Dynamic Time Warping in the study of ERPs in dyslexic children

Silvia Casarotto
Summary of the Awarded Paper

Introduction. Recording ERPs (Event-Related Potentials) is important for studying cerebral functions in vivo with high temporal resolution and in a non-invasive way. The study of ERPs is based on the quantification of their morphological characteristics (in particular latency and amplitude of the most relevant components). In order to investigate the neurophysiological bases of reading processes, ERPs were recorded in children with, and without developmental dyslexia.

Developmental dyslexia is a neuropsychological disorder primarily characterised by reading difficulties despite average intelligence, adequate education and normal sensory acuity. This disorder is sometimes reversible when specific rehabilitative therapies are led on young children. Therefore, it is important to investigate the reading processes, to diagnose dyslexia as early as possible and to work out efficacious and targeted therapies. Electrical brain responses to appropriate stimulation paradigms can help to attain these aims.

ERPs recorded in children during the execution of cognitive tasks are characterised by a great inter-individual morphological variability. Due to this variability, the quantification of ERPs morphology has to be performed manually and therefore it is time-consuming and greatly influenced by subjective judgement of the measuring technician. Furthermore, the identification of normal "patterns" of activation is difficult; the usual approach is to compute grand averages that should represent the mean morphological characteristics of the ERPs recorded in homogeneous groups of subjects. Inter-individual variability greatly affects grand averages reliability because it leads to a smoothing of the cerebral components.

Aim of the work. In this work a non-linear alignment method based on Dynamic Time Warping (DTW) technique is proposed to compute reliable templates of ERPs recorded from homogeneous groups of subjects and to automate the quantification of ERPs morphology. The method is applied to ERPs recorded from normal and dyslexic children and the differences between these two groups of subjects are analysed in terms of ERP morphology.

Materials and Methods. DTW is a non-linear alignment algorithm that reduces the temporal differences between morphologically similar signals through local compressions and extensions of their temporal axes. Given two equally sampled signals, made of I and J samples respectively, it is possible to represent a correspondence between their samples in a Cartesian plot by means of a Warping Function (WF). The WF has some specific properties and has to respect some constraints that regulate flexibility and sensitivity of the alignment process. The signals have to be aligned according to a criterion of morphological similarity. Therefore, in order to compute the Warping Function, it is necessary to define a measure of the morphological distance between the samples of the signals. This measure is called dissimilarity function. In this work the dissimilarity function is computed as the sum of the absolute differences between the amplitude and slope of the samples of the signals. While satisfying the constraints imposed to the WF, it is possible to compute the dissimilarity function at every point of the warping plane and to organize these values into a matrix of size (I,J). Applying the Dynamic Programming (DP) equations this matrix is transformed in another matrix, whose elements represent a cost function connected to alignment. According to a criterion of morphological similarity, the best alignment between the two signals is obtained by searching in this cost matrix the path with the minimum cost reaching the point (I,J) from the point (1,1).

The template is obtained by averaging the two signals on the basis of the morphological rather than temporal correspondence between their samples, as represented by the WF. The computation of the template on more than two signals is realised by iteratively applying the method to paired signals following a binary tree structure. After the computation of a template on a set of signals, the alignment of the template with each of the original signals produces a morphological correspondence between the samples of the template and the signals. This approach realises the automatic quantification of ERP morphology.

We recorded ERPs from sixteen 8-year-old children, half of whom did not suffer from dyslexia and half with dyslexia. Stimuli consisted in the visual presentation of Italian alphabetic letters. Each child performed two tasks. The Letter Presentation task consisted in the passive observation of letters. The Letter Recognition task consisted in reading aloud the letters. We recorded 10 EEG channels located according to the 10-20 system and some other muscular and ocular electrical activities. We applied the alignment method to the first 700 ms recordings after stimulus presentation.

Results. Before describing results in detail, it is suitable to introduce some basic considerations on reading-related potentials. Looking at ERPs morphology, it is possible to identify a set of components. Their exact functional meaning is not precisely known: however, it is possible to subdivide them into three distinct categories. The short-latency components (N0, P0, N1, P1) belong to the pre-lexical period: they mainly result from the sensorial processing of stimuli and from attention mechanisms; they generally spread over the scalp in an occipital-frontal direction. Among these components, P1 is particularly evident in Oz, because it is related to the primary visual cortex activation in correspondence to the presentation of the stimulus. The middle-latency components (N2, PmaxA, PmaxB, N3) belong to the lexical period: they are mainly concerned with the stimulus categorization and some control mechanisms; they generally spread over the scalp in a frontal-occipital direction. The long-latency components (P4, P600a, N4) belong to the post-lexical period: they presumably reflect long-term memory faculties and feedback processes.

We applied the non-linear alignment method in order to obtain ERP templates from the control group and the group with dyslexia in the two different conditions. The relevant ERP components were automatically identified. A two-sided t-test analysis was performed to compare the latencies of the relevant ERP components in the two groups of children for each task separately.

Considering the Letter Presentation task, we noticed that the latency of ERP components is always greater in dyslexics compared to controls. Significant differences between controls and dyslexics are present in many ERP components, in particular N1, N2, N3, N4, P600a; these differences are particularly located in the right precentral regions and in the parietal and temporal regions.

Considering the Letter Recognition task, we noticed that the latency of ERP components is always greater in dyslexics compared to controls. Significant differences between controls and dyslexics are present in N1, P1, N2, P4 components; these differences are particularly located in the temporal regions.

Discussion. We found that templates are more reliable than grand averages because the non-linear alignment reduces inter-individual variability.

Reading is a complex task that involves visual processing, attention mechanisms and memory faculties. The left hemisphere is usually associated with language, but some studies highlighted the participation of the right hemisphere in processing of phonological information. Our results agree with these findings because we noticed significant differences between controls and dyslexics in the right pre-central region. The finding of significant differences between controls and dyslexics in the pre-lexical components suggests that the low-level processing of visual information and the attention mechanisms are disrupted in dyslexia. The finding of significant differences between controls and dyslexics in the post-lexical components suggests that dyslexic subjects have difficulties in storing and recalling to memory the correct correspondence between graphemes and phonemes. This observation recalls the phonological theory about the origins of dyslexia. Some fMRI studies highlighted that parietal regions are greatly involved in the processing of phonological information. Our results agree with these studies because, even if we used a partially different stimulation protocol and a different biological signal, we found significant differences between controls and dyslexics in the EEG channels corresponding to the parietal lobe.

In conclusion, our results suggest that dyslexia is a pathology in which not only the reading processes, but also more general functions, at different time scales, are compromised