Interpretable Prediction of Recovery from Stroke with Machine Learning

The PhD project will involve using modern machine learning to improve prediction of stroke patients' deficit and the trajectory that their recovery will take. Prediction will be made on the basis of structural MRIs and behaviour assessments collected shortly after patients' stroke. We will be collaborating with Professor Cathy Price's Ploras team at the Wellcome Centre for Human Neuroimaging at University College London. The Ploras project has provided one of the largest data bases of stroke patients in the world - currently around 1100 patients.Deep learning is beginning to be applied to classifying and predicting stroke patients recovery from structural MRI. However, a clear difficulty with these methods is that they are effectively black-boxes and do not provide easily comprehended explanations of their classification of prediction output. This makes them very difficult to use in practical clinical contexts, in which, for example, a clinician will need to explain their decisions to patients and carers.A first objective for the studentship is to apply emerging interpretable machine learning methods to the Ploras data set. A second objective is to refine these methods to the specific requirements of stroke data. More specifically, the student will build from preliminary findings from a collaboration between the Ploras group and the Turing institute, which has applied deep convolution neural networks to the Ploras data set [Roohani et al; 2018]. This work provided an intriguing result, which suggests that participants who recover exhibit compensatory changes in grey matter volume in homologous regions to their stroke lesion. However, without further understanding of how the deep convolution neural network arrives at its classification, this finding remains intriguing, but difficult to "sell" to the stroke community.This student's project will initially focus on the Roohani et al finding and apply methods such as, Layer-wise Relevance Propagation, sensitivity analysis, decision trees, as well as methods arising from neuro-symbolic research, where the supervisor has expertise (Besold et al; 2017). This investigation will provide insight into the developments that need to be made of current machine learning methods.Other lines of research will be applying variational Bayesian methods, where our collaborator Friston at the Wellcome Center for Human Neuroimaging can provide expertise [Friston et al; 2017]. These methods involve fitting Bayesian networks, from which an explanation of a classification or prediction can be extracted. The manner in which Friston's variational framework applies to stroke data will require foundational development of the framework, in order to respond to the limitations arising from the mean field and Laplace (Gaussian) approximations. This could involve use of resampling techniques during the fitting of the approximate posterior distribution.