Machine Learning Differentiation of Autism Spectrum Sub-Classifications

Area of Science:

• Neuroscience
• Computational Psychiatry

Background:

• Autism spectrum disorder (ASD) presents with diverse challenges in communication and daily functioning.
• Early identification of ASD is crucial for effective intervention.
• Changes in diagnostic criteria (DSM-IV to DSM-5) complicate ASD diagnosis.

Purpose of the Study:

• To evaluate the efficacy of machine learning in classifying individuals with autism spectrum disorder.
• To differentiate between three specific ASDs under DSM-IV criteria and non-spectrum cases.

Main Methods:

• Machine learning algorithms were applied to a large retrospective dataset of 38,560 individuals.
• Analysis incorporated clinical, demographic, and assessment data.

Main Results:

• The machine learning model demonstrated high performance with Area Under the Receiver Operating Characteristic Curves (AUROCs) from 0.863 to 0.980.
• The algorithm achieved an overall correct classification rate of 80.5%.
• A misclassification rate of 12.6% occurred between different autism spectrum disorders.

Conclusions:

• Machine learning offers a viable tool for classifying individuals with autism spectrum disorder versus non-spectrum.
• The models can effectively utilize limited data inputs for diagnostic classification.