Extended Invariant Information Clustering Is Effective for Leave-One-Site-Out Cross-Validation in Resting State Functional Connectivity Modeling.

Naoki Okamoto,Hiroyuki Akama

doi:10.3389/fninf.2021.709179

Abstract

Herein, we propose a new deep neural network model based on invariant information clustering (IIC), proposed by Ji et al., to improve the modeling performance of the leave-one-site-out cross-validation (LOSO-CV) for a multi-source dataset. Our Extended IIC (EIIC) is a type of contrastive learning; however, unlike the original IIC, it is characterized by transfer learning with labeled data pairs, but without the need for a data augmentation technique. Each site in LOSO-CV is left out in turn from the remaining sites used for training and receives a value for modeling evaluation. We applied the EIIC to the resting state functional connectivity magnetic resonance imaging dataset of the Autism Brain Imaging Data Exchange. The challenging nature of brain analysis for autism spectrum disorder (ASD) can be attributed to the variability of subjects, particularly the rapid change in the neural system of children as the target ASD age group. However, EIIC demonstrated higher LOSO-CV classification accuracy for the majority of scanning locations than previously used methods. Particularly, with the adjustment of a mini-batch size, EIIC outperformed other classifiers with an accuracy >0.8 for the sites with highest mean age of the subjects. Considering its effectiveness, our proposed method might be promising for harmonization in other domains, owing to its simplicity and intrinsic flexibility.

Highlights

Machine learning discrimination has been widely applied to resting-state functional connectivity (RSFC) datasets which cover a wide range of neural diseases such as schizophrenia (Shen et al, 2010), mild cognitive impairment (Chen X. et al, 2016), and autism spectrum disorder (ASD) (Guo et al, 2017)
In the wake of previous studies pertaining to the LOSOoriented Autism Brain Imaging Data Exchange (ABIDE) analysis, we proposed an information clustering (IIC)-based deep neural network model based on the original IIC of Ji et al This model was developed for discriminating the labeled data of ASD and TC in the ABIDE dataset only with the RSFC data obtained from almost all imaging locations
We proposed a new algorithm for multi-site harmonization of an RSFC dataset, derived from different sources

Summary

Introduction

Machine learning discrimination has been widely applied to resting-state functional connectivity (RSFC) datasets which cover a wide range of neural diseases such as schizophrenia (Shen et al, 2010), mild cognitive impairment (Chen X. et al, 2016), and autism spectrum disorder (ASD) (Guo et al, 2017). Monk et al (2009) demonstrated a difference in the strength of connectivity withinExtended Invariant Information Clustering the default mode network between ASD and control groups; the magnitude of this difference was correlated with the severity of symptoms. Guo et al (2017) used a support vector machine (SVM) as a classifier and successfully differentiated between the two groups with >80% accuracy for data derived from a single imaging location in the Autism Brain Imaging Data Exchange (ABIDE) collection. Machine learning performs quite poorly on combined datasets derived from different imaging sites This can be attributed to the effects of data heterogeneity, such as differences in the characteristics of MRI scanner suppliers, scan parameters ( the length of the repetition time [TR]), and the cohort setting, wherein variability in phenotype, a deviation in the age and gender group composition, is unavoidable. We experienced a considerable drop in the accuracy on excluding the overall data of each single site to create an independent testing set and featuring the remaining sites as the provider of the modeling set This leave-one-site-out cross-validation (LOSO-CV) technique has certain complications, in the case of ASD, such as the diverse age distribution for selecting the subjects. Before running the LOSO-CV, they preselected the age bracket in the range of 12–18 years and removed the sites with

Methods

Results

Discussion

Conclusion