Using Predictive Performance from an Elastic Net Regression to Classify Developmental Language Disorder (DLD).

Area of Science:

• Developmental Psychology
• Linguistics
• Computational Neuroscience

Background:

• Developmental disorders are increasingly understood through a multidimensional lens, acknowledging complex, variable contributing factors.
• Developmental Language Disorder (DLD) presents a unique challenge due to its heterogeneous nature and the continuum of language abilities.

Purpose of the Study:

• To determine if individual variability in high-density measures can identify developmental language disorder (DLD).
• To leverage a multidimensional approach to classify children with DLD from typically developing peers.

Main Methods:

• Utilized repeated elastic net logistic regression on 71 high-density measures from 223 children (ages 7-11).
• Employed a two-study approach: Study 1 trained/tested on data subsets; Study 2 used the complete dataset for model derivation.
• Evaluated model performance using Area Under the Receiver Operating Characteristic Curve (AUROC).

Main Results:

• Study 1 achieved an average AUROC of 0.88 (training) and 0.85 (holdout) for discriminating DLD and typically developing (TD) groups.
• Study 2 reported an average AUROC of 0.87 for discriminating groups.
• Model-derived probability scores significantly correlated with language severity measures.

Conclusions:

• Successfully developed a predictive model classifying children with DLD using a multidimensional dataset and an elastic net algorithm.
• Findings provide indirect support for conceptualizing DLD as a multidimensional disorder.
• Discussion includes challenges in data-driven model derivation and comparative analysis of methodologies.