Governing AI-Assisted Early Recognition of Postpartum Hemorrhage in Viet Nam: Objective Blood Loss Measurement, Timely Clinical Response and Maternal Safety Monitoring

Governing AI-Assisted Early Recognition of Postpartum Hemorrhage in Viet Nam: Objective Blood Loss Measurement, Timely Clinical Response and Maternal Safety Monitoring

A woman has just given birth vaginally at a hospital. Her initial recovery appears uneventful, but the attending midwife notices that bleeding is continuing. The measured blood loss has reached 320 mL, and her pulse has increased to 108 beats per minute.

Earlier in labor, an artificial intelligence (AI) system classified her as having a relatively low predicted risk of postpartum hemorrhage. No new high-risk alert has appeared.

Should the clinical team wait for the system to update its prediction?

No. The combination of objectively measured blood loss and an elevated pulse meets the early recognition criteria in current international guidance and is consistent with Viet Nam's updated national approach to recognizing hemorrhage after vaginal birth. The clinical team should assess the woman and initiate an appropriate response under the hospital's approved obstetric hemorrhage protocol.

The situation illustrates a distinction that matters in maternal care. AI may help identify women who could be at increased risk of postpartum hemorrhage, but preventing serious harm depends on recognizing bleeding as it occurs and delivering appropriate treatment without delay.

For healthcare organizations in Viet Nam, the question is not simply whether an AI model can predict postpartum hemorrhage accurately. It is whether the technology can contribute useful information without weakening established practices for measuring blood loss, monitoring maternal condition and responding to emergencies.

Even a well-performing prediction model has limited clinical value if significant bleeding is not recognized or the care team cannot respond when treatment is needed.

Why Postpartum Hemorrhage Remains a Maternal Safety Priority

Postpartum hemorrhage (PPH) remains a major cause of maternal morbidity and mortality worldwide. Effective preventive and treatment interventions are available, yet serious complications continue to occur when excessive bleeding is recognized late or appropriate care is delayed.

Severe hemorrhage can lead to circulatory shock, coagulopathy, organ injury, emergency surgery, hysterectomy or death. Women who survive may experience anemia, prolonged recovery, psychological distress and other consequences affecting their health and well-being.

The underlying causes vary. Uterine atony is common, but genital tract trauma, retained placental tissue and coagulation disorders are also important. More than one cause may be present, and a woman's condition can deteriorate rapidly when bleeding continues.

Recognition therefore requires more than recording a single blood loss measurement. Clinicians need to consider cumulative blood loss, the rate of ongoing bleeding, physiological changes and the likely underlying cause.

Postpartum hemorrhage also cannot always be predicted before birth. Some women experience severe bleeding despite having no obvious antenatal risk factors.

Risk assessment can support preparation, including reviewing delivery arrangements, monitoring needs and the availability of appropriate clinical expertise. However, separating women into high-risk and low-risk groups cannot replace routine surveillance and emergency readiness for every birth.

A prediction made before delivery represents an estimate based on the information available at that time. It cannot reliably determine how the woman's condition will evolve during labor, delivery or the immediate postpartum period.

For hospitals, some of the most immediate opportunities to improve maternal safety may involve recognizing abnormal bleeding earlier, maintaining reliable clinical observations and reducing delays in treatment.

Any proposed AI application should be evaluated against those practical needs.

Viet Nam's Updated National Guidance Provides the Clinical Foundation

An important development for maternal care in Viet Nam is Ministry of Health Decision No. 1098/QD-BYT, issued on 21 April 2026.

The Decision introduced the national professional guidance National Guidelines for Reproductive Healthcare Services: Obstetric Hemorrhage (Hướng dẫn Quốc gia các dịch vụ chăm sóc sức khỏe sinh sản về nội dung Chảy máu sản khoa).

The guidance applies to healthcare examination and treatment facilities nationwide. It replaces specified sections of the national reproductive healthcare guidance issued under Decision No. 4128/QD-BYT in 2016, rather than replacing the earlier Decision in its entirety.

The updated document addresses postpartum hemorrhage alongside other obstetric bleeding conditions. It provides recommendations concerning recognition, assessment, staged management, treatment according to the underlying cause and multidisciplinary coordination when necessary.

For vaginal birth, the guidance incorporates objective blood loss measurement and early recognition criteria involving blood loss of at least 500 mL, or at least 300 mL accompanied by abnormal hemodynamic findings.

These findings include a pulse above 100 beats per minute, an elevated shock index, systolic blood pressure below 100 mmHg or diastolic blood pressure below 60 mmHg.

Particular attention is given to the first two hours after birth, although the guidance addresses hemorrhage occurring within the first 24 hours.

The national document also distinguishes vaginal birth from cesarean delivery. Its introductory section refers to a blood loss threshold of at least 1,000 mL in the cesarean context, while the detailed postpartum hemorrhage section emphasizes earlier recognition using measured blood loss and hemodynamic findings.

