Page 13 - AIH-2-1

Artificial Intelligence in Health AI in higher medical education

radiology presents serious challenges to those engaged study of 44 students aimed at validating VR-based medical
62
in curriculum and syllabus design. Moreover, there is an training, Pedram et al. not only found a user acceptance
75
external dimension of malign intention represented by level of 75% but also an outperformance by those using VR
cybersecurity threats. Medicine is under increasing attack, of the control group that did not. These studies reinforce
and practically no field is more greatly exposed than the view that there is a greater ethical risk in a sluggish
radiology. Cyber-attacks can range from malicious insider implementation of AI in medical education than in a rapid
63
activity to data theft, credential harvesting, and phishing. one. Slow implementation will result in inferior education.
They can occur at various points in the radiological In turn, this will lead to slower and possibly deficient
landscape, including medical devices, wireless systems, deployment of AI in the clinic, with consequently worse
data warehouses, and social networks, and the increasing patient outcomes. While the fast deployment of AI in
use of AI on both sides has created vulnerabilities. medical education will bring lower ethical risk, another
63
Moreover, there is also no clear overview of approved aspect of risk may be avoided, that of the vulnerability of
AI-based medical devices. This leads to inconsistency data. In a clinical setting, real patient data will be used.
and increased ethical risk. However, the problem is In a VR scenario, simulated data are sufficient. Mergen
recognized, and investigations are currently underway by et al. 76,77 have developed a project tool entitled “medical
the Food and Drug Administration in the USA, 64,65 and the tr.AI.ning,” an immersive VR learning platform based on
Medicines and Healthcare Products Regulatory Agency, AI that generates simulated patient data, thus obviating
65
which is developing guidelines for such devices. Here, ethical concerns.
66
5
broadly understood, cybersecurity is an important issue. Regarding ChatGPT and other potential GenText
AI systems are vulnerable to adversarial attacks, such as engines, there are many points of ethical risk in medical
67
the introduction of minor modifications to input data in education. Once more, the output quality depends on
changing training labels that lead to invalid predictions. the input datasets. Very often, data, especially medical
Each such attack is a breach of sensitive patient information, data, is burdened with various types of bias. There is
46
and any wrong decision in the medical field has potentially also a further question of whether ChatGPT is biased as
disastrous consequences. This vulnerability extends not a collection of algorithms or whether algorithmic bias
only to patient data but also to student data. Tsai and could be introduced unethically. If bias can occur at these
78
Lin proposed a procedure to evaluate the resistance of two levels, there is a further systemic ethical threat in the
68
AI models based on medical images against these attacks. vulnerability of GenText and other engines to jailbreak,
There are various techniques to defend against adversarial where an AI system acts outside the restrictions placed on
attacks, including data augmentation, adversarial training, it by its designers. Further alarming consequences may
79
and robust optimization. However, establishing effective arise when an AI that has broken free can create other AIs
protection protocols remains a challenge. 69 that may produce harmful output, such as producing a
In the development of two further contrasting and set of instructions for synthesizing methamphetamine.
80
specific areas of AI, there is evidence of AI being used in However, it is at the day-to-day level that ChatGPT causes
education, VR, and GenText. In the case of VR, although a great deal of concern: at face value, ChatGPT can be
70
some evidence of adoption has been found to be sparse, as used to generate substantial amounts of convincing text.
in the database search by Lie et al. covering November Such text can be used for framework and content infill for
71
and December 2021, a subsequent more extensive literature curricula and syllabi (by course managers and designers),
search study in the period January 2017 to March 2022 teaching material (by educators), and assignments (by
demonstrated a rapid and increasing take-up perhaps in the students). However, where there is a risk that the content
latter part of this period, although this is not stated in the generated is at risk of being out of date (depending at
research. Moreover, students trained using VR produce least on the latency of input protocols and difficulty in
72
better results than those conventionally taught. Kim and ensuring the provision of the latest academic material
Kim identified and examined 24 studies and a sub-group (due to secrecy concerns) and that ChatGPT is capable
73
of 18 on the use versus non-use of VR in medical education of hallucinating with it comes to references, the validity
81
and found that “there was a significant improvement of the output will be variable and at times questionable.
in the VR group’s skill and satisfaction levels, and that In addition, Májovský et al., 2023 considered ChatGTP
82
less immersive VR was more efficacious for knowledge as a tool for the generation of fake medical papers. The
outcomes than fully immersive VR” (ibid., p. 13). Greater whole process took an hour, and it turned out that the text
student satisfaction in using AI is also confirmed by Leng, looked convincing. Although references and specific errors
74
who found that in the case of learning anatomy, ChatGPT raised doubts, these errors could only be detected by an
has increased student engagement. Then, in a small-scale experienced reader, here a medical doctor. This creates a

Volume 2 Issue 1 (2025) 7 doi: 10.36922/aih.3276