Constructing a Predictive Model for Children with Autism Spectrum Disorder Based on Whole-Brain MR Radiomics: A...

Current autism spectrum disorder (ASD) diagnosis depends on subjective behavioral assessments causing delays and variability. Structural MRI shows brain abnormalities but traditional analysis lacks sensitivity. This study aims to construct a prediction model based on whole-brain imaging radiomics analysis by using machine learning for identifying children with ASD.

This study used imaging and clinical data from 223 subjects in the Autism Brain Imaging Data Exchange database, including 120 patients diagnosed with ASD. These data were randomly divided into training and test sets in a 7:3 ratio. An independent external test set comprising 87 participants (38 with ASD) from the Georgetown University and University of Miami ASD data set was also utilized. Then, quantitative radiomics features of white matter, gray matter, and CSF were extracted from the whole-brain MR structural images of each subject, and feature dimensionality reduction was performed based on the training set data to construct radiomics signature. Afterward, multivariate logistic regression was used to screen independent predictors of ASD from clinical features and then, combined with radiomics signature, multiple machine learning models were constructed to predict ASD. Finally, the optimal model was selected, and the receiver operating characteristic curve was used to evaluate the training, test, and external test sets data. Simultaneously, the model divides the data set into low-risk and high-risk subgroups, comparing the actual number of individuals with ASD between the 2 subgroups to evaluate the clinical efficacy of the model.

The areas under the curve (AUCs) of radiomics markers in the training set, test set, and external test set were 0.78 (95% CI, 0.71-0.85), 0.75 (95% CI, 0.67-0.83), and 0.74 (95% CI, 0.64-0.83), respectively. Multivariate logistic regression showed that the Verbal Intelligence Quotient (VIQ) was a predictor of ASD. The joint model constructed by the decision tree algorithm with radiomics markers performed best, with AUC of 0.87 (95% CI, 0.81-0.92), 0.84 (95% CI, 0.76-0.91), and 0.83 (95% CI, 0.74-0.91), a sensitivity of 0.89 (95% CI, 0.82-0.95), 0.84 (95% CI, 0.73-0.93), and 0.86 (95% CI, 0.72-0.96), and the specificity of 0.70 (95% CI, 0.60-0.79), 0.63 (95% CI, 0.52-0.74), and 0.66 (95% CI, 0.49-0.80), respectively. The low-risk subgroup and high-risk subgroup classified according to the cutoff value of 0.4285 of the model showed statistically significant differences in the actual number of patients with ASD in the training set (χ² = 21.325; P < .05), the test set (χ² = 5.379; P < .05), and the external test set (χ² = 21.52; P < .05).

Radiomics signature for identifying ASD can be constructed based on whole-brain MRI imaging features. The performance of identifying ASD can be improved by adding VIQ data and the decision tree algorithm model, which can provide an adaptive strategy for clinical practice.