Towards interpretable approaches for online healthcare consultation

Student thesis: PhD Thesis

Abstract

Healthcare consultation on the Internet has become an emerging phenomenon. Over the past decades, computerized techniques to automatically understand medical information (smart healthcare) have attracted significant attention. Despite increasing scientific evolution of computer algorithms, solutions given by prevailing machine learning based methods remain as "black boxes" to human experts. In addition to performance measures such as accuracy, precision, and recall, the interpretability of a model is also important, especially in medical domain. This thesis aims to seek comprehensive methods to achieve smart healthcare with both reasoning ability and interpretability. Two specific research problems are formed under the online healthcare consultation scenario: automatic clinical triage and automatic question answering (QA). Specifically, for automatic clinical triage, the benefits of rule-based classifiers have been demonstrated and a novel hybrid method combining both rule-based and neural network-based techniques is developed for effective and interpretable medical text classifications. For automatic question answering, a novel context-related representation technique whose basis terms can be updated by human experts is devised and a multi-level coarse-to-fine sentence matching framework is developed to provide standard answer suggestions to online doctors. In summary, the works described in this thesis incorporate artificial intelligence (AI) based computerized techniques into existing online healthcare consultation systems with high-quality and interpretable solutions. In addition to academic contributions, this research is also of high practical and business values. The proposed models have been successfully deployed to real-world online healthcare consultation applications and demonstrate promising performances in large-scale industry practice.
Date of AwardJul 2022
Original languageEnglish
Awarding Institution
  • University of Nottingham
SupervisorRuibin Bai (Supervisor), Zheng LU (Supervisor) & Uwe Aickelin (Supervisor)

Keywords

  • Natural Language Processing
  • medical text mining
  • interpretability

Cite this

'