The tale of my first hackathon
I recently competed in my first hackathon — the HealthAI Hackathon @ Stanford.
As some may know, I have developed a passion for healthcare-focused use cases of artificial intelligence and machine learning, and I am working to gain both business knowledge and technical expertise. Outside the UC Berkeley School of Information’s applied machine learning course and an internship with medical imaging AI startup Visla Labs, I am constantly looking for new opportunities to learn and grow.
So on a Monday when my friend and San Francisco techie pathfinder, Akansh Murthy, told me about a 2-day weekend hackathon that upcoming weekend, I immediately cleared my calendar and signed up. My fiancé, Rika Sukenik, competed in (and won!) multiple blockchain hackathons, but I had always been skeptical and honestly a bit intimidated. But it’s early 2019: a time for transformation and new goals.
I arrived at the hackathon early and was welcomed by a surprisingly energic Ben Busby PhD, a hackathon coordinator for the National Institutes of Health (“NIH) and National Center for Biotechnology Information (“NCBI”). Recalling a similar feeling to a past life as a single man at a bar, I urged myself to smile and timidly introduce myself to as many people as possible. After meeting a few folks, it became easier, and I found that there were many diverse backgrounds, not only developers.
The hackathon coordinators — Lukasz Kidzinski PhD and Dr. Olga Afanasiev — kicked off programming with keynote speakers from Google Brain, NIH / NCBI, and Stanford. Then, of the 100 hackers, 20 pitched ideas to work on, and a courting ritual commenced, forming diverse teams– including developers, statisticians, clinicians, designers, business people, PhDs, and Masters students.
I hit it off with two guys who were already working on a startup, called Brain Key, that is part of Y Combinator’s Winter 2019 class. Owen Phillips PhD was a Stanford postdoc in Neuroscience and Ben Kotopka is wrapping up his Stanford PhD in Bioengineering. They had access to a bunch of brain MRI scans and were looking to build a predictor for their business. In addition to myself, Owen and Ben were able to recruit some awesome folks, including Yann Le Guen PhD (Stanford postdoc in Neuroscience), Rudra S Bandhu PhD (AI and ML enthusiast), Daria Czerniawko (deep learning enthusiast), and Bartosz Zieba (Fudo Security tech support and systems administrator).
We started hacking at 1:00pm and ending around 1:00am. Here’s what happened…
Owen and Ben kicked things off by explaining the data and predictors that we could possibly develop. Based on team member skill sets, data availability, time, and perceived stakeholder value, we landed on a tool that was fed 3D MRI scans and determined patients’ “brain age”. We thought that this would be similar to the outputs of other health risks assessments. By knowing your “brain age”, you can better know your brain health and then possibly associate that with neurodegenerative disease prognoses, such as Alzheimer’s Disease. So appropriately, we named our team Brain Age.
The development strategy was to use machine learning to fit a multivariate regression model on age and image intensity histograms. These histograms were derived from 2D slices of the 3D MRI scans. Owen and Ben knew to build on existing research from James Cole PhD on associating brain age with mortality.
After seeing that our development infrastructure was set up, I asked the team to go over the game plan: what activities needed to happen and what was the sequence, who was capable of doing each activity, and who would specifically own each activity. I definitely didn’t know the answers but instead could ask the questions that got us all on the same page; plus, I brought an assortment of sticky notes, sharpies, and whiteboard markers. Once on the same page, the team was better able to divide and conquer.
Yann and others started to preprocess the 3D MRI scans, cutting them into slices and removing unwanted information (e.g., skull). Once this happened, almost everyone started setting up the ML architecture. There were some issues with compatibility with our infrastructure as well as limited TensorFlow experience. Rudra, Daria, Bartosz, Ben, and Yann each were hacking away at it, but after a while, the team pivoted to PyTorch.
In parallel, Owen and I were storyboarding the pitch — my specialty after being a strategy consultant for five years and a summer associate in HealthTech venture capital. I pulled from my human-centered design days and framed the problem by looking at a specific use case around a 68-year-old woman named Mary. With big pictures and few words, we showed how Mary would have been better off knowing about her brain health. Thanks to my social networks on Facebook, Twitter, and LinkedIn; I was connected with neurologists, radiologists, and PM&R specialists — each helping to make the story, business model, and predictor more believable.
