Artificial intelligence chatbots: Are they a reliable source for patients and families in pediatric neuro‐oncology?

Abstract

The recent increase in popularity of artificial intelligence (AI) chatbots has brought some new implications to the oncology field. Multiple publications have shown that full papers can be written using these tools,1, 2 and these language models can certainly help physicians get some ideas for writing a scientific article. However, these tools can also be used for getting advice on how to manage patients and certain diseases.3 Also, patients and families are able to use these open-access tools for the purpose of getting help and recommendations for treating their illnesses. Nevertheless, the extent of how much of this information is truly accurate is a matter of debate right now.4 How many treatment recommendations are provided? None, one, or more than one. Are the recommended treatments in concordance with standard clinical practice or nationally accepted treatment protocols? All, some but not all, or none. If “Some but not all” to above, are any correct in their entirety with standard clinical practice or nationally accepted treatment protocols? This question refers to the generality and specificity of the answers provided by ChatGPT. None or at least one. Are any recommended treatments hallucinated? None or at least one. If yes to 4, is the hallucinated treatment now a recommended treatment in the most current versions of treatment protocols? None or at least one. Then, SC compiled all the replies and gave the average result for each question (Table 2 and Figure 1). A full table with results is available upon request. Overall, ChatGPT's recommendations were considered very general, with few specificities. The advice given usually included multiple types of treatment such as surgery, radiotherapy, and chemotherapy, but with scarce references to radiotherapy dosing or chemotherapy drugs. Sixty-one percent of the replies consisted of multiple types of management suggestions, with the other 29% deemed to be a single, sequential treatment proposal. Furthermore, 80% of the answers were contemplated to be in full concordance with the standard clinical practice. Occasionally, there were references to the order in which each treatment needed to be administered (for instance, ChatGPT correctly replied that surgery was the initial step in treating standard-risk pediatric medulloblastoma, and after aiming for complete resection the next step was to give radiotherapy followed by chemotherapy). There were also some striking responses that were hallucinated (14% of the responses), suggesting recommendations not currently accepted worldwide, such as targeted therapies and immunotherapy for treating high-risk medulloblastoma. There were no mentions at all concerning age for giving craniospinal radiotherapy, therefore with no remarks to situations in which radiotherapy replacement with high-dose chemotherapy was considered useful. Regarding germinoma and non-germinomatous germ cell tumors, the absence of references to the correct order in which treatment needs to be given resulted in a hallucinated response. ChatGPT's reply consisted of a list with 5 items separated by bullet points. Thus, the absence of a correct sequence implied that radiotherapy was supposed to be given after surgery, which is not always the standard of care. Besides the content of the responses, we also tested the reading ease score. We calculated the Flesch Kincaid score, with the grade level needed for understanding ChatGPT's responses ranging consistently between college and college graduate levels. It is important to note that in conducting our study we believe that the questions were initially formulated in a very general way. To try to ensure that this was not a bias, on a second occasion, we also asked about specific cases of pediatric brain tumors. We inquired the chatbot about management recommendations for certain patients, but responses were also given in a vague, nonspecific way. Again, ChatGPT's response consisted of a list with several items, with few references to the correct sequence of treatments. Conversely, although at the moment ChatGPT can accurately work as a large language model by providing large amounts of text in relation to a certain question, it cannot provide an evidence-based response. We hold the perspective that the responses generated by the system lack reliability, rendering it unsuitable as a robust platform for acquiring information intended for patients and their families. It could also be harmful to the doctor–patient relationship, giving on some occasions untrue responses, which can confuse families. Nowadays, we live in a digital era, and all the information can be found at the touch of a button, and that is one of the problems we need to deal with as a medical community. As healthcare providers, we believe it is important to send a strong message and to emphasize that medical advice should be given by doctors and health and care workers. Thus, information will be reliable and trustworthy. While it is true that in the future, these models may be helpful for taking medical decisions, at the moment replies given by chatbots would need to be refined so that they can really compete with human-given responses. However, we remain optimistic that advances in these models will occur progressively. With each subsequent update, AI chatbots will likely offer enhanced and more accurate treatment recommendations. The authors declare no conflicts of interest.

Ähnliche Arbeiten