Classification of ADHD patients on the basis of independent ERP components using a machine learning system

Background

In the context of sensory and cognitive-processing deficits in ADHD patients, there is considerable evidence of altered event related potentials (ERP). Most of the studies, however, were done on ADHD children. Using the independent component analysis (ICA) method, ERPs can be decomposed into functionally different components. Using the classification method of support vector machine, this study investigated whether features of independent ERP components can be used for discrimination of ADHD adults from healthy subjects.

Methods

Two groups of age- and sex-matched adults (74 ADHD, 74 controls) performed a visual two stimulus GO/NOGO task. ERP responses were decomposed into independent components by means of ICA. A feature selection algorithm defined a set of independent component features which was entered into a support vector machine.

Results

The feature set consisted of five latency measures in specific time windows, which were collected from four different independent components. The independent components involved were a novelty component, a sensory related and two executive function related components. Using a 10-fold cross-validation approach, classification accuracy was 92%.

Conclusions

This study was a first attempt to classify ADHD adults by means of support vector machine which indicates that classification by means of non-linear methods is feasible in the context of clinical groups. Further, independent ERP components have been shown to provide features that can be used for characterizing clinical populations.

     

Discriminating between ADHD adults and controls using independent ERP components and a support vector machine: a validation study

Background

There are numerous event-related potential (ERP) studies in relation to attention-deficit hyperactivity disorder (ADHD), and a substantial number of ERP correlates of the disorder have been identified. However, most of the studies are limited to group differences in children. Independent component analysis (ICA) separates a set of mixed event-related potentials into a corresponding set of statistically independent source signals, which are likely to represent different functional processes. Using a support vector machine (SVM), a classification method originating from machine learning, this study aimed at investigating the use of such independent ERP components in differentiating adult ADHD patients from non-clinical controls by selecting a most informative feature set. A second aim was to validate the predictive power of the SVM classifier by means of an independent ADHD sample recruited at a different laboratory.

Methods

Two groups of age-matched adults (75 ADHD, 75 controls) performed a visual two stimulus go/no-go task. ERP responses were decomposed into independent components, and a selected set of independent ERP component features was used for SVM classification.

Results

Using a 10-fold cross-validation approach, classification accuracy was 91%. Predictive power of the SVM classifier was verified on the basis of the independent ADHD sample (17 ADHD patients), resulting in a classification accuracy of 94%. The latency and amplitude measures which in combination differentiated best between ADHD patients and non-clinical subjects primarily originated from independent components associated with inhibitory and other executive operations.

Conclusions

This study shows that ERPs can substantially contribute to the diagnosis of ADHD when combined with up-to-date methods.     

Single-event-related potential analysis by means of fragmentary decomposition

 A recently developed fragmentary decomposition method is employed to analyse single-trial event-related potentials (ERPs), thereby extending the traditional method of averaging. Using a conventional auditory oddball paradigm with 40 target stimuli, single-trial ERPs in 40 normal subjects were analysed for midline scalp (Fz, Cz and Pz) recording sites. The normalization effect, reported in our previous study of eye blink EMGs and proposed to be a characteristic property of a wide class of non-stationary physiological processes, was found to apply to these single-trial ERPs. Fragmentary decomposition of single-trial ERPs may be regarded as re-statement of the normalization effect. This allows both pre-stimulus EEGs and post-stimulus ERPs to be regarded as overlapping generic mass potentials (GMPs), with a characteristic Gaussian amplitude spectrum. On theoretical and empirical grounds we uniquely deduce a model GMP using an introduced “bud” function, and physically support it by the resting and transient conditions. The model takes into account the shape of the component, which suggests a simple relationship between the peak latency and the time of the component onset. Given that GMPs may be manipulated and sorted out, we present principles of the fragmentary synthesis, i.e. probabilistic ERP reconstructions on the basis of individual and ensemble properties of its identified components. Summarizing the component quantification in the form of the dynamic model provides for the first time the opportunity to quantify all significant components in single-trial ERPs. This method of single-trial analysis opens up new possibilities of exploring the dynamical ERP changes within a recording trial, particularly in late component “cognitive” paradigms. Received: 29 August 2000 / Accepted in revised form: 5 February 2001     

