How University College Dublin uses SAS Viya and AI to drive better COVID manageability

This is a use case of how the exploitation of one piece of research can, through astute application of AI technology combined with medical experience, produce a potentially huge jump in the way senior managers - in this case clinicians - plan the management of resources in order to contain and resolve a wide range of previously intractable problems.

Dr Paulina Szklanna, Manager of the AI Healthcare Hub at University College Dublin (UCD)  had not long successfully completed her PhD studies in biochemistry, working as part of a team which was studying and identifying the proteins that can be found carried on blood platelets. The team was not only developing the equipment needed to physically find the proteins on platelets, but also ways of using AI and machine learning to help classify and identify individual proteins from the data the equipment generated:

I'm extremely passionate about the democratisation of AI. By that I mean we want to make AI accessible to non-coding researchers so that they can accelerate projects and make data-driven decisions to deliver meaningful and translational results that help solve difficult real world problems, such as the COVID-19 pandemic.

The need for strong interdisciplinary teamwork is key to what the Hub is setting out to achieve. It requires the ability to bridge the taxonomy barriers that exist between electronic engineering, biomedical research and clinical healthcare management and operations. This requirement has developed out of the fundamental discoveries that Professor Maquire’s team has made.

Evidence served on a plate(let)

This is, stripped to its essence, that blood platelets are carriers of a range of proteins on their surface, and there is, they have discovered, what can be termed a standard `set’ of  them that define a normal, healthy human individual. They can be considered as markers of normal human processes and activities. But one thing that has been known for some time is that there are other, less common, proteins and enzymes carried in the blood that are produced by processes or events in the human body that one might hope had not happened. It is, for example, already a standard part of the diagnostic protocol for heart attacks and helps cardiac specialists understand the severity of an attack a patient has suffered.

Extending this basic approach to examining platelets led to the realisation that the appearance of anomalous proteins would likely be the markers of a source of disease or disorder, and the team decided to apply their own resources, and those of SAS, to start to identify them. Working with consultant haematologist Dr Barry Kavan at St James Hospital, the first target selected was pre-eclampsia:

It is a devastating pregnancy complication that affects one in 10 pregnant women, and it actually causes the death of 50,000 women and 500,000 babies annually worldwide. Five million babies are born prematurity due to slow growth, and this prematurity in itself carries long term, even lifelong, complications for those babies. Even though pre-eclampsia has been known since the times of ancient Egyptians, there is no screening test available for this disease. And the only cure for pre-eclampsia is the delivery of the premature baby.

The important advantage that comes from the ability to classify, identify and count the number of individual proteins riding on a blood platelet is that these markers of disease and other troubles can be found before any symptoms emerge. So while it is not possible to go so far as to identify a predisposition towards any particular illness, it can provide a screening service that shows that disease is present before symptoms emerge.

The second target the Hub team took on was multiple sclerosis, explains Szklanna:

It is an umbrella disorder presenting with different symptoms and having different disease cores in different patients. Therefore, it is very difficult to diagnose and it requires multiple MRI scans, blood tests and spinal taps to eventually diagnose it.

Once again the team set out to identify the novel biomarkers for the disease, then analyse and classify them. This produced similar results to the pre-eclampsia work, producing 96% accuracy when diagnosing multiple sclerosis, with a greatly reduced diagnostic time. Further interrogation of that data found that one of the biomarkers actually correlated with disease severity. Indeed, it became possible actually to predict how severe multiple sclerosis will get for different patients, with nearly 97% accuracy.

The interesting side issue with the team’s approach is that becomes possible to help clinicians with known diseases in both speeding diagnosis and helping them plan and manage their resources in order to provide the right level of care and medical intervention for individual patients, by already knowing which ones are likely to create the most problems. For the first time they can be proactive and prepared in managing each patient’s care, rather than reactive and effectively chasing the development of the disease.

COVID – one disease, several outcomes

According to Szklanna, this is already starting to prove particularly helpful at Dublin’s St James Hospital and its management of COVID-19 patients. The methodology has been used to classify and identify markers not just for COVID itself, but also for the range of serious effects it can have on patients, causing a range of serious, often fatal secondary diseases to add to its prime effect, pneumonia. These include serious liver damage, and blood clotting causing heart attacks, embolisms and deep vein thrombosis:

The COVID-19 journey is very difficult for patients, but it's also difficult for doctors. So any early warning of the onset of one of them would be a valuable, and often life-saving advantage. when a patient condition deteriorates, a whole team of healthcare professionals, including ICU staff, consultant clinicians and nurses have to gather to decide on the best course of action for this patient. And often, this decision is whether or not to admit this patient. This is a very difficult decision to make.

The Hub team started work on this in May during the first UK lockdown, so much of the initial work was conducted remotely. This included significant levels of liaison, technical support and guidance from data scientists working at SAS. The hospital helped recruit COVID-positive patients to the program who could donate blood samples. These were from across the spectrum of symptoms and severity. A group of currently non-severe patients were also recruited and once the COVID marker proteins were identified and classified, the object became to establish if the process could determine the level of severity those patients would suffer. In particular, how many would need ICU treatment?

The data was uploaded to SAS Viya, running in a Microsoft Azure cloud, where it was analysed to identify each patient’s likely level of severity. By September they were achieving quantifiable results. Szklanna says:

Thanks to the collaboration with the SAS team, we were able to identify three key parameters from the routine blood test that can predict ICU admission at day zero of patient care in the hospital, the day of the cold positive swab for a patient. Utilising those three parameters, we have a decision tree classifier and using this we were able to correctly determine ICU admission in 96% of cases. This can potentially be done within a few hours of patient admission to the hospital.

This gives the clinicians a significant head start in planning their treatment programs, such as making appropriate medial interventions far earlier or proactively planning for a known future ICU workload.

Szklanna is full of praise for both the ease of use and intuitive nature of the Viya analytical toolset, not least because it allowed  her to work with her own biochemical skillset and her typical level of self-taught IT skills, without having to first become an expert data scientist:

I understood the principles and the basics of data science, and I was actually able to apply my understanding and my knowledge to my data, without having to spend eternity coding. It's super user-friendly, and super intuitive as well.

As an example, she pointed to the range of pre-coded graphs and models that are available, which allowed her to test out different analyses and then compare the results to establish which gave the best solution.

The next step is, of course, obvious. The Hub is now working with the University’s Technology Transfer Unit to commercialise it and bring it out to the global healthcare marketplace where it can do the most good.