The 24-hour Medical Data Hackathon organised by Deloitte and Amsterdam University Medical Hospital (Amsterdam UMC) saw data scientists working with AI students on real data from intensive care. In the process, the teams came up with real world solutions.
When one walked in during the final hour of this hackathon, one immediately recognised that some serious work had taken place over the last 23 hours. Ten teams, all diverse and with an equal mix of genders, were scurrying to finish up.
“The first team to walk in at the start brought air mattresses and sleeping bags. That made me realise people were in it to win,” observed one of the organisers Dr Lucas Fleuren, a PhD candidate in artificial intelligence and resident at the intensive care unit (ICU) at Amsterdam UMC.
As clinically relevant as possible…
On 5 and 6 March 2020, the 24-hour Medical Data Hackathon took place at De Nieuwe Poort in Amsterdam’s business centre Zuidas. The event brought together data scientists from Deloitte with students following a masters in artificial intelligence at VU Amsterdam.
For 24 hours, the teams worked with machine learning models using ICU data from the Amsterdam UMC – who recently made headlines as the first hospital in Europe to make data on ICU patients available for research. Doctors from the UMC were also on hand to ensure the questions were as clinically relevant as possible.
The challenge: improving care for ICU patients
Amsterdam UMC collects billions of data points from their ICU patients and now hopes to discover patterns in this vast amount of data to help improve the treatment of patients and reduce the number of complications. To focus the hackathon teams, UMC doctors formulated three challenges faced by intensivists:
- Can we predict atrial fibrillation in the intensive care unit ahead of its onset?
- Should we treat an increase of lactate for a specific patient in the intensive care unit?
- Can we predict whether we can stop antibiotic treatment in a specific patient in the intensive care unit?
Let’s get coding
After the kick-off, the teams were quick to organise themselves and begin interrogating the doctors to get as much relevant medical information out of them as possible. “It was great to see the doctors help out the teams to understand the problem and the data,” says Fleuren. “And also vice versa: the teams were explaining their approaches.”
Indeed, a member of the I-Cure team openly admitted: “I had no clue about lactate before yesterday.” And some of the doctors were also on a steep learning curve as they were confronted with terms such as: logistic regression, gradient boosting, random forest classification and death flags.
“We came with many hopes and dreams,” said a member of the Tomato Soup team. “To be honest, the challenges were filled with alien words. Luckily lots of doctors were on hand to explain. And I think a lot of people needed a while to get their head around Databricks…”
The winners: complete with app dashboard
After a long tallying process, a winner was announced: The Beta Blockers. They had taken on the third challenge: how to help doctors determine the best moment to stop antibiotics – in the name of saving money and staving off microbial resistance – while not sacrificing on patient outcomes.
Not only did The Beta Blockers get promising results, they were already able to translate their results to a clinically useful application for doctors – complete with a dashboard that showed which patients could have their antibiotics reduced or stopped without influencing outcome. “It was very impressive. It really makes sense for daily practice,” says Fleuren.
Going to the next level
The team was comprised of Amelie Schuler (age 28, data preparation/modelling), Leidy Molina (age 28, modelling/communication link between frontend and backend), Moreen van der Mooren (age 23, problem solver/dashboard development), Akin Ipek (age 26, data and machine learning engineering) and Peter Hoogendoorn (age 26, project management).
“As a team, we were able to provide proof of concept,” says Ipek. “Live dashboards and patient-by-patient predictions can really push healthcare units to the next level. With more time and data access, we’re sure we will be able to help make even more accurate decisions.” While the team came to the hackathon feeling confident that they had the right combination of skills, they were unsure of winning until the final announcement. One key to their success: they each got to sleep a few hours.
Everyone a winner
Fleuren was also impressed by the other teams. “Some teams presented the clinical problem very accurately, showing they had really listened to the doctors. Other teams had working models with decent accuracy. Lastly, some teams had no results to show, but their approach was creative and original.”
“One of the hostesses from the venue I think summarised it best,” says Fleuren. “As I walked out in the middle of the night to get a couple hours of sleep, she said: ‘You guys are craaazy, but it’s also really cool’.”