Functional Organization of the Brain during the Operation of Working Memory

Event-related potentials (ERPs) recorded from various cortical areas during matching of two consecutive pictures were analyzed. Reflecting the process of trace fixation, the ERP to the reference stimulus was characterized by an increase in components P150 and P300 in the occipital and temporo-parieto-occipital areas and components N300 and N400 in the precentral areas as compared with the ERP elicited by the warning stimulus. The ERP to the test stimulus, which reflected trace retrieval and matching with current information, was characterized by a generalized increase in the late positive complex in the interval 300–600 ms. Similarity and/or dissimilarity of the test and reference stimuli was reflected in the parameters of the ERP to the test stimulus. The results testify to the difference in functional and topographic organization of the brain cortex at the initial and late stages of operation of the working memory.     

A new method to detect event-related potentials based on Pearson’s correlation

Event-related potentials (ERPs) are widely used in brain-computer interface applications and in neuroscience.  Normal EEG activity is rich in background noise, and therefore, in order to detect ERPs, it is usually necessary to take the average from multiple trials to reduce the effects of this noise.  The noise produced by EEG activity itself is not correlated with the ERP waveform and so, by calculating the average, the noise is decreased by a factor inversely proportional to the square root of N, where N is the number of averaged epochs. This is the easiest strategy currently used to detect ERPs, which is based on calculating the average of all ERP’s waveform, these waveforms being time- and phase-locked.  In this paper, a new method called GW6 is proposed, which calculates the ERP using a mathematical method based only on Pearson’s correlation. The result is a graph with the same time resolution as the classical ERP and which shows only positive peaks representing the increase—in consonance with the stimuli—in EEG signal correlation over all channels.  This new method is also useful for selectively identifying and highlighting some hidden components of the ERP response that are not phase-locked, and that are usually hidden in the standard and simple method based on the averaging of all the epochs.  These hidden components seem to be caused by variations (between each successive stimulus) of the ERP’s inherent phase latency period (jitter), although the same stimulus across all EEG channels produces a reasonably constant phase. For this reason, this new method could be very helpful to investigate these hidden components of the ERP response and to develop applications for scientific and medical purposes. Moreover, this new method is more resistant to EEG artifacts than the standard calculations of the average and could be very useful in research and neurology.  The method we are proposing can be directly used in the form of a process written in the well-known Matlab programming language and can be easily and quickly written in any other software language.     

Effects of the menstrual cycle on auditory event-related potentials

Gonadal steroids (estradiol and progesterone) can alter neuronal functioning, but electrophysiological evidence in women is still sparse. Therefore, the present study investigated event-related potentials (ERPs) to neutral stimuli over the course of the menstrual cycle. In addition, associations between ERPs and salivary estradiol and progesterone concentrations were investigated. Eighteen young healthy women were tested at three different phases of their menstrual cycle (menses, and follicular and luteal phases). ERPs (i.e., the N1 and P2 components, reflecting cortical arousal and the orienting response, the N2, P3, and the Slow Wave (SW), reflecting controlled processing) were measured using two different paradigms. In the luteal phase, early ERPs reflecting the cortical arousal response were diminished in the first stimulus block indicating an attenuated orienting response. These changes were significantly correlated with estradiol as well as progesterone levels. As to the later ERP components, the N2 latency was shorter during menses compared to the other two phases. No menstrual cycle-associated changes were apparent in other late ERP components. In sum, this study documents changes in auditory ERPs across the menstrual cycle with the most prominent changes occurring during the luteal phase. Future ERP studies therefore need to be more attentive to the issue of menstrual phase when studying female subjects or female patients.     

