Harmonizing External PK Datasets

Harmonization as a Data-Meaning Check

A locked population PK model can only be applied meaningfully when its inputs represent the same things in the development and external datasets. Matching column names is not enough. A column called “time” might mean clock time, time since admission, or time since the most recent dose; “concentration” might refer to different analytes or units. Harmonization makes those meanings explicit and maps external records to the definitions the locked model expects.

The task is not to make two datasets look identical or to improve the model’s fit to the external data. It is to use prespecified, scientifically justified transformations that preserve the meaning of each record. If a difference cannot be resolved without guessing, the appropriate outcome is to flag the record or the dataset, not to invent a conversion. The model’s input definitions serve as the reference contract: they specify which events, units, analytes, covariate definitions, and time conventions the model uses.

Before transforming data, write down the model-input contract. For each field, specify its definition, units, allowed values, time reference, and any encoding rules. Keep the external source value as well as the harmonized value, and record the transformation and its rationale. This lets a reviewer trace a model input back to its original record.

Time and Dosing Records

PK models use times in relation to events, not merely dates or clock readings. The relevant origin might be the first dose, a particular dose, or another specified study event. External records must be mapped to the same origin and time unit. If the model uses elapsed time in hours since the first dose, then both sample times and dosing-event times must be expressed on that scale. A sample without a reliable event history may not have a defensible model time.

For timestamps in the same time zone and on an unambiguous clock, elapsed time can be calculated as the difference between timestamps. If \(T\) denotes a timestamp and \(T_0\) the agreed origin, then elapsed time in hours is the timestamp difference converted to hours. The conversion must account for the date, not just the clock face: a sample at 08:00 on the following day is 24 hours after an 08:00 origin, not at time zero. Daylight-saving transitions, time-zone differences, and local versus standard time can make a simple clock subtraction wrong; resolve these using documented timestamp conventions before calculating elapsed times.

This expression assumes the timestamp difference is represented in seconds on a consistent, unambiguous time scale. Each dose event also needs a consistent time and amount. Check for duplicate or impossible event times, missing dose amounts, implausible intervals, and sample times that precede the stated origin. A valid-looking numerical time does not establish that the event history is correct.

Dose records need the same care. Confirm that amount, route, formulation, and event time are interpretable under the model’s conventions. A dose recorded as administered must not be silently treated as prescribed if administration is uncertain. Nor should a dose be assigned to a sample solely because it is the nearest recorded event. Where the event history is incomplete or internally inconsistent, retain the discrepancy for review.

Units, Analytes, and Concentration Records

A numerical concentration is not interpretable without its unit and analyte definition. Unit conversion is appropriate when two records measure the same quantity on compatible scales. For example, a mass concentration in mg/L can be expressed in ng/mL using a fixed conversion: \(1\ \mathrm{mg/L}=1000\ \mathrm{ng/mL}\). The conversion changes the numerical value, not the underlying concentration.

Analyte identity is a scientific definition, not a label to normalize casually. Parent drug, metabolite, total drug, and unbound drug are not interchangeable measurements. Even when two reported concentrations use the same units, they may differ in what chemical species or fraction was measured. Molecular-weight conversions may sometimes express the same substance on a different mass basis, but they do not turn a metabolite concentration into a parent-drug concentration, or total concentration into unbound concentration. Document the analyte and matrix definitions and flag unresolved differences.

Check whether values are raw, dilution-corrected, or otherwise transformed according to their data dictionary. Applying a correction twice can create a systematic error while leaving the results numerically plausible. Also distinguish a true zero from a missing result, a below-quantification result, or a value recorded using a special code. Do not replace these states with one another as a convenience. Their treatment depends on the model and analysis plan; a conversion step should preserve the original status and make any recoding explicit.

Covariates and Categorical Encodings

Covariates require harmonization of both units and meaning. A model using weight in kilograms cannot be supplied a weight in pounds simply because both columns are numeric. A model using a particular clinical definition of renal function cannot necessarily accept a differently calculated value with a similar label. Confirm the measurement definition, units, timing, and any categorization or coding rule used by the locked model.

Continuous covariates may also require checking the date on which they were measured. A weight measured long before the relevant dosing period may not represent the model-defined covariate. For categorical covariates, create an explicit mapping table and check every source category, including uncommon values. A category absent from the mapping is a review issue, not a reason to assign the most convenient code.

A Practical Audit Before Model Application

A harmonization workflow should make its decisions reproducible and expose records that cannot safely be mapped. The following sequence separates definition, transformation, and quality checks.

A useful audit table reports, for each variable, the source definition, target definition, transformation, assumptions, and unresolved issues. Automated checks can identify impossible values or unrecognized codes, but they cannot establish scientific equivalence by themselves. For example, a concentration range check might find a unit error, but a plausible value does not prove that the sample measured the right analyte.

Harmonization is distinct from several tempting alternatives. A prespecified conversion is preferable to ad hoc adjustment because it is reproducible and does not use external outcomes to tune the model. Dropping every record with a discrepancy may avoid some input errors but can discard usable data and change which population remains represented; exclusions need transparent rules. Conversely, forcing every field into a target format can conceal a genuine incompatibility. Harmonization does not resolve differences in study design, measurement practice, or population. It prepares inputs; it does not establish that the model is suitable for the external setting or determine how its predictions perform.

Common Pitfalls and Limits

The most common mistake is to treat a field as harmonized once its units or label match. A time column in hours is still wrong if its origin differs. A concentration in ng/mL is still wrong if it measures a metabolite while the model expects parent drug. A binary covariate coded 0 and 1 is still wrong if the reference category is reversed. Each mapping must be checked against the underlying definition.

• Converting units without checking the quantity. Verify both the units and the analyte, matrix, and concentration basis before applying a numerical conversion.
• Using clock time as elapsed time. Calculate elapsed time from an explicit origin, using dates and consistent time-zone information.
• Collapsing missing and unknown values. Preserve distinct data states rather than assigning an unsupported category or numerical value.
• Silently repairing event histories. Record the source, rule, and reason for any correction; flag ambiguities that cannot be resolved from available evidence.

Some discrepancies cannot be repaired through data transformation. The external study may not have recorded a required covariate, may use an incompatible analyte definition, or may lack sufficiently precise dose times. In such cases, document the limitation and determine whether the locked model can accept the available information under its established input rules. Any decision to change the model or its assumptions is not merely harmonization and should be treated separately.