spider-directed-connectivity-stitching

name	spider-directed-connectivity-stitching
description	SPIDER: Non-parametric frequency-domain framework for recovering directed brain connectivity from incomplete asynchronous recordings. Stitches power-spectra across sessions. Activation: effective connectivity, directed information flow, SPIDER, brain connectivity stitching, 脑连接拼接, 有效连接
tags	["neuroscience","brain-networks","effective-connectivity","frequency-domain","multi-session"]

SPIDER: Stitched Power-spectra for Inferring Directed Information Flow

arXiv: 2606.22695
Authors: Yisi S. Zhang, Daniel Y. Takahashi
Date: 2026-06-21 (updated 2026-07-04)

Core Methodology

Problem Statement

Mapping directed information flow (effective connectivity) between brain regions is central to understanding brain function, but large-scale recordings sample only a fraction of the brain at a time. Sessions, animals, and laboratories cover different, partially overlapping regions, usually without a shared temporal reference. Established methods (Granger causality, DCM, PDC) require all regions recorded simultaneously with a common clock.

Key Innovation

SPIDER (Stitched Power-spectra for Inferring Directed information flow from incomplete and asynchronous Experimental Recordings):

• Non-parametric, frequency-domain framework
• Stitches local power-spectral estimates from overlapping channel subsets into a global spectral matrix
• Obtains frequency-resolved directed interactions via canonical spectral factorization and PDC
• No temporal alignment required
• Nuclear-norm completion fills in never-co-observed region pairs

Technical Pipeline

1. Local spectral estimation: compute power-spectra from overlapping channel subsets
2. Stitching: assemble global spectral matrix from local estimates
3. Completion: nuclear-norm completion for unobserved region pairs
4. Factorization: canonical spectral factorization
5. Direction inference: Partial Directed Coherence (PDC)

Mathematical Framework

Spectral Matrix Stitching

S_global(f) = stitch({S_local^(i)(f)})

where each S_local^(i) comes from overlapping channel subset i.

Nuclear-Norm Completion

For never-co-observed region pairs:

min ||S||_* subject to S_observed = S_global_observed

Directed Connectivity via PDC

From spectral matrix factorization:

PDC_{i→j}(f) = |A_{ij}(f)|² / Σ_k |A_{kj}(f)|²

where A(f) is the transfer function from spectral factorization.

Experimental Results

Validation Datasets

1. Simulations: ground-truth directed connectivity
2. Two-photon calcium imaging: mouse brain
3. IBL Neuropixels dataset: 50 brain areas from 43 sessions across 12 laboratories (never recorded together)
4. Human intracranial EEG: 43 patients with non-overlapping coverage

Key Findings

• Brain-wide spontaneous flow is largely recurrent
• Theta band: significant feedforward hierarchy with hippocampal formation at source
• Cross-species consistency: same theta-band hierarchy recovered in mouse and human
• Cross-modality consistency: calcium imaging and electrophysiology agree

Scale Achievement

• Recovered directed flow among 50 areas from 43 sessions in 12 laboratories
• Previously impossible with traditional methods

Practical Applications

For Brain Network Analysis

1. Multi-session connectivity: combine data across experimental sessions
2. Cross-laboratory integration: merge datasets from different labs
3. Cross-species comparison: compare connectivity patterns across species
4. Large-scale brain mapping: whole-brain effective connectivity from partial observations

Implementation Guidelines

When to Use

• Multiple recording sessions with overlapping but incomplete coverage
• Cross-laboratory data integration
• Cross-species connectivity comparison
• Scenarios where simultaneous whole-brain recording is impossible

Requirements

• Overlapping channel subsets across sessions
• Frequency-domain data (power spectra)
• Sufficient overlap for stitching

Pitfalls

• Requires overlapping channels between sessions
• Nuclear-norm completion assumes low-rank structure
• Frequency resolution affects temporal precision
• Validation requires ground-truth or strong assumptions

Algorithm Steps

Step 1: Local Spectral Estimation

for session in sessions:
    S_local[session] = compute_power_spectrum(data[session])

Step 2: Stitching

S_global = stitch_local_spectra(S_local_list, overlap_info)

Step 3: Completion

S_completed = nuclear_norm_completion(S_global, observed_mask)

Step 4: Factorization & PDC

A = spectral_factorization(S_completed)
PDC = compute_pdc(A)

Comparison with Traditional Methods

Method	Requires Simultaneous Recording	Temporal Alignment	Handles Incomplete Data
Granger Causality	Yes	Yes	No
DCM	Yes	Yes	No
PDC	Yes	Yes	No
SPIDER	No	No	Yes

Related Concepts

• Effective connectivity
• Granger causality
• Dynamic causal modeling (DCM)
• Partial directed coherence (PDC)
• Spectral factorization
• Nuclear norm minimization
• Multi-session data integration

References

• Paper: https://arxiv.org/abs/2606.22695
• PDF: https://arxiv.org/pdf/2606.22695