Detecting autism from picture book narratives using deep neural utterance embeddings.

Area of Science:

• Computational linguistics
• Psychiatric diagnostics
• Machine learning in healthcare

Background:

• Social communication deficits, including storytelling, are characteristic of autism spectrum disorder (ASD).
• Machine learning (ML) offers potential for aiding psychiatric diagnosis and treatment decisions.
• Current ML applications in ASD detection primarily use screening questionnaires or image data.

Purpose of the Study:

• To evaluate deep neural networks (DNNs) for detecting ASD from textual narratives.
• To assess the efficacy of specific text encoders (ELMo, USE) and classifiers (XGBoost, SVM, DNN) in ASD detection.
• To compare computational model performance against human expert ratings and standardized diagnostic instruments.

Main Methods:

• Utilized Embeddings from Language Models (ELMo) and Universal Sentence Encoder (USE) for text representation.
• Employed XGBoost, support vector machines, and dense neural network layers for classification.
• Classified 50 participants (25 ASD, 25 typical development) based on Autism Diagnostic Observation Schedule, Second Edition (ADOS-2) picture book task narrations.
• Compared model performance with two psychiatrists' classifications and standardized ASD screening tools.

Main Results:

• Computer-based models demonstrated superior sensitivity, specificity, and predictive values compared to human raters.
• Model performance was lower than established instruments like ADOS-2 and the Social Communication Questionnaire (SCQ).
• ELMo and USE text encoders showed promising performance metrics for ASD detection.

Conclusions:

• Deep neural network-based text models can augment ASD diagnosis and screening.
• ELMo and USE show potential as effective text encoders for ASD detection from narrative data.
• Page-level embeddings are valuable for representing utterances in tasks like the ADOS-2 picture book narrative.