CDT Postgraduate Studentship in Healthcare Innovation (RCUK Digital Economy Programme grant number EP/G036861/1)
Text Prediction for Therapeutic Interfaces
Human computer interaction for patients with severely degraded motor skills including but not limited to ALS, MS, Stroke, and other sufferers is impossible using traditional user interfaces. Companies, techniques, and specialized interfaces have been developed to serve these communities. At the core of all is prediction, in particular prediction of either an intended mouse-like input or an intended text continuation input (not unlike T9 or the text prediction used on a modern smart-phone). There is both formal and practical room to improve the predictive performance of such systems. This project has two key components: one a mathematical component, namely, research to improve the predictive performance of natural language text models; second, engineering research to integrate such models into clinically relevant therapeutic computer interface systems.
A review of a subset of available text-prediction language models has been performed. Research will include understanding, leveraging, and improving these models for therapeutic applications and more. In particular we propose to investigate ways to include contextual information extracted from past discourse to improve conversation-specific prediction (e.g. using past email communication to improve predictive accuracy in an on-going conversation). This is a fundamental machine learning research task formalized as discrete conditional distribution estimation with side knowledge.
Preliminary contact has also been made with both a commercial vendor of therapeutic computer interface software (Applied Human Factors http://newsite.ahf-net.com/) and a single ALS patient who uses said software. A second, applied component of this research project would be to develop the software infrastructure required to translate the research from the first component into a real-world clinical application.