AI-Generated Discharge and Aftercare Instructions in Viet Nam: Governing Medication Accuracy, Warning Signs and Patient Understanding
PRACTICAL FRAMEWORKS & TOOLS
AI-Generated Discharge and Aftercare Instructions in Viet Nam: Governing Medication Accuracy, Warning Signs and Patient Understanding
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 9 October 2026
A patient is preparing to leave the hospital after several days of treatment. The medical team has changed some medications, arranged a follow-up appointment and explained which symptoms should prompt the patient to seek further care. Before discharge, an AI application uses information from the medical record to generate a personalized set of instructions.
The document looks helpful. It explains the diagnosis in everyday language, organizes medications into a schedule and summarizes what the patient should do at home. However, one medication that was stopped during hospitalization appears in the list of medicines to continue. The instructions also advise the patient to watch for worsening symptoms without explaining which changes require urgent medical attention.
The document may be easy to read, but it is not necessarily safe to follow.
This example illustrates an important challenge for healthcare organizations introducing generative AI into discharge planning and patient education. AI can make complex clinical information more accessible, but it may also omit essential details, misrepresent medication changes or provide advice that does not reflect the patient's actual treatment plan.
For hospitals and other healthcare providers in Viet Nam, the question is not simply whether AI can prepare discharge instructions more quickly. It is whether the information patients receive is accurate, safe, understandable and sufficiently specific to help them manage their care after leaving the healthcare facility.
Discharge instructions are different from a summary of hospital treatment
Discharge is an important transition in patient care. Once a patient leaves the hospital, many activities previously supported by healthcare professionals become the responsibility of the patient, family members or other caregivers.
The patient may need to start new medications, stop medicines used before admission, monitor symptoms, care for a wound, follow dietary or activity restrictions, attend follow-up appointments or arrange further investigations. Some instructions may be straightforward, while others require considerable knowledge and confidence to carry out safely.
The information provided at discharge therefore serves a different purpose from a clinical discharge summary. A discharge summary primarily records relevant aspects of the hospital stay and supports continuity of care between healthcare professionals. Patient-facing discharge and aftercare instructions explain what the patient or caregiver needs to know and do after leaving the facility. Prescriptions, discharge certificates and patient education materials may contain related information, but they have different functions and applicable requirements.
AI may help convert an approved clinical discharge plan into understandable language. However, it should not independently determine medication changes, follow-up requirements or other clinical instructions that have not been authorized by the responsible healthcare professionals.
This distinction matters because an AI system may receive only part of the relevant information. A discharge summary might record medicines administered during hospitalization without identifying which ones should continue at home. Laboratory results may still be pending, or a follow-up appointment may have been discussed without being entered into the information system.
When the source information is incomplete or inconsistent, generating a polished document does not resolve the underlying problem. It may simply make the uncertainty less visible. Healthcare organizations should therefore distinguish between using AI to explain an established discharge plan and allowing AI to create clinical recommendations that are not supported by that plan.
What international evidence tells us about AI-generated discharge instructions
Research on AI-assisted discharge communication has expanded in recent years. Studies have examined readability, information accuracy, medication safety and patient comprehension, with findings that are encouraging in some areas but concerning in others.
Stanceski and colleagues published a study in npj Digital Medicine in November 2024 evaluating patient-centered discharge instructions generated by GPT-3.5 from 100 hospital discharge summaries. The AI-generated instructions were generally shorter and used simpler language than the original documents. However, potentially harmful safety issues attributable to the AI-generated responses were identified in 18% of cases, including hallucinated information in 6% and newly introduced medications in 3% [1].
The findings illustrate a practical concern. A document may become easier to read while introducing information that was never part of the source clinical plan.
Tang and colleagues examined medication-related risks more closely in a study published in the International Journal of Medical Informatics in April 2026. Using 140 discharge summaries, the researchers evaluated patient-centered medication instructions generated by GPT-4o. All medications from the source summaries were included in 95% of the generated instructions, yet at least one potential safety issue was identified in 69.3% of cases. The study also found differences in the occurrence of safety issues across certain patient groups, including older patients and those with greater clinical complexity [2].
