Multistage fusion approaches based on a generative model and multivariate exponentially weighted moving average for diagnosis of cardiovascular autonomic nerve dysfunction

Accepted Manuscript

Multistage fusion approaches based on a generative model and multivariate exponentially weighted moving average for diagnosis of cardiovascular autonomic nerve dysfunction Mohammad Mehedi Hassan, Shamsul Huda, John Yearwood, Herbert F. Jelinek, Ahmad Almogren PII: DOI: Reference:

S1566-2535(17)30479-7 10.1016/j.inffus.2017.08.004 INFFUS 891

To appear in:

Information Fusion

Received date: Revised date: Accepted date:

31 December 2016 1 July 2017 3 August 2017

Please cite this article as: Mohammad Mehedi Hassan, Shamsul Huda, John Yearwood, Herbert F. Jelinek, Ahmad Almogren, Multistage fusion approaches based on a generative model and multivariate exponentially weighted moving average for diagnosis of cardiovascular autonomic nerve dysfunction, Information Fusion (2017), doi: 10.1016/j.inffus.2017.08.004

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Highlights • Multistage fusion approach for diagnosis of autonomic nerve dysfunction • Independent Component Analysis and statistical process control are used for fusion

CR IP T

• Body sensor data from ECG and blood chemistry are used for fusion approach • Decision fusion has been proposed for diagnosis by using a multi-classifier system

AC

CE

PT

ED

M

AN US

• Proposed fusion approach achieves high performance for diagnosis of nerve dysfunction

1

ACCEPTED MANUSCRIPT

Multistage fusion approaches based on a generative model and multivariate exponentially weighted moving average for diagnosis of cardiovascular autonomic nerve dysfunction Mohammad Mehedi Hassanb,∗, Shamsul Hudaa , John Yearwooda , Herbert F. Jelinekc , Ahmad Almogrenb a School of IT, Deakin University, Australia of Computer and Information sciences, King Saud University, Riyadh 11543, Saudi Arabia c school of community health, Charles Sturt University,NSW, Australua

CR IP T

b College

Abstract

PT

ED

M

AN US

Like many medical diagnoses, clinical decision support system (CDSS) is very essential to diagnose the cardiovascular autonomic neuropathy (CAN). However, diagnosis of CAN using traditional ‘Ewing battery test’ becomes very difficult due to the inherent imbalanced and incompleteness condition in the collected clinical data. This influences the health professionals to investigate other related diagnostic reports of patients, including Electrocardiogram (ECG) data from ECG sensors, blood chemistry, podiatry and endocrinology features. However, additional components increase the dimensionality of the feature set as well as its heterogeneity and modality in the clinical data which may limit the applications of traditional data mining approaches for an accurate diagnosis of CAN dysfunction in the CDSS. To address the aforementioned problem, in this paper, we have proposed, a novel multistage fusion approach based on a generative model and a statistical process control (SPC) technique to diagnose CAN more accurately. Proposed approach develops two different generative models by using a shared and a separated Independent Component Analysis (ICA) to overcome the incompleteness and modality of the data. Due to the heterogeneous and non-normality features, statistical correlations and multivariate control limits in relation to the CAN diagnosis parameters are determined by fusioning of a series of exponentially weighted moving average (MEWMA) control processes. Fusioned features from both component analyses and SPC are applied in an ensemble classification system. Proposed multistage fusion approach is experimentally verified to justify its performance by using a large dataset collected from the diabetes screening research initiative (DiScRi) project at Charles Sturt University, NSW, Australia. Our comprehensive experimental results show that proposed fusion approach performs better than standard classifier for both ‘Ewing’ feature set and ‘Ewing and additional feature set’ with significant improvement in accuracy.

AC

CE

Keywords: Fusion of multiple statistical process control techniques, Multivariate Exponentially Weighted moving average, autonomic nerve dysfunction classification, blind source separation, fusion of features and decisions 1. Introduction Body Sensor Networks (BSNs) [17], [18] is a specific class of wireless sensor networks which are ∗ Corresponding

author Email addresses: [email protected] (Mohammad Mehedi Hassan), [email protected] (Shamsul Huda), [email protected] (John Yearwood), [email protected] (Ahmad Almogren) Preprint submitted to Elsevier

emerging as a noteworthy unobtrusive technology to collect and process different vital signs of a patient for the purpose of managing chronic diseases and detecting health anomalies. BSNs are typically equipped on the human body as tiny patches or hidden in users’ clothes or even implanted in the human body [18]. These sensors have the capability to collect real-time data of various physiological parameters (e.g. heart rate (HR), the rate of breathing (RR), blood pressure (BP), pulse, body temperAugust 4, 2017

ACCEPTED MANUSCRIPT

not be suitable for elderly patients who are hard to diagnosis either due to the insensitive response to the tests or having an impaired mobility. This may lead to incomplete datasets of clinical CAN diagnosis feature and can affect the performances of CDSS for CAN diagnosis. Researchers are investigating complementary features including ECG and blood chemistry that may help to overcome these aggravating test conditions. However, additional features pose data analysis challenges for CDSS including high dimensionality, heterogeneity, multimodality to some extent and incompleteness in the data. In this paper, we address this data analysis challenge in CDSS for CAN diagnosis to achieve a high performance detection of CAN. High dimensionality have been addressed in CAN diagnosis with a small data sets (only 291 patients) in 2010 by the co-authors [22] using a wrapper-filter approach which achieved up to 82% accuracy using ‘Ewing’ features. Relevance of blood chemistry with CAN have been recently studied in a number related researches [38] , [14], particularly for glucose level among diabetes patients. Relationship with lipid profile with CAN have also been studied recently in several article in [34] and [33]. These studies [38] , [14], [34] and [33] show that blood chemistry has strong relationship with CAN. However recent studies [38] , [14], [34] and [33] are limited to an independent parameter based study and did not consider their combined effect including all CAN features. Also earlier work [22] uses repeated evaluation of Artificial Neural Network (ANN) [21] in a backward elimination process which increases the computational complexity [26] at a very high level due to the training time of ANN [21] and is not suitable for a large scale datatsets. Therefore, an extensive study with the combined challenge of incompleteness, heterogeneity and modality including an additional blood chemistry feature with a large number of patients is of utmost importance in order to develop a robust and high performance CDSS for CAN diagnosis. This the main focus of this paper. In this paper, we propose a multistage fusion approach through the fusion of an independent component analysis (ICA) based generative model and multivariate exponentially weighted moving average (MEWMA) based SPC technique. Two different generative models have been developed using a shared ICA and separated ICA of CAN features. Then the extracted components have been passed through a multilevel fusioned MEWMA processes.

AC

CE

PT

ED

M

AN US

CR IP T

ature, blood oxygen intensity (SPO2) and electrocardiogram (ECG) ) [17], [18], [23]. The monitoring of patient health conditions helps in preventing terminal illness, monitoring the progression of chronic disease, and enhancing emergency services, especially for elderly and physically impaired people [1]. The data from sensor nodes are collected at local personal devices such as mobile phone and PC. These data are then can be used for real-time monitoring and long term remote storage for diagnostic analysis. Often chronic diseases such as cardiovascular autonomic neuropathy (CAN) is very hard to determine if it is not monitored carefully at early stages. In these cases, BSN are very useful for collecting patients’ physiological data which can be later used for analysis and detection [18], [23]. Cardiovascular autonomic neuropathy (CAN) [15], [4] is directly associated with cardiac arrhythmia and many cardiovascular related diseases which increases the unexpected death rate [8], [11]; particularly for diabetes patients. Therefore appropriate monitoring and early detection of CAN is very essential in clinical diagnostic and management systems. Often, clinical decision support systems (CDSS) are very helpful for appropriate management of chronic disease such as diabetes and CAN for accurate monitoring and early diagnosis of diseases. However, data analysis and intelligent processing of medical data in CDSS are quite challenging. Since many medical data are heterogeneous, multimodal, imbalance and high dimensional due to the complexity in the data collection and production processes from their sources in the medical systems. The complexity may arise from many different reasons including cost-sensitiveness, high risks with side effects of diagnostic tests and operating dangerous equipments for which diagnosis tests are not always completed for the patients unless it is strictly required. This can lead to an incompleteness in the collected data [36], [36] which makes the diagnosis of the diseases difficult for health professionals and data analysis tasks in CDSS. This in turn may results into a complete failure or an inaccurate diagnosis of the diseases. Similar to other medical data, successful diagnosis of CAN using conventional ‘Ewing battery tests’ [15], [4] is complex and dependent on the capability of the patients to undergo all the tests. Often many patients are unable to go through all of the ‘Ewing’ tests; for example, one of the tests require a movement of patients from a position ‘Lying’ to ‘standing’ or vice versa. Some other tests also may

3

ACCEPTED MANUSCRIPT

CR IP T

the components of CAN and used in a ensemble classification. A large dataset of CAN from the diabetes screening research initiative (DiScRi) project at Charles Sturt University, NSW, Australia has been used to justify the performance of the proposed multistage fusion framework. The rest of the paper is organized as follows. Section 2 discusses related work in CAN and different existing techniques for CAN identification. Section-3 explained the proposed multistage fusion approaches. Description of the data collection method and method of pre-processing data have been discussed in subsection 3.1. Section 3.3 and Section 3.4 describe the proposed two multistage fusion models including the fusioning of MEWMA processes, fusion of feature and decision models. The experiment results are presented in Section 4. Conclusions from this study and references are presented in the last two sections.

AN US

Figure 1: QRS complex in a heart beat’s, an electrocardiogram (ECG)

Identified upper control limits with patients’ multivariate characteristics features by SPC are fusioned with the components of original CAN feature. These features are applied on a ensemble classifier for CAN classification. The novelties and contributions of the proposed approach are described below which include:

2. Related work

AC

CE

PT

ED

M

1. A novel multistage fusion approach has been developed using a generative model and multivariate exponentially weighted moving average for CAN diagnosis. 2. An unsupervised statistical model has been developed to determine the multivariate corelations and corresponding statistical upper control limit in order to distinguish the incontrol and out-of-control patients by fusioning a series of MEWMA processes. 3. A feature based fusion and ensemble decision model has been developed by using the independent component analysis (ICA) and MEWMA to minimize the effect of nonnormality, heterogeneity, high dimensional and multimodal challenge of CAN data.

ICA based generative models successfully identify sources of input features using the inherent blind source separation technique which also deals with the high dimensionality. MEWMA has been used to identify multivariate co-relations among heterogeneous data which is also able to identify joint upper control limit of the CAN parameters through the fusion of a series of MEWMA with varying average run length (ARL). Identified multivariate characteristics from MEWMA process are fusioned with 4

CAN is a complication of diabetes mellitus which involves a severe damage to the autonomic nerve fibres and is directly associated with increased levels of systemic inflammation and a high risk of cardiovascular disease [8], [11]. Conventional method of CAN diagnosis requires five simple autonomic function tests known as ‘Ewing battery tests’ [15], [4]. The tests include the measurements of variations in the heart rate (HR) and blood pressure (BP) for different situation while patients perform either a movement of body from one position to another or perform some specified activities. These measurements are: 1) Changes in HR while patients attempt an exhalation against a closed airway (valsalva manoeuvre) 2) HR variation during deep breathing, 3) HR variation from lying to standing 4) variations of BP from lying to standing 5) variations of BP associated with hand grip. In [15], [4], test results of Ewing are categorized into three different ranges which are abnormal, normal and borderline. Then CAN is determined based on how many of the test results goes either borderline or abnormal which is used to classify CAN into normal, mild, moderate, severe and atypical. Often all tests of CAN may not be possible to undertake by the patients due to elderly situations such as inability in performing handgrip appropriately for patient with arthritis or changing position from lying to standing due to mobility problems. Recent studies show that [38] , [14], [34], [33] many other

ACCEPTED MANUSCRIPT

CR IP T

Figure 4: QRS complex detection from an ECG signal

AN US

Figure 3: RR Interval in a QRS complex

Ensemble Decision fusion layer

ED

Multivariate Analyses: MEWMA

M

Fusion of Features

Generative Model: Component Analyses

PT

Feature selection and fusion layer

CE

Feature extraction

ECG sensors and blood samples

AC

Sensing Layer

Figure 2: A three layer architecture of multistage data fusion approach for cardiovascular autonomic neuropathy diagnosis

