MemoryMate - Early Dementia Detection Through AI-Powered Conversations

💡 Inspiration

Watching a teammate's grandmother struggle with undiagnosed dementia was heartbreaking. Her family noticed small changes—forgetting words, repeating questions—but didn't know these were warning signs until it was too late. By the time she was diagnosed, critical intervention windows had closed.

We learned that early detection can improve dementia outcomes by up to 40%, yet seniors rarely get screened. Current methods require expensive clinical visits, and by the time symptoms are obvious to families, years of subtle decline have already occurred.

The gap: Over 55 million people worldwide live with dementia, but existing AI solutions either provide unexplainable "black box" scores or focus only on text analysis. None give families the transparent, actionable insights they need to advocate for their loved ones.

We built MemoryMate to bridge this gap—an AI companion that detects cognitive decline early through natural daily conversations and gives families explainable, clinically-meaningful insights.

🎯 What It Does

MemoryMate monitors cognitive health through simple daily check-ins via SMS or voice calls. Unlike existing solutions, it provides:

Multi-Modal Analysis

Analyzes both what seniors say and how they say it
Tracks voice patterns (hesitation, word-finding pauses, speech rate)
Monitors text patterns (grammar, vocabulary complexity, repetition)
Detects behavioral changes (response times, engagement levels)

Explainable AI

Instead of cryptic scores, families see specific evidence:

"Used generic words like 'thing' instead of nouns 4 times this week (baseline: 0.5x/week)"
"Average pause before answering increased from 1.2s to 3.8s"
Direct links to concerning conversation snippets

Clinical-Grade Output

Doctor-ready PDF reports with decline trajectory graphs
Multi-dimensional scoring: memory, language fluency, executive function
Exportable data formatted for healthcare providers
Predictive modeling: "At current rate, clinical symptoms likely in 8-12 months"

Actionable Family Support

Plain-language explanations of what changes mean
Specific next steps: when to schedule appointments, what to ask doctors
Local resource finder for memory clinics and specialists
Educational content about dementia stages

🛠️ How We Built It

Architecture

Backend: Node.js + Express REST API
Database: PostgreSQL with time-series optimization for longitudinal tracking
Frontend: React dashboard with Recharts for data visualization
Communication: Twilio API for SMS/voice calls
AI Analysis: Anthropic's Claude API for conversation analysis
Voice Processing: Deepgram for speech-to-text with diarization
Report Generation: Puppeteer for PDF export

Technical Implementation

1. Conversation Engine

// Daily automated check-ins with contextual follow-ups
const questions = generateContextualQuestions(userHistory);
await twilioClient.messages.create({
  body: questions[0],
  to: seniorPhoneNumber,
  from: twilioNumber
});

2. Cognitive Marker Detection We implemented multi-dimensional analysis:

Lexical Diversity: $\text{TTR} = \frac{\text{unique words}}{\text{total words}}$ (Type-Token Ratio)
Semantic Coherence: Embedding similarity between responses
Temporal Patterns: Response latency, time-of-day engagement

3. Baseline Comparison

// Compare current metrics to rolling 30-day baseline
const decline_score = (current_metric - baseline_mean) / baseline_std;
if (decline_score < -2.0) {
  generateExplainableAlert(conversation_id, metric_type);
}

4. Explainable Alerts We built a custom evidence extraction system:

Flags specific responses that deviate from baseline
Highlights linguistic patterns (word substitutions, repetitions)
Generates human-readable explanations using Claude API

5. Predictive Modeling Linear regression on time-series data: $$\hat{y}_t = \beta_0 + \beta_1 t + \epsilon_t$$ Where $y_t$ represents cognitive score at time $t$, projecting future trajectory.

🧗 Challenges We Faced

1. Balancing Sensitivity vs. False Positives

Challenge: Too sensitive = anxious families, too lenient = missed cases
Solution: Implemented confidence thresholds and trend-based alerts (requires 3+ weeks of decline)

2. Privacy & Ethics

Challenge: Handling sensitive health data, avoiding surveillance concerns
Solution: End-to-end encryption, explicit consent flows, data deletion options, transparent AI explanations

3. Natural Conversation Flow

Challenge: Making daily check-ins feel like friendly chat, not interrogation
Solution: Claude API with memory of past conversations, contextual questions ("How's your garden doing?"), varied question patterns

4. Explainability Without Overwhelming Users

Challenge: Technical metrics mean nothing to families
Solution: Two-tier reporting: simple alerts for families, detailed reports for doctors

5. Time Constraints

Challenge: Building multi-modal analysis in 12 hours
Solution: Started with SMS (simpler), voice analysis as stretch goal, focused on MVP with real clinical value

🎓 What We Learned

Technical Skills

Time-series database optimization for longitudinal health tracking
Prompt engineering for healthcare-grade AI analysis
Building explainable AI systems (not just black-box predictions)
Real-time voice processing pipelines

Domain Knowledge

Clinical markers of early dementia (from research papers)
Ethical considerations in health monitoring AI
The difference between correlation and clinical significance
How families actually interact with aging parents (user research)

Hackathon Lessons

Scope ruthlessly: We cut video analysis and focused on SMS/voice
Demo-driven development: Built the most compelling features first
Real problem validation: Talked to 3 families during the event—all said they'd use this

🏆 What's Next for MemoryMate

Immediate Improvements

Integration with wearable devices (sleep, activity patterns)
Multi-language support for immigrant families
Caregiver burnout detection (monitor the family too)

Clinical Validation

Partner with memory clinics for pilot study
Validate against Montreal Cognitive Assessment (MoCA)
Collect IRB approval for research dataset

Long-Term Vision

FDA clearance as a medical device
Insurance reimbursement for remote monitoring
Open-source the explainability framework for other health AI tools
Expand to other neurodegenerative conditions (Parkinson's, ALS)

📊 Impact Potential

55 million people worldwide living with dementia
63 million family caregivers in the US who need support
$3.4 billion lost annually to financial scams targeting seniors with cognitive decline
40% improvement in outcomes with early detection

MemoryMate isn't just a monitoring tool—it's a lifeline for families navigating one of the hardest journeys they'll ever face.