PLORAS challenge: Predicting language outcome and recovery after stroke

Introduction

Globally, stroke is the second most common cause of death, and the third most common cause of disability. More people in the UK are surviving after experiencing a stroke, but many leave hospital with complex difficulties with communicating and understanding language, both written and spoken.

The Predicting Language Outcome and Recovery After Stroke (PLORAS) project funded by Wellcome and based at the UCL Institute of Neurology is a database of post-stroke structural MRI scans, and demographic and behavioural data. It aims to empower stroke survivors and clinicians with more and better information about a patient’s likely course of recovery. Using the demographic data and scans that the PLORAS team provided, Turing Data Study Group researchers were challenged to develop a clinical tool that could accurately predict whether and how quickly individual patients might recover language skills impaired after stroke. A breakthrough in this domain would allow clinicians to create better personalised treatment plans for each survivor and gain deeper scientific understanding of how language works in the brain and how the brain adapts to damage.

Interview

Dr Tom Hope explains: “Someone suffers a stroke roughly every three minutes in the UK alone, and a lot of stroke survivors suffer language difficulties, which lead to social isolation and depression and make it difficult to return to work. These patients want to know whether and when they might hope to recover, but current medicine cannot answer those questions. This is a problem both for the patients and for the health system itself, because we currently have no way to predict what ongoing care or support people will need. And because we don’t know which patients will or can recover spontaneously (without much therapy), we also cannot test therapeutic interventions properly. To start to answer these questions, you’ve got to collect the right data, and that’s what the PLORAS project is doing.”

A diverse team of 15 researchers with backgrounds in quantitative science, machine learning, glaciology, medical imaging, chemistry, biology, data science, computer science and neuroscience worked intensively on the challenge over the five days. To Tom’s great satisfaction they were able to deliver promising results with the potential to lead to some important clinical and scientific breakthroughs. Reflecting back on the week Tom says:

“I’ve been researching this area since 2011 and was sceptical that in a week they would be able to get to something I haven’t already tried. I also have a background in AI, so I know quite a bit about machine learning. This was helpful as it meant I could direct them away from things I had already tried, which shifted their focus.”

This approach worked and led the researchers to bypass the brain data that Tom had been using for many years consisting of three dimensional legion images which showed where the patient’s brain had been damaged. He explains: “They decided to use the scans of the brains instead (with no explicit measurement of what regions were damaged) and this led to a result that seems better than anything I’ve seen before. They used a deep neural network to encode the most important features from the scans, then used those features to make what looked like highly accurate predictions of patients’ language outcomes after a stroke.

“At this point further work needs to be done on this to ensure that the method and results are solid, but it is definitely worth pursuing. If the predictions are really as good as they seem, we will be closer than ever to a tool that can actually be used in practice.”

“A researcher opted to pursue an approach I thought was a mad idea, not likely to work. But I was proved wrong.”

Dr Tom Hope, UCL

Another positive result emerging from the challenge may contribute to providing answers to the scientific aspects relating to how language is organised in the brain. Toms says: “A lot of researchers have claimed that language is a lot less complicated than it seems: that all of the apparently separable skills like reading and writing still broadly get impaired and recover together, in concert. One of the group’s researchers produced a very simple but powerful demonstration that this is wrong: that more complex language skills (like understanding a sentence) recover in a different way to more simple skills (like repeating a single word). I’m not really surprised that this true, but I am gratified that we can now show that it’s true.

“When this researcher opted to pursue an approach which explains rather than predicts I thought it was a mad idea, just not likely to work or produce results we could really use. But I was proved wrong and we will explore this further in a paper.”

These were two key outcomes from the Data Study Group that Tom will follow up, but he says there were actually five leads that could potentially be explored further.

Summing up and thinking back to the highlight of the week he says: “The week was very worthwhile and exceeded my expectations. It was also a personal education for me where I learned how to direct and guide people and how to explain and prepare the data in a way that non-experts could understand.

“The key highlight was on day two when they stopped talking. That was when I knew that they understood the data well enough and had started to properly engage with the problem. At that point I felt confident that something constructive was happening and that they might produce something.”

In terms of advice to any organisation or individual considering bringing a challenged to Turing Data Study Group, Tom says: “Do it, but definitely assign someone to be there all day every day. Also find a version of the data that minimises any technical restrictions (like confidentiality), this will make it easier.”