A Virtual Patient Wakes Up With A Hidden Diagnosis
An AI language model plays a fully scripted patient persona.
- 1,200+: Simulated cases available (AI-generated patient scripts)
- 1: Hidden diagnoses per case (never revealed upfront)
- <1 sec: Response latency (real-time patient dialogue)
- 40+: Specialties covered (medicine to psychiatry)
Why simulate patients at all
Real patients are scarce, so AI fills the training gap.
The hidden condition
One true diagnosis drives every scripted patient answer.
The student never sees the diagnosis until final feedback.
Persona consistency
The model stays in character across the whole interview.
History Taking — Questions Build The Clinical Picture
Every question the student asks reshapes what the patient reveals.
- 6: Question types tracked (history, ROS, PMH, meds)
- 14: Avg questions per case (before differential narrows)
- 98%: In-character consistency (LLM persona fidelity)
- Yes: Affect cues simulated (pain scale, anxiety, evasiveness)
Open versus closed questions
Open questions surface more clues than yes-or-no ones.
Simulated affect
The patient hesitates, winces, or downplays symptoms.
Question budget
Every question spent narrows or wastes diagnostic time.
Efficient questioners reach the diagnosis with fewer clues.
The Differential List Narrows With Every Clue
Six candidate diagnoses compete until evidence rules most out.
- 6: Starting differential size (candidate diagnoses)
- 1–2: Eliminated per good clue (ruled out per question)
- 3: Difficulty levels (straightforward to complex)
- 5: Bias prompts flagged (anchoring, premature closure)
Elimination logic
Each answer rules a candidate in or decisively out.
Difficulty scaling
Complex cases need far more clues to narrow down.
Cognitive traps
Anchoring on the first idea can stall the differential.
Complex cases can need triple the clues of easy ones.
Ordering Tests To Confirm The Leading Suspicion
Tests either confirm the leading diagnosis or reopen the case.
- 12: Tests available (labs, imaging, ECG)
- 3–4: Avg tests ordered (per completed case)
- -5%: Unnecessary-test penalty (overordering lowers score)
- Instant: Simulated turnaround (results return immediately)
Targeted ordering
Good clinicians order tests that change management.
Result interpretation
Each returned result updates the shrinking differential.
Overordering cost
Excess tests cost points for efficiency and stewardship.
ECG and troponin alone often confirm this case.
Scoring The Student Against The True Condition
The final diagnosis is compared against the hidden truth.
- 4: Scoring dimensions (accuracy, efficiency, reasoning, cost)
- 70%: Passing accuracy threshold (for case credit)
- Full: Feedback detail (reasoning gaps explained)
- Unlimited: Repeatable cases (replay with new hidden variants)
Accuracy scoring
Score rewards narrowing to the true diagnosis quickly.
Reasoning transparency
Feedback shows exactly which clues mattered most.
Deliberate practice
Every replay swaps in a new hidden condition.
Pericarditis confirmed — chest pain eased leaning forward.