Fallback Dosing When Model Output Is Unusable

Why a Fallback Must Be a Defined Action

A dosing model may fail to return a recommendation, return a result that cannot safely be interpreted, or encounter a patient for whom its use is not supported. A safety policy must specify what happens in each situation before the situation occurs. Otherwise, a user may improvise, reuse an old recommendation, or assume that a missing result means “no change.” None of those responses is necessarily safe.

A fallback is not simply any action taken when the model fails. It is a pre-specified, independently governed alternative dosing approach, together with rules for deciding whether that approach applies. Depending on the medicine and clinical context, the alternative might be a standard protocol, a label-based regimen, or another approved dosing pathway. The particular approach must be chosen and validated for the intended setting; the model must not invent it at the moment it fails.

Definition: A fallback dosing policy specifies (1) the conditions that make a model output unavailable, invalid, or unsupported; (2) a named standard approach to use when those conditions occur and that approach is applicable; and (3) what to do when no safe, applicable fallback is available. A fallback policy is a rule for routing a dosing decision, not a guarantee that every fallback dose is safe for every patient.

The central distinction is between the model output being unusable and the patient being eligible for the fallback. Failure of the first does not establish the second. A standard regimen may have its own exclusions, required measurements, dose limits, or patient-specific conditions. A contraindication screening gate therefore still applies: a fallback is not a way to bypass a contraindication.

Define the Routing Rule Before Use

For a particular dosing decision, define a set of checks that must all pass before a model recommendation is actionable. These checks should reflect the intended use and the model’s validated operating conditions. At a minimum, they usually cover whether the model returned a result, whether the result is interpretable and internally consistent, and whether the patient and inputs are within the supported scope.

Definition: Let \(V\) mean that a recommendation is available and passes the required output checks, and let \(A\) mean that the specified standard approach is applicable to this patient and decision. The routing rule is: use the model recommendation only when \(V\) is true; use the specified standard approach when \(V\) is false and \(A\) is true; if both \(V\) and \(A\) are false, do not generate an automatic dose from this policy. The unresolved decision requires the appropriate clinical process rather than an improvised fallback.

“Available” should mean more than a populated field. A recommendation can be unavailable because computation failed, the service timed out, or a required result was not returned. It can be invalid because it is malformed, ambiguous, inconsistent with its own stated units or schedule, or fails predefined plausibility and safety checks. It can be unsupported because the patient or decision lies outside the model’s specified population or operating conditions. These categories help identify different system failures, but they lead to the same routing question: may this output be acted on?

The checks should be operational rather than subjective. For example, a policy might require a dose amount, unit, route, and interval; a valid patient and encounter identifier; an allowed dose range; and a model status indicating that the result is supported. “Looks reasonable” is not a reproducible validity criterion. Likewise, a plausible dose does not repair a missing unit or establish that the model was used within its supported population.

Why the Fallback Is Conditional

The routing rule can be justified by separating the failure of the model from the suitability of the alternative. If the model output is unusable, it cannot serve as the basis for an automated recommendation under the policy. A standard approach is a candidate alternative, but it is acceptable only if its own applicability conditions pass. If they do not, neither candidate has been established as a safe automatic action.

\text{automatic action} = \begin{cases} \text{model recommendation}, & V = 1,\\ \text{specified standard approach}, & V = 0 \text{ and } A = 1,\\ \text{no automatic dose from this policy}, & V = 0 \text{ and } A = 0. \end{cases}

This rule does not claim that the fallback is risk-free or superior for every individual. It makes a narrower, auditable claim: a model recommendation is used only after defined checks pass, and a pre-specified alternative is used only when its own conditions pass. If neither condition is met, the policy avoids presenting an unsupported dose as though it were a valid recommendation.

The standard approach should be specified in enough detail to reproduce the action: the regimen or protocol version, dose units, route, schedule, relevant patient inputs, eligibility criteria, and any necessary dose calculations or limits. It should also state how conflicts are handled—for example, when required data are missing or the patient meets a listed exclusion. A name such as “usual care” is not enough if it does not determine what action to take.

Worked Example: Model Service Unavailable, Standard Regimen Applicable

Hypothetical example. A fictional medicine has a pre-specified adult standard regimen: give a loading dose of 5 mg/kg, capped at 400 mg, followed by 2.5 mg/kg every 24 hours. The protocol applies to adults with a recorded current weight and no listed exclusion. For this example, the patient is an adult weighing 72 kg, has no listed exclusion, and all required information for the protocol is present. The model service times out and returns no recommendation.

The output is unavailable, so \(V=0\). The standard regimen’s conditions are met, so \(A=1\). The policy therefore selects the standard regimen rather than treating the timeout as a dose recommendation or reusing a prior model result.

The loading-dose calculation is:

5\ \text{mg/kg} \times 72\ \text{kg} = 360\ \text{mg}.

The uncapped dose is 360 mg, which is below the 400 mg cap, so the loading dose is 360 mg. The maintenance calculation is:

2.5\ \text{mg/kg} \times 72\ \text{kg} = 180\ \text{mg}.

Under the stated hypothetical protocol, the fallback is therefore a 360 mg loading dose followed by 180 mg every 24 hours. The arithmetic does not itself establish clinical appropriateness; applicability follows from the assumed protocol criteria and patient information. If an exclusion had been present, \(A\) would not be true and this routing rule would not authorize the regimen.

