Diagnosing autism spectrum disorder using brain entropy: A fast entropy method

Abstract Background and objective Previous resting-state fMRI-based diagnostic models for autism spectrum disorder (ASD) were based on traditional linear features. The complexity of the ASD brain remains unexplored. Methods To increase our understanding of the nonlinear neural mechanisms in ASD, entropy (i.e., approximate entropy (ApEn) and sample entropy (SampEn)) method was used to analyze the resting-state fMRI datasets collected from 21 ASD patients and 26 typically developing (TD) individuals. Here, a fast entropy algorithm was proposed through matrix computation. We combined entropy with a support-vector machine and selected "important entropy" as features to diagnose ASD. The classification performance of the fast entropy method was compared to the state-of-the-art functional connectivity (FC) method. Results The area under the receiver operating characteristic curve based on FC was 0.62. The areas under the receiver operating characteristic curves based on ApEn and SampEn were 0.79 and 0.89, respectively. The results showed that the proposed fast entropy method was more efficacious than the FC method. In addition, lower entropy was found in the ASD patients. The ApEn of the left postcentral gyrus (rs = −0.556, p = 0.009) and the SampEn of the right lingual gyrus (rs = −0.526, p = 0.014) were both significantly negatively related to Autism Diagnostic Observation Schedule total scores in the ASD patients. The proposed algorithm for entropy computation was faster than the traditional entropy method. Conclusions Our study provides a new perspective to better understand the neural mechanisms of ASD. Brain entropy based on a fast algorithm was applied to distinguish ASD patients from TD individuals. ApEn and SampEn could be potential biomarkers in ASD investigations.