Prediction of stroke outcome using brain imaging machine-learning
Optimizing prediction of stroke outcome using brain imaging machine-learning
Imperial College London
16517, UKCRN ID; 204393, IRAS ID
Duration of Study in the UK
3 years, 0 months, 1 days
Thrombolysis (tPA) is the most efficacious and widely-used treatment for acute ischaemic stroke, but suffers a major complication rate of 2-10%, due to symptomatic intracranial haemorrhage (SICH). Developing methods that improve prediction of SICH can reduce harm, whilst also potentially increasing treatment opportunity for patients previously believed to be unsuitable. Current prognostication methods show only moderate accuracy (60-70% c-statistic) at predicting SICH; and are prone to high inter-rater variability, due to the need for radiological interpretation of acute brain (CT/MRI) images.
Since the challenge of SICH prediction amounts to mapping a large number of high-dimensional inputs (imaging plus clinical variables), onto a dichotomous outcome, the problem lends itself well to computerized "machine learning" approaches that can optimally and automatically classify complex patterns. A well-known example of this technology is face, iris or fingerprint recognition scanners in airports. However machine learning, especially of images requires large datasets to train a model successfully for new cases, explaining why this study requires collection of thousands of previous cases from hospitals around the country.
This project represents a cross-faculty, computer science – stroke medicine collaboration, to see whether machine learning can predict SICH as well as, or better than, current prediction methods. We aim to : 1) develop a computerized algorithm that relates early brain images to outcome (measured both functionally e.g. independence; and whether bleeding i.e. SICH occurred); and 2) test this algorithm on independent datasets.
We envision that the software developed could be used in stroke centres around the world to stratify acute stroke patients for treatment more effectively than existing methods.
Wales REC 3
Date of REC Opinion
23 Nov 2016
Further Information Favourable Opinion