Imported from CUHK-AIM-Group/NeuroDiscovery (skills/cpm/SKILL.md). Install upstream with npx skills add CUHK-AIM-Group/NeuroDiscovery --skill cpm. Copyright stays with the author (MIT).

CPM Skill

Overview

cpm is the canonical NeuroClaw implementation of Connectome Predictive Modeling. Edge selection is repeated independently inside every training fold.

Installation

pip install numpy pandas scipy scikit-learn joblib

Workflows

1. Prepare data

connectomes.npz:

X:          [subjects, nodes, nodes] or [subjects, edges]
subject_id: [subjects]

labels.csv contains the same subject IDs and a target column.

2. Regression

python skills/cpm/scripts/train_reference.py \
  --connectomes connectomes.npz \
  --labels labels.csv \
  --target cognitive_score \
  --subject-col subject_id \
  --task regression \
  --p-threshold 0.01 \
  --folds 5 \
  --output-dir run_models_output/cpm

3. Classification

python skills/cpm/scripts/train_reference.py \
  --connectomes connectomes.npz \
  --labels labels.csv \
  --target diagnosis \
  --task classification \
  --p-threshold 0.01 \
  --output-dir run_models_output/cpm_classification

If p-threshold is tuned, use nested validation or training-only selection.

Input / Output Summary

Testing

pytest models/tests/test_extended_models.py -q
python skills/cpm/scripts/train_reference.py --help

Directory Reference

models/cpm/
├── cpm.py              fold-local CPM estimator
└── train.py            cross-validated CLI

skills/cpm/
├── SKILL.md
└── scripts/train_reference.py

Reference

• Finn et al. functional connectome fingerprinting and connectome-based prediction framework, Nature Neuroscience (2015).

Created At: 2026-07-29 HKT Last Updated At: 2026-07-29 HKT Author: chengwang96