Late divergence of target and nontarget ERPs in a visual oddball task

Different mental operations were expected in the late phase of intracerebral ERPs obtained in the visual oddball task with mental counting. Therefore we searched for late divergences of target and nontarget ERPs followed by components exceeding the temporal window of the P300 wave. Electrical activity from 152 brain regions of 14 epileptic patients was recorded by means of depth electrodes. Average target and nontarget records from 1800 ms long EEG periods free of epileptic activity were compared. Late divergence preceded by almost identical course of the target and nontarget ERPs was found in 16 brain regions of 6 patients. The mean latency of the divergence point was 570+/-93 ms after the stimulus onset. The target post-divergence section of the ERP differed from the nontarget one by opposite polarity, different latency of the components, or even different number of the components. Generators of post-divergence ERP components were found in the parahippocampal gyrus, superior, middle and inferior temporal gyri, amygdala, and fronto-orbital cortex. Finding of late divergence indicates that functional differences exist even not sooner than during the final phase of the task.     

Molecular cloning and functional properties of two early-stage encapsulation-relating proteins from the coleopteran insect, Tenebrio molitor larvae.

Encapsulation is a major defensive reaction against foreign materials that are too large to be phagocytosed by individual hemocytes; however, the biochemical process of encapsulation is still obscure. To isolate and characterize the early-stage encapsulation-relating protein (ERP), we used the coleopteran insect, Tenebrio molitor larvae, injecting three differing kinds of bead or inserting pieces of surgical suture into the abdomen of T. molitor larvae. The resulting proteins from the injected beads or the inserted pieces of surgical suture were recovered 10 min after injection or insertion, and were analyzed on SDS/PAGE under reducing conditions. Four different proteins (86, 78, 56 and 48 kDa) were enriched compared with the crude hemolymph. Among them, we purified 56-kDa and 48-kDa ERPs to homogeneity and raised polyclonal antibodies against each protein. Immunoblotting analysis showed that the affinity-purified antibodies of the 56-kDa and 48-kDa ERPs cross-reacted with the 48-kDa and 56-kDa ERPs, respectively. Analysis of the cDNA of 56-kDa ERP consisted of 579 amino acid residues and showed a novel glutamine-rich protein. Positive clones of the 48-kDa ERP showed the same DNA sequence as 56-kDa ERP. Interestingly, the chemically determined N-terminal amino acid sequence and the three partial amino acid sequences of the 48-kDa protein were found in the 56-kDa ERP, suggesting that the 48 kDa ERP was produced by the cleavage of Arg101-Gly102 of the 56-kDa ERP by a limited proteolysis. Western blotting analysis showed that these ERPs were detected exclusively on membrane fractions of hemocytes. Also, when the early-stage encapsulated beads were coated with both the 56-kDa and 48-kDa ERP antibodies and re-injected into larvae, no further encapsulation reaction was observed. However, when the early-stage encapsulated beads were incubated with 56-kDa ERP antibody, 48-kDa ERP antibody or nonimmunized rabbit IgG and re-injected into larvae, further encapsulation did occur.     

ERPs to subclasses of nouns and verbs

The present findings show the existence of significant differences in latency and amplitude of some waves of ERPs to three subclasses of nouns or verbs. Latency LP3 of ERPs to action verbs was shorter than latency of the same wave to abstract verbs. In nouns the relation was a similar (manipulable objects versus abstract ones). The amplitude of early positive components (P1, P2) and BN3 wave also depended on semantic attributes of nouns and verbs. Some waves of ERPs to motion verbs in our experiment had significantly higher amplitude than the same waves of ERPs to nonmanipulable objects. Also revealed was interaction between some ERP features and the subject's gender. It was primarily the amplitude of BP2, the size of which depended on gender in all cases--BP2 amplitude was significantly higher in females then in males. In the other components (BP1, BP4, BN2 and BN4) there were fewer significant differences and if they do occurred, then their amplitude was higher in males than in females. In some cases, gender also affected latency of waves LP2, LN1 and LN4 of ERPs to the same noun or verb subclass--latencies of these waves were shorter in females.     

Functionally independent components of early event-related potentials in a visual spatial attention task.

