Safe Decisions with Missing or Delayed Measurements

Why Measurement Availability Is a Safety Question

A dosing policy can be well calibrated and still make an unsafe decision if it treats unavailable information as reassuring. A missing concentration is not a low concentration; a pending laboratory result is not a normal result; and a once-valid measurement may no longer describe the patient’s current state. The policy therefore needs an explicit response to measurement status, not just a rule for interpreting measured values.

The issue is especially important when a dose decision depends on a safety laboratory value or a measured drug concentration. A policy must determine whether a result is valid for the current decision, what to do if it is not, and how to prevent an automated recommendation from proceeding on the basis of an assumption disguised as data. The specific clinical action depends on the drug, the measurement, the patient, and the approved protocol; the general safety principle is that unknown status must remain unknown until valid evidence or an explicitly approved fallback resolves it.

Define Freshness Before Applying a Dosing Rule

Freshness is not a universal number of hours. It is a policy specification for a particular measurement and decision context. A rapidly changing laboratory value may require a shorter validity window than a relatively stable measurement; the appropriate interval also depends on clinical circumstances and the intended dosing action. A policy should state the interval in advance rather than deciding after seeing a value.

Record at least two times when they are available: sample time, when the specimen was collected, and result time, when the result became available to the system or clinician. The sample time determines the age of the biological information. A result delivered now from a specimen collected yesterday is not a measurement of the patient now. The result time still matters operationally: a long reporting delay may mean the patient’s decision was made without information that was eventually reported.

For measurement \(j\), let \(t_{s,j}\) be its sample time and let \(\tau_j\) be the policy-defined maximum age for using it. Its age at decision time is

The value meets the time criterion only when \(0\leq\Delta_j\leq\tau_j\). Passing this test alone is not enough: identity, units, specimen quality, and any relevant context must also be valid. If the sample time is absent or unreliable, the system cannot establish that the measurement is current. It should not silently substitute the result time or the time the record was entered.

A practical policy treats measurement handling as a sequence of checks: confirm that a result exists, establish its provenance and sample time, assess quality and units, compare its age with the applicable validity window, then decide whether the dosing rule may use it. If any required check fails, the policy routes the case to a specified response rather than passing an unverified value downstream.

Reasoning From an Old Measurement to Current Uncertainty

A previous value can sometimes inform a range of plausible current values, but it does not automatically identify the current value. To illustrate, suppose a hypothetical protocol has a safety boundary \(c\), and a measured value \(y\) was obtained \(\Delta\) hours before the decision. If the protocol has justified a maximum rate of change \(r\) and a measurement-error bound \(e\), then a conservative range for the current value is

This range follows from the assumptions \(|x(t_d)-x(t_s)|\leq r\Delta\) and \(|y-x(t_s)|\leq e\), where \(x(t)\) is the underlying value at time \(t\). Combining these inequalities gives

Thus the interval represents uncertainty under specified bounds; it is not a confidence interval, and it is only as credible as the assumptions that support \(r\) and \(e\). If the range includes states on both sides of a safety boundary, the old value does not establish that the patient is currently on the safe side. If no defensible change bound is available, this calculation cannot be used to certify current safety.

This leads to a useful policy principle: allow an automated action only when its safety conditions hold across the states the policy considers plausible, or when an approved clinical pathway explicitly permits a different response. If the available information cannot distinguish safe from unsafe states, the policy should not treat the favorable possibility as known fact. The appropriate response might be repeat measurement, clinician review, or a pre-specified interim plan; it is not automatically the same action for every drug or patient.

Specify Responses to Missing, Stale, and Pending Results

A safe policy makes the response explicit for each status. The response should distinguish what the system may do automatically from what requires a clinician, a repeat measurement, or an established fallback pathway. It should also specify how long an interim state may continue and how the case is closed once valid information becomes available.

The interim action is context-dependent. For one protocol, an unresolved result might block an automated dose increase while allowing a separately approved plan to continue. For another, the safe response may require prompt clinician review. A policy should not assume that “hold,” “continue,” or “repeat” is universally safest; each action has possible harms, and the approved pathway must address those trade-offs.

A pending result deserves particular care because its eventual value is unknown at the decision time. A system may record that a sample was collected and estimate when a result is expected, but neither fact supplies the result. If a dose is due before reporting is complete, the policy needs an explicit plan for that interval rather than an implicit assumption that the value will be acceptable.

Diagnostics, Failure Modes, and Alternatives

A policy should be tested not only with valid measurements but also with realistic failures in the measurement workflow. Useful checks include whether the system uses collection time rather than entry time, whether time zones and clock errors are handled consistently, and whether a corrected or rejected result can be mistaken for a final report. Test cases should include absent timestamps, duplicate results, mismatched units, delayed transmission, and a result that arrives after the decision has already been made.

Also inspect how often each status occurs and what happens next. Frequent missing or delayed results may indicate a workflow problem rather than an unusual patient pattern. If cases labelled “current” cluster around data-entry delays or missing sample times, the freshness logic may be ineffective. Review whether the policy reliably blocks unsupported automated decisions and whether escalation reaches an accountable person within the required timeframe. These are operational diagnostics, not evidence that a particular dosing action is clinically correct.

Several tempting alternatives are unsafe unless specifically justified and governed. Carrying the last observation forward makes an old value appear current. Replacing a missing result with a typical or population-average value hides individual uncertainty. Assuming a pending result will be in range turns an expectation into an observation. A validated model-based estimate may be considered only if the policy permits it, its inputs and uncertainty are appropriate for the decision, and its use has been evaluated for the relevant setting. It should be identified as an estimate, not presented as a measured result.

No freshness rule eliminates uncertainty. A result inside its time window can still be wrong, affected by a collection or processing problem, or inadequate for a patient whose condition changed rapidly. Conversely, a stale result may remain informative in some contexts, but it cannot be treated as current merely because it is the latest available value. Freshness criteria are one safety gate among the broader checks described in the earlier material on input validation, plausibility checks, contraindication screening, fallback dosing, and escalation.

Common Mistake

Wrong reasoning: “The last recorded laboratory value was acceptable, and there is no new result showing a problem, so the patient is probably still safe to dose automatically.”

Correction: The absence of a new adverse result is not evidence of a current acceptable value. Check the sample time, validity window, result status, and data quality. If current safety cannot be established under the policy’s rules, use the specified interim pathway instead of interpreting unknown information as reassuring.