MariUncs

LINK TO GOOGLE SLIDES https://docs.google.com/presentation/d/1E8xqnLaYQdgBe3wTLMobkt2ECZcZ2QtZ-HUCLj7xh3s/edit?usp=sharing

Inspiration

Our inspiration came from the growing need for early detection of neurodegenerative diseases such as Alzheimer’s Disease (AD) and Frontotemporal Dementia (FTD). These conditions are often difficult to diagnose early because symptoms can overlap with normal aging or other neurological disorders.

Electroencephalography (EEG) provides a non-invasive and relatively inexpensive way to monitor brain activity, making it a promising tool for early detection.

We wanted to explore whether machine learning and deep learning could identify subtle patterns in EEG signals that differentiate healthy individuals from patients with dementia.

What it does

MariUncs is a deep learning system that analyzes EEG brain signals to classify neurological conditions.

The system:

Processes EEG recordings from 19 brain electrodes Splits the signals into 30-second windows Uses a Convolutional Neural Network (CNN) to detect patterns in the signals Aggregates predictions across windows to produce a subject-level diagnosis

The model can classify:

Alzheimer’s Disease (AD) Cognitively Normal (CN)

It also performs targeted binary tasks:

AD vs CN

This approach helps evaluate how well EEG signals can be used for automated neurological screening.

How we built it

We trained the model using:

PyTorch CrossEntropyLoss Adam optimizer

Predictions were aggregated across windows to produce subject-level predictions.

Challenges we ran into

One of the biggest challenges was working with EEG data, which is noisy and high-dimensional.

Key challenges included:

Handling long time-series recordings Converting EEG signals into meaningful input windows Ensuring correct label types for multiclass classification Avoiding overfitting due to the small dataset Implementing subject-level prediction instead of window-level prediction

Another challenge was debugging deep learning issues such as:

datatype mismatches in PyTorch model convergence validation instability due to small sample size.

Accomplishments that we're proud of

We are proud that we successfully:

Built a full deep learning pipeline from raw EEG signals Implemented multiclass classification (AD vs CN) Implemented binary diagnostic tasks (AD vs CN) Achieved meaningful classification performance using only EEG data Designed a model capable of subject-level neurological prediction

We are also proud that our solution demonstrates the potential of ML-assisted neurological diagnosis.

What we learned

Through this project we learned:

How to process biomedical EEG signals How to build deep learning architectures for time-series data The importance of data preprocessing and window segmentation How to evaluate models using subject-level metrics How to debug real-world machine learning pipelines

We also learned how challenging and rewarding it is to apply machine learning to healthcare data.