We focused on Humana challenge to minimize unnecessary provider interactions, and as a result lower costs. A lot of times patients simply lack resources to answer their health questions, or go to a doctor with a minor illness. Pulse allows you to self-service such requests. Our tool is especially useful to users who don't have access to a healthcare professional.

What it does

Pulse allows you to self-service some of the medical requests you might have:

  • Given a list of symptoms, we provide the most likely cause.
  • We answer questions you might have about diseases.
  • Voice interface which uses NLP and NLU to identify the likely disease given the symptoms.
  • Personal insights page from insurance provider.

How we built it

  • Question answering about diseases is done using BERT - a former SOTA model for such task. It is deployed on a separate server (needs GPU to run fast enough), with fine-tuning on data used. For this purpose we wrote a web crawler to index pages about diseases from Mayo Clinic.
  • Symptom-to-cause is done with a custom algorithm, "Disease-Symptom Knowledge Databse" dataset from Columbia University was used.
  • Voice interface built using BERT (over SQUAD dataset) and TTS modules.
  • Web part is done is Flask.
  • Data is stored in IRIS database.

Accomplishments that we're proud of

  • We've leveraged state of the art NLP technologies to build a useful tool.
  • Developed our own algorithm to provide real value from an educational dataset.
  • Used web crawling / scraping to get a new dataset.
  • Picked up some skills on interacting with a new database (IRIS) using pyodbc.

What's next for Pulse:

With more polishing this could be a useful tool for many, especially those who can't access a medical professional at given time.

Where to find us

Come find us at table 60 or at :)

Share this project: