- Our work
- Corra
Corra: a symptom tracker that finds patterns in one person's own data
Corra is for people living with a chronic condition. They log symptoms, food, sleep, medication and more, and the app looks for patterns in that person's own history, such as a food, a short night or the weather that tends to come before a symptom. We built the Flutter app for iOS and Android, the Node.js API, the Python analysis pipeline and the admin panel.

- Industry
- Healthcare, chronic condition tracking
- Platforms
- iOS and Android (Flutter), web admin panel
- Our scope
- Mobile app, API, analysis pipeline, admin panel
- Flutter app delivered
- Within 3 months
What made it hard
Four constraints shaped the app, the API and the analysis.
Health data comes in many shapes
People log symptoms, pain, medication, food, sleep, vitals, cycle, journals and appointments. Each type has its own screen and its own data, and the analysis has to read all of them.
Wearable data arrives while the app is closed
Apple Health and Fitbit data is pulled in the background, using background isolates and scheduled work, so the logs fill in without the person opening the app.
Patterns belong to one person
A food that sets off symptoms for one person says nothing about the next. The pipeline fits a separate model for each person, so one person's data never feeds another person's results.
A suggestion has to be earned
Corra does not suggest anything until there is enough history to support it. A person can also dismiss a suggested pattern for good.
How does Corra find patterns in one person's data?
Every person's data goes through the same five steps, and each person gets a model of their own.
Hourly job
A scheduled job on the Node.js server runs every hour. It takes the users whose ID matches the hour (user ID modulo 24), so the work is spread across the day.
Collect
Each person's logs and weather data are written to S3 as a dataset for that person alone.
Model
A Python Lambda screens the factors with elastic-net regression, fits a logistic regression and drops weak factors one at a time. It then estimates each factor's effect by comparing predictions with and without it.
Store
A Node.js Lambda saves the results to the MySQL database on Amazon RDS.
Deliver
Results show up in the app and arrive as a push notification through Firebase Cloud Messaging. A separate endpoint renders a PDF report for a date range with Puppeteer, to share with a clinician.
What has to be true before Corra suggests anything
- At least 30 days since the first log, and at least 30 days with something logged.
- More than 10 days on which the symptom itself was logged.
- Each factor has to appear on at least 3 days to be considered.
What we built
The system has four parts. The app lets people log these twelve kinds of entries:
- Symptoms
- Pain by body part
- Medication
- Supplements
- Food and water
- Sleep
- Vitals
- Bowel movements
- Menstrual cycle
- Daily check-in
- Journals
- Appointments
Flutter app for iOS and Android
Logs for everything listed above, plus reminders for medication and supplements. People sign in with email, Google, Apple or Facebook. Apple Health and Fitbit data syncs in the background. The app has sleep charts, a guided tutorial, subscriptions through Qonversion, and five build flavors from one codebase.
Node.js API
Koa and TypeScript on MySQL. Every request carries a Firebase ID token that the API verifies, and request bodies are checked against Joi schemas. The code is layered as router, service, database and model, and the hourly job lives in the same codebase.
Analysis pipeline
A Python Lambda builds the model for one person at a time from their logs and weather data. A second Lambda stores the results. The same service renders the clinician PDF report.
React admin panel
A web panel behind a login with a dashboard and screens for users, offers and symptoms, so the Corra team can manage the data the app depends on.
Handling health data
What the system does today
Every API request is authenticated with a Firebase ID token. Each person's data is analysed on its own and never pooled with other users, and a person can dismiss a suggested pattern permanently.
We designed the output as a lead for the person and their clinician to talk about, not as a diagnosis or treatment advice. Running a health product in a regulated setting takes more than architecture, and our HIPAA architecture guide covers the mistakes we see most often.
Building a health or symptom tracking app?
Tell us what people will log and what you want to learn from it. We will tell you what is realistic in Flutter and what needs a backend. The client-facing write-up is in the original Corra case study.
