03 · What You Need to Know
Why the Absence of a Diagnosis Code Is Not Equivalent to the Absence of Disease
Distinguish Clinical Disease Status from Recorded Diagnosis Status
A patient's underlying medical condition and the information recorded about that condition are related but different.
Clinical disease status concerns whether the patient actually has the condition according to an appropriate clinical or diagnostic reference. Recorded diagnosis status concerns whether the available database contains information satisfying a specified identification rule.
A person may have diabetes without a corresponding diagnosis code in the accessible records. Another patient may have a diagnosis code entered provisionally before the condition is confirmed.
Thus, neither the presence nor the absence of a code necessarily provides perfect classification.
No recorded diagnosis
No qualifying diagnosis was found within the accessible data source and specified observation window.
Condition clinically absent
Available evidence supports the conclusion that the patient does not have the condition, according to an appropriate assessment or reference standard.
The first is an observation about a database. The second is a conclusion about a person's health status.
Researchers should not substitute one for the other without sufficient justification.
Why a Patient with a Medical Condition May Have No Diagnosis Code
Several processes can produce an absent diagnosis code despite the presence of disease.
The condition has not been diagnosed. Some diseases remain asymptomatic, develop gradually, or require tests that have not been performed. A patient may have the condition without being aware of it.
The condition was diagnosed outside the available system. A hospital database may not contain diagnoses established by community clinics, other hospitals, or providers whose records are not connected.
The condition was not relevant to the documented encounter. Routine documentation may emphasize problems addressed during a particular visit rather than every condition the patient has.
The diagnosis exists in unstructured information. Clinical notes, scanned documents, referral letters, or historical summaries may contain information that was not translated into a structured diagnostic code.
The coding algorithm does not recognize the available evidence. A narrow code list, insufficient observation period, or requirement for repeated diagnoses may fail to identify some genuine cases.
Documentation or extraction is incomplete. Relevant information may be absent because of local workflows, delayed entry, migration errors, or limitations in the research extract.
These explanations have different implications. Some reflect undiagnosed disease, others incomplete observation, and others limitations in the operational definition.
Understand the Difference Between a Closed-World and an Open-World Assumption
Database reasoning sometimes distinguishes between two approaches to interpreting information that is not present.
Closed-world assumption
Information not recorded in a sufficiently complete system may be treated as false or absent under the system's defined rules.
Open-world assumption
Information not recorded remains unknown because the available system may not contain all relevant facts.
Routine clinical databases frequently require an open-world interpretation for disease status because they capture only part of a patient's clinical history and healthcare experience.
Nevertheless, some carefully defined administrative outcomes can support a more restricted closed-world interpretation. For example, a database designed to capture every qualifying event within a particular insured population may provide strong evidence that no such recorded event occurred during a period of complete coverage.
Even then, the inference applies to the event the system is designed to capture. No recorded hospitalization does not necessarily mean no illness, and no recorded diagnosis does not necessarily mean no disease.
How the Observation Window Affects Negative Classification
The period during which records are available strongly influences whether a diagnosis can be identified.
Consider two patients without a recorded diagnosis. One has ten years of continuous records containing multiple relevant clinical assessments. The other has only one encounter recorded during the previous month.
The same absence of a diagnosis code provides different amounts of evidence in these situations.
A longer observation period may increase the opportunity to detect previously recorded conditions. However, longer follow-up does not guarantee complete ascertainment, especially when patients receive care outside the system or the condition is rarely assessed.
Researchers should define a lookback period appropriate to the disease, expected healthcare utilization, and available data coverage.
They should also distinguish the first recorded diagnosis from true disease onset. A condition first documented during the study may have existed for years beforehand.
How Healthcare Utilization Influences Diagnosis Detection
Patients who interact frequently with healthcare services have more opportunities for conditions to be assessed, diagnosed, and recorded.
Those with fewer encounters may have fewer recorded diagnoses even when their underlying disease burden is similar.
This creates a potential ascertainment problem. The probability of detecting a condition can depend on healthcare utilization rather than disease status alone.
For example, patients receiving regular monitoring for a chronic illness may be more likely to have additional conditions documented than patients who rarely seek care.
If healthcare utilization also differs between exposure groups, the resulting diagnostic ascertainment may be differential.
Researchers should examine encounter frequency, observation duration, care settings, and other relevant indicators of recording opportunity. Adjusting for utilization may sometimes help characterize differences, but it requires care because healthcare utilization can itself be affected by exposure or disease and may act as a mediator or collider.
What False Negatives Mean for Disease Classification
A false negative occurs when a patient who truly has the condition is classified as not having it by the record-based algorithm.
False negatives can arise because the diagnosis was never made, the relevant record is unavailable, or the algorithm fails to recognize the recorded evidence.
Two validation measures are particularly useful for understanding negative classification:
-
Sensitivity: the proportion of people who truly have the condition and are correctly identified as positive.
