NIH Grant Activity Classifier

Multi-class classification of NIH award activity series (F/T/K/R) from structured NIH RePORTER signals.

Predict NIH grant activity series (F/T/K/R) from structured award signals (e.g., total cost and grant length) using NIH RePORTER data (FY2018–FY2023, 118k+ records).

Best baseline: Histogram-based Gradient Boosting
Validation metric: ~0.93 macro-F1

Why this matters

NIH awards include an Activity code (e.g., R01, K08, T32).
This project frames a multi-class classification problem to infer the activity series:

• F: Fellowship
• T: Training
• K: Career Development
• R: Research

Data

• Source: NIH RePORTER (public)
• Window: FY2018–FY2023
• Scale: 118,355 records (after filtering/alignment)
• Repo includes: sample CSVs in data/sample/ (full processed datasets excluded due to GitHub size limits)
• To reproduce full-scale training, download NIH RePORTER exports for FY2018–FY2023 and place the raw files under data/raw/ (ignored by git), then run the notebook/pipeline steps.

Baseline results

Best baseline: Histogram-based Gradient Boosting (~0.93 macro-F1 on validation)
Benchmarked against: Decision Tree, MLP, Logistic Regression.

---

Project artifacts

Notebook: notebooks/NIH_Grant_Activity_Classifier.ipynb
Slides: slides/NIH_Grant_Activity_Classifier_Slides.pdf
Project report: reports/report.md

Quickstart

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Open and run: NIH_Grant_Activity_Classifier.ipynb

---

Repository structure

• notebooks/ analysis + training notebook
• src/ reusable utilities (data cleaning / feature helpers)
• data/sample/ small sample dataset (tracked)
• data/raw/ full dataset (NOT tracked)
• reports/ write-up
• slides/ presentation slides

---

Roadmap

• Add stronger features (e.g., IC/Institute, FY, direct vs indirect costs)
• Evaluate with macro-F1; handle class imbalance
• Add explainability (feature importance / SHAP)
• Ship a minimal demo (Streamlit) and/or API (FastAPI)

Full dataset note

This repository includes sample CSVs under data/sample/. Full processed datasets are excluded due to GitHub file size limits.
Full-data reproduction can be done by downloading NIH RePORTER exports (FY2018–FY2023) and running the notebook to regenerate splits locally.

License

MIT