What It Means for a Dosing Service to Travel
A dosing service that performs well in one hospital does not automatically work in another, even when both settings treat the same condition with the same medicine. The target site may record a different dose event, collect concentrations at different times, use different prescribing practices, or care for patients whose characteristics fall outside the model’s experience. Any of these differences can make a model input mean something different—or make a required input unavailable.
Transportability is therefore a question about a specific source setting, target setting, patient population, and clinical workflow. It is not a general claim that a model “works everywhere.” External validation and prediction-error interpretation provide relevant evidence about model performance; assessing transportability also asks whether the target setting supplies data and follows practices that make the model’s assumptions applicable.
The assessment should compare what the model expects with what the target service can actually provide. A variable with the same label at two sites may have different definitions. A concentration recorded in the same units may have been collected at a different point in the dosing interval. A prescribed dose may not be the dose administered. The comparison must therefore cover definitions, timing, prescribing practice, and model assumptions, not just column names or units.
A Structured Comparison
Begin by specifying the intended use: which patients, medicine, decisions, and point in care are in scope? Then document the source definition and target equivalent for each input and workflow step that can affect a recommendation. Distinguish a true mismatch from a difference that can be converted reliably. For example, converting pounds to kilograms is usually a deterministic unit conversion; substituting a historical weight for a current one is a change in the measurement’s meaning.
| Comparison area | Questions to ask | Possible transportability barrier |
|---|---|---|
| Data definitions | Does each field represent the same clinical quantity, event, and time? | A dose-order field is used where the model expects an administered dose. |
| Measurement schedules | Are measurements collected at comparable points relative to dosing and clinical events? | A concentration is treated as a trough although its actual collection time is unknown. |
| Prescribing practices | Are dose, interval, infusion, and interruption practices compatible with the model’s inputs and assumptions? | The target setting frequently changes the interval, but the model receives only a nominal regimen. |
| Model assumptions | Are patient characteristics, data quality, and dosing history within the conditions for which the model is intended? | The target population or observed covariate range is not represented in the model’s evidence. |
| Workflow and action | Can required information arrive in time, and can the recommendation be applied as specified? | A recommendation is generated before a delayed administration record is available. |
For each difference, record its likely consequence and evidence: a definition comparison, data-quality summary, schedule distribution, or documented clinical workflow. A useful classification is harmonizable (a stable conversion or mapping preserves meaning), manageable with a specified change (such as a workflow or input-collection change), or unresolved (the difference prevents reliable interpretation or use). These labels organize the assessment; they do not replace clinical and technical review.
Worked Example: A Dose Field That Does Not Mean the Same Thing
Suppose a hypothetical source-site model requires the actual administered dose and its administration time. At a target site, the available field is the prescribed dose; administration records are stored separately. In a review of 100 target-site dosing episodes, 18 have no usable administration record, and another 12 have an order but no administration time available to the service when a recommendation is requested. Assume these groups are distinct.
Define the mismatch. The target’s order field describes an intended dose, not necessarily a dose received. The model’s required input describes a dose administered. Treating the two as interchangeable would be an unsupported change in data definition.
Quantify the operational gap. Episodes lacking a usable administration record or a timely administration time total \(18+12=30\). Thus \(30/100=0.30\), or 30.0% of the reviewed episodes, do not have the required dose history available at the decision point. The remaining \(100-30=70\), or 70.0%, meet this particular availability check; that does not establish that all other model inputs or assumptions are met.
Assess the implication. A deterministic mapping from “prescribed” to “administered” cannot resolve whether a dose was delayed, omitted, or changed. A reasonable response is to test whether the administration feed can be made timely and complete, define how corrections or omissions are represented, and limit use to episodes with verified administration data until the gap is addressed. The 70.0% figure is a readiness measure for this one requirement, not evidence of model accuracy in the target setting.
Measurement Timing and Model Meaning
Timing differences can matter even when units, assay, and field names match. A concentration is interpreted relative to the dosing history and the time at which the sample was taken. If the model expects a sample at a defined point in the interval but the target site collects it later, the measured value may differ for pharmacokinetic reasons rather than because the underlying patient or model has changed.
A simple one-compartment elimination relationship illustrates the issue. After an input has ended, assume concentration declines as \(C(t)=C_0e^{-kt}\), where \(C_0\) is concentration at the reference time, \(k\) is a positive elimination-rate constant, and \(t\) is elapsed time. This illustration assumes the same patient, dose history, and elimination rate across the compared times, with no additional dose or other process affecting concentration during the interval. It is not a general model for every dosing situation.
Worked Example: Comparing Two Sample Times
Consider a hypothetical medicine for which the model uses an 8-hour half-life in a simple post-dose elimination interval. A target-site concentration is collected 14 hours after the reference time, while the source-site comparison is at 12 hours. Assume no dose or other concentration-changing event occurs between 12 and 14 hours, and use the exponential relationship above.
The time difference is \(\Delta t=14-12=2\) hours. The predicted ratio is
If \(C(12)=8.00\ \text{mg/L}\), then
Under the stated assumptions, the later sample is about \(1.273\ \text{mg/L}\) lower, or about 15.9% below the 12-hour value. This calculation does not establish that either site’s measurements are wrong. It shows why the target site’s actual collection time and the model’s expected timing must be compared before treating the values as equivalent. If dosing or sampling conventions are more complex, the applicable model and its assumptions—not this simple illustration—must guide the assessment.
