harmonization-tool

Harmonization Tool Skill (Cross-Site Feature Alignment Layer)

Overview

harmonization-tool is the NeuroClaw cross-cutting layer that sits between dataset skills (ABIDE, ADHD-200, ABCD, HCP, UKB, ...) and model skills (BrainGNN, BNT, IBGNN, LGGNN, BrainNetCNN, FM-APP, SVM, SpaceNet, ...).

Its job: take subject-level features extracted by dataset skills and remove technical / batch variance introduced by site, scanner, field strength, sequence, or dataset, while preserving biological variance (age, sex, diagnosis, ...).

This is the prerequisite for any honest mega-analysis that pools individual-participant data (IPD) across sites or datasets.

This skill follows NeuroClaw hierarchy:

• Defines WHAT to do, not low-level implementation details.
• Does not execute direct shell commands itself.
• Delegates all execution via claw-shell.

Research use only.

---

When to Use This Skill

Trigger this skill when the user asks for any of:

• "harmonize features across sites / scanners / datasets"
• "ComBat / ComBat-GAM / CovBat / neuroHarmonize / neuroCombat"
• "remove site effect / scanner effect / batch effect"
• "mega-analysis on ABIDE / ADHD-200 / ABCD / multi-site"
• "leave-site-out cross-validation"
• "site-stratified split"
• "cross-site generalization"
• "IPD pooling across cohorts"

Do NOT trigger for:

• Single-site, single-scanner studies (no batch variable exists)
• Pure preprocessing requests (delegate to fmri-skill / smri-skill)
• Model training itself (delegate to model skills via run_models)

---

Position in the NeuroClaw Pipeline

[dataset-skill]  →  feature matrix + meta (site, scanner, age, sex, dx)
                                |
                                v
                    [harmonization-tool]   ← this skill
                                |
                                v
              harmonized feature matrix + same meta
                                |
                                v
                   [model-skill via run_models]