This content has been downloaded from IOPscience. Please scroll down to see the full text.

Download details:

IP Address: 193.140.151.85

This content was downloaded on 04/11/2014 at 17:26

Please note that terms and conditions apply.

Centre-specific multichannel electrogastrographic testing utilizing wavelet-based

decomposition

View the table of contents for this issue, or go to the journal homepage for more

2006 Physiol. Meas. 27 569

(http://iopscience.iop.org/0967-3334/27/7/002)

Home Search Collections Journals About Contact us My IOPscience

INSTITUTE OF PHYSICS PUBLISHING PHYSIOLOGICAL MEASUREMENT

Physiol. Meas. 27 (2006) 569–584 doi:10.1088/0967-3334/27/7/002

Centre-specific multichannel electrogastrographictesting utilizing wavelet-based decomposition

I V Tchervensky1, R J de Sobral Cintra2, E Neshev1, V S Dimitrov1,D C Sadowski3 and M P Mintchev1,3

1 Department of Electrical and Computer Engineering, University of Calgary, Calgary,Alberta T2N 1N4, Canada2 Department of Statistics, Federal University of Pernambuco, Recife, Pernambuco, Brazil3 Faculty of Medicine, University of Alberta, Edmonton, Alberta T6G 2B7, Canada

E-mail: mmintche@ucalgary.ca

Received 20 December 2005, accepted for publication 27 March 2006Published 27 April 2006Online at stacks.iop.org/PM/27/569

AbstractAlthough the principles of electrogastrography (EGG) have been known foryears, the clinical utility of EGG has not been clearly demonstrated, and EGGrecording and analysis techniques have not been fully standardized. The aim ofthis study was to develop a multichannel EGG technique for detecting abnormalgastric motility using an EGG database specifically designed for a particulartesting centre, maximizing the sensitivity and the specificity of the test. Eighthealthy volunteers formed a reference group to which 4 study groups (17 gastro-oesophageal reflux disease (GORD) patients, 7 functional dyspepsia patients,8 post-fundoplication patients and 12 healthy volunteers) were compared.Eight-channel EGG was recorded in the postprandial and fasting states for30 min each. The recorded signals were wavelet compressed and the resultingerror (per cent root mean square difference (PRD)) after the compression wasutilized to compare the study groups to the reference group. A threshold in thenumber of channels with significantly different PRD values was introduced.Sensitivity (SE), specificity (SP) and correct classification rate (CC) of the testin recognizing each clinical condition in the study groups for several channelthresholds and compressions were calculated, and were maximized. Increasingthe compression and channel threshold levels improved the specificity, butdecreased the sensitivity of the multichannel EGG test. An optimal combinationregion was identified based on a centre-specific adjustment of the channelthreshold and the wavelet compression. The achieved maximum sensitivity,specificity and correct classification for this region in our test centre were asfollows: GORD—SE 82.4%, SP 83.3%, CC 82.8%; functional dyspepsia—SE 100%, SP 75%, CC 84.2%; post-fundoplication—SE 75.0%, SP 83.3%,CC 80.0%. The utilization of a wavelet-based decomposition technique toprocess multichannel EGG signals can be a very effective method for enhancing

0967-3334/06/070569+16$30.00 © 2006 IOP Publishing Ltd Printed in the UK 569

570 I V Tchervensky et al

the clinical utility of EGG, provided it is specifically developed for a giventesting centre.

Keywords: electrogastrography, specificity, sensitivity, correct classificationrate, centre-specific testing, uncoupling, wavelets

1. Introduction

The motor function of the stomach involves receiving ingested food, churning and grindingfood particles into a size suitable for efficient processing, and finally propelling gastric contentfurther along the gastrointestinal (GI) tract for nutrition absorption (Mayer 1987). Thisfunction requires substantial and well-coordinated gastric motility. Gastric myoelectricalactivity (GMA) controls frequency, propagation and velocity of gastric contractions(Szurszewski 1987). At a macro level, GMA consists of two distinct components—slowwave, which in humans is a regular sinusoid-like signal with a frequency of about 3 cycles perminute (cpm), and spikes which are action potentials short in duration, superimposed on topof the slow wave and superseding strong gastric contractions. Abnormalities in GMA maylead to gastric motor dysfunctions resulting in a variety of gastrointestinal disorders, such asgastroparesis (Chen et al 1996, Brzana et al 1998), functional dyspepsia (FD) (Pfaffenbachet al 1997, Lin and Chen 2001, van der Voort et al 2003) and gastro-oesophageal refluxdisease (GORD) (Cucchiara et al 1997, Soykan et al 1997, Orr et al 2000). Therefore, aneasy, reliable and non-invasive test would be very beneficial for detecting GMA alterations,and the associated gastric motility abnormalities.

Several techniques exist for acquiring gastric myoelectrical activity data. Recordingelectrodes can be implanted directly into the stomach wall either serosally or mucosally,or can be positioned cutaneously on the abdomen. Cutaneous recordings of GMA aremost often referred to as electrogastrography (EGG) (Alvarez 1922). Because of its non-invasiveness and relative simplicity, EGG is an attractive alternative to internal methods forGMA recording. Good correlation between cutaneously and serosally recorded gastric signalshas been demonstrated by many researchers (Smout et al 1980, Hamilton et al 1986, Chenet al 1994).