These findings need careful interpretation. The 69.3% figure represents cases with potential safety issues identified during expert evaluation, not patients who experienced medication-related harm. The study did not establish that these errors would necessarily reach patients or cause adverse clinical outcomes. It also used a different model and assessment approach from the earlier research, so the two percentages should not be directly compared.
Other studies suggest that AI-generated patient-facing documents can improve certain aspects of comprehension.
In a prospective randomized crossover study published in the Journal of Medical Internet Research in February 2026, Holderried and colleagues compared conventional discharge letters with GPT-4-generated patient-centered versions. The study involved 48 trained standardized patients who reviewed documents relating to selected clinical conditions. The AI-generated letters improved participants' ability to recall and explain predefined safety-relevant information, particularly medication instructions and organizational arrangements. However, improvements were smaller for understanding complication prevention and lifestyle-related aspects of recovery [3].
This distinction matters. A patient may remember the name of a medication or the date of an appointment while remaining uncertain about how to respond when their condition deteriorates.
A systematic review by Wang and colleagues, published in npj Digital Medicine in January 2026, examined 20 studies involving large language models in preoperative and discharge education. Some studies reported improvements in anxiety or selected aspects of patient satisfaction, while findings for other outcomes were inconsistent. The review also identified gaps in how model reliability, trustworthiness and clinical outcomes were evaluated [4].
Research in emergency care has raised similar concerns about language and completeness. Hernández and colleagues evaluated discharge instructions generated by several large language models for 20 emergency department diagnoses in a study published in The Journal of Emergency Medicine in 2026. Readability varied considerably between models, and many instructions exceeded commonly recommended English-language health-literacy targets despite prompts requesting simpler language. Although physician ratings were favorable in several domains, approximately one-quarter of the assessed GPT-4 instructions received lower scores for completeness or understandability [5].
These studies do not establish that AI-generated discharge instructions are inherently unsafe or ineffective. They demonstrate why improvements in writing efficiency, readability or patient satisfaction cannot be treated as sufficient evidence of clinical reliability.
Most of the available research also comes from settings outside Viet Nam. Its findings should not automatically be generalized to Vietnamese patients, clinical documentation practices or healthcare delivery arrangements. Local evaluation remains necessary before routine implementation.
Medication accuracy must come before language simplification
Medication information is one of the most consequential parts of a discharge plan.
During hospitalization, a patient's medicines may change several times. Some are started for a new diagnosis, others are temporarily withheld, and some are discontinued or prescribed at a different dose. Medicines used during hospitalization may not be intended for continued use after discharge.
A patient may also have medicines at home that are not fully reflected in the hospital information system. These may include prescriptions from another healthcare provider, over-the-counter medicines or products purchased directly from a pharmacy. If AI prepares instructions from an incomplete medication history, it may reproduce an outdated list or incorrectly suggest that a patient should resume a treatment that has been stopped.
The clinical medication plan should therefore be established and relevant discrepancies resolved before AI is used to prepare patient-facing instructions.
Medication reconciliation is particularly important at this stage. The process involves comparing the medicines the patient was using before admission with those prescribed during hospitalization and those intended for use after discharge. The aim is to identify and resolve clinically important discrepancies so that the final medication plan is accurate and clear.
The World Health Organization identifies medication safety during transitions of care as a priority because changes in care settings create opportunities for medication discrepancies and preventable harm. Its 2019 technical report emphasizes structured processes, reliable information and the involvement of patients and families [7].
For AI-assisted discharge communication, the practical implication is that the system should work from the current, professionally authorized medication plan rather than attempting to reconcile conflicting information independently.
The final instructions should distinguish medicines to start, continue, change or stop. Where relevant, they should accurately communicate the medication name, strength, prescribed dose, route, frequency, duration and important administration precautions.
Particular attention is needed when tablet strength differs from the prescribed dose. Confusing the number of tablets with the amount of active ingredient can produce clinically significant errors. Similar problems arise when an AI system changes a dosing interval, omits a stop date or converts an instruction to use a medicine only when necessary into a routine daily schedule.
