TRV-2026-2979Version 1 · Published

Saved · see the current record

Why this version exists

Certified into the record

The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2979
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-17T00:02:13.227612Z
status: published
lens: trace
sector: health
headline: Medical Interview Performance Using an AI-Based Standardized Patient vs a Human Standardized Patient: Randomized Crossover Trial
dek: Background: In Japan, standardized patients (SPs) have traditionally been recruited to teach medical interview competencies. Recent advances in large language models have enabled increasingly natural conversational interactions, raising the possibility of using AI-based SPs (AI-SPs) for medical interview training. However, medical interview performance during interactions with AI-SPs compared with human SPs remains insufficiently understood.
Objective: This randomized single-center crossover trial aimed to compa…
gain_title: In a randomized crossover trial of 48 Chiba University medical students, self-rated active engagement in history taking was higher with a GPT-4o simulated patient (mean 6.2) than with a human simulated patient (mean 5.8).
problem_title: In a randomized crossover trial of 48 Chiba University medical students, interview scores were lower with a GPT-4o simulated patient than with a human simulated patient (mean 22.7 vs 23.9 on selected Mini-CEX items; P=.001).
trace_subject: GPT-4o simulated patient for medical interview training of fifth-year medical students at Chiba University Hospital
gain_reading: In a randomized crossover trial of 48 Chiba University medical students, self-rated active engagement in history taking was higher with a GPT-4o simulated patient (mean 6.2) than with a human simulated patient (mean 5.8).
gain_evidence: Active engagement in history taking was significantly higher in the AI-SP condition (human SP: mean 5.8, SD 1.1; AI-SP: mean 6.2, SD 0.7; t47=2.7; P=.009).
problem_reading: In a randomized crossover trial of 48 Chiba University medical students, interview scores were lower with a GPT-4o simulated patient than with a human simulated patient (mean 22.7 vs 23.9 on selected Mini-CEX items; P=.001).
problem_evidence: The Mini-CEX-based mean total score was significantly higher for the human SP than for the AI-SP (23.9, SD 1.7 vs 22.7, SD 1.7, respectively; mean difference 1.24; t47=3.7; P=.001).
quick_read: A randomized crossover trial at Chiba University Hospital in Japan compared a text-based simulated patient built on GPT-4o with a human simulated patient. Forty-eight fifth-year medical students interviewed both, and two general medicine specialists rated them. Students scored 23.9 on selected Mini-CEX items with the human patient and 22.7 with the AI patient.

Scores for history taking and overall clinical competence were higher in the human condition. Students rated learning satisfaction and rapport about the same in both conditions, but said they were more actively engaged in history taking with the AI patient (6.2 vs 5.8 on a 7-point scale).

The authors suggest the AI patient could complement human simulated patients for interview practice. They note the study measured performance in simulated encounters, not later learning or performance with real patients.
limitation: The study assessed performance during simulated encounters rather than later learning gains or performance with real patients, and the AI patient was text-based with no nonverbal cues.
tag: Dual reading
key_points: Participants were fifth-year medical students on clinical clerkships at Chiba University Hospital; 48 were analyzed. | Each student completed two cases, depression and appendicitis, with both the GPT-4o text-based patient and a human simulated patient, in randomized order with a washout of at least 3 days. | The mean Mini-CEX-based total score was 23.9 (SD 1.7) with the human patient versus 22.7 (SD 1.7) with the AI patient, a difference of 1.24 (P=.001). | Learning satisfaction was similar: 5.9 for the human patient and 6.0 for the AI patient on a 7-point scale. | Perceived rapport was similar: 5.4 for the human patient and 5.3 for the AI patient. | Self-rated active engagement in history taking was higher with the AI patient: 6.2 versus 5.8 (P=.009). | Interrater reliability between the two raters was 0.76 (95% CI 0.66-0.83).
rundown: The trial used a single-center randomized crossover design. Students were randomized to the AI patient first or the human patient first, and each did two cases in both conditions. AI interviews were text chats with a physician giving feedback after both cases; human interviews were video recorded with feedback from the physician portraying the patient immediately after each interview.

Two general medicine specialists independently rated performance using selected Mini-CEX items. The human patient condition scored higher on total score, history taking and overall clinical competence, with interrater reliability of 0.76.

Students' self-ratings differed from the specialist ratings. Learning satisfaction and rapport did not differ significantly, while active engagement in history taking was higher with the AI patient. The authors say the text-only format and missing nonverbal cues may have contributed to the performance gap.
sources:
- peer_reviewed | JMIR Medical Education | https://doi.org/10.2196/111030 | 2026-11-16
prev: 0000000000000000000000000000000000000000000000000000000000000000
SHA-256 hash
453bac8ee46c9f0f657614e1a4acdd0978776568f8915199e688b6765ab762ad
Previous version’s hash
0000000000000000000000000000000000000000000000000000000000000000
Check this record hasn’t been altered
How to check it yourself, without trusting this page

Download any version from /api/record/TRV-2026-2979 and hash its canonical field yourself, for example with shasum -a 256. You should get its content_hash. Each version’s text also ends with prev:and the hash of the version before it (version 1 points to 64 zeros). If anyone had changed a single character of any version after it was saved, those hashes wouldn’t line up.