Although possessing the significant advantage of not affecting the underlying gastricmotility process, EGG comes with several drawbacks. Extracting morphological patterns anduseful information from visual inspection of EGG waveforms has been unsuccessful mainlydue to the nature of the gastric electrical field, the low amplitude of the EGG signal and thehigh level of noise present (Abell and Malagelada 1988). However, since the introductionof running spectrum analysis (RSA) (van der Schee et al 1982, van der Schee and Grashuis1987), EGG started to gain interest in the scientific community. The main parameters extractedfrom the Fourier analysis, which is the core of this technique, are dominant frequency andits dynamics, as well as signal amplitude dynamics. Since the amplitude of EGG recordingsdepends on many factors, e.g. the position and the size of the electrodes, as well as the distancebetween the stomach and the electrodes (Verhagen et al 1999, Mintchev and Bowes 1996,1998), usually only the relative change before and after meal ingestion is considered usefulfor inter-subject comparison. Deviations from the normal frequency range of gastric slowwaves are termed tachygastrias (higher than normal rhythm) or bradygastrias (lower thannormal rhythm) (Chen and McCallum 1994). The usefulness of EGG amplitude as a reliablediagnostic parameter has been disputed (Mintchev et al 1993, Pfaffenbach et al 1995a), as has

Centre-specific electrogastrography 571

been the appropriateness of using the terms tachygastria and bradygastria, since it has beenshown that EGG dysrhythmias are not necessarily related to dysrhythmias in the underlyingGMA (Mintchev and Bowes 1997).

Motion artefacts, myoelectricity produced by other GI organs, and the fact that the EGGsignal is not ideally sinusoidal can easily cause misinterpretations and identification of non-existent pathological patterns, especially when an automated computer analysis is employed(Verhagen et al 1999). Several studies have demonstrated that postprandial increases of theamplitude in EGG recordings could be related to gastric distension (Mintchev et al 1993).Recently, randomness in EGG signals has been investigated, and it has been suggestedthat EGG is a non-stationary and non-deterministic signal, thus questioning the traditionaltechniques utilizing Fourier analysis (Mintchev et al 1998, Price et al 2005). For this reasonrefined analysis methods such as wavelet decomposition have been applied (Liang et al 1996,Qiao et al 1998, Akin and Sun 1999, Ryu et al 2002, de Sobral Cintra et al 2004b, Karaet al 2005). The low signal-to-noise (S/N) ratio of EGG signals has prompted the utilizationof other advanced analysis techniques (Irimia and Bradshaw 2005), offering alternatives toeliminating sections of EGG recordings which have been deemed ‘noisy’, usually based onsubjective visual inspection.

However, in order for these and future analysis techniques to be successful and reliable,comprehensive information on EGG dynamics at acquisition is needed. This can be achievedby recording from an adequate number of different EGG channels. Presently, most EGG testsare performed utilizing only a single recording channel, or even when several channels areavailable, only the one with the highest signal-to-noise ratio is used for analysis, thus discardingpotentially valuable information contained in the rest of the EGG channels. Multichannel EGGwas introduced in various configurations in the second half of the 20th century (Sobakin andPrivalov 1976, Mirizzi and Scafoglieri 1983, Chen et al 1989). Using this technique, itwas demonstrated that EGG can recognize gastric electrical uncoupling which is the lack ofelectrical synchronization in different parts of the stomach (Mintchev et al 1997, de SobralCintra et al 2004b). It has been suggested that localized gastric electrical uncoupling is a muchmore common phenomenon than a global gastric tachygastrias affecting the entire stomach(Mintchev and Bowes 1997).

In addition, EGG still lacks standardization of the recording technique between studycentres (e.g. electrode positioning, number of channels, meal content, etc). Therefore, it is notsurprising that inconsistent or even controversial results have been reported in many studieson the clinical utility of EGG for diagnostic purposes (Simonian et al 2004, van der Voort et al2003, Holmvall and Lindberg 2002, Liang et al 2000, Leahy et al 1999, Lin et al 1998, Kaueret al 1999, Chen et al 1996, Pfaffenbach et al 1995b, Parkman et al 2003). Interestingly, noneof these studies has been optimized for maximizing the sensitivity (SE) and the specificity(SP) of the test.

Recent studies reporting consistent sensitivity and specificity values for EGG have beenrelated to abnormal gastric emptying tests. van der Voort et al (2003) reported 87% SE and 94%SP of predicting delayed gastric emptying. They found differences only in the postprandial-fasting (P/F) power increase, and failed to detect any significant frequency changes in the EGGsignals of 40 patients compared to 22 healthy volunteers. Liang et al (2000) used an artificialneural network based on a genetic cascade correlation algorithm in a study of 152 patientswith delayed gastric emptying, reporting 84% SE, 82% SP and 83% correct classification rate.In both of these studies, however, optimization of the technique for maximizing the SE andSP was not considered.

In a comprehensive study, Leahy et al (1999) studied 170 patients with functionaldyspepsia, 70 patients with irritable bowel syndrome (IBS) and 20 GORD patients. Although

572 I V Tchervensky et al

the specificity was high (93%), the sensitivity was only 36%, 25% and 10% for the FD, IBSand GORD groups respectively, and the correct classification rate was not reported. Morerecently, Holmvall and Lindberg (2002) studied patients with functional dyspepsia. Theyreported the inability to use changes in the dominant frequency or P/F power increase asreliable parameters for discriminating 10 patients from 20 healthy volunteers. Simonianet al (2004) studied sensitivity and specificity of a single versus multichannel EGG in patientswith broadly defined upper gastrointestinal dysmotility problems (nausea, vomiting, anorexia,early satiety, abdominal bloating or distension, and upper abdominal discomfort/pain). Theyconcluded that multichannel EGG rendered superior sensitivity and specificity (43% SE and89% SP when using single channel EGG, versus 60% SE and 89% SP with a 4-channel EGGsystem), although their definition of specificity was non-standard. Nevertheless, the correctclassification rate was not calculated and the obtained sensitivity was relatively low even formultichannel EGG.

The aim of this study was to address these deficiencies by developing a multichannelEGG testing technique for detecting abnormal gastric motility using a centre-specific databaseof multichannel recordings for the purpose of optimizing the sensitivity and the specificity ofthe test by maximizing the correct classification rate for several gastric disorders.

