A machine learning approach using auditory odd-ball responses to investigate the effect of Clozapine therapy.

Clinical Neurophysiology xxx (2014) xxx–xxx

Contents lists available at ScienceDirect

Clinical Neurophysiology journal homepage: www.elsevier.com/locate/clinph

A machine learning approach using auditory odd-ball responses to investigate the effect of Clozapine therapy Maryam Ravan a,⇑, Gary Hasey b,1, James P. Reilly a, Duncan MacCrimmon b,1, Ahmad Khodayari-Rostamabad a a b

Department of Electrical and Computer Engineering, McMaster University, Hamilton, ON, Canada Department of Psychiatry and Behavioural Neurosciences, McMaster University, Hamilton, ON, Canada

a r t i c l e

i n f o

Article history: Accepted 7 July 2014 Available online xxxx Keywords: Brain source localization EEG signals Odd-ball auditory evoked potentials Machine learning Schizophrenia Clozapine treatment

h i g h l i g h t s A machine learning algorithm is used to identify a set of ‘‘features’’, from odd-ball auditory evoked

potentials, that can simultaneously discriminate between clinically important conditions. This discrimination capability concludes that the brain function associated with these features

normalizes in responding patients as a result of Clozapine treatment. The proposed approach can help in our understanding of the changes in brain behavior due to

Clozapine and its therapeutic effect in schizophrenia.

a b s t r a c t Objective: To develop a machine learning (ML) methodology based on features extracted from odd-ball auditory evoked potentials to identify neurophysiologic changes induced by Clozapine (CLZ) treatment in responding schizophrenic (SCZ) subjects. This objective is of particular interest because CLZ, though a potentially dangerous drug, can be uniquely effective for otherwise medication-resistant SCZ subjects. We wish to determine whether ML methods can be used to identify a set of EEG-based discriminating features that can simultaneously (1) distinguish all the SCZ subjects before treatment (BT) from healthy volunteer (HV) subjects, (2) distinguish EEGs collected before CLZ treatment (BT) vs. those collected after treatment (AT) for those subjects most responsive to CLZ, (3) discriminate least responsive subjects from HV AT, and (4) no longer discriminate most responsive subjects from HVs AT. If a set of EEG-derived features satisfy these four conditions, then it may be concluded that these features normalize in responsive subjects as a result of CLZ treatment, and therefore potentially provide insight into the functioning of the drug on the SCZ brain. Methods: Odd-ball auditory evoked potentials of 66 HVs and 47 SCZ adults both BT and AT with CLZ were derived from EEG recordings. Treatment outcome, after at least one year follow-up, was assessed through clinical rating scores assigned by an experienced clinician, blind to EEG results. Using a criterion of at least 35% improvement after CLZ treatment, subjects were divided into ‘‘most-responsive’’ (MR) and ‘‘least-responsive’’ (LR) groups. As a first step, a brain source localization (BSL) procedure was employed on the EEG signals to extract source waveforms from specified brain regions. ML methods were then applied to these source waveform signals to determine whether a set of features satisfying the four conditions outlined above could be discovered. Results: A set of cross-power spectral density (CPSD) features meeting these criteria was identified. These CPSD features, consisting of a combination of brain regional source activity and connectivity measures, significantly overlap with the default mode network (DMN). All decrease with CLZ treatment in responding SCZs. Conclusions: A set of EEG-derived discriminating features which normalize as a result of CLZ treatment was identified. These discriminating features define a network that shares significant commonality with

⇑ Corresponding author. Address: 1280 Main Street West, Hamilton, ON L8S 4K1, Canada. Tel.: +1 832 769 2245. 1

E-mail address: [email protected] (M. Ravan). Also with Mood Disorders Program, St. Joseph Hospital, Hamilton, ON, Canada.

http://dx.doi.org/10.1016/j.clinph.2014.07.017 1388-2457/Ó 2014 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.

Please cite this article in press as: Ravan M et al. A machine learning approach using auditory odd-ball responses to investigate the effect of Clozapine therapy. Clin Neurophysiol (2014), http://dx.doi.org/10.1016/j.clinph.2014.07.017

2

M. Ravan et al. / Clinical Neurophysiology xxx (2014) xxx–xxx

the DMN. Our findings are consistent with those of previous literature, which suggest that regions of the DMN are hyperactive and hyperconnected in SCZ subjects. Our study shows that these discriminating features decrease after treatment, consistent with portions of the DMN normalizing with CLZ therapy in responsive subjects. Significance: Machine learning is proposed as a potentially powerful tool for analysis of the effect of medication on psychiatric illness. If replicated, the proposed approach could be used to gain some improved understanding of the effect of neuroleptic medications in treating psychotic illness. These results may also be useful in the development of new pharmaceuticals, since a new drug which induces changes in brain electrophysiology similar to those seen after CLZ could also have powerful antipsychotic properties. Ó 2014 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.

1. Introduction Schizophrenia is a severe psychotic disorder affecting approximately 1% of the world’s population (Kelly, 2006). Among the many anti-psychotic drugs used in the treatment of schizophrenia, Clozapine (CLZ) is recognized to have superior therapeutic effectiveness in the treatment of chronic, medication-resistant syndromes, commonly estimated to constitute at least one-third of all cases (e.g., Essali et al., 2009). However, CLZ is potentially toxic, producing agranulocytosis, adverse metabolic changes and myocarditis in some patients (Schulte, 2003; Meltzer, 2012). Studies investigating the neuro-physiological actions of CLZ and other agents are of particular interest as they may provide clues to both underlying pathophysiology and mechanism of action of this exceptionally effective drug. Quantitative electroencephalography (QEEG or EEG) has shown some promise in this regard. QEEG abnormalities in schizophrenic (SCZ) subjects and changes due to CLZ therapy have been the focus of several clinical studies (see e.g., Gunther et al., 1993; Malow et al., 1994; Freudenreich et al., 1997; Hughes and John, 1999; Knott et al., 2000, 2001, 2002; Adler et al., 2002; Gross et al., 2004; Birca et al., 2006; Coburn et al., 2006; Dunki and Dressel, 2006; Oikonomou et al., 2006; Sakkalis et al., 2006; Boutros et al., 2008; MacCrimmon et al., 2012). Knott et al. (2001) have used conventional statistical analytical methods in a study to determine the effect of chronic and acute CLZ therapy on the SCZ brain. They evaluated coherences (connectivity) from resting EEGs and identified changes in coherence patterns between normal vs. pre-treatment SCZ subjects, and then also between the pre-treatment state and both the acute and chronic post-treatment states after CLZ therapy. Takahashi et al. (2013) used exact low-resolution electromagnetic tomography analysis (eLORETA) (Pascual-Marqui et al., 2006) to identify neural sources of mismatch negativity (MMN) and P3a. They found MMN deficits in SCZ that were associated with reduced activations in discrete medial frontal brain regions. Na et al. (2002) used information theoretic constructs such as average cross mutual information (A-CMI) to characterize connectivity patterns in SCZ subjects. They concluded that SCZ patients were subject to left temporal lobe deficits and inter- and intra-hemispheric overconnectivity. Although there are numerous such studies identifying and characterizing deficits of schizophrenia, few have identified and characterized aspects of the brain that are normalized in subjects who respond to neuroleptic treatment in general, and to CLZ therapy specifically. Our objective is therefore to determine what features if any in the SCZ brain normalize as a result of CLZ treatment. We use machine learning (ML) methods to identify EEG-derived features that satisfy the following set of four discriminating conditions. That is, we wish to find features that simultaneously

