brain-spi

Comparison of pairwise statistics for fMRI connectivity estimation, and a pipeline for finding group differences across them.

brain_spi wraps the SPI connectivity-analysis workflow into a small, importable library. It lets you run the full pipeline in one call and then inspect either the aggregated cross-SPI result or any individual SPI's intermediate artifacts.

Try it now in Colab (no install, nothing to clone):

What it does

• Connectivity — derive functional network connectivity (FNC) matrices from multivariate fMRI time series using many pairwise statistics (SPIs), via pyspi plus a small custom set.
• Group differences — per SPI, find edges that differ between two groups using a Welch t-test (Bonferroni-corrected) and a Random-Forest importance mask, then intersect them.
• Aggregate — average the per-SPI masks into a consensus matrix: the fraction of SPIs that flag each edge.
• Robustness — subject-resampling (bootstrap) and label-shuffle permutation nulls.
• Plots & caching — heatmaps with network guides, per-SPI triptychs, the aggregate panel; on-disk caching of the (slow) pyspi computations.

Install

pip install brain-spi

I recommend a separate Python environment: brain-spi relies on pyspi, which requires numpy < 2 and can conflict with other packages. For development, clone the repo and pip install -e . instead.

Quickstart

import numpy as np
from brain_spi import BrainSPI

# data: (B subjects, T timepoints, C channels/ROIs);  labels: (B,) with 2 unique values
pipe   = BrainSPI(group_names=('HC', 'patient'))   # sensible defaults (9 curated SPIs)
result = pipe.fit(data, labels)

# headline result — cross-SPI aggregate (computed lazily, cached on first access)
result.aggregate.mean_and        # (C, C) float: fraction of SPIs flagging each edge
result.aggregate.plot()          # one-shot figure

# inspect a single SPI
result['kendalltau'].mean_matrix()       # (C, C) mean connectivity across subjects
result['kendalltau'].and_mask            # significant-p AND RF-important edges
result['kendalltau'].plot_triptych()     # 3-panel: sig-p / RF / AND

# not sure what's available? every result object has a repr + .help()
result.help()

Choosing SPIs

BrainSPI()                                   # default 9-SPI curated set
BrainSPI(spis='spis_all')                    # the broad pyspi set
BrainSPI(spis='/path/to/my_spis.yaml')       # a config file
BrainSPI(spis=['kendalltau', 'spearmanr'])   # an explicit list

SPIs are computed one at a time across all subjects, with a progress bar and a per-SPI wall-time log (enable with logging.basicConfig(level=logging.INFO)), so it's easy to spot and drop slow ones.

Robustness

boot = result.bootstrap(n=20, frac=0.66)     # 20 subject-resampled cross-SPI AND maps
boot.mean                                     # average of the 20 (resampling-smoothed aggregate)
boot.survival_rate()                          # how often each edge is flagged (reproducibility)

null = result.label_shuffle(n=100)            # permutation null
null.p_value(result.aggregate.mean_and)       # per-edge permutation p-values (low = significant)

Saving results

The first fit is slow (pyspi dominates); subsequent fits on the same data are near-instant thanks to the on-disk cache (~/.cache/brain_spi/, override with BrainSPI(cache_dir=...)). To save a finished result:

result.to_npz('result.npz')              # portable — open with plain numpy.load, no package needed
result.to_pickle('result.pkl')           # exact object

import brain_spi
result = brain_spi.load_npz('result.npz')   # or load_pickle(...)

The .npz is a flat collection of arrays with a self-describing README key inside.

Example notebook

examples/colab_quickstart.ipynb runs end-to-end on public data — ABIDE (controls vs. autism) by default, or COBRE (schizophrenia) via a one-line switch. Part 1 computes SPIs and inspects their mean connectivity; Part 2 runs the significant-differences pipeline. Datasets download automatically via examples/datasets.py.