Spatial visual attention modulates the first negative-going deflection in the human averaged event-related potential (ERP) in response to visual target and non-target stimuli (the N1 complex). Here we demonstrate a decomposition of N1 into functionally independent subcomponents with functionally distinct relations to task and stimulus conditions. ERPs were collected from 20 subjects in response to visual target and non-target stimuli presented at five attended and non-attended screen locations. Independent component analysis, a new method for blind source separation, was trained simultaneously on 500 ms grand average responses from all 25 stimulus-attention conditions and decomposed the non-target N1 complexes into five spatially fixed, temporally independent and physiologically plausible components. Activity of an early, laterally symmetrical component pair (N1aR and N1aL) was evoked by the left and right visual field stimuli, respectively. Component N1aR peaked ca. 9 ms earlier than N1aL. Central stimuli evoked both components with the same peak latency difference, producing a bilateral scalp distribution. The amplitudes of these components were no reliably augmented by spatial attention. Stimuli in the right visual field evoked activity in a spatio-temporally overlapping bilateral component (N1b) that peaked at ca. 180 ms and was strongly enhanced by attention. Stimuli presented at unattended locations evoked a fourth component (P2a) peaking near 240 ms. A fifth component (P3f) was evoked only by targets presented in either visual field. The distinct response patterns of these components across the array of stimulus and attention conditions suggest that they reflect activity in functionally independent brain systems involved in processing attended and unattended visuospatial events.     

Stimulus dependency of object-evoked responses in human visual cortex: an inverse problem for category specificity

Many studies have linked the processing of different object categories to specific event-related potentials (ERPs) such as the face-specific N170. Despite reports showing that object-related ERPs are influenced by visual stimulus features, there is consensus that these components primarily reflect categorical aspects of the stimuli. Here, we re-investigated this idea by systematically measuring the effects of visual feature manipulations on ERP responses elicited by both structure-from-motion (SFM)-defined and luminance-defined object stimuli. SFM objects elicited a novel component at 200-250 ms (N250) over parietal and posterior temporal sites. We found, however, that the N250 amplitude was unaffected by restructuring SFM stimuli into meaningless objects based on identical visual cues. This suggests that this N250 peak was not uniquely linked to categorical aspects of the objects, but is strongly determined by visual stimulus features. We provide strong support for this hypothesis by parametrically manipulating the depth range of both SFM- and luminance-defined object stimuli and showing that the N250 evoked by SFM stimuli as well as the well-known N170 to static faces were sensitive to this manipulation. Importantly, this effect could not be attributed to compromised object categorization in low depth stimuli, confirming a strong impact of visual stimulus features on object-related ERP signals. As ERP components linked with visual categorical object perception are likely determined by multiple stimulus features, this creates an interesting inverse problem when deriving specific perceptual processes from variations in ERP components.     

视觉空间注意事件相关电位的协同学分析

在用事件相关电位(event-related potentials,ERP)研究视觉空间注意问题时,直接观察ERP数据就可得出,空间注意的主要作用是对视觉信息处理的调制,它出现在刺激开始后大约80～250ms,主要表现为枕叶的P1、N1和P2波有明显的增强但它们的潜伏期没有变化。采用基于协同学的时空模式分解方法,把视觉空间注意ERP分解为3个模式成分。结果表明,注意不仅使模式1的第一个正波成分(P11)、第一个负波成分(N11)以及第二个正波成分(P12)增强,还使模式3的第一个正波成分(P31)的潜伏期缩短。用探照灯模型对这些现象作了初步解释,说明该方法是研究注意ERP的一种有潜力的新方法。     

Processing of visually presented verbal stimuli in evoked potentials

The ERPs to familiar, unfamiliar, and recently learned words were recorded in 20 right-handed students who have recently studied a course in sensory physiology during performance of two tasks. The first task consisted in processing of different types of words, the second task required recognition of terms from sensory physiology. In the first task, three types of words were presented to a subject: familiar words related and unrelated to a given category and unfamiliar words. Subjects had to press the appropriate button responding to the questions whether they considered the presented word be related to the given category and whether they were sure in their answer. In the second task, terms from recently studied physiological material were presented to the students who had to detect whether those terms were related or unrelated to the given sensory system. Significant differences between late ERP components were revealed by Student test for matched samples. In the first task, component N400 of the ERP to unfamiliar words was higher than that of the ERP to familiar words. It was also higher in ERPs to words unrelated to the given category than in ERPs to related words. However, this component did not differ between ERPs to well and poorly acquired words in the test on sensory physiology. On the contrary, the late positive component was more expressed in ERPs to familiar and better acquired words in both tasks. Brain mechanisms of processing of different types of words are discussed.     

