Integrating GPT-Based AI into Virtual Patients to Facilitate Communication Training Among Medical First Responders: Usability Study of Mixed Reality Simulation
Bibliographic record
Abstract
BACKGROUND: Training in social-verbal interactions is crucial for medical first responders (MFRs) to assess a patient's condition and perform urgent treatment during emergency medical service administration. Integrating conversational agents (CAs) in virtual patients (VPs), that is, digital simulations, is a cost-effective alternative to resource-intensive human role-playing. There is moderate evidence that CAs improve communication skills more effectively when used with instructional interventions. However, more recent GPT-based artificial intelligence (AI) produces richer, more diverse, and more natural responses than previous CAs and has control of prosodic voice qualities like pitch and duration. These functionalities have the potential to better match the interaction expectations of MFRs regarding habitability. OBJECTIVE: We aimed to study how the integration of GPT-based AI in a mixed reality (MR)-VP could support communication training of MFRs. METHODS: We developed an MR simulation of a traffic accident with a VP. ChatGPT (OpenAI) was integrated into the VP and prompted with verified characteristics of accident victims. MFRs (N=24) were instructed on how to interact with the MR scenario. After assessing and treating the VP, the MFRs were administered the Mean Opinion Scale-Expanded, version 2, and the Subjective Assessment of Speech System Interfaces questionnaires to study their perception of the voice quality and the usability of the voice interactions, respectively. Open-ended questions were asked after completing the questionnaires. The observed and logged interactions with the VP, descriptive statistics of the questionnaires, and the output of the open-ended questions are reported. RESULTS: The usability assessment of the VP resulted in moderate positive ratings, especially in habitability (median 4.25, IQR 4-4.81) and likeability (median 4.50, IQR 3.97-5.91). Interactions were negatively affected by the approximately 3-second latency of the responses. MFRs acknowledged the naturalness of determining the physiological states of the VP through verbal communication, for example, with questions such as "Where does it hurt?" However, the question-answer dynamic in the verbal exchange with the VP and the lack of the VP's ability to start the verbal exchange were noticed. Noteworthy insights highlighted the potential of domain-knowledge prompt engineering to steer the actions of MFRs for effective training. CONCLUSIONS: Generative AI in VPs facilitates MFRs' training but continues to rely on instructions for effective verbal interactions. Therefore, the capabilities of the GPT-VP and a training protocol need to be communicated to trainees. Future interactions should implement triggers based on keyword recognition, the VP pointing to the hurting area, conversational turn-taking techniques, and add the ability for the VP to start a verbal exchange. Furthermore, a local AI server, chunk processing, and lowering the audio resolution of the VP's voice could ameliorate the delay in response and allay privacy concerns. Prompting could be used in future studies to create a virtual MFR capable of assisting trainees.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.004 | 0.016 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.000 |
| Insufficient payload (model declined to judge) | 0.002 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".