5

regular diagnostic tests which are accomplished as part of diabetes or cardiovascular management can be useful in CAN diagnosis. In [14], Rumyana et. al. finds a high prevalence of CAN with higher glucose level. Melissa et. al.[13] show that impaired glucose tolerance (IGT) represents a sign of early stage of CAN. Catherine et. al [33] measured the cardiac autonomic nerve function at the Pittsburgh Epidemiology of Diabetes Complication and find that CAN is associated with arterial stiffness measures and high-density lipoprotein cholesterol(HDL) cholesterol. Catherine et. al [33] also find that HDL and triglyceride(TG) are associated to the prevalence of CAN in a Chinese population sample. Addition of ECG and blood chemistry [33], [14] data could be potential direction of CAN diagnosis. However, this increases the heterogeneity and dimensionality in the data. In [39], statistical process control techniques (SPC) are also used in monitoring the patients. Akshay Sood et. al [37] applied SPC chart to monitor the safety of patients at the training stage of robotic kidney transplantation (RKT). A systematic review of the application of control chart on a biomedical process has been presented in [39]. In [20], Hirotsugu Hashimoto et. al used multivariate SPC chart for epileptic seizure monitoring based on heart rate variability (HRV). Suresh Pujar et. al [35] demonstrated SPC was an essential clinical tool to asses the effectiveness of drug intervention and variability in seizure frequency. Multimodal analysis of medical data using blind

ACCEPTED MANUSCRIPT

els, ICA and MEWMA processes where explanatory source variables are applied for determining the outof-control observations in the MEWMA subsystem. Both ICA and MEWMA are unsupervised [28] and therefore takes the benefit of any class imbalanced situation of CAN data or incompleteness due to the mobility challenge of elderly patients while extracting statistical correlations and multivariate upper control boundaries of the CAN parameters. The framework for proposed multistage fusion has been presented in Fig. 2. The proposed fusion approach are described in the following subsections.

CR IP T

source separation (BSS) has been used in many CDSSs which also can deal with the dimensionality and existing heterogeneity with different domain value ranges. Christoph Hoog Antink et. al [3] have used fusion technique with different sensors data obtained from ballisto-cardiographic signal obtained from the seat of a chair to estimate beat to beat heart rate. Fusion of different machine learning (ML) techniques [5], [40], [9] have also been applied in complex classification and prediction problems. For example Jigar patel et. al [30] used fusion of ML techniques for a stock market prediction problem. A recent review article [19] clarifies the significance of fusion techniques from sensor data in physical activity recognition. BSS based data fusion techniques [12] have been applied for complex EEG and fMRLI signals analysis for a lower-dimensional representation of selected brain activity. Liu J et. al [24] used an Independent Component Analysis (ICA) based feature-level fusion for identifying relationships between neuroimaging data types. In [29], C. Orphanidoua et. al. used data fusion technique with ECG for estimating respiratory rate. Data fusion techniques [32] also used to minimize the trade-off for heterogeneity between different user and related metrics in a sensor network dataset. In [10], ML methods including support vector machine (SVM) [10] also have been used to distinguish diabetics and severeness of CAN.

ED

M

AN US

3.1. Data collection method and datasets The dataset was collected and prepared by coauthor Herbert Jelinek, at Charles Sturt University, NSW, Australia. The data was collected from a cohort of patients and healthy people who attended the diabetes complications screening research initiative (DiScRi) project located at Charles Sturt University, NSW, Australia. The ethics application was submitted to the university and approved by the ethics in human research committee of Charles Sturt University before the start of the project. A media advertisement was made at the local community to invite the participants and to attend for screening in which participants were requested to abstain from smoking, alcohol and coffee for twenty four hours before they attend at the clinic. They were also requested to fast from midnight till the test completed while the different clinical measurements were conducted from 9:00 AM to 12:00PM. Tests were conducted on a number of days and a total of 1167 participants attended at the clinic. Two sets of measurements were taken from the participants. First set was blood chemistry which is the cholesterol profile and blood glucose level (BGL). The second set was ‘Ewing tests’ [15] which consists of five tests and their electrocardiogram (ECG) report. The five Ewing tests are defined based on the changes in heart rate and blood pressure of patients for certain positions. These include changes in heart rate and blood pressure when a person changes his position from lying to standing, changes in heart rate for deep breathing and valsealva manoeuvre, and changes in blood pressure associated with hand grip.

PT

3. Proposed Methodology:Multistage fusion approach based on a generative model and multivariate process control technique

AC

CE

SPC [20],[35] techniques are used to determine the quality of a process by using the distribution of the quality characteristics in many multivariate processes [39], [7]. This also can be used to monitor the bio-medical processes. SPC approach such as multivariate exponentially weighted moving average(MEWMA) charts can find any abrupt change or variations in the observed medical data, at the same time can evaluate unanticipated aberrations in the data. Changliang ZOU et. al. [39] also have shown in their paper and stated: ‘in many situations, the quality of a process may be better characterized and summarized by the relationship between the response variable and one or more explanatory variables’. This theory of process control [39] influences us to employ fusion of two generative mod-

3.2. ECG Feature extraction and data preprocessing Collected electrocardiogram(ECG) have been further pre-processed for feature extraction. An 6

ACCEPTED MANUSCRIPT

Ewing battery

Blood

Generative Model: Component Analyses Components

Fusion Level‐1

CR IP T

ECG

Components

Multivariate Analyses: MEWMA In Control/out  control statistics

Fusion Level‐2

AN US

Fusion of features

Fusion of Classification (Ensemble)

Fusion Level‐3

Figure 5: Multistage fusion fusion architecture based on shared independent component analyses and multivariate exponentially weighted moving average (SCA-MEWMA)

ED

M

electrocardiogram (ECG) records the heart’s activity in terms of the electrical manifestation of the heart muscles movements. ECG can be represented as a set of twelve views of the electrical impulses generated by the heart muscles. Among this twelve views, six left-half of the ECG are: (I, II, III, aVR, aVL, and aVF) which are produced through the electrodes placed on the arms and legs of human body. The rest six views are taken from the right half (V1 through V6) which are produced by the electrodes placed on the chest.

as in Fig. 3. Heart rate is calculated from the R-R interval.

PT

A common structure of detecting the QRS complex (developed by Pan and Tompkins, 1985) [25] from ECG is presented in Fig. 4 which has been followed in the proposed approach. The process uses bandpass filter, differentiator, squaring function, moving integral component and decision component to find the QRS complex. Please refer to (Pan and Tompkins, 1985) algorithm for detail in [25]. The morphological features were collected from the 12-lead ECG, including QRS, PQ, RR, ST, QTc and QTd intervals, heart rate and QRS axis. The extracted blood chemistry, ewing tests and ECG features are listed in the following two tables. Table 1 describes the Ewing tests and ECG features and Table 2 presents the blood chemistry (glucose and cholesterol) features. In Table 1, ‘LABPI’ refers to Ankle brachial pressure index is determined by blood pressure in ankle divided by blood pressure in arm indicates vascular disease and ‘R-ABPI’ refers to the same as ‘L-ABPI’, but for right leg.

