Mr Matt Hacket1,2, Kirsty Sim1, Duangmanee (Fon) Seedam1, Adithi Ramachandra1, Katherine Berry1, Stephanie Curtis1,2, Pauline Bass1, Andrew Stewardson1,2, Mr Philip Rawson-Harris1,2
1Alfred Health, Melbourne, Australia, 2Monash University, Melbourne, Australia
Biography:
Matt is a HIV Data Analyst in Infectious Diseases at Alfred Health whilst supporting Infection Prevention and is specialising in AI.
Kirsty is a AMS Data Analyst in Infectious Diseases at Alfred Health. She has a background in pharmacy and Computer Science.
Abstract:
Introduction
Surveillance and auditing of reportable healthcare-associated infections are likely to benefit from artificial intelligence (AI), but early adopters face multiple challenges and nascent external guidance.
Methods
We are in the exploratory phase of developing an AI Clostridioides difficile surveillance tool to automate some of the mandated external infection prevention reporting. Five challenges experienced during this work were documented for discussion.
Results
(1) Organisational governance frameworks for AI are being developed and cybersecurity structures outlined in parallel with this project. Whilst allowing us to consult on these items it was a rate limiting step. (2) Application of natural language processing (NLP) libraries to clinical shorthand (e.g. “BO++”) was hard for the model to recognise. (3) Jurisdictional guidelines require infections to be “Reviewed by a nurse”. There is discussion around whether this includes AI models built on nurse created training sets. (4) Automation bias was expected if an augmented AI classification tool was used to suggest classifications for each isolate to nurses. Instead, an AI model deciding the classification of most cases was chosen, with a human-in-the-loop reviewing smaller subsets and lower confidence classifications to check validity/minimise drift. (5) The costs and benefits of implementing NLP/AI surveillance need to be assessed against existing available methods.
Conclusion
The development of AI models is en vogue, however governance, and guidelines are playing catchup. Common issues such as positive reinforcement and NLP identification were experienced whilst the upskill and access to more complex traditional data programming provides us the constant question about use.