(1) discriminate all before treatment (BT) SCZ from healthy volunteer (HV) subjects,2 (2) discriminate the BT condition from the after treatment (AT) condition in most responsive (MR) SCZ subjects, (3) discriminate AT least responsive (LR) SCZ subjects from HV subjects, (4) specifically do NOT discriminate AT most responsive (MR) SCZ subjects from HV subjects. We conclude that a set of features simultaneously satisfying these discriminating conditions normalizes as a result of treatment; i.e., that the brain function associated with such a set of features becomes indistinguishable from normal. This study demonstrates that it is indeed possible to identify such a feature set using the methodology explained in this paper. This finding may be of significant value in the development of new psycho-pharmaceutical agents, since a new drug, which demonstrates this same effect on the brain could have psychotropic properties similar to those of CLZ, potentially without the toxicity. Machine learning, which includes pattern recognition and dimensionality reduction paradigms, is finding increasing application in psychiatry, particularly when multi-dimensional, noisy, highly complex data or multi-modal data sets are analyzed together, (see e.g., Gallinat and Heinz, 2006). For example, support vector machine (SVM) techniques that select spectro-temporal patterns from multichannel magnetoencephalogram (MEG) data collected during a verbal working memory task have been used to distinguish SCZ from control subjects (Ince et al., 2008). ML algorithms analyzing structural brain magnetic resonance images (MRIs) (Fan et al., 2007), functional MRI (fMRI) data (Guo et al., 2008) and combined genomic and clinical data (Struyf et al., 2008) have been employed to separate SCZ, bipolar and healthy control subjects. KhodayariRostamabad et al. (2010, 2013) have used ML methods to predict response to CLZ treatment and SSRI medications for depression. QEEG technology is considerably less expensive and more readily available than genetic analyses, MEG and fMRI. One of the most reliable brain source localization (BSL) methods to localize the EEG signal is the source montage approach (Scherg et al., 2002; Miller et al., 2007). The source montage approach is a BSL method that localizes the EEG signal and estimates composite brain source activity emanating from predefined low-resolution regions throughout the cortex from the signals received from the scalp electrodes. An important feature of the beamforming class of BSL methods, to which the source montage approach belongs, is that it explicitly incorporates a spatial filtering procedure that suppresses interference from sources in neighboring regions (Ille 2 This condition implies that the selected features are capable of diagnosing schizophrenia. However, since this capability is not directly related to the objectives of this study (the objectives being the identification of brain changes induced by CLZ treatment), this finding is not discussed any further in this paper.



et al., 2002). An advantage of the source montage approach is that it provides useful results (although with low resolution) from a relatively small number of electrodes. In this study, features are extracted from auditory odd-ball responses that have been localized using the source montage approach. This approach is expected to yield features of enhanced salience and therefore improve our chances of finding features which satisfy our four discriminating conditions. Since odd-ball responses appear to distinguish SCZ from normal subjects (e.g., Takahashi et al., 2013; Linden, 2005), features derived from these signals might offer useful information regarding the pathophysiology of schizophrenia. BSL is also relevant to generating features of enhanced salience, since a waveform extracted directly from a brain source affected by schizophrenia will contain more salient information (i.e., have a higher signal-to-noise ratio) than one received from surface electrodes, due to contamination from volume conduction by sources irrelevant to our purposes. Furthermore, BSL methods give us some capability to extract waveforms from deep brain regions, thought to be generators of odd-ball responses. There have been many studies that have characterized the SCZ brain state using various measures of brain connectivity. Coherence of EEG activity between different brain regions is considered a measure of connectivity. Abnormalities of EEG coherence are present in patients with SCZ with both reduced and increased coherence compared with healthy control subjects (Medkour et al., 2010). Winterer et al. (2003) observed altered fronto-temporal and temporo-parietal cP300 coherence in SCZ patients compared with healthy control subjects. Knott et al. (2002) used spectral coherence values as a means of differentiating SCZs from normals. Canuet et al. (2011) used the lagged phase synchronization technique as a connectivity measure, whereas Na et al. (2002) used mutual information analysis for the same purpose. Thus there are many valid techniques for measuring connectivity. In the present study, we propose the use of the cross-power spectral density (CPSD) function, which is a spectral measure closely related to coherence. Consider the crosscorrelation function Rab(s) between the signals xa(t) and xb(t) received from two EEG electrodes. This function is defined as

Rab ðsÞ ¼

Ns 1X xa ðtÞxb ðt þ sÞ: N t¼1

ð1Þ

The CPSD function Sab(f) is the Fourier transform (i.e., frequency representation) of Rab(s). In comparison, the magnitude coherence function Cab(f) is related to the CPSD function as follows:

jSab ðf Þj C ab ðf Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; Sa ðf ÞSb ðf Þ

ð2Þ

where || denotes magnitude, and Sab(f) and Sb(f) are the conventional power spectral density values corresponding to signals xa(t) and xb(t) respectively, at frequency f. Thus the coherence function is a normalized version of the CPSD. In effect, the CPSD is a combination of the magnitude coherence between the signals xa(t) and xb(t) (through the term Cab(f)), and the (square-root) power spectral density values of these signals. Thus CPSD is a combination of the activity of two specified brain regions and the connectivity between them. CPSD values are chosen for this study, since they have been shown to offer enhanced discrimination over that obtained using coherence values alone and also because they simplify comparisons with other studies. 2. Methods 2.1. Subjects Auditory odd-ball evoked potentials were collected from 47 SCZ subjects both BT and AT with CLZ, as well as from 66 HV controls.

3

All subjects were unpaid volunteers. SCZ subjects were recruited from St. Joseph’s Hospital, Centre for Mountain Health Services, Hamilton, Ontario. All SCZ subjects met both DSM-IV criteria for schizophrenia and the Kane et al. (1988) criteria for treatment resistance. Patients meeting these criteria are both chronically and severely symptomatic. Although our SCZ subjects were not yet on CLZ at the time of the BT QEEG collection, these subjects were receiving one or more conventional antipsychotics with a mean daily dose in chlorpromazine equivalents of 628 mg (range 40–2740 mg.). AT QEEG data was collected after a mean duration of CLZ treatment of 1.4 years (range 0.4–4.6 years). The mean daily dose of CLZ employed was 347 mg (range 50–800 mg). Demographic characteristics of the SCZ participants are: age [years]: avg. = 37.3, std = 9.44, min = 22, max = 57, at start of treatment, gender: 29 male subjects (61.7%) and 18 female subjects (38.3%). The HV demographics are: age [years]: avg. = 37.1, std = 15.8, min = 18, max = 74, gender: 36 male subjects (54.5%) and 30 female subjects (45.5%). EEG data was collected only once from the HV subjects. The clinical rating procedure was devised in the context of an unrelated earlier naturalistic retrospective study of treatment-resistant SCZ patients being considered for CLZ treatment and is described in Khodayari-Rostamabad et al. (2010). The subjects in this present study were included in the previous study (Khodayari-Rostamabad et al., 2010). An experienced clinician reviewed all the available clinical descriptive information of the patient’s symptomatology prior to beginning a course of CLZ (BT). Symptoms, corresponding to those described in the Positive and Negative Symptoms of Schizophrenia (PANSS) severity rating scale, were rated as: ‘‘present’’, ‘‘moderate’’ or ‘‘severe’’ on a one to six point scale. Only explicitly described symptoms were scored and the clinical rater was instructed not to infer the presence of potential symptoms. The same rating procedure was repeated AT, based on case records and clinician interviews, at the time when a decision would be made regarding further on-going long-term maintenance CLZ therapy. Subjects who demonstrated at least 35% reduction in severity score from baseline were designated ‘‘most responsive’’ (MR) while those with severity score reduction of less than 35% were designated ‘‘least responsive’’ (LR). There were 20 of the former and 27 of the latter. 2.2. EEG recordings The EEG data were recorded using a QSI-9500 system. Electrodes, referenced to linked ears with impedances below 5 kO were placed in standardized locations according to the International 10/ 20 system. There were 19 data electrodes (in positions Fp1, Fp2, F7, F3, Fz, F4, F8, T7, C3, Cz, C4, T8, P7, P3, Pz, P4, P8, O1, O2) as well as 4 additional muscle, vertical and horizontal eye movement auxiliary artifact channels. An auditory odd-ball paradigm was used, consisting of two different 50 ms, 91.5 dB tone pip stimuli, which were delivered in random order over stereo headphones. The ‘‘common’’ stimulus is a 1000 Hz tone (80% of all presentations), while the ‘‘rare’’ (also referred to as ‘‘target’’ or ‘‘odd-ball’’) stimulus is a 2000 Hz tone (20% of all presentations). The inter-stimulus interval varied from 1.25 to 3 s on a pseudo-random basis. The subject was seated comfortably with eyes opened and fixated and instructed to indicate by finger lift when a rare stimulus was heard. The recording session persisted until a total of 120 rare responses were collected. The subject’s performance was monitored continuously for accuracy with breaks as required to maintain attention. Accuracies in identifying the rare tones were maintained at or above a 90% level. Epochs with duration of 1024 ms were sampled at 250 Hz. The system separately recorded the average of the ‘‘common’’ and ‘‘rare’’ epochs at each electrode until the total of 120 successful rare epochs were recorded. The signals were band pass filtered