These passages need to be interpreted together. Blood loss during cesarean delivery presents different measurement challenges, and concerning clinical findings may require action well before a higher volume threshold is reached.

Another detail deserves attention when translating the document into clinical software. The postpartum hemorrhage definition uses a shock index threshold of at least 1, while the staging section uses a threshold greater than 1. The accompanying appendix also discusses the interpretation of shock index values.

Such differences may have little impact on an experienced clinician's response to an obviously deteriorating patient, but they matter when numerical comparisons are converted into automated alert rules.

Hospitals should therefore establish clear, professionally approved clinical criteria before incorporating them into electronic decision support.

Importantly, the national guidance recognizes that rapidly continuing bleeding, bright-red blood loss or patient-specific clinical concerns may justify intervention even when the recorded volume has not reached a conventional threshold.

An AI system should not discourage clinical action in these circumstances.

The Ministry of Health's guidance provides the national professional foundation for care. AI may be evaluated as a supporting technology, but it should not establish a competing set of clinical criteria or independently determine when emergency treatment is necessary.

What the 2025 WHO–FIGO–ICM Guidelines Recommend

The consolidated guidelines published by the World Health Organization (WHO), International Federation of Gynecology and Obstetrics (FIGO) and International Confederation of Midwives (ICM) on 5 October 2025 represent a major international update.

They bring together 51 evidence-based recommendations covering prevention, diagnosis, treatment, supportive care and health-system implementation.

One important change concerns the criteria for recognizing postpartum bleeding that warrants first-response treatment.

The guidelines recommend acting when objectively measured blood loss reaches at least 300 mL together with any specified abnormal hemodynamic sign, or reaches at least 500 mL, whichever occurs first within 24 hours after birth.

The specified abnormalities include a pulse above 100 beats per minute, shock index above 1, systolic blood pressure below 100 mmHg or diastolic blood pressure below 60 mmHg.

Particular vigilance is recommended during the first two hours after birth.

These criteria are intended to support earlier recognition than approaches that rely exclusively on the traditional 500 mL threshold.

They are informed by international evidence examining the relationship between blood loss, physiological findings and severe maternal outcomes.

A WHO individual participant data meta-analysis published in The Lancet in 2025 analyzed information from 312,151 women across 12 datasets.

The analysis found that combining lower blood loss thresholds with abnormal hemodynamic findings improved the identification of women at risk of maternal death or severe morbidity compared with relying on blood loss volume alone.

This provides an important evidence base for the updated international approach.

However, the thresholds should not be interpreted as instructions to delay assessment until a numerical criterion is satisfied.

Rapid bleeding, concerning symptoms, incomplete measurements or evidence of deterioration may require immediate clinical action.

Applying these recommendations during cesarean delivery also requires consideration of surgical circumstances and the limitations of blood loss measurement in the operating room.

The practical implication is that objective measurement, physiological assessment and professional judgment must function together.

An AI prediction should support this process where it can demonstrate value, rather than replace the clinical criteria that determine when a woman requires urgent attention.

Objective Blood Loss Measurement Comes Before AI

A hospital does not need artificial intelligence to make one of the most important improvements in postpartum hemorrhage recognition.

It can begin by measuring blood loss more reliably.

Visual estimation is often inaccurate. Blood may be absorbed into pads, collected beneath the woman or spread across delivery surfaces. During cesarean delivery, blood may also be mixed with amniotic fluid or irrigation fluids.

These circumstances can make clinically significant bleeding difficult to recognize.

A review published in The Lancet in July 2026 reported that subjective visual estimation missed approximately 52% of postpartum hemorrhage diagnoses following vaginal birth in the evidence analyzed.

Objective measurement provides a more reliable basis for recognizing excessive bleeding, although no practical method is perfectly accurate.

For vaginal birth, a calibrated blood-collection drape can help quantify postpartum blood loss. Weighing blood-soaked materials may provide additional information when the method is standardized and the dry weight of the materials is accounted for.

During cesarean delivery, assessment may involve measuring suctioned fluid, accounting for amniotic and irrigation fluids and weighing blood-soaked surgical materials.

The method needs to reflect the clinical setting, and its results must be interpreted alongside the woman's physiological condition.

Measurement should also continue beyond the initial observation.

A woman may lose a moderate amount of blood immediately after birth and then experience further bleeding during recovery. If subsequent blood loss is not recorded, the cumulative volume may be underestimated.

This creates an important responsibility during handover between the delivery room, operating room, recovery area and postpartum ward.

Relevant information should remain available to clinicians responsible for ongoing monitoring and treatment.

These processes require suitable equipment, staff training, clear documentation and reliable communication. They do not depend on machine learning.

AI could potentially help analyze blood loss trends, identify inconsistencies or combine measurements with other clinical information. However, such functions require evaluation, and their usefulness depends on the quality and timeliness of the data.

If an additional 200 mL of bleeding is not entered into the clinical record, a prediction system may continue using an outdated measurement even though the woman's condition has changed.

The problem in that situation may lie in the measurement and documentation process rather than the predictive algorithm.

Before introducing AI, hospitals should establish whether blood loss is being measured, recorded and communicated reliably enough to support timely clinical decisions.

Predicting Hemorrhage Risk Is Different From Recognizing Active Bleeding

AI applications related to postpartum hemorrhage serve different purposes, and those functions should not be confused.

Some models estimate risk before delivery using maternal characteristics, obstetric history, laboratory findings and other clinical information.

Their potential role is to support preparation. A woman with an elevated predicted risk may require additional planning around monitoring, clinical expertise or the availability of blood products and emergency resources.

Other systems analyze changing information during labor or immediately after birth. They may incorporate new observations to identify patterns suggesting an increasing probability of hemorrhage or physiological deterioration.

A third category focuses on retrospective surveillance. These applications may examine clinical records to identify possible hemorrhage events, incomplete documentation or cases requiring quality review.

Each function has different evidence requirements and patient-safety implications.

A model developed for antenatal risk stratification should not automatically be used as a real-time emergency detection system. Likewise, software designed to identify cases from completed medical records cannot be assumed suitable for issuing time-critical clinical alerts.

It is also important to distinguish AI from conventional electronic decision rules.

A system that displays an alert whenever blood loss reaches a specified threshold may be clinically useful, but it does not necessarily involve artificial intelligence.

Hospitals should avoid attributing improvements resulting from established monitoring procedures or electronic threshold alerts to AI unless the additional contribution of the AI component has been evaluated separately.

The distinction becomes especially important when bleeding has already begun.

A woman who received a low-risk prediction before delivery may subsequently develop uterine atony or genital tract trauma.

At that point, observed bleeding and physiological condition must guide the response. The earlier prediction does not provide reassurance that the developing hemorrhage is harmless.

A high-risk prediction may support preparation, but a low-risk prediction must never reduce the standard of postpartum monitoring.

A Clinical Alert Matters Only When the Team Can Respond

An alert about possible postpartum hemorrhage has value when it reaches the appropriate professionals and contributes to timely clinical action.

When significant bleeding is suspected, the responsible clinicians need to assess the woman, verify relevant observations and initiate treatment according to the hospital's approved protocol.

Emergency care should not depend on a lengthy sequence of electronic acknowledgments.

The WHO–FIGO–ICM guidelines recommend a coordinated first-response treatment bundle commonly described through the MOTIVE approach.

The bundle includes uterine massage, appropriate uterotonic treatment, tranexamic acid, intravenous fluids, examination to identify the cause of bleeding and escalation when required.

These interventions are intended to be delivered promptly and in coordination, according to clinical indications, contraindications, professional responsibilities and applicable protocols.

The guidelines also recommend early intravenous tranexamic acid for postpartum hemorrhage within three hours of birth when indicated.

Discussing these interventions does not mean that an AI system should prescribe medication or independently direct invasive treatment.

Instead, the bundle demonstrates what a healthcare organization needs to be capable of delivering when hemorrhage is recognized.

A functioning response requires trained personnel, accessible medicines, equipment for intravenous access and monitoring, clear communication and appropriate escalation arrangements.

More severe hemorrhage may require anesthesiology, resuscitation, laboratory testing, blood and blood products, operative intervention or transfer to a higher-level facility.

An AI-generated warning cannot compensate for unavailable medicines, delayed blood products or unclear emergency responsibilities.

For hospitals, this is why clinical readiness should be examined before software procurement.

The objective is to reduce the interval between clinically important bleeding, recognition and appropriate treatment, rather than simply generate an earlier electronic notification.

What Strong Clinical Evidence Already Demonstrates

One of the most important studies of postpartum hemorrhage recognition and treatment is the E-MOTIVE trial, published in The New England Journal of Medicine in 2023.

The international cluster-randomized trial involved 80 secondary-level hospitals in Kenya, Nigeria, South Africa and Tanzania, with 210,132 women undergoing vaginal delivery.

The intervention combined objective blood loss measurement using a calibrated collection drape with a standardized first-response treatment bundle and an implementation strategy.

The primary outcome was a composite of severe postpartum hemorrhage, laparotomy for bleeding or maternal death caused by bleeding.

Among participants with available primary outcome data, this composite occurred in 1.6% of the intervention group compared with 4.3% of the usual-care group.

Hemorrhage detection and adherence to the treatment bundle were also substantially better in the intervention group.

These findings provide strong evidence that coordinated early detection and treatment can reduce serious maternal complications in the settings studied.

However, E-MOTIVE was not a trial of artificial intelligence.

Its results cannot be presented as evidence that AI prediction reduces severe postpartum hemorrhage, hysterectomy or maternal mortality.

The intervention involved objective measurement, standardized clinical treatment and changes in how care was organized and delivered.

This distinction is particularly important for hospitals considering digital technologies.

Evidence that a clinical pathway improves outcomes does not automatically establish that adding AI to that pathway provides further benefit.

A new prediction system should be assessed against the existing evidence-based process, rather than evaluated as though the hospital had no established method of recognizing hemorrhage.

The 2026 Lancet review on postpartum hemorrhage similarly emphasizes that early diagnosis and effective treatment are closely connected.

Delays may occur at several points, including blood loss measurement, recognition of abnormal findings, availability of medication, access to blood products and definitive control of bleeding.

These delays provide practical targets for improvement.

AI should be considered where it may help address a clearly defined clinical or operational problem and where its additional benefit can be meaningfully evaluated.

What the Latest AI Research Shows

Research on AI-assisted postpartum hemorrhage prediction has expanded rapidly, but the evidence remains concentrated in model development and retrospective evaluation.

A systematic review by AlSaad and colleagues, published in Frontiers in Global Women's Health in June 2026, included 33 eligible studies published between 2015 and 2025.

Approximately 91% of the studies were retrospective, 55% were conducted at a single site and 76% relied primarily on internal validation.

Most models focused on anticipatory risk stratification, while fewer examined intrapartum or immediate postpartum warning systems.

The authors identified substantial variation in outcome definitions, data sources, prediction targets and validation methods.

These limitations are important because a model trained to predict hemorrhage using one definition may not perform similarly when evaluated against another.

A model based on visually estimated blood loss, for example, may not predict the same outcome as one trained using objective measurements. Diagnostic codes, transfusion records and treatment interventions can also produce different outcome labels.

The review concluded that progress toward clinical readiness requires more consistent outcome definitions, stronger external validation and prospective evaluation.

Research published during 2026 offers further insight into these challenges.

A study by Kovacheva and colleagues, published online in December 2025 and included in the January 2026 issue of Pregnancy, evaluated the external performance of a previously developed machine learning model across eight hospitals in the northeastern United States.

The study included 87,662 deliveries.

The original model achieved an area under the receiver operating characteristic curve (AUROC) of 0.60, indicating substantially reduced discrimination compared with expectations based on its earlier development.

When the researchers refitted a model using local data, discrimination improved to an AUROC of 0.75. However, the revised model overestimated risk across a range of predicted probabilities.

The study defined postpartum hemorrhage using estimated blood loss of at least 1,000 mL at delivery or blood transfusion within 24 hours after birth.

These findings provide a practical example of why performance demonstrated during model development cannot be assumed to transfer reliably to another healthcare system.

Another study, published by Akazawa and colleagues in the Journal of Obstetrics and Gynaecology Research in 2026, examined machine learning risk stratification using a nationwide Japanese cohort of 1,224,990 women undergoing vaginal delivery.

The researchers compared predictions made before delivery with updated estimates incorporating information obtained during labor.

The study used blood loss of at least 1,000 mL as its postpartum hemorrhage outcome, which differs from the earlier clinical recognition thresholds discussed in current WHO guidance.

In temporal validation, the predelivery model achieved an AUROC of 0.676, increasing to 0.703 when intrapartum information was incorporated.

At the selected alert threshold, the updated model identified approximately five to six additional hemorrhage cases per 1,000 deliveries, with a modest increase in the number of women classified as high-risk.

The results suggest that updating risk estimates during labor may improve detection, but the gain was limited and important cases remained unidentified.

The study did not demonstrate that deploying the model in routine clinical practice reduced severe maternal outcomes.

These findings help place AI prediction in a realistic clinical context.

A model may offer useful risk information without detecting every case, and improvements in statistical performance do not necessarily translate into better clinical outcomes.

Even external validation cannot establish that clinicians will respond appropriately to alerts or that treatment will occur sooner.

For Vietnamese hospitals, the available evidence supports careful evaluation of AI applications, not an assumption that a commercially available prediction system is already a proven maternal safety intervention.

Evaluating AI in Vietnamese Obstetric Settings

A model developed in another healthcare system may behave differently when introduced into a Vietnamese hospital.

Patient characteristics, referral patterns, delivery practices, laboratory availability, monitoring routines and electronic clinical records can differ substantially.

The information available at a specialized maternity hospital may also differ from that collected at a provincial hospital, general hospital or smaller maternity facility.

These differences affect both the model's inputs and the events it is expected to predict.

A model trained predominantly on vaginal births should not automatically be extended to cesarean deliveries. Likewise, performance demonstrated in one maternal population cannot be assumed to apply to women with substantially different clinical characteristics.

Before local validation, the hospital should establish a clinically reliable reference standard for postpartum hemorrhage.

This may require reviewing objectively measured blood loss, physiological observations, clinical findings and treatment records.

Diagnostic codes alone may be insufficient when recognition and documentation practices are inconsistent.

Data timing also deserves careful attention.

A prediction should be evaluated using only information that would actually have been available at the intended prediction time.

A retrospective model may appear highly accurate if it includes laboratory results obtained after hemorrhage began, medications ordered in response to bleeding or documentation completed after the diagnosis was already apparent.

In such cases, the algorithm may be recognizing an event that is already known rather than predicting it early enough to influence care.

Local validation should therefore preserve the timing of clinical information and reflect the workflow in which the system would operate.

Discrimination, calibration, sensitivity, specificity and performance at clinically meaningful alert thresholds should be considered together.

False-negative predictions are especially important in a time-sensitive condition. A model that misses serious hemorrhage may create additional danger if staff begin to rely on its low-risk classifications.

False-positive alerts also have consequences. Frequent unnecessary warnings can increase workload, reduce attention to important findings or contribute to inappropriate clinical interventions.

Hospitals should evaluate how many alerts would be generated, how many would be clinically meaningful and whether the clinical team has the capacity to review them.

An equally important question is whether AI identifies risks that existing monitoring processes would otherwise miss.

If a conventional rule based on measured blood loss and vital signs identifies the same women at the same time, a more complex model may provide limited additional value.

The appropriate comparison is therefore not simply AI versus no prediction. It is the AI-supported workflow compared with the clinical process already available.

Hospitals should also examine how model performance may change over time. Modifications to documentation, referral patterns, patient populations or the electronic record system can alter the information available to an algorithm.

Validation before deployment does not guarantee continued reliability.

Human Oversight During an Obstetric Emergency

Human oversight is essential when AI may influence clinical decisions, particularly during emergencies.

However, oversight should not be understood as requiring clinicians to complete unnecessary administrative checks before taking appropriate action.

When postpartum hemorrhage is suspected, appropriately qualified healthcare professionals must be able to act immediately on the woman's clinical condition.

An algorithm should not prevent escalation because its predicted probability remains below a predefined threshold.

Nor should it independently initiate medication administration or invasive procedures without the necessary professional authorization and clinical safeguards.

The appropriate form of oversight depends on the function of the system.

A model used for antenatal risk assessment may allow time for structured clinical review and preparation.

A system intended to support real-time hemorrhage recognition requires a faster process, including clear responsibility for receiving and reviewing alerts.

The hospital should also determine what happens when the AI output and bedside assessment disagree.

If a midwife identifies continuing bleeding and abnormal vital signs despite a low-risk prediction, the clinical findings should guide the immediate response.

When an AI alert is not supported by the available clinical evidence, the team should assess the woman and manage the concern under the approved protocol.

Repeated discrepancies may indicate problems with data completeness, model calibration, alert timing or the suitability of the model for the local population.

These cases should inform further evaluation.

The purpose of human oversight is to ensure that AI contributes to clinical decisions without weakening professional responsibility or delaying necessary treatment.

Emergency Preparedness, Teamwork and Referral

Postpartum hemorrhage management depends on coordinated action among professionals with different responsibilities.

Midwives and nurses are often the first to observe continuing bleeding, changes in vital signs or deterioration during immediate postpartum care.

Obstetric clinicians assess the likely cause and direct management within their professional responsibilities.

Anesthesiology, resuscitation, laboratory, blood transfusion and operating room teams may become involved when bleeding becomes severe.

The effectiveness of this response depends on responsibilities being established before an emergency occurs.

Healthcare organizations should define how hemorrhage is recognized, who can activate the emergency response, how additional assistance is obtained and how treatment is coordinated.

These arrangements need to reflect the facility's staffing, equipment, professional scope of practice and referral capabilities.

Multidisciplinary simulation can help teams examine how the pathway functions under time pressure.

A simulation may reveal that calibrated blood collection equipment is unavailable in certain delivery rooms, that emergency medicines are difficult to access or that communication with the blood bank takes longer than expected.

Such weaknesses should be addressed regardless of whether AI is introduced.

The WHO implementation guide published in February 2026 emphasizes systematic preparation, stakeholder engagement, workforce capacity, availability of essential commodities, monitoring and continuous improvement.

These principles are relevant to Vietnamese healthcare organizations.

Facilities with limited emergency resources may need to rely on referral to higher-level hospitals for advanced treatment.

In these circumstances, reliable recognition and timely transfer arrangements may provide greater immediate benefit than sophisticated predictive technology.

An AI-generated probability should not independently determine whether a woman requires transfer.

The decision must reflect her clinical condition, available treatment, transport safety and the receiving facility's capacity.

Relevant information about measured blood loss, physiological observations, interventions already performed and continuing concerns should accompany the referral.

The referring facility remains responsible for appropriate care and stabilization within its capabilities while transfer arrangements are being made.

Digital technology can support communication, but it cannot replace the clinical responsibilities of the referring and receiving teams.

Measuring Whether Maternal Safety Improves

The introduction of AI should be accompanied by a credible evaluation of whether care becomes safer.

The starting point is the reliability of existing measurement and response processes.

Hospitals may examine whether objective blood loss measurement is performed consistently, whether cumulative values are recorded accurately and whether postpartum physiological monitoring identifies concerning changes.

Once hemorrhage is suspected or recognized, attention should shift to the clinical response.

Relevant measures may include the interval between recognition and the first indicated intervention, timeliness of escalation, adherence to the approved treatment bundle and availability of necessary resources.

These measures should reflect what actually happens during care.

An electronic alert acknowledgment does not necessarily mean that treatment has begun.

