How Artificial Intelligence Can Transform Mortality Surveillance in the United States

Once data are processed, public health agencies produce reports on mortality rates, leading causes of death, and trends by geography, age, and other characteristics. These outputs guide everything from highway safety regulations to cancer screening guidelines. They are also critical during emergencies — for example, detecting excess deaths during a pandemic or heatwave — and for long-term trend monitoring, such as the rise in drug overdose deaths.

The system is called “vital statistics,” because it is so important to public policy. But these systems also have important weaknesses. Death reporting and recording is often delayed by weeks or months. Cause-of-death certificates can be vague or inaccurate. Errors occur in coding, i.e., turning the information from a patient’s record into digital death entered into a health database. And the capacity to analyze and disseminate mortality data in a timely, actionable format varies widely between states. These challenges mean public health agencies sometimes detect important changes in mortality only after they have already caused substantial harm.

Artificial intelligence could help strengthen each stage of mortality surveillance — from registration to cause-of-death determination to reporting and dissemination — if deployed carefully, with proper safeguards for privacy, accuracy, and equity.

Challenges in recording who died in a public health database

The registration process begins with the death certificate, a legal document that must be completed before burial or cremation. In the U.S., electronic death registration systems (EDRS) are now more common, but their adoption, design, and functionality vary by state. Physicians, funeral directors, and medical examiners log into these systems to enter demographic details and certify the cause of death.

Delays occur when physicians are unavailable or unfamiliar with the system, when legal investigations require more time, or when technical barriers slow completion. In some rural areas, internet access is limited, and in certain jurisdictions, parts of the process are still paper-based. Inaccuracies can creep in from simple typing errors, incorrect demographic entries, or confusion over which data fields are required.

AI’s potential to improve death recording and registration

Natural language processing (NLP) and AI-assisted data entry could make registration faster and more accurate. For example, AI could auto-complete demographic fields from linked hospital or clinic records, flag inconsistencies (such as mismatched dates or implausible age values), and prompt users to correct them before submission.

For funeral directors or physicians working in settings with low connectivity, AI-enabled mobile apps could allow offline completion of certificates with automatic synchronization once a connection is available. Voice-to-text systems, trained on medical terminology and death certification vocabulary, could let certifiers dictate information directly into the EDRS, reducing typing errors and improving accessibility for less tech-savvy users.

AI could also act as a quality control layer. Algorithms could compare new entries with historical records to detect duplicates or anomalies, such as multiple certificates for the same individual or improbable demographic patterns. These alerts could be sent to vital records staff for review, reducing the workload and shortening the time between death and official registration.

Challenges in recording causes of death in a population

Assigning a cause of death is one of the most challenging aspects of mortality surveillance. Physicians are trained to complete a cause-of-death statement in a specific sequence: the immediate cause (the final disease or injury leading to death), the underlying cause (the disease or injury that initiated the chain of events), and any contributing conditions. This information is then translated into ICD codes.

The quality of these statements varies widely. Some physicians provide precise, clinically supported details (“acute myocardial infarction due to coronary artery disease”), while others write vague terms (“cardiac arrest”) that describe the mechanism of death rather than the cause. In cases of injury, poisoning, or suspected foul play, medical examiners and coroners may require toxicology tests, autopsy results, or investigative findings before certifying the cause.

Even when statements are complete, the translation into ICD codes is complex. Automated coding systems exist, but they often require human review for ambiguous terms, conflicting information, or unusual combinations of conditions. Human coders are in short supply, and backlogs can delay the availability of final mortality statistics.

AI’s potential to improve tracking causes of death

AI can enhance this process in several ways:

