Perfect — here’s a clean, polished version you can paste directly into a Datathon submission form.

Inspiration

Stress is one of the most measurable yet least understood drivers of long-term health. Cortisol follows a daily rhythm that reflects sleep quality, recovery, and physiological load — but most people don’t have access to continuous cortisol measurements.

We were inspired by the question:

Can we estimate meaningful cortisol trends using consumer wearable data like heart rate, steps, and sleep?

CortisolTracker explores whether everyday Fitbit data can be transformed into a real-time physiological signal that helps people understand their stress patterns.

What It Does

CortisolTracker predicts and visualizes an individual’s estimated cortisol levels using wearable data.

Specifically, it:

Uses heart rate intensity, steps, and sleep metrics
Feeds engineered features into a trained Ridge regression model
Predicts wake cortisol (normalized and converted to µg/dL)
Models a realistic daily decay curve
Computes:
- Current cortisol estimate
- Average cortisol for today
- Peak cortisol
- Current vs national average
- Whether the user is within a time-adjusted normal range
Visualizes results with an interactive gauge and dashboard

It turns raw wearable data into an interpretable physiological stress profile.

How We Built It

Data

MMASH dataset (multi-modal wearable + salivary cortisol data)
22 users
Actigraph HR
Steps
Sleep metrics
Saliva-based wake cortisol

Feature Engineering

For each day:

Avg heart rate intensity (z-score vs baseline)
Previous-day steps
Previous-night sleep duration
Sleep efficiency

Model

Ridge regression
Trained on normalized wake cortisol
MAE ≈ 0.076 (normalized scale)
Converted to µg/dL using assumed population parameters

Backend

Python model inference via joblib
Next.js API route for prediction
Parallel Fitbit data fetching
Real-time curve modeling

Frontend

React + Next.js
Dynamic dashboard
Custom-built SVG cortisol gauge
Real-time percent-vs-average visualization

Challenges We Ran Into

Data sparsity: Only 16 valid training rows after filtering.
Small dataset modeling: Risk of overfitting with limited samples.
Time alignment: Matching previous-day wearable features to next-morning cortisol.
Real-time inference speed: Spawning Python processes inside a Next.js API route required optimization.
Visualization math: Building a properly oriented semi-circular gauge required careful SVG arc geometry.

We also had to make decisions about assumed national averages and daily decay modeling in the absence of continuous cortisol data.

Accomplishments We're Proud Of

Successfully trained a working physiological prediction model.
Built a full-stack pipeline from wearable data → ML inference → real-time UI.
Designed a biologically realistic cortisol decay curve.
Created a dynamic gauge reflecting current vs time-adjusted average.
Optimized backend performance to reduce 40s inference times to near-real-time.

Most importantly, we translated complex physiological modeling into an intuitive consumer-facing dashboard.

What We Learned

Wearable signals can approximate physiological markers when engineered thoughtfully.
Even simple models (Ridge regression) can perform well with meaningful features.
Visualization clarity is just as important as model accuracy.
Full-stack ML integration introduces engineering challenges beyond training the model.
Small datasets require careful normalization and interpretation.

We also learned how to move from static ML experiments to real-time production-style deployment.