2. Methods

2.1. Experimental setup

Eight cutaneous electrogastrographic channels were recorded from 20 healthy volunteers and32 patients with different gastric disorders. The body mass indices of the healthy volunteersand the patients were not statistically significantly different (p < 0.05). The normal subjectswere randomly assigned to two groups. Eight healthy volunteers (3 M and 5 F; median age 25,range 19–30 years) formed the reference group and twelve (2 M and 10 F; median age 24,range 19–36 years) composed the control group. Seventeen gastro-oesophageal reflux disease(GORD, 9 M and 8 F, median age 49, range 32–70 years), seven functional dyspepsia (FD,1 M and 6 F, median age 48, range 26–81 years) and eight post-fundoplication (PF, 2 M and6 F, median age 43, range 42–58 years) patients composed the three pre-diagnosed gastricdisorder groups.

From each subject, a 30 min basal EGG recording was obtained after a 12 h overnightfast. During signal acquisition, the subjects lay comfortably in a supine position in a quietdark room, and were asked to remain as calm as possible. After a standard meal (500 kCal,52% carbohydrates, 19% proteins and 29% fat), a second 30 min recording in the postprandialstate was acquired from each subject. Abdominal skin was cleaned with isopropyl alcoholswabs, and then lightly abraded with ECG & EEG skin paste (D O Weaver & Co, Aurora,CO, USA). Five active and one ground ECG electrodes (Neotrode 1720-003, ConMed, Utica,NY, USA) were attached to the abdominal surface. The first active electrode was positionedat the patient’s left side, 5 cm distally from the xiphoid along the rib cage. The fifth activeelectrode was affixed on the right side at the point of intersection of the midclavicular line andthe lower rib. The rest of the electrodes were equidistantly spaced in between, thus mappingthe abdominal wall along the cutaneous projection of the gastric axis. The ground electrodewas attached to the patient’s left hip. Electrode positioning is presented in figure 1. Table 1shows the electrode combinations utilized to produce 8 bipolar EGG channels.

To eliminate the influence of external noises (e.g. respiration artefacts and cardiacelectrical activity), a 0.02–0.2 Hz band-pass, first-order Butterworth active filter wasemployed for EGG signal conditioning. Following an appropriate amplification, sampling

Centre-specific electrogastrography 573

Figure 1. EGG electrode positions on the abdominal wall.

Table 1. Electrode combinations forming the eight bipolar EGG channels.

Electrode combination Channel

1–2 12–3 23–4 34–5 41–3 52–4 61–4 72–5 8

and digitization were performed using a sampling frequency of 10 Hz and an 8-channelanalogue-to-digital converter (Labmaster 20009, Scientific Solutions Inc, Mentor, OH, USA).

Due to the relatively long duration of the acquisition, some intermittent artefacts werepresent in the recordings. Saturation, loss of signal and motion artefacts were carefullyidentified and excluded. Thus, a visually clean 25.6 min interval from each of the 8 EGGchannels of each subject in each state (fasting and postprandial) was considered for furtheranalysis.

2.2. Signal analysis

In our previous study (de Sobral Cintra et al 2004b) we described in detail the principlesof an EGG-targeting wavelet-based compression technique. With this method, the analysedsignal is wavelet-decomposed up to several scales or levels. Then only a certain number ofwavelet transform coefficients having the largest absolute values are kept for reconstructingthe signal back into the time domain, while the rest of the coefficients are zeroed. Thenumber of preserved coefficients is determined by a parameter called compression ratio.It is calculated as the ratio between the number of the retained and the original waveletdecomposition coefficients. After the compression is performed in the wavelet domain, thesignal is reconstructed in its time-domain form (figure 2). The original and the reconstructedEGG signals are compared in the time domain, and the resulting error after the compressionis calculated. This error, known as per cent root mean square error difference (PRD), is an

574 I V Tchervensky et al

(a)

(b)

Figure 2. Typical artefact-free episode of an original EGG tracing recorded from a normalvolunteer overlaid with a compressed version of the same EGG interval. Three levels of waveletdecomposition and a compression ratio of 9 were used in the processing of the original EGG signal(a). Time-domain difference between the original and the compressed signals (b). The waveletcompression technique acted as an effective filter for high-frequency noise components in the EGGsignal.

indicator of how well the utilized wavelet can represent the analysed signal. Lower errorvalues would result if most of the energy of the signal were concentrated in a small numberof large wavelet coefficients, and thus the remaining energy in the discarded coefficientswere insignificant. By comparing the reconstruction errors for a fixed compression ratio, werecently demonstrated (de Sobral Cintra et al 2004b) that a normal canine EGG signal can bediscriminated from an electrically uncoupled canine EGG recording. Daubechies-3 standardwavelet was shown to be optimal for normogastric EGG (de Sobral Cintra et al 2004a).

In the present study, each 25.6 min EGG recording for each channel was divided into30 consecutive segments with an equal length of 512 samples. Each segment was waveletdecomposed, compressed and reconstructed. Post-compression distortion was measured bycalculating the PRD values between the original and the reconstructed compressed versionsof each segment. Thus, for each EGG channel recorded from each subject under fasting andpostprandial conditions, a set of 30 PRD values was formed.

2.3. Statistical analysis

The reference group of healthy volunteers was utilized to identify the subjects from the controland pre-diagnosed groups as normal or having a gastric disorder. All PRD values of thewavelet-compressed reference signals were arranged according to the recording channel andcondition. Thus, for each EGG channel, a reference set of 240 PRD values was obtained.

The sets of PRD values of the subjects from the control and disorder groups were comparedto the reference PRD sets for a given recording channel and condition. When the evaluated setsof PRD values had normal distributions, a statistical analysis was performed using Student’st-test; otherwise Wilcoxon signed rank test was utilized (Snedecor and Cochran 1967).Lilliefors (Kolmogorov–Smirnov) test for normality (Gonzalez et al 1977) was employedto verify the statistical distribution of the PRD sets.