4


between 0.5 and 80 Hz and notch filtered at 60 Hz during the recording. The signals were then digitally band pass filtered after recording between 0.5 and 20 Hz to partially mitigate the effects of eye movement and muscle artifacts. Only the averages of the odd-ball responses from each of the electrodes were used in subsequent analysis. 2.3. De-artifacting The recording system automatically rejected epochs with signals exceeding 100 uV peak to peak on any channel. The subject was given frequent ‘‘blink holidays’’ to alleviate eye blinking during recording. Muscle artifacts were alleviated by seating the subject in a comfortable reclining chair with support with pillows as appropriate. The subject was continuously observed by a trained technician during recording to monitor and identify epochs containing eye and muscle artifacts. Any such identified epochs were removed from the recordings. Any residual eye or muscle artifacts were identified and removed off-line by visual inspection of the recorded signal, corroborated by indication of strong signal on any of the 4 auxiliary artifact channels. In addition, in some cases, muscle artifacts are further suppressed by the inherent averaging operation implicit in the evaluation of the CPSD parameters, as indicated in Eq. (1). Muscle artifacts appearing on any electrode may be assumed uncorrelated with the true EEG signal component appearing on any electrode. In many cases, muscle artifact signal components appearing on a pair of electrodes are uncorrelated with each other (particularly where the electrodes are widely separated in space). On the other hand, the signal components common to the two electrodes of any selected CPSD parameter may be assumed coherent. Thus, in the averaging process of Eq. (1), the artifact in these cases adds incoherently, while the signal component adds coherently. This phenomenon tends to reduce the level of the artifact relative to that of the desired signal component. 2.4. Brain source localization We use the BESA 5.1.8 EEG review and analysis program (MEGIS Software, Gräfeling Germany) for filtering and source montage processing (http://www.besa.de). The BESA source montage method (Mosher et al., 1992; Scherg, 1990; Scherg et al., 2002) identifies source waveforms (montages) from 15 pre-defined regions of the brain: midline fronto-polar (FpM); frontal left (FL), frontal midline (FM) and frontal right (FR); anterior temporal left (TAL) and anterior temporal right (TAR); central left (CL), central midline (CM) and central right (CR); posterior temporal right (TPR) and posterior temporal left (TPL); parietal left (PL), parietal midline (PM) and parietal right (PR); and midline occipito-polar (OpM) areas. These regions are depicted in Fig. 1a. The source montage method is appropriate for use in the case of a relatively small number of electrodes (Scherg et al., 2002). Composite source activity in each region is modeled as a single regional source in the respective region. The sources are modeled as a dipole, specified by three mutually orthogonal coordinates. These are the horizontal, vertical, and radial components in the spherical coordinate system. Therefore the number of brain source montages is 3 15 = 45 for 15 regions. Regional sources provide more stable fits than discrete dipoles and can account for any activity in an extended brain region, independent of the current orientations. 2.5. An overview of the machine learning procedure A necessary component of the ML process is the existence of a set of training patterns (training set). For this study, the training

FpM FL

FR FM

TAL

TAR CL

CM

CR

TPL

TPR PL

PM

PR

OpM Fig. 1. The configuration of regions that are resolvable using the source montage brain source localization method (BESA, Scherg et al., 2002; Miller et al., 2007). The colors distinguish the left, medial and right regions. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

set consists of the EEG responses for each of the Mt = 113 subjects. Also included are the labels that indicate the various conditions used in this study; i.e., the BT vs. AT, MR vs. LR, and SCZ vs. HV conditions corresponding to each subject. A somewhat more detailed explanation of ML in the clinical context is available in Khodayari-Rostamabad et al.(2010, 2013) and Ravan et al. (2011). There are four phases in the design of a machine learning procedure. They are the feature extraction, feature selection, classification and evaluation phases. The first step is to extract candidate features from all of the BT and AT source montage data. In our study, the set of candidate features consisted of the auto-power spectral density (auto-PSD) values of 15 extracted source waveforms and the cross-power spectral density (CPSD) values between all pairs of source waveforms, at various frequencies. The sampling frequency of 250 Hz and the duration of 1024 ms imply that the odd-ball source montage signal from each region extends over an interval of 256 samples. Therefore the one-sided PSD function has (256/2 + 1) = 129 samples. There were 15 auto-PSD functions from the 15 regions and 105 cross-PSD functions between all pairs of the 15, in each direction. Hence the total number of candidate features is Nc = 129 (105 + 15) 3 = 46440 The CPSD functions were evaluated using the MATLAB built-in function ‘‘CPSD’’, which uses a modified Welch periodogram technique (Welch, 1967) to compute the spectral estimates. Default settings were used. This process divides the available time window into a specified number of (possibly overlapping) segments, applies a Hamming window on each segment, and then evaluates the magnitude-squared Fourier transform for each segment. The resulting magnitude squared functions are then averaged over the segments. All CPSD values were transformed into a decibel (dB) scale using the transformation 10 log(|CPSD|), where || denotes magnitude. All features were then normalized using their corresponding z-score value before being fed to the feature selection and classifier processes. The z-score value was evaluated by calculating the means li and the standard deviations ri of each feature f i ; i ¼ 1; 2; . . . ; N c where Nc is the number of candidate features. These features vary across all locations and frequencies. The means li and variance r2i for each of these features are obtained by averaging the ith feature value over all subjects. The normalized z-score value for the ith feature and jth subject is then calculated as



