Page 713 - ISC PROCEEDINGS 21.4

to pharmacy education [1, 2]. Pharmacoepidemiology helps evaluate the real-world safety
and effectiveness of medicines beyond controlled clinical trials [3]. The
pharmacoepidemiological indicator system encompasses adjusted Odds Ratio (aOR), Risk
Ratio (RR), Hazard Ratio (HR), Incidence Rate Ratio (IRR),
Absolute Risk Reduction (ARR), Number Needed to Treat (NNT), and the
heterogeneity index I². These indicators differ in meaning, calculation, and application,
which can make them difficult for students to interpret [4].
A direct survey conducted at Thanh Do University revealed a competency gap: only
27% of fourth-year students correctly interpreted the meaning of aOR/IRR, 19% could
distinguish the statistical differences among aOR, RR, HR, and IRR, and only 11% were
able to calculate or interpret NNT/ARR in the context of a clinical trial. These figures fall
substantially below the competency level expected in the learning outcomes of the
Pharmacoepidemiology course. This gap suggests that many students encounter
difficulties when interpreting pharmacoepidemiological indicators in real research
contexts. Three interrelated barriers appear to underlie this gap. Most
pharmacoepidemiological literature is published in technical English, posing an immediate
language challenge. Students also lack sufficient biostatistical grounding to work through
regression-based models. Beyond these two cognitive obstacles, conventional classroom
formats rarely provide enough guided practice across multiple study designs.
The emergence of generative AI chatbots such as ChatGPT (OpenAI) and Gemini
(Google) represents a breakthrough in accessible learning support tools. Unlike basic
translation tools, chatbots such as ChatGPT and Gemini can walk students through multi-
step reasoning, adjust explanatory depth to the learner's level, and respond to follow-up
questions in real time — functions that closely approximate the role of an on-demand
tutor [6, 7]. Kung et al. (2023) demonstrated that ChatGPT achieved passing scores on
USMLE Steps 1-3, with approximately 60% accuracy, without specialized fine-tuning [8].
Nori et al. (2023) further showed that GPT-4 achieved performance comparable to or
exceeding specialized medical models on the MedQA and PubMedQA benchmarks,
establishing a foundation for evaluating large language models in medical education [9].
However, evidence on the use of AI chatbots to teach pharmacoepidemiological indicator
interpretation in Vietnamese pharmacy education remains limited.
This study is grounded in Vygotsky’s (1978) Zone of Proximal Development (ZPD)
theory [10], operationalized through the scaffolding mechanism proposed by Wood,
Bruner, and Ross (1976) [11]. In this context, the AI chatbot can serve as a form of
learning support by providing step-by-step explanations when students encounter
unfamiliar pharmacoepidemiological indicators. This process may help students move
from basic definitions to clinical interpretation at a level appropriate to their current
understanding. At the cognitive level, the structured sequence of conversational tasks
may have reduced cognitive load by breaking complex reasoning into smaller, more
manageable steps [12]. To address this gap, the present study pursued three objectives:
evaluating whether a structured AI chatbot intervention could improve students'
pharmacoepidemiological competency; comparing outcomes between ChatGPT and
Gemini; and assessing changes in self-confidence and analysis speed
2. Materials and methods
2.1. Study design
The study employed a one-group pretest-posttest design (before-after study)
without a parallel control group, a design that is recognized as appropriate in fixed

712