Aidan MANNION

ABSTRACT:
The widespread adoption of Electronic Health Records (EHRs) in clinical databases, alongside the rapid advances in machine learning (ML) and natural language processing (NLP) technologies in recent years, has expanded the opportunities for innovative and exploratory applications of sophisticated computational techniques to provide practical aid to medical practitioners. This thesis explores the potential applicability of automated decision support tools to unscheduled general consultations, an area which remains under-explored by the medical AI research community. In the context of a multi-disciplinary project uniting NLP research, EHR system developers, and medical practitioners, we bridge the gap between biomedicaL NLP research and general medical practice by using a database of SOS Médecins consultation records to build, evaluate and implement two different ML-based tools that are capable of Carrying out real-time predictive inference based on textual information entered by the practitioner during a consultation. Beginning in an NLP research framework, we first developed and evaluated various biomedical BERT models, using a variety of information sources including scientific text, structured ontological knowledge from the UMLS metathesaurus, and proprietary text from SOS Médecins consultation records. We then adapted these models to two novel document classification applications, designed in collaboration with a medical practitioner: 1) binary classification for hospitalisation and/or serious diagnosis risk, and 2) diagnosis prediction. Having developed and validated these classification algorithms, we then integrated them into the SOMELOG clinical data management software used by a number of SOS Médecins practices. This enabled us to evaluate the gap between experimental evaluation and practical utility, and also to collect feedback on the perception of this kind of tool among practitioners. We find that many technical and conceptual challenges remain to be addressed before technologies like these can be fully integrated into generalist clinical workflows in a useful way. In summary, the main contributions of this work are 1) software tools and language models adapted for the domain of unscheduled general consultations and 2) quantitative studies of the applicability of artificial intelligence tools for clinical decision support in the SOS Médecins context. We hope that our work will provide useful experimental frameworks and practical insights for future work in this area.