Cognitive Measures in Developmental Language Disorder Classification in Monolingual and Bilingual Children: A Machine Learning Approach

Purpose: We used machine learning (ML) models to investigate the accuracy of a cognitive assessment battery to differentiate developmental language disorder (DLD) and typical development (TD) in monolingual and sequential bilingual children. Additionally, we tested how the model trained on monolingual children can classify bilingual children and examined the relative importance of the different linguistic and nonlinguistic tasks in the classifications. Method: The participants were 4- to 7-year-old monolingual and sequential bilingual children with DLD ( n = 167) or TD ( n = 127) from the Helsinki longitudinal SLI study. The assessment battery included standardized tasks used to measure different domains of language and other cognition. To investigate the ability of the tasks to classify mono- and bilingual children into having DLD or TD, we used a random forest ML classification model. Results: The cognitive assessment battery classified DLD/TD well in the monolingual group (91.3%) and fairly well in the bilingual group (84.7%). However, the model trained with monolingual data was not accurate in the bilingual group (66.0%). The best tasks for classifying DLD and TD reflected language processing and verbal reasoning in both mono- and bilingual children. The nonlinguistic tasks did not considerably improve the classification. Conclusions: This study is among the first to employ ML methods for DLD classification and presents a cognitive assessment battery for detecting DLD in mono- and bilingual children. The current results show that bilingual children's performance should not be compared to monolingual standards. The role of the nonlinguistic functions remains unclear. Supplemental Material: https://doi.org/10.23641/asha.31069621