Making Recommendations Actionable in the EHR

From a Recommendation to an Order

A dose recommendation is not yet a medication order. It may describe a target amount or regimen, but an order must also tell the EHR—and the clinician reviewing it—what product to use, how to administer it, and when it applies. If any of those details must be guessed or re-entered, the handoff can change the intended treatment.

This tutorial focuses on making that handoff explicit. The goal is not to have a model silently prescribe. It is to transform its output into a structured, reviewable order draft that preserves the intended regimen and makes necessary checks visible before a clinician acts. The interpretation of model uncertainty is a separate part of the recommendation; this page addresses how to carry the chosen recommendation into an order without losing its meaning.

Definition: An actionable EHR recommendation is a model-informed recommendation presented with enough structured information for an authorized clinician to verify the patient, medication, dose, formulation, route, schedule, timing, and applicable review instructions, and—if appropriate—to approve or modify an order. A recommendation is not itself an authorized medication order.

A useful test is whether a clinician can answer, without calculation or interpretation: what medication, what amount per administration, by what route, on what schedule, starting when, and what must be reviewed? “Give 7.5 mg/kg” does not answer all of those questions. Even a value such as “140 mg” is incomplete if the selected formulation, route, schedule, or timing is unclear.

Preserve Meaning Across the Handoff

An EHR recommendation typically starts with model output and patient context, then maps to a medication catalog item and a proposed order. Each transformation must preserve the intent. The EHR’s display should distinguish the model’s calculated result from any value changed through rounding, catalog selection, or clinician editing.

Key principle: Preserve the recommendation’s clinical meaning across the handoff: medication and formulation, dose amount and units, route, schedule, and temporal anchor must remain consistent. If the system cannot make a safe, unambiguous mapping to an available orderable, it should show the unresolved issue rather than silently substitute or guess.

Before drafting an order, check that the recommendation refers to the correct patient and encounter, uses the intended inputs, and is still applicable at the time of review. These checks rely on the earlier material on clinical data quality checks, missing inputs, identity reconciliation, and aligning doses, samples, and clinical events in time. Here, the practical consequence is that the order must carry enough context to make its assumptions and timing verifiable.

A structured order should make the following information explicit when applicable:

Medication identity: the intended ingredient and a matching orderable, including strength and formulation. A liquid and a tablet are not interchangeable fields merely because they contain the same ingredient.
Dose: the amount for one administration, with units. If a weight-based result is converted to a total amount, preserve the calculation or source value so the conversion can be checked.
Route and administration details: for example, oral or intravenous, and any relevant infusion duration or delivery instruction.
Schedule and timing: frequency, start time or next due time, and duration or stop condition when specified. A frequency alone does not establish when the first or next dose is due.
Review instructions: what information needs checking, when it should be checked, and who should review it, where that is part of the recommendation.

The orderable selected from the EHR catalog must match the intended ingredient, strength, dosage form, and route. Display both the dose in the units used by the order and any calculated administration volume when that helps prevent a conversion error. Do not present a volume without its concentration, or a dose without units. Use explicit, unambiguous schedule wording rather than shorthand that could be interpreted differently across interfaces.

A Structured Translation Process

Confirm context.
Verify the patient and encounter, the recommendation’s inputs and time context, and whether it is intended to apply now or at a later scheduled dose.

Resolve the medication.
Match the intended ingredient, strength, formulation, and route to an available orderable. Stop for review if there is no unambiguous match.

Calculate and check the order amount.
Convert the recommendation into the total dose and, if needed, the administration volume or infusion rate. Show the units and check the arithmetic.

Specify the regimen.
State route, schedule, start or next-dose time, and any duration or administration instruction required for the order.

Attach review information and verify the display.
State the relevant review step, then check that the displayed order matches the recommendation and does not introduce an unsupported substitution.

The process depends on assumptions that should be checked, not taken for granted: the patient and encounter are correctly identified; inputs used by the model are appropriate and current for this decision; the model output is valid for the intended use; the order catalog mapping is correct; and the EHR can represent the requested dose and schedule without loss of meaning. If an assumption fails, the safe output is a clearly identified issue for resolution—not a plausible-looking but mismatched order.

Worked Order Translations

Worked Example: Converting a Weight-Based Oral Recommendation

Hypothetical training example only. A model recommends 7.5 mg/kg per administration for a fictional medication. The verified weight for this example is 18.4 kg. The available oral suspension is 40 mg/mL. Assume the approved example protocol permits rounding the calculated dose to the nearest 10 mg, and that the schedule is every 12 hours by mouth, beginning at 08:00 today.

Calculate the unrounded dose:

7.5\ \frac{\text{mg}}{\text{kg}}\times 18.4\ \text{kg}=138\ \text{mg}

The units cancel to mg. Applying the stated rounding rule gives 140 mg per administration. The corresponding volume is:

\frac{140\ \text{mg}}{40\ \text{mg/mL}}=3.5\ \text{mL}

The order draft should therefore identify the matching oral suspension, state 140 mg (3.5 mL) by mouth every 12 hours, starting at 08:00 today, and display the concentration used in the volume conversion. A suitable review instruction in this hypothetical setup is to verify the weight and selected formulation before approving the first dose. The calculated 138 mg should not be displayed as though it were the final rounded order amount; showing both the calculation and the rounding makes the change transparent.

This mapping assumes the weight and concentration are correct and the example’s rounding rule applies. If the catalog offers a different concentration, the volume must be recalculated from that concentration; the system must not carry over 3.5 mL automatically.

Worked Example: Representing an Intravenous Dose and Infusion

Hypothetical training example only. For a fictional medication, the model recommends 0.8 mg/kg for a patient weighing 62.5 kg. The intended route is intravenous infusion over 30 minutes. In this example, the prepared final infusion volume is specified as 50 mL, with a final concentration of 1 mg/mL.

