Diagnosis Predictor - Patient Platform

Your smart patient platform to provide accurate diagnoses at the click of a button!

Comment

What inspired us

Our main inspiration for this project came from the the many stories we heard about AI becoming intelligent and fleshed-out enough in the near future to be able to accurately predict all the problems a person could have, whether those problems are current or are years down the road. Our project seeks to emulate that "futuristic" level of accurate, universal diagnosis.

What we learned

We gained increased knowledge in how to work on large datasets, neural networks, google cloud, front-end design, and others in the industry with differing knowledge and experience.

How we built the project

Ideation was the pivotal stage in our development process, where we narrowed down our list of ideas down to Diagnosis Predictor. First we obtained the data from Synthea and analyzed it manually to determine possible use-cases and potential data cleansing. Then, we made a script that prepared the data by excluding unnecessary information and formatting it to pass to a recurrent neural network. We trained a preliminary model using TensorFlow and Keras. Meanwhile, we also started bulk uploading the data using the GCloud commandline toolkit and also prepared the AI Platform to export the model to be used with the RESTful API calls from the Flask python micro-framework, which is something that we plan to build upon in the future due to its significance in our original ideation phase. Overall, with Google Cloud, we were able to successfully build a working prototype for our AI application that intends to serve people and bring medical diagnoses to them at their convenience.

Challenges we faced

One of the major issues was working with large datasets, as we had over 300GB of data in total of synthetic patient data (leaving only 1GB of free space on my computer :/). However, after persevering through, we were able to upload a good part of the data to the cloud storage to be able to train the network. Preprocessing the data also required a considerable amount of time, as we needed to figure out how to assimilate all the information from a patient's medical history into something the network could use. Finally, a great hurdle was working with Google Cloud and identifying all the quirks that come with the 100+ products in its cloud suite. We were, however, successful in implementing the neural network and exporting the model for future client-side use.

Accomplishments that we're proud of

Some things we are proud of is having build a preliminary functioning neural network that could be trained to accurately predict the diseases that a person could have. We also managed to dive into working on the front-end website aspect of the project with minimal experience.

Built With

Share this project:

Updates