ComBat Harmonization

ComBat uses an Empirical Bayes approach to pool information across features, providing more stable estimates of batch effect parameters when the number of samples per batch is small. This is critical in radiomics where multi-center cohorts often have limited cases per scanner.

• Borrows strength across all features to estimate batch parameters
• Outperforms naive location-scale adjustments when sample sizes are unbalanced
• Reduces over-correction risk compared to fully Bayesian methods

The algorithm models batch effects as additive (location) and multiplicative (scale) shifts in feature values. It estimates and removes these shifts to align the means and variances of feature distributions across different imaging centers or scanners.

• Corrects systematic shifts in feature mean values
• Adjusts for variance inflation or deflation per batch
• Preserves the rank order of biological differences within each batch

A defining strength of ComBat is its ability to preserve biological covariates of interest during harmonization. By specifying clinical variables such as tumor grade or survival status in the model matrix, the algorithm ensures that variability attributable to these factors is not removed as unwanted noise.

• Protects disease-related signal during correction
• Prevents harmonization from masking true group differences
• Requires careful pre-specification of the biological model design

Standard ComBat assumes normally distributed feature values. For radiomic features that deviate from normality, non-parametric ComBat (ComBat-NP) uses empirical quantile mapping to harmonize distributions without parametric assumptions.

• Standard ComBat: Assumes Gaussian feature distributions
• ComBat-NP: Harmonizes arbitrary distributions via quantile matching
• M-ComBat: Robust variant using median-based estimators for outlier resistance

Originally developed by Johnson, Li, and Rabinovic (2007) for microarray gene expression data, ComBat has been extensively validated for radiomic feature harmonization. The method translates naturally because both domains share the problem of high-dimensional feature matrices confounded by technical acquisition variability.

• Proven in genomics: Corrects lab-to-lab expression variation
• Validated in radiomics: Harmonizes CT texture features across scanner vendors
• Requires feature-wise application with identical feature extraction pipelines

ComBat assumes that batch effects are consistent across the feature space and that the biological and batch effects are independent. It also requires that each batch contains samples from all biological subgroups to avoid confounding.

• Cannot correct batch effects confounded with biology
• Sensitive to extreme outliers in small batches
• Does not address non-linear batch effects without extensions