Enhancing Autism-Linked Language Disorder Diagnosis Using Large Language Models

Overview

This study demonstrates that large language models (LLMs) can improve the diagnosis of language disorders associated with autism spectrum disorder (ASD) by increasing sensitivity and positive predictive value by over 10% in zero-shot learning settings. The approach identifies key linguistic features such as echolalia and pronoun reversal, offering a promising supplementary tool to traditional clinical assessments.

Background

Autism spectrum disorder is characterized by social communication difficulties and restricted behaviors, with language impairments being a significant challenge, especially in adults. Traditional diagnostic methods like the ADOS-2 rely heavily on clinical expertise and are resource-intensive, limiting scalability and consistency. Machine learning approaches have been explored but face challenges due to high data requirements and difficulty in capturing subtle language disorder patterns. Large language models, trained on vast text corpora, offer potential advantages in analyzing ASD-related communication behaviors with less reliance on labeled data.

Data Highlights

Key Findings

• LLMs achieved over 10% improvement in sensitivity and positive predictive value compared to baseline models in zero-shot learning configurations.
• Ten key linguistic features associated with autism-related language disorders were identified, including echolalia, pronoun reversal, and atypical language usage.
• LLMs can detect subtle and variable language disorder patterns that traditional ML models struggle to reliably extract.
• The approach reduces reliance on large labeled datasets, addressing a major limitation of conventional machine learning methods.
• LLMs offer a scalable, efficient, and objective supplementary tool to enhance clinical ASD diagnostic processes.

Clinical Implications

Incorporating LLM-based analysis into ASD diagnostic workflows can improve detection of language impairments, enabling earlier and more accurate identification of communication difficulties. This facilitates personalized therapeutic interventions tailored to specific linguistic features, potentially improving social and occupational outcomes for adults with ASD. Additionally, the reduced dependence on extensive labeled data may increase accessibility and consistency of assessments across clinical settings.

Conclusion

Large language models represent a promising adjunct to traditional ASD diagnostic methods by enhancing sensitivity and specificity in identifying autism-related language disorders. Their ability to recognize key linguistic markers supports more precise and personalized clinical interventions.

References

1. OpenAI 2023 -- GPT-3.5 and GPT-4o Development
2. Lord et al. 2012 -- Autism Diagnostic Observation Schedule, Second Edition (ADOS-2)
3. Recent ML Studies 2019-2023 -- Machine Learning in ASD Diagnosis