A diagnosis of diabetes does not come lightly; it forces a drastic change in lifestyle and leads to the incredibly terrifying feeling of having no control. At any point in time, while studying, gardening, eating or even sleeping, a sudden spike in blood sugar levels could leave you comatose. This constant worry of being blindsided, and by your own body at that, inspired us to develop an app that would give diabetics back the steering wheel to their lives.
What it does
It actively reads a patients data and predicts with 91% accuracy what their blood glucose level will be in 20 minutes. This allows patients to react and prevent any further damage.
How we built it
So we initially had 100 data points to work with to build a neural network that would make accurate predictions, however we needed more data to process, and managed to convince the Johnson and Johnson sponsors to provide us with 11000 data points. Then we implemented a creative algorithm that used machine learning to find the most effective configuration for the predictive machine learning - "machine-ception" if you will. A neural network that configured itself, all written from scratch!
Challenges we ran into
Initially we were only given a very limited amount of data, that led to a highly inconsistent algorithm. It would have wild swings from 72% accuracy down all the way to 40%. However, the real struggle came once we actually obtained an abundance of data. In order to process the 11000 data points, we had to find out how to run our machine learning on a multi-threaded Linode server. It took a great deal of configuration to make our algorithm compatible with the sever, and even more config to get the multi-threading to work.
Accomplishments that we're proud of
Working with a team with a very diversified skillset, we achieved a predictive accuracy of 91%! Every member of the team had an opportunity to not only apply what they already knew, but grow into the necessary skills by the end of the hack.
What we learned
Creative machine learning algorithms, and how to analyse data using cross validation, test, and training sets as well as implementing a variety of pre-existing algorithms and our own algorithms to improve performance from an initial 11% to an incredible 91% final accuracy!
What's next for Live Free
The implementation of RNNs instead of ANNs to increase accuracy.