AC

CE

Each heart beat forms a complex waveform which is known as a QRS complex as presented in Fig. 1. A QRS complex has three components (P wave, QRS wave and T wave) produced due to the different parts of heart’s activities. P Wave is generated due to atrial depolarisation, QRS wave is generated due to ventricle depolarisation and T wave is generated due to ventricles repolarisation. The segment of an ECG between the QRS complex and the T wave is defined as an ST segment. The segment from the beginning of the P wave to the beginning of the QRS complex is defined as a PR interval. The maximum amplitude of a QRS wave is known as a peak R point and the duration between two consecutive R points is defined as a R-R interval 7

ACCEPTED MANUSCRIPT

Table 1: Features collected from ewing tests and ECG

AN US

QRS 10sec

Description Systolic blood pressure measured for lying patient. Diastolic blood pressure measured for lying patient. The average of ‘L-ABPI’ and ‘R-ABPI’ Heart rate change due to lying to sanding Heart rate change due to deep breathing Heart rate change due to valsalva manoeuvre Heart rate change due to Standing to lying Blood pressure change due to hand grip Blood pressure change due to lying to standing PQ is interval on ECG time for electrical conduction between point P and Q on 10 second recording. The width of the QRS interval in milliseconds long interval over 100msec is abnormal and shows bad conduction through ventricle. Corrected QT interval. QT dispersion, i.e., interval differences between recording leads. Axis of QRS shows abnormal conduction of electrical impulses.

CR IP T

Extracted feature from ewing tests and ECG Lying SBP average Lying DBP average ABPI average LSHR DBHR VAHR SLHR HGBP LSBP PQ 10sec

ED

M

QTc 10sec QTd 10sec QRS axis (degree) 10sec

Ewing battery

PT

ECG

CE

Generative Model: Component  Analyses

Blood Generative Model: Component  Analyses Components

AC

Components

Multivariate Analyses: MEWMA

Fusion Level‐1

In Control/out  control statistics Fusion of features

Fusion Level‐2

Fusion of Classification (Ensemble)

Fusion Level‐3

Figure 6: Multistage fusion architecture based on separated independent component analyses and MEWMA

8

ACCEPTED MANUSCRIPT

Table 2: Features collected from patients’ blood chemistry

Description Total cholesterol in blood. The level of triglyceride in blood. High density lipoprotein in blood. Low density lipoprotein in blood. Ratio of total cholesterol to high density lipoprotein. Blood glucose level

3.3. Proposed fusion architecture-1: Fusion of shared component analyses and multivariate exponentially weighted moving average (SCAMEWMA)

below:

CR IP T

Extracted feature from blood TC(mmol/L) Triglyceride(mmol/L) HDL(mmol/L) LDL(mmol/L) TC/HDL ratio Glucose(mmol/L)

x1i = a11 s1i + a12 s2i + .... + a1m sm i

x2i = a21 s1i + a22 s2i + .... + a2m sm i .. .

In this fusion strategy, an independent component analysis(ICA) [41], [31] based generative model is fusioned with a series of multivariate exponentially weighted moving average subprocess. The architecture for a three layer multistage fusion approach based on a shared-ICA generative model and MEWMA is presented in Fig 5.

AN US

xvi = av1 s1i + av2 s2i + .... + avm sm i

Which can be written in a matrix from as X = AS

ED

M

3.3.1. A generative model:independent component analysis(ICA) Finding the relationship between different data modalities can be accomplished by transforming data and then by increasing the homogeneity of within patients and modalities and decreasing the components that minimally contributed to the disease. Independent Component Analysis (ICA) [41] are used for this purpose which can find a linear representation of non-gaussian data so that latent components become statistically as independent as possible [41] in a form of source variables. ICA [41] was originally proposed for a problem similar to the ‘cocktail-party’ problem. In the ‘cocktailparty’ problem, many people speaks simultaneously in which multiple microphones record mixtures of the speaker’s voices. However, ICA can separate mixed voices into the voice of an individual. We have proposed a generative model based on ICA. Let us consider that a set of latent non-gaussian T source components as s = {s1i , s2i , s3i ....sm i } of the multimodal data of CAN patients dataset X = {x1 , x2 , x3 ....xn } which has the diagnosis label as y = {y1 , y2 , y3 ...yn }. Here xi is the observed component of the data for a patient and i = 1, 2, 3, ..., n = total patients. A generative model of the data can be represented based on ICA as

(1)

Let us assume observation variable is: Xe = {xei1 , xei2 , xei3 ....xeiv }T for ECG and observation variable is: Xb = {xbi1 , xbi2 , xbi3 ....xbiu }T for blood chemistry. In the proposed method, for the shared sources generative model, observed components with different modalities and patients are fusioned as below, Xeb = {xei1 , xei2 , ...xeiv , xbi1 , xbi2 , ...xbiu }T

(2)

CE

PT

Then the sources can be computed using the inverse of A where A is the mixing matrix and W is the un-mixing matrix. S eb = A−1 Xeb

(3)

S eb = WXeb

(4)

where shared source components are computed as:

AC

T S eb = {s1i , s2i , s3i ....sw i }

(5)

3.3.2. Statistical process control (SPC): Fusion of a series of multivariate exponentially weighted moving average (MEWMA) In the shared analysis based fusion architecture (Shared-ICA), once the common sources of data S eb are computed, it is fed to a multivariate statistical process control monitoring subsystem to monitor the out-of-control observations using a multivariate exponentially weighted moving average (MEWMA) 9

ACCEPTED MANUSCRIPT

technique [7]. The monitoring quantity Ti2 in the chart is computed as: Ti2 = Z´i Σ−1 (6) zi Zi

3.4. Proposed fusion architecture-2: Fusion of an individual component analyses and MEWMA (ICA-MEWMA)

where Zi is computed from the sources variables Seb and i is the index of patients in the training data. Zi = λseb (7) i + (1 − λ)Zi−1 = 0

In the second fusion architecture, the sources component of observation variables from two different datasets are computed separately. The architecture for a three layer multistage fusion approach based on a separated-ICA generative model and MEWMA has been presented in Fig 6.