Brand and generic names also require care. A patient in Viet Nam may recognize a medicine by its commercial name while the hospital records its active ingredient. If the instructions do not make this relationship clear, the patient could unintentionally use two products containing the same active ingredient.
Medication changes deserve greater emphasis than simply reproducing a complete list. A person who has taken a medicine for several years may need a particularly clear explanation when its dose changes or when it must be stopped.
Higher-consequence medications, such as anticoagulants, insulin and certain medicines with a narrow therapeutic range, may justify additional pharmacist or clinician verification according to the patient's circumstances.
AI should not independently invent a replacement medicine, calculate a new dose, resolve a conflicting prescription or recommend restarting a discontinued treatment without an authorized clinical basis. When medication information is missing or inconsistent, the appropriate response is to seek professional clarification rather than produce an apparently complete schedule.
Warning signs must tell patients when and how to seek help
Aftercare instructions often advise patients to return if symptoms worsen. Although this may be intended as a safety precaution, it can be too vague to support timely decisions.
Patients may not know which symptoms are expected during recovery, which should prompt contact with the treating service and which require immediate medical attention.
For example, some discomfort after a procedure may be anticipated, while severe or worsening pain, significant bleeding or other condition-specific findings may require urgent assessment. The relevant warning signs depend on the diagnosis, treatment, procedure and individual clinical circumstances.
AI-generated instructions should therefore be checked for the accuracy, completeness and clarity of warning signs.
A useful document should help patients distinguish between symptoms requiring immediate emergency assessment, problems requiring prompt professional advice and concerns that can reasonably be discussed during scheduled follow-up.
It should also explain how to obtain assistance. Telling a patient to contact the hospital is of limited value if the instructions do not identify the appropriate service, provide a verified contact method or explain what to do outside normal working hours.
This is particularly important when a patient lives far from the treating hospital or returns to an area where specialist services may not be readily available. Instructions should make clear when the patient needs to seek care at an appropriate nearby healthcare facility rather than wait to return to the original hospital.
AI can generate lists of complications associated with a diagnosis, but a generic list is not necessarily appropriate for an individual patient. Some warnings may be irrelevant, while others specific to the patient's treatment or procedure may be missing.
An AI system may also provide excessive reassurance or suggest that a patient wait when the clinical situation requires urgent assessment.
For these reasons, the responsible clinical team should determine which warning signs and escalation instructions are relevant. AI can help express that information clearly, but it should not independently establish clinical thresholds for seeking care.
The distinction between urgency levels must remain understandable to the patient. A technically accurate description of a complication is not enough if the person reading it cannot determine what action to take.
Aftercare instructions must be practical enough to follow at home
Medication accuracy and warning signs are central to discharge safety, but patients may also need guidance about everyday recovery.
Depending on the clinical situation, this may include wound care, physical activity, nutrition, rehabilitation exercises, home monitoring, use of medical equipment or support from family members and community healthcare providers.
An instruction such as “continue wound care” may be meaningful to a healthcare professional but insufficient for someone who has never changed a dressing. Similarly, advising a patient to monitor blood pressure may not be useful unless they understand how and when to measure it, which results matter and what to do when a reading is outside the range specified by the clinical team.
AI-generated instructions should therefore be assessed for whether the patient can act on them, not just whether the information is complete or easy to read.
Where a task requires practical skills, a demonstration may be more useful than written instructions alone. Patients learning to use an inhaler, administer an injection or manage a device may need direct teaching and an opportunity to demonstrate the procedure.
Follow-up arrangements also need to be clear. The patient should understand whether an appointment has already been scheduled or still needs to be arranged, which healthcare service is responsible and whether additional investigations or test results remain outstanding.
AI should not convert a proposed appointment into a confirmed booking or suggest that a pending test result is normal because no abnormality appears in the available information. Where follow-up responsibilities are unclear, the gap should be resolved through the discharge process rather than concealed by AI-generated wording.
