Comparison of Human and GPT-Generated Concept Maps in a Clinical Reasoning Collection of Educational Virtual Patients

This study investigates the potential of a generative pre-trained transformer (GPT) model for creating clinical reasoning concept maps for virtual patient cases to compare these maps with those generated by clinicians for 20 diverse clinical scenarios. To evaluate the model's alignment with clinicians' approach, precision, and recall metrics were calculated. For concepts, the recall was between 46%-74%, while precision was between 16%-50%. The custom GPT model identified a higher number of medical concepts than clinicians. The obtained results substantiate its potential as a valuable tool for supporting the creation of educational concept maps. GPT-generated maps can enhance the process of map creation by introducing additional concepts that assist clinicians and medical educators in considering alternative diagnoses, tests, or treatments to facilitate student feedback.