AstroWatch

• 

  AstroWatch!

• 

  Notifies you whenever your vitals are irregular and you should check the app.

• 

  The default homepage. Easy access to your respiratory rate and heart rate-- two important vitals to be aware of, especially as an astronaut.

• 

  Cardio Recommendations

• 

  Cognitive Function Recommendations

• 

  Stress Recommendations

Inspiration

Astronauts experience cognitive degradation, stress, and physical deconditioning during missions, but NASA still lacks a personalized way to detect changes before they become dangerous.

What it does

AstroWatch takes vitals from the Apple Watch, runs them through a predictive model, and instantly flags early warning signs in cardio, cognitive, and stress decline. Then we pass that output through an LLM to contextualize it in plain language that astronauts can understand and act on.

How we built it

We trained a Random Forest model on 5,000 biologically-correlated samples from Kaggle using an 80/20 split. We ran our backend on an ARM-based AWS Graviton EC2 instance for cheap, fast inference. We pull the latest Apple Watch logs (since real-time streams aren’t provided) and feed them into our model to predict the astronaut’s health scores, then contextualize those predictions with a Mistral-7B-Instruct LLM. On the frontend, we built dashboards to visualize cardio, cognitive, and stress indicators in a clean way.

Challenges we ran into

Apple Watch data isn’t real-time, so we had to engineer around limited sampling frequency. We also had to learn ARM-based cloud hosting, model serving, and finding the right dataset with the variables we needed. It was also difficult to get the EC2 server to communicate with SwiftUI as the Apple ecosystem is very closed off.

Accomplishments that we're proud of

We trained a predictive model on 5,000 samples and got it running efficiently on an ARM-based AWS backend. We also integrated an LLM to turn raw model numbers into easy-to-understand insights. And we built a clean interface that lets people actually see and interact with those predictions in a simple, visual way.

What we learned

ARM compute can be insanely cost-efficient. Random Forest is great for structured vitals. And LLMs become way more useful when you constrain them and give them tight boundaries.

What's next for AstroWatch

We want to keep adding more health signals so the predictions get even more accurate. We also want to test this with real Apple Watch users over time to see how well it holds up. And eventually, we’d like to make this useful not just for astronauts, but for everyday people under stress here on Earth too.

Built With

• arm
• ec2
• gravitron
• healthkit
• mistral-7b-instruct
• swift