Use of artificial intelligence for veterinary triage: a promising tool but with limitations in the emergency setting

IN recent years, concepts once only spoken of in the context of science fiction have rapidly become a reality in our daily lives. For example, asking a computer to link ideas and draw conclusions from the data entered is something we used to see only on Star Trek. However, the use of artificial intelligence (AI) now forms part of many people's daily routines. With the increasing use of AI in everything from drafting emails to data searches and summarising research papers, AI has inevitably made its way into the veterinary medical field. Indeed, a quick search of the veterinary literature shows AI being used for a wide range of applications, including radiograph interpretation,1 development of diagnostic methods in oncology,2 screening for endocrine disease3 and medical record transcription.4 In each of these situations, the use of AI has strong potential to improve or augment existing human-based techniques, but it also has limitations and still requires human oversight. Research in the field of human medicine has demonstrated the potential of AI tools to support triage processes. However, the use of AI tools in veterinary triage has been little investigated. To this end, a recent study by Wong and colleagues, summarised on p 84 of this issue of Vet Record, evaluated the performance of AI models (specifically ChatGPT-3.5 and ChatGPT-4.0) for triaging canine emergency cases.5 The findings of this study raise several concerns, the primary of which is that if the AI model misses between 7 and 17 per cent of cases that require immediate evaluation and is no more accurate than a trained veterinary nurse, then there does not seem to be a strong reason to use it in place of standard triage techniques. In response to this, Wong and colleagues argue that AI's best place might be as a complementary tool to augment veterinary nurses' triage and improve detection of the most urgent cases.5 A further concern is that Wong and colleagues' findings5 would be very difficult to apply to a real-world emergency clinic. In contrast to a real-world setting, where rapid decision making is crucial, the individuals retrospectively performing triage in this investigation did not have a time limit; this immediately calls the findings into question. Would these individuals have triaged differently ‘in the moment’ than they did when retrospectively evaluating information from the medical records? Additionally, the veterinarians and veterinary nurses participating in this study were operating from a set of standard written findings.5 In a real-world setting, individuals may evaluate clinical findings differently (eg, a dog might be deemed obtunded by one operator and quiet-alert-responsive by another). Also, factors such as how the patient interacts with their environment and the history being given simultaneously by the owner during triage may influence clinical assessments. However, these factors could not be considered in this study due to its retrospective nature. Finally, it is unclear how AI triage would be applied in a real-world setting. Would a client enter data into the computer, followed by further data entry from a triage staff member? Would a veterinary nurse perform triage and then enter data to see whether the AI agreed with them? In cases that require immediate treatment, such as animals in respiratory distress, the time required to manually enter the information may potentially have severe consequences for the animal's welfare. In conclusion, the AI models used in Wong and colleagues' paper do not appear to significantly improve triage performance relative to triage by trained veterinary professionals.5 I also believe that the interface currently required to use the AI (ie, typing data into a questionnaire) is not practical within a busy emergency clinic; the time taken to enter the data could literally be the difference between life and death for a severely sick animal. Does this mean that AI will not continue to evolve and that a functional holographic doctor (straight out of Star Trek) who can assess a patient and apply an algorithm to not only triage but also treat a patient is not on the horizon? No. However, at this point in time, while AI has potential, it is not yet a tool that is ready for routine use in veterinary emergency medicine.