What 102 AI patient cases show about history taking
Updated 20 October 2026
In 102 scored practice cases played by 43 learners on Diagnosica between 15 September and 19 October 2026, learners asked the AI patient a median of 12 questions and covered 63% of the history questions each case expected. In 75% of cases they left at least one question the case marks as critical unasked.
History taking was the strongest of the four scored areas (median 16 out of 25) and the management plan the weakest (median 4 out of 25). Learners named the main diagnosis correctly in 49% of cases.
Diagnosica is an AI clinical simulator for medical students and junior doctors. You take a history from an AI patient on video, by voice or text, order labs and real imaging, commit to a diagnosis and management plan, and get a competency scorecard.
We publish these numbers because we haven't found another clinical practice tool that does, and because they describe what learners actually did with an AI patient. They aren't a measure of what doctors do with real patients. Every figure gives its sample size; it's a young product, so the samples are small and the confidence ranges are wide.
| Measure | Result | Based on |
|---|---|---|
| Questions to the patient per full case (median) | 12 (middle half: 6 to 19) | 102 cases, 43 learners |
| Cases with fewer than 5 questions | 16% (95% range 10 to 24%) | 102 cases |
| Expected history questions that learners asked | 63% (61 to 66%) | 1,307 rubric questions in 102 cases |
| Critical history questions left unasked | 59% (52 to 65%) | 210 critical questions |
| Cases with at least one critical question unasked | 75% (66 to 83%) | 102 cases |
| Main diagnosis named correctly | 49% (40 to 59%) | 102 cases |
| Played full cases that were finished | 64% (57 to 71%) | 166 played cases, 59 learners |
Which parts of the history went unasked most often
Each case comes with a list of the history questions it expects, written by our clinical team. After the case, an AI grader reads the transcript and marks each one asked or not asked, quoting the learner's words as evidence. We grouped those questions into history domains. "Cases" counts the cases whose list included at least one question of that kind.
| History domain | Cases that expected it | Cases where learners left at least one unasked | Expected questions left unasked |
|---|---|---|---|
| Fever, weight loss, night sweats, appetite | 58 | 31 (53%) | 35 of 82 (43%) |
| Past medical history and previous episodes | 70 | 36 (51%) | 42 of 93 (45%) |
| Onset, timing and progression | 91 | 38 (42%) | 44 of 133 (33%) |
| Medications and allergies | 66 | 27 (41%) | 29 of 91 (32%) |
| Character, site, severity and triggers of the main symptom | 56 | 20 (36%) | 25 of 80 (31%) |
| Smoking, alcohol, drugs, work, travel, diet | 59 | 19 (32%) | 21 of 87 (24%) |
| Associated symptoms and other systems | 102 | not comparable (about 7 questions per case) | 273 of 695 (39%) |
Family history and the patient's own ideas and concerns appeared in too few cases to report (fewer than 10 unasked, or fewer than 10 cases).
Scores by competency
Every finished full case is scored out of 100: four areas worth 25 each.
| Area | Median out of 25 | Middle half | Mean |
|---|---|---|---|
| History taking | 16 | 11 to 19 | 14.9 |
| Investigations | 11 | 8 to 16 | 11.4 |
| Diagnosis | 14 | 0 to 15 | 10.1 |
| Management | 4 | 0 to 8 | 4.8 |
| Total out of 100 | 41 | 28 to 56 | 41.2 |
102 cases from 43 learners. Diagnosis splits in two because the main diagnosis carries 15 of its 25 points: a wrong one pulls the area close to zero.
More questions went with higher scores, up to a point:
| Questions asked | Cases | Mean total out of 100 | Mean history taking out of 25 |
|---|---|---|---|
| Fewer than 6 | 20 | 28.0 | 7.7 |
| 6 to 11 | 28 | 42.9 | 14.6 |
| 12 to 17 | 27 | 46.0 | 16.3 |
| 18 or more | 27 | 44.3 | 19.1 |
The total score levelled off after about 12 questions, while the history score kept climbing. That's a correlation in a small sample, so it doesn't show that asking more questions causes a higher score.