zi;j ¼

f i;j li

ri

ð3Þ

Feature values with absolute z-values larger than 5 are clipped so that their value is in the range 5 6 z 6 5, in order to suppress outliers. After extracting candidate features, the second step in the machine learning process is feature reduction, or ‘feature selection’, which is critical to the performance of the corresponding classifier. We wish to identify only a set of Nr most salient features from the extensive list of candidate features which are relevant to distinguishing between the four discriminating conditions listed in Section 1. Specifically, we wish to find a common feature set that simultaneously discriminates between HV and SCZ subjects BT (condition 1), and BT vs. AT in MR SCZ subjects (condition 2). In this study, we found empirically that if the feature selection algorithm was implemented to select a feature set that satisfied only conditions 1 and 2, then the same set serendipitously satisfied conditions 3 and 4 as well. The proposed procedure for feature selection uses the ‘‘regularized feature selection’’ method of Peng et al. (2005). The output of the feature selection process is a set of indices that identify which of the candidate features are to be included in the set of Nr most relevant features. In order to avoid choosing features that are dominant in just a few patterns, a leave-one-out (LOO) procedure (Hastie et al., 2009) is used to select the best Nr features. The LOO procedure is an iterative process, where in each iteration all the data associated with one particular subject is omitted from the training set. The iterations repeat until all subjects have been omitted once. In the proposed feature selection scheme, for each iteration, a list of the best kNr, k > 1 features is determined using Peng’s method (Peng et al., 2005). For this study the value of k is chosen to be 2. After all iterations are completed, the Nr features with the highest number of repetitions (probability of appearance) among the available lists were selected as the final set of selected features. In this study, Nr = 5. This value was determined on the basis that it is the lowest value which gave adequate performance. It is desirable to use as small a number as possible so as to avoid over-fitting, as described later in the paper. The third step in the ML procedure is the specification of the classifier. Here the job of the classifier is to input a reduced feature vector and output the corresponding class label. In our study, we used the fuzzy c-mean (FCM) algorithm to implement the classifier. This method is developed by Dunn (1973) and improved by Bezdek (1981). This algorithm is an iterative classification method having some advantages with respect to other classifiers, the most prominent of which is its high generalization capacity for a reduced number of training trials. The final step in the proposed ML procedure is performance evaluation. This task is executed using a second LOO procedure. In each iteration (fold) of this second LOO process, the set of features corresponding to one particular subject is again omitted from the total available training set. The classifier is trained using the remaining training set and the resulting structure is tested using the omitted subject. The test result is compared to the known result (which is provided by the training set). The process iterates over the training set, each time using a different omitted subject, until all subjects have been omitted/tested once. The same set of previously-identified features is used in each fold. In this way, considering the small size of our available training set, we can obtain an efficient estimate of the performance of the process. LOO cross validation is useful because it does not waste data and provides an asymptotically unbiased estimate of the averaged classification error probability over all possible training sets (Theodoridis and Koutroumbas, 2008).

5

3. Results A set of typical averaged odd-ball responses for the BESA source montage regions FM, CM and PM are shown in Fig. 2. The three traces on each sub-figure correspond to the response from a single BT SCZ, AT SCZ and normal subject. The post-stimulus interval between 0 and 200 ms was used to calculate the baseline value for each trace that was then subtracted from the entire epoch. A set of 5 features satisfying all four discriminating conditions of section 1 have been identified in Table 1. These are now referred to as discriminating features in the sequel. A ‘‘wiring diagram’’, showing corresponding topography of the selected feature locations in the brain, is depicted in Fig. 3, where a line between two regions indicates the cross-power spectrum function between these regions has been selected as a discriminating feature. The values of the discriminating features for distinguishing HV vs. all SCZ subjects BT (condition 1) is shown in Table 2, along with their corresponding t- and p-values. The p-values are only valid under the assumption that the features are Gaussian distributed. Note that the average feature values are uniformly lower in this case for the HV compared to the SCZ group. Similarly, discriminating feature values corresponding to each of the discriminating conditions 2–4 are given in Tables 3–5 respectively. From Table 3, it may be observed that each of the selected discriminating features is significantly distinct between the BT and AT conditions. This implies these features have changed due to the CLZ treatment, in a direction which normalizes their values (as seen with the help of Table 2). Table 4 shows that the feature values for LR SCZ subjects remain significantly different from those of normals, whereas the feature values for MR SCZ subjects are indistinguishable after treatment from the values of normals, as indicated in Table 5. Thus, our results indicate the selected feature set indeed satisfies the four discriminating conditions. The t-statistic values in Tables 2, 4 and 5 were produced using a Welch t-test (which is the conventional Student t-test generalized for unequal sample sizes and variances between the groups (Welch, 1947), whereas in Table 3 (BT vs. AT in MT SCZ), a paired Student t-test for unequal variances was used. Caution must be exercised when assessing the discriminating qualities of each feature individually. If there are significant correlations between the features, then the joint behavior of the features can be more discriminating than each feature individually. However, it may be noted from the tables that in this case, each feature is individually discriminative for conditions 1, 2 and 3, and non-discriminative for condition 4, as desired. Contingency the respective classifier performances in testing each of the four discriminating conditions are shown in Tables 6–9. We see that the classifiers are capable of accuracies of 81.4%, 85.0%, 79.6% and 48.8% respectively, corresponding to the four conditions. We see that the accuracies for each of the first three conditions are close to or above the 80% level, whereas the accuracy corresponding to condition 4 is 48.8%, which is the level expected if there is to be no discernible distinction between the MR SCZ subjects AT, and HV subjects, as desired. Thus, we see that the proposed ML procedure for satisfying each of the four conditions is indeed conclusive in its level of performance. As an additional demonstration of performance, we show the clustering behavior of the discriminating features corresponding to the 20 MR subjects in both of the BT and AT conditions in Fig. 4 (corresponding to discriminating condition 2). This figure is generated by projecting the 40 total training samples in the 5 dimensional feature space onto the two major nonlinear principal components of the feature space, using the kernelized principle component analysis (KPCA) method with a Gaussian kernel (Muller et al., 2001). The KPCA method is used for visualization


6

M. Ravan et al. / Clinical Neurophysiology xxx (2014) xxx–xxx Table 1 The Nr = 5 discriminating features which jointly satisfy all four discriminating conditions of section 1. Ci(A, B) is the log of absolute cross-PSD between regions A and B (shown in Fig. 1) in the ith direction, where i = h (horizontal), v (vertical), or r (radial). Feature #

Feature

1 2 3 4 5

Ch(FpM, TAR) at f = 19 Hz Ch(FpM, PR) at f = 19 Hz Ch(FpM, FM) at f = 16 Hz Ch(CM, PR) at f = 14 Hz Cv(OpM, PR) at f = 2 Hz

FpM

FM TAR CM

PR OpM Fig. 3. Topography of the selected feature locations on the brain. A line between two regions indicates the respective cross-PSD function between these regions at a particular frequency was selected as a feature.

4. Discussion and conclusion

Fig. 2. Averaged typical odd-ball responses from various regions corresponding to the BESA source montage method (as shown in Fig. 1): (a) the FM, (b) the CM, and (c) the PM regions. The three traces in each figure show the response from a normal subject, and from a single MR SCZ subject with BT and AT data being displayed.

purposes only. Note that even though excellent performance is demonstrated with this 2-dimensional representation, better overall performance is obtained in the Nr = 5 dimensional feature space.