Centre-specific electrogastrography 575

Statistically significant difference (p < 0.05) between the PRD values of the referencegroup and the control and disorder groups for a given channel and condition was sought. Agiven EGG recording was declared abnormal if it exhibited statistically significant differencesin more than a predetermined number of channels, a parameter labelled channel threshold.

2.4. Evaluation indices used for assessing the performance of the test

The performance of the developed test was estimated by computing three evaluation indices(EI) routinely used for validating medical and clinical tests: specificity (SP), sensitivity (SE)and correct classification rate (CC) (Kraemer 1992):

SP = [TN/(TN + FP)] · 100%, (1)

SE = [TP/(TP + FN)] · 100%, (2)

CC = [(TP + TN)/N] · 100%, (3)

where N is the total number of patients in a given disorder group plus the healthy volunteersin the control group, TP is the number of true positives, TN is the number of true negatives,FN is the number of false negatives and FP is the number of false positives.

2.5. Optimization of the test

The optimization of the test was performed by modifying the compression ratio used in thewavelet compression technique and the channel threshold employed in the EGG statisticalanalysis for a given maximum level of wavelet decomposition. The core of the optimizationprocess involved maximizing the evaluation indices with respect to the compression ratio andthe channel threshold. Since the maximum level of decomposition plays a fundamental role inthe wavelet processing technique, its influence on the performance of the test was investigatedalso. An optimal value of the level of decomposition was sought such that the evaluationindices were maximized globally.

For each of the evaluation indices, a pseudo three-dimensional (3D) surface EI =EI(compression, threshold) was created by mapping the value of a particular index to a pointfrom the plane defined by the compression ratios and channel thresholds. A coordinate pair(compression, threshold) in this plane was identified for which the CC index had a maximum,CCmax. Since CCmax is a localized point in the compression-threshold plane, and dependsstrongly on the choice of proper compression and threshold, an evaluation parameter giving abroader perspective of the performance of the test was needed. For this reason, a region in theplane where CC exhibited values �65% was defined. This value was selected as the medianbetween a CC = 50% (flip of the coin) and the averaged value of all CCmax rates. The relativearea (areaCC�65%) of this region with respect to the total area of the compression-thresholdplane, and the averaged CC index for the same region were determined.

CCmax = max[CC(compression, threshold)], (4)

areaCC�65% = K

M· 100%, (5)

CCavg =∑K

i=1 CC(compressioni , thresholdi )

K, CC � 65%, (6)

where K is the number of points in the compression-threshold plane for which CC � 65%,and M is the number of all points in the plane. Compression ratios ranged between 1.5

576 I V Tchervensky et al

and 20, calculated with a step of 0.25. Channel thresholds were integers 1, 2, . . . , 8 (eightwas the total number of recorded EGG channels).

The product between areaCC�65% and the averaged correct classification for the sameregion was also calculated. An optimal maximum level of wavelet decomposition for whichthe product areaCC�65%CCavg had a maximum was determined for each gastric disorder group.

The test was repeated for five different maximum levels of wavelet decomposition,and results were compared. Due to the short length of the segments, the coarsest waveletdecomposition scale was limited to five. The test was performed separately on fasting andpostprandial EGG recordings.

3. Results

3.1. Sensitivity and specificity of the test

Figure 3 illustrates the pseudo-3D surfaces of two evaluation indices, sensitivity (plots a, cand e) and specificity (plots b, d and f ), calculated for three different maximum levels ofwavelet decomposition: 1, 3 and 5. The signals are from the GORD group of patients inthe fasting state. For better visualization, the colours of the points from the surfaces weremapped to a greyscale according to the values of the EI at the corresponding points, thus lightercolours represented higher specificity or sensitivity. For the same reason, the X (threshold)and Y (compression) axes of the plots were both inverted on the SE and SP graphs. Figure 3(c)demonstrates that the sensitivity increased as both the threshold and compression decreased.On the other hand, the specificity had a minimum at low thresholds and compressions, whichcan be seen in the adjacent figure 3(d).

It is interesting to note that for small compressions and sub-maximal decompositionlevels, the SE and SP exhibited a steep step-like transition (see figures 3(a) and (b)). With theincrease of the decomposition level, this step-like change became less abrupt and movedtowards higher compression ratios to eventually diminish or disappear (see figures 3(e)and (f )).

3.2. Correct classification rate of the test

Figure 4 depicts greyscale 2D plots of the correct classification rate of the test for the samesignals and parameters used in the calculation of the SE and SP shown in figure 3. The greygradient to the right of each graph indicates the scale of the plots in per cent. The positionof maximum correct classification point, CCmax, is denoted by a white arrow. The value ofthe CCmax for each plot is provided on the top of the graphs. The region where CC � 65%is surrounded by a white line. It can be noted that the position of this region changed fordifferent maximum decomposition levels, sliding towards higher compressions and channelthresholds. At a single level of wavelet decomposition it occupied only a narrow strip ofcompressions between 1.25 and 2.75 at several thresholds (see figure 4(a)), while at threelevels of decomposition it encompassed a much broader area (see figure 4(b)). When fouror five decomposition levels were utilized, the graphs of the correct classification rate werealmost identical (see figure 5).

