Using machine learning to diagnose disease

Profiling the immune system in paediatric arthritis patients offers hope for improved diagnosis and treatment

A team of scientists from VIB and KU Leuven has developed a machine learning algorithm that identifies children with juvenile arthritis with almost 90% accuracy from a simple blood test. The new findings, published this week in Annals of the Rheumatic Diseases, pave the way for the use of machine learning to improve diagnosis and to predict which juvenile arthritis patients may respond best to different treatment options. The work was led by Professor Adrian Liston, a group leader at the Babraham Institute in Cambridge, UK and at VIB and KU Leuven in Leuven, Belgium.

Juvenile idiopathic arthritis is the most common rheumatic disease in children, but it presents in many different severities and forms. This diversity makes clinical assessment and patient classification difficult.

A team of researchers at Belgian research organisations VIB, KU Leuven and UZ Leuven undertook a detailed biological characterisation of the immune system of hundreds of children with and without juvenile arthritis to help the diagnosis or treatment decisions for this disease.

“Essentially, we took blood samples from more than 100 children, two thirds of whom had childhood arthritis,” explains Erika Van Nieuwenhove (VIB-KU Leuven), and first author of the study. “We analysed their immune system at a greater level of detail than was ever done before for this disease, and simply using this data we then used machine learning to see if we could tell which children had arthritis.”

The results were quite remarkable: the algorithm was about 90% accurate at identifying the children with the disease. “Using only information on the immune system, and no clinical data at all, we could design a machine learning algorithm that was about 90% accurate at spotting which kids had arthritis,” says Professor Adrian Liston (Babraham Institute, Cambridge, UK and VIB-KU Leuven). “This result is a proof-of-principle demonstration that immune phenotyping combined with machine learning holds huge potential to diagnose disease. Similar approaches could be applied to improve patient selection for treatments and clinical trials.”

The researchers are hopeful about the impact of this research in improving patient outcomes. “The tool needs further validation but otherwise there are no scientific barriers to this approach being quickly translated to the clinic,” comments Professor Carine Wouters (UZ Leuven), who was the clinical lead for this study. “Down the line, we could use this kind of detailed classification information—and machine learning analysis—to identify which patients will respond best to specific treatment options.”

Interview with "The Optimist" magazine

Read the article at The Optimist

Thinking back to your PhD, how would you describe this experience.

I quite enjoyed my PhD. The key success in a PhD is to find a match between supervisor and student. I only spoke to my supervisor every 3-4 months, and it was always about concepts and strategy rather than trouble-shooting. For me, I loved the independence that this gave me, and the amazing post-docs in the lab gave me more than enough technical advice. However, some of the other students around me did not like this approach to mentoring, and would have preferred weekly meetings going into the detail of their experiments. This was pure luck on my behalf - I could easily have ended up in a lab that I found stifling, because I didn't ask the right questions going in. This independence let me mould my PhD to my strengths. I learned just a few basic techniques and then applied them to novel questions. It was an approach that let me generate data and papers quickly, and led to an "easy PhD". 

To ask the famous question, is there anything you would like to change retrospectively regarding your PhD?

The flip side of having an "easy PhD" is that I never really had to leave my comfort zone. Since I didn't spend months (or years) painfully learning and optimising new techniques, I never became as technically skilled as the other PhD students around me. Science is so fast moving, that the best strategy is to learn how to learn, which you only get the hard way. Instead, I had my couple of techniques and I had learned how to plan experiments and writing papers. This made my post-doc really difficult - I didn't have the versatility or skill of other post-docs around me, who were picking up and using the latest techniques with trained ease, earned by blood, sweat and tears earlier on. Now as a PI, this deficit is not so important, since my job is all planning and writing, but even now I regret never learning to become a great experimentalist.   

 Which advice or tips would you give us PhDs on our way?

1. Analyse experiments as you go. I started the habit very early on of always analysing experiments once I finished them. By this I mean a full analysis, including a publication level graph, a figure legend and a few lines of text describing the result. It takes a little time, but it means you get real-time feedback on the quality of your experiments - give you have all the right controls, were the numbers high enough to make conclusions, are my conclusions solid enough to plan the next stage, etc. It also made writing papers and a thesis very simple - I just cut and paste my analysed data in, and I was half-way there. Since editing is much less intimidating than writing, I never developed that writing paralysis that some students get.

2. Don't stress about careers! The infamous "bottleneck" in the academic career is mostly illusional. In Flanders, perhaps 15-20% of PhD students go on to a long-term academic career, but even in countries with lower rates (2-5% would be normal) this this not due to a bottleneck. A PhD in biomedical science is one of the most desirable training programs possible for a modern career. The vast majority of people who leave academia are not pushed out; instead, biomed PhDs are leaving academic because of pull-factors - they find highly desirable jobs that they believe they will enjoy more. When I look around at the PhDs that I trained with or that I have trained, I can honestly say that not one has had a career failure. Yes, very few are now research professors, but that is because almost all of them found something else they liked more. Doctor, CEO, start-up company, scientific writer, senior public servant - all great jobs. Very, very few of the 100+ academic careers that I have followed have ended with someone getting pushed out of academia (i.e., timing out of the post-doc fellowship system), and those that did landed on their feet and found a great career that they now say is better suited to them. So.... don't stress about your future career. Concentrate on doing well in your PhD, and start planning your career a year in advance of any decision, but don't make yourself unhappy about uncertainty in which successful career path you will end up taking.

EMBO Young Investigator meeting

Great meeting with great people

Punting on the Cam

Visiting the original lab books of Rosalind Franklin