Normative Measures Harmonization for Pathway Dependent Connectivity Measures in Multiple Sclerosis
Poster No:
1497
Submission Type:
Abstract Submission
Authors:
Mark Lowe1, Ajay Nemani1, XUEMEI HUANG2, Katherine Koenig1
Institutions:
1The Cleveland Clinic, Cleveland, OH, 2Cleveland Clinic, Cleveland, OH
First Author:
Co-Author(s):
Introduction:
We recently validated a statistically straightforward, connectivity-based metric as a marker for multiple sclerosis (MS) disease status and progression [1]. This metric is composed of pathway-dependent structural and functional connectivity measures and normalized using values derived from sex- and age-matched healthy controls (HC). The stability and robustness of this metric depends on having a stable HC distribution from which to derive normative values. The applicability of the metric depends on ensuring that the measures are not biased by the measurement systems used to collect functional and structural connectivity data. In this study, we apply a harmonization procedure intended to produce device-independent normative values using data obtained from a multi-MRI scanner platform dataset known as the Cleveland Clinic Brain Study (CCBS)[2].
Methods:
The structural and functional connectivity index (SFCI) is calculated using resting state functional connectivity (rsfMRI) and diffusion-based structural connectivity (DTI). In this study, we use rsfMRI and DTI data acquired by the CCBS, a lifespan study of older healthy adults[2]. Data for this study are produced on two MRI scanners at Cleveland Clinic, a Siemens Skyra 3T and a Siemens Prisma 3T MRI. We selected two sets of 10 age- and sex-matched subjects from each scanner. The same protocol is run on both scanners, so any biases are device-dependent and not protocol-dependent. Identical processing pipelines were performed on all data. Structural connectivity was taken to be mean pathway-dependent radial diffusivity and functional connectivity was taken to be the z-score-converted Pearson correlation of seed regions at the endpoints of the desired pathways. Details of the SFCI and calculation of component measures is described in Koenig et al.[1].
ComBat data harmonization[3, 4] is a popular batch correction tool that removes inter-site variability while preserving inherent biological variability. We applied harmonization separately to the structural and functional connectivity results. ComBat was applied with default parameters using the three underlying pathways described above as features across subjects. The SFCI was then recalculated on these harmonized data.
ComBat data harmonization[3, 4] is a popular batch correction tool that removes inter-site variability while preserving inherent biological variability. We applied harmonization separately to the structural and functional connectivity results. ComBat was applied with default parameters using the three underlying pathways described above as features across subjects. The SFCI was then recalculated on these harmonized data.
Results:
Paired t-tests between original SFCI and ComBat harmonized SFCI showed no significant differences. Figure 1 shows a comparison of the harmonized results on component and combined metrics. Comparison of Figures 1c and 1d show that ComBat is effective at removing scanner-dependent effects such that norms can be calculated across data acquired from both scanners, resulting in normalization that is equivalent to using within-scanner normalization (1d). Figure 2 shows the SFCI calculated for each of the 20 subjects, with each column representing a different calculation method for the normative values used for z-scoring. Column 1 uses unharmonized values normed across both scanners. Column 2 uses unharmonized values normed separately for each scanner. Column 3 uses norms calculated using harmonized data across all scanners. The fact that there is very little difference between SFCI calculated in columns 2 and 3 indicates that harmonization has effectively removed scanner-dependent biases from the pooled calculation.
Conclusions:
ComBat is increasingly utilized in imaging studies because of its wide availability and effectiveness at removing site- or system-dependent biases from multi-platform studies. Here, we demonstrate that ComBat can be used in a combined DTI and rsfMRI context to effectively remove scanner biases. Using harmonized versions of our metrics, we can take advantage of a very large pooled dataset to produce stable normative values for SFCI calculation in individual MS patients. By incorporating more platforms, the goal would be to produce norms that would be applicable to connectivity measures obtained on any MRI scanner.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 1
Methods Development 2
Keywords:
Data analysis
FUNCTIONAL MRI
Other - Harmonization
1|2Indicates the priority used for review
Provide references using author date format
2. ; Available from: https://my.clevelandclinic.org/departments/neurological/research-innovations/brain-study.
3. Fortin, J.-P., et al., Harmonization of multi-site diffusion tensor imaging data. NeuroImage, 2017. 161: p. 149-170.
4. Johnson, W.E., C. Li, and A. Rabinovic, Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics, 2006. 8(1): p. 118-127.