Detailed contribution information

Back to list

Contribution title	Using cognitive measures to identify developmental language disorder in mono- and bilingual children: insights from machine learning
Contribution code	D3.126
Authors	Jade Plym University of Helsinki and Helsinki University Hospital, Finland Presenter Federico Màlato University of Eastern Finland, Finland Pekka Lahti-Nuuttila Sini Smolander University of Eastern Finland, Finland Eva Arkkila University of Helsinki and Helsinki University Hospital Sari Kunnari University of Oulu, Finland Rosa González-Hautamäki University of Eastern Finland, Finland Marja Laasonen University of Eastern Finland, Finland
Form of presentation	Poster
Topic	T37 - Speech and language
Abstract	Aims. Developmental language disorder (DLD) is a highly prevalent neurodevelopmental disorder defined by a range of persistent difficulties in language acquisition. Accurate identification of DLD is challenging, especially in populations with diverse language environments, such as sequential bilingual children. The purpose of this study is to investigate the accuracy of a cognitive task set to differentiate DLD and typical development (TD) in monolingual and bilingual children, and to assess which tasks are best in the differentiation. Moreover, we aim to examine how Random Forest (RF), a machine learning (ML) technique, compares to traditional statistical methods. Methods. The study was part of the Helsinki longitudinal SLI study. The participants were 294 4–7-year-old Finnish-speaking monolingual children (n=150) with DLD and TD, and sequential bilingual children (n=144) with DLD and TD. The assessment battery included 22 cognitive tasks conducted in Finnish language, measuring different domains, for example reasoning, language processing and memory. The classification accuracy in the mono- and bilingual groups was evaluated using RF and traditional statistical methods. Results and conclusions. Preliminary results showed that using RF models, DLD can be identified with cognitive tasks accurately in monolingual children (classification accuracy: 91.3%) and with fair accuracy in bilingual children (84.7%). Language processing and verbal reasoning tasks were best in both language groups. The mono- and bilingual models were different, discouraging the use of monolingual reference population when assessing bilingual children. Traditional statistical methods suffered from overfitting, which demonstrates that ML approach can be more viable when studying numerous tasks.