Run Searchlight Analysis

Execute a searchlight analysis using multivariate pattern analysis.

Usage

run_searchlight(
  model_spec,
  radius,
  method = c("standard", "randomized", "resampled"),
  niter = 4,
  backend = c("default", "shard", "auto"),
  ...
)

# S3 method for class 'feature_rsa_model'
run_searchlight(
  model_spec,
  radius,
  method = c("standard", "randomized", "resampled"),
  niter = 4,
  backend = c("default", "shard", "auto"),
  ...
)

Arguments

model_spec

A mvpa_model instance containing the model specifications

radius

The searchlight radius in millimeters

method

The type of searchlight, either 'randomized' or 'standard'

niter

The number of searchlight iterations (used only for 'randomized' method)

backend

Execution backend: "default" (standard pipeline), "shard" (shared-memory backend), or "auto" (try shard and fall back to default).

...

Extra arguments passed to specific searchlight methods. Currently supported:

• batch_size: Integer specifying how many searchlights (ROIs) are grouped into a single batch for processing by mvpa_iterate. Each batch is launched sequentially, while ROIs within a batch are processed in parallel (using the active future/furrr plan). The default is 10% of the total number of searchlights. Smaller values lower peak memory usage and can improve load balancing, at the cost of more scheduling overhead; larger values reduce overhead but require more memory per worker. As a rule of thumb, start with the default, decrease batch_size if you hit memory limits, and increase it if you have many ROIs, ample RAM, and see low CPU utilization.

Value

A named list of NeuroVol objects containing performance metrics (e.g., AUC) at each voxel location

Progress reporting

When verbose = TRUE is passed via ... and the progressr package is installed, real-time per-ROI progress updates are emitted from parallel workers. To enable a progress bar:

  library(progressr)
  handlers(global = TRUE)               # once per session
  result <- run_searchlight(mspec, radius = 8, verbose = TRUE)

Without progressr, only coarse batch-level log messages are shown.

Examples

# \donttest{
  # Generate sample dataset with categorical response
  dataset <- gen_sample_dataset(
    D = c(5,5,5),           # 5x5x5 volume
    nobs = 60,              # 60 observations
    response_type = "categorical",
    data_mode = "image",
    blocks = 3,             # 3 blocks for cross-validation
    nlevels = 2             # binary classification
  )
  
  # Create cross-validation specification using blocks
  cval <- blocked_cross_validation(dataset$design$block_var)
  
  # Load the SDA classifier (Shrinkage Discriminant Analysis)
  model <- load_model("sda_notune")
  
  # Create MVPA model
  mspec <- mvpa_model(
    model = model,
    dataset = dataset$dataset,
    design = dataset$design,
    model_type = "classification",
    crossval = cval
  )
  
  # Run searchlight analysis
  results <- run_searchlight(
    mspec,
    radius = 3,            # 3mm radius
    method = "standard"    # Use standard searchlight
  )
#> INFO [2026-10-16 11:46:41] searchlight engine: sda_fast
  
  # Run with custom batch size for memory management
  # results <- run_searchlight(
  #   mspec,
  #   radius = 3,
  #   method = "standard",
  #   batch_size = 500      # Process 500 searchlights per batch
  # )
  
  # Results contain performance metrics
  # Access them with results$performance
# }