A Novel Feature Extraction Method for Signal Quality Assessment of Arterial Blood Pressure for Monitoring Cerebral Autoregulation Pandeng Zhang12, Jia Liu12, Xinyu Wu12, Xiaochang Liu12，Qingchun Gao3 1
Shenzhen Institute of Advanced Technology, Shenzhen, China,
2
Chinese Academy of Sciences/The Chinese University of Hong Kong, HongKong, China 3 The Second Affiliated Hospital of Guangzhou Medical College, Guangzhou, China
[email protected],
[email protected],
[email protected],
[email protected],
[email protected]
Abstract—In this paper, we proposed a novel method of signal quality assessment of arterial blood pressure for monitoring Cerebral Autoregulation (CA). This method is based on algorithm of signal abnormality index (SAI). Two simple and effective features-end diastole slope sum (EDSS) and slow ejection slope sum (SESS), were proposed to identify abnormal beats from ABP as CA input in real-time. The methods of cumulative distribution function (CDF) and receiver operating characteristic (ROC) analysis were used to select best feature and confirm the parameter of the feature. Using the best feature with SAI model, we can directly estimate the signal quality of ABP in CA assessment. It has been tested in the data of CA assessment experiment and compared to an expert annotator, the algorithm’s sensitivity is 93.95%, and specificity is 84.87%. Keywords-arterial blood pressure; cerebral autoregulation; receiver operating characteristic; cumulative distribution function; signal quality
I.
INTRODUCTION
Cerebral autoregulation (CA) is a control mechanism that maintains cerebral blood flow at a relatively constant level despite the changes of arterial blood pressure (ABP) [1]. The status of this control mechanism has been shown to be correlated with the outcome of a number of cerebrovasular diseases including ischemic stroke, vasospasm secondary to subarachnoid hemorrhage， internal carotid artery occlusion, and etc. [2]. Therefore, the assessment of CA is of great importance in terms of the treatments of these diseases [3]. In order to assess CA, the control mechanism can be modeled as a single input-single output (SISO) system [4-7] where ABP and cerebral blood flow velocity (CBFV) are considered as the input and the output, respectively. Hence, the signal quality and model selection may determine the robustness of the assessment. In this work, we paid special attention to the signal quality of ABP (artifacts were frequently observed during patient movement and instrument calibration), as continuous assessment of CA has become a major focus of recent studies [8-11] and there is currently no specially designed method for real-time signal quality assessment (SQA) of ABP for the study of CA. Zong et al proposed a method of fuzzy logic to assess ABP signal quality in intensive care unit (ICU) to reduce the false alarm rate [12]. In their method, they used fifteen features of
ABP. Though the method was effective in terms of identification, the structure of it was complicated due to nested rules without detailed explanation, and it has six short-time averaged features which need a learning time more than 10 seconds to initialize. Sun et al proposed an algorithm with a simple structure named signal abnormality index (SAI) to assess the quality of ABP for estimating cardiac output [13]. The algorithm flags ABP beats by setting constraints on physiology, noise/artifact, and beat-to-beat variation values. However, this method was designed for estimating cardiac output in ICU. When we applied SAI to signal quality assessment of ABP for our study, we found that it could not identify the artifacts of square waves [14] generated during the instrument calibration, as the SAI features extracted from the artifacts and normal signals were similar. Therefore, we set out to investigate other possible features which may differentiate between the square waves and normal signals and thus could be used to improve the SAI algorithm. In this paper, we defined two simple and effective features named end diastole slope sum (EDSS) and slow ejection slope sum (SESS), based on the morphology and physiology of ABP. In order to compare the performance of these two features with features- maximum positive BP slope (MPPS) and maximum duration up-slope duration (MUSD) proposed by Zong et al, we applied cumulative distribution function (CDF)[19] and receiver operating characteristic (ROC) analysis[16] to carry out best feature selection and the parameter of the selected feature confirmation. Finally, we tested our methods in the data described below. II.
METHODS
A. Data Acquisition (Subjects and Measurement) The measurement was carried out at the Second Affiliated Hospital of Guangzhou Medical College. Ethical approval was received from the clinical research ethics committees/authorities of the hospital. The data from this experiment was originally designed for assessing cerebral autoregulation with entidal CO2 changes [15]. Seventy five subjects aged from 12 to 82 (53.28±17.54, 23 female and 52 male) were recruited for the experiment in China. The
978-1-4244-4713-8/10/$25.00 ©2010 IEEE
measurements were undertaken with subjects lying supine. CBFV was measured by using 2 MHz transcranial Doppler ultrasound (DWL MultiDopT) to insonate the middle cerebral artery. ABP was recorded non-invasively by a Colin tonometer. Data were recorded including 4 phases: rest (3-5 minutes), hyperventilation (2 minutes), rest again (6-10 minutes), hypercapnia by re-breathing, inhalation of previously exhaled gases (2 minutes). The re-breathing procedure may result in progressively decreasing concentrations of oxygen and progressively increasing concentrations of carbon dioxide in the blood. End-expiratory CO2 partial pressure (PCO2) was measured simultaneously by using a Capnograph (Hewlett Packard 47210A) connected to the face mask, and provided an estimate of arterial PCO2. The sampling rate of signals is 100 Hz. After data checking, four records eliminated due to short length. Seventy eight records were thus left from the experiment for our study. We defined a beat as normal: 1) clean waveform and clear features; 2) consistent with beat before and after in shape; 3) consistent with the referenced signal in timing and shape. According to this definition, all beats in these 78 records were annotated by two bio-signal experts. Excluding the normal beats, we obtained abnormal signals in our database. The main of them are as follows: 1) square wave; 2) signal before and after square wave; 3) movement artifact; 4) missing signal. See figure below.
