DeepODIN

Inspiration

We are Materials Science & Engineering PhD students at the University of Pennsylvania interested in using our quantitative skills to make an impact beyond our research. Thinking about problems affecting our community here in Pennsylvania, we came across the Overdose Information Network (ODIN) Data and became interested in building a data science solution to one of the most prominent and tragic public health crises in America - the opioid epidemic.

What it does

DeepODIN is a neural network built with TensorFlow 2.0 that attempts to use the ODIN database and US census data to predict mortality rate for overdose incidents based on factors like Naloxone usage, and the time and place where the overdose occurred.

How we built it

We collected data from opendataPA and factfinder.census.gov, and did data cleaning with python/pandas. The deep neural net was constructed and trained on GPUs with TensorFlow 2.0 in Google Colab.

Challenges we ran into

The imbalance in survival rates, as well as lack of differentiation between overdose incidents made network training difficult.

Accomplishments that we're proud of

Exploring a public health problem outside of our expertise and thinking about ways to use data science to make an immediate positive impact in our community.

What we learned

We learned practical skills about using TensorFlow 2.0 and GPU resources in Google Colab. More broadly, we got to learn just a bit about the many challenges and complications that arise when dealing with real-world data and issues that affect many people on a deeply personal level.