Similarly, a documented medication order does not establish that the medicine was administered, and a transfer request does not prove that the patient reached the receiving facility safely.

Outcome monitoring may consider severe postpartum hemorrhage, blood transfusion requirements, emergency surgical interventions, hysterectomy, intensive care admission and maternal mortality, where appropriate.

The selection of measures should reflect the organization's services, available data and the frequency of relevant outcomes.

Because some severe events are uncommon, individual facilities may require longer observation periods or collaborative evaluations involving multiple hospitals.

Changes in reported incidence also need careful interpretation.

When a hospital introduces objective blood loss measurement and earlier recognition criteria, more hemorrhage cases may be documented.

An increase in recorded cases may therefore reflect better detection rather than a genuine increase in severe bleeding.

Conversely, a lower reported incidence may reflect incomplete surveillance or documentation rather than improved patient safety.

Outcome trends should be examined alongside patient characteristics, delivery methods, staffing, treatment resources and concurrent quality-improvement activities.

Hospitals should also monitor unintended consequences of AI use, including excessive alerts, inappropriate interventions, increased staff workload and disruption of clinical routines.

Respectful maternal care remains important during emergencies. Communication with the woman and her family, whenever clinically feasible, should remain part of the response.

When AI is introduced alongside improved staff training, revised protocols or new measurement equipment, the hospital should distinguish the contribution of these changes from any additional benefit attributable to AI.

A reduction in severe outcomes following implementation does not, by itself, establish that the prediction system caused the improvement.

Learning From Severe Hemorrhage and Near-Miss Events

Maternal safety improves when healthcare organizations examine what happened during actual emergencies.

A review following severe postpartum hemorrhage should consider when bleeding began, how it was measured, when abnormal findings were recognized and how the clinical response unfolded.

The review should examine whether necessary personnel, medicines, equipment and blood products were available, and whether delays occurred during communication, treatment or transfer.

Where AI was involved, reviewers should determine what information was available to the model, when the prediction was generated and whether its output influenced clinical decisions.

An absent AI alert does not necessarily mean that the model caused a delayed response. Established clinical monitoring should remain capable of recognizing hemorrhage independently.

Similarly, a high-risk prediction does not prove that a subsequent complication was necessarily preventable.

Near-miss events may provide particularly useful information.

A woman may experience substantial bleeding but avoid severe harm because the clinical team recognized deterioration quickly and delivered effective treatment.

Understanding which safeguards worked can help hospitals reinforce good practices.

A constructive review process should encourage clinicians to report inaccurate predictions, confusing alerts, communication difficulties and operational limitations.

The objective is to identify opportunities for improvement, not to assign responsibility based solely on retrospective algorithmic outputs.

Legal and Professional Responsibilities in Viet Nam

AI-assisted postpartum hemorrhage recognition must be considered within Viet Nam's legal framework governing healthcare practice, artificial intelligence, medical devices and personal data protection.

The Law on Medical Examination and Treatment No. 15/2023/QH15 provides the broader legal basis for professional healthcare practice, patient rights, medical records and the responsibilities of healthcare organizations.

Ministry of Health Decision No. 1098/QD-BYT, dated 21 April 2026, provides directly relevant national professional guidance for obstetric hemorrhage.

Healthcare organizations should ensure that their approved clinical procedures reflect applicable Vietnamese guidance, authorized professional responsibilities and the facility's emergency response capabilities.

The introduction of AI does not transfer responsibility for clinical assessment, treatment or emergency management away from appropriately qualified healthcare professionals.

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on 1 March 2026.

Decree No. 142/2026/ND-CP, effective from 1 May 2026, provides detailed implementing provisions.

Decision No. 33/2026/QD-TTg, effective from 15 August 2026, establishes the current list of high-risk AI systems.

The healthcare entries in that list concern specified AI-assisted surgical and robotic surgical systems. AI-assisted postpartum hemorrhage risk prediction is not separately identified by name.

However, the absence of a specifically named entry does not automatically establish that every such application is exempt from applicable AI requirements or presents little clinical risk.

A product-specific legal assessment must consider its intended purpose, actual functions, degree of automation, deployment arrangements and applicable statutory criteria.

Formal regulatory classification and clinical risk are related but distinct questions.

A system that provides inaccurate information during an obstetric emergency may create significant patient-safety consequences regardless of how it is formally categorized.

Healthcare organizations should also determine whether a particular product falls within the applicable medical device regulatory framework.

Decree No. 98/2021/ND-CP on medical device management, as amended, and Ministry of Health Circular No. 24/2026/TT-BYT may be relevant depending on the intended medical purpose and characteristics of the product.

An application used solely for retrospective quality reporting may have a different regulatory position from software promoted for real-time clinical prediction or treatment decision support.

The presence of AI alone does not determine whether software is a medical device. Equally, describing a system as a dashboard or administrative tool does not automatically place it outside medical device regulation.

The actual intended medical purpose and functions require assessment.

Personal data protection is another significant consideration.

