Robust deep learning pipeline to improve access to autism triage and diagnosis

Timely access to autism spectrum disorder (ASD) diagnosis represents a major clinical burden, especially in children. Robust and quantitative triage tools to refer children at high risk of ASD for rapid evaluation are not available. We investigate a robust artificial intelligence pipeline combining anatomic and resting state MRI scans with demographics to triage children into high and low risk of an ASD diagnosis. The ABIDE I and II datasets were used to train deep learning (DL) models on both the anatomic and resting state MRI scans, which were then combined with demographics to render a prediction for an ASD diagnosis. The pipeline was formally tested using a train (70%, n = 1555), validation (20%, n = 439) and test (10%, n = 222) data split ensuring robust testing and assessment of generalizability. DL model outputs were combined with demographics in a generalized linear model with Lasso regularization, and was then optimized using area-under-the-curve (AUC) analysis. The test, validation and train groups had similar demographics in which 46–52% of the population had an ASD diagnosis, 80–83% were male, with a median age of 13–14 and 71–75% reported eyes open during the resting state MRI. The AI pipeline successfully distinguished between ASD and non-ASD with suspicion scores of 0.803 [0.461–0.964] compared to 0.377 [0.044–0.873], p < 0.0001 and an AUC of 71.1%. Under AUC optimization, an accuracy of 62.6% with a sensitivity of 58.9% and specificity of 66.1% was achieved. The artificial intelligence pipeline described here is a promising method for accurately distinguishing between ASD and non-ASD participants.