Title: MRI biomarkers extraction, teachings from an autism-prediction challenge
Abstract: Can MRI be useful to extract practical biomarkers of brain pathology? We recently ran a blind challenge of prediction of Autism spectrum disorder diagnostic status. The 10 winning solutions scored on average 0.81 area under the ROC curve, which is a very good prediction score. This confirms our experience that combining machine learning with MRI can lead to robust biomarkers.
I will talk about what these challenges teaches us: tackling heterogeneity and overfit, and the importance of resting-state fMRI. I will show that its findings are consistent with other studies of biomarkers extractions.
Moving back in time, I will summarize what we have learned over the years on biomarker extraction from MRI, in particular from resting-state fMRI.