Transfer learning for clinical applications
COLLABORATORS: Prof David Llewellyn (CMH and Turing Fellow), Dr Neil Oxtoby (UCL), Dr Janice Ranson (CMH), Dr Charlotte James (CMH), Razvan Marinescu, (Postdoctoral Researcher in Medical Image Computing, MIT)
IDSAI Research Fellow: Dr Bertrand Nortier
Description: One barrier to the development and implementation of robust unbiased algorithms is the heterogeneous nature of clinical, population-based and experimental data. Attempts to enhance healthcare have been hampered by the use of single datasets to develop algorithms that are overfitted and don’t perform well in other datasets, contexts or populations. Models developed for rare conditions are particularly problematic due to limited training data and opportunities for external validation. Transfer learning may be a solution to these difficulties, allowing for knowledge learnt from one dataset to be transferred to a second, where the data is similar. The primary objective of the project is to investigate the utility of transfer learning for clinical application. The project will assess whether a machine learning model trained to predict two-year dementia risk can be transferred between clinical samples. As a secondary objective the project will evaluate the utility of Disease Knowledge Transfer (https://arxiv.org/abs/1901.03517) to extend the way in which transfer learning can be used with clinical data.