3.3. PRD as a function of the compression rate and the maximum level of decomposition

Figure 5 illustrates the average PRD values obtained from EGG signals from healthy volunteersin the fasting state at several maximum levels of wavelet decomposition and different

Centre-specific electrogastrography 577

(a) (b)

(c) (d)

(e) (f )

Figure 3. Sensitivity ((a), (c), (e)) and specificity ((b), (d), (f )) of the test calculated with 1, 3and 5 levels of wavelet decomposition of the signals from the GORD patients in the fasting state.Lighter colours represent higher values. Note the inverse directions of the X and Y axes of SE andSP. Increasing the threshold leads to decrease in SE but increase in SP. This trend is present for alllevels. Increasing the compression results in decrease in SE but increase in SP. This behaviour isprominent at lower decomposition levels, but diminishes at levels 4 and 5.

compression ratios. The curves were produced by averaging across (a) all subjects in thereference group, (b) all eight EGG recording channels and (c) all slices comprising eachchannel. The obtained PRD values were comprised of two clearly distinguishable sections,an initial slow-changing section, followed by more rapid PRD dynamics for certain maximumlevels of decomposition. All curves demonstrated an increase when the compression ratiosincreased. The lower compressions zeroed low-amplitude wavelet coefficients only. Becausethese coefficients were generally a result of high-frequency noise in the EGG signal, the low

578 I V Tchervensky et al

Figure 4. Correct classification rate as a function of the compression ratio and the channelthreshold calculated for the same signals and utilizing the same wavelet and levels of decompositionemployed for determining the sensitivity and specificity shown in figure 3. Areas with CC � 65%are highlighted. The arrow marks the point of maximum CC.

Figure 5. Average PRD values as a function of compression and maximum level of decomposition,calculated from EGG signals from healthy volunteers in the fasting state. The graphs are producedby averaging across all subjects, EGG channels and slices in each channel.

compression ratios served as an effective noise-reducing filter, thus gradually enhancing theability of the test to detect abnormal signals. In the second section, each curve started to

Centre-specific electrogastrography 579

Table 2. Sensitivity (SE), specificity (SP) and correct classification (CC) of the test calculated forseveral levels of decomposition for three gastric disorders—GORD, functional dyspepsia and post-fundoplication, and two food consumption statuses—fasting and postprandial. The optimizationof the test was achieved by determining the wavelet decomposition level, compression (CMP) andthreshold (THR) values that maximized the correct classification (CCmax).

Fasting Postprandial

Level CMP THR SE (%) SP (%) CCmax (%) CMP THR SE (%) SP (%) CCmax (%)

GORD1 1.75 7 64.7 83.3 72.4 2.25 2 100.0 25.0 69.02 4.75 4 70.6 83.3 75.9 8.75 2 82.4 58.3 72.43 10.75 4 82.4 83.3 82.8 7.75 7 52.9 75.0 62.14 11.00 6 76.5 75.0 75.9 4.75 4 88.2 25.0 62.15 12.25 6 82.4 75.0 79.3 4.5 4 88.2 25.0 62.1

Functional dyspepsia1 2.50 3 71.4 83.3 78.9 14.75 3 57.1 83.3 73.72 4.75 4 71.4 83.3 78.9 4.25 7 71.4 75.0 73.73 6.00 7 100.0 75.0 84.2 4.50 8 57.1 91.7 78.94 10.25 6 100.0 75.0 84.2 4.50 8 57.1 100.0 84.25 9.75 6 100.0 75.0 84.2 4.75 8 57.1 100.0 84.2

Post-fundoplication1 2.25 4 62.5 83.3 75.0 12.25 3 75.0 83.3 80.02 4.75 4 75.0 83.3 80.0 5.25 5 75.0 91.7 85.03 18.50 3 75.0 83.3 80.0 4.25 8 87.5 91.7 90.04 19.75 5 87.5 75.0 80.0 4.50 8 87.5 100.0 95.05 11.75 7 75.0 83.3 80.0 4.50 8 87.5 100.0 95.0

increase more rapidly due to the discarding of more significant wavelet coefficients pertainingto the EGG signal.

3.4. Performance of the test

Tables 2 and 3 summarize the results of the test for the three gastric disorders separatelyfor the fasting and the postprandial states. Table 2 gives the maximum values of correctclassification achieved for different maximum decomposition levels. The corresponding pairsof compression ratios and channel thresholds for which these maxima have been calculated,as well as the sensitivity and specificity for the same parameters, are also shown.

Table 3 shows the normalized areas of the regions in the compression-threshold planesin which CC � 65%. The average correct classification rate for these regions was comparedto CCmax. Product areaCC�65%CCavg was useful in evaluating the test in a broader range ofdecomposition parameters, and for optimizing the level of decomposition. Because CCmax

for the GORD group in the postprandial state was less than 65%, CCavg, areaCC�65% andareaCC�65%CCavg were not calculated.

4. Discussion

Contemporary techniques for analysis of EGG recordings are mainly based on runningspectrum short time Fourier transform analysis (Chang 2005). Terms like tachygastria andbradygastria are introduced to explain deviations from the normogastric frequency range ofthe EGG signals. Norms have been set to the percentage of time the EGG signal can exhibit

580 I V Tchervensky et al

Table 3. Maximum correct classification rate compared to the average CC computed for thecompression-threshold plane regions in which CC � 65%. The normalized area of theseregions is also shown. The product areaCC�65%CCavg was used for optimization of the levelof decomposition.

