ICD-10 coding and BHYT payment data meet in the same hospital record, but they do not answer the same question. ICD-10 classifies the diagnosis or documented health condition. A BHYT record also carries the insurance card, facility, encounter, services, clinicians, dates, and other data needed for management and payment. A valid diagnosis code cannot establish those other facts.
This distinction sets the first abstention rule for AI. A system may help find an ICD-10 code. It must not turn that code alone into a conclusion about BHYT eligibility or payment.
The code must not exceed the clinical record
Circular 06/2026/TT-BYT now governs Vietnam's ICD-10 list and coding rules. It took effect on 1 July 2026. The circular defines the principal condition at the end of the encounter. When several conditions are present, the one that used the most resources is selected. When no definitive disease can be diagnosed, the documented main symptom, disorder, or abnormal finding may become the principal condition. Its transition clause makes the new list controlling when older Ministry coding rules conflict and applies it to encounters ending on or after 1 July 2026.
That is a precise limit on automation. If the record supports only a symptom, AI may surface a symptom-code candidate. It should abstain from supplying a disease diagnosis that the practitioner did not establish. Circular 32/2023/TT-BYT requires medical-record entries to be accurate, truthful, and complete. Coding software can search and check the record; it cannot repair missing clinical judgment by inventing it.
The Ministry's June 2026 implementation letter tells hospitals to update their HIS with the Circular 06 list while continuing to use the BHYT output fields amended by Decision 3176/QĐ-BYT. For the summary record, the healthcare facility records one principal ICD-10 code and up to 12 accompanying codes. Those codes must correspond to diagnoses confirmed at the end of the encounter. Suspected, monitored, or differential diagnoses that were excluded by discharge belong in the relevant narrative fields, not the confirmed-code fields.
Abstention belongs at five evidence boundaries
First, the system should not assign a final disease code while the encounter is still open and the diagnosis is changing. It should also stop when different notes conflict and no discharge or encounter summary resolves them.
Second, selecting the principal condition may require information that the available extract does not contain. When several conditions brought the patient to hospital, the coding guidance can depend on which condition used the most resources. Without that context, a model cannot reliably choose the principal code.
Third, every candidate must resolve against the Ministry list and the version in force. A code remembered from training data, a foreign code set, or an old local table is not enough.
Fourth, a suspected or differential diagnosis that was later excluded must not survive as if it were confirmed. When no disease was established but the record contains a documented main symptom, the system can narrow its output to that symptom. That narrower output is the answer the evidence permits.
Fifth, the system must refuse to infer BHYT payment from ICD-10 alone. The claim context is a separate data problem with its own fields and rules.
A refusal should move the case forward
An empty field and a low-confidence label do not tell a reviewer what to do. A useful abstention names the unresolved issue, identifies the missing or conflicting evidence, links back to the relevant chart text, and assigns the next review. If it shows candidate codes, each candidate should remain bounded by a diagnosis, symptom, or finding already documented in the record.
Hospitals may set confidence or agreement thresholds that route cases to human review. Those thresholds need local validation and should reflect the consequence of each error. Ministry guidance defines coding and data structure; it does not publish one AI confidence cutoff for every hospital.
The reviewer also depends on the question. A physician confirms clinical content. A clinical coder checks the ICD-10 rule. BHYT staff verify the payment-data context. A single approval button does not merge these responsibilities.
A 2026 npj Digital Medicine deployment study evaluated an ICD-10-CM assistant for 13 weeks with ten certified coding specialists at two institutions in Taiwan. AI assistance reduced coding time while maintaining accuracy, but satisfaction varied by user experience and certification. This does not validate Vietnamese ICD-10 or BHYT mapping. It changes the hospital evaluation method by making reviewer experience, review time, uptake, overrides, and unresolved cases part of the result.
Meddies does not yet implement this mapping workflow
The current Meddies source defines an ICD reference-table schema, an authenticated search endpoint, and a manual autocomplete that attaches ICD-10 tags to documents. Its seed script contains four example codes, and its clinical evaluation code includes an exact-match verifier for ICD cases. This establishes an early structure. It does not prove that any separately managed production database contains the complete Ministry list.
The repository does not currently implement chart-to-ICD generation, mapping to the Decision 3176 BHYT schema, claim submission, or production abstention behavior for this task. These components establish early lookup and evaluation infrastructure, not a hospital-ready automated coding system.
The abstention rules in this article are therefore a design contract for future work, not a description of a shipped Meddies feature. A safe system fills only fields supported by evidence and sends the rest to a named human reviewer with the reason.