"To transfer research to clinical practice, we have to engage a wide range of people"

 

There is currently a great deal of hype around artificial intelligence (AI), and the ADaS Lab (Applied Data Science Lab) at the Universitat Oberta de Catalunya (UOC) is working on projects that will soon be of use to patients and help medical professionals be more precise and efficient in their work. Ferran Prados and Jordi Casas Roma, leaders of this group, which is part of the eHealth Center, and members of the university's Faculty of Computer Science, Multimedia and Telecommunications, talked about how the research they are carrying out will impact the provision of healthcare and about the daily challenges they face.

 

What is it that motivates you the most in your research?

There are two main factors that motivate us. Firstly, we like the technical side of things, including image processing, the collection and integration of medical and health data, the design of data pre-processing and machine learning pipelines, the integration of AI models and the assessment and validation of results. Secondly, we are driven by the chance to 'do our bit' to improve the provision of healthcare and, as a result, people's lives. So it is the transfer of our research to the medical sector that gives us great personal satisfaction. 

What are the challenges you face in your work?

There are a number of them. One of the main issues is the lack of data. This is because, in the medical field, organizations do not tend to share them due to the legal and privacy implications. It's especially difficult to obtain data on the control group (i.e. healthy people), which we need to compare with patient data.

Transferring research knowledge to clinical practice is also complicated because it calls for the engagement of a wide range of people, which makes it very difficult to implement some research results.

Lastly, it's not easy to work with multidisciplinary teams and projects with the involvement of widely differing specialists such as computer scientists, physicists, mathematicians, statisticians, doctors, biologists, etc. To progress with the project, it's essential that we all work together, from start to finish, defining the goals and work plans, and assessing the results.

What's the source of the data you use?

Most of the data we're using in our research come from agreements with institutions and hospitals, such as Barcelona's Hospital Clínic or the Agency for Health Quality and Assessment of Catalonia (AQuAS).

When necessary and when they're available, we also use open data, which are relatively easy to access to carry out research. Although things are still far from ideal, there is an increasing number of initiatives on sharing this kind of data, such as different countries' and regions' biobanks, like the pioneering UK Biobank.

A great number of institutions and companies state that they are now working with AI. Do you think that the term is overused?

We're seeing a boom similar to that experienced a few years ago with 'big data'. Firstly, there is very significant interest on the part of companies and institutions in the use and application of AI in many fields, from healthcare to agriculture or driverless vehicles, transport networks, logistics, etc. Secondly, we have society's needs to overcome important challenges, such as the robust and efficient diagnosis and prognosis of diseases, and the automation of complex tasks entailing different levels of decision-making.

A substantial number of models and algorithms are being implemented to resolve a range of issues across all sectors. However, as is always the case with a boom, the term 'AI' is being overused and there tends, in some cases, to be a wish to apply the technology without there being a real need to do so. AI should always be used to satisfactorily resolve a problem affecting people or society. It shouldn't be used simply for the sake of it.

Which of the projects the ADaS Lab is working on is closest to reaching the clinical world and benefiting patients?

The detection and automatic quantification of new multiple sclerosis lesions with our AI tool will soon be able to be used in clinical practice, and will be of great use to neurologists and neuroradiologists in gauging the evolution of people with this disease and deciding upon subsequent treatment. This task currently takes up a lot of time and is not always carried out under optimal conditions, with the negative impact this has for people with multiple sclerosis. The solution designed by the ADaS Lab will cut this time and standardize criteria, benefiting people with this condition and the healthcare system in general.

Do you think that research into neurodegenerative diseases with artificial intelligence tools can change the approach taken and have a positive impact upon patients?

Indeed we do. AI is helping with early diagnosis and individualized patient treatment, and will do so much more in the short and medium term. There are numerous research groups around the world working on improving the diagnosis and treatment of people with neurodegenerative diseases. Progress is slowly being made, but the complexity and our lack of understanding of the human brain, together with a lack of common patterns and open data, make it difficult to achieve results.

Unfortunately, we will still have to wait a few years before we see clinical practice adopting some of the things we can already see in scientific publications.

What have been the results of the project on diagnosing COVID-19 with AI techniques?

After testing many AI models developed and trained to detect COVID-19 based on chest X-rays, we had unexpected and disappointing results: all models display poor robustness with regard to the data employed for training. In other words, a model trained with data from the same medical centre and with a specific X-ray machine works well with data from that same centre, but performance takes an abrupt hit when it has to work with data from a different one.

This means that, at the moment, we don't have a model that works well for a heterogeneous set of data from different centres. This problem is becoming a huge scientific challenge for the AI research community. Our current goal is to design a metamodel that combines results and the interpretation of different models to increase robustness and final performance.