Machine Learning Challenge

At Bio Conscious, we build machine learning tools to help people.  That mission requires constant refinement, experimentation, and testing of new algorithms, and a singular focus on beating our current production model.  

Do you want to try?

Instructions

In this challenge, you will try to create a blood sugar estimation algorithm which functions like the Diabits app.  You will be provided with simulated data, and will train, validate, and test an algorithm which can predict a trend for blood sugar 60 minutes into the future.  Your objective is to make an algorithm which is able to predict blood sugar as accurately as possible.

1. Learn about the Diabits app

We recommend you learn more about what the Diabits app does, before attempting the project.  Learn more at www.diabits.com.  It may also help to learn a little about diabetes in general.

 

2. Download the file included below, and review the contents of the three spreadsheets.  

In this file you will find blood glucose, activity, and heart rate data.  The blood glucose data contains a blood glucose measurement in mg/dL, and a timestamp.  The heart rate data contains a value, and a timestamp. The distance activity data contains a value, a timestamp, and the name of the device which made the measurement.  

 

3. Using this data as your inputs, build a machine learning algorithm which predicts blood sugar.  

You may use any machine learning method you like.  

The output of your model should be a predicted blood sugar trend, for 60 minutes into the future, with one predicted point every 5 minutes.  The prediction should, of course, only take into account past data. When you give it a new blood sugar value, it should generate a new predicted trend, taking into account whatever secondary inputs you choose to use.

When building the model, we recommend using no more that 2 weeks of data for training, less than a week for validation and 5 weeks for testing. 

4. Test your algorithm, focus on prediction accuracy. 

We will assess your accuracy with RMSE, MARD, and A + B region accuracy on the Parkes Error Grid, which is an iteration of the Clarke Error Grid.  

 

5. When you’re finished, summarize your results and submit your project.

Materials

Download this file, it contains everything you need. 

In order to download, you must be signed into a Google account, and hit the "request access" button on the link.

If you do not have a google account, please email us at info@bioconscious.tech to be sent a copy.

 

Submission

Please send your completed algorithm, and a summary of your results, to info@bioconscious.tech with the subject line “Machine Learning Challenge - YOUR NAME.”

Attach your algorithm as a .zip file, and your results summary as a PDF.  

The best submissions will come with clear, concise results summaries and competent data visualization.