Large language models deconstruct the clinical intuition behind diagnosing autism.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Clinical Psychology

Background:

• Current diagnostic methods for autism spectrum disorder (ASD) using genome-wide assays or brain scans have yielded limited success.
• Clinical intuition of healthcare professionals remains the primary method for diagnosing ASD, highlighting a gap in objective diagnostic tools.

Purpose of the Study:

• To deconstruct and analyze the logic behind expert clinician intuition in diagnosing autism using deep learning.
• To identify key indicators within clinical reports that contribute to accurate autism diagnoses.

Main Methods:

• Leveraged large language models (LLMs) pre-trained on general text and fine-tuned on over 4,000 clinical health records.
• Employed an explainability strategy within the LLM architecture to pinpoint salient sentences driving diagnostic decisions.

Main Results:

• The fine-tuned LLM successfully distinguished between confirmed and suspected autism cases.
• The framework identified stereotyped repetitive behaviors, special interests, and perception-based behaviors as critical diagnostic indicators for autism.
• The findings challenge the current diagnostic emphasis on deficits in social interaction.

Conclusions:

• Deep learning analysis of clinical reports can effectively capture and interrogate expert diagnostic intuition.
• The study suggests a need to revise current diagnostic criteria for autism, particularly the Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (DSM-5), to better incorporate behavioral and perceptual factors.