Effects of the ADHD Syndrome on the Frequency Composition of ERPs Revealed by Independent Component Analysis

In this study, we investigated the frequency characteristics of independent components (ICs) of event-related potentials (ERPs) for persons suffering from attention deficit/hyperactivity disorder (ADHD) and normal adults while performing a continuous performance test (CPT). A group of 50 participants (10 ADHD subjects and 40 ones with no attention disorders) was examined. Independent component analysis was applied to the recorded signals. For ERP extraction, averages for each group of ICs, which were time-locked to the onset of stimuli, were calculated. Several frequency characteristics (704 items) were extracted from different ERPs in each IC. Eight features of the brain signals had a significant (P < 0.001) correlation with the participants’ clinical presentation, which is consistent with the results of previous studies. The revealed promising relation can be used for further evaluation of the sustained attention level.     

Individual Differences in the Recognition of Facial Expressions: An Event-Related Potentials Study

Previous studies have shown that early posterior components of event-related potentials (ERPs) are modulated by facial expressions. The goal of the current study was to investigate individual differences in the recognition of facial expressions by examining the relationship between ERP components and the discrimination of facial expressions. Pictures of 3 facial expressions (angry, happy, and neutral) were presented to 36 young adults during ERP recording. Participants were asked to respond with a button press as soon as they recognized the expression depicted. A multiple regression analysis, where ERP components were set as predictor variables, assessed hits and reaction times in response to the facial expressions as dependent variables. The N170 amplitudes significantly predicted for accuracy of angry and happy expressions, and the N170 latencies were predictive for accuracy of neutral expressions. The P2 amplitudes significantly predicted reaction time. The P2 latencies significantly predicted reaction times only for neutral faces. These results suggest that individual differences in the recognition of facial expressions emerge from early components in visual processing.     

ACTH 4-9 effects on the human visual event-related potential

Three oral doses (5, 10 and 20 mg) of an analog of ACTH 4-9 were compared with a vehicle control and d-amphetamine (10 mg). In a double-blind procedure, five men and five women were tested at weekly intervals with each treatment. In each session, four visual event-related potentials (ERPs) were recorded at hourly intervals. Visual ERPs were averaged from the electroencephalogram recorded from the left and right hemisphere. Dosage, time after administration, hemisphere of the brain and sex of the subject influenced the ERP. The ACTH 4-9 analog decreased amplitude of P100 but increased integrated activity of the ERP. This effect peaked at 60 minutes then "recovered." The effects of the peptide were more pronounced with doses of 5 and 10 mg, in the right hemisphere of men and in the left hemisphere of women. The findings indicated that the ACTH 4-9 analog influenced components of ERP commonly related to the perceptual/attentional state of the organism in a sexually dimorphic manner.     

P3 latency jitter assessed using 2 techniques. I. Simulated data and surface recordings in normal subjects

Latency variability measurement using cross-correlational techniques has the drawback of alignment to background noise not related to ERP activity. We compared latency jitter estimation in simulated and real P3 recordings using Woody's algorithm and a non-cross-correlational technique, the maximum likelihood technique (MLT). Simulated ERPs (with introduced latency jitter) were generated using either a 1/2 cycle 2 Hz sine wave or an averaged P3 ERP with 1 of 3 added noise types in 5 signal to noise ratios (SNRs): (i) white noise; (ii) a 10 Hz sine wave; (iii) a 7.5 Hz sine wave. Jitter measurement accuracy was assessed using mean square error (MSE) for 1 iteration of the Woody method and each of 4 iterations of the MLT. Lowest MSEs occurred for higher SNRs and 1 iteration of the MLT. The MLT and Woody method were applied to P3 ERPs of 13 subjects with SNRs greater than 0.4. P3 latency jitter was significantly lower for the MLT. Latency jitter (both methods) did not differ between homologous electrodes and was highest in posterior electrodes. In the latency corrected ERP data of subjects with persistent alpha activity periodic components occurred in the Woody corrected average (not seen in the conventional or the MLT corrected averages). Our data indicate that the MLT is the more accurate method for determining latency jitter.     