AI-assisted maternal care systems may process identifiable information concerning pregnancy, obstetric history, laboratory findings, physiological observations, diagnoses and treatment.

The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/ND-CP, both effective from 1 January 2026, establish relevant requirements for personal data processing and protection.

Decree No. 102/2025/ND-CP on health data management, effective from 1 July 2025, is also relevant to the governance, protection, processing and use of health data.

Healthcare organizations should determine the applicable legal grounds and purposes for processing patient information and establish safeguards appropriate to the actual activities involved.

These may include access restrictions, information security controls, data retention arrangements, contractual responsibilities, incident response procedures and applicable assessment or documentation requirements.

Where an external technology provider processes clinical information, the hospital should clarify responsibilities for access, storage, maintenance, security incidents and permitted uses of patient data.

Secondary use of clinical records for model development, retraining, research or commercial purposes requires separate consideration of the applicable legal and ethical requirements.

Authorization to process information for patient care should not automatically be interpreted as authorization for every subsequent purpose.

Cross-border data processing, overseas hosting and access by external service providers also require review under relevant legal provisions.

These matters should be addressed before identifiable patient information is transferred to an external AI platform.

International clinical guidelines may inform professional practice, but they do not automatically become binding Vietnamese legal requirements.

Similarly, organizational safeguards recommended for responsible AI implementation should be distinguished from specific statutory obligations unless the applicable legal provisions make them mandatory.

A Practical Implementation Approach for Vietnamese Hospitals

A hospital considering AI-assisted postpartum hemorrhage recognition should begin with a review of its existing maternal safety process.

The review can examine recent deliveries to determine whether blood loss was measured appropriately, whether cumulative measurements were maintained and whether concerning clinical signs were recognized.

It should also consider how promptly treatment began, whether staff followed approved hemorrhage procedures and whether necessary medicines, equipment and blood products were available.

The findings may reveal problems that can be addressed without AI.

Inconsistent objective measurement, incomplete documentation, delayed communication or limited emergency resources may be more important barriers than the absence of a prediction model.

Once the baseline is understood, the organization can define a specific function for AI evaluation.

One possible use case is supporting preparation before delivery by identifying women who may need additional resources or monitoring.

Another is examining whether updated risk predictions provide useful information during labor or the immediate postpartum period beyond what is already available from objective blood loss measurement and routine physiological assessment.

The intended patient population, prediction timing, data sources, clinical response and evaluation criteria should be specified before testing begins.

An initial retrospective or appropriately governed silent evaluation may help determine how a model performs using representative local data without allowing its predictions to direct live clinical decisions.

Established monitoring and emergency care should continue unchanged.

Such an evaluation may reveal delayed data entry, inaccurate predictions, excessive alerts or cases where important clinical findings were present despite a low-risk score.

If the results support further investigation, the hospital may consider a limited supervised prospective pilot under an approved governance arrangement.

The pilot should examine whether predictions reach clinicians at useful times, whether they lead to appropriate assessments and whether implementation affects the timing and reliability of care.

It should also identify unintended consequences, including increased workload, unnecessary interventions or confusion about responsibilities.

Criteria for restricting, modifying or suspending the system should be established before clinical deployment.

Staff education needs to address not only how to use the software but also what it cannot reliably determine.

Obstetric clinicians, midwives, nursing leaders, anesthesiology, quality management and information technology personnel should participate in evaluation and implementation.

Any expansion to additional departments or facilities should depend on demonstrated safety, clinical usefulness and operational readiness.

Material changes to the model, input data, intended patient population or clinical workflow may require further evaluation.

The objective is not to create a maternal safety process that depends on proprietary prediction technology. It is to determine whether AI adds meaningful value to an established system of hemorrhage recognition, emergency treatment and continuous quality improvement.

From Earlier Recognition to Safer Maternal Care

Postpartum hemorrhage illustrates why responsible AI implementation must begin with the clinical problem rather than the technology.

AI may help clinicians prepare for women at increased risk, analyze changing clinical information or support quality surveillance. Research in these areas is advancing, but evidence that AI prediction systems consistently improve maternal outcomes in routine care remains limited.

At the same time, strong evidence already supports objective blood loss measurement, earlier recognition, coordinated first-response treatment and systematic preparation for obstetric emergencies.

Viet Nam's updated national guidance issued in April 2026 provides an important foundation for clinical practice. The WHO–FIGO–ICM recommendations and the 2026 implementation guide offer additional evidence and practical approaches for strengthening maternal safety.

The appropriate role of AI is to support these established processes where it can demonstrate additional clinical value.

A useful system should help clinicians recognize important information without weakening bedside assessment, delaying treatment or creating uncertainty about professional responsibility.

It should also help healthcare organizations understand whether improvements in recognition lead to timely clinical action and better maternal outcomes.

The measure of success is not the number of high-risk predictions generated or the statistical accuracy of a model in isolation.

It is whether women experiencing dangerous bleeding are recognized promptly, receive appropriate treatment without avoidable delay and remain within a reliable, accountable maternal safety system throughout the emergency.

