Page 61 - AIH-1-3
P. 61

Artificial Intelligence in Health                                       ChatGPT in writing scientific articles



            (cardiology, oncology, and remote therapy). We evaluated   •   It is necessary to indicate the exact number of references
            the result of ChatGPT’s work on two aspects: (i) The   •   It is necessary to specify that ChatGPT can use the
            quality of the content obtained and (ii) the reliability of the   Internet  to  search  for  references.  Although  only
            literature references cited by ChatGPT.               version  4 has this feature, to achieve equal initial
                                                                  conditions  for  both  versions  of  ChatGPT,  a  similar
            2. Data and methods                                   addition was used with version 3.

            2.1. ChatGPT versions in use                         It is also important to add that a new chat was created to
            In this article, versions of ChatGPT 3.5, released on   generate a response to each prompt so that ChatGPT would
            May 28, 2020, and ChatGPT 4, released on March 14,   not be tied to its past responses. No custom instructions
            2023, were used.  Compared to the 45 terabytes of data   or other custom settings were used with ChatGPT. Third-
                         32
            for GPT-3, the latest iteration of OpenAI’s GPT-4 has   party plugins and other experimental new features were
            1 petabyte of training data. 33,34  As a result, GPT-4 can   disabled for ChatGPT 4.
            produce significantly more accurate results than GPT-3.   Table 1 presents examples of prompts for two versions
            In additionally the latest version of the neural network   of ChatGPT, using the case of writing a scientific article on
            has about 100 trillion parameters, compared to 175 billion   the topic of biotelemetry in cardiology. To study the impact
            for GPT-3.  This feature has enabled the creation of more   of prompt formulation on the sources obtained, three types
                    35
            precise  formulations  and  coherent  text,  which  is one  of   of prompts were formulated, each of which complements
            the key parameters for evaluating the performance of the   and refines the previous type.
            neural network.
                                                               2.4. Parameters for the evaluation of articles
            2.2. Branches of medicine
                                                               Literature sources for the generated articles were considered
            According to 2022 statistics, the primary causes of   according to the following criteria:
            mortality among the population in the United States of   •   Reality of the source: This criterion assesses the
            America (USA) are cardiovascular diseases, cancer, and   correspondence of the title, author, and journal to the
            COVID-19 infections.  Specifically, about 699,000 people   real scientific publication. The search was conducted
                              36
            died from conditions related to the cardiovascular system.   on the largest academic databases
            Similar statistics were reported for 2020 in Europe, where   •   Impact  factor  (IF) of  the  journal: The  value  of
            out of 5.18 million deaths, 33% were  due to circulatory   this parameter was obtained using the platform
            diseases. 37,38  For  this reason, fields such as cardiology,   scijournal.org. 39
            oncology, and clinical medicine, particularly in therapy,
            were chosen to explore the potential of writing scientific   The assessment of the reality of the source requires
            publications using a neural network. The following specific   a separate explanation. We categorized all the sources
            topics were studied:                               obtained into three major groups:
            •   “Biotelemetry in cardiology”                   (i)  Reliable sources (“blue”): These sources are fully
            •   “Biotelemetry in oncology”                        reliable sources with correct authors, titles, and
            •   “Remote medical examination.”
                                                               Table 1. Prompts on the topic of biotelemetry in cardiology
            2.3. Prompts used and their engineering features   for the two versions of ChatGPT
            The primary interface for interacting with the ChatGPT   No       Prompt for GPT-3 and GPT-4
            chatbot involves submitting a prompt that entails a detailed   1  Write on your own a short research paper on “biotelemetry in
            description of the requirements for the neural network.   cardiology.” Do not send me an outline, I need a paper. Be sure
            In the study, specific prompts were selected for each of   to add references to sources as you write. The total number of
            the chosen fields of medicine. Each of those prompts was   sources should be at least 10.
            tailored for the ChatGPT versions being used. To generate a   2  Write on your own a short research paper on “biotelemetry in
            prompt for ChatGPT, the following points were considered:  cardiology.” Do not send me an outline, I need a paper. Be sure
            •   It is necessary to write a research paper          to add references to sources as you write. The total number of
                                                                   sources should be at least 10. Use only real existing sources
            •   It is necessary to write a research paper “on your own”  (you can search for Internet resources).
            •   It is necessary to write a research paper that is not an   3  Write on your own a short research paper on “biotelemetry in
               outline for this paper, and the paper should be short;   cardiology.” Do not send me an outline, I need a paper. Be sure
               otherwise, ChatGPT outputs only a part of a paper   to add references to sources as you write. The total number of
            •   It is necessary to include references – without specifying   sources should be at least 10. Use only real existing sources
               this phrase, ChatGPT might not include them         (you can search for Internet resources) with a high-impact factor.


            Volume 1 Issue 3 (2024)                         55                               doi: 10.36922/aih.2592
   56   57   58   59   60   61   62   63   64   65   66