Effects of early high-dose levothyroxine treatment on auditory brain event-related potentials at school entry in children with congenital hypothyroidism

AIMS: We tested whether brain event-related potentials (ERPs) are normal in children with congenital hypothyroidism (CH) after early high-dose levothyroxine treatment. METHODS: Auditory ERPs were recorded in 33 normal controls and in 15 children with CH at 5 years 9/12. Based on bone maturation at diagnosis, the CH group was divided into severe (n = 8) and moderate (n = 7) subgroups. CH patients were treated at a median age of 14 days with a mean initial dose of levothyroxine of 11.6 microg/kgxday. Two ERP components (N100 and N200) were measured and clinical follow-up variables collected. RESULTS: The functional anatomical and cognitive organisation of the auditory system, as revealed by the analyses of ERP measures, did not differ between CH and controls, or between severe and moderate CH subjects. However, N200 latency was globally longer in the CH than in the control group (p = 0.01) and was positively correlated with the over-treatment index (r = 0.61; p < 0.05) and verbal IQ. N200 amplitude was negatively correlated with initial dose (r = -0.74; p < 0.005). CONCLUSION: These data suggest that sensitive tools such as ERPs can reveal differences between CH and controls and relate these differences to the adequacy of treatment of CH.     

Functional biomarkers in the diagnostics of mental disorders: Cognitive event-related potentials

The article deals with recent research in the field of fundamental knowledge about the mechanisms of information processing in the human brain for the diagnosis of mental disorders performed in the laboratory of neurobiology for action programming of the Bechtereva Institute of the Human Brain of the Russian Academy of Sciences. These studies were connected with the analysis of functional components for cognitive event-related potentials (ERPs) obtained under various behavioral conditions. The main goal of this fundamental approach is the decomposition of multi-channel ERPs into functionally different components. These components are generated in various cortical areas, have different temporal dynamics, and reflect a variety of mental operations. The main methodology we used is the independent component analysis, applied to a large set of ERPs (from hundreds of people) obtained by varying functional conditions in the psychological test. In particular, components related to psychological processes, such as the comparison of sensory signals with a trace in working memory, inhibition of current activity, and monitoring of conflict, were identified in the GO/NOGO test. In the European project, a normative database was constructed for the components described above, and this allowed us to compare the data obtained from large groups of patients (including patients with attention deficit disorder, schizophrenia, obsessive disorders, depression, autism, dyslexia, brain trauma, and dementia) with the healthy subjects. This article presents data from patients with a diagnosis of attention deficit disorder and schizophrenia.     

The possibility of using biofeedback for training the amplitude of independent components of event-related potentials

The possibility of voluntary training of the independent component of event-related potentials (ERPs) by the method of biological feedback (BFB) using a modified two-stimulus test of Go/NoGo paradigm was studied. Ten subjects had to increase the amplitude of the P3b component of ERPs induced by the Go stimulus during one training session. This component is generated in the parietal cortex and is associated with an operation of action engagement. The study consisted of three sessions that included determination of the average amplitude, training, and placebo. It was found that the amplitude of the P3b component did not significantly change during the training and placebo sessions as compared to the session of threshold measurement, whereas the amplitude of the SPD component, which was also observed in ERPs in response to the Go stimulus, significantly decreased in both sessions. According to the sLORETA data, the source of the SPD component is located in the parahippocampal area. The results obtained help to determine further pathways to study the possibility of the use of BFB for correction of components of cognitive evoked potentials.