The performance of these two algorithms is validated usingquantitative evaluation criteria. The ICH is forecasted by the SVM classification systemand the system performance is evaluated by control experiment. The sensitivityachieved by the system is84%and the specificity is96%. To our knowledge, this studyis the first attempt of ICH forecastion with a time span of5minutes. The preliminaryresults indicate its effectiveness and the potential practicality and provide experimentalfoundation for the clinical application.