Page 8 - AIH-2-1

Artificial Intelligence in Health AI in higher medical education

One important area in healthcare involves electronic population or reflects historical biases and inequalities,
health records, which can serve as input data for AI and be AI can learn and perpetuate these biases. For example,
processed quickly. However, such datasets not only contain a language model trained on text from certain online
sensitive content but also constitute ethical risks, especially communities may accidentally learn and replicate the
when data collection is subject to various forms of bias biases expressed in that community. A lack of diversity in
2
and is exposed to a large number of hostile attacks. More the ethical standpoint of AI researchers may also contribute
3
concerning is that most medical researchers treat AI as to bias issues. Moreover, the algorithms themselves may
a black box, leaving its ethical risks concealed. A strong introduce or amplify algorithmic errors due to their
4
argument can be made that the successful application of AI inherent operational principles.
in medical practice will depend on addressing legitimate Another challenge related to data is security. Compared
concerns about misunderstandings of its principles and data to traditional statistical methods, AI-based algorithms are
5
risks, in accordance with evolving bioethical principles. In more susceptible to adversarial attacks that exploit security
a field such as medicine, which is critically related to issues vulnerabilities, such as sensitivity to even low noise in the
of life and health, it is particularly important to explain input data. Traditional methods are more deterministic,
6
10
and address the impact of AI on its essence and principles making them more resistant to such attacks.
in medical educational programs, both in terms of how it
works and its underlying ethical assumptions. For medical In this paper, we analyze the technical and ethical
practitioners to use AI-based solutions effectively in their risks associated with certain AI applications in medical
work, they must first learn how to use them correctly education, exploring the potential benefits and risks of
during their training. these technologies in practice, the awareness of students
and practitioners regarding these issues, and the latest
Moreover, AI-based solutions may be more vulnerable scientific research in this area.
to attacks compared to other approaches, such as statistical
methods. It is also worth stressing that, especially in the 2. An overview of current research activity
7
field of medicine, deep neural networks with many layers in AI, medical education, and ethics
(such as highly complex architectures) are commonly
applied. This may contribute to AI models being more In this paper, we conducted a systematic review of
susceptible to overfitting, where the neural network research on AI, medical education, and ethics based on
memorizes the training data rather than generalizing from the PRISMA academic review process and its extensions,
11
it. In this context, statistical methods are composed of including PRISMA-S. Resources written in English from
simpler models with fewer parameters, which may lead to the Web of Science (WoS) database were considered,
easier interpretation of the model. 8 excluding PhD theses and any material not related to AI or
education. Our searches for the terms “AI,” “education,” and
A significant limitation of AI is its dependence on data. “medicine” yielded 488 resources, of which 34 addressed
9
In particular, the essence of AI, comprising algorithms ethical issues. Figure 1 presents the participation rate in
for learning complex patterns and making accurate % of individual areas of the world in research relating to
predictions, has a core sensitivity feature: the quality and AI in medical education (Figure 1A) and AI in medical
representativeness of the training data. Inaccuracies in the education, taking into account ethical issues (Figure 1B).
training data significantly affect the efficiency and accuracy These results highlight both the very low participation of
of the results obtained, potentially skewing outcomes and low-income countries in research and a lack of focus on
leading to ethical consequences that oppose the institution’s ethics. However, the study also included searches involving
goal. Indeed, it can be said that the quality and output of the search terms “artificial intelligence,” “medicine,” and
AI algorithms are directly dependent on the medical data “ethics” (AI+med+ethics), which yielded 328 results, giving
used to develop, test, and validate them. Therefore, a key a higher result when education as a whole is considered.
issue in using AI in medicine is the reliability of biomedical The sources included were selected to answer the research
data obtained from patients, which must be compiled and question, “What multi-criteria impact will AI have on
categorized in an ethical manner. Unlike AI-based models, higher education in the field of medicine?” First, duplicate
statistical methods can work with smaller datasets, and the records in the database were excluded. In the second step,
optimal selection of data may help minimize data errors records whose titles and abstracts were not related to the
more efficiently. The heavy data dependence on AI-based subject of the analysis were excluded. Then, records that
solutions also makes them vulnerable to developing were not accessible were disabled. In the final stage of the
learning patterns based on biased and faulty training data. search, records without information concerning the topic
If the input data is not representative of the real-world of consideration were excluded from the analysis. Finally,

Volume 2 Issue 1 (2025) 2 doi: 10.36922/aih.3276