1. Real-time guidance during certification
   An AI tool embedded in the EDRS could analyze the physician’s draft statement in real time, comparing it against structured clinical data from the patient’s health records. It could prompt for missing details, flag nonspecific terms, and suggest more precise language based on the medical history and clinical findings. For example, if a physician writes “respiratory failure,” the system could prompt: “Was this due to pneumonia, chronic obstructive pulmonary disease, COVID-19, or another cause?”
2. Automated cause-of-death coding
   Advanced NLP models, trained on millions of historical death certificates and ICD-coded records, could assign codes with higher accuracy and speed than current systems. These models could also quantify uncertainty, flagging cases for human review when confidence is low or when the case includes rare or conflicting details.
3. Integration with forensic and laboratory data
   AI systems could link death certificates to toxicology results, autopsy reports, or police incident records, enabling more accurate cause-of-death determination in injury and overdose cases. For example, an algorithm could match postmortem lab results showing fentanyl in the bloodstream with the narrative description of an overdose, ensuring consistent coding across jurisdictions.
4. Handling disagreements between AI and human coders
   Discrepancies will arise when AI-generated coding differs from human assessment. Policy frameworks will need to specify how these disagreements are resolved. In some cases, AI could serve as a “second reader,” prompting coders to reconsider ambiguous cases. In others, human judgment may override AI suggestions, with systems designed to learn from these overrides to improve future performance.

Enhancing reporting, analysis, and dissemination of death records

Once death records are complete and coded, state health departments and the NCHS produce statistical summaries and analytic reports. Standard products include monthly and annual mortality tables, dashboards of leading causes of death, and special reports on emerging threats. However, there is often a substantial lag — sometimes up to a year — between the date of death and the publication of final data.

During emergencies, provisional death counts can be released more quickly, but these are often incomplete and lack the detail needed for precise public health action. Data visualizations may not be tailored to the needs of policymakers, journalists, or community leaders, and some jurisdictions have limited capacity to create interactive, user-friendly tools.

AI’s potential to improve reporting, analysis, and dissemination of mortality surveillance data

AI could make mortality data more timely, useful, and accessible:

1. Near-real-time analysis
   By continuously ingesting provisional death records, AI models could estimate mortality trends with statistical adjustments for missing data. These “nowcasting” techniques, already used in infectious disease surveillance, could be adapted for all-cause and cause-specific mortality, providing policymakers with early warnings of spikes in deaths from emerging threats.
2. Automated analytics and visualization
   AI-powered platforms could generate customizable dashboards that allow users to filter mortality data by age, race, geography, or cause, with automatic updates as new data arrive. These systems could also detect unusual patterns — for example, a sudden increase in deaths from a specific substance in one county — and push alerts to relevant officials.
3. Tailored communication
   Generative AI could create reports suited to different audiences from the same dataset. Policymakers might receive a concise briefing with key statistics and recommendations, while community groups could get a plain-language summary with culturally relevant examples. Journalists could be provided with ready-to-use charts and explanatory text that reduce the risk of misinterpretation.
4. Linking mortality to broader health surveillance
   AI could integrate mortality data with hospital discharge records, emergency medical services reports, and public health surveillance systems to provide a more comprehensive picture of health threats. This would allow faster identification of upstream causes — for example, linking increased deaths from heatstroke to climate data on extreme heat events.

Policy and ethical considerations specific to mortality surveillance

• Accuracy vs. speed: The pressure to produce real-time mortality estimates must be balanced against the risk of errors that could mislead policymakers or the public.
• Bias and inequity: If AI models are trained on historical death records that undercount certain populations — for example, due to disparities in access to medical investigation or misclassification of race — the models may perpetuate those inequities.
• Transparency in disagreements: When AI coding diverges from human coding, the resolution process should be transparent, with documented reasons for the final decision.
• Privacy: Even though mortality data are public in many jurisdictions, linking them with clinical, forensic, or geospatial data raises new risks of re-identification, requiring strict controls on access and use.
• Public trust: Mortality statistics are often central to debates about health policy and government performance. AI’s role must be communicated clearly to maintain trust, particularly if AI systems are used to estimate mortality during contentious events like pandemics or environmental disasters.

Improving Morality Surveillance With AI

Timely, accurate mortality data are the foundation for understanding a nation’s health, guiding prevention strategies, and holding systems accountable. AI can help strengthen this foundation: streamlining death registration, making cause-of-death identification more precise, and accelerating the analysis and communication of results. Doing so will require sustained investment in both technology and the human expertise that must guide it. As with all public health innovations, the goal is to augment, rather than replace, human operations, allowing scarce human resources to focus on tasks that require judgment, empathy, and local knowledge. The future of mortality surveillance should be one in which technology accelerates the flow of reliable information, and where that information drives faster, more effective action to save lives.