One of the largest problems facing the medical field is patient compliance. This is especially true for mental illnesses and disorders . According to a 2013 study, one of the leading causes of non compliance in psychiatric treatment is irregular attendance to clinic accounting for about 55%. Another cause is lack of education about medication and ignorance about side effects of medication at 61%. It was also observed that social supervision was very poor in non-compliant patients whereas 49.4% of compliant patients had very good family support.

What it does

Provides psychiatrist with a dashboard to visualize patient sentiment between patient visits.

How we built it

Three steps: implementing a chrome extension with Javascript to scrape user input from the Messenger web application, then sending it to a server through HTTP POST; training an LSTM on labelled tweets to extract emotion from variable-length text, and designing a responsive React.js dashboard to allow telemonitoring of patient mental states and physician annotation of the patient timeline.

Challenges we ran into

Messenger extension: Scraping the data from just messenger because they seem to obfuscate many of the dom elements. Adding event listeners required a lot of testing to understand how different dom elements interacted with others. For example, an event listener for a send button could only be added once text was added to a textbox, requiring an eventlistener for an eventlistener.

For the LSTM, making variable length encoding, coding up the neural net with keras, and finding extensive training sets was very difficult.

Accomplishments that we're proud of

We are very proud of the Model we made, we tested with several sentences and it seems to work very well. We all walked out with a much better understanding of word embeddings and how LSTM's take input.

What we learned

We learned a lot about natural language processing and the unsolved challenges behind it. We found it was difficult to impossible to diagnose the disease rather than just evaluate emotion. We learned how to spin up servers, how to use react with fetch to retrieve information. We learned how to create a nodejs application to listen for get and post requests.

What's next for Sunrise

We hope to explore more emotions and diseases. NLP has the potential to learn a lot of nuance about people also through personalized machine learning, where each person has a model for them. We hope to one day be able to market to real physicians to create real impact.

Share this project: