HealthPulse AI

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for HealthPulse AI

Inspiration

Healthcare is increasingly data-driven, yet most people track their health metrics across disconnected apps with no unified analysis. We wanted to build a tool that takes raw health data and turns it into actionable intelligence — without requiring cloud services, API keys, or compromising privacy.

What it does

HealthPulse AI analyzes 11 vital health metrics (heart rate, blood pressure, sleep, steps, SpO2, stress, and more) using an ensemble ML model. It provides:

• Risk Assessment: A soft-voting ensemble combining Random Forest and Logistic Regression produces calibrated risk scores across Low, Moderate, and High categories
• Interactive Dashboards: Six Plotly visualization tabs — trend lines with healthy range bands, correlation heatmaps, animated risk gauges, distribution charts, weekly comparisons, and multi-metric overlays
• Personalized Recommendations: Context-aware health advice based on current values, 7-day trends, and deviations from clinically-informed healthy ranges
• Sleep Quality Analysis: Composite scoring of duration adequacy, schedule consistency, and optimal range percentage
• Health Calculators: BMI with WHO categories and Mifflin-St Jeor calorie estimation for 5 activity levels
• PDF Reports: Professional multi-page reports with color-coded risk badges, per-metric analysis tables, and numbered recommendations

How we built it

The backend uses a VotingClassifier ensemble (Random Forest at 60% weight + Logistic Regression at 40%) trained on 2,000 synthetic population samples at runtime. Feature extraction computes rolling averages, standard deviations, and trend indicators from the most recent 7 days. Risk scoring combines per-metric deviation analysis (40%) with ML probability outputs (60%).

The frontend is built with Streamlit, using Plotly for all interactive charts. PDF generation uses fpdf2 with custom headers, footers, and formatting. A synthetic data generator with weekly seasonality patterns lets users try the app immediately without uploading any data.

Challenges we ran into

Balancing model complexity with runtime performance — the ensemble trains on every request since there's no persistent storage. Getting clinically meaningful healthy ranges for all 11 metrics required research across multiple medical sources. Making the PDF reports visually professional with fpdf2's limited layout engine was also non-trivial.

What we learned

Building ML-powered health tools that run entirely offline is very feasible. Ensemble methods provide more robust predictions than single classifiers for health risk assessment. Privacy-first design (zero cloud, zero APIs) doesn't have to mean compromised functionality.

What's next

Adding model caching for better performance, supporting wearable device data import (Apple Health, Fitbit CSV), and expanding to more health metrics like HRV and glucose monitoring.

Built With

• fpdf2
• plotly
• python
• scikit-learn
• streamlit