That is where AI-assisted early recognition has the potential to make a meaningful contribution to obstetric care.

Selected References and Regulatory Sources

Vietnamese Clinical Guidance

Ministry of Health of Viet Nam. Decision No. 1098/QD-BYT, dated 21 April 2026, issuing the National Guidelines for Reproductive Healthcare Services: Obstetric Hemorrhage (Hướng dẫn Quốc gia các dịch vụ chăm sóc sức khỏe sinh sản về nội dung Chảy máu sản khoa).

International Clinical Guidelines and Implementation Resources

World Health Organization, International Federation of Gynecology and Obstetrics, and International Confederation of Midwives. Consolidated Guidelines for the Prevention, Diagnosis and Treatment of Postpartum Haemorrhage. Geneva: World Health Organization; 2025. ISBN: 978-92-4-011563-7.

World Health Organization, International Federation of Gynecology and Obstetrics, International Confederation of Midwives, Jhpiego, and United Nations Population Fund. Consolidated Guidelines for the Prevention, Diagnosis and Treatment of Postpartum Haemorrhage: Implementation Guide. Geneva: World Health Organization; 2026. ISBN: 978-92-4-011611-5.

Begum F, Nieto-Calvache AJ, Schlembach D, et al. FIGO Recommendations on Objective Measurement of Blood Loss After Birth for Early Detection of Postpartum Hemorrhage. International Journal of Gynecology & Obstetrics. 2025;171(3):933–950. DOI: 10.1002/ijgo.70523.

Clinical Evidence and AI Research

Gallos I, Devall A, Martin J, et al. Randomized Trial of Early Detection and Treatment of Postpartum Hemorrhage. New England Journal of Medicine. 2023;389(1):11–21. DOI: 10.1056/NEJMoa2303966.

Gallos I, Williams CR, Price MJ, et al. Prognostic Accuracy of Clinical Markers of Postpartum Bleeding in Predicting Maternal Mortality or Severe Morbidity: A WHO Individual Participant Data Meta-Analysis. The Lancet. 2025;406(10514):1969–1982. DOI: 10.1016/S0140-6736(25)01639-3.

AlSaad R, Yazbek F, Farrell T, Albasha S, Al-Marri A, Thomas R. Artificial Intelligence for Postpartum Hemorrhage: A Systematic Review. Frontiers in Global Women's Health. 2026;7:1806220. DOI: 10.3389/fgwh.2026.1806220.

Coomarasamy A, Devall AJ, Bell S, et al. Diagnosis and Treatment of Postpartum Haemorrhage: A Race Against Time. The Lancet. 2026;408(10549):83–92. DOI: 10.1016/S0140-6736(26)01031-7.

Kovacheva VP, Kleinlein R, Wheeler N, et al. External Validation of a Machine Learning Model to Predict Postpartum Hemorrhage in a US Northeastern Healthcare System. Pregnancy. 2026;2(1). Published online 12 December 2025. DOI: 10.1002/pmf2.70200.

Akazawa M, Hashimoto K, Ueno M, et al. Risk Stratification for Postpartum Hemorrhage in Vaginal Delivery: A Nationwide Cohort Study of Predelivery and Intrapartum Risk Assessment. Journal of Obstetrics and Gynaecology Research. 2026;52(8). DOI: 10.1111/jog.70417.

Relevant Vietnamese Legal and Regulatory Framework

Law on Medical Examination and Treatment No. 15/2023/QH15.

Law on Artificial Intelligence No. 134/2025/QH15; Decree No. 142/2026/ND-CP; Decision No. 33/2026/QD-TTg.

Law on Personal Data Protection No. 91/2025/QH15; Decree No. 356/2025/ND-CP.

Decree No. 102/2025/ND-CP on health data management.

Decree No. 98/2021/ND-CP on medical device management, as amended; Ministry of Health Circular No. 24/2026/TT-BYT.

Editorial and Clinical Notice

This article is intended for professional education, responsible healthcare AI governance and healthcare quality-improvement discussion. It draws on Vietnamese professional guidance, international clinical recommendations, published research and relevant legal and regulatory sources.

It does not establish a mandatory AI implementation standard, provide patient-specific obstetric advice, authorize automated treatment decisions or constitute a legal determination of any AI product's regulatory status.

The clinical examples are illustrative. Recognition, assessment, treatment and emergency management of postpartum hemorrhage must remain under the responsibility of appropriately qualified healthcare professionals, consistent with applicable Vietnamese requirements, current clinical guidance and locally approved procedures.

The evidence discussed does not establish that AI-assisted postpartum hemorrhage prediction or early recognition necessarily reduces severe maternal morbidity, hysterectomy or maternal mortality in Vietnamese healthcare settings. Claims of clinical effectiveness require appropriate supporting evidence, including evaluation relevant to the intended patient population, clinical function and deployment environment.

Digital Medicine Vietnam (DMVN) | Evidence and regulatory review: 1 November 2026.