ScreeningPaL: LLM-NLP Enabled Early Autism Detection Method from Caregiver's Free-Text Input.

Area of Science:

• Developmental Pediatrics
• Computational Linguistics
• Artificial Intelligence in Healthcare

Background:

• Early identification of autism spectrum disorder (ASD) traits is crucial for improving long-term quality of life.
• Current screening methods may lack accessibility or require specialized tools like speech samples.
• Caregiver-reported behavioral descriptions offer rich, albeit unstructured, data for potential early risk assessment.

Purpose of the Study:

• To develop and evaluate a text-driven approach for early autism risk detection using natural language processing (NLP).
• To assess the performance of advanced language models in analyzing free-text behavioral descriptions.
• To explore the impact of data augmentation on model generalization and performance.

Main Methods:

• Utilized synthetic free-text data generated from validated screening items to train language models.
• Employed fine-tuned transformer models, comparing their performance against other models like GPT and Gemini, and traditional NLP baselines.
• Evaluated model generalization on an external benchmark dataset (TASD) under domain shift conditions.
• Investigated the effect of augmenting training data with noisy, realistic text.

Main Results:

• Fine-tuned transformer models achieved 90% accuracy in early autism risk detection.
• These models outperformed GPT, Gemini, and conventional NLP baselines.
• Noise-aware data augmentation enhanced model performance, particularly improving recall in traditional pipelines.
• The methodology demonstrated effective generalization on an external dataset.

Conclusions:

• A text-driven NLP approach enables low-cost, accessible early autism risk assessment without structured questionnaires or speech samples.
• This method provides valuable early cues to support specialist evaluations and proactive developmental monitoring.
• Advanced language models, especially fine-tuned transformers, show significant promise for analyzing unstructured behavioral data in developmental screening.