Fasting Postprandial

CCmax CCavg Area CCavg · area CCmax CCavg Area CCavg · areaLevel (%) (%) (%) (%) (%) (%) (%) (%)

GORD1 72.4 68.6 1.5 1.0 69.0 65.9 0.8 0.52 75.9 67.2 6.3 4.2 72.4 67.9 3.7 2.53 82.8 69.6 16.5 11.5 62.1 – – –4 75.9 67.9 16.0 10.9 62.1 – – –5 79.3 68.1 16.0 10.9 62.1 – – –

Functional dyspepsia1 78.9 69.4 17.2 11.9 73.7 69.0 37.8 26.12 78.9 70.6 27.0 19.1 73.7 68.9 22.5 15.53 84.2 73.1 39.3 28.7 78.9 70.6 17.5 12.44 84.2 76.2 36.3 27.7 84.2 73.5 4.2 3.15 84.2 74.6 30.3 22.6 84.2 75.1 3.0 2.3

Post-fundoplication1 75.0 66.4 12.7 8.4 80.0 72.0 71.0 51.12 80.0 66.5 20.0 13.3 85.0 72.3 65.0 47.03 80.0 69.4 42.2 29.3 90.0 71.1 40.8 29.04 80.0 69.9 26.5 18.5 95.0 68.7 13.5 9.35 80.0 70.1 16.5 11.6 95.0 68.8 13.3 9.2

abnormally high/low frequency, or exhibit complete absence of dominant frequency. The ratiobetween the standard deviation of the dominant EGG frequency and its mean value, labelledinstability coefficient, has also been utilized (Chen and McCallum 1994). Another morecommon parameter used in the conventional EGG analysis is the postprandial-to-fasting powerincrease of the cutaneously recorded GMA signal (Chen and McCallum 1994). However,these attempts at parameter-based quantification failed to convert the otherwise attractivenon-invasive EGG test into a routine clinical diagnostic procedure.

In the present study we propose a centre-specific multichannel testing procedure based onthe wavelet compression technique, utilizing an 8-channel EGG recording system on variousgroups of normal volunteers and pre-diagnosed patients. Instead of selecting only the channelwith highest signal-to-noise ratio for further analysis and discarding the rest, we determinedthe number of statistically different channels and used it as a parameter for optimizing theperformance of the test. We assumed that an abnormality could easily be missed in a singlechannel but could be present in several others, depending on the severity of the GI disorder.By utilizing five active cutaneous electrodes positioned over the abdominal projection of thegastric axis, we practically mapped the stomach abdominally. The bipolar configuration ofthe EGG channels ensured low noise and reduction of the artefacts.

Most commonly employed EGG testing procedures worldwide include acquiring thediagnostic parameters based on frequency analysis, and comparing the values of theseparameters to previously published or to reference results. The still non-standardizedpositioning of the recording electrodes and the recording procedure itself could be a preventivefactor for such comparison. In the present study, we propose an approach in which a

Centre-specific electrogastrography 581

multichannel database of EGG recordings is utilized that is specific for a particular EGGtest centre. Thus, the test can be optimized for a given local medical practice and the influenceof subjective factors, such as the recording technique or the test meal, could be eliminated.

In our study we used two parameters for optimizing the test, the compression ratioand the channel threshold. A combination of these two parameters was sought so that thecorrect classification rate was maximized. In general, the sensitivity and the specificitywere found to be inversely related to each other, i.e. when one was maximized, the otherwas minimized. Obviously, when the channel threshold was increased, more channels wereexpected to exhibit abnormality, thus more patients were missed as having abnormal EGG. Thisalso resulted in improving the specificity, since more healthy volunteers were recognized ashaving normal EGG recordings. The influence of the compression ratio was not that obvious.Lower levels of compression meant zeroing only a few noise-related and insignificant waveletcoefficients. Higher compressions resulted in discarding a significant amount of valuableinformation in the higher-amplitude coefficients, starting to affect the EGG signal directly.This rendered the patients indistinguishable from normal reference subjects, i.e. lowered thesensitivity. Similarly, the difference between control and reference subjects diminished, thusincreasing the specificity. Finding the maximum of the correct classification ensured thatoptimal simultaneously high values for SE and SP were achieved for given compression ratioand channel threshold.

The influence of the maximum level of decomposition was also evaluated. The maximumcorrect classification in the fasting state for all GI disorders was achieved with three ormore levels of wavelet decomposition (see table 2). Three levels proved to be optimalwhen areaCC�65%CCavg product was maximized (see table 3). In the postprandial state,the maximum correct classification for functional dyspepsia and post-fundoplication patientgroups was attained when four or five decomposition levels were utilized. Although95% CCmax with 87.5% SE and 100% SP in detecting PF subjects seems extremely promising,the regions of CC � 65% occupied relatively small areas of the compression-threshold plane(see table 3). The average correct classification rates for these regions were considerably belowthe CCmax, indicating a narrow localization of the CC extremes. Both patient groups showed asteadily declining areaCC�65%CCavg product when the maximum level of decomposition wasincreased. Thus, for these GI disorders the areaCC�65%CCavg product had a maximum whena single level of wavelet decomposition was utilized. However, the test sensitivity in bothpatient groups (functional dyspepsia and post-fundoplication) was not very high. Only fourout of seven (57.1% SE) FD patients were recognized as abnormal when one decompositionlevel was used. The GORD subjects in the postprandial state showed maxima of both thecorrect classification rate and the areaCC�65%CCavg product when two wavelet decompositionlevels were employed. Contrary to our expectations, areaCC�65% for GORD was very small.

The fasting state analysis exhibited a well-defined and consistent pattern across the groups.The maximum correct classification was around 80%–82.8%, 84.2% and 80.0% for GORD,FD and PF respectively. Average correct classification rates in the regions where CC � 65%were around 70%. The areas of these regions for GORD patients were considerably lowerthan those in FD and PF patients—16.5% versus 39.3% and 42.2%. This could be due to thelarger and statistically more representative GORD group compared to the FD and PF patientgroups.

It should be mentioned that postprandial CCmax rates and particularly the areaCC�65%

values differed from those in the fasting state for some disorders (see tables 2 and 3). Forexample, in functional dyspepsia fasting and postprandial rates of CCmax for a given level ofdecomposition were similar, while areaCC�65% differed, suggesting that the performance ofthe test was better in the fasting state. It should be emphasized, however, that these findings

582 I V Tchervensky et al

are sensitive to the level of decomposition, which can be optimized further to maximizethe performance of the test for a given disorder. Simultaneously high values of correctclassification and areaCC�65% (i.e. high areaCC�65%CCavg values) are always desirable but rarelyachievable. The optimum maximum level of decomposition for best correct classification orhighest areaCC�65%CCavg is usually different. Optimization of the level of decomposition bymaximizing areaCC�65%CCavg product is preferable to optimization using CCmax only when atest with more uniform correct classification is needed.