Prescribing Practices and Model Assumptions
A model may use a regimen as more than a label: dose amount, interval, infusion duration, dose interruptions, and the timing of changes may all affect how an observation is interpreted. Compare the prescribing practices in the target setting with the regimen information the model accepts. Check whether the model represents the relevant event history, and whether routine target-site practices create patterns not covered by its intended use.
Likewise, compare the target population with the model’s intended population and evidence base. Relevant questions include whether required covariates are measured with compatible definitions, whether values fall within a supported range, and whether important subgroups are common in the target setting but sparse or absent in the source data. A range overlap does not by itself prove suitability, and a difference in patient mix does not by itself prove failure. The point is to identify what evidence or adaptation is needed before deployment.
Worked Example: Checking Compatibility Across Several Requirements
A hypothetical service is being considered for a target ward. Reviewers examine 100 recent episodes against three pre-specified requirements: an administration timestamp available at decision time, a concentration with a documented collection time, and a regimen represented by the model. The data show that 82 episodes meet the first requirement, 90 meet the second, and 76 meet the third. A record-level review finds that 61 episodes meet all three requirements.
The fraction meeting all requirements is \(61/100=0.610\), or 61.0%. This is the relevant joint fraction for the three checks; the separate percentages cannot simply be multiplied or treated as the joint fraction because the requirements may co-occur in the same episodes. The review also finds 39 episodes that fail at least one check, since \(100-61=39\).
These figures identify a material operational gap for this hypothetical review, but they do not show that the model is inaccurate in the 61 eligible episodes or that it is unusable in every target episode. The next assessment should determine which failures reflect remediable data capture, which reflect incompatible prescribing practice, and which fall outside the model’s supported use. Any eligibility rule should be explicit and implementable at the time the recommendation is produced.
Choosing a Response and Detecting Weak Transportability
No single response fits every difference. The principal alternatives have different implications:
- Use the same model unchanged. This is defensible only when input meanings, timing, prescribing practices, population, and workflow are sufficiently aligned for the intended use, with appropriate evidence in the target setting.
- Harmonize inputs or modify workflow. A stable unit conversion, clear event mapping, or earlier capture of administration times may address a difference if the transformation preserves clinical meaning and is tested end to end.
- Restrict eligibility. The service may be limited to cases with complete, compatible inputs, provided exclusions are detectable in time and do not silently remove important patient groups.
- Adapt or rebuild the model. If target data indicate a genuine population or model-relationship difference, adaptation requires suitable data and renewed evaluation; relabeling an input or changing a threshold alone is not evidence that the model has transferred.
- Defer deployment. If the required information or practice cannot be reconciled, withholding recommendations may be safer than presenting outputs whose assumptions are unmet.
Diagnostics should be planned around the suspected barriers. Compare field definitions with clinical owners; inspect missingness and availability at the actual decision time; summarize sample times relative to dose events; and examine regimen patterns and covariate ranges against the model’s intended use. Where target data permit, evaluate predictions and calibration in the target context, including relevant subgroups. The earlier material on external validation, temporal validation, calibration, and prediction error patterns provides the corresponding performance-evaluation concepts; this assessment determines whether the target data and care process make those evaluations interpretable.
A common mistake is to conclude that transportability is established because the target site uses the same laboratory units and the model’s software runs successfully. That reasoning checks representation and technical execution, not whether the data represent the same events, the samples align with the model’s timing assumptions, or actual prescribing practices are supported. The correction is to verify meaning and timing first, then assess population and workflow compatibility, and finally evaluate model performance in the target context.
Transportability has limits even after careful assessment. Care settings change, data may be incomplete in ways that are hard to observe, and a pilot or retrospective review may not represent future practice. A favorable compatibility assessment supports a defined use under documented conditions; it is not a guarantee of performance in every patient or future workflow. Reassessment is warranted when data definitions, measurement schedules, prescribing practices, or the intended population change.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A target system records concentration in the same units as the source, but not the time since the last dose. Is the field necessarily equivalent? Explain.
- Under the exponential relationship, a half-life is 6 hours and two measurements are 3 hours apart. What is the later-to-earlier concentration ratio, to three significant figures?
- In a review of 80 episodes, 50 satisfy all pre-specified data and regimen requirements. What fraction satisfies all requirements, and what does that fraction not establish?
- Give one example of a harmonizable difference and one example of a difference that cannot be resolved by a simple conversion.
Solutions. 1. No. The units match, but without dose-relative timing the measurement’s position in the interval is unknown, so equivalence to a source-site sample schedule has not been established. 2. The ratio is \(2^{-3/6}=2^{-0.5}\approx0.707\). Under the stated elimination assumptions, the later concentration is about 70.7% of the earlier one. 3. \(50/80=0.625\), or 62.5%. This describes eligibility for the specified checks in those reviewed episodes; it does not establish model accuracy, representativeness, or future performance. 4. Converting a consistently recorded weight from pounds to kilograms can be harmonizable. Replacing an administered dose with an ordered dose is not a simple conversion because the order does not establish what the patient received.