Finding the forest in the trees: Using machine learning and online cognitive and perceptual measures to predict adult autism diagnosis

TL;DR

This study uses online cognitive and perceptual tasks with machine learning to accurately predict autism diagnosis in adults, especially those who are hard to identify.

Contribution

The study introduces a machine learning approach combining performance-based measures and questionnaires to improve autism prediction accuracy.

Findings

Performance-based measures predicted autism in a late-diagnosed population with high accuracy.

Combining these measures with a popular screening questionnaire increased predictive accuracy to 92%.

Variables without significant group differences still contributed to prediction, indicating complex relationships.

Abstract

Traditional subjective measures are limited in the insight they provide into underlying behavioral differences associated with autism and, accordingly, their ability to predict diagnosis. Performance-based measures offer an attractive alternative, being designed to capture neuropsychological constructs more directly and objectively. However, due to the heterogeneity of autism, differences in any one specific neuropsychological domain are inconsistently detected. Meanwhile, protracted wait times for diagnostic interviews delay access to care, highlighting the importance of developing better methods for identifying individuals likely to be autistic and understanding the associated behavioral differences. We administered a battery of online tasks measuring multisensory perception, emotion recognition, and executive function to a large group of autistic and non-autistic adults. We then used machine learning to classify participants and reveal which factors from the resulting dataset were most predictive of diagnosis. Not only were these measures able to predict autism in a late-diagnosed population known to be particularly difficult to identify, their combination with the most popular screening questionnaire enhanced its predictive accuracy (reaching 92% together). This indicates that performance-based measures may be a promising means of predicting autism, providing complementary information to existing screening questionnaires. Many variables in which significant group differences were not detected had predictive value in combination, suggesting complex latent relationships associated with autism. Machine learning’s ability to harness these connections and pinpoint the most crucial features for prediction could allow optimization of a screening tool that offers a unique marriage of predictive accuracy and accessibility.

Figures