A cross-dataset study on automatic detection of autism spectrum disorder from text data

Area of Science:

• Computational linguistics
• Neurodevelopmental disorders
• Artificial intelligence

Background:

• Autism spectrum disorder (ASD) diagnosis can be challenging.
• Text data holds potential for identifying ASD biomarkers.
• Advancements in machine learning offer new avenues for diagnostic tools.

Purpose of the Study:

• To evaluate machine learning models for detecting ASD from transcribed statements.
• To compare model performance across datasets from different diagnostic tools (TLC and ADOS-2).
• To explore knowledge transfer and data augmentation for improving ASD detection.

Main Methods:

• Fine-tuning a BERT-based transformer model (HerBERT).
• Utilizing OpenAI's text embeddings with a classifier.
• Applying methods to two distinct ASD diagnostic datasets (TLC and ADOS-2).
• Conducting cross-dataset experiments, including zero-shot and transfer learning scenarios.

Main Results:

• Models achieved very good performance on TLC data but average results on ADOS-2 data.
• No significant benefits were observed from knowledge transfer between datasets.
• Data augmentation using back translation showed poor performance, potentially obscuring ASD-specific signals.

Conclusions:

• Machine learning model performance for ASD detection from text is improving.
• Model accuracy is significantly influenced by the input data and the specific diagnostic tool used.