Exploring White Matter Abnormalities in Young Children with Autism Spectrum Disorder: Integrating Multi-shell Diffusion Data and Machine Learning Analysis.

This study employed tract-based spatial statistics (TBSS) to investigate abnormalities in the white matter microstructure among children with autism spectrum disorder (ASD). Additionally, an eXtreme Gradient Boosting (XGBoost) model was developed to effectively classify individuals with ASD and typical developing children (TDC). Multi-shell diffusion weighted images were acquired from 62 children with ASD and 44 TDC. Using the Pydesigner procedure, diffusion tensor (DT), diffusion kurtosis (DK), and white matter tract integrity (WMTI) metrics were computed. Subsequently, TBSS analysis was applied to discern differences in these diffusion parameters between ASD and TDC groups. The XGBoost model was then trained using metrics showing significant differences, and Shapley Additive explanations (SHAP) values were computed to assess the feature importance in the model's predictions. TBSS analysis revealed a significant reduction in axonal diffusivity (AD) in the left posterior corona radiata and the right superior corona radiata. Among the DK indicators, mean kurtosis, axial kurtosis, and kurtosis fractional anisotropy were notably increased in children with ASD, with no significant difference in radial kurtosis. WMTI metrics such as axonal water fraction, axonal diffusivity of the extra-axonal space (EAS_AD), tortuosity of the extra-axonal space (EAS_TORT), and diffusivity of intra-axonal space (IAS_Da) were significantly increased, primarily in the corpus callosum and fornix. Notably, there was no significant difference in radial diffusivity of the extra-axial space (EAS_RD). The XGBoost model demonstrated excellent classification ability, and the SHAP analysis identified EAS_TORT as the feature with the highest importance in the model's predictions. This study utilized TBSS analyses with multi-shell diffusion data to examine white matter abnormalities in pediatric autism. Additionally, the developed XGBoost model showed outstanding performance in classifying ASD and TDC. The ranking of SHAP values based on the XGBoost model underscored the significance of features in influencing model predictions. [ABSTRACT FROM AUTHOR]