AI4AD: Predicting Mild Cognitive Impairment from Plasma Proteomics

McGill AI Lab · February 2026

A machine learning pipeline to classify Cognitively Normal (CN) vs Mild Cognitive Impairment (MCI) participants using plasma protein levels from the ADNI NULISA CNS panel. The goal is to evaluate whether blood-based biomarkers can improve early MCI screening beyond standard clinical assessments.

---

What This Project Does

We build and compare three models:

1. Clinical Baseline — logistic regression on age, sex, APOE ε4, and MMSE score
2. Proteomics Model — logistic regression on ~113 plasma protein features
3. Multimodal Model — clinical covariates + proteomics combined

Models are evaluated using stratified 5-fold cross-validation with AUC-ROC as the primary metric. We also plan to run a survival analysis (Cox model) on MCI-to-AD conversion and a SHAP-based interpretability analysis to identify the most biologically relevant proteins.

---

Project Structure

alzheimer-s-detection/
├── notebooks/              # Exploratory notebooks
├── plots/                  # Generated plots
├── reports/
│   └── figures/            # Final figures
├── scripts/
│   ├── models/             # Model scripts
│   ├── utils/              # Utility functions
│   ├── 00_inspect_columns.py
│   ├── 00_sanity_check.py
│   ├── 01_build_cohort.py
│   ├── 02_eda.py
│   └── 03_model.py
├── src/                    # Shared source modules
├── .gitignore
├── requirements.txt
└── README.md

Raw ADNI data goes in a data/raw/ folder locally — it is not committed to the repo.

---

Getting Started

git clone <repo-url>
cd alzheimers-detection
pip install -r requirements.txt

Place your raw ADNI CSVs in data/raw/, then run the scripts in order starting from 01_build_cohort.py.

---

Data accessed under the ADNI data use agreement. Raw data is not stored in this repository.