All computers closed at around 1:00am… besides for the virtual ones that were running the models.
By 8:00am on Sunday, the team started shuffling into the team room and unfortunately found some hiccups. Miscommunication due to rapid decision making, led to some errors early in the process and much of the overnight processing had to be redone. However, with somewhat fresher minds, the team was able to pull together an get to a decent predictor using three estimating approaches — Lasso, Ridge, and ElasticNet.
At 12:59pm, we submitted the presentation on time and the tech was done. Then, eight of the 20 original teams presented to a panel of judges, including AI researchers, VCs, clinicians, and entrepreneurs. Ben and I jammed a ton of information into our two-minute limit, and the team fielded various follow up questions.
What felt like an eternity, the next round was announced and our team was top five! We would be presenting in front of the entire 150-person hackathon. This was super exciting for everyone, and we refined the now five-minute presentation based on feedback from the previous round.
After a keynote from Stanford physician and AI researcher, Dr. Jonathan Chen, the final presentations began. Our team presented in the middle after some awesome ideas — clinical trial matching, clinical photo de-identification, precision medicine for immunotherapy, and stroke prediction. Ben and I knocked out the five-minute presentation, and the whole team came together to answer follow up questions. At the end, I was immensely proud of both what I had accomplished and what our team had accomplished. Everyone demonstrated growth.
The judges went away for what was supposed to be 30 minutes; but due to intense debate, it turned into an hour. The room was tense when Lukasz and Olga began talking, given that a lot on the line — automatic interviews with the StartX accelerator, an interview with LEO Pharma’s corporate VC fund and access to the LEO Innovation Lab, and NVIDIA Titan V GPUs.
Sadly, our team, Brain Age, did not win. Stroke De-coded took first place for using AI to predict time of stroke from brain MRIs — a similar yet distinct technology from ours. Despite being a little bummed, our team hung out a little longer, had dinner and a few laughs, and exchanged contact information and job ideas.
This was a fantastic first immersion in the technical side of healthcare AI/ML, and I look forward to going deeper.
Here is our pitch, if you are interested to learn more!
Here are some learnings, from the lens of a less technical individual…
1. Not all hackathon participants are developers. At this event, there were clinicians, non-technical students, designers, and businesses-focused individuals. All were important and contributed in their own way.
2. A minimum viable product (“MVP”) is necessary but insufficient. Hackathons are primarily focused on building new technology and are skewed toward people with those specific skill sets. However, I was happy to find that communicating the problem statement, explaining how the solution works, and outlining a business plan were also critical.
3. Organization is a critical skill for developer-heavy teams. Having someone to press pause and facilitate communication and planning increased efficiency, accuracy, and trust. It would have been easy to run off in the wrong direction due to the stress imposed by limited time
4. Awesome startups have a presence in hackathons. Numerous startups led groups, looking to get free help, fresh ideas, trials runs with potential hires, and possibly prizes. This was a great space to get a job if you were looking
5. Awesome startups are born out of hackathons. Numerous teams, especially the winners, are inspired to continue developing their ideas. A few from previous hackathons came to speak as keynotes and share what is possible as a next step after a hackathon weekend
6. Hackathons are competitive but not cut throat. Despite being a competition and everyone wanting to win, each person was friendly and supportive — helping other teams and chatting during down time. It is a great space to make new friends and come together around a niche that you all enjoy
Now that I have caught the hackathon bug, I am searching for any other health and/or AI/ML-related hackathons. Below is a running list that I will continue to update; and if you have any tips, please do share!
● Texas Healthcare Challenge (Feb 8–9) — Dallas, TX
● Hack Mental Health Yale (Feb 22–24) — New Haven, CT
● Hack Mental Health UCSF (Mar 23–24) — San Francisco, CA
● Grand Hack MIT Hacking Medicine (May 3–5) — Boston, MA
● Mount Sinai Health Hackathon (Oct 2019?) — New York, NY
● Jefferson Health Hack (Nov 1–3) — Philadelphia, PA
● Health Plus Plus (Nov 2019?) — Palo Alto, CA
● Miami Healthcare Hackathon (December 2019?) — Miami, FL
Thanks to MIT Hacking Medicine for getting things started with your healthcare hackathon database!