These details become especially important for patients with multiple conditions, those living alone and people who require assistance with daily activities or ongoing treatment.
Clear Vietnamese is not necessarily understandable Vietnamese
Generative AI may help convert complex medical terminology into more familiar language. This could be useful in Viet Nam, where patients differ in educational background, health literacy, preferred language and experience using healthcare services.
However, simplifying or translating medical information can introduce new errors. A technical term may have a precise clinical meaning that changes when it is replaced with an everyday expression. Medication instructions can become ambiguous when timing, frequency or duration is simplified incorrectly. Warning signs may lose their urgency if the wording becomes too reassuring.
A study by Martos and colleagues, published in JAMA Network Open in September 2025, provides particularly relevant evidence. The researchers compared machine translation with professional translation of pediatric discharge instructions in four languages, including Vietnamese. The AI system evaluated in this study used a neural machine translation approach rather than the same type of generative language model used to draft discharge instructions [6].
In the evaluated US healthcare setting, AI translations into Vietnamese, Simplified Chinese and Somali performed worse than professional translations across the assessed measures of fluency, adequacy, meaning and error severity. For Vietnamese, errors classified as clinically impactful occurred in 60 of 148 AI-translated sections, approximately 41%, compared with 20 of 148 professionally translated sections, approximately 14% [6].
These results should not be interpreted as the expected error rates of AI systems used in Vietnamese hospitals. They concern one evaluated translation workflow and do not establish the performance of all contemporary generative AI models. Nevertheless, they demonstrate why language quality and clinical meaning require explicit verification.
Healthcare organizations in Viet Nam should evaluate patient-facing AI content in the language patients will actually use. A system tested primarily in English should not be presumed equally reliable when generating or translating Vietnamese instructions.
The assessment should consider whether medical terminology is appropriate, whether the original clinical meaning has been preserved and whether patients with different levels of health literacy can understand the information.
Common English-language readability scores should not be applied mechanically to Vietnamese text. They were developed around characteristics of English and cannot independently establish whether Vietnamese patients can understand or act on the instructions.
Comprehension also depends on presentation. Font size, document structure, the amount of information presented at once and the clarity of medication schedules can influence how easily patients use written instructions.
For older patients, people with visual or cognitive difficulties, patients with limited literacy and those who speak a language other than Vietnamese, additional communication support may be necessary.
A concise document written in familiar language is often preferable to a lengthy explanation filled with medical terminology. However, simplifying the text should never remove essential medication precautions, warning signs or follow-up requirements.
Patient understanding must be checked, not assumed
Receiving a discharge document does not necessarily mean the patient understands it. This is true whether the instructions were prepared by a healthcare professional, produced from a standard template or drafted using AI.
A patient may appear to understand an explanation but later take a medicine incorrectly, miss a follow-up appointment or remain uncertain about what to do when symptoms develop.
The teach-back method offers a practical way to identify misunderstandings before the patient leaves the healthcare facility. Rather than simply asking whether the patient understands, a healthcare professional invites the patient or caregiver to explain the essential instructions in their own words.
For example, the professional may ask which medications have changed, how the patient plans to take them at home or what they would do if an important warning sign occurs.
The purpose is not to test the patient's memory or intelligence. It is to assess whether the information has been communicated clearly enough for the person to use it safely. If a misunderstanding is identified, the professional can explain the information differently and check again.
For tasks requiring practical skills, demonstration or show-back may be more appropriate. A patient may correctly repeat instructions for using an inhaler or measuring blood glucose but still need help performing the task.
The US Agency for Healthcare Research and Quality recommends teach-back and related approaches in its Health Literacy Universal Precautions Toolkit. Its Re-Engineered Discharge Toolkit also emphasizes understandable written care plans, medication review, warning signs, follow-up arrangements and assessment of patient understanding [8,9].
These are international quality-improvement resources, not mandatory Vietnamese AI-discharge procedures. Their value lies in providing practical methods for making discharge communication more effective.
For healthcare organizations using AI-generated instructions, the final quality check should not end when the document is professionally approved. It should also consider whether the patient or caregiver can understand and carry out the essential actions.