CR IP T

The Σzi is the the covariance matrix of Zi and defined as follows λ [1 − (1 − λ)2i ]Σ > L (8) Σzi = 2−λ Where λ = diag(λ1 , λ2 , λ3 ....λm ) and 0 < λ ≤ 1 and Σ is the covariance matrix which is computed from the given in-control data. The upper control limit (UCL) L is computed for a specified in-control limit.

e S e = A−1 e X

(12)

S e = We Xe

(13)

b S b = A−1 b X

(14)

S b = Wb Xb

(15)

The joint components from two ICA processes are applied in the MEWMA process where the plotted quantity Ti2 is determined based on the new sources as below and equation (16).

AN US

AC

CE

PT

ED

M

3.3.3. Fusion of multiple MEWMAs to compute the final value of out-of-control measurement: The control status of each patients for a particular ARL is computed by using following function in equation (9) where ‘False=in-control (0)’ and ‘True=out-control(1)’. ( 0, if Ti2 < U CL eb M EW M AControl(si ) = 1, otherwise (9) The ARL of MEWMA process is varied for a range of ARL values ARLj . Then Ti2 is computed to find out the in-control and out-of-control patients for each of the ARL values (ARLj >= 200 and ARLj <= 700 , j = 1, 2...J). A voting mechanism is applied on the values of Ti2 for each patient as following in equation (10).    J  X    if  (Tij2 < U CL)     j=1  0,   J M EW M AF inal(seb ) = i X   >  (Tij2 > U CL)     j=1    1, otherwise (10) Here ‘0=in-control’ and ‘1=out-control’ of each patient’s final measurement which is determined using the above voting process. Average value of Ti2 is also calculated as below (Ti2Average ): Ti2Average =

J 1X 2 T J j=1 i

(11) 10

Zi = λsi + (1 − λ)Zi−1 = 0

(16)

where sources is S = {S e , S b }. Then fusion of MEWMA precesses follows equation (10). Fusion of components and MEMWMA decision follows equation (11) and equation (10). 3.5. Fusion of features and decision using ensemble classifier At this stage, features collected from the generative model and statistical process controls are fusioned and passed to the classifier in order to identify whether a participant has a ‘CAN’. We proposed a multi-classifier ensemble of decision fusion approach. Often a multiple classifier system performs better than a single classifier system which can be more robust [27]. A general structure of an ensemble classifier from different classifier’s decisions is presented in Fig. 7. In an ensemble classifier, a set of classifiers Cl where l = 1, 2, 3.. are constructed from the training data. Then the test data are classified by aggregating the decisions made by multiple classifiers. Here a support vector machine (SVM) [2] classifier has been used as a base classifier. Let us assume that the fusioned feature eb 2 Φi = {seb i , M EW M AF inal(si ), TiAverage } for proposed fusion architecture-1. We have used an ‘AdaBoost’ [27] ensemble classification. In ‘AdaBoost’ [27], a set of classifiers Cl are constructed where

ACCEPTED MANUSCRIPT

CR IP T

Figure 7: Decision fusion using ensemble

l = 1, 2, 3.., the error rate l for a classifier Cl is computed for fusion architecture-1 as below: n

l =

1X weighti f (Cl (Φi ) 6= yi ) n i

For ICAs of blood chemistry, all six components have been considered. The computed ICAs of two sets of features (Ewing plus ECG features, blood chemistry ) are presented in Fig. 11 and Fig. 12. Fig. 11 presents the ICAs of separated features (ewing plus ECG features) and Fig. 12 presents the ICAs of separated features for blood chemistry. Normality tests for ICAs from both feature sets of Fig. 11 and Fig. 12 have been accomplished and then it was transformed to normal distribution using ‘Johnson Transformation’. Examples of normality tests and corresponding transformation for component-3 of blood-ICAs are presented in Fig. 13 and Fig. 14. Examples of normality tests and corresponding transformation for component-7 of Ewing-ECG-ICAs are presented in Fig. 15 and Fig. 16. For shared ICAs of Ewing plus ECG features, total of 16 components has been taken. The computed ICAs of shared features (ewing plus ECG features and blood chemistry ) are presented in Fig. 17. Normality tests for ICAs from shared feature sets have been accomplished and then it was transformed to normal distribution using ‘Johnson Transformation’. Examples of probability plot and corresponding transformation for component-3 of shared ICAs are presented in Fig. 18 and Fig. 19. Examples of normality tests and corresponding transformation for component-7 of shared-ICAs are presented in Fig. 20 and Fig. 21. The components from both fusion approaches have been normalized and MEWMA have been applied on both sets of ICAs separately in order to find their status based on the UCLs. At this stages ARL has been varied from 300 500 while weight factors have been varied between 0 to 1. It is seen that the in-control and out-of-control percentage changes for different values of ARL and weight factor parameter. The effect on out-of-control due to the variation of ARL and weight factor for both separated and shared ICAs from two sets of features are presented in Table 3. Example of control chart for separated-ICA for fusion architecture-2 for ARL=300 and weight=0.5 has been presented in Fig. 22 and for ARL=500 and weight=0.9 has been presented in Fig. 23. Example of a control chart for shared-ICA for fusion architecture-1 for ARL=300 and weight=0.5 has been presented in Fig. 25 and for ARL=500 and weight=0.9 has been presented in Fig. 24. It is observed in Table 3 that there is a relationship between the out-of-control (OOC) list, ARL and weight factor. For separated-ICAs, percent of OOC does not change with the increase of ARL, how-

(17)

AN US

where f (Cl (Φi )) is the decision of classifier Cl and yi is the known diagnosis. Classifier’s significance factor is computed as below:   1 1 − l γl = ln (18) 2 l

γl f (Cl (Φi ) = yi )

(20)

l=1

CE

i

T X

PT

Clf inal (xi ) = arg max

ED

M

the weighti is computed as below where π l is the normalization factor: ( weightli exp−γl , if (Cl (Φi ) = yi l+1 weighti = πl if (Cl (Φi ) 6= yi expγl , (19) The final decision for a patient xi is determined as equation (20):

4. Experiment results and discussion

AC

After feature extraction from collected data, a normality test is performed for all variables. Most of the features are found to be non-normal, an example of normality test for glucose level and LSHR have been presented in Fig. 8 and Fig. 9. Latent variable components have been computed by using shared-ICA and separated-ICA as mentioned in the methods described in earlier sections. For separated-ICAs of Ewing plus ECG features, total of 10 components has been taken based on their eigenvalues of co-variance matrix. The eigenvalues which have very small value, corresponding components have been discarded as mentioned in Fig. 10. 11

CR IP T

ACCEPTED MANUSCRIPT

ED

M

AN US

Figure 8: Normality test for Glucose level

AC

CE

PT

Figure 9: Normality test for Lying to standing HR

Figure 10: Eigenvalues of covariance matrix for components of ewing plus ECG features (Separated)

