Doctor's dashboard view of a patient. We assess 6 sectors of the brain with tasks classified among short/long term recall, and cognition
Our overall architecture
Doctor dashboard for querying with natural language. This allows them to quickly find at-risk patients, compare patients, and do even more!
Patient facing video call. The user can practice different recall exercises in a natural, conversational manner as the model learns them

Mind Mate

An AI companion that builds a unique patient baseline to catch mental degradation early, providing actionable risk assessment to doctors long before symptoms become obvious.

🧠 Inspiration: The 80% Problem

We all know Alzheimer's is a devastating problem, not just because there is no cure, but because of its timelines. You can slow and treat the disease, but the critical window for treatment is just two years after symptoms first appear.

Here is the statistic that motivated our project:

80% of people with Alzheimer's miss this window and miss out on effective treatment altogether.

Evidently, even our best models, MRIs, and doctors cannot overcome this time problem.

Yearly check-ups are simply not frequent enough and lack a deep, personal baseline.
Insurance often covers expensive MRIs only after symptoms have progressed to a noticeable amount. By then, it's too late.

✨ What It Does

We solve this problem by offering a platform that does two things:

For Patients: It provides pleasant memories, engaging conversations, and frequent, low-stress check-ins.
For Doctors: It learns a patient's unique cognitive baseline and provides actionable, explorable insights and a clear risk assessment.

Our system builds a rich collection of a patient's history by assessing:

Irregular speech patterns
Memory recall during conversations
Loss of interest or distractibility
Analysis of brain scans (when available)

This allows us to catch the subtle, early-stage degradation that doctors, and even family members, might miss.

🛠️ How We Built It: An AI-Centric Pipeline

The core of our project is an AI-centric pipeline that a doctor can query with natural language.

A doctor doesn't have to dig through charts. They can simply ask:

"Show me any changes in Patient X's speech patterns or memory recall over the last 30 days."

To achieve this, we integrated:

Intelligent Chain-of-Thought (CoT) Reasoning: To allow the AI to "think" step-by-step and provide clear, auditable risk assessments.
Patient Data Caching & Semantic Search: Securely builds a vector database of all patient interactions, conversations, and scan data, allowing for instant and intelligent querying.
Multi-Modal Analysis: Our models analyze not just text, but also audio for irregular speech and image data from scans.
Risk Assessment Module: A final model that synthesizes all data points into a single, easy-to-understand risk score for the provider.

🏃 Challenges We Ran Into

Our biggest challenge was effectively coordinating among different team members each with different data sources, models, systems, and model training/api calls. We solved this by implementing a vector database with semantic search. When a patient interacts with the AI, we generate embeddings for that conversation and store them. When a doctor queries the system, we first find the most relevant pieces of data using semantic search and run our intensive AI reasoning models on that small, relevant subset. This along with model switching, chain-of-thought reasoning, risk analysis, and short-term memory caching made the model extremely cost-effective, intelligent, and well-balanced for different tasks.

🚀 What's Next

Real-time Alerts: Developing a system to proactively alert doctors if a patient's risk score crosses a critical threshold, rather than waiting for the doctor to query.
Family/Caregiver Portal: Creating a separate view for approved family members to provide their own observations ("Mom seemed confused today") and receive non-clinical updates.
HIPAA Compliance: Integrating with compliant databases and APIs to prepare the platform for real-world pilot testing.

🏆 Accomplishments That We're Proud Of

Successfully integrating four different AI models into a single, cohesive pipeline.
Creating a user interface that is simple and calming for elderly patients.
Developing a natural language query system for doctors that actually works.