Where patients rely on family caregivers, the organization should identify who will actually manage medications or provide care at home and involve that person appropriately, while respecting the patient's rights and confidentiality.
Patient understanding is especially important when discharge involves several medication changes, new medical equipment or a recovery plan that the patient has never followed before.
What Viet Nam's legal framework means for AI-assisted discharge instructions
Viet Nam's Law on Medical Examination and Treatment No. 15/2023/QH15 provides a direct legal foundation for safe discharge communication.
Article 9 establishes patients' rights to receive information and explanations about their health and treatment, including guidance on self-monitoring, care and prevention of complications. Article 44 includes the obligation of practitioners to provide information and counseling in accordance with those rights [10].
Article 77 is particularly relevant to inpatient discharge. Under Clause 6, when a patient leaves the healthcare establishment, the establishment is responsible for completing the medical record, providing self-care guidance, arranging outpatient treatment when necessary, settling costs and providing a discharge certificate. Article 78 also applies the provisions of Article 77 to day treatment [10].
These provisions directly support the importance of appropriate discharge guidance. They should not, however, be presented as prescribing an identical documentation process for every outpatient consultation or ambulatory procedure.
Medication-related responsibilities are addressed in Articles 62 and 63. These provisions require appropriate professional conduct in prescribing and using medicines, including accurate prescribing, relevant medication information and verification during dispensing [10].
For prescriptions involving chemical drugs and biological products in outpatient treatment, Circular No. 26/2025/TT-BYT, effective from 1 July 2025, provides more specific requirements. It also addresses prescribing arrangements for patients who need to continue medication after inpatient discharge [11].
AI-generated explanations or medication schedules should therefore remain consistent with the applicable authorized prescriptions and clinical instructions. They should not be treated as substitutes for prescriptions or professional treatment decisions issued under the relevant requirements.
Circular No. 13/2025/TT-BYT, effective from 21 July 2025, provides guidance on implementing electronic medical records. Where discharge-related documents are incorporated into an electronic medical record, the applicable requirements for documentation, authentication, management and storage must be considered [12].
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 has been effective since 1 March 2026. Article 4 establishes human-centered principles, while Article 6 requires stricter risk management in essential sectors directly affecting life and health, including healthcare. Decree No. 142/2026/NĐ-CP, effective from 1 May 2026, further details the regulatory framework, including risk classification, transparency and the responsibilities of relevant participants [13,14].
The application of these requirements depends on the actual AI system, its intended purpose, functionality, legal classification and the roles of those providing or deploying it. An application should not automatically be described as legally high-risk simply because it helps prepare patient education materials, although errors in those materials may still carry significant clinical consequences.
Where patient information is processed through AI applications, healthcare organizations must also assess their responsibilities under the Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP, both effective from 1 January 2026 [15,16].
These requirements are particularly relevant when medical records are supplied to external AI providers, processed through cloud-based services or used to generate documents distributed through digital channels. Healthcare organizations need to understand which data are processed, the lawful basis and purposes of processing, who can access the information and what protections apply.
The current legal framework does not establish one universal AI-generated discharge instruction checklist or identical AI-specific verification procedure for all Vietnamese healthcare organizations.
The safeguards discussed in this article should therefore be understood as practical approaches to clinical quality and patient safety, implemented alongside applicable legal and professional requirements rather than presented as newly prescribed Ministry of Health procedures.
From AI drafting to instructions patients can safely use
Introducing AI into discharge communication should begin with an understanding of how the organization currently prepares, verifies and provides discharge information.
The first requirement is a sufficiently complete and authorized clinical discharge plan. AI should generate patient-facing instructions from verified information, including the reconciled medication plan, follow-up arrangements, important warning signs and approved aftercare recommendations.
Where information is missing, conflicting or clinically uncertain, the system should support referral to the responsible professional rather than generate advice that appears definitive.