We have identified a set of 5 ‘‘discriminating features’’ that distinguish pre-treatment from post-treatment odd-ball auditory responses in a group of 20 SCZ subjects who respond well to CLZ therapy. These discriminating features are listed in Table 1. These same discriminating features are also found to distinguish these most responsive SCZ subjects from HV before CLZ treatment but after CLZ treatment they no longer do so. Since these features discriminate BT from AT conditions in MR subjects, it follows that they must have changed BT vs. AT. Moreover the lack of differentiation between the most responsive SCZ and the HV normal after treatment suggests that the therapeutic effect of CLZ is associated with a shift of brain source activity to more closely resemble that of a normal state. One of the issues in any machine learning application is overtraining or over-fitting. Over-fitting happens when the feature selection and classifier design processes over-adapt to the specific training set, with the result that the resulting structure performs well with the given training set, but does not generalize well to new samples. Over-training can be controlled by keeping the number of selected features low in comparison to the number of training samples. In Table 3 (MR SCZ, BT vs. AT) we have 40 training samples and 5 features, for a feature-to-subject ratio of 12.5%, which is sufficiently small to avoid major problems with overtraining. With regard to the results of Table 2 (BT, all SCZ vs. HV) we have available 47 SCZ + 66 HV = 113 training samples, with 5 features used. Thus, the ratio of features to subjects is 4.4% in this


7


Table 2 Average values (±std) of the discriminating features for the HV vs. all SCZ subject BT (discriminating condition 1). It may be observed that the average feature values are uniformly lower for the HV group. In this and the following tables, the p-values correspond to a Gaussian distribution on the feature values. Feature #

Feature

Average (±std) for HV

Average (±std) for (R + NR) BT

t-Statistic values

p-Value

1 2

Ch(FpM, TAR) at f = 19 Hz Ch(FpM, PR) at f = 19 Hz

0.80 (±1.27) 0.71 (±1.31)

0.58 (±0.92) 0.51 (±1.03)

6.6982 5.5355

1 107

3

Ch(FpM, FM) at f = 16 Hz

0.40 (±1.30)

0.37 (±1.20)

3.2468

7:78 104

4

Ch(CM, PR) at f = 14 Hz

0.25 (±1.33)

0.63 (±1.02)

3.9780

5

Cv(OpM, PR) at f = 2 Hz

0.45 (±1.11)

0.92 (±0.72)

7.9498

5:75 105 0

0

Table 3 Average values (±std) of the discriminating features for most responsive subjects, BT vs. AT. (discriminating condition 2). The average feature values are uniformly lower AT than BT. Feature #

Feature

Average (±std) for BT (MR)

Average (±std) for AT (MR)

t-Statistic values

p-Value

1

Ch(FpM, TAR) at f = 19 Hz

0.70 (±1.30)

0.70 (±0.92)

3.9228

2 3 4 5

Ch(FpM, PR) at f = 19 Hz Ch(FpM, FM) at f = 16 Hz Ch(CM, PR) at f = 14 Hz Cv(OpM, PR) at f = 2 Hz

0.60 0.55 0.45 0.35

0.60 0.55 0.30 0.30

(±0.88) (±1.28) (±1.30) (±1.30)

3.5427 3.1650 1.9687 1.7786

4:0 104 0.001 0.0025 0.031 0.045

(±1.23) (±0.89) (±1.10) (±0.99)

Table 4 Average values (±std) of the discriminating features for HV subjects vs. least responsive SCZ subjects AT (discriminating condition 3). The average feature values are uniformly lower for HVs than SCZ subjects. Feature #

Feature


Average (±std) for AT (LR)

t-Statistic values

p-Value

1 2

Ch(FpM, TAR) at f = 19 Hz Ch(FpM, PR) at f = 19 Hz

0.80 (±1.27) 0.71 (±1.31)

0.95 (±0.80) 0.67 (±0.88)

7.9759 5.9014

6:0 108

3

Ch(FpM, FM) at f = 16 Hz

0.40 (±1.30)

0.78 (±1.01)

4.6868

7:74 106

4

Ch(CM, PR) at f = 14 Hz

0.25 (±1.33)

0.44 (±1.05)

2.6532

5:11 103

5

Cv(OpM, PR) at f = 2 Hz

0.45 (±1.11)

0.73(±0.78)

5.8133

9:0 108

0

Table 5 Average values (±std) of the discriminating features for HV subjects vs. most responsive SCZ subjects AT (discriminating condition 4). The average feature values are not significantly different between the HV and SCZ groups, indicating the feature values have normalized with treatment. Feature #

Feature


1 2 3 4 5

Ch(FpM, TAR) at f = 19 Hz Ch(FpM, PR) at f = 19 Hz Ch(FpM, FM) at f = 16 Hz Ch(CM, PR) at f = 14 Hz Cv(OpM, PR) at f = 2 Hz

0.80 0.71 0.40 0.25 0.45

(±1.27) (±1.31) (±1.30) (±1.33) (±1.11)

Average (±std) for AT (MR) 0.70 0.60 0.55 0.30 0.30

Table 6 Classification performance discriminating all SCZ from HV subjects using BT data (condition 1). There are a total of 47 SCZ subjects and 66 HV subjects. In this and the following tables, TCA refers to ‘‘total classification accuracy’’. Classes

Predicted SCZ

Predicted HV

% correct

TCA

SCZ HV

41 15

6 51

87.2 77.3

81.4%

Table 7 Classification performance discriminating between BT and AT in the 20 most responsive (MR) subjects with SCZ (condition 2). Classes

Predicted BT (MR)

Predicted AT (MR)

% correct

TCA

BT (MR) AT (MR)

17 3

3 17

85 85

85%

case. Furthermore, for the cases of Tables 4 and 5 (AT least responsive vs. HV, and AT most responsive vs. HV, respectively), we have 93 and 86 subjects and the same 5 features, for feature to subject ratios of 5.4% and 5.8%, respectively. These low percentages in each

(±0.92) (±0.88) (±1.28) (±1.30) (±1.30)

t-Statistic values

p-Value

0.3870 0.4324 0.4574 0. 499 0.4670

0.350 0.334 0.325 0.441 0.322

Table 8 Classification performance discriminating least responsive (LR) SCZ from HV using AT data (condition 3). There are 27 SCZ subjects and 66 HV subjects. Class

Predicted AT (LR)

Predicted HV

% correct

TCA

AT (LR) HV

23 15

4 51

85.2 77.3

79.6%

Table 9 Classification performance discriminating the 20 most responsive (MR) SCZ from the 66 HV subjects using AT data (condition 4). Class

Predicted AT (MR)

Predicted HV

% correct

TCA

AT (MR) HV

9 33

11 33

45 50.0

48.8%

case suggest that the likelihood of over-training occurring in these results is low. The use of brain-source localization methods is critical in this study. BSL techniques (in comparison to using the raw signals from the electrodes) allow us to isolate the waveform emanating from a


