AI-Based Computational Modeling of Language and Communication Difficulties among School-Age Learners

Abstract

Language and communication difficulties among school-age learners are often overlooked in multilingual, resource-constrained education systems such as Pakistan, where access to speech-language specialists remains limited. This study developed and evaluated an AI-based screening framework using a nine-domain, 54-item rating instrument completed by teachers and parents for 428 learners aged 6–14 years from seven provinces and three types of educational institutions. The instrument demonstrated excellent internal consistency and satisfactory inter-rater agreement, identifying 25.7% of learners as being at risk of language and communication difficulties. Five machine-learning classifiers were evaluated, with random forest showing the strongest predictive performance (AUC=0.906). Expressive communication, multilingual and code-switching difficulties, and academic impact emerged as the most influential predictors, and their contributions were further supported by logistic regression analysis. K-means clustering identified four distinct learner profiles, including a multilingual-specific profile representing 17.8% of the sample that was not adequately captured by composite risk scores. Difficulty patterns also varied according to school type and medium of instruction, with Urdu–English exposure exerting the strongest influence on code-switching difficulties. These findings support the Multidomain Linguistic Difficulty Computational Screening Model (MLD-CSM) as a reliable and scalable framework for identifying learners who may require further language and communication assessment, particularly in multilingual educational settings with limited specialist resources.