Predicting the Onset of Dementia Using Neural Networks

A Deep Learning Approach to Neuroimaging and Multimodal Analysis

Research: https://github.com/abrizu/480-NN-Project-MRI/blob/mri/docs/Predicting_the_Onset_of_Dementia_Using_Neural_Networks.pdf

---

Overview

DementiaNet is a multimodal deep-learning system that predicts Alzheimer's Disease / Dementia risk across three independent data modalities, each backed by its own trained neural network:

Modality	Model Type	Classes / Output
MRI	CNN (_MRI_CNN)	Non-Demented · Very Mild · Mild · Moderate
Biomarkers	Feedforward NN (_Biomarker_NN)	No Dementia · Dementia (probability)
Speech	(in progress)	Low Risk · High Risk

Results from individual models are fused via a weighted ensemble to produce a combined dementia risk score.

---

Features

• Full in-app training — no terminal required. Train the MRI and Biomarker models directly from the Streamlit UI with configurable epochs and learning rate.
• Named checkpoint management — save multiple trained weights with custom names and notes. Switch between checkpoints at any time without restarting the app.
• Checkpoint registry — a checkpoints/registry.json file tracks every saved model's accuracy, hyperparameters, date, and notes, persisting across sessions.
• Demo-ready loading — on demo day, open the app, go to the Model Manager, and load your best checkpoint in one click.
• Weighted ensemble — the Ensemble tab combines MRI (45%), Biomarker (25%), and Speech (30%) predictions into a single risk gauge.
• Live training feedback — per-epoch loss and validation accuracy stream into the UI in real time during training.

---

Project Structure

480-NN-Project-MRI/
├── app.py                        # Streamlit application (main entry point)
├── run.sh                        # One-command launcher (venv + deps + app)
├── requirements.txt              # Python dependencies
│
├── functions/
│   ├── mri.py                    # MRI CNN model definition and training script
│   ├── biomarker.py              # Biomarker NN model definition and training script
│   ├── speech.py                 # Speech model (in progress)
│   ├── multimodal.py             # Ensemble / multimodal utilities
│   └── main.py                   # CLI entry point
│
├── checkpoints/
│   ├── registry.json             # Named checkpoint registry (auto-generated)
│   ├── mri_best.pt               # Active MRI weights (auto-generated)
│   └── biomarker_best.pt         # Active Biomarker weights (auto-generated)
│
└── Data/
    ├── MRIData/
    │   └── Data/
    │       ├── Mild Dementia/
    │       ├── Very mild Dementia/
    │       ├── Moderate Dementia/
    │       └── Non Demented/
    └── BMData/
        └── main.csv

---

Setup

1. Clone the repository

git clone https://github.com/abrizu/480-NN-Project-MRI.git
cd 480-NN-Project-MRI

2. Download datasets

Datasets are not included due to size. Download and extract them manually:

Dataset	Source	Extract to
MRI Images (4-class)	Kaggle – Alzheimer MRI Dataset	Data/MRIData/Data/
Biomarker CSV	Kaggle – Comprehensive Health & Brain Imaging	Data/BMData/main.csv
Speech Dataset	(insert link)	Data/SPData/

After extraction, your Data/MRIData/Data/ folder should contain exactly four class subfolders:
Mild Dementia/, Very mild Dementia/, Moderate Dementia/, Non Demented/

3. Launch the app

bash run.sh

run.sh automatically:

• Detects your Python version (3.9+ required)
• Creates a .venv virtual environment if one doesn't exist
• Installs all dependencies
• Starts the Streamlit app at http://localhost:8501

---

Using the App

Training a model

1. Open http://localhost:8501 in your browser.
2. Navigate to the MRI or Biomarkers tab.
3. Expand the ⚙️ Model Manager panel.
4. Click the 🏋️ Train New Model sub-tab.
5. Adjust epochs and learning rate using the sliders.
6. Click 🚀 Start Training — live loss/accuracy updates stream into the UI.
7. When training completes, enter a name and optional notes for the checkpoint (e.g. demo_run, high_lr_v2).
8. Click 💾 Save & Activate — the checkpoint is saved, registered, and immediately loaded.

Loading a saved checkpoint (e.g. on demo day)

1. Open the ⚙️ Model Manager panel in the relevant tab.
2. Click 📁 Saved Checkpoints.
3. Select the checkpoint from the dropdown and click ⬆️ Load & Activate.

Running inference

• MRI: Upload a .jpg / .png brain MRI scan → click Analyze MRI → view class probabilities.
• Biomarkers: Fill in the patient form (age, comorbidities, imaging scores) → click Analyze Biomarkers → view the risk gauge.
• Ensemble: After running at least one modality, visit the 📊 Ensemble tab for a combined weighted risk score.

---

Training from the CLI (optional)

# MRI model
python functions/mri.py

# Biomarker model
python functions/biomarker.py

Manually save weights to checkpoints/mri_best.pt or checkpoints/biomarker_best.pt after CLI training to use them in the app.

---

Dependencies

All installed automatically by run.sh:

torch · torchvision · torchaudio
numpy · pandas · scikit-learn · scipy · matplotlib
streamlit · plotly
librosa · soundfile
Pillow · nibabel · kaggle · tqdm · requests

---

Disclaimer

⚠️ Research use only.
DementiaNet is an academic research project and is not a validated clinical diagnostic instrument. Do not use it for actual medical decision-making.