Abstract:

Developmental dyslexia is the most prevalent form of specific learning disorder with a neurobiological basis that not only restricts an individual's academic achievement and career development, but also negatively impacts an individual's psychological and social adjustment substantially. If children with developmental dyslexia are not timely identified and intervened, this negative impact may persist from early childhood into adulthood. Therefore, efficient early screening and effective early intervention are critical to the development of dyslexic children. Machine learning allows “machines” (i.e., computers) to learn and extract patterns from large amounts of data to make predictions or decisions. Recently, machine learning has been gradually applied to the early screening of dyslexic children due to its powerful data processing and mining capabilities. This review paper aims to outline possible development paths and ideas for machine learning research in dyslexia by integrating the recent advances, main applications, and possible future directions of machine learning in dyslexia screening. We searched for studies using machine learning to identify dyslexia since 2016 and ultimately selected 25 articles based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) protocol. These studies have mainly collected data with standardized psychoeducational tests, eye tracking, game testing, and brain imaging techniques. More researchers have begun to not limit to a single modality of data collection. They integrated survey data, behavioral data, and neuroimaging data in an attempt to improve the accuracy of detecting dyslexia and its biomarkers. In the selection of algorithms, Support Vector Machine (SVM) is the most widely used in the field of dyslexia. Recently, there is an increasing tendency for researchers to find the best algorithm among multiple algorithms to obtain optimal parameters. The main trend is to move from a single traditional machine learning algorithm to deep learning algorithms and to compare many different types of algorithms. The evaluation performances (accuracy) of the machine learning models are summarized as follows: standardized psychoeducational tests between 68% and 94.1%, eye tracking tests between 81.25% and 99%, game tests between 74% and 99.9%, brain imaging data (EEG) between 89% and 90%, and brain imaging data (fMRI) between 65% and 94.83%. In practical applications, machine learning can contribute to identify predictors of dyslexia, allowing us to effectively detect children at risk for dyslexia and intervene timely, thereby reducing the likelihood of reading failure after literacy and even in adulthood. Second, machine learning is being used to assist in clinical screening and automate the identification of dyslexic children, which not only incorporates a large number of objective classifiers to improve accuracy, but also is convenient and reduces waiting costs. Third, identifying children at high risk for dyslexia at an early age will enable early prevention and intervention. This early predictive function can be achieved by training machine learning predictive models. This review summarizes the advantages of using machine learning for early screening of dyslexia. For example, machine learning can identify complex nonlinear relationships between variables, providing more accurate screening and prediction of dyslexia. In addition, machine learning avoids the effects of subjective understanding bias on the one hand, and achieves higher accuracy and reproducibility in less time than human methods of identifying dyslexia on the other. Finally, machine learning has powerful high-dimensional data processing capabilities that can detect abnormalities in tiny brain imaging that may reflect important pathophysiological mechanisms that are not visible to the human eye. However, there are still some limitations in machine learning researches on dyslexia, such as small sample size, low clinical transformation rate, insufficient combination of multi-modal data, and threats to data security and privacy protection. Moreover, there is a lack of studies on the optimal intervention period for groups of children, and early screening of dyslexic children is not really achieved. Future researches should first focus on risk identification in preschool children. Second, as dyslexia is not specific to a region, language, or culture, language-independent data collection methods need to be developed to create a uniform standard database of dyslexia. Finally, we need to collect data from multiple sources (e.g., scales, behavioral tests, brain imaging, etc.), mix multiple models, and consider multimodal deep learning frameworks to improve the predictive power of machine learning, continuously optimize the constructed dyslexia screening models, and eventually achieve widespread use in clinical practice.

Key words: dyslexia, machine learning, early screening, children