12

AN US

CR IP T

ACCEPTED MANUSCRIPT

AC

CE

PT

ED

M

Figure 11: Components of Ewing plus ECG features (Separated)

Figure 12: Components of blood chemistry features (Separated)

13

CR IP T

ACCEPTED MANUSCRIPT

ED

M

AN US

Figure 13: Normality tests for component-3 of blood-ICAs (Separated)

AC

CE

PT

Figure 14: Johnson Transformation of component-3 of blood-ICAs (Separated)

Figure 15: Normality tests for component-3 of Ewing-ECG shared-ICAs

14

CR IP T

ACCEPTED MANUSCRIPT

ED

M

AN US

Figure 16: Johnson Transformation of component-7 of Ewing-ECG shared-ICAs

AC

CE

PT

Figure 17: Components of ewing plus ECG and blood features Shared-ICAs

Figure 18: Normality tests for component-3 of shared-ICAs

15

CR IP T

ACCEPTED MANUSCRIPT

ED

M

AN US

Figure 19: Johnson Transformation of component-3 of shared-ICAs

AC

CE

PT

Figure 20: Normality tests for component-7 of shared-ICAs

Figure 21: Johnson Transformation of component-7 of shared-ICAs

16

ACCEPTED MANUSCRIPT

Table 3: Effect on out of control due to the variation of ARL and weight for both separated-ICAs and shared-ICAs from two sets of features. In the table ‘out’ refers to percentage of out-of-control and ‘wt’ refers to weight factor of MEWMA control chart

wt=0.3 1092 1086 1086 1098 1091 1086

out(%) 93.6 93.1 93.1 94.1 93.5 93.1

wt=0.5 1007 988 975 994 982 970

out(%) 86.3 84.7 83.5 85.2 84.1 83.1

wt=0.7 864 839 829 849 824 808

out(%) 74.0 71.9 71.0 72.8 70.6 69.2

wt=0.9 633 608 592 642 625 599

M ED

AC

CE

PT

Figure 22: MEWMA control chart for separated-ICAs for ARL=300 and weight=0.5

Figure 23: MEWMA control chart for separated-ICAs for ARL=500 and weight=0.9

17

out(%) 54.2 52.1 50.7 55.0 53.6 51.3

CR IP T

ARL 300 400 500 300 400 500

AN US

Type Separated-ICAs Separated-ICAs Separated-ICAs Shared-ICAs Shared-ICAs Shared-ICAs

ACCEPTED MANUSCRIPT

Table 4: Confusion matrix

p n Total

Predicted class

ever for a particular ARL, when weight factor is increased, percent of OOC is decreased. In contrast for shared-ICAs, percent of OOC is also does not change a lot as the ARL is increased. However for a particular ARL, when weight factor is increased, percent of OOC is decreased more than the corresponding OOC percentage of separated-ICA. This indicates a clear differences between the two processes, a larger values of ARL is required in sharedICAs for a large mean shift while a large value of weight factor is required for a large mean shift for separated-ICAs.

Actual class p n TP FP FN TN P N TP + TN P +N

CR IP T

Accuracy =

(23)

AN US

The fusioned feature from both stages (Fusion level-1 and Fusion level-2) are used for identifying CAN. As mentioned earlier section 3.5 that a multi-classifier decision fusion system has been used where classification is accomplished by following equation-(20). The different metrics used are computed based on the classification results by using equations-(21),(22),(23). Following proposed four fusion approaches have been extensively verified with the data and compared with its counterpart-(No fusion approaches) including(Blood chemistry Feature( No fusion),Blood chemistry Feature ( No fusion)+ MEWMA,ECG and Ewing test( No fusion), ECG and Ewing test ( No fusion)+MEWMA),ECG and Ewing and Blood( No fusion),ECG and Ewing and Blood( No fusion)+MEWMA) to justify the performance of the following approaches. The results are presented in Table 5.

4.1. Performance metrics and evaluation

ED

M

While justifying the performances of the proposed approaches to diagnose the CAN, many metrics have been used, These are mentioned in Table 4 which presents the classification categories. If the patient does not have any CAN and he/she is identified as a ‘no-CAN’, this is counted as a true positive (TP) in equations (21) and (23). However, if the the patient has a CAN and he/she is identified as a ‘no-CAN’, this is counted as a false positive (FP). If the patient has a CAN and he/she is classified as a CAN, this is counted as a true negative (TN) in equation (23). However, if the patient has a ‘no-CAN’ and he/she is identified as a ‘CAN’, this is counted as false negative(FN). In equations (21) and (23), P and N represent the total number of patients for ‘CAN’ and ‘no-CAN’ respectively. This is explained in Table 4 using a confusion matrix. The metrics which are used for performance analysis include true positive (TP) rate, false positive (FP) rate, and area under the receiver operating characteristics (ROC) graph [16]. ROC graph (AUC) area has been calculated as mentioned in the article-[6]. Values of AUC span from the range (0 5 AU C 5 1).

• Shared ICA(Proposed approach)

PT

• Shared ICA +MEWMA(Proposed approach) • Separated ICA(Proposed approach)

CE

• Separated proach)

TP P

(21)

F alse positive(F P ) rate(F P rate) =

FP N

(22)

ap-

It is seen from the experimental results that fusion with MEWMA and ICA performs better than without the fusion approach. Better performance is achieved with a shared-ICA than a separated-ICA. This is due to a better component analysis which is reflected in ARL variation results. In ARL variation, comparatively less percent of OOC is achieved in shared-ICA. Best performance is achieved with Shared ICA +MEWMA(Proposed approach) which achieves an accuracy of 94%. This is much higher than standard ‘Ewing’ and ‘ECG and Ewing’ features.

AC

T rue positive(T P ) rate(T P rate) =

ICA+MEWMA(Proposed

18

AN US

CR IP T

ACCEPTED MANUSCRIPT

AC

CE

PT

ED

M

Figure 24: MEWMA control chart for Shared-ICAs for ARL=500 and weight=0.9

Figure 25: MEWMA control chart for Shared-ICAs for ARL=300 and weight=0.5

19

ACCEPTED MANUSCRIPT

5. Conclusion

which variable is more sensitive and cannot be used as a rule for CAN progress monitoring. The variable selection task can be further extended from the current work. In this case, an individual univariate EWMA process can be run for every individual feature with computed UCL and lower control limits (LCL). Then these can be mapped to the MEWMA T 2 statistics directly by using the proposed fusion approach and its ensemble classification technique for determining sensitive variables and their individual UCL and LCL. The corresponding UCLs and LCLs of individual variable from EWMA can be used to monitor the progress of CAN development either for an individual patient or for a cohort of patients. This can be accomplished in a future work as the extension of current work.