The next requirement is professional verification. Before the instructions are released, an appropriately authorized healthcare professional should confirm that the AI-generated content accurately reflects the intended plan. Medication information, warning signs, follow-up actions and patient-specific precautions deserve particular attention. Pharmacists, nurses and other professionals may contribute according to their responsibilities and the organization's workflow.
The instructions should then be explained in a form suitable for the patient. This may involve a printed document, a secure digital version or both. The patient or caregiver should have an opportunity to ask questions and demonstrate understanding of the most important actions before leaving the facility.
The organization also needs to manage changes in the discharge plan. A prescription may be revised or a follow-up arrangement confirmed after an AI draft has already been prepared. The final instructions should reflect the current approved information, not an earlier version.
If a clinically significant mistake is discovered after the document has been issued, correcting the electronic record may not be sufficient. The healthcare team should consider whether the patient or caregiver has already acted on the original instructions and arrange appropriate communication of the correction.
This is particularly relevant when documents have been printed, downloaded or shared through messaging applications. An updated version in the hospital system may not automatically replace a copy held by the patient.
A controlled process should therefore connect the authorized clinical plan, AI drafting, professional review, patient education and management of important corrections without creating unnecessary administrative burden.
Implementing AI-assisted discharge instructions in Vietnamese healthcare settings
For healthcare organizations in Viet Nam, a practical starting point is to select a limited clinical service where discharge planning is reasonably standardized, the necessary information is available and professionals can review AI-generated instructions before they reach patients.
The initial application could focus on preparing plain-language explanations from clinician-approved aftercare plans. It does not need to involve AI independently deciding medication changes or generating new treatment recommendations.
Before deployment, the organization should evaluate the application using representative local discharge cases. These should include routine situations as well as cases involving medication changes, multiple chronic conditions, significant warning signs, pending test results and incomplete information.
Testing should establish whether the AI accurately preserves medication names, strengths, doses, timing, duration and instructions to discontinue treatment. It should also assess whether the correct warning signs are included, whether the recommended actions match the approved clinical plan and whether the language is suitable for the intended patients.
Cases with conflicting or missing information are particularly useful. The organization needs to know whether the system identifies uncertainty or produces unsupported advice when the source material is inadequate.
Where feasible, evaluation should involve patients or caregivers who are representative of the intended users. Healthcare professionals may consider a document clear while patients still struggle to understand what they should do at home.
After implementation, the organization can examine clinically significant medication discrepancies, missing or inaccurate warning signs, errors in follow-up instructions and the proportion of patients who can correctly explain essential actions through teach-back.
Time saved in drafting may also be measured, but it should be considered alongside the time required for professional verification and patient education. A faster documentation process is not necessarily an improvement if it creates additional confusion or risk after discharge.
Particular attention should be given to patients with complex medication regimens, older people who depend on caregivers and those requiring language or communication assistance. The same AI system may not perform equally well for every patient group.
Monitoring should distinguish minor wording problems from errors that could affect treatment, delay urgent care or contribute to medication harm. Clinically significant errors should be reviewed to understand whether they originated in inaccurate source information, AI generation, professional verification or the communication process.
Healthcare organizations need to remain realistic about their resources. A smaller hospital may not have a dedicated AI governance team or sophisticated monitoring infrastructure. Nevertheless, it can establish a clearly identified source of approved clinical information, professional review responsibilities, a method for resolving discrepancies and a manageable process for learning from significant errors.
These arrangements can be incorporated into existing discharge planning, medication safety, nursing education and quality-improvement activities.
From discharge information to a plan patients can safely follow
AI-generated discharge and aftercare instructions may help healthcare professionals reduce repetitive documentation and communicate clinical information in a clearer, more accessible way. For patients, this could mean receiving a better organized explanation of their medicines, recovery activities and follow-up arrangements.
However, the quality of these instructions cannot be judged simply by their readability, length or professional appearance. A document written in friendly language may still contain an incorrect medication dose, omit a warning sign or fail to explain what the patient should do when recovery does not proceed as expected.