Calculate the dose:

0.8\ \frac{\text{mg}}{\text{kg}}\times 62.5\ \text{kg}=50\ \text{mg}

At 1 mg/mL, the amount in 50 mL is:

1\ \frac{\text{mg}}{\text{mL}}\times 50\ \text{mL}=50\ \text{mg}

The dose and prepared volume are consistent. If the infusion is to run evenly over 30 minutes, then 30 minutes is 0.5 hours, and the corresponding pump rate for this specified final volume is:

\frac{50\ \text{mL}}{0.5\ \text{h}}=100\ \text{mL/h}

The proposed order should identify the matching intravenous orderable and state 50 mg in a final volume of 50 mL, infused intravenously over 30 minutes. If the EHR requires a pump rate, it is 100 mL/h under the example’s stated final-volume assumption. Including only “50 mg IV” would omit the infusion instruction; including “100 mL/h” without the final volume and duration would leave the intended administration unclear.

The calculation depends on the final infusion volume being 50 mL. If that volume is not specified or the selected preparation has a different final volume, the system cannot safely infer this rate. It must use the correct preparation information or flag the missing detail for resolution.

Worked Example: Anchoring the Next Dose to the Medication Record

Hypothetical training example only. A model recommendation specifies a fictional medication at 200 mg orally every 24 hours. The medication record shows the most recent administration was at 06:00 today. Assume the record is verified, no intervening dose is documented, and the recommendation is intended to continue the same interval.

The interval is 24 hours, so the next dose is due 24 hours after the recorded administration:

06{:}00\ \text{today}+24\ \text{h}=06{:}00\ \text{tomorrow}

The order draft should state 200 mg by mouth every 24 hours, next dose due at 06:00 tomorrow, rather than “start now.” That timing is conditional on the administration record being complete and accurate. The review step is to confirm the most recent administration and due time before approval. If a dose was given but is missing from the record, or the record’s time zone or date context is wrong, the calculated next time may be unsafe. In that case, the system should not present the derived time as verified.

Diagnostics, Alternatives, and Limitations

A recommendation can look complete while still failing as an order. Check the displayed result against the model output and the source context. Useful diagnostics include verifying that dose units agree before and after conversion, that the selected concentration produces the displayed volume, that the route matches the formulation, and that the timing is consistent with the documented administration history. Review the order summary as a clinician would see it; do not rely solely on a correct calculation hidden in a technical detail panel.

Warning signs include a dose with no unit, a volume without a concentration, a route that does not match the selected formulation, a schedule without a clear first or next dose time, and an orderable that differs from the intended medication. Another warning sign is a value that changed during mapping or rounding without a visible explanation. These are not cosmetic issues: each can alter what is administered or when.

A structured draft is generally more actionable than a free-text recommendation because its fields can be checked and transferred into the EHR order without asking the clinician to infer them. A free-text-only approach may still communicate rationale, but it can leave dose units, timing, or product selection ambiguous. Manual transcription gives the clinician control but requires re-entry and creates opportunities for copying or unit errors. Structured fields do not remove the need for review: they make the proposed order easier to inspect and reduce avoidable translation steps.

There are limits to what automated mapping can establish. A calculation cannot verify that a dose is clinically appropriate merely because its arithmetic is correct. Nor can an order draft resolve contradictory source records, missing administration history, uncertain formulation equivalence, or a recommendation outside its intended use. The system should preserve those issues for clinical review rather than hide them behind a completed-looking order. A clinician remains responsible for reviewing and authorizing the order according to the applicable workflow.

Common Mistakes

Treating the model output as the order. A calculated dose does not specify all order fields. Add the correct formulation, route, schedule, and timing before presenting an actionable draft.
Showing a volume without its basis. “3.5 mL” alone cannot be checked. Display the concentration and corresponding dose, and recalculate if the formulation changes.
Using frequency as a substitute for timing. “Every 24 hours” does not establish whether the next dose is due now or tomorrow. Anchor timing to the verified medication record when required.
Silently substituting an orderable. A similar name or strength does not guarantee a matching formulation or route. Flag a mismatch instead of assuming equivalence.
Hiding rounding or conversions. A changed total dose or derived volume should be traceable to its inputs and stated rule, so the reviewer can check it.

Key takeaway: Make the recommendation actionable by translating it into explicit, checkable order fields—not by asking the clinician to reconstruct the regimen. Preserve medication identity, dose and units, route, schedule, timing, and necessary review context; flag any mapping that cannot be verified.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A hypothetical model recommends 6 mg/kg per dose for a 25 kg patient. What is the calculated total dose, and what further details are needed before it becomes an order?
A liquid contains 25 mg/mL. A hypothetical order is for 75 mg per administration. What volume corresponds to that dose?
A verified medication record shows a dose at 14:00 on Monday. A hypothetical regimen is every 8 hours. What is the next scheduled time if no intervening dose or scheduling change applies?
Why is “100 mg IV” not a complete infusion instruction if the recommendation specifies a 30-minute infusion?

Solutions. (1) \(6\ \text{mg/kg}\times25\ \text{kg}=150\ \text{mg}\). The order still needs a matching medication and formulation, route, schedule, applicable start or next-dose time, and any required administration or review instructions. (2) \(75\ \text{mg}\div25\ \text{mg/mL}=3\ \text{mL}\). (3) Eight hours after 14:00 Monday is 22:00 Monday. This depends on the record and interval being correct. (4) The statement gives an amount and route but not the infusion duration or, where needed, final volume and rate. The order must carry the specified administration details so the infusion is not left for interpretation.