The proposed method envisions the establishment of acceptable correct classification ratestogether with optimal compression ratios and channel thresholds within a given diagnosticcentre using groups of pre-diagnosed patients with specific gastric motility abnormalities andhealthy volunteers. Once these parameters are determined using the described methodology,EGG tests on upcoming non-diagnosed symptomatic patients may be performed as a pre-screening tool to indicate the possible presence of a given abnormality. It remains to beestablished whether the obtained correct classification rates specific for a given clinical centrewould be adequate to define EGG tests as the sole basis of reliable diagnosis, or the patientswould be streamlined to more specialized tests.

5. Conclusion

Our study suggests that the proposed wavelet compression-based testing can be a valuable toolin enhancing the clinical utility of EGG. The proposed centre-specific approach overcomesthe current recording-related problems, while the multichannel EGG signal acquisition andanalysis enhances the performance of the test.

Acknowledgments

This research is supported in part by the Natural Sciences and Engineering Research Councilof Canada and the Gastrointestinal Motility Laboratory, University of Alberta Hospitals,Edmonton, Alberta.

References

Abell T L and Malagelada J R 1988 Electrogastrography: current assessment and future perspectives Dig. Dis. Sci.33 982–92

Akin A and Sun H H 1999 Time frequency methods for detecting spike activity of stomach Med. Biol. Eng. Comput.37 381–90

Alvarez W C 1922 The electrogastrogram and what it shows J. Am. Med. Assoc. 78 1116–8Brzana R J, Koch K L and Bingaman S 1998 Gastric myoelectrical activity in patients with gastric outlet obstruction

and idiopathic gastroparesis Am. J. Gastroenterol. 93 1803–9Chang F Y 2005 Electrogastrography: basic knowledge, recording, processing and its clinical applications

J. Gastroenterol. Hepatol. 20 502–16Chen J, Vandewalle J, Sansen W, van Cutsem E, Vantrappen G and Janssens J 1989 Observation of the

propagation direction of human electrogastric activity from cutaneous recordings Med. Biol. Eng. Comput. 27538–42

Chen J D, Lin Z, Pan J and McCallum R W 1996 Abnormal gastric myoelectrical activity and delayed gastric emptyingin patients with symptoms suggestive of gastroparesis Dig. Dis. Sci. 41 1538–45

Chen J D, Schirmer B D and McCallum R W 1994 Serosal and cutaneous recordings of gastric myoelectrical activityin patients with gastroparesis Am. J. Physiol. 266 G90–8

Chen J Z and McCallum R W 1994 Electrogastrography, Principles and Applications (New York: Raven Press)Cucchiara S, Salvia G, Borrelli O, Ciccimarra E, Az-Zeqeh N, Rapagiolo S, Minella R, Campanozzi A and

Riezzo G 1997 Gastric electrical dysrhythmias and delayed gastric emptying in gastroesophageal reflux diseaseAm. J. Gastroenterol. 92 1103–8

Centre-specific electrogastrography 583

de Sobral Cintra R, Tchervensky I, Dimitrov V and Mintchev M 2004a Optimal wavelets for electrogastrography Conf.Proc. 26th Ann. Int. Conf. IEEE Engineering in Medicine and Biology Society, EMBC 2004 (San Francisco,CA, 1–5 Sep.) (Piscataway, NJ: IEEE) pp 329–32

de Sobral Cintra R J, Tchervensky I V, Dimitrov V S and Mintchev M P 2004b Wavelet analysis in a canine modelof gastric electrical uncoupling Physiol. Meas. 25 1355–69

Gonzalez T G, Sahni S and Franta W R 1977 An efficient algorithm for the Kolmogorov–Smirnov and Lilliefors testsACM. Trans. Math. Software 3 60–4

Hamilton J W, Bellahsene B E, Reichelderfer M, Webster J G and Bass P 1986 Human electrogastrograms: comparisonof surface and mucosal recordings Dig. Dis. Sci. 31 33–9

Holmvall P and Lindberg G 2002 Electrogastrography before and after a high-caloric, liquid test meal in healthyvolunteers and patients with severe functional dyspepsia Scand. J. Gastroenterol. 37 1144–8

Irimia A and Bradshaw L A 2005 Artifact reduction in magnetogastrography using fast independent componentanalysis Physiol. Meas. 26 1059–73

Kara S, Dirgenali F and Okkesim S 2005 Estimating gastric rhythm differences using a wavelet method from theelectrogastrograms of normal and diabetic subjects Instrum. Sci. Technol. 33 519–32

Kauer W K, Stein H J, Balint A and Siewert J R 1999 Transcutaneous electrogastrography: a non-invasive method toevaluate post-operative gastric disorders? Hepatogastroenterology 46 1244–8

Kraemer H C 1992 Evaluating Medical Tests: Objective and Quantitative Guidelines (Newbury Park, CA: SagePublications)

Leahy A, Besherdas K, Clayman C, Mason I and Epstein O 1999 Abnormalities of the electrogastrogram in functionalgastrointestinal disorders Am. J. Gastroenterol. 94 1023–8

Liang J, Cheung J C and Chen J D Z 1996 Noise detection and denoising on electrogastrography using nonorthogonalmultiresolution wavelet analysis Proc. 1996 18th Ann. Int. Conf. IEEE Engineering in Medicine and BiologySociety part 3 (Amsterdam, 31 Oct–3 Nov) (Piscataway, NJ: IEEE) pp 1039–40

Liang H, Lin Z and McCallum R W 2000 Application of combined genetic algorithms with cascade correlation todiagnosis of delayed gastric emptying from electrogastrograms Med. Eng. Phys. 22 229–34

