CPM Skill

Overview

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

Task	Input	Output
Classification	subject FC matrices/vectors	class and probability
Regression	subject FC matrices/vectors	continuous prediction

---

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

Item	Format
Connectomes	.npz with X, subject_id
Labels	CSV keyed by subject ID
Predictions	predictions.csv
Fold membership	fold_assignments.csv
Metrics	metrics.json
Fold models	checkpoint.joblib
Provenance	config.json, run_manifest.json

---

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