Identifying overlapping and distinctive traits of autism and schizophrenia using machine learning classification

Area of Science:

• Neuroscience
• Psychiatry
• Computational Psychiatry

Background:

• Autism spectrum disorder (ASD) and schizophrenia spectrum disorder (SSD) share overlapping symptomatology.
• Differentiating these conditions is crucial for accurate diagnosis and treatment.
• Existing research screeners, such as the Autism-Spectrum Quotient (AQ) and Schizotypal Personality Questionnaire - Brief Revised (SPQ-BR), are widely used in non-clinical and subclinical populations.

Purpose of the Study:

• To investigate if the AQ and SPQ-BR can identify non-overlapping symptoms between ASD and SSD using machine learning.
• To determine the extent of symptom overlap and identify distinctive features between the two conditions.
• To provide guidance on utilizing these screeners for improved diagnostic accuracy and biomarker identification.

Main Methods:

• A cohort of 1,397 undergraduates completed both the SPQ-BR and AQ.
• Random forest classification models were employed to assess predictive relationships between item scores and factors of the AQ and SPQ-BR.
• Models were trained using all item scores, and subsequently, the least and most important features were analyzed.

Main Results:

• Robust trait overlap was confirmed, enabling prediction of AQ scores from SPQ-BR scores and vice versa.
• AQ item scores predicted disorganised and interpersonal factors of the SPQ-BR.
• SPQ-BR item scores predicted communication and social skills factors of the AQ.
• Crucially, classification model failures indicated that AQ item scores could not predict the SPQ-BR cognitive-perceptual factor, and SPQ-BR item scores could not predict imagination, attention to detail, or attention switching factors of the AQ.

Conclusions:

• The SPQ-BR and AQ effectively measure overlapping symptoms, with some factors being distinguishable.
• Model failures highlight specific, non-overlapping symptom domains unique to each condition.
• Findings offer a framework for leveraging existing screeners to prevent misdiagnosis and advance the identification of specific biomarkers for ASD and SSD.