Lin X and Chen J Z 2001 Abnormal gastric slow waves in patients with functional dyspepsia assessed by multichannelelectrogastrography Am. J. Physiol. Gastrointest. Liver Physiol. 280 G1370–5

Lin X, Levanon D and Chen J D 1998 Impaired postprandial gastric slow waves in patients with functional dyspepsiaDig. Dis. Sci. 43 1678–84

Mayer J H 1987 Motility of the stomach and gastroduodenal junction Physiology of the Gastrointestinal Tracted L R Johnson 2nd edn, vol 1 (New York: Raven Press) pp 613–30

Mintchev M P and Bowes K L 1996 Impact of external factors on the stability of human electrogastrogramsMed. Biol. Eng. Comput. 34 270–2

Mintchev M P and Bowes K L 1997 Do increased electrogastrographic frequencies always correspond to internaltachygastria? Ann. Biomed. Eng. 25 1052–8

Mintchev M P and Bowes K L 1998 Computer simulation of the effect of changing abdominal thickness on theelectrogastrogram Med. Eng. Phys. 20 177–81

Mintchev M P, Kingma Y J and Bowes K L 1993 Accuracy of cutaneous recordings of gastric electrical activityGastroenterology 104 1273–80

Mintchev M P, Otto S J and Bowes K L 1997 Electrogastrography can recognize gastric electrical uncoupling in dogsGastroenterology 112 2006–11

Mintchev M P, Stickel A and Bowes K L 1998 Dynamics of the level of randomness in gastric electrical activityDig. Dis. Sci. 43 953–6

Mirizzi N and Scafoglieri U 1983 Optimal direction of the electrogastrographic signal in man Med. Biol. Eng. Comput.21 385–9

Orr W C, Zhang M, McClanahan J, Sloan S and Chen J D 2000 Gastric myoelectric activity in older adults treatedwith cisapride for gastro-oesophageal reflux disease Aliment. Pharmacol. Ther. 14 337–43

Parkman H P, Hasler W L, Barnett J L and Eaker E Y 2003 Electrogastrography: a document prepared by thegastric section of the American Motility Society Clinical GI Motility Testing Task Force Neurogastroenterol.Motil. 15 89–102

Pfaffenbach B, Adamek R J, Bartholomaus C and Wegener M 1997 Gastric dysrhythmias and delayed gastric emptyingin patients with functional dyspepsia Dig. Dis. Sci. 42 2094–9

Pfaffenbach B, Wedmann B, Adamek R J and Wegener M 1995a The significance of electrogastrographicallydetermined amplitudes—is there a correlation to sonographically measured antral mechanical contractions?Z. Gastroenterol. 33 103–7

Pfaffenbach B, Wegener M, Adamek R J, Schaffstein J, Lee Y H and Ricken D 1995b Antral myoelectric activity,gastric emptying, and dyspeptic symptoms in diabetics Scand. J. Gastroenterol. 30 1166–71

584 I V Tchervensky et al

Price C N, Westwick D T and Mintchev M P 2005 Analysis of canine model of gastric electrical uncoupling usingrecurrence quantification analysis Dig. Dis. Sci. 50 885–92

Qiao W, Sun H H, Chey W Y and Lee K Y 1998 Continuous wavelet analysis as an aid in the representation andinterpretation of electrogastrographic signals Ann. Biomed. Eng. 26 1072–81

Ryu C Y, Nam K C, Kim S C and Kim D W 2002 Comparison of digital filters with wavelet multiresolutionfilter for electrogastrogram Proc. 2002 IEEE Engineering in Medicine and Biology 24th Ann. Conf. and the2002 Fall Meeting of the Biomedical Engineering Society (BMES/EMBS) (23–26 Oct) (Houston, TX: IEEE)pp 137–8

Simonian H P, Panganamamula K, Chen J Z, Fisher R S and Parkman H P 2004 Multichannel electrogastrography(EGG) in symptomatic patients: a single center study Am. J. Gastroenterol. 99 478–85

Smout A J, van der Schee E J and Grashuis J L 1980 What is measured in electrogastrography? Dig. Dis. Sci.25 179–87

Snedecor G W and Cochran W G 1967 Statistical Methods 6th edn (Ames, IO: The Iowa State University Press)Sobakin M A and Privalov I A 1976 Multichannel recording of gastric electric activity from the surface of the human

body Biull. Eksp. Biol. Med. 81 506–7Soykan I, Lin Z, Jones S, Chen J and McCallum R W 1997 Gastric myoelectrical activity, gastric emptying and

correlations with dyspepsia symptoms in patients with gastroesophageal reflux J. Investig. Med. 45 483–7Szurszewski J H 1987 Electrical basis for gastrointestinal motility Physiology of the Gastrointestinal Tract

ed L R Johnson 2nd edn, vol 1 (New York: Raven Press) pp 383–422van der Schee E J and Grashuis J L 1987 Running spectrum analysis as an aid in the representation and interpretation

of electrogastrographic signals Med. Biol. Eng. Comput. 25 57–62van der Schee E J, Smout A J P M and Grashuis J L 1982 Application of running spectrum analysis to

electrogastrographic signals recorded from dog and man Motility of the Digestive Tract ed M Weinbeck (NewYork: Raven Press) pp 241–50

van der Voort I R, Osmanoglou E, Seybold M, Heymann-Monnikes I, Tebbe J, Wiedenmann B, Klapp B F andMonnikes H 2003 Electrogastrography as a diagnostic tool for delayed gastric emptying in functional dyspepsiaand irritable bowel syndrome Neurogastroenterol. Motil. 15 467–73

Verhagen M A, Van Schelven L J, Samsom M and Smout A J 1999 Pitfalls in the analysis of electrogastrographicrecordings Gastroenterology 117 453–60