The use of a chatbot based on artificial intelligence to collect patients reported outcomes in oncology practice. A pilot study with real-world data.

Abstract

e19119 Background: There is growing interest in enhancing symptoms monitoring during routine cancer care using patient-reported outcomes (PROs). The quality of evidence that demonstrates clinical benefits of this type of assistance is also increasing. However, the best method for this approach is evolving and there are many barriers to implement these tools in the real world scenario. We aim to describe a pilot study about a chatbot with artificial intelligence developed to collect PROs and to optimize adherence to systemic cancer treatment. Methods: This is a case series that reports the first patients in a private oncology clinic in Belo Horizonte, Southeast Brazil who consecutively underwent regular conversations with a chatbot based on artificial intelligence. Data was collected from February the 23th to December the 3rd 2019. The virtual assistant interacted with the patient in four different ways, being the first three of these an active search from the chatbot: a) search for adverse events, b) adherence to oral treatment c) screening for depression d) spontaneous patient demand (not an active search). Results: Interaction with the chatbot was offered to 193 patients. A total of 107 patients were included. Of these 74% were female, 55% older than 60 years and 22% had at least 7 years education.194 protocols of treatment were analysed, 66% of these being chemotherapy regimens and 23% hormone therapy. Oral drugs corresponded to 23% of the protocols. The main adverse events reported were fatigue 20%, nausea 13%, pain 11%, diarrhea 10%, lack of appetite 6%. Adverse effects were classified by patients as grade III or IV approximately 24% of the time. For patient safety the system runs a script twice a day to detect any adverse effect and send it to the service attended. A total of 3883 dialogues were initiated, the majority of which (3772) was carried out by the machine. Only 3% of the dialogues were spontaneously initiated by the patients. Once the conversation started, adherence was considerably satisfactory since engagement was 73% for questions about adherence to oral medications and 76% of people reported at least one adverse event. Conclusions: An initial barrier must be surpassed since the chatbot was offered to 193 patients and 86 (44%) did not register for use. Once the contact started, we understand that the use of AI is promising since the engagement rates were very good. It is important to highlight the potential capacity for early identification of symptoms since most dialogues were initiated by the virtual assistant.