as ejection (rapid and slow), rapid inflow, isovolumic relaxation, diastasis, and atrial systole. According to physiological knowledge above and the main problem of SAI, we proposed two new features after comparing the square wave artifact with normal signal in morphology. as:
The first one is end diastole slope sum (EDSS). We defined
EDSS =
∑ Δy
i
(1)
i:c → d
Where c is the position around
2 T (where T is period 3
length of a heart beat), d is the position at the onset of the next beat, Δyi = yi − yi −1 , y i is the ABP value at position i. EDSS represents arterial pressure gross change in the end period of diastole. The second feature is slow ejection slope sum (SESS). We defined as:
SESS =
∑ Δy
i
(2)
i:a → b
Where a and b is the position of systolic peak (SP) and dicrotic notch (DN) [17] of the ABP, respectively. SESS reflects arterial pressure gross change in the period of slow ejection. See schematic illustration below. In practice, we can (a)
choose b as the position around
1 T. 2
(b)
(c) Figure 2.
(d) Figure 1. Examples of main abnormal signals in ABP: (a)square wave, (b) signal before and after square wave, (c)movement artifact, (d)missing signal. The referenced signal is CBFV of left middle cerebral artery (lMCA) in the figure.
B. Proposed Features Using signals of the vital signs, including ABP, electrocardiograph (ECG), phonocardiogram (PCG), we can divide a cardiac cycle into several phases in physiology, such
Schematic illustration of period of slow ejection and end diastole
In Figure 2, we can easily find that EDSS or SESS has a very low value near zero where the beat is a square wave artifact. Both of them can distinguish normal signal and square wave. Besides, EDSS can identify part of movement artifact and missing signal, SESS can identify part of square wave before and after signal. C. Method of Feature Selection and Parameter Confirmation The procedure of method about feature selection and parameter confirmation is as follows:
•
To put the whole database as a test set, which was divided into two parts by experts, the part set of normal signal beats and the part set of abnormal signal beats;
•
To use the open-source onset detection algorithm and waveform feature extraction routine to compute the features of ABP;
•
To obtain statistics of features in two sets, and draw corresponding cumulative probability distribution[19] of them, respectively (see Figure 3);
•
To compute forecast value of sensitivity and specificity[16] with different values of candidate features to identify two sets, respectively: 1) MPPS, MUSD:
sensitivity = Pnormal,f = k
(3)
specificity = 1 − Pabnormal,f = k 2) EDSS, SESS:
sensitivity = 1 − Pnormal,f = k
(4)
specificity = Pabnormal,f = k
Figure 4. Schematic illustration of ROC analysis for feature selection and parameter confirmation, the curve is discrete for limitation of feature precision
D. Abnormal/artifact Identification Using New Feature Our approach was based on beat-to-beat waveform features. The process before signal quality assessment included a low-pass filter, beat detection using an open-source onset detection algorithm [18] and feature extraction routine. Features were then assessed by a series of criteria, including SAI abnormal criteria and our feature criteria. Finally, the output of all criteria were combined by the logical OR operation. See figure below.
Where P is the cumulative probability, f is the value of candidate feature. •
To draw receiver operating characteristic (ROC) curve, and carry out ROC analysis to obtain the best feature and its parameter (see Figure 4). According to the ROC analysis, we can obtain the best performance feature among these four features and its parameter with making area under the ROC curve in Figure 4 to be a maximum, distance between point p and the diagonal to be a maximum. abnormal set 1
0.9
0.9
0.8
0.8
0.7
0.7
cumulative probability
cumulative probability
normal set 1
0.6 0.5 0.4 0.3
Figure 5.
Block diagram. Input is an ABP, output is a binary string (normal=0, abnormal=1) to each beat of ABP
E. Algorithm Evaluation To verify the performance of our algorithm, we tested the effect in our whole records by comparing the algorithm’s performance with two experts in detecting abnormal beats. The algorithm should be in perfect consistent with the annotation of experts. III.
RESULTS
A. Result of Feature Selection and Parameter Selection
0.6
ROC 1
0.5 0.4
X: 0.2393 Y: 0.9577
0.9
EDSS SESS MPPS MUSD
X: 0.5039 Y: 0.9672
X: 0.2852 Y: 0.9515
0.3 0.8
0.2 0.1
0.7
0.1
0
200 feature x
(a)
400
0
0.6
0
500 feature x
1000
(b)
Figure 3. Example of cumulative probability distribution for feature x: (a) normal set cumulative distribution (b) abnormal set cumulative distribution, feature x could be anyone of four features mentioned above
sensitivity
0
X: 0.4476 Y: 0.8116
0.2
0.5 0.4 0.3 0.2 0.1
0
0
0.1
0.2
0.3
Figure 6.
0.4
0.5 1-specificity
0.6
0.7
ROC curve of features
0.8
0.9
1
From the figures above, we can find that EDSS is the best feature which has max area of the curve in figure below. Combined the corresponding cumulative probability distribution, we can obtain the best parameter of each feature. The parameters of EDSS, SESS, MPPS, and MUSD are 6.1, 16.2, 19.35, and 151 respectively. B. Result of Artifacts Identification Table I shows performance of different methods in whole records. Note that SAI add SESS is the best combination which has the same effect of the Zong’s method. TABLE I.
PERFORMANCE OF DIFFERENT METHODS IN WHOLE
[2] [3]
[4] [5]
[6]
[7]
RECORDS
Methods
N Sensibility Specificity (Beat) (%) (%) Zong’s method 128342 97.45 82.77 SAI 128342 97.75 18.38 SAI+MUSD 128342 79.43 65.17 SAI+MPPS 128342 93.11 82.48 SAI+SESS 128342 94.99 60.20 SAI+EDSS 128342 93.95 84.87 |MUSD|>151, |MPPS|>19.35, |SESS|
