Background: Patients with rare conditions such as Vasculitis often have many questions about their condition in the early days after diagnosis. They, and their loved ones, are understandably anxious and scared about their prognosis and what their diagnosis will mean for their quality of life. While healthcare providers can answer these questions and reassure them, the opportunities to do this are limited to infrequent appointments where a patient will not always remember or know what to ask. Subsequent online searches can lead to misinformation and contextless data, increasing patient anxiety. One GPA patient told us that he had Googled his condition and was told that he had only months to live. This project addresses the need for a patient-centered solution using generative AI to provide accurate, curated, and empathetic responses without the need to wait for one’s next appointment. However, it is crucial to note that it will never be the intention to provide medical advice. This tool is designed to relay generic information that is not patient specific and can be found in publicly available documents and web pages. To date, most applications of AI in medicine are in areas such as disease detection and diagnosis, personalized treatment, medical imaging, clinical trials and accelerated drug development [1] The goal here is to increase the quality of patient support and doctor-patient engagement, by supporting the patient outside precious clinical interactions.

Objectives: Our objective was to develop and embed an AI-powered chatbot into the Vasculitis Ireland Awareness (VIA) website to provide patients and their families with accessible and trustworthy information about Vasculitis, including symptoms, prognosis, treatments, and support resources. Newly diagnosed patients will usually start by Googling their condition. Unfortunately, search engines can sometimes prioritize unreliable or sponsored content, which can mislead patients and increase anxiety. For example, searches for GPA may emphasize historically high mortality rates without context, creating unnecessary fear. A purpose-built chatbot offers curated, accurate responses that can give a patient or a family member the answers they want in a conversational and more importantly empathetic manner. The information used to train the chatbot can be fully controlled, so the answers returned are limited to only the data used to train the chatbot.

Methods: The chatbot was built using a 3rd party “no-code” AI tool and trained using curated data from reputable vasculitis websites, PDFs, and other resources. Additional features were built using other no-code automation tools and some custom Javascript extensions. The chatbot uses sentiment analysis to detect the user’s mood and tailor its answers. For example, if a user’s question indicates anxiety, the bot recognizes this and provides a reassuring answer about how VIA and other organizations can help, advising the user to contact them directly. If the conversation takes a darker turn and the user exhibits signs of depression or suicidal thoughts, the bot provides links to the Samaritans. The chatbot was initially deployed on a test page of the VIA website. Feedback was collected from board members, patients, and clinicians who tested the chatbot. The chatbot did not collect any personally identifiable information, and users were encouraged not to provide any during their conversation. Performance was monitored through conversation transcripts and direct user interviews to refine its capabilities and identify requirements for additional information or features. There is also the option to use voice input mode on the chatbot, which ensures the bot can be used by those with decreased mobility and those with learning or reading difficulties.

Results: At the time of abstract submission 52 conversations with patients and medical professionals had taken place. The most common topics were: