‘Inspired AI’ and the future of health

On Thursday, 12 November, World Summit AI 2020 presented the online ‘Inspired AI’ programme ‘The Future of Health’. Smart Health Amsterdam’s director of programming Femke Blokhuis acted as chairperson to a dizzying array of presentations – from Chronolife’s AI-powered t-shirt that can detect and predict worsening heart failure, to SkinVision’s app that identifies skin cancer as effectively as an experienced dermatologist.

AI adds precision to the balancing act of cancer treatment

On the main stage, the panel for ‘Clinical applications of AI in Oncology’ turned into more of “a fireside chat”, according to the moderator Jeroen Maas, the Challenge Lead Health at Smart Health Amsterdam. As irony would have it, one of the scheduled participants had a minor accident and had to go to the hospital.

Maas took the opportunity to introduce the international audience to Amsterdam’s ambitious ‘AI Technology for the People’ initiative that involves investing a billion euros over the next 10 years to encourage collaboration with leading institutions around the world, as well as public partners and members of the business community. One of the partners is Netherlands Cancer Institute (NKI) where guest and University of Amsterdam professor Jan-Jakob Sonke leads a research group on adaptive radiotherapy.

When it comes to medical imaging, AI can be trained to analyse MRI and CT images better and more quickly than the most specialised radiologists. As a result, you can analyse more patients and get better diagnoses in less time. And thanks to machine learning, the process has been sped up enough to track changes during treatment and adjust the course daily if necessary – reducing the side effects of unnecessary treatments by either length, dosage or location. Everyone wins.

Sonke explained how such innovations can be applied quickly in practice since NKI is both a research and treatment facility. Recent NKI breakthroughs include automatic pulmonary embolism detection – a potentially deadly side effect to certain treatments. “Radiologists would discover the signs in images a few days after a patient died – and feel guilty they didn’t see it earlier. So, this is a particularly beneficial workload and workflow enhancer.”

But Sonke also recommends patience. “At NKI we have a lot of data, but we know there are biases, so it needs to be validated against other high-quality data – to check if the models remain robust. Luckily this sense for collaboration is strong in Amsterdam. Check in over five years, and we’ll be able to talk a lot longer.”

Sonke also doesn’t expect any self-driving digital surgeons anytime soon. “It’s about supervised algorithms in the short-term. Unsupervised is somewhat further in the future,” he says dryly.

The AI art of patient treatment strategy selection

During the session, ‘Using AI to support patient treatment strategy selection’, three young Amsterdam researchers presented their findings. This session was hosted in collaboration with Amsterdam Data Science and Amsterdam Medical Data Science. The first talk again emphasised how cancer treatment is often a balancing act.

With ‘Evolutionary intelligent bi-objective treatment planning for prostate cancer’, PhD candidate Anton Bouter explained how he designed and applied a model-based evolutionary algorithm to better control High-Dose Rate (HDR) brachytherapy. For this internal cancer treatment, the key is to ensure the prostate and seminal vesicles receive a sufficient dose, while minimising the damage to surrounding organs such as the bladder, rectum and urethra.

The resulting novel treatment planning method BRIGHT generates numerous personalised treatment plans showing the different trade-offs between tumour coverage and sparing organs – that a doctor can use to decide the best course. Already, deemed superior to previous plans, the system is now being applied at Amsterdam UMC.

From ‘oh shit’ to ‘oh wow’

Moving beyond oncology, ‘AI-driven optimal medical treatment protocol for mechanically ventilated patients with COVID-19 pneumonia’ had Luca Roggeveen, a resident at Amsterdam UMC, talk about his ICU experiences during the first wave of COVID-19.

“Halfway through March, we thought no worries, but two weeks later: ohhh shit. We saw how a significant portion died and a significant portion were discharged,” recalls Roggeveen. “We also saw flawed research such as with hydroxychloroquine, and we saw some success stories such as with dexamethasone.”

So, the question became how do we create more good news? His fellow doctor-meets-data-scientist colleague Lucas Fleuren said: “Let’s just collect all the data from all the ICUs.”

Due to numerous issues – from technical to ethical – this is a near-impossible task. “We were all still working in silos with each hospital having their own system,” says Roggeveen. But in times of crisis, even the medical ethics committees of 60 different hospitals were on board – so a process that usually takes two years could start within three days.

Lots of beautiful data

“I’m here to talk about what I’ve been doing with all this beautiful data Luca mentioned,” says Lucas Ramos to open his presentation ‘Predicting mortality of individual COVID-19 patients: Can we improve decision-making?’.

“I’m so glad I did not have to organise the collection of all this data, but it’s very rich,” says Ramos who is a PhD at the department of biomedical engineering and physics at Amsterdam UMC. He researches combining heterogeneous data (images, signals and tabular data) using machine learning to predict potential outcomes and complications.

Ramos was disconcerted by how in Spain and Italy, overwhelmed ICUs had to decide who would get treated and who would die – with age often being the single deciding factor. “So how can you make this decision less crude and discriminatory?” wondered Ramos. “The research question became: How can we predict 21-day mortality of COVID-19 patients?”

Using 80 different variables from Dutch ICU data, the research indeed confirmed that age is the main variable, but there are other factors such as the number of medications used by the patient, blood nitrogen levels and cardiac history. In addition, he discovered he could get the same results using only 10 different variables – streamlining the process considerably.

“Of course, we now have to see if this actually adds value to triage and how this lines up in other countries using other protocols,” says Ramos. “But meanwhile I can only recommend to everyone to use this beautiful data for research. You can request access at covidpredict.org.”