ComGamHarm

Adam Lang

• Overview
• Installation
• ComGamHarm
  • Arguments
• Example
  • Cross-Sectional Data
  • Harmonize
• Applying Harmonization To New Data
  • Example
• Plots
  • Plots Before Harmonization
    • ROI1
    • ROI2
    • ROI3
    • ROI4
  • Plots After Harmonization
    • ROI1
    • ROI2
    • ROI3
    • ROI4

Overview

Based on ComBatHarmonization and neuroHarmonize

ComGamHarm estimates and removes site-effect for each feature while preserving biological variance. If Empirical Bayes is specified, ComGamHarm can pool information across features and estimate site effect with more accuracy. ComGamHarm utilizes gam models to allow for preservation of non-linear biological variance. This is often important in AD biomarker data as there is a known non-linear relationship between AD biomarkers and age.

ComGamHarm requires cross sectional data in order to estimate site effect. If a user wants to harmonize longitudinal data, they can first run ComGamHarm on the cross sectional data and use ApplyHarm to then harmonize their longitudinal data. Both the longitudinal data and cross sectional data must have the same batch labels. This also works with any new data containing the same batch labels

If you are using this package please cite the follwing papers:

	Cite	Link
ComBat for multi-site DTI data	Jean-Philippe Fortin, Drew Parker, Birkan Tunc, Takanori Watanabe, Mark A Elliott, Kosha Ruparel, David R Roalf, Theodore D Satterthwaite, Ruben C Gur, Raquel E Gur, Robert T Schultz, Ragini Verma, Russell T Shinohara. Harmonization Of Multi-Site Diffusion Tensor Imaging Data. NeuroImage, 161, 149-170, 2017	Here
ComBat for multi-site cortical thickness measurements	Jean-Philippe Fortin, Nicholas Cullen, Yvette I. Sheline, Warren D. Taylor, Irem Aselcioglu, Philip A. Cook, Phil Adams, Crystal Cooper, Maurizio Fava, Patrick J. McGrath, Melvin McInnis, Mary L. Phillips, Madhukar H. Trivedi, Myrna M. Weissman, Russell T. Shinohara. Harmonization of cortical thickness measurements across scanners and sites. NeuroImage, 167, 104-120, 2018	Here
Original ComBat paper for gene expression array	W. Evan Johnson and Cheng Li, Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics, 8(1):118-127, 2007.	Here
Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan	Pomponio, R., Shou, H., Davatzikos, C., et al., (2019). “Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan.” Neuroimage 208.	Here

Installation

Both ComGamHarmFunction.R and ComGamHarmFunctionHelpers.R must be downloaded. The following packages must be installed:

install.packages("ggplot2")
install.packages("mgcv")
install.packages("dplyr")
install.packages("reshape2")
install.packages("matrixStats")
install.packages("ebbr")
install.packages("BiocParallel")
install.packages("furniture")

ComGamHarm

Arguments

ComGamHarm requires the following arguments:

• feature.data data.frame data.frame with features for harmonization

• covar.data data.frame data.frame with covariates for harmonization. All covariates included in covar.data will be preserved during harmonization. covar.data must contain a column STUDY corresponding to batch

• eb logical whether or not to perform Empirical Bayesion estimation of site effect. TRUE/FALSE defaults FALSE

• parametric logical parametric EB adjustments defaults TRUE

• smooth.terms character vector of non-linear covariates for harmonization Ex: c(“Age”)

• k.val numeric vector of smooth parameters for non-linear covariates for harmonization. Must be ordered in respect to smooth.terms

• verbose logical print model fit progress defaults TRUE

• model.diagnostics logical return gam fit diagnostics defaults FALSE

Example

Cross-Sectional Data

#data generated in DataSimulation.R
#not real medical data. 

#first 5 subjects in each Site
knitr::kable(full.simulated.data.cs[c(1:5, 301:306, 601:606),])

ROI1	ROI2	ROI3	ROI4	Age	Sex	ApoE	Education	STUDY	id
6560.067	1031.1221	36833.68	15209.550	72.19762	2	1	18.47753	A	1
5845.162	1044.4538	35690.31	12440.928	73.84911	1	1	17.81281	A	2
4601.741	1009.8177	33542.62	12279.165	82.79354	2	2	16.75394	A	3
6633.555	1210.9749	33618.14	11469.827	75.35254	2	2	17.23873	A	4
4696.089	1412.6101	33241.65	12656.505	75.64644	2	1	17.58859	A	5
5340.044	1507.8670	39532.46	15977.043	72.19762	2	1	18.47753	B	301
7050.704	1477.6788	32607.81	16950.297	73.84911	1	1	17.81281	B	302
4768.223	1367.9741	33102.17	15598.596	82.79354	2	2	16.75394	B	303
5027.141	1409.2130	34729.24	16162.799	75.35254	2	2	17.23873	B	304
5470.245	1299.2436	31844.18	15610.725	75.64644	2	1	17.58859	B	305
5004.993	1345.2858	29191.05	13891.303	83.57532	1	1	11.39115	B	306
4927.961	1112.7856	33460.39	11613.298	72.19762	2	1	18.47753	C	601
5268.513	1183.4759	32266.08	13217.589	73.84911	1	1	17.81281	C	602
3887.168	1079.3900	29715.05	9979.098	82.79354	2	2	16.75394	C	603
5162.835	1049.1165	35851.78	12106.939	75.35254	2	2	17.23873	C	604
4852.780	1208.9107	33630.76	11994.121	75.64644	2	1	17.58859	C	605
3818.349	893.1338	31108.71	10500.881	83.57532	1	1	11.39115	C	606

Harmonize

#Feature Data
feature.data <- full.simulated.data.cs[,c("ROI1", "ROI2", "ROI3", "ROI4")]

#Covariate Data
covariate.data <- full.simulated.data.cs[,c("Age","Sex", "ApoE", "Education", "STUDY")]


#Harmonize
harmfeats <- ComGamHarm(feature.data = feature.data,
                        covar.data   = covariate.data,
                        eb           = TRUE,
                        parametric   = TRUE,
                        smooth.terms = c("Age"),
                        k.val        = 5, 
                        verbose      = FALSE)

Harmonized features are contained in harmfeats[[“harm.results”]]. They are in matrix format with dimension n_features X n_observations and must be transposed to match the original dimensons.

harmonized.features <- as.data.frame(t(harmfeats$harm.results))

We can then recombine our new harmonized features with the rest of our data

harm.data <- cbind(harmonized.features, covariate.data)

Applying Harmonization To New Data

Using ApplyHarm we can apply pre-trained harmonization models to new data under the condition the new data comes from the same sites.

ApplyHarm requires the following arguments:

• feature.data data.frame data.frame with features for harmonization

• covar.data data.frame data.frame with covariates for harmonization. All covariates included in covar.data will be preserved during harmonization. covar.data must contain a column STUDY corresponding to batch

• comgam.out list output from ComGamHarm function

Example

harm.longitudinal <- ApplyHarm(feature.data   = feature.data.long,
                               covariate.data = covariate.data.long,
                               comgam.out     = harmfeats)

harm.longitudinal <- as.data.frame(t(harm.longitudinal))

harm.longitudinal <- cbind(harm.longitudinal, covariate.data.long)

Plots

Plots Before Harmonization

ROI1

ROI2

ROI3

ROI4

Plots After Harmonization

ROI1

ROI2

ROI3

ROI4