Open Sourcing Science and AI

Introducing the Prediction Lab by Kartik Thakore

A few weeks ago at the SF Healthcare Data Science Meetup I presented a sneak peek at the MVF (minimal viable fun!) of the Prediction Lab. The inspiration for Prediction Lab came from blocks d3.js project. Blocks is an exciting project that showed a community of visualization geeks what could be possible with d3.js. Additionally, blocks have an amazing website encouraging the community to share their visualizations for bragging rights. The best sort of rights :)

D3.js Popular Blocks

At doc.ai we want to democratize access to health data and make it easier for doctors, data scientists and engineers to participate in pushing the frontier, as Medical/AI practitioners. Predictions are a way that Medical/AI practitioners can provide insights to patients and the population about their medical data. In this regard one of our goals is to build a community of healthcare data geeks and enthusiasts that want to collaborate and build insights (predictions) on top of self/community collected medical data. Essentially we need a portal that facilitates:

  1. Real-time feedback from health care data and prediction models
  2. Ability to sandbox the result and the code in the same place
  3. Community Building: Allowing other data hackers to comment on the model code

Real-time Feedback

At doc.ai we have built several integrations to many different types of health data sources which have since then open sourced (github.com/doc.ai). You are more than welcome to write your own integration to health data sources you feel is valuable following our importer architecture, and use it in prediction labs in the future. However, as many of you know working with medical data is tricky and difficult. It was quite difficult to onboard open source developers to show them what how they can build insightful predictions!

Open source code have enabled amazing tools and platforms to develop. This has certainly been possible with the expo snack project with application development. Snack.Expo.Io is a phenomenal project that allows users to quickly build React Native mobile apps in the browser. We need the same or something similar in the medical data field!

In the Prediction Lab you can run python notebooks against our Medical Selfie and use the data against you own prediction!

Import your Selfie and Predict risks!

Sandboxing: Result and Code in one place

One of the most valuable features of Expo is that you can update code and see the results in the same place! This facilitates good flow that reduces the friction for users to get started. Traditionally, getting this working requires a lot of data engineering work. By simplifying the steps needed to feed new data into a model it allows users to play by tightening up feedback loops. The beauty of this solution is that you and users of your model can bring your own dataset to perform analysis with!

Community Building

Finally we want to build a community and that means sharing and getting comments. For this reason we are playing with Disqus and incorporated that into our UI.

In the future we want to bring in forking and sharing features built into our notebooks.

Next Steps

We are gearing up to launch https://predictionlab.doc.ai/ soon. Today this is in alpha state and our team is working hard to make it a production ready product. If you are interested in taking a sneak peak go ahead! (caveat emptor: this is alpha and might be rough around the edges). Today we are support importing data with our first importer — Medical Selfie which takes your selfie and predicts your BMI information in a structured way, imported right into your jupyter notebook!

Our immediate next steps are to move to a JupyterHub enabled platform on Kubernetes which will use github oAuth to have personal provisioned environments.

Our future is to have a PredictionLab mobile app, similar to Expo, with several client side importers, including our exciting New Medical Selfie.

The users of this app will be the first people to see our decentralized importers, models and storages. Additionally, with the release of TensorIO wrapper for TensorFlow lite (created by our mobile machine learning expert Philip Dow) we may be able to allow users to deploy their models directly on the mobile app!

Keep your eyes peeled! More to come :)