DUAL-MODEL CLINICAL SUPPORT SYSTEM FOR ALZHEIMER'S & DEMENENTIA DETECTION

Submission for Track 2: Machine Learning / AI

Inspiration

As of 2025, over 7.2 million Americans are living with Alzheimer's. In recent years, early-onset dementia diagnoses spiked by 373% for Millennials & Gen X. Yet, dementia diagnoses currently takes on average 3.5 years from symptom onset. MRIs and PET scans are the standard for Alzheimer’s & dementia detection, but they are expensive and inaccessible. The Montreal Cognitive Assessment (MoCA) is reliable, low-cost, and widely used, but the cognitive insights are usually buried in isolated records. I built NeuralTrack to bridge the gap between accessibility and expensive imaging by using low-cost screening assessments to provide the long-term insights.

What it does

NeuralTrack is a clinical dashboard that automates the conversion of MoCA and subdomain scores into Clinical Dementia Rating (CDR) stages. It uses two XGBoost models to process scores: one identifies the patient's current stage (CDR staging) and the other predicts their status in 12 months. By analyzing longitudinal data (assessments taken over several years) it helps healthcare providers see the subtlest warning signs and distinguish between healthy aging and decline.

How we built it

The platform is built with the MERN stack (MongoDB, Express, React, Node) integrated with machine learning using XGBoost. The models are trained and validated on the Washington University's OASIS-3 dataset, which contains 30 years of longitudinal data from 1,300+ participants. The models focus on the rate of decline between visits and interaction features, how variables like age and specific memory drops influence each other, so it analyzes a patient's clinical history rather than just a single assessment.

Challenges we encountered

The main technical hurdle was preventing data leakage because the dataset had multiple entries per person, I used GroupKFold cross validation to ensure the models were tested on new patients it hasn't seen before so that it is representative of new clinical cases.

Accomplishments that we're proud of

Using just MoCA assessments and a patient age, the models reached 83.2% accuracy in predicting the patient's current CDR stage and 82.5% accuracy for 12-month cognitive projections.

What we learned

Building the predictive model was only half of the challenge. The other half was making it actually work for healthcare providers. I realized that clinicians are already overwhelmed with raw data. They do not need more numbers, they need a clear way to see who to help first. I spent a lot of time refining the UI to make the data easy to read and the web app intuitive to navigate.

What's next for NeuralTrack

The goal is to move the prototype into a production-ready platform. This includes refining our models to improve predictive accuracy, adding support for more types of cognitive tests, and building an API service so other healthcare developers can utilize our models and automative prediction into their own clinical systems.

AI Usage Statement

Gen AI was used to accelerate the development of NeuralTrack, by providing coding assistance, XGBoost script optimizations, and to refine the documentation. OpenAI was used for image generation.

Built With

• machine-learning
• mongodb
• python
• react