cat-cog

Cognitive assessment and visual scoring powered by LLMs.

cat-cog provides LLM-powered evaluation of hand-drawn images for neuropsychological testing. It builds on cat-stack, the shared classification engine for the CatLLM ecosystem.

Installation

pip install cat-cog

This automatically installs cat-stack (the LLM classification engine).

CERAD Constructional Praxis Scoring

Score hand-drawn shapes (circle, diamond, overlapping rectangles, cube) using LLM vision models. The function sends images to the LLM, classifies drawing features, then applies CERAD scoring rules.

from catcog import cerad_drawn_score

# Score circle drawings (max score: 2)
results = cerad_drawn_score(
    shape="circle",
    image_input="./circle_drawings/",
    api_key=OPENAI_KEY,
    user_model="gpt-4o",
)
print(results[["image_file", "score"]])

Supported shapes

Using with multiple models

All catstack.classify() parameters are available via **kwargs:

results = cerad_drawn_score(
    shape="cube",
    image_input="./cube_drawings/",
    api_key=OPENAI_KEY,
    models=[
        ("gpt-4o", "openai", OPENAI_KEY),
        ("claude-sonnet-4-20250514", "anthropic", ANTHROPIC_KEY),
    ],
)

Future

• Fine-tuned vision models for shape quality assessment (circle classifier, etc.)
• Additional cognitive screening instruments (clock drawing tests, etc.)

Ecosystem

License

GPL-3.0-or-later