Most medical AI is built English-first. When teams build for another language, they usually scrape Western datasets and run them through a translator. In healthcare, that shortcut is dangerous.
Why translation loses clinical signal
How a patient describes pain is clinical signal, not just wording. Vietnamese carries distinctions that English collapses. Đau nhức is a dull, deep ache. Đau buốt is a sharp, piercing pain. Đau thắt is a tight, gripping pain, the word a patient reaches for when describing the chest pain a doctor needs to act on.
Flatten those three into "aching," "sharp," and "tight" and you lose the distinctions a clinician triages by. A model trained on translated data learns how English maps those three words, then meets a Vietnamese patient and quietly misreads them.
How Vietnamese patients report symptoms
Symptom language is only half of it. Patients here report differently. Many patients here are stoic by habit, so they understate what they feel or describe it in terms a Western-trained model has never seen. A system that does not expect that under-reporting will take a calm description at face value and miss the severity underneath.
These are not edge cases. They are the ordinary Vietnamese consultation. A clinical model that cannot read them is not a slightly worse model. It is the wrong tool.
Why we generate the data natively
Meddies Consultant is generated natively in Vietnamese, not translated into it, so the nuance survives into the training data instead of being lost in a pre-processing step. Vietnamese has more than 100 million native speakers, and a clinical tool meant to serve them has to start from how they actually speak.
Translated data will always carry the accent of the language it came from. For a tool that has to understand a patient in a Vietnamese clinic, that accent is a defect, not a detail.