Introduction

Developmental Dyslexia (DD) is a common neurodevelopmental disorder characterized by reading deficit in spite of adequate cognitive and sensorial abilities¹. According to the magnocellular (M) deficit theory, deficits in the visual M functions predict reading difficulties and are often associated with DD². Here, we applied a Multi-Variate Pattern Analysis (MVPA) approach to investigate whether functional activations associated with the M functions discriminate between Normal Readers (NR) and children with DD.

Materials and methods

Dataset: 22 NR (age 13.2±1.7 yo) and 22 children with DD (age 14.1±1.5 yo) were enrolled in the study (full demographics and cognitive assessment are reported in table 1).
MRI protocol: each subject underwent a 3T MRI scan session including a T1-weighted 3D morphological sequence (TE=8.2ms, TE=3.8ms, flip angle=8°, voxel size 1x1x1 mm³) and two fMRI task sequences (TR=2s, TE=35ms, voxel size 3x3x3.5 mm³).
The first fMRI task (Task 1) consisted of a block-design stimulation of the M and Parvocellular (P) pathways³. M and P stimuli were both full-field sinusoidal grating stimulations with sinusoidal counter-phase flicker. Figure 1 reports an example of the M and P stimuli. The baseline condition was a gray screen of mean luminance with a white fixation point. The task protocol consisted of 8 M blocks, 8 P blocks and 12 baseline blocks presented in pseudo-random order. Subjects were asked to perform a target detection task during the M and P blocks to encourage attention during the acquisition.
The second fMRI task consisted of an event related design task of sensitivity to Coherent Dots Motion (CDM)⁴. An example of the CDM stimulus is reported in figure 2. We used three different ratios of coherently moving dots (CMR), i.e. 6%, 15% and 40% ⁴. After each CDM stimulus, subjects had 4s to report dots motion direction. Finally, a 4.25s gray screen block (baseline condition) was placed between the answer block and the subsequent CDM stimulus. The protocol included 48 stimuli (8 repetitions for each combination of coherence level and motion direction) administered in pseudo-random order. Subjects were instructed to maintain fixation throughout the run, and were actively engaged in performing a motion detection task.
Data preprocessing: morphological images were processed using the recon-all pipeline of the FreeSurfer software package (v6.0) to extract the white matter and the pial surfaces⁵. Subject’s fMRI data were preprocessed following the FreeSurfer Functional Analysis Stream (FS-FAST v6.0) which includes motion and slice-timing corrections, functional-anatomical registration, mask creation, intensity normalization, sampling to common space, spatial smoothing and first level analysis. Five contrast maps were computed for each subject, two from the grating task (M Vs baseline and P Vs baseline) and three from the CDM task (CMR-6% Vs baseline, CMR-15% Vs baseline, CMR-40% Vs baseline), and sampled over the HCP-MMP 1.0 atlas⁶, which included 180 cortical Region Of Interest (ROI) for each hemisphere.
Machine Learning: we performed a classification experiment predicting the subject class (i.e. NR Vs DD) combining all fMRI contrast maps (i.e. Task 1-M, Task 1-P, CMR-6%, CMR-15% and CMR-40%). In particular, we used the Group Lasso implementation of the Multiple Kernel - Support Vector Machine (MK-SVM) algorithm⁷. We assigned a different linear kernel to each atlas ROI, each characterized by a feature vector of five elements (i.e. the average value of each contrast map). The MK-SVM algorithm computes a weight for each kernel, thus providing an estimate of the importance of each ROI for the classification task. A leave-one cross-validation procedure was applied to test the classifier and a nested 10-fold cross validation procedure was used for the hyper-parameter optimization. The classifier performances were assessed computing the classification accuracy and the Area Under the Receiver Operating Characteristic (ROC) curve. A 10'000 label permutation experiment was performed to investigate whether the classifier performance was significantly different from the random choice.

Results

Figure 3 showed the population activation maps for the grating and for the CDM tasks. Both tasks involved the visual processing areas, with the CDM recruiting also frontal areas usually associated with attentive functions.
The MK-SVM algorithm reached a classification accuracy of 65.9% and an AUC of 0.647, significantly over the random choice (p=0.043). Figure 4 shows the kernel weights associated with each ROI. Few ROIs (15/360) account for more than 50% of the overall classification weights and are located mainly in the visual areas, in the dorsal stream, including the neighboring parietal areas, and in the inferior frontal area.

Discussion and conclusion

By combining the activation patterns in two well-established visual tasks, we discriminated between NR and children with DD. Our findings suggest that the M functions may discriminate NR and subjects with DD and support the M deficit theory of DD. Indeed, while the phonological deficit hypothesis remains the most widely accepted⁸, recent studies focused on the amodal temporal deficits for both auditory and visual rapid stimuli (cf. the M deficit theory) as causal factors of the temporal deficits in DD (sluggish attentional shifting⁹).
The MVPA approach successfully combined multi-domain data and provided a tool to effectively analyze a complex spatial disease more likely involving complex processing networks rather than single punctual regions of the brain.

Acknowledgements

This work was supported by the Italian Ministry of Health ("Ricerca corrente" funds)