Online Program

Return to main conference page
Friday, May 18
Bioinformatics/Biomedical
Fri, May 18, 10:00 AM - 10:45 AM
Regency Ballroom B
 

Diagnostic Prediction of Autism in Resting-State Functional Mri Using Conditional Random Forest (304589)

Ralph-Axel Muller, San Diego State University 
Brian Faires, San Diego State University 
Juanjuan Fan, San Diego State University 
*Afrooz Jahedi, San Diego State University 
Chanond Art Nasamran, San Diego State University 

Keywords: autism spectrum disorder, resting-state fMRI, intrinsic functional connectivity, machine learning, diagnostic prediction, conditional random forest

Background Autism spectrum disorder (ASD) is characterized by social and behavioral impairments. Although it is a neurodevelopmental disorder, no unique brain biomarkers for ASD are known. Previous research has used machine learning and computational statistics to mine MRI data for ASD biomarkers. Chen et al. (NICL 2015) achieved 90.8% diagnostic accuracy using the ensemble learning technique, random forest (RF). However, RF is known to have an intrinsic variable selection bias (Strobl et al., BMC Bioinformatics 2007). In order to eliminate this bias and to increase the interpretability of the results, we developed a conditional random forest (CRF) ASD diagnostic prediction model for resting-state functional MRI data (rs-fMRI).

Methods Rs-fMRI data from 252 patients (126 ASD and 126 TD) were selected from the Autism Brain Imaging Data Exchange (ABIDE). Preprocessing techniques and participant selection criteria were adopted from Chen at al. to allow for direct comparisons between the RF and CRF models. Connectivity matrices for each participant were created using 220 regions of interest (ROI) from Power et al. (Neuron 2011). The dimensionality of the dataset was reduced to eliminate noise and for computational feasibility.

Results The CRF model achieved a diagnostic accuracy of 92.5-95% (in two runs with different seeds) from 180 most informative connections between ROIs. Most informative connections (normalized based on total number of possible connections per network) were heavily represented in somatosensory/motor (especially mouth region), ventral attention, salience, cingulo-opercular, memory retrieval, and cerebellar networks. Raw numbers for default mode network (before normalization) were also high.

Discussion/Conclusions CRF reached very high, though not perfect, diagnostic prediction accuracy based on a complex set of 180 functional connectivities. Reduced variable selection bias compared to the earlier RF study (Chen et al., 2015) resulted not only in