NeuroDetect: Early Alzheimer’s Risk & Interpretability Model

This project began with my curiosity about how machine learning can help detect diseases like Alzheimer’s early, especially when symptoms are subtle and often missed. I wanted to understand whether simple computational tools—accessible even to students—could uncover early risk patterns in a transparent way. Since most biomedical datasets are restricted or hard to use, I set out to build a model that anyone could run and understand.

What Inspired Me

Alzheimer’s research usually needs advanced labs, expensive imaging, or genetics data. As a student without access to those resources, I wondered if there was a way to still contribute to this field. That idea—making computational health accessible—became the core inspiration. I wanted to show that meaningful insights don’t require huge models or complex tools, just thoughtful design and interpretability.

What I Learned

I learned how important it is to balance accuracy with explainability. In healthcare, predictions mean little unless we can understand what drives them. Working with multimodal features also taught me how different signals—like cognitive scores, biomarker-like values, and speech features—can complement each other. I also gained confidence in building reproducible notebooks and presenting results clearly.

How I Built the Project

Since real datasets often require permissions, I created a synthetic multimodal dataset modeled after real patterns from OASIS-3 and DementiaBank. I trained a Random Forest classifier because it is reliable, runs easily on any device, and supports interpretability. I evaluated the model using AUC and accuracy, and then used permutation feature importance to find which features influenced predictions the most. Finally, I organized the entire workflow into a clean, reproducible notebook and a structured report.

Challenges I Faced

One challenge was designing synthetic features that still reflect realistic Alzheimer’s patterns. Another was keeping the model simple enough for beginners while still meaningful for evaluation. The biggest challenge was ensuring interpretability—making sure the model not only predicts but also explains why. Throughout the process, I learned how to communicate results in a way that is both understandable and scientifically grounded.

Built With

• google-colab
• python