Vit Novacek. AI Research Lead at Data Science Institute, NUI Galway
- What is the role of the NUI within the Clarify project?
We are responsible for the AI and Machine Learning backbone for the predictive models in CLARIFY, working closely with our partners from Accenture and UCL. The key contribution we hope to bring to the field is using knowledge graph techniques and statistical relational learning for computing the predictions and augmenting the predictive models with bespoke explainability features. Ultimately, this is envisioned to support clinicians in making truly informed and evidence-based decisions when it comes to patient stratification based on their risk of developing long-term complications of oncological treatments.
- Can you tell us about the research you are currently working on?
The methods we are exploring now are standard survival analysis models, off-the-shelf binary classification techniques and some cutting-edge statistical relational learning (i.e. link prediction) methods, all applied to predicting the risk of lung cancer relapse. What we’re most interested in is how these methods can be combined, or stacked up, so we could deliver state of the art results in this area. The next step is to augment the predictive models with explanation capabilities and to test them with more diseases and their complications. To this end, it will be crucial to incorporate research outcomes from other partners such as UCD (mechanistic predictive models based on systems biology), UPM (data mining pipelines) and TIB (knowledge graphs integrating the patient data with open linked data).
- In your opinion, why is this research important?
Even if we manage to achieve our least ambitious goals, it will still leave the clinicians with a brand new set of methods they can use for making their decisions anchored in a much broader biomedical context than possible before. And if we do anything beyond that baseline, we’re already talking about immediate and tangible impact on the quality of life of cancer patients due to targeted screening for possible complications and more timely prevention of those risks. And then there is yet another, sort of cultural dimension to CLARIFY – it is a perfect vehicle for bridging the gap between the clinicians and computer scientists who have much to offer to each other, but still seldom truly work together on addressing important challenges in oncology and beyond. The benefits of this may not be so visible during the project itself, but I’m sure the experience will live on and flourish within the follow-up work of all the partners, which is something I find rather exciting.
- What applications do you foresee for your research?
I’d love to see our AI toys saving lives in the clinic, period. I realise it may still be a long way towards routine and provably game-changing adoption of predictive models like those we’re working on in the daily life of clinicians, not the least due to various regulatory challenges that need to be overcome even if the techniques themselves are efficient and reliable. But if we can contribute to that process, no matter how little, I would be very happy.
- How did you get specialized in this area? What would you say to the new generations?
Already since my early PhD stages I saw life sciences as an interesting and challenging field for testing my AI research ideas “in the wild”, and I suppose I’ve just kept learning how to do it better ever since. The only major change is that the more I know about biology and medicine the more humble and sometimes even desperate I become. But it doesn’t prevent me from doggedly trying on and on to develop some truly actionable piece of biomedical AI one day. And things like CLARIFY that have happened along the way are perhaps a proof that I’m not doing it completely wrong. So, my advice to someone just starting off would be follow your passion full on but prepare to be patient as well. The field has been taking off for a while, but we’re still far from making a truly major impact. Also, if you’re a computer scientist, talk to the doctors (and the other way around). The sooner you do that the better but be prepared to be humiliated before you really start getting each other. I’d say it’s only natural, and it sucks, but the rewards can be amazing if you survive feeling like a powerless fool most of the time (I’ll never forget the feeling when I learned some of our predictions in protein signalling got confirmed in living human cells – that was one hell of a kick for a computer scientists who hadn’t quite believed in the actual power of his air castles).