8

M. Ravan et al. / Clinical Neurophysiology xxx (2014) xxx–xxx 3 MR(BT) MR(AT) 2

PC2

1

0

-1

-2

-3 -2.5

-2

-1.5

-1

-0.5

0 PC1

0.5

1

1.5

2

2.5

Fig. 4. Compressed subject-wise scatter plot of the feature space showing BT (blue circles) vs. AT (red squares) of the most-responsive subjects, projected onto the first two major nonlinear principal component directions. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

specified brain region, and as such, suppress spurious irrelevant and interfering signals from other nearby regions, thereby increasing the signal-to-noise ratio of the desired signal. BSL methods can also suppress signals received on the electrodes due to volume conduction, which may give rise to spurious connectivities between regions. Furthermore BSL in conjunction with ML methods allow us to better isolate which brain regions are most affected by the illness. It is interesting to compare fundamental differences between conventional approaches to biomarker (feature) identification, and the machine learning approach proposed in this paper. The conventional approach to the problem is based on statistical significance tests conducted over an available training set. A set of EEG parameters (features) is identified as those which individually have significant variations between two groups (classes), i.e., SCZ and normal. The integrity or performance of the resulting features is established due to the fact that the probability over the training set that a feature will not predict its class is inherently low, by design. In the conventional case when multiple features are considered, evaluation of the joint effectiveness of a feature set must take into account any possible correlations between the features. This requires multivariate statistical tests (such as the Hotelling t-test), which require a stable estimate of the feature covariance matrix. This is a difficult prospect in the small-sample case. But most significantly, the conventional approach to selecting a multivariate set from a large number of candidate features is an ‘‘np-hard’’ problem, meaning that computer execution times can become intractable. In contrast, the ML approach formulates a rudimentary prediction model, again based on a training set. The feature selection process is of fundamental relevance to the ML paradigm. The feature selection algorithms identify a set of features that jointly are (nearly) optimally statistically dependent, i.e., predictive, of the class, in a computationally tractable manner. The features are then fed into a classifier (the model) which outputs the prediction. Integrity or performance in this case is established by monitoring the prediction accuracy of the system, over the training set. In brief, the fundamental advantage of the ML approach is the improved efficiency of the feature selection process. It has long been recognized that CLZ, which is known to lower seizure thresholds, has had an obvious effect on EEG data. Recent work (MacCrimmon et al., 2012) confirms previous observations

of increased low-frequency activity frontally and adds the novel observation of a right-sided anterior shift of alpha activity. Knott et al. (2002) measured QEEG coherence in 17 patients’ pre-and six weeks post-CLZ. They concluded that CLZ appeared to alter reduced local coupling in temporal regions thereby shifting power spectrums towards more normative values. Studies exploring CLZ effects using P300 responses are relatively few (note that P300 responses are a component of the overall odd-ball response, and hence the two phenomena are closely related in the present context). Most of those that exist to date limit observations to midline electrodes, without source localization, and consider only power spectral values (i.e., activity, rather than connectivity levels). They report normalization (i.e., increases) of P300 waveform amplitudes with CLZ. P300 studies of conventionally treated SCZ populations are more numerous and generally show abnormalities of either reduced amplitudes or aberrant topography. For example, Bolsche et al. (1996) in one of the few studies using both uni- and bilateral stimuli and a 10/ 20 montage, report atypical right anterior hemispheric activity with unilateral left hemispheric stimulation in un-medicated SCZ. More recently, Pae et al. (2003) studied 20 conventionally medicated subjects with SCZ and 21 matched controls using low resolution electromagnetic tomography (LORETA) imaging (Pascual-Marqui et al., 1994) of P300 generators. The SCZ subjects showed significant P300 current density reduction in the left temporal area and left inferior parietal areas while both left prefrontal and right orbitofrontal areas were relatively activated. Consistent with the preceding observations, our discriminating features primarily involve frontal and right temporal regions. Contemporary neuroanatomical studies indicate separate P300 generators in the anterior subiculum and posterior hippocampus (Ludowig et al., 2010) and implicate frontal involvement in SCZ symptomatology (Goghari et al., 2010), all of which support the above idea that the regions we have identified might be normalized by CLZ. Others have demonstrated changes in P300 with antipsychotic medication treatment which are associated with improvement (Gallinat et al., 2001; Sumiyoshi et al., 2009). In a manner congruent with our own findings, the P300 changes noted after CLZ were apparent only in the SCZ who showed improvement (Gallinat et al., 2001). Of the discriminating features, listed in Table 1, that have been identified in the present study as normalizing, all decrease with CLZ treatment in MR SCZ subjects, as shown in Table 3. These features are predominantly in the beta-band, and measure joint activity (CPSD values) between a medial and a right-hemispheric region. The most dominant features are the CPSD values between the pre-frontal and the right parietal or right anterior temporal regions. Several studies (e.g., Canuet et al., 2011) report intraand inter-hemispheric over-connectivity in SCZs relative to healthy subjects. Even though these studies use resting EEGs, this result is nevertheless consistent with those of the present study, since our discriminating features, which incorporate connectivity measures, are initially elevated in SCZ relative to HV subjects (Table 2), but then decrease (normalize) in MR subjects after treatment (Table 3). The fact that the p-values listed in Tables 2–4 are so small is worthy of note. From these tables we see that the discriminating features are individually highly effective at distinguishing between the two states for each respective condition. Although we are primarily interested in the aggregate behavior of multiple features over multiple conditions, the data of Table 2 suggest that each of our five discriminating features could potentially be individually useful as biomarkers in the diagnosis of schizophrenia. Full investigation of this concept remains a topic for future study. Activity in the DMN in the brain may represent underlying physiological processes during rest (Raichle and Snyder, 2007). This network is typically suppressed with task-related activity. It



is characterized by coherent neuronal oscillations at a rate lower than 0.1 Hz in areas such as the medial temporal lobe, part of the medial prefrontal cortex (MPFC), the posterior cingulate cortex (PCC), the adjacent ventral precuneus, and the parietal cortex (PC). Hyperactivity, hyperconnectivity, and reduced task-related suppression of activity in the DMN have been associated with disturbances of thought in schizophrenia by Whitfield-Gabrieli et al. (2009) who used fMRI to study brain activity and connectivity in SCZ subjects, with particular attention to the DMN. During rest and task, they found abnormally high functional connectivity in the DMN in SCZ subjects (and their first-degree relatives), particularly in the MPFC and the PCC. Whitfield-Gabrieli’s findings are in partial agreement with those of the present study. The discriminating features of Table 1 (CPSD values, which consist of a combination of both activity and connectivity) overlap neuroanatomically with the DMN. Electrophysiologically the DMN operates at very low frequencies (0.01 Hz), whereas the discriminating features identified in the present study are in the low beta-band. This difference is due in part to the fact that our EEG signals have been high-pass filtered during recoding above 0.5 Hz. Also Whitfield-Gabrieli et al. (2009) studied the DMN using fMRI, whereas the present study uses EEG odd-ball responses. Despite these differences, Whitfield-Gabrieli et al. (2009) clearly implicate hyperconnectivity and hyperactivity in the MPFC as being distinctive of schizophrenia. This area shows up in the present study as the ‘‘FpM’’ source montage region, associated with our BSL technique, shown in Fig. 1. It is interesting to note that the first 3 discriminating features of Table 1 include the FpM (i.e., MPFC) and an associated region. These associated regions are the (1) TAR (temporal anterior right) source montage region (which includes the medial temporal lobe, part of the DMN), (2) the PR (parietal right) region, also part of the DMN, and (3) the FM (frontal medial) region, which is apparently not included in the DMN. The PR region is also included with features 4 and 5 in Table 1, although the associated regions are not included in the DMN. Thus, it appears there is a strong association between the regions of the DMN and the discriminating features identified by the present study. Further, these features have been seen to be initially hyperactive and then decreasing after treatment (Table 3), facts which are consistent with these parameters normalizing in responsive individuals. Whitfield-Gabrieli et al. (2009) have also identified that negative correlation between the dorsolateral pre-frontal cortex (DLPFC) and the MPFC diminishes in SCZs during difficult (2-back) working memory tasks, compared to HV. However, the DLPFC (which would show up as the label ‘‘FL’’ or ‘‘FR’’), is not implicated in the present study. This may be due to the fact that the stimulus used in this study to generate the odd-ball responses do not place sufficient load on the DLPFC, which is activated with working memory tasks. A limitation of our study is that our clinical ratings were semistructured in nature. However these ratings were done well before the present analysis was conducted and so could not have been biased by knowledge of the present results. Another limitation is the use of concurrent medications. As is typical of conventional clinical practice, our patients were receiving a range of other psychotropic agents both before and while on CLZ. However our sample of 47 subjects is relatively large compared to most of the literature and has the advantage of a lengthy follow-up which can be expected to maximize the likelihood of detecting CLZ effectiveness (Meltzer et al., 1989). In addition, the naturalistic use of concurrent medication may allow our findings to be more easily extrapolated into the typical clinical setting. Our data suggest that combining electrical source mapping with ERP and ML analytic techniques may provide insights into the pathophysiology of SCZ and potentially, mechanism and site of action of CLZ and thus may have utility in the development of