CR IP T

Early and accurate monitoring and diagnosis of cardiovascular autonomic neuropathy (CAN) can reduce the risk of cardiovascular disease related death rate significantly. Conventional ‘Ewing battery test’ are often difficult for elderly patients to undertake accurately or cannot be undertaken by them at all. In this situation ‘Ewing test’ results which can lead to an incomplete CAN dataset. Blood biochemistry and different morphological features from ECG have been considered as a complementary features of ‘Ewing tests’ in this paper. However these feature brings a number of dataanalysis challenges including dimensionality, heterogeneity in the domain values and multimodality to some extent. A multistage fusion approach based on a generative model and statistical process control (SPC) techniques have been proposed to address the CAN diagnosis challenge. Two different generative models have been proposed including a shared-ICA and separated-ICA based approach which are fusioned with multiple MEWMA processes with varying ARLs and weight factors. Fusioned latent components with their multivariate T 2 statistics computed by MEWMA are used in a multi-classifier ensemble classification system. Experimental results demonstrate that ICA based generative model extracts the hidden sources using blind source separation approach which performs better than the existing ewing-test features. When applied this proposed component based features in an unsupervised multivariate process control technique, it can find the inherent statistical co-relations among different CAN features with UCLs showing the expected boundary value of multivariate determining the out-of-control lists. The achieved unsupervised knowledge by SPC when fusioned with latent components, this accelerates the performance of an ensemble classification for both generative models. Shared-ICA performed better than separated-ICA models and both outperform the conventional ‘Ewing battery test’ approach. It is experimentally proved that complementary features can be used along with ‘Ewing features’ for patients with mobility challenge for achieving a high performance diagnosis of CAN by using our proposed ‘multistage fusion approach’. In the current approach computed statistical upper control limit (UCL) of fusioned MEWMA processes is a multivariate joint quantity which cannot be directly used for identifying

AN US

Acknowledgement

AC

CE

PT

ED

M

The authors would like to extend their sincere appreciation to the Deanship of Scientific Research at King Saud University for funding this research through Research Group Project No. RGP- 143735

20

References [1] Andreoli1, A., Gravina, R., Giannantonio, R., Pierleoni1, P., and Fortino, G. (2010). Spine-hrv: A bsnbased toolkit for heart rate variability analysis in the timedomain. Lecture Notes in Electrical Engineering, 75:369– 389. [2] Ankishan, H. (2013). Comparison of svm and anfis for snore related sounds classification by using the largest lyapunov exponent and entropy. Computational Mathematical Methods Medicine. [3] Antink, C. H., Gao, H., Bruser, C., and Leonhardt, S. (2015). Beat-to-beat heart rate estimation fusing multimodal video and sensor data. Biomedical optics express, 6(8). [4] Association, A. D. and of Neurology, A. A. (1988). Report and recommendations of the san antonio conference on diabetic neuropathy (consensus statement). Diabetes, 37:1000–1004. [5] Baraka, S., Arjmandc, A., and Ortobelli, S. (2017). Fusion of multiple diverse predictors in stock market. Information Fusion, 36:90–102. [6] Bradley, A. (1997). The use of the area under the roc curve in the evaluation of machine learning algorithms. Pattern Recognition, 30(7):1145–1159. [7] Changliang Zou, F. T. . Z. W. (2007). Monitoring general linear profiles using multivariate exponentially weighted moving average schemes. Technometrics, 49(4):395–408. [8] Chanudet, X., Bonnevie, L., and Bauduceau, B. (2007). Coronary heart disease and cardiovascular autonomic neuropathy in the elderly diabetic. Diabetes and Metabolism, pages 19–31.

ACCEPTED MANUSCRIPT

Table 5: Experiment results: accuracies and other performances metrics of proposed fusion approaches and their comparison with the experiments that were accomplished without the proposed approaches

TP Rate 0.765 0.766 0.816 0.828 0.832 0.835 0.865 0.875 0.895 0.943 0.889 0.938

M

ED

PT

CE

AC

21

FP Rate 0.324 0.322 0.25 0.264 0.259 0.254 0.205 0.168 0.143 0.059 0.144 0.069

ROC area 0.823 0.814 0.872 0.782 0.787 0.79 0.830 0.919 0.941 0.981 0.93 0.974

mation Fusion, 35(5):60–80. [20] Hashimoto, H., Fujiwara, K., Suzuki, Y., Miyajima, M., Yamakawa, T., Kano, M., and Maehara, T. (2013). Epileptic seizure monitoring by using multivariate statistical process control. In 12th IFAC Symposium on Computer Applications in Biotechnology, The International Federation of Automatic Control, pages 16–18. [21] Hsu, C., Huang, H., and Schuschel, D. (2016). The annigma-wrapper approach to fast feature selection for neural nets. IEEE Trans. on Systems, man and cybernetics-Part B: Cybernetics, 32(2). [22] Huda, S., Jelinek, H., Ray, B., Stranieri, A., and Yearwood, J. (2010). Exploring novel features and decision rules to identify cardiovascular autonomic neuropathy using a hybrid of wrapper-filter based feature selection view document. In Sixth International Conference on Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP). [23] Iyengar, S., Bonda, F. T., Gravina, R., Antonio Guerrieri, G. F., and Sangiovanni-Vincentelli, A. (2008). A framework for creating healthcare monitoring applications using wireless body sensor networks. BODYNETS, 8:1–5. [24] J, L. and et.al, P. G. (209). Combining fmri and snp data to investigate connections between brain function and genetics using parallel ica. Human and Brain mapp, 30(1):214–255. [25] J., P. and J., T. W. (1985). A realtime qrs detection algorithm. IEEE Transaction on Biomedical Engineering, 32:230–236. [26] Kohavi, R. and John, G. Wrappers for feature subset selection. Artificial Intelligence, 97(1-2):273–324. [27] Mayhua-Lpez, E., Gmez-Verdejo, V., R., A., and Figueiras-Vidal (2015). A new boosting design of support vector machine classifier. Information Fusion, 25:6371. [28] Mitra, P., M. C. and Pal, S. (2002). Unsupervised feature selection using feature similarity. EEE Transaction on Pattern Analysis and Machine Intelligence, 24:301– 312. [29] Orphanidoua, C., Fleminga, S., Shaha, S., and Tarassenkoa, L. (2013). Data fusion for estimating respiratory rate from a single-lead ecg. Biomedical Signal Processing and Control, 8(1):98–105. [30] Patel, J., Shah, S., Thakkar, P., and Kotecha, K.

