The people who work at Unai are highly creative. After all, data science is a creative process - but that creativity is bound, to an extent, by the projects they’re working on.
So, just for a day, we wanted to change it up for our team and give them free reign. In hackathon-style, we gave everyone the freedom to form teams of six or so and compete to create a winning idea.
They would have two hours to plan and one day to build it – code and all – so they could show off a demo of it to a judging panel (consisting of our Chief Data Officer, Matt Oates) and a captive audience of colleagues.
And we were completely blown away by the results.
It was the first time we’d ever done anything like this as a company, so we weren’t entirely sure what to expect. Though, judging by the usual calibre of our teams, maybe we should have seen this coming.
Six hours to change the world
As with algorithms, so with competitions: there were rules.
Each team’s hack would be based upon Ruleau, Unai’s own decision engine. Since Ruleau is designed to be transparent, auditable and interactive, teams would receive points based on how well they stayed true to these principles.
There would be bonus points for ideas with a social impact, for being practical and familiar, for having good documentation and for being viable in the market (i.e. we could actually use it as a company.) For instance, one team put together a way to automatically check that passengers on airlines complied with safety checks.
In other words, the challenge was to take Ruleau and repurpose it for a viable market using new algorithms created in a single day. Six hours to be precise.
Since Ruleau is so new within Unai, most of those hacking on the day had never used it before. In fact, none of the winning team had ever used it outside of a few training sessions. As Graham, a winning team member said, “The hackathon was a much more interesting way to get to grips with the tech.”
And speaking of the winning team, let’s introduce them.
The Dream Team
The winning team chose the name Creme Ruleau – and they were a talented lineup.
Graham Parsons
Generating patient data in Synthea
Writing ruleset in Ruleau
Lorena Paskeviciute
Writing ruleset in Ruleau
Joe Bryan
Generating SNOMED codes
Documentation & tests of ruleset
Luke Merrett
Planning
Research for & writing presentation
Laura Sobola
Project idea & planning
Writing ruleset in Ruleau
Before they created their winning hack – which we’ll get to in a second – they had to figure out what on earth they were going to do. The teams would need to spend all their energy hacking on the day, so we gave them two hours of company time beforehand to bash out some ideas.
“The team’s first ideas didn’t fly,” says Graham, “and I was skeptical about whether we’d come up with something we’d be excited about. But then Laura brought up an idea which evolved, and it became something we were all keen to work on.”
Laura remembered that, two years ago, one of Unai’s potential clients had asked whether it was possible to resolve a medical triage issue with automation. Originally an early-pandemic problem, the need for a solution was now less urgent. But if Creme Ruleau pulled it off, their idea could still help a lot of people.
As for the idea itself... well, imagine that you’re in charge of an intensive care unit...
Who lives and who dies?
Here’s how Creme Ruleau pitched the problem to us:
On average, an ICU unit in the UK has 11 beds; that’s one bed to every 21,300 people in the country.
Triage around who gets prioritised can often be a difficult, ethically charged process, especially during pandemics such as COVID-19. It can even be hard in flu seasons.
To help make these crucial decisions, standardised rulesets do exist (such as the Charlson Comorbidity Index). But there are always exceptions that fall outside these provisions.
At the start of each day, doctors have to check how many beds there are, how many staff and which patients can be discharged. Most difficult of all, they need to consider who has the best chance of surviving in an ICU environment.
Doctors may later be asked questions. So they need to be able to trace back why certain patients were given precedence during triage. It’s fair to say it’s a weighty burden to bear.
The solution to triage
The hackathon team saw that a semi-automated process could provide incredible relief and empower ICU staff’s decisions – reducing both admin and the mental strain of the ethical dilemma. It could also speed up in-patient decisions and help create better patient outcomes.
And of course, since the solution would use Ruleau and follow its principles, all the decisions suggested would be traceable and auditable. ICU staff could check the Ruleau user interface to see how an in-patient decision was made – and a doctor would still have the power to override the decision to admit, or refuse admittance, if they deemed it right. The responsible human is always kept in the loop.
Simply put, the solution takes the existing triage process and improves it to be faster, clearer and traceable, without losing the value of human input.
What makes the solution even more special, is that it’s replicable. Because it uses open rules, the underlying logic can be shared across different regions. That’s particularly important in a field where the term “heart attack” can be written a myriad of different ways on a patient record. So while Creme Ruleau’s solution can be customised, no ICU unit would need a data scientist to write a new ruleset for their hospital from scratch. It would almost work out of the box.
The race against time
You might wonder how a team of six people could possibly create and demo such a sophisticated idea in a single day.
“The idea was perfect for Ruleau because it was based on rules,” says Laura. “We knew how many rules we wanted to have and how many we could realistically include in six hours. After that, the key was really good delegation.”
Graham notes that they were also a bit lucky: “It worked because of the mix of people. Luke was good at organising and chairing the meeting in a soft way – getting us out of rabbit holes when we needed to. And Laura had experience with these kinds of ideas, so the plan even before we began was great.”
Laura says that everyone was interested, experienced and happy with what they brought to the table. And they were able to collaborate and move flexibly between tasks throughout the day.
“We were super careful not to overextend ourselves,” says Laura. “It was most important to determine the ICU cutoff decision, so we focused on that and stopped at five algorithms. The end result was super polished.”
Lorena adds, “Once the day began, everyone was good at getting on with it to hit the deadline. The fifth and final algorithm was complex though, so an extra hour would have helped!”
Who should you trust with your life?
Of course, although a hackathon involves a fun sense of competition, the winning team chose a very sensitive issue to solve.
“We hoped that by adding in a piece of machine separation, we could alleviate the burden of triage,” says Laura. “There’s not a lot of difference as to whether a human or a doctor is weighing up the factors that contribute to a final decision. So our solution is very close to what is already happening in ICUs, except that it’s semi-automated.”
The doctor could still examine the decision-making and override it. But there would only be one useful override – to admit the patient or not. A doctor can’t magically make a patient younger, but sometimes they can see flexibility that the machine can’t account for, so the override is important.
“It’s valuable where we are right now in our engagement with AI,” says Graham. “Medical ethics has been evolving for a couple of hundred years, and it’s mostly been trial and error. It’s also very robust and everyone agrees with it, while AI ethics is often more hypothetical.”
“Humans make mistakes all the time,” Graham says. “We’re clinging onto the idea that their neural networks are more trustworthy. But the purely human route might be the least humane.”
Lightening the mood, Lorena says it was really fun to work together on something different. Even if some of the discussions were a bit morbid.
The future of ICU triage
The winning hackathon team recognised that their working demo still had limitations, as you might expect from a solution made in a single day.
Right now, their solution doesn’t consider the conditions of patients already admitted to the ICU, it only considers patients on a case by case basis, based on their chance of survival, the number of beds and the number of staff. So the next stage would be to build in the capability to analyse the entire ICU, considering all patients at once to achieve a fairer and better optimised system.
Reflecting on the hackathon, Creme Ruleau realised that open rule sets are something of a game changer. It enables any expert in any field to quickly make a decision engine like Ruleau useful to their industry.
Right now at Unai, we’re onboarding Ruleau partners. So if there is a rules-based problem in your industry where our decision engine can make a difference, we can work with you in partnership to design and implement a solution – just as our winning team of hackers did with ICU care. To get the conversation started, simply get in touch.