For healthcare organizations in Viet Nam, three questions deserve particular attention. Does the document accurately reflect the patient's approved medication and treatment plan? Does it explain which symptoms require further medical attention and how to obtain help? Can the patient or caregiver understand and carry out the essential instructions?
Answering these questions requires reliable clinical information, appropriate professional verification, clear communication and confirmation of patient understanding.
AI can support this process, but it should not replace the clinical decisions and professional responsibilities on which safe discharge planning depends.
The goal is not simply to give patients better-written discharge instructions. It is to help them leave the healthcare facility knowing which medicines to take, what to do at home, when to seek help and how to manage their recovery as safely as possible.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, prescribing, legal, regulatory, personal-data-protection, accreditation or quality-management advice.
The approaches described are practical organizational recommendations, not a prescribed Vietnamese Ministry of Health AI-discharge framework, mandatory AI verification checklist or international certification standard.
Applicable responsibilities depend on the healthcare service, the AI system's intended purpose and actual functionality, professional scope of practice, prescription and documentation requirements, patient circumstances and relevant legal provisions.
The international studies discussed involve different models, clinical settings, source documents and evaluation methods. Their findings should not be generalized automatically to other AI applications or healthcare organizations, or interpreted as estimates of actual patient harm in Viet Nam.
Recommendations concerning medication reconciliation, clinical verification, warning-sign accuracy, patient-friendly language, teach-back, document updates and communication of corrections are intended to support existing clinical and patient-safety responsibilities.
References
Stanceski K, Zhong S, Zhang X, et al. The Quality and Safety of Using Generative AI to Produce Patient-Centred Discharge Instructions. npj Digital Medicine. 2024;7:329. DOI: 10.1038/s41746-024-01336-w.
Tang M, Markova K, Stanceski K, et al. Assessing the Safety of Patient-Centred Discharge Medication Instructions Generated by an AI Model. International Journal of Medical Informatics. 2026;209:106272. DOI: 10.1016/j.ijmedinf.2026.106272.
Holderried F, Sonanini A, Stegemann-Philipps C, et al. Impact of GPT-4–Generated Discharge Letters on Patients' Medical Comprehension: Prospective Crossover Study. Journal of Medical Internet Research. 2026;28. DOI: 10.2196/81243.
Wang M, Ma H, Piao M. Effectiveness of Large Language Models in Preoperative and Discharge Education: A Systematic Review Based on an Evaluation Framework. npj Digital Medicine. 2026;9:122. DOI: 10.1038/s41746-025-02302-w.
Hernández P, Rodriguez G, Fischetti C, et al. Evaluating the Performance of Large Language Models for Generating Emergency Department Discharge Instructions. The Journal of Emergency Medicine. 2026;85:45–55. DOI: 10.1016/j.jemermed.2026.02.042.
Martos M, Fields B, Finlayson SG, et al. Accuracy of Artificial Intelligence vs Professionally Translated Discharge Instructions. JAMA Network Open. 2025;8(9). DOI: 10.1001/jamanetworkopen.2025.32312.
World Health Organization. Medication Safety in Transitions of Care. Geneva: World Health Organization; 2019. WHO/UHC/SDS/2019.9.
Agency for Healthcare Research and Quality. Health Literacy Universal Precautions Toolkit. Third edition. Edited by Cindy Brach. Rockville, MD: AHRQ; 2024. AHRQ Publication No. 23-0075.
Agency for Healthcare Research and Quality. Re-Engineered Discharge (RED) Toolkit. Rockville, MD: AHRQ. Professional discharge planning and patient education resources.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted 9 January 2023; effective 1 January 2024.
Ministry of Health of Viet Nam. Circular No. 26/2025/TT-BYT on Prescriptions and Prescribing of Chemical Drugs and Biological Products in Outpatient Treatment at Medical Examination and Treatment Establishments. Issued 30 June 2025; effective 1 July 2025.
Ministry of Health of Viet Nam. Circular No. 13/2025/TT-BYT Providing Guidance on the Implementation of Electronic Medical Records. Issued 6 June 2025; effective 21 July 2025.
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. Issued 30 April 2026; effective 1 May 2026.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.