Medical Data + Pizza #17: playing Cupid for better patient outcomes

The 17th Medical Data + Pizza took place on 15 February 2022 in the new O2 building next to the VU location at Amsterdam UMC. While the organising Amsterdam Medical Data Science (AMDS) network has co-hosted various online events in the past couple of years, interaction had been somewhat limited, since they largely took place via Zoom chat.

Meanwhile, with the network now boasting 1,850 members, data science has obviously never been sexier. As the opening speaker Giovanni Cinà noted: COVID-19-motivated projects such as The Dutch ICU Data Warehouse highlight the power of collaboration in elevating knowledge and improving patient outcomes.

In short: let’s mingle.

Looking for shortcuts in finding effective COVID-19 treatments

Cinà works in two worlds, which AMDS actively seeks to unite. He represents the business world as research lead at Pacmed, a company bringing cutting-edge AI technologies to the ICU. And he’s also in academia, as the assistant professor in Responsible Medical AI at the Institute for Logic, Language and Computation at the University of Amsterdam.

In his presentation ‘Practical causal inference – average treatment effect estimation in critically ill patients with Covid-19’, Cinà presented Pacmed’s work in developing guidelines on how to quickly apply observational data and AI in the improved care of COVID-19 patients.
With a newly emerging illness such as COVID-19, frontline ICU practitioners cannot rely on the gold standard of long-term clinical research using randomised trials to decide what treatment works best. They work in the now, and the sooner they get a grasp on what medical intervention is the most effective, the better. Cinà and his team believe practical causal inference can be an answer.

Pacmed’s first study related to estimating the treatment effect of the proning manoeuvre on COVID-19 ICU patients. This act of turning patients on their stomach to improve respiration is already routinely used for ARDS (acute respiratory distress syndrome) patients. Hence, the team emulated a recent randomised control trial for ARDS using data from COVID patients.

“While the results are promising to clinicians using the treatment, a true randomised control trial will be needed to confirm effectiveness in a definitive way,” Cinà said.

“So, you may ask: why is this research useful? Because sometimes we do need to act. And arguably, having partial information is better than having no information,” said Cinà. “The insights can also help to plan the future trials – helping to cut time and costs.”

Your Tinder-like app for medical literature

According to dr. Jan Willem Plaisier of the Amsterdam UMC, “as doctors, we want to deliver the best possible care to our patients. That’s why we need to stay up to date with scientific literature.” However, it can be difficult to keep up. “Each day, thousands of articles are published in medical journals. Finding the meaningful ones turns out to be a hard job.” In the presentation ‘Daily Dose – the Tinder app for doctors to find the best medical papers’, a solution was sketched out that combines AI with a social platform. Daily Dose starts with scraping the abstracts of all the new research papers published on PubMed. Then algorithms work to make an initial selection of relevant articles in a particular field of interest. Doctors “swipe right” on those articles they think are particularly relevant, which they can then share and discuss with colleagues.

Lars de Ruiter of the Amsterdam Data Collective was on hand to explain how the recommendations are personalised using neuro-linguistic programming (NLP) – and to stress that the project is still in a very early stage. Currently, Daily Dose is being tested by a group of volunteers, so the challenge remains how to develop a model and measure performance without a user base.

In time, the Amsterdam Data Collective hopes to start comparing users instead of just articles, and apply any user ‘like’ as a reinforcement for the model. They may even consider switching over to another already existing model.

During the discussion period, the audience latched on to the project’s social media aspect. People wondered if the team was worried about the notorious ‘echo chamber effect’ or the well-document bias for flashy research results. As Lars noted, these were “future challenges,” but reassured the audience that “we will want to look to more reinforcement learning to prevent these sorts of black holes.”

A hush fell as the crowd moved to another room for pizza. But a few slices in, the volume increased. But a few slices in, the volume increased. The co-mingling had begun.