Worked Decisions Across Failure Types

Worked Example: Invalid Output Despite a Plausible Patient

Hypothetical example. A fictional adult dosing service is specified to return a dose in milligrams and an interval in hours. Its output must be within 100–600 mg per dose and must include both fields. A standard protocol, applicable to this patient, specifies 300 mg by mouth every 24 hours. The model returns “900 mg, every 24 hours.” The patient otherwise meets the protocol’s eligibility conditions.

The output contains a dose and an interval, but the dose is outside the allowed 100–600 mg range. The predefined validity check therefore fails: \(V=0\). It is not appropriate to clip the model result to 600 mg unless such a transformation was explicitly validated and specified as part of the policy; silently altering an output creates a new, unvalidated dosing rule.

The standard protocol applies, so \(A=1\). The fallback action is 300 mg by mouth every 24 hours, as specified by that protocol. The fact that 900 mg is numerically interpretable does not make it valid. Conversely, the fallback is not chosen merely because it is lower; it is chosen because the model failed a defined check and the separate standard-protocol criteria pass.

Worked Example: Unsupported Population With an Applicable Standard Pathway

Hypothetical example. A fictional model is validated only for adults and receives data for a 14-year-old patient weighing 32 kg. Its output is marked “unsupported.” The approved standard pediatric pathway uses weight bands: patients weighing 20 to less than 40 kg receive 150 mg once daily, provided no listed exclusion applies. Assume the patient has no such exclusion.

The patient is outside the model’s supported population, so the model recommendation is not actionable and \(V=0\), even if the software happens to display a numerical result. The pediatric pathway is applicable: \(32\) kg satisfies \(20 \leq 32 < 40\), and the stated exclusion check passes. Thus \(A=1\), and the specified fallback is 150 mg once daily.

This decision does not extend the adult model to a pediatric patient. It routes the decision to a separately specified pathway whose eligibility conditions include this patient’s age and weight band. If the weight were missing, or if the patient met an exclusion, the example’s conditions would no longer establish \(A=1\); the fallback must not be guessed from the nearest band.

Implementation, Diagnostics, and Alternatives

A fallback policy is only as reliable as its implementation. The system should record why the model output was rejected, which standard approach was selected, whether its applicability checks passed, and the resulting dose and schedule. The record should identify the relevant protocol version and preserve the input values used in any calculation. These details make it possible to distinguish a genuine fallback from a model recommendation, a manual change, or a software defect.

Useful diagnostics include the frequency and reasons for fallback use, the proportion of fallback cases that lack required standard-protocol inputs, and whether fallback calculations conform to the protocol. Results can be reviewed by relevant patient groups and by failure type. A high rate of timeouts, invalid outputs, or unsupported cases may indicate a technical or deployment problem rather than a need to normalize frequent fallback use. The safety of the overall policy should also be evaluated using outcomes relevant to the dosing decision; merely observing that a fallback dose was issued does not establish that it was safe or effective.

A common alternative is to leave the dose unchanged, reuse the last model recommendation, or substitute a default number whenever the model fails. These are not equivalent to a specified standard approach. A prior recommendation may be stale, based on different inputs, or no longer applicable; a default may not reflect the patient’s regimen or eligibility. Automatic fallback is also different from a clinician deciding to use another approach: the former follows pre-authorized criteria, while an individual clinical decision requires its own appropriate process.

The fallback policy has important limits. It cannot compensate for incorrect patient data, a poorly designed standard regimen, or an unrecognized contraindication. A protocol may be unsuitable for a patient even when its recorded eligibility fields appear to pass. A fallback also cannot guarantee continuity of the model’s intended treatment effect, because the model and standard approach may make different assumptions or use different information. Finally, a policy that routes every failure to the same regimen may be unsafe if the model has multiple intended uses with different appropriate alternatives.

Common Mistake

Common mistake: “The model output is missing, so use the usual dose.” This reasoning treats model failure as proof that a default is appropriate. The correction is to identify the named standard approach and check its applicability conditions independently. If those conditions are not met, the fallback rule does not authorize an automatic dose.

Another mistake is to repair an invalid model output informally—for example, by rounding an implausible number, supplying a missing interval from memory, or clipping a dose to a limit. Unless that exact transformation is specified and validated, it is a new dosing policy, not a safe interpretation of the model result.

Key takeaway: A safe fallback is defined in advance, distinguishable from the model recommendation, and conditional on its own eligibility criteria. If the model output is unusable and no specified standard approach applies, the policy should not manufacture an automatic dose.

Check Your Understanding

For each question, apply the routing rule: first assess whether the model output is actionable, then assess whether the specified standard approach applies.

A model returns no result. The standard protocol applies only when a current weight is recorded, but weight is missing. Does the fallback policy authorize an automatic dose?
A hypothetical standard regimen is 4 mg/kg, capped at 250 mg. For a patient weighing 70 kg, calculate the dose if the regimen applies.
A model is explicitly unsupported for a patient’s age group, but it displays a plausible dose. Is that dose actionable under the policy described here?
Why is clipping an out-of-range model dose to the maximum allowed dose not automatically a valid fallback?

Solutions. (1) No. The model is unavailable, but the required weight for the standard protocol is absent, so the information given does not establish that the fallback applies. The policy does not authorize guessing a dose. (2) The uncapped calculation is \(4\ \text{mg/kg} \times 70\ \text{kg}=280\ \text{mg}\). Because 280 mg exceeds the 250 mg cap, the specified dose is 250 mg. (3) No. A plausible number does not remove the unsupported-population status; the output fails the supported-use condition. A separately specified and applicable standard pathway would need to be assessed. (4) Clipping changes the model output. Unless that exact transformation is pre-specified and validated, the clipped amount is neither an approved model recommendation nor necessarily the specified standard fallback.