6 min read

GE Healthcare's Precision Challenge 2018

My experiences from my first hackathon and tips for students on how to approach them.

Cover Image for GE Healthcare's Precision Challenge 2018

We had 8 crushed cans of Red Bulls, 4 packets of half-eaten chips, two pairs of drowsy red eyes and 8 hours left. I hadn’t slept in days and all we did was stare at our screens. This was our first hackathon at GE Healthcare’s JFWTC campus, Bengaluru.

For students, hackathons are like ice cream trucks — selling swag, goodies and a carnival ticket. They romanticize the world of code and silicon, glittering it with ridiculous dumps of food and cash. The effectiveness of such large scale hackathons which rely on extrinsic factors of motivation to attract students and incubate innovation, can be debated. Or one could argue that the goal is not to innovate, but to rather encourage students to indulge in this field. Either way, we were sold on the idea. We were desperately looking for one and luckily fell upon GE Healthcare’s precision challenge. There was an ideation phase followed by a 24-hour on-site implementation challenge. We submitted our proposal within hours and were surprisingly amongst 36 teams all over India invited for the 24-hour coding sprint.

Collecting and maintaining medical data while respecting all privacy concerns is difficult. Under the US Health Insurance Portability and Accountability Act (HIPAA), there are 18 identifiers that must be treated with special care while collecting medical data. For many cases, collecting facial information becomes necessary, especially in the case of dermatology. Keeping records of patients who were diagnosed with acne, melanoma, rosacea, Eczema, skin tan etc. usually requires collection of images of the faces of the patients. In order to make the data non-identifiable, the images are usually cropped to show a specific part of the face, usually the forehead, cheeks and the mouth region. One other way to remove personal identifiers is to apply a censor bars on the eyes and mouth. Both ways need a significant manual preprocessing time.

Our idea was to automate this process while creating a parseable dataset with metadata. This will help researchers save data preparation time and easily share it for research and learning purposes.


We made our first mistake by going to the event with very little preparation, and It’s virtually impossible to complete a project with a front-end in a measly 24 hours. Nevertheless, we woke up that morning, barely showered and grabbed an Uber to JFWTC.

The place was ridiculous. They had this pantry setup with continental breakfast, a mini bakery, a juice bar and an ice bowl with a million Red Bulls just waiting to get you diabetic. After a small introductory session, we all got our workstations for the rest of the night.

4 hours into the challenge

4 hours into the challenge.

Our solution was an electron app where the user can dump multiple images and we clean it and parse it into a dataset. and we weren’t familiar with electron. I managed to bootstrap together an image cleaning pipeline by midnight and started working on a presentable demo while my teammate worked on getting the app to work.

We had 8 hours left and we were trying to figure out how to display our final outputs in the app. After 2 hours of going through documentation and tutorials, I had given up on my dreams of holding a giant cheque and took a nap. When I finally did wake up, my teammate had figured out a solution (which I didn't hear thanks, to my hearing loss. Damn you ATH-M50s) and went for his nap shift. Still in the impression that we are screwed, I started preparing our pitch to the judges.

Below is an illustration of how our solution would clean images and bundle them into a dataset. We used dlib's Resnet architecture to identify facial features. We then masked the features using Pillow. The textual information were identified using EAST text detection which we later blurred out.

Flow chart of how our idea functions

Fast forward to 9AM, we were done with our prototype and the pitch decks. There were multiple panels set up for screening all the 36 teams. We had to pitch our idea and present a live demo of the prototype to 3 judges in our panel, and we aced it. I could see the amusement in the judges eyes when we displayed our output. Once the screenings were over, we were asked to be seated in the main hall for the results. After toying around with us for an hour they finally started announcing the finalists, and when the emcee struggled pronouncing one of the team names, we knew we made it to the finals.

Our team name was GLaDOS.SHoDAN, a weird portmanteau of GLaDOS aka Genetic Lifeform and Disk Operating System, the infamous cynical AI from the Portal game series, which is the name I go by on various online platforms, and SHoDAN aka Sentient Hyper-optimized Data Access Network, another AI from System Shock game series, which was my teammate's choice. He just wanted to choose something cool.

I was in the impression that we were done, they'll give us something and call it a day. But for the finals, the top 5 teams had to compete in a shark tank style pitch to the audience and the panelists. We went in pretty early for our pitch and did pretty ok, though we didn't get any questions, unlike others who were being grilled, and it's generally a good thing to get a lot of questions. Looking at other teams, who worked ideas far more innovative than ours, I had lost all hopes on getting through. I went and sat in the corner in dismay for the rest of the event. But as it turned out, we actually won! 2nd runner up. Bagged 50K cash, an internship and got to hold the cheque! Although there were far more exciting projects, they weren't really executable.

Our team holding a cheque of 50,000 rupees

You don't need to see our faces, we didn't bath for a week.

Hackathons are intimidating. I wouldn't have even thought about participating two years back. But to be honest, they aren't that hard to crack. As I write this article, I have participated in around six competitions and what I've learned is, they are all about the art of pitching. We need to sell the idea, not its technical prowess.

We usually get lost in finding the technical novelty when finding the right product-solution and market fit is the key. One extremely effective way of figuring that out is by observing our environment. There are a hundred different interactions and events happening around you, many of them being suboptimal or riddled with issues. Note them down on a notepad or an app, even if you don't know how you can solve it. The next time you participate in a hackathon, you'll have an ample number of problems to work on, and you'll be cashing in on giant cheques like hotcakes.

Michael Scott from The Office giving cheque to a stripper

I love this episode.