Prediction of Children's Math Learning Disability Using Longitudinal Brain Data and Machine Learning

Project Information

MathDevBML

Grant agreement ID: 101023033

DOI

10.3030/101023033

Project closed

EC signature date 22 April 2021

Start date 12 May 2022

End date 11 May 2024

Funded under

• EXCELLENT SCIENCE - Marie Skłodowska-Curie Actions

Total cost

€ 184 707,84

EU contribution

€ 184 707,84

Coordinated by

INSTITUT NATIONAL DE LA SANTE ET DE LA RECHERCHE MEDICALE
France

Periodic Reporting for period 1 - MathDevBML (Prediction of Children's Math Learning Disability Using Longitudinal Brain Data and Machine Learning)

Reporting period: 2022-05-12 to 2024-05-11

Summary of the context and overall objectives of the project

Mathematics is the fundamental basis of modern science and technology. However, mathematical skills differ among individuals, and 5-7% of the population suffers from mathematical learning difficulty (MLD). MLD may result in poor professional and personal outcomes, including severe financial difficulties. Early detection of potential MLD in young children is thus essential for providing appropriate support, preventing the worsening of difficulties, and achieving financial well-being. In addition, research on the biological basis associated with mathematical difficulties and the development of diagnostic techniques for these difficulties would be beneficial to society, as recent developments in the field of artificial intelligence have brought more attention to the importance of mathematical skills.
Previous studies used noninvasive brain measurement techniques such as functional magnetic resonance imaging (fMRI) to reveal brain activity differences between MLD and typically developing children. Other studies further applied machine learning techniques to neuroimaging data to predict MLD. However, participating children in most of these studies had already experienced math education in elementary school. Moreover, previous studies used model-free methods such as support vector classification, and model-based machine learning approach, which facilitates interpretation of brain representations based on latent features, has been largely limited.
The project aims to address these gaps based on three objectives. The first objective is to evaluate the neural basis of MLD using longitudinal functional magnetic resonance imaging (fMRI) data collected from preschool children. The second objective is to apply model-based machine learning methods to classify MLD. The third objective is to evaluate the developmental relationship between mathematical difficulty and other cognitive abilities.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

The project yielded the following three main scientific results. First, to clarify the development of number-related brain functions that may influence MLD, the researcher collected fMRI data from 5- and 8-year-old children and examined age-related changes in children’s number representations based on these cross-sectional data. Brain activity was measured while children passively perceived nonsymbolic and symbolic number stimuli. Cross-format decoding analysis revealed a format-independent neural representation of quantity in the right parietal cortex for 5-year-olds, but not for 8-year-olds. This indicates that, after three years of formal education, brain representations of symbolic numbers are likely to become more independent from nonsymbolic quantity. This work has been published in PLOS Biology (Nakai et al., PLOS Biol 2023).
Second, the researcher performed model-based machine learning analysis using voxel-wise encoding models, based on fMRI data from eight participants solving math word and expression problems. Mathematical problems with different formats had similar cortical organization in the intraparietal sulcus, indicating that mathematical problems are represented in the brain in a format-invariant manner. Moreover, based on the same dataset, the researcher further constructed encoding models using artificial neural networks (ANNs) and found shared representations between ANN and brain activity patterns for mathematical problem solving. These results demonstrate that it is possible to perform model-based machine learning analysis using ANN features of mathematics. This work has been published in European Journal of Neuroscience (Nakai and Nishimoto, Eur J Neurosci 2023) and NeuroImage (Nakai and Nishimoto, NeuroImage 2023).
Third, the researcher performed a systematic review of brain-based machine learning applications on mathematical ability and other cognitive functions. The researcher reviewed articles with the cross-sectional and longitudinal designs in the literacy and numeracy domains, and described how they can be coupled with regression and classification approaches. The researcher argued that the field needs a standardization of methods, as well as a greater use of accessible and portable neuroimaging methods that have more applicability potential than lab-based neuroimaging techniques. This work has been published in Imaging Neuroscience (Nakai et al., Imag Neurosci 2024).
Overall, the project produced eight peer-reviewed articles and two preprints currently under-review, as well as participation in four international conferences, three workshops, one press release, and two articles in a general magazine. These results indicate that the project has reached beyond its initial objectives and has led to international and interdisciplinary collaborations, generating further research in related areas.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

The current project provides the following progress beyond the state-of-the-art. First, the result supports the so-called symbolic estrangement hypothesis, which suggests that extended exposure to symbolic mathematics in formal education changes the brain’s representation of symbolic numbers and renders them more independent of non-symbolic numbers. This hypothesis had been claimed primarily based on behavioral data, but the project provided the first neural evidence.
Second, the project was the first attempt to use voxel-wise encoding models to mathematical cognition. This technique is at the heart of the computational model approach to human neuroimaging but has not been applied to higher cognitive functions such as mathematics. The current project successfully opens the possibility of a model- based approach to mathematical cognition.
Third, the systematic review provides an overview of machine learning applications to predict academic achievement. The current work will be a useful resource for many researchers interested in the educational neuroscience field to learn about state-of-the-art methods and findings.
To fully accomplish the project's initial objectives, the researcher and members of the hosting organization will continue recruiting the 8-year-old longitudinal samples for their second session (two and a half years after the first session). After the recruitment process is completed, the researchers will conduct a predictive model analysis of MLD using longitudinal data.
Thus, the researcher will combine model-based machine learning methodology already developed during the project with longitudinal fMRI data. This analysis would enable building a prediction model of MLD with higher precision than model-free methods.　Furthermore, by extracting different cognitive components via neural networks, the researcher will compare prediction performance and the brain organization of different cognitive factors related to mathematical skills.
The project has potential for future industrial applications in terms of artificial intelligence application to neuroscience (neurotechnology). Neurotechnology has recently attracted significant attention, but its application to developmental data and education was largely limited. The current findings, in particular those showing potential prediction of academic achievement from brain data, could attract industry interested in building higher precision neurotechnology systems.

figure.jpg