Multivoxel Pattern Analysis in R • rMVPA

A methodologically rigorous toolbox for decoding and representational analysis of neuroimaging data. One model spec runs regionally or via searchlight. Cross-validation respects fMRI run structure. RSA, MS-ReVE contrast decomposition, model-space connectivity, ERA cross-decoding, and domain-adaptive REMAP-RRR are all first-class — no one-off scripts in a graveyard of unmaintained lab repos.

📖 Documentation: https://bbuchsbaum.github.io/rMVPA/

The mental model

mvpa_dataset  ──►   mvpa_design   ──►   model_spec   ──►   engine     ──►   result
(voxels x time      (response,         (rsa_model,        (run_regional      (regional_mvpa_result
 + mask)             block_var,         mvpa_model,        run_searchlight)   searchlight_result)
                     splits)            feature_rsa, …)

Three S3 layers — dataset → design → model_spec — and two engines (run_regional / run_searchlight). The same model_spec works in both. Cross-validation, parallelism (via future), error handling, and per-ROI diagnostics come for free.

What you can build with it

You want to …	Reach for	Vignette
Decode condition labels from local activity (8-way category, condition pair, etc.)	`mvpa_model()` + `run_searchlight()` / `run_regional()`	`Searchlight_Analysis`, `Regional_Analysis`, `Haxby_2001` ⓡ
Test whether a region’s pattern geometry matches a model RDM	`rsa_model()` + `rsa_design()` (or `pair_rsa_design()`)	`RSA`, `Kriegeskorte_92_Images` ⓡ
Ask which regions share representational geometry through your models	`model_space_connectivity()` on `rsa_model(..., return_fingerprint = TRUE)`	`Model_Space_Connectivity`
Decompose representational structure by signed contrasts (MS-ReVE)	`contrast_rsa_model()`	`Contrast_RSA`
Cross-decode between encoding & retrieval (ERA), or perception & memory	`era_rsa_model()`, `naive_xdec_model()`, `remap_rrr_model()`, the ReNA family	`ERA_RSA_Cross_Decoding`, `REMAP_RRR`, and the `Naive_Cross_Decoding` glossary
Plug in a custom per-ROI / per-sphere function	`run_custom_regional()` / `run_custom_searchlight()`	`CustomAnalyses`, `Plugin_Development`

ⓡ = real public data, not synthetic.

A 30-second taste

library(rMVPA)

# Synthetic 6 x 6 x 6, 80 trials, 4 runs, 2 categories
data <- gen_sample_dataset(D = c(6, 6, 6), nobs = 80, blocks = 4, nlevels = 2)

dset <- mvpa_dataset(data$dataset$train_data, mask = data$dataset$mask)
cval <- blocked_cross_validation(data$design$block_var)
mod  <- load_model("sda_notune")

mspec <- mvpa_model(mod, dataset = dset, design = data$design,
                    crossval = cval,
                    tune_grid = data.frame(lambda = 0.01, diagonal = FALSE))

# Per-voxel decoding map
sl <- run_searchlight(mspec, radius = 4, method = "standard")
names(sl$results)   # "Accuracy", "AUC"

The Get Started vignette walks the same example, then routes you to classification, RSA, and cross-domain workflows.

Real-data demonstrations

The catalogue includes two non-synthetic vignettes that reproduce canonical findings on public data, end-to-end:

Kriegeskorte 2008 (vignette) — 92 object images, 4 subjects × 2 sessions of human-IT RDMs. Recovers the published model-RDM ranking (monkey-IT and animacy lead; low-level vision lags) and shows model-space connectivity across subjects.
Haxby 2001 (vignette) — Subject 1 ventral-temporal patterns. Reaches 91.7 % 8-way category accuracy at 12.5 % chance with sda_notune. Reproduces the canonical face / house / object confusion structure.

The bundled data lives under inst/extdata/; the data-raw/*.R scripts regenerate it from upstream sources.

Installation

From within R:

# install.packages("pak")
pak::pak("bbuchsbaum/neuroim2")    # required, not on CRAN yet
pak::pak("bbuchsbaum/rMVPA")

Or with devtools:

library(devtools)
install_github("bbuchsbaum/neuroim2")
install_github("bbuchsbaum/rMVPA")

Command-line interface

rMVPA ships packaged CLI wrappers for searchlight and regional analyses:

# After installing, copy the wrappers onto your PATH
rMVPA::install_cli("~/.local/bin", overwrite = TRUE)

rmvpa-searchlight --help
rmvpa-regional   --help

See vignette("CommandLine") for input formats, configuration files, and parallelisation.

API stability

Public-API surface is intentionally narrow and explicit. Use rmvpa_api_lifecycle() to inspect lifecycle tiers and rmvpa_stable_api() to list the stable entry points intended for scripts, extensions, and downstream packages.

Citation

If you use rMVPA in your research, please cite:

Buchsbaum, B. (2026). rMVPA: Multivoxel Pattern Analysis in R. R package version 0.1.2. https://github.com/bbuchsbaum/rMVPA

rMVPA — multivoxel pattern analysis in R