Table 1 Variants of ComBat and their recommended use cases.

Longitudinal

ComBat²⁶

Extends ComBat to longitudinal data using linear mixed-effects models to preserve within-subject changes over time.

Longitudinal MRI studies with repeated measures per subject (e.g., disease progression, development), where standard ComBat would remove true within-subject changes.

ComBat-GAM²⁹

Incorporates generalized additive models to capture non-linear covariate effects (e.g., age trajectories).

Large multi-site datasets with wide covariate ranges (e.g., lifespan studies) where covariate effects are non-linear and need preservation.

AutoComBat²⁷

Automates ComBat by inferring batch labels from data (metadata, quality metrics) when they are unknown or incomplete.

When site/scanner labels are unavailable, inconsistent, or when harmonizing unseen batches without predefined batch IDs.

CovBat¹⁴

Extends ComBat to harmonize not just mean and variance but also covariance between features.

Multivariate MRI features (e.g., radiomics, connectivity) where site differences in feature correlations could harm downstream ML models.

GMM-ComBat¹⁶

Uses Gaussian mixture models to split heterogeneous/multi-modal feature distributions before applying ComBat.

Sites with multi-modal feature distributions (e.g., mixed protocols, subpopulations) where a single Gaussian assumption is violated.

Cluster-ComBat³⁰

(Federated ComBat)

Groups sites into clusters and harmonizes in a distributed way, allowing new sites to be integrated without retraining.

Federated/multi-center setups with privacy constraints or frequent addition of new sites, where central pooling is not possible.