Connecting AI to the ICU
On 16 March 2021, the Data plus Pizza meetup – temporarily redubbed ‘Data without Pizza’ for its online incarnation – returned for the second time since the coronavirus outbreak. It’s a special occasion, since many of the organisers behind the Amsterdam Medical Data Science (AMDS) network are usually busy treating COVID-19 patients at UMC’s intensive care unit.
The mission of the AMDS network – now with 1,813 members and counting – remains unchanged: to connect data scientists and medical professionals to help bring artificial intelligence to the clinical setting. And, in a bittersweet twist, the COVID-19 pandemic seems to accelerating these aspirations.
As host Dr Lucas Fleuren observes: “The combination of AI and healthcare seems to be really taking flight at the moment.” In other words, medical professionals are becoming more open and flexible to the idea of using data science to improve patient outcomes.
Preventing unnecessary interventions through machine learning
Bacteraemia is a blood infection caused by bacteria – often brought on by invasive instruments such as IVs, catheters and other foreign bodies entering arteries or veins during surgery or other medical treatments. And with blood being blood, these infections can often spread to other parts of the body. Bacteraemia can often be fatal.
Bacteraemia is tested for using blood cultures. These are attained during a rather invasive procedure that can cause further infections. In addition, the results of blood cultures take 24-72 hours to process, and only 11-15% turn out to be positive – with half of those being false positives. With the rise of antibiotic resistance, it’s also ill-advised to administer a pre-emptive run of antibiotics. So, what to do?
A predictive model could come to the rescue. In their presentation ‘Predicting blood culture outcomes in the emergency department using machine learning’, PhD candidates Roos Boerman and Michiel Schinkel from Amsterdam UMC introduced a model they have developed that can identify ER patients with a low risk of bacteraemia. By applying this model, ICUs can potentially reduce unnecessary diagnostics, antibiotic overuse, hospital length of stay, and medical costs.
While the results look promising, Schinkel says there’s still room for improvement. Due to privacy restrictions, no free-text data could be included. “And this is a real shame,” he explains, “because it’s in these fields we could find out whether, for example, a patient had received antibiotics earlier by their GP, or were immunosuppressive. Knowing this kind of information could really sharpen the results.”
But Boerman believes they already have “a valid extra test to help doctors make their decision on whether to prescribe antibiotics. It may make them wait and see – and thereby avoid unnecessary antibiotics.” Meanwhile, the pair also hope to improve the model so it can predict whether a certain infection is bacterial or viral.
“I’d also be interested to learn more about what would convince doctors to actually use our model in the real-world: the best way to present this as a tool that would be helpful for them,” says Schinkel. Is it perhaps time for a whole new algorithm to predict a doctor’s flexibility for embracing the new?
Randomness to improve healthcare logistics
To open his presentation ‘Modelling COVID-19 hospital admissions and occupancy in the Netherlands’, Dr Rene Bekker apologised to the data scientists in the crowd for not bringing in a lot of numbers. But, as an associate professor at VU’s Department of Mathematics specialising in performance analysis and randomness theory, his fears were unfounded. He had plenty of numbers and formulas to share.
As part of a wide-ranging collaboration, Dr Bekker helped develop a model that is currently being used by the National Coordination Centre for Patient Distribution (LCPS) to proactively spread patients to other hospitals around the Netherlands.
The model not only takes into account the daily number of patient intakes and discharges, but also everything from policy changes (such as when new lockdown measures are put in place) to the Inspection Paradox (how fewer patients may actually result in longer hospital stays)
Perhaps counter-intuitively, the model does not take into account any shifts in the number of positive tests. “These numbers are unreliable for a lot of reasons – symptomless carriers, changes in testing behaviour – and are simply not parallel with an admission spike,” says Dr Bekker.
Currently, the model is also being used to look at minimising the impact of COVID-19 on deferred care. “For the first wave alone, it’s estimated that 34,000 to 50,000 healthy life years have been lost due to cancellations.”
“We also want to look at the potentially even larger long-term impact on well-being and the economy,” says Dr Bekker. “And we should look more carefully at the healthcare system as a whole. There is an amazing system in place – very well organised. But what COVID-19 has shown is that we need to factor in the random. We need to become more flexible in dealing with rare events.”
About Amsterdam Medical Data Science
The Medical Data plus Pizza meeting aims to bridge the gap between health professionals and data scientists by bringing both together in an informal setting for presentations and pizza. At Amsterdam UMC data scientists, medical professionals and researchers discuss how AI (Artificial Intelligence) and medical data can be used to benefit patients and improve medical practices. Since it’s launch in August 2018 it has steadily grown in popularity with more and more people attending to hear the latest news and innovations in medical data and spark new collaborations.
For more reports from previous Medical Data plus Pizza Meet-ups, click here.
The Amsterdam Medical Data Science Group meetings are supported by The Right Data Right Now consortium, which includes Amsterdam UMC, OLVG, Vrije Universiteit, Pacmed, Smart Health Amsterdam and the Amsterdam Economic Board.