In November of 2016, GO sponsored Stanford University’s first healthcare-focused hackathon. Health++ was a weekend event hosted by student group Stanford Health Innovations in Future Technologies (SHIFT) and attended by more than 200 students and professionals of various disciplines from around the globe. The goal was to tackle problems of healthcare affordability, and to prototype real and applicable solutions. Read the following account from Doron Reuven, student at UC Berkeley and member of winning team, R-Net, recipient of GO’s first-place prize.
Big Data. You often hear this word in the slogan of the latest Silicon Valley company, but what is it? Big Data is the act of leveraging large amounts of information to find insights or patterns in data. These insights would often be near impossible to discover using our raw human intellect alone.
My team and I sought to demonstrate the applicability of Big Data to oncology at a health hackathon at Stanford University. A hackathon is an event where engineers, designers, and innovators collaborate during an extremely short amount of time, in our case 2 days, to solve a problem and develop a prototype of that solution.
We created a project we named R-Net. R-Net is a system designed to help radiologists make quicker, more accurate decisions and to help train radiology students. The system read tens of thousands of anonymized reports from the University of California, San Francisco hospital and is able to automatically determine which reports are most similar to one another based off of the report description. In our prototype, the radiologist enters a report description and immediately the five most similar reports are shown. Each report shown to the radiologist is given a relevance score on a scale from 0 to 1 in order of increasing relevance (shown below right; images are placeholders but the descriptions are real)
One of the members on our team is a radiology resident at UCSF. In his experience, such a system could aid radiology students when they aren’t quite sure how to classify a report. Radiologists might use such a system to check the validity of their diagnosis or to expedite the process of making a decision.
The biggest challenge in this project was how to compare the “closeness” of reports. This task is difficult since each radiology report is written in natural language, meaning there is no structure to the text in the report description. To overcome this, we used an algorithm called “term frequency-inverse document frequency”. This algorithm essentially relates documents by the number of words they have in common weighted by how uncommon the words they have in common are. For instance, the word “the” would do little to make two reports more similar whereas having the word “pneumonia” in common would.
This project was especially meaningful for me because my mother passed away from breast cancer. For many years I have imagined what the experience of a cancer patient will be like in the future and what systems would have to be engineered in order to make that future a reality. Although many more improvements can be made, it was exciting and rewarding for me to see what could be accomplished in just two days. My team and I are grateful for having won the Global Oncology prize at the hackathon.