How we think about clinical AI in Vietnam.
A hospital pilot needs a prespecified decision contract covering intended use, local cases, reviewers, failure thresholds, data paths, human control, stopping rules, and rollback.
Read moreA proposed protocol for testing whether prescription alerts are correct, relevant, actionable, and worth interrupting a clinician. No Meddies benchmark result is claimed.
Read moreA proposed hospital protocol for finding clinically important omissions in AI chart summaries with source maps, bidirectional review, consequence labels, and blocking gates.
Read moreVietnamese healthcare facilities can report a suspected ADR before causality is resolved. The workflow keeps care, documentation, deadlines, submission, follow-up, and AI assistance separate.
Read moreAn exploratory audit of 24 synthetic Vietnamese symptom utterances found six translation problems. Five received conservative post-edits; one source ambiguity requires clarification. Clinical validation is pending.
Read moreClinical de-identification must reduce identifiability without erasing the facts a declared use needs. Meddies PII detects seven entity families; it does not redact or certify compliance.
Read moreAI should abstain when a chart cannot support a final ICD-10 code or when BHYT payment context is missing. Meddies has early lookup and evaluation seams, not a shipped mapper.
Read moreVietnam’s EMR mandate creates infrastructure for clinical AI, not clinical intelligence. Regional fit still needs evidence from local records, medicines, sources, and workflows.
Read moreVietnam's EMR mandate sets infrastructure requirements that integrated CDSS could use; it does not guarantee interoperability, clinical readiness, or proven benefit. Evidence from testing in 2013 to a five-hospital assessment in 2025.
Read moreA July 2026 documentation audit of six tools that solve different clinical jobs, covering public source boundaries, workflow placement, and open validation questions.
Read moreA meta-analysis estimated a 90% physician override rate across 11 studies, with extreme variation. It identifies a problem to measure, not proof that every override was unsafe.
Read moreCircular 13 sets two EMR deadlines and preserves transition rules for paper records. Electronic does not automatically mean complete, interoperable, or ready for clinical AI.
Read moreA supported clinical summary can still miss relevant facts. Accuracy and coverage require two audits: output to source for hallucinations, then source to output for omissions.
Read moreVietnam’s AI Law replaced the DTI Law’s AI chapter on 1 March 2026. Hospital AI now raises legal questions about risk, transparency, and human control.
Read moreA CDSS uses patient context to support a decision at the right time. It may be an alert, summary, reminder, order set, or diagnostic aid, and need not use AI.
Read moreA citation can be real and still fail to support the claim beside it. What Meddies can trace today, what it cannot, and what a verifiable clinical answer requires.
Read moreQA benchmarks test answers in a fixed context. A consultation must track state and find missing information. Meddies Consultant separates these tasks; model results remain unpublished.
Read moreSynthetic generation can avoid unnecessary access to patient records. It does not prove zero leakage, formal privacy, or safe release without provenance and output checks.
Read moreVietnamese text is only one part of clinical localization. Language, patient context, and dataset design each need separate evidence; current Meddies releases do not complete all three.
Read moreMeddies Consultant publishes consultation, QA, and question-only configs. Their structure is inspectable; model behavior and clinical safety still need evaluation.
Read moreReview the Meddies product direction and one synthetic workflow example.
Book a demo
