NLP advances to better understand drivers of antibiotic use in veterinary care of companion animal electronic health records.

Antimicrobial resistance (AMR) is a critical challenge for human and animal health that requires a coordinated endeavour across the disciplines of clinical and data science. The project is a collaboration with the Institute of Infection, Veterinary and Ecological Sciences at the University of Liverpool specifically with the SAVSNET group, a veterinary bioinformatics lab who have been collating electronic health records at the point of consultation since 2014. Today, the dataset has amassed over 7 million records with each record containing entries such as age, sex, species, pharmacological prescriptions and crucially a free text narrative for the veterinary practitioners to further detail the consultation. Previous works from this group have arisen from supervised strategies using regular expressions, although this has been successful for highly specific targeting of a singular disease or condition it is not appropriate for a wider analysis. Mr Sean Farrell has joined Dr Noura Al Moubayed lab within the Innovative Computing Group at Durham University to achieve this integration of novel Natural Language Processing solutions and distilling them into the emerging field of veterinary bioinformatics. The project seeks to embody this interface to recognise the features and signals that might be available in a large companion animal clinical records dataset and to develop and apply cutting-edge machine-learning methodologies to derive important insights.
To combat AMR, we need to understand the factors that influence antimicrobial prescription by veterinary clinicians. The project aims to use Natural Language Processes in capitalising on this large dataset which may hold the key to reducing antimicrobial usage and ultimately to stagnate the development of AMR. The project begins with supervised learning approaches, using a generic label applied by veterinary clinicians during the consultations to formulate a tuned BERT model with specific speciality in identifying broad disease classifiers for any veterinary clinical narrative with a potential possibility to no longer requiring veterinary practitioners to self-label. Dependent on the success, the resolution of these classifiers could be increased beyond the scope of these simple labels into indicating specific diseases and conditions. Explainability of the models is therefore important to understand what clinical features resulted in the model selecting a label over another, and theoretically could uncover new symptoms of diseases not previously associated with them. It is important to understand why a veterinary practitioner felt it were necessary to prescribe an antibiotic and whether their decision is justified, it will become apparent if certain diseases are unnecessarily overprescribed over others and may shape future guidelines on when antibiotics are necessary or when they should be avoided. This broader analysis may also uncover new risk factors resulting in disease requiring antimicrobials and can view diseases as a compilation of steps leading to an event rather than considering all infections as being acute.
It is highly important in this increasingly antimicrobial resistance world that any prescription of an antibiotic is both necessary and without other means, it is a possibility that the models produced here may formulate into future tools to provide a second opinion to vets to whether the condition justifies the prescription or whether an alternative solution is more appropriate. Antimicrobials are an invaluable tool that has changed the course of history, however the return to a pre antimicrobial world is a not-so-distant reality. Veterinary practitioners are as equally responsible in maintaining good antimicrobial stewardship as human medicine and we would hope this project can increase awareness and reduce unnecessary prescriptions.