Genes and Screens: Machine Learning Reveals What Drives Dyslexia Risk in Children

Developmental dyslexia, one of the most common learning disorders worldwide, has long been understood as a condition shaped by both nature and nurture. A new multi-center study from China, published in Translational Psychiatry, has now brought these two forces together in an unusually comprehensive analysis, combining machine learning on environmental risk factors with cutting-edge polygenic risk scoring across several psychiatric and neurodevelopmental disorders. The results offer one of the clearest pictures yet of how home environment and inherited genetic vulnerability jointly shape a child’s risk of struggling to read.

The research team, led by Haoxue Wang and Ranran Song of the Department of Maternal and Child Health at Tongji Medical College, Huazhong University of Science and Technology, screened 29,820 primary school students in grades two through six across four Chinese cities. From this large population, they enrolled 948 children who met diagnostic criteria for developmental dyslexia and matched them with 948 controls of the same age and sex. This case-control design, anchored in a very large screening base, gives the study considerable statistical power and reduces the chance that observed associations are artifacts of sampling variation.

To disentangle the many environmental variables that could plausibly contribute to reading difficulties, the researchers turned to eXtreme Gradient Boosting, or XGBoost, a machine learning algorithm that excels at modeling complex, nonlinear relationships among predictors. Rather than relying on traditional regression alone, which can miss interactions between variables, XGBoost builds an ensemble of decision trees that collectively rank the importance of each factor. The team then applied Shapley additive explanations, known as SHAP, a technique borrowed from cooperative game theory that quantifies how much each individual feature contributes to the model’s predictions for each child. This combination allowed the investigators to move beyond simple correlation and identify which environmental exposures carry the most predictive weight.

The findings on the environmental side are striking. The most influential risk factors for dyslexia were the child’s proactive learning habits, the age at which the child first acquired literacy, maternal education level, parental out-migration for work, and prolonged screen exposure during pregnancy. The home literacy environment emerged as the single most important domain, accounting for the largest share of variance in dyslexia status within both the case and control groups. In practical terms, this suggests that the richness of books, reading routines, and verbal engagement in the household may be a central lever in either buffering or amplifying a child’s underlying risk.

Several of these factors deserve closer attention. The age of literacy onset is particularly notable: children who began reading later appeared at elevated risk, which aligns with a substantial body of evidence that early exposure to printed language builds the phonological and orthographic foundations on which fluent reading depends. Maternal education level, a well-established socioeconomic proxy, likely captures a bundle of advantages, from greater access to reading materials to more enriched language input in the earliest years. Parental out-migration, a widespread phenomenon in rapidly urbanizing economies, can disrupt caregiving continuity and reduce the day-to-day supervision of learning habits. And prenatal screen exposure, a more novel finding, points to the possibility that maternal habits during pregnancy may influence neurodevelopment in ways that only large, well-powered studies can detect.

While environmental analysis formed one pillar of the study, the second pillar reached into the genome. The researchers calculated polygenic risk scores, or PRS, which aggregate the effects of thousands of common genetic variants identified in large-scale genome-wide association studies. Each variant contributes a tiny amount of risk on its own, but summed across the genome, the resulting score captures an individual’s inherited liability to a given condition. The team computed PRS not only for dyslexia itself but also for attention deficit hyperactivity disorder, autism spectrum disorder, schizophrenia, major depressive disorder, and bipolar disorder, allowing them to test whether the genetic architecture of dyslexia overlaps with that of other neurodevelopmental and psychiatric conditions.

The genetic results were compelling. Children in the highest quartile of dyslexia polygenic risk had a 2.13-fold higher odds of dyslexia compared with those in the lowest quartile, after adjusting for environmental risk factors, with a 95 percent confidence interval of 1.08 to 4.18. This means that even after accounting for the home environment, parental education, and screen exposure, inherited genetic liability remained a substantial and independent predictor of reading difficulty. The dose-response pattern across quartiles further strengthens the inference that the relationship is not a statistical fluke.

More provocatively, the study found cross-disorder genetic overlap. Children in the highest quartile of autism spectrum disorder PRS had 2.13-fold higher odds of dyslexia, with a 95 percent confidence interval of 1.14 to 3.98, and those in the highest quartile of schizophrenia PRS had 2.41-fold higher odds, with a confidence interval of 1.23 to 4.72. These findings suggest that dyslexia shares a measurable portion of its genetic underpinnings with conditions that, on the surface, seem quite different. Such pleiotropy, where the same genetic variants influence multiple traits, is increasingly recognized as a common feature of neurodevelopmental disorders and may reflect shared pathways in brain development, synaptic function, or neural migration.

The implications of this dual approach are significant for both science and policy. By demonstrating that environmental factors and polygenic risk each contribute independently to dyslexia, the study supports a model in which genetic vulnerability sets a baseline of risk that environmental conditions can either exacerbate or mitigate. The identification of modifiable factors, particularly the home literacy environment and early literacy onset, offers concrete targets for intervention. Screening programs that assess both family reading practices and, eventually, genetic risk profiles could help identify children who would benefit most from early, intensive literacy support before academic failure takes its toll on self-esteem and psychosocial development.

The authors emphasize that their findings offer novel insight into the etiology of dyslexia and contribute to the advancement of early prevention strategies. As machine learning methods like XGBoost and SHAP become more common in epidemiological research, and as polygenic risk scoring matures with ever-larger genomic datasets, studies of this kind are likely to multiply. For a condition that affects millions of schoolchildren worldwide and often goes undetected until years of struggle have already accumulated, the message from this research is clear: both the books in the home and the variants in the genome matter, and understanding how they interact may be the key to catching dyslexia before it takes root.

Subject of Research: Environmental and genetic risk factors for developmental dyslexia in primary school children

Article Title: Environmental factors and polygenic risk score for psychiatric and neurodevelopmental disorders in children with dyslexia

Article References: Wang, H., Fan, H., Liu, R., Zhu, K., Xiang, Z., Zhao, S., Chen, Q., Zhang, J., Wang, T., Liang, X., & Song, R. (2026). Environmental factors and polygenic risk score for psychiatric and neurodevelopmental disorders in children with dyslexia. Translational Psychiatry. https://doi.org/10.1038/s41398-026-04429-0

Image Credits: AI Generated

DOI: 10.1038/s41398-026-04429-0

Keywords: dyslexia, polygenic risk score, machine learning, XGBoost, SHAP, home literacy environment, neurodevelopmental disorders, genetics, children, screen exposure, Translational Psychiatry, early prevention