Emergency-department data are a window, not a census

#Question

What can emergency-department data support in public-health surveillance, and what should they not be interpreted as measuring?

#Answer in brief

Emergency-department data are unusually valuable because they can arrive quickly and contain symptoms, diagnoses, demographics and encounter information useful for syndromic surveillance. CDC's National Syndromic Surveillance Program uses ED and other health-care data for near-real-time detection and monitoring of health threats.^[1]

But an ED dataset is a dataset of encounters among people who reached emergency care. It is not automatically a census of disease in the community. The route from illness to ED attendance is shaped by severity, access, health-seeking behaviour, primary-care availability, local service organization, referral practices and changing thresholds for admission or discharge. Treating utilization as incidence silently changes the epidemiological question.

#The key distinction: event, encounter, and population

Suppose a dashboard shows a 30% increase in ED visits for an influenza-like syndrome.

Three different statements might be proposed:

1. ED encounters meeting the syndrome definition increased by 30%.
2. community incidence of the syndrome increased by 30%.
3. severe disease in the population increased by 30%.

Only the first follows directly from the data.

The second requires assumptions about who seeks ED care and whether that probability remained stable. The third additionally requires a valid relationship between the surveillance definition and severity.

This is the central interpretive discipline for ED data: the observed unit is an encounter, while the public-health question often concerns people, disease events, or population risk.

#Why ED data are powerful

CDC describes syndromic surveillance as a timely system for detecting, understanding and monitoring health threats, with ED information often available rapidly after an encounter.^[1] This speed makes ED data useful for questions that can tolerate imperfect specificity in exchange for early situational awareness.

Examples include:

• detecting an unusual rise in respiratory or gastrointestinal presentations;
• monitoring heat-related or injury-related health-care use;
• identifying abrupt changes in overdose-related encounters;
• comparing current activity with an expected seasonal baseline;
• understanding pressure on emergency services;
• providing one signal among several during an unfolding event.

The key phrase is one signal among several. The value of syndromic surveillance often comes from temporal sensitivity, not from pretending that one care setting captures every event in the population.

#The selection mechanism is part of the data

A person with the same underlying condition may:

• remain at home;
• call a general practitioner;
• use an out-of-hours service;
• attend an urgent-care centre;
• attend an ED;
• be referred directly to hospital;
• be admitted through another pathway.

The probability of each path can change with health-care access and service organization. A systematic review of ED syndromic surveillance found substantial variation in ED surveillance systems and health-care provision across settings, reinforcing that these systems monitor people who reach emergency care rather than an unselected population.^[2] Holidays, service reorganization, public messaging or changing care-seeking can therefore alter the observed encounter stream.

When that mechanism changes, an ED trend can change even if underlying incidence does not.

This does not make the data "biased" in the colloquial sense of being bad. It means the data are generated by a selection process that must be part of the interpretation.

RECORD was developed specifically because routinely collected health data require transparent reporting of selection, data sources, algorithms and limitations.^[3] A careful ED analysis should make the encounter-generating process visible rather than treating the database as a neutral window onto the whole population.

#Coding is not a fixed truth layer

Syndromic systems can use chief complaints, diagnostic terms and coded discharge information. CDC's current respiratory ED guidance documents explicit code and clinical-term definitions used to extract diagnosed respiratory illnesses from syndromic data.^[4]

That matters because a "case" in an ED surveillance query can depend on:

• which field is used;
• when that field becomes available;
• local coding practice;
• changes in EHR templates;
• terminology mapping;
• whether a diagnosis is preliminary or final;
• missingness and free text;
• revisions to the syndrome definition.

A sudden rise can therefore be epidemiological, operational, semantic, or some combination.

#Timeliness versus specificity

ED data illustrate a recurring surveillance trade-off.

Chief complaints may be available early but can be vague. Final diagnoses can be more specific but arrive later or be incomplete. A syndromic definition may deliberately prioritize sensitivity, accepting false positives to detect changes quickly.