9

better therapies for the SCZ psychoses. In particular, pharmaceuticals that induce EEG changes similar to those observed for CLZ may also have antipsychotic properties similar to CLZ, but, potentially, less toxicity. However, in light of the limitations discussed above, our results should be interpreted with caution pending further replication or supportive evidence from other investigative methodologies. Acknowledgements The authors gratefully acknowledge the financial support of the Natural Science and Engineering Research Council of Canada (NSERC), MITACS Canada, Magstim Ltd., Carmarthenshire, Wales, and the Ontario Technology Fund. All authors except MR hold equity in Digital Medical Experts Inc., and all authors except MR hold the US patent: ‘‘Expert system for determining patient treatment response’’, No. 8655817. References Adler G, Grieshaber S, Faude V, Thebaldi B, Dressing H. Clozapine in Patients with Chronic Schizophrenia: Serum Level, EEG and Memory Performance. Pharmacopsychiatry 2002;35:190–4. Bezdek JC. Pattern recognition with fuzzy objective function algorithms. New York: Plenum Press; 1981. Birca A, Carmant L, Lortie A, Lassonde M. Interaction between the flash evoked SSEPs and the spontaneous EEG activity in children and adults. Clin Neurophysiol 2006;117:279–88. Bolsche F, MacCrimmon DJ, Kropf S. The effect of laterality of stimulus presentation on auditory P300 topography in schizophrenia. J Psychiatry Neurosci 1996;21:83–8. Boutros NN, Arfken C, Galderisi S, Warrick J, Pratt G, Iacono W. The status of spectral EEG abnormality as a diagnostic test for schizophrenia. Schizophr Res 2008;99:225–37. Canuet L, Ishii R, Pascual-Marqui RD, Iwase M, Kurimoto R, Aoki Y, et al. Restingstate EEG source localization and functional connectivity in schizophrenia-like psychosis of epilepsy. PLoS One 2011;6:e27863. Coburn KL, Lauterbach EC, Boutros NN, Black KJ, Arciniegas DB, Coffey CE. the value of quantitative electroencephalography in clinical psychiatry: a report by the Committee on Research of the American Neuropsychiatric Association. J Neuropsychiatry Clin Neurosci 2006;18:460–500. Dunki RM, Dressel M. Statistics of biophysical signal characteristics and state specificity of the human EEG. Physica A 2006;370:632–50. Dunn JC. A fuzzy relative of the ISODATA process and its use in detecting compact well-separated clusters. J Cyber 1973;3:32–57. Essali A, Haj-Hasan NA, Li C, Rathbone J. Clozapine versus typical neuroleptic medication for schizophrenia. Cochrane Database Syst Rev 2009;1:CD000059. Fan Y, Shen D, Gur RC, Gur RE, Davatzikos C. COMPARE: classification of morphological patterns using adaptive regional elements. IEEE Trans Med Imaging 2007;26:93–105. Freudenreich O, Weiner RD, McEvoy JP. Clozapine-induced electroencephalogram changes as a function of Clozapine serum levels. Biol Psychiatry 1997;42:132–7. Gallinat J, Riedel M, Juckel G, Sokullu S, Frodl T, Moukhtieva R, et al. P300 and symptom improvement in schizophrenia. Psychopharmacology 2001;158:55–65. Gallinat J, Heinz A. Combination of multimodal imaging and molecular genetic information to investigate complex psychiatric disorders. Pharmacopsychiatry 2006;39:S76–9. Goghari M, Sponheim SR, MacDonald AW. The functional neuroanatomy of symptom dimensions in schizophrenia: a qualitative and quantitative review of a persistent question. Neurosci Biobehav Rev 2010;34:468–86. Gross A, Joutsiniemi SL, Rimon R, Appelberg B. Clozapine-induced QEEG changes correlate with clinical response in schizophrenic patients: a prospective, longitudinal study. Pharmacopsychiatry 2004;37:119–22. Gunther W, Baghai T, Naber D, Spatz R, Hippius H. EEG alterations and seizures during treatment with Clozapine: a retrospective study of 283 patients. Pharmacopsychiatry 1993;26:69–74. Guo Y, Bowman FD, Kilts C. Predicting the brain response to treatment using a Bayesian hierarchical model with application to a study of schizophrenia. Hum Brain Mapp 2008;29:1092–109. Hastie T, Tibshirani R, Friedman J. The elements of statistical learning: data mining, inference, and prediction. 2nd ed. Springer; 2009. Hughes JR, John ER. Conventional and quantitative electroencephalography in psychiatry. J Neuropsychiatry Clin Neurosci 1999;11:190–208. Ille N, Berg P, Scherg M. Artifact correction of the ongoing EEG using spatial filters based on artifact and brain signal topographies. Clin Neurophysiol 2002;19:113–24. Ince NF, Goksu F, Pellizzer G, Tewfik A, Stephane M. Selection of spectro-temporal patterns in multichannel MEG with support vector machines for schizophrenia classification. Conf Proc IEEE Eng Med Biol Soc 2008;2008:3554–7.