How long a case takes
| Measure | Median | Middle half | Based on |
|---|---|---|---|
| Full case, start to scorecard | 17.1 minutes | 11.2 to 27.5 | 102 cases, 43 learners |
| Full case, first to last question to the patient | 8.0 minutes | 3.9 to 14.8 | 100 cases, 43 learners |
| Quick case, opened to answered | 66 seconds | 18 to 175 seconds | 130 answers, 19 learners |
Start to scorecard includes reading results and any time the app sat open, so it's an upper bound. The question span covers the history only. The quick case figure leans on a small group of regular players.
Do scores go up with practice?
For the 12 learners who finished three or more scored full cases, the average score on their first case was 33.5 out of 100 and on their third case onwards 47.1. The average gain per learner was 13.6 points (95% range 6.0 to 21.2), and 10 of the 12 improved.
Treat that with care. Twelve people is a small group, learners who keep playing may differ from those who stop, each person played different cases, and part of any first-case gap is getting used to the app. Only 5 learners have finished five or more scored cases, too few to report.
What gets played
| Specialty | Case starts | Share |
|---|---|---|
| Renal | 39 | 15% |
| Surgery | 34 | 13% |
| Paediatrics | 28 | 11% |
| Acute medicine | 27 | 11% |
| Gastroenterology | 25 | 10% |
| Endocrinology | 19 | 7% |
| Neurology | 16 | 6% |
| Ophthalmology | 13 | 5% |
| Cardiology | 13 | 5% |
| Eight other specialties (under 10 starts each) | 40 | 16% |
254 case starts from 71 learners. The mix reflects which cases the app features as much as what learners chose.
14% of full case starts used voice (27 of 196; 95% range 10 to 19%); the rest were typed. Free accounts get 30 minutes of voice a month, so that share is partly a limit.
Methods
- Window. Cases started from 15 September 2026 to 19 October 2026 inclusive (UTC). Earlier data comes from a version of the product that no longer exists, so it's left out.
- Who's included. Registered learner accounts only. We removed our own team, test and review accounts, friends and family, and app-store review accounts. Cases played without an account are left out because we can't separate them from our own testing.
- Which cases are scored. Full cases with a complete score on the current marking scheme (102 cases from 43 learners).
- How marking works. AI grading against a case-specific checklist written by our clinical team. It's formative feedback and hasn't been validated against human examiners. A learner may have covered a point in words the grader didn't credit.
- History domains. We grouped checklist questions by keyword. A hand check of 60 questions found about 55 grouped sensibly.
- Small numbers. We don't report any figure resting on fewer than 10 cases or 10 learners. The ranges shown are 95% confidence intervals; one learner often plays several cases, so the true uncertainty is somewhat wider.
- Privacy. Everything here is aggregated. No individual learner can be identified.
These are practice cases with an AI patient, written for teaching. None of it is clinical advice or a statement about how doctors treat real patients. We'll refresh the numbers as more cases come in.
Common questions
How many questions do learners ask an AI patient?
In 102 scored practice cases on Diagnosica between 15 September and 19 October 2026 (43 learners), the median was 12 questions to the patient per full case, with the middle half asking between 6 and 19. 16% of cases had fewer than 5 questions.
What do learners most often leave out of a practice history?
Across those 102 cases, learners left at least one question the case marks as critical unasked in 75% of cases. Among cases that expected them, at least one question about past medical history or previous episodes went unasked in 51%, and about fever, weight loss, night sweats or appetite in 53%.
Can I quote these numbers?
Yes. Please cite "Diagnosica practice data, diagnosica.com/ai-patient-practice-data, updated 20 October 2026" and keep the sample size with the figure (for example, "in 102 practice cases from 43 learners"). The figures describe practice with an AI patient, not real clinical care.
How was the history marked?
Each case has a checklist of history questions written by our clinical team. An AI grader reads the transcript and marks each question asked or not, quoting the learner as evidence. It is formative feedback, not a validated exam mark.
How much does Diagnosica cost?
Diagnosica is free to start, with no card. A free account gets 5 cases in your first week, then 3 cases a week, plus unlimited 2-minute quick cases, with every case in the library open. Standard is $14.99 a month or $119.99 a year (UK: £15 a month or £129 a year) and allows 50 cases a month. Prices as of 19 October 2026.
Is Diagnosica legit, and who makes it?
Diagnosica is a small, independent company operated by CC Global Group LLC, live on the web, the App Store and Google Play. Cases are written for teaching and clinically reviewed before they go live. It is a learning tool, not a medical device, and it is not affiliated with any exam board.