Metalearning for Health AI

点击翻译按钮获取中文摘要

In its infancy is the development of new AI methods for "learning how to learn"; that is, improving the learning process involved in training very complex models using huge healthcare datasets. Healthcare AI is characterised by having access to very large, mostly "unlabelled" datasets, in the sense that it is infeasible to ask a domain expert to label data - this contrasts with traditional areas of AI (typically imaging), in which labels can be obtained from, for example, crowdsourcing or other means of engaging non-experts. The complexity of healthcare time series are such that often the only labels that exist are "hard end-points", such as the eventual death of a patient - this lack of labels occurs because labelled clinical time series (upwards of 100-variate) is not at all trivial, even for a clinical expert. For example, clinical collaborators find it difficult to characterise the many possible manifestations of disease trajectory, because of the considerable inter- and intra-patient variability that exists in this challenging domain. Further exacerbating the labelling problem is that a particular application may often focus on a disease or a patient group for which very little data have been collected. The fact that a patient may be healing, or attempting to respond to treatment, adds yet further complexity to the labelling problem. Existing methods of unsupervised / semi-supervised / active learning typically perform poorly as a result of the above limitations. While transfer learning has proven successful in many imaging problems (because conventional deep learning methods can boost performance significantly on a small, target task by having access to data from a larger imaging task), it typically fails in the non-imaging healthcare AI domain, due to the complexity and variability of patient time-series data.This theme of research within the doctoral programme will propose novel "oracle-free" methods of active learning, recognising the clinical reality that labels are either entirely absent or that, at best, a clinician can only give an opinion on time-series data form a very small subset of patients, and typically for a small subset of the total number of variables available. Building on our past experience of curriculum learning (ICML 2020, in which committees of "student" complex networks are overseen by "teacher" networks), we will propose student-teacher-principal approaches. This hierarchy corresponds to students learning the required clinical task, while teachers learn transferable summaries of latent representations of the students. (This will initially use the massively-multivariate state of the student networks as input to the teachers, from which latent representations of the students will be learned.) The "principal" is the highest level of the hierarchy, which will oversee the transfer of these latent summaries between healthcare tasks, with the aim of allowing us to exploit the value of large publicly-available and patient-confidential datasets in synchrony. This is a critical new direction for Health AI, where typically every new dataset involves "starting again" ab initio. With such new methods, tasks with similar dynamics (not necessarily from analogous physiological conditions or acquisition modalities) can be identified automatically, and where the principal will encourage learning across teachers trained on tasks of similar dynamics.Secondly, datasets are owned by hospital networks; while we have gained access to data on a hospital-by-hospital basis, this is often only appropriate for constructing proof-of-principle models. For AI to make a real impact at national scale, technologies must be developed for operating on data in a federated manner, transmitting models / model updates between a "hub" and the various "spokes" throughout This project falls within the EPSRC Digital Economies, Healthcare Technologies, and ICT research areas. collaborations.