Generative AI in Scenario-Based Healthcare Education: A Systematic Review of Applications, Validation Practices, and Pedagogical Integration

Synthesis: Neto and colleagues present a PRISMA 2020 systematic review synthesizing 23 empirical studies (from 1,151 initial records across five databases) on Generative AI applications in scenario-, case-, problem-, and simulation-based learning within healthcare education. Using the Mixed Methods Appraisal Tool (MMAT) for quality appraisal and thematic synthesis, they identify six cross-cutting themes centered on a core claim: prompt design in educational contexts functions as a form of instructional specification, encoding the cognitive targets and quality criteria implicit in expert authoring. The review argues that treating prompt design as methodology, standardizing multi-stage validation, and formalizing Human AI Collaboration are the prerequisites for responsible integration of GenAI into Medical and Health Professions Education and Simulation.

Key Findings

• Only 34.8% of studies aligned generated content with established instructional frameworks such as Bloom's Taxonomy — a gap of design culture rather than technical capability, since specifying cognitive targets within prompts is straightforward yet most implementations treated generation as a text-production rather than an instructional-design problem.
• An equal 34.8% reported prompting strategies in sufficient detail for reproduction, and among those that did, the highest-rated content came from iterative refinement with structured templates — supporting Prompt Engineering and transparent prompt architecture as a methodological norm.
• GPT-4 dominated implementations (44.4%), with OpenAI models underpinning roughly two-thirds of all model instances; Open Source alternatives were under-explored, creating a practical monoculture that compounds reproducibility problems.
• Hybrid human–AI co-creation models outperformed fully automated generation, achieving the highest expert ratings and student satisfaction, with educators positioned as collaborators throughout generation rather than final reviewers only.
• Educational outcomes were generally positive for higher-order cognitive skills (critical thinking, clinical reasoning) but inconsistent for factual knowledge acquisition, suggesting GenAI-assisted learning preferentially supports reasoning over recall.
• Scalability was the primary adoption driver yet efficiency gains were largely unquantified — only one study measured authoring time (a ~33% reduction) — while documented quality failures (16.67% hallucination rate in ophthalmology cases, 50% accuracy agreement in pharmacy PBL) persisted across the review period, indicating a structural rather than temporary problem.

What this means for practice

• Curriculum designers. Treat prompt design as a methodological element on par with instrument selection, disclosing system prompts, refinement steps, and evaluation criteria so that instructional design is not rendered invisible by unreported prompting.
• Curriculum designers. Specify cognitive targets against an established taxonomy and structure learning progressions before deployment — only 34.8% of the 23 reviewed studies aligned generated content with instructional frameworks.
• Medical educators. Standardize multi-stage validation through the proposed four-stage pipeline of automated checks, clinical expert review, pedagogical alignment, and learner pilot testing, because single-point validation cannot catch the failure modes the review documents.
• Medical educators. Keep educators inside the generation loop rather than at the final review stage: hybrid human–AI co-creation models achieved the highest expert ratings and student satisfaction.
• Researchers. Report authoring time and cost alongside content quality, since scalability was the primary adoption driver yet only one study quantified the efficiency gain (a ~33% reduction).

Limitations

• The evidence base is 23 empirical studies from 1151 initial records across five databases searched on 9 November 2025, restricted to peer-reviewed work published from January 2023, with 67 duplicates removed and 1014 records excluded at title and abstract screening.
• Gray literature was excluded and five studies were dropped because full texts were inaccessible, which the authors note may underrepresent emerging findings or null results.
• The inclusion criteria restricted designs to quantitative and mixed-methods studies, excluding purely qualitative work that the authors acknowledge captures complementary process-oriented implementation evidence.
• The corpus is temporally and geographically thin: 19 of the 23 studies appeared in 2025 and Asia–Pacific institutions account for 56.5% (n = 13), following Europe (17.4%, n = 4), North Africa and the Middle East (13.0%, n = 3), and North America (8.7%, n = 2), and one included result (Cohen's d = 3.14 for long-term retention) is flagged by the authors as a likely methodological artifact.

Citation

Neto, M., Pinto, R., Reis, J., & Antão, L. (2026). Generative AI in scenario-based healthcare education: A systematic review of applications, validation practices, and pedagogical integration. Computers and Education: Artificial Intelligence, 11, 100654.