10


Kane J, Honigfeld G, Singer J, Meltzer H. Clozapine for the treatment-resistant schizophrenic: a double-blind comparison with chlorpromazine. Arch Gen Psychiatry 1988;45:789–96. Kelly D. Treatment considerations in women with schizophrenia. J Womens Health (Larchmt) 2006;15:1132–40. Khodayari-Rostamabad A, Hasey GM, MacCrimmon DJ, Reilly JP, de Bruin H. A pilot study to determine whether machine learning methodologies using pretreatment electroencephalography can predict the symptomatic response to Clozapine therapy. Clin Neurophysiol 2010;121:1998–2006. Khodayari-Rostamabad A, Reilly JP, Hasey GM, de Bruin H, MacCrimmon DJ. A machine learning approach using EEG data to predict response to SSRI treatment for major depressive disorder. Clin Neurophysiol 2013;124:1975–85. Knott V, Labelle A, Jones B, Mahoney C. EEG hemispheric asymmetry as a predictor and correlate of short-term response to Clozapine treatment in schizophrenia. Clin Electroencephalogr 2000;31:145–52. Knott V, Labelle A, Jones B, Mahoney C. Quantitative EEG in schizophrenia and in response to acute and chronic Clozapine treatment. Schizophr Res 2001;50:41–53. Knott VJ, LaBelle A, Jones B, Mahoney C. EEG coherence following acute and chronic Clozapine in treatment-resistant schizophrenics. Exp Clin Psychopharmacol 2002;10:435–44. Linden DE. The P300: Where in the brain is it produced and what does it tell us? Neuroscientist 2005;11:563–76. Ludowig E, Bien CG, Elger CE, Rosburg T. Two P300 generators in the hippocampal formation. Hippocampus 2010;20:186–95. MacCrimmon D, Brunet D, Criollo M, Galin H, Lawson JS. Clozapine augments delta, theta and right frontal EEG alpha power in schizophrenic patients. ISRN Psychiatry 2012;2012:1–8. Malow BA, Reese KB, Sato S, Bogard PJ, Malhotra AK, Su TP, Pickar D. Spectrum of EEG abnormalities during Clozapine treatment. Electroencephalogr Clin Neurophysiol 1994;91:205–11. Medkour T, Walden AT, Burgess AP, Strelets VB. Brain connectivity in positive and negative syndrome schizophrenia. Neuroscience 2010;169:1779–88. Meltzer HY, Bastani B, Kwon KY, Ramirez LF, Burnett S, Sharpe J. A prospective study of Clozapine in treatment-resistant schizophrenic patients. I. Preliminary report. Psychopharmacology 1989;99:S68–72. Meltzer HY. Clozapine: balancing safety with superior antipsychotic efficacy. Clin Schizophr Relat Psychoses 2012;6:134–44. Miller JW, Kim W, Holmes MD, Vanhatalo S. Ictal localization by source analysis of infra slow activity in DC-coupled scalp EEG recordings. Neuroimage 2007;35:583–97. Mosher JC, Lewis PS, Leahy RM. Multiple dipole modeling and localization of spatiotemporal MEG data. IEEE Trans Biomed Eng 1992;39:541–57. Muller KR, Mika S, Ratsch G, Tsuda K, Scholkopf B. An introduction to kernel-based learning algorithms. IEEE Trans Neural Networks 2001;12:181–201. Na SH, Jin SH, Kim SY, Ham BJ. EEG in schizophrenic patients: mutual information analysis. Clin Neurophysiol 2002;113:1954–60. Oikonomou T, Sakkalis V, Tollis IG, Micheloyannis S. Searching and visualizing brain networks in Schizophrenia. Springer Lecture Notes in Computer Science. Biol Med Data Anal 2006;4345:172–82. Pae JS, Kwon JS, Youn T, Park HJ, Kim MS, Lee B, et al. LORETA imaging of P300 in schizophrenia with individual MRI and 128-channel EEG. Neuroimage 2003;20:1552–60.

Pascual-Marqui RD, Michel CM, Lehmann D. Low resolution electromagnetic tomography: a new method for localizing electrical activity in the brain. Int J Psychophysiol 1994;18:49–65. Pascual-Marqui RD, Pascual-Montano AD, Lehmann D, Kochi K, Esslen M, Jancke L, et al. Exact low resolution brain electromagnetic tomography (eLORETA). Neuroimage 2006;31:S86. Peng H, Long F, Ding C. Feature selection based on mutual information: Criteria of max-dependency, max-relevance, and min-redundancy. IEEE Trans Pattern Anal Mach Intell 2005;27:1226–38. Raichle ME, Snyder AZ. A default mode of brain function: a brief history of an evolving idea. NeuroImage 2007;37:1083–90. Ravan M, Reilly JP, Trainor LJ, Khodayari-Rostamabad A. A machine learning approach for distinguishing age of infants using auditory evoked potentials. Clin Neurophysiol 2011;122:2139–50. Sakkalis V, Oikonomou T, Pachou E, Tollis I, Micheloyannis S, Zervakis M. Timesignificant wavelet coherence for the evaluation of Schizophrenic brain activity using a graph theory approach. Conf Proc IEEE Eng Med Biol Soc 2006;1:4265–8. Scherg M. Fundamentals of dipole source potential analysis. In: Grandori F, Hoke M, Romani GL, editors. Auditory evoked magnetic fields and electric potentials 1990;vol. 6. Basel: Karger; 1990. p. 40–69. Scherg M, Ille N, Bornfleth H, Berg P. Advanced tools for digital EEG review: virtual source montages, whole-head mapping, correlation, and phase analysis. J Clin Neurophysiol 2002;19:91–112. Schulte P. What is an adequate trial with Clozapine?: therapeutic drug monitoring and time to response in treatment-refractory schizophrenia. Clin Pharmacokinet 2003;42:607–18. Struyf J, Dobrin S, Page D. Combining gene expression, demographic and clinical data in modeling disease: a case study of bipolar disorder and schizophrenia. BMC Genomics 2008;9:531. Sumiyoshi T, Higuchi Y, Itoh T, Matsui M, Arai H, Suzuki M, et al. Effect of perospirone on P300 electrophysiological activity and social cognition in schizophrenia: a three-dimensional analysis with sloreta. Psychiatry Res 2009;172:180–3. Takahashi H, Rissling AJ, Pascual-Marqui R, Kirihara K, Pela M, Sprock J, et al. Neural substrates of normal and impaired preattentive sensory discrimination in large cohorts of nonpsychiatric subjects and schizophrenia patients as indexed by MMN and P3a change detection responses. Neuroimage 2013;66:594–603. Theodoridis S, Koutroumbas K. Pattern recognition. 4th ed. Academic Press; 2008. Welch BL. The generalization of ‘‘Student’s’’ problem when several different population variances are involved. Biometrika 1947;34:28–35. Welch PD. The use of fast Fourier transform for the estimation of power spectra: a method based on time averaging over short, modified periodograms. IEEE Trans Audio Electroacoust 1967;15:70–3. Whitfield-Gabrieli S, Thermenos HW, Milanovic S, Tsuang MT, Faraone SV, McCarley RW, et al. Hyperactivity and hyperconnectivity of the default network in schizophrenia and in first-degree relatives of persons with schizophrenia. Proc Natl Acad Sci USA 2009;106:1279–84. Winterer G, Coppola R, Egan MF, Goldberg TE, Weinberger DR. Functional and effective frontotemporal connectivity and genetic risk for schizophrenia. Biol Psychiatry 2003;54:1181–92.


Efficient design of meganucleases using a machine learning approach.

Duration perception of visual and auditory oddball stimuli: does judgment task modulate the temporal oddball effect?

Classifying kinase conformations using a machine learning approach.

A machine learning approach to improve contactless heart rate monitoring using a webcam.

A Machine Learning Approach to Evaluating Illness-Induced Religious Struggle.

A machine learning approach to computer-aided molecular design.

Identifying novel oncogenes: a machine learning approach.

Learning to Classify Organic and Conventional Wheat - A Machine Learning Driven Approach Using the MeltDB 2.0 Metabolomics Analysis Platform.

A machine learning approach to estimate Minimum Toe Clearance using Inertial Measurement Units.

Beyond where to how: a machine learning approach for sensing mobility contexts using smartphone sensors.

The other-race effect in face learning: Using naturalistic images to investigate face ethnicity effects in a learning paradigm.

Effect of Subliminal Lexical Priming on the Subjective Perception of Images: A Machine Learning Approach.

On the effect of subliminal priming on subjective perception of images: a machine learning approach.

Using an integrative approach to investigate the evolution of behaviour.

An approach to investigate intracellular protein network responses.

Bioinformatics algorithm based on a parallel implementation of a machine learning approach using transducers.

Using Blood Indexes to Predict Overweight Statuses: An Extreme Learning Machine-Based Approach.

Improved spatial accuracy of functional maps in the rat olfactory bulb using supervised machine learning approach.

Machine learning approach for the prediction of protein secondary structure.

A machine learning approach for viral genome classification.

Detecting false positive sequence homology: a machine learning approach.

A machine learning approach for predicting methionine oxidation sites.

Analysis of microRNA expression using machine learning.

Machine learning approaches to investigate the impact of PCBs on the transcriptome of the common bottlenose dolphin (Tursiops truncatus).