A field note from building clinical data in a language the medical-AI world mostly ignores.
The word for the pain
Early on, we kept hitting the same wall. A model trained on translated Western data would meet a Vietnamese patient describing their pain and quietly get it wrong, not because the medicine was hard, but because the words did not survive the trip through English.
Vietnamese patients do not say "I have pain." They tell you which pain. Đau nhức is the dull, deep ache that settles into a joint or a tooth. Đau buốt is the sharp, piercing pain that flashes and fades. Đau thắt is the tight, gripping pain, the phrase that should make a clinician think of the chest and pay attention.
Three distinct clinical pictures, three words a patient chooses without thinking. Translate all of them into "aching," "sharp," and "tight" and you have not just lost some flavor. You have erased the distinction the triage depends on. The English-trained model reads a flattened symptom and gives a flattened answer.
Stoicism is data, too
The language is only the first layer. How patients here report is its own signal. There is a cultural stoicism that runs through many Vietnamese consultations: pain is understated, fear is hidden, the real complaint sometimes arrives only after the third question.
A model that expects a Western-style patient, who names the symptom plainly and rates it out of ten on request, will take a calm, downplayed account at face value. It will miss the severity sitting underneath the understatement. In a clinic, that is not a rough edge. That is a missed call.
Why we generate, not scrape
This is why we stopped trying to adapt borrowed datasets and started building natively. The synthetic patients in meddies-consultant are designed to behave like real ones: anxious, indirect, prone to rambling, reluctant to overstate. We program that inner state deliberately so the model that learns from them is ready for the patient who does not arrive as clean data.
It also forces the AI doctor into the right posture. Faced with a vague or stoic complaint, the model that does well is the one that asks another question. The model that does badly is the one that answers too soon. We built the data to reward the first behavior and punish the second.
The lesson
Clinical AI for Vietnam is not English clinical AI with the labels swapped. The pain has a different name, the patient tells a different story, and a tool that cannot hear either one is the wrong tool in the room.
So we build for consultation, not for answering. The goal is a model that knows it does not yet know enough, and asks. In a language that has three words for pain, that humility is not optional. It is the job.