Dr P-J Noble BSc(hons VetSci), BVM&S, PhD, MRCVS.

Every day, veterinary surgeons across the country write notes about their patients into electronic health records using a variety of practice management systems (PMS). These notes provide a potential resource for developing our understanding of companion animal health but, additionaly, contain the most timely information about changes of clinical syndromes that might occur as part of a disease outbreak. The Small Animal Surveillance Network (SAVSNET) has designed a system that can be implemented in the PMS to allow these clinical notes to be sent immediately to the University of Liverpool for analysis.

To date SAVSNET is collecting data from over 400 veterinary premises from across the UK and the data sent includes the free text that the veterinary surgeon wrote during the patient visit. This text is typed free hand and contains a variety of spelling, grammar and punctuation variations that challenge modern text-mining software but when the software approaches are developed in close collaboration with clinicians conversant in veterinary language, this potent partnerships provides an efficient environment to improve the accuracy of text mining.

SAVSNET has made good use of simple test mining tools such as key word recognition for words like 'Tick' or 'Diarrhoea' but we are also using more complex pattern matching tools to filter records for specific features. For instance in records where the word 'vomiting' is mentioned, this may actually allude to vomiting being absent (e.g. if the vet wrote 'no vomiting seen') to correct for this a more complex filter (a regular expression) is used to screen records. Keyword and regular expression searches are incorporated into our text mining suite (SAVSNET Datalab) which allows students to develop and test regular expressions, select a filtered set of records and then read and manually annotate them for relevance to their given project prior to collating the text-mining findings with other patient data (breed, species, sex, age, geographical data). This has been used for numerous undergraduate and post-graduate projects and publications.

We are in the process of implementing various machine-learning methods into our text mining that include neural word embedding, Neural network analysis and topic modelling.