An fMRI study lives as thousands of files: NIfTI images, events.tsv files, confound regressors, JSON sidecars. The BIDS H5 archive captures an entire study in one compressed HDF5 file that is queryable by subject, task, session, and run. You download one file and start analyzing immediately.

library(fmridataset)

Which compression mode should you use?

The archive supports two modes. Pick the one that matches your analysis:

Mode	What it stores	Reconstruction	Best for
`parcellated`	Cluster averages `[T, K]`	Parcel-level only	ROI analyses, connectivity
`latent`	Basis `[T, K]` + loadings `[V, K]`	Full voxel resolution	Searchlight, fine-grained spatial

Both modes store events, confounds, censor vectors, and BIDS metadata alongside the data. The standard fmridataset API works on either.

How do you create an archive?

compress_bids_study() reads a BIDS directory and streams scans one at a time into the archive. Only one NIfTI is held in memory at once.

Parcellated mode

library(bidser)

bids <- bids_project("/path/to/my_study")
atlas <- neuroim2::read_vol("schaefer_400.nii.gz")

study <- compress_bids_study(
  bids,
  file = "my_study.h5",
  mode = "parcellated",
  clusters = atlas
)

The result is a bids_h5_study_dataset you can use immediately.

Latent mode

Latent mode uses fmrilatent::encode() to compress each scan into a low-rank basis + loadings representation that can be reconstructed back to voxel resolution.

library(fmrilatent)

study <- compress_bids_study(
  bids,
  file = "my_study_latent.h5",
  mode = "latent",
  encoding = spec_time_dct(k = 30),
  mask = brain_mask
)

For multi-subject studies that share a spatial atlas, a shared template stores the loadings once and only keeps per-scan coefficients, significantly reducing file size:

tpl <- parcel_basis_template(parcellation, basis_spec = basis_slepian(k = 10))

study <- compress_bids_study(
  bids,
  file = "my_study_template.h5",
  mode = "latent",
  template = tpl,
  mask = brain_mask
)

How do you open an archive?

One function, one file:

study <- bids_h5_dataset("my_study.h5")
study

#> <bids_h5_study_dataset>
#>   File: my_study.h5
#>   Mode: parcellated (400 parcels)
#>   Subjects: 20 | Tasks: nback, rest | Sessions: pre, post
#>   Scans: 80 | Total timepoints: 24000 | TR: 2s

Exploring the study

BIDS metadata is directly accessible:

participants(study)
tasks(study)
sessions(study)
scan_manifest(study)

scan_manifest() returns a tibble with one row per scan:

#>   scan_name                       subject task  session run n_time
#>   sub-01_ses-pre_task-nback_run-01 01     nback pre     01  300
#>   sub-01_ses-pre_task-rest_run-01  01     rest  pre     01  200
#>   ...

Subsetting by task, subject, or session

Use subset_bids_h5() to carve out the data you need. This returns a new bids_h5_study_dataset backed by the same file:

nback <- subset_bids_h5(study, task = "nback")
sub01 <- subset_bids_h5(study, subject = "01")
pre_nback <- subset_bids_h5(study, task = "nback", session = "pre")

Subsetting is cheap — it selects scan backends from the shared H5 connection, no data is copied.

Accessing data

The standard fmridataset API works on the result:

mat <- get_data_matrix(nback)
dim(mat)

#> [1] 6000  400

That is [total_timepoints, K] where K is the number of parcels (or components in latent mode). Per-subject data:

mat01 <- get_data_matrix(nback, subject_id = "01")
dim(mat01)

#> [1] 600 400

Events

The event table combines all scans with task, session, subject_id, and both the BIDS run label and an internal run_id:

head(nback$event_table)

#>   onset duration trial_type run run_id subject_id task  session
#>   0.0   2.0      face       01  1      01         nback pre
#>   4.0   2.0      house      01  1      01         nback pre
#>   ...

Confounds

conf <- get_confounds(study, scan_name = "sub-01_ses-pre_task-nback_run-01")
head(conf)

Chunked iteration

Memory-efficient processing via data_chunks():

chunks <- data_chunks(nback, nchunks = 10)
while (!is.null(chunk <- iterators::nextElem(chunks))) {
  # Process chunk$data [T, K_chunk]
}

Group operations

Convert to fmri_group for per-subject analyses:

group <- study_to_group(nback)

Latent-mode extras

When working with a latent-mode archive, three additional accessors are available:

info <- encoding_info(study)
info$encoding_family
info$n_components
info$has_shared_template

loadings <- get_loadings(study, scan_name = "sub-01_task-nback_run-01")
dim(loadings)

recon <- reconstruct_voxels(study,
  scan_name = "sub-01_task-nback_run-01",
  rows = 1:10,
  voxels = roi_indices
)
dim(recon)

reconstruct_voxels() computes basis %*% t(loadings) + offset on the fly, so you only materialize the slice you need.

Parcellation metadata

For parcellated archives, parcellation_info() gives you the cluster mapping:

pinfo <- parcellation_info(study)
pinfo$n_parcels
pinfo$cluster_ids

This returns NULL for latent-mode archives.

Next steps

vignette("fmridataset-intro") for the core dataset API
vignette("study-level-analysis") for multi-subject workflows without BIDS
vignette("backend-development-basics") if you want to write a custom backend
?compress_bids_study, ?bids_h5_dataset, ?subset_bids_h5 for full parameter docs

BIDS H5 Archive: Compressing a Study into a Single File