For early warning this can be rational. For estimating disease-specific burden it may not be.

The error is not using a broad syndrome. The error is using a broad syndrome and interpreting it as if it were a validated disease count.

#A synthetic example: respiratory activity

Imagine a region with five EDs. During week 3, visits matching a respiratory syndrome rise from 400 to 540.

At first glance, the increase appears to indicate sharply rising community incidence.

Further checks show:

• a large primary-care service closed temporarily;
• one hospital introduced a new triage phrase that maps into the syndrome definition;
• two EDs increased completeness of discharge-diagnosis reporting;
• laboratory positivity also rose, but more modestly;
• age distribution shifted toward older adults.

The ED signal is still useful. It correctly shows increased respiratory-related emergency activity. But the magnitude cannot be assigned wholly to disease incidence.

A defensible interpretation might be:

Respiratory-related ED activity increased substantially, alongside a smaller rise in laboratory positivity. Changes in primary-care availability and ED coding/reporting may have amplified the observed encounter increase. The signal supports heightened situational awareness but does not by itself quantify community incidence.

That sentence is less dramatic and more useful.

#What to check before interpreting a trend

#1. Coverage

Which facilities contribute data? Did participation change? Does the dataset include the relevant geography and care settings?

#2. Completeness

Are chief complaint, diagnosis, age, residence and disposition fields stable over time? A field becoming more complete can look like a health event.

#3. Definition

Is the syndrome based on symptoms, diagnosis codes, text, laboratory results, or combinations? Has the definition changed?

#4. Care-seeking

Could access, public messaging or service reorganization have changed the probability of attending an ED?

#5. Case mix

Did age, severity, referral source or disposition change? Stable counts with a changing case mix can still be epidemiologically important.

#6. Comparison data

Do laboratory, primary-care, ambulance, mortality, pharmacy or other sources show compatible patterns? Multisource agreement does not prove causation, but disagreement is often diagnostically useful.

#7. Denominator

Are you reporting counts, proportions of ED visits, rates per population, or standardized measures? Each answers a different question.

#When rates per population still need caution

Dividing ED encounters by resident population creates a population-based rate, but it does not remove selection.

The numerator still depends on ED use. Visitors can cross geographic boundaries. Residents can attend facilities outside the surveillance network. The probability of ED attendance may differ by deprivation, age, transport, insurance or service availability.

A population denominator can make a quantity comparable while leaving the causal pathway into the numerator unchanged.

#Research use requires explicit provenance

When ED data are used for observational research rather than routine situational awareness, RECORD recommends transparent reporting of routinely collected data sources and the algorithms used to identify study populations and variables.^[3]

That transparency is especially important when an apparently clinical variable is actually an administrative or coding construct. A reader should be able to tell whether "influenza" means a laboratory result, a discharge code, a syndrome definition, or a combination.

#Practical interpretation language

Prefer statements such as:

• "ED visits meeting the syndrome definition increased."
• "The proportion of ED encounters with this diagnosis rose."
• "The signal is consistent with increased activity but may also reflect changing care-seeking."
• "These data monitor emergency-care utilization and syndromic activity; they do not directly estimate all community cases."

Avoid silently upgrading the observed object:

• ED visits → infections;
• diagnosis codes → confirmed disease;
• participating facilities → the whole region;
• change in encounters → causal effect of an intervention.

#Limitations and uncertainty

The examples here are synthetic and do not describe any specific health system. The selection mechanisms affecting ED data differ across jurisdictions and over time. CDC's NSSP documentation describes a large US surveillance infrastructure and is used here to illustrate how contemporary syndromic surveillance employs ED data; its coverage and operational details should not be generalized to other countries.^[1] RECORD is a reporting guideline rather than a design-quality instrument and does not by itself establish validity.^[3] Disease-specific analyses require validated local definitions, knowledge of data-generation processes and, where appropriate, linkage with other surveillance sources.