Cross-Field Consistency Checks in Pharmacometric Workflows

Why Check Fields Together?

A record can pass every individual-field check and still be incoherent. A dose amount may be positive, a rate may be positive, and a duration may be positive, while the three values contradict one another. A concentration may be a valid number and a censoring flag may be a valid category, yet the flag may not agree with the reported value and assay limit. These are cross-field consistency problems: the concern is not whether each field is acceptable alone, but whether their combination satisfies the meaning assigned to them.

Such checks are important before pharmacometric data are used to construct dosing histories, observations, and covariates. They do not establish which conflicting value is correct. They identify records that need review under the data specification and source documentation. A check should therefore state its condition, scope, expected result, and action when the condition fails.

A useful rule is conditional. For example, “when the record is an infusion dose and both amount and duration are recorded, the rate must agree with amount divided by duration within the declared tolerance” is more precise than “check dose fields.” The condition prevents the rule from being applied to bolus doses or records for which duration is not defined. It also makes the expected behavior reviewable and reproducible.

Specify Rules Before Running Them

Before implementing a check, define the dataset convention it relies on. A field name alone is not enough: the meaning of an event category, the treatment of missing values, the unit of a dose, and the encoding of a below-quantification result must be explicit. Earlier material on dataset schemas and required-field validation establishes how fields and allowed values are specified; cross-field rules add relationships among those specified fields.

A rule specification should include the applicable record type, the fields used, the logical or numerical condition, any tolerance, and the outcome for a failure. It should also make clear whether a failure blocks dataset use or generates a review warning. An apparent inconsistency can result from an unrecorded convention or legitimate exception, so do not silently overwrite a source value to make a record pass.

Dose Amount, Rate, and Duration

For an infusion recorded with a total amount \(A\), a rate \(R\), and a duration \(T\), all in compatible units, the defining relationship is \(A=RT\), or equivalently \(R=A/T\) when \(T>0\). If amount is in milligrams and duration is in hours, rate must be in milligrams per hour. This relationship is appropriate only if the fields refer to the same dose and the rate is constant over the stated duration. A bolus dose does not have an infusion duration under this convention; a missing rate for a bolus should not fail an infusion-specific rule.

The tolerance is not a universal pharmacometric constant. It should reflect the source precision, transformation and rounding conventions, and intended use. If a recorded rate is rounded, equality to many decimal places can produce spurious failures. Conversely, a tolerance that is too wide can conceal meaningful contradictions. Check units before evaluating the arithmetic: a thousand-fold difference can be a unit mismatch rather than a dose calculation error.

Event Type and Observation Status

An event type describes what a row represents under the dataset’s convention; an observation flag indicates whether the row contributes a measured response. Their valid combinations are schema-dependent. A dose row might be designated as a dosing event and not as a concentration observation, while a concentration-sample row might be designated as an observation. Other row types, such as covariate updates or administrative records, may have their own flag behavior.

Do not assume that a particular numeric code universally means “dose” or “observation.” Instead, use the documented category mapping and write the permitted combinations as a table or logical rule. Also distinguish a row with no observed concentration from a concentration value that is missing for a different reason. The relevant rule concerns the meaning of the row and its flags, not merely whether a numeric field happens to be populated.

Censoring Status and Assay Limits

An assay limit gives meaning to certain reported results. For example, a result below the lower limit of quantification (LLOQ) may be flagged as below quantification. The dataset must specify how such results are stored: a numeric value may be missing, set to the LLOQ, or represented by another convention, with a separate censoring indicator. These encodings are not interchangeable, and a cross-field check must follow the declared one.

A useful rule examines the combination of censoring status, reported result, and the applicable limit for that assay and sample. It can flag a record marked as quantified when its value is below the applicable LLOQ, or a record marked below the LLOQ whose stored value contradicts the declared encoding. Such findings are conditional on the limit being matched to the correct analyte, assay, and unit. The check does not turn a censored result into an exact concentration.

Covariate Values and Units

A covariate value has meaning only together with its unit and, where relevant, its measurement basis. A value of 70 could represent kilograms, pounds, or another quantity; comparing values without checking units can create false discrepancies or misleading covariates. Cross-field checks can verify that a unit label is allowed for the variable, that a conversion uses the declared factor, and that two representations intended to describe the same measurement agree within an appropriate tolerance.

For a conversion check, keep the raw value and unit available, compute the converted value using the declared factor, and compare it with the corresponding standardized value. The tolerance should account for recorded precision and any documented rounding. A passing comparison supports internal consistency, not the truth of either measurement. A failed comparison is a review signal, not permission to replace one field with another.

Diagnostics, Alternatives, and Limitations

A good validation report makes it possible to understand why a row failed. Include the rule identifier, record key, relevant field values and units, the expected relationship, and the result of the comparison. Review patterns as well as individual findings: many rate failures may indicate a unit conversion or schema misunderstanding, while failures concentrated in one assay may point to an incorrect limit mapping. Such patterns guide investigation but do not establish a cause.

Cross-field checks complement, rather than replace, other data-quality approaches. Range and distribution checks assess individual values and patterns across records; they may miss a pair of individually plausible but contradictory values. A cross-field rule tests a declared relationship, but it cannot establish source accuracy. Manual review can use contextual judgment but is harder to apply consistently at scale. These approaches answer different questions and are strongest when their roles are explicit.

Several failure modes deserve particular attention:

• Applying a rule to the wrong row type. An infusion relationship does not apply to a bolus unless the specification defines one.
• Comparing incompatible units. Arithmetic can be internally correct and still be meaningless if amount, rate, duration, or covariate units do not match.
• Using an undocumented encoding assumption. Censoring values and event codes vary by dataset; a rule must follow the current specification.
• Using excessive precision or tolerance. Unrounded floating-point comparisons can create false alarms, while a loose tolerance can mask material conflicts.
• Automatically “fixing” failures. A logical conflict identifies a need for review; it does not identify the correct field to change.

Cross-field checks are limited by the quality of their definitions and inputs. They cannot detect every plausible but incorrect record, resolve conflicting sources, or infer the intended value from internal consistency alone. Their purpose is narrower and valuable: make important relationships explicit, apply them consistently, and expose records that do not satisfy them.