AN US

[9] Chena, C., Dantcheva, A., and Ross, A. (2016). An ensemble of patch-based subspaces for makeup-robust face recognition. Information Fusion, 32:80–92. [10] Chiu1, C.-C., Yeh, S.-J., Hu, Y.-H., and Liao, K. Y.-K. (2013). Svm classification for diabetics with various degrees of autonomic neuropathy based on cross-correlation features. Journal of Medical and Biological Engineering, 34(5):495–500. [11] Cohena, J. A., Estaciob, R. O., Lundgrenc, R. A., Esler, A. L., and Schrierd, R. W. (2003). Diabetic autonomic neuropathy is associated with an increased incidence of strokes. Autonomic Neuroscience: Basic and Clinical, 108:73 78. [12] Correa, N. M., Adali, T., Li, Y.-O., and Calhoun, V. D. (2010). Canonical correlation analysis for data fusion and group inferences: Examining applications of medical imaging data. IEEE Signal Process Mag., 27(4):39–50. [13] Cortex, M., Singleton, J. R., and Smith, A. G. (2014). Handbook of Clinical Neurology, volume 126, chapter Glucose intolerance, metabolic syndrome, and neuropathy. Elsevier. [14] Dimova, R., Tankova, T., Guergueltcheva, V., Tournev, I., Chakarova, N., Grozeva, G., and Dakovska, L. (2016). Risk factors for autonomic and somatic nerve dysfunction in different stages of glucose tolerance. Journal of Diabetes and Its Complications. [15] Ewing, D. J. and Clarke, B. F. (1982). Diagnosis and management of diabetic autonomic neuropathy. British Medical Journal, 285:916–918. [16] Fawcett, T. (2006). An introduction to roc analysis. Pattern Recognition Letters, 27:861–874. [17] Fortino, G., Giannantonio, R., Gravina, R., Kuryloski, P., and Jafari, R. (2013). Enabling effective programming and flexible management of efficient body sensor network applications. IEEE Trans. Human-Machine Systems, 43(1):115–133. [18] Gravina, R. and Fortino, G. (2016). Automatic methods for the detection of accelerative cardiac defense response. IEEE Transactions on Affective Computing, 7(3):286– 298. [19] Gravinaa, R., Alinia, P., Ghasemzadeh, H., and Fortino, G. (2017). Multi-sensor fusion in body sensor networks: State-of-the-art and research challenges. Infor-

Accuracy 76.52% 76.61% 81.58% 82.78% 83.20% 83.46% 86.54% 87.48% 89.46% 94.34% 88.86% 93.83%

CR IP T

Proposed approaches and without proposed fusion Blood chemistry Feature( No fusion) Blood chemistry Feature ( No fusion)+MEWMA Ewing test( No fusion) Ewing test ( No fusion)+MEWMA ECG and Ewing test( No fusion) ECG and Ewing test ( No fusion)+MEWMA ECG and Ewing and Blood( No fusion) ECG and Ewing and Blood( No fusion)+MEWMA Shared ICA(Proposed approach) Shared ICA +MEWMA(Proposed approach) Separated ICA(Proposed approach) Separated ICA+MEWMA(Proposed approach)

ACCEPTED MANUSCRIPT

AC

CE

PT

ED

M

AN US

CR IP T

(2015). Predicting stock market index using fusion of machine learning techniques. Experts systems with Applications, 42:2162–2172. [31] Pearlmutter, B. A. and Parra, L. C. (1996). A contextsensitive generalization of ica. In In Proc. of International Conference on Neural Information Processing, Hong Kong, September 24 - 27,. [32] Pinto, A., Montez, C., Araujo, G., Vasques, F., and Portugal, P. (2014). An approach to implement data fusion techniques in wireless sensor networks using genetic machine learning algorithms. Information Fusion, 15:90– 100. [33] Prince, C. T., Secrest, A. M., Mackey, R. H., Arena, V. C., Kingsley, L. A., and Orchard, T. J. (2010). Cardiovascular autonomic neuropathy, hdl cholesterol, and smoking correlate with arterial stiffness markers determined 18 years later in type 1 diabetes. DIABETES CARE, 33(3):73 78. [34] Prince, C. T., Secrest, A. M., Mackey, R. H., Arena, V. C., Kingsley, L. A., and Orchard, T. J. (2016). An association analysis of lipid profile and diabetic cardiovascular autonomic neuropathy in a chinese sample. Lipids Health Dis., 122(15). [35] Pujar, S., Calvert, S., Cortina-Borja, M., China, R. F., Smitha, R. A., and et. al, J. H. C. (2010). Statistical process control (spc)a simple objective method for monitoring seizure frequency and evaluating effectiveness of drug interventions in refractory childhood epilepsy. Epilepsy Research, 91:205213. [36] Sibravaa, N. J., Boisseaub, C. L., Jane L. Eisenb, M. C., Mancebo, and Rasmussenb, S. A. (2016). An empirical investigation of incompleteness in a large clinical sample of obsessive compulsive disorder. Journal of Anxiety Disorders, (42):45–51. [37] Sood, A., Ghani, K. R., and et. al, R. A. (2014). Application of the statistical process control method for prospective patient safety monitoring during the learning phase: Robotic kidney transplantation with regional hypothermia (ideal phase 2ab). EUROPEAN UROLOGY, 66(1):371378. [38] Spallone, V., Maiello, M. R., Cicconetti, E., and Menzinger, G. (1997). Autonomic neuropathy and cardiovascular risk factors in insulin-dependent and non insulindependent diabetes. Diabetes Research and Clinical Practice, 34:169–179. [39] Tennant, R., Mohammed, M. A., Coleman, J. J., and Martin, U. (2007). Monitoring patients using control charts: a systematic review. International Journal for Quality in Health Care, 19(4):187 – 194. [40] Tsai, C.-F. (2014). Combining cluster analysis with classifier ensembles to predict financial distress. Information Fusion, 16:46–58. [41] Z, K. and Tichavsk P, O. E. (2006). Efficient variant of algorithm fastica for independent component analysis attaining the cramr-rao lower bound. IEEE Transaction on Neural Networks, 17(5):1265–1277.

22