-
Negative predictive value: the proportion of people classified as negative who truly do not have the condition.
High sensitivity reduces the proportion of true cases missed by the algorithm. High negative predictive value indicates that negative classifications are usually correct in the population being evaluated.
However, negative predictive value depends partly on disease prevalence. A rare condition can produce a high negative predictive value even when the identification algorithm misses a substantial proportion of actual cases.
Therefore, researchers should not rely on negative predictive value alone when evaluating an algorithm's ability to identify disease-free patients.
Why Disease Prevalence Changes the Meaning of a Negative Result
Consider an algorithm applied to a population where the condition is uncommon. Most patients genuinely do not have the condition, so many negative classifications may be correct even if the algorithm is imperfect.
In a high-prevalence population, the same sensitivity and specificity can produce a lower negative predictive value because a larger proportion of patients have the condition and may be missed.
This means that validation results from a general population may not transfer directly to a specialized clinic or high-risk cohort.
When using published validation evidence, assess the reference standard, population, disease prevalence, observation period, and recording environment.
For an algorithm with sensitivity Se, specificity Sp, and disease prevalence P, the negative predictive value is:
This calculation assumes that the stated sensitivity, specificity, and prevalence appropriately characterize the population and classification process. It should not be used as a substitute for empirical validation.
Should Patients Without Diagnosis Codes Be Classified as Controls?
Researchers often construct comparison groups by identifying patients who do not meet the case definition.
Such individuals may appropriately be described as algorithm-negative or without a recorded qualifying diagnosis. Whether they can be described as disease-free controls depends on the validity of that classification.
A control group drawn from patients with no relevant codes may contain undiagnosed cases, patients with diagnoses recorded elsewhere, and patients whose conditions were missed by the algorithm.
This contamination can affect estimated associations. The direction and magnitude of bias depend on the misclassification process, exposure distribution, and study design.
Possible approaches include requiring adequate prior observation, using validated negative definitions, incorporating supporting clinical evidence, or distinguishing confirmed negatives from patients with uncertain disease status.
Researchers should avoid excluding patients merely because their classification is uncertain without considering whether the resulting selection process could introduce additional bias.
Can Laboratory Results or Other Clinical Evidence Strengthen Negative Classification?
Depending on the condition, researchers may use laboratory results, imaging findings, procedures, clinical assessments, or other evidence to supplement diagnosis codes.
For example, a study of diabetes might consider relevant laboratory measurements alongside diagnostic codes and medication records.
However, a normal laboratory result does not always establish the absence of disease. Interpretation depends on the test's diagnostic properties, timing, clinical context, and whether testing was performed under appropriate conditions.
Similarly, the absence of a disease-specific medication is not proof of disease absence because some patients are untreated or receive alternative management.
Combining evidence sources may improve classification, but the resulting algorithm still requires evaluation.
When developing operational definitions for exposures and outcomes, researchers should specify what qualifies as positive, negative, and indeterminate evidence rather than assume that every record can be classified reliably into two categories.
When Is an Explicit Unknown Category Appropriate?
Sometimes the available records do not support a defensible positive or negative classification.
For example, a patient with limited observation history and no relevant diagnostic assessment may have unknown disease status rather than a confirmed absence of disease.
An explicit unknown or indeterminate category can preserve this distinction during data preparation and descriptive reporting.
However, adding an unknown category does not automatically solve missing-data or selection problems. Its use in inferential analysis requires a method consistent with the research question and the likely mechanism producing uncertainty.
Researchers should distinguish uncertainty about underlying disease status from a missing value for a variable that was intended to be measured.
These distinctions are particularly relevant when incomplete documentation affects routine records.
Validate Negative Classifications Where Feasible
Validation should assess whether patients classified as negative genuinely lack the condition according to an appropriate reference standard.
Potential approaches include reviewing clinical records, examining independent registries, using relevant laboratory evidence, or evaluating a representative sample against an adjudicated reference.
Validation restricted to algorithm-positive cases cannot directly establish the algorithm's sensitivity or negative predictive value. Evaluating false negatives requires information about at least some patients classified as negative.
When validation is feasible only for a subset, sampling procedures and the representativeness of that subset should be documented.
Where no suitable reference standard exists, researchers may compare alternative algorithms and conduct quantitative bias analyses under plausible assumptions. Such analyses characterize uncertainty but do not establish true disease status.
Report Exactly What the Data Establish
Terminology matters when presenting results.
If the study identifies diagnoses through claims or electronic health record codes, phrases such as "patients without a recorded diagnosis" or "algorithm-negative patients" may be more accurate than "healthy patients" or "patients without the disease."
The stronger wording may be justified when an adequately validated definition supports that interpretation.
Explain the observation period, qualifying code sets, available clinical evidence, validation performance, and limitations of record coverage.
For studies using routinely collected health data, the RECORD statement provides relevant reporting guidance concerning code-based definitions and the characteristics of the data source.
