Objective:

To improve early autism diagnosis by integrating automated diagnostic labeling and parental input using a BioBERT machine learning model.

Approach:

• Model Training: Trained a BioBERT model to label individual autism behavioral descriptions based on DSM-5 criteria.
• Performance Evaluation: Evaluated model performance on lay and clinical behavior descriptions, comparing diagnostic utility across dimensions.

Key Findings:

• BioBERT achieved higher precision (69%) on clinical descriptions and higher recall (83%) on lay descriptions.
• Transferring models between data types resulted in performance drops.
• Training on clinical data yielded the best-performing diagnostic models.
• Lay behavioral descriptions provided diagnostically valuable information comparable to clinical observations.

Interpretation:

The integration of lay information into diagnostic workflows could enhance the speed of autism diagnosis without compromising clinical utility.

Limitations:

• Performance differences were not explained by sample size.
• AI-generated summaries were only moderately representative of examples.

Conclusion:

The study indicates that lay narratives may provide valuable information in autism diagnosis.