Prediction Modelling for Clinical Medicine Symposium – a PhD perspective
Elsie Horne, PhD Student based at the University of Edinburgh
Held on Tuesday 24th April, the Prediction Modelling for Clinical Medicine Symposium was organised and chaired by Luke Daines, a fellow Asthma UK Centre for Applied Research (AUKCAR) PhD student. The symposium was held in Edinburgh, and while it was well attended by Edinburgh colleagues it also attracted attendees and speakers with special interests in prediction modelling from across the UK and Ireland.
In the first session, we heard from three statisticians from the University of Edinburgh’s Usher Institute. Professor Steff Lewis’ talk on statistical models for prediction concluded with an important message: always use your brain and don’t let an analysis package tell you what to think! Dr Stephanie Read followed by outlining seven steps for developing valid prediction models. The session was rounded off with some useful examples of model evaluation given by Dr Rebecca Pillinger.
The second session focused on two highly topical areas in medical research: the use of routine clinical data for research and machine learning. Dr Tim Wilkinson described some of the advantages and challenges of using routine clinical data when deriving prediction models. A key message from Dr Wilkinson was that clinical prediction models should be built on data that is easily obtainable in a clinical setting, rather than relying on expensive or intrusive measurements that are rarely collected in routine care. Dr Chris Newby then compared and contrasted some traditional statistical techniques with those from machine learning.
The first two sessions were followed by a lively panel discussion surrounding statistical techniques, and how these may have to be adapted to big data and to address challenges of routine clinical data. It was great to see that this discussion was by no means dominated by the statisticians, and many clinicians had queries and suggestions on how these challenges could be addressed. An example was the issue of missing data, and the fact that clinical insight into why these data may be missing is crucial to how it is handled in statistical models.
The final session featured two clinicians, both with a wealth of experience implementing and evaluating prediction models in clinical practice. Professor Keith Fox used the GRACE risk score to illustrate the risk-treatment paradox in cardiology (in which high-risk patients are less likely to receive treatment). Professor Fox stressed that while a risk score can be used to identify high-risk individuals, they are equally important as tools to avoid over-treating low risk individuals. This idea was built on by Professor Tom Fahey, as he discussed the challenge of applying thresholds to risk scores in order to achieve an acceptable balance between sensitivity and specificity.
This symposium kicked off from a statistical viewpoint, and rounded off with a clinical perspective. However, every talk took both statistical and clinical issues into consideration. The take home message for me was that the development of robust clinical decision rules relies heavily on a close collaboration between statisticians and clinicians. Professor Fahey also highlighted the fact that communicating risk score with patients is a key step in the process of shared decision making, and therefore their interpretation of the score must also be taken into consideration during development.
The symposium was an excellent opportunity for clinicians, statisticians and other researchers to share ideas on the topic of clinical prediction modelling. It was particularly relevant to me, as my PhD project concerns both routine clinical data and the prediction of asthma outcomes. I finished the day with a number of fresh ideas as well as new contacts. Along with my fellow attendees, I’d like to thank Luke for organising such a fantastic and informative symposium, and the Asthma UK Centre for Applied Research for providing the funding.