Clinical evidence

Missing Is Not Zero: Designing Clinical Software That Keeps Uncertainty Visible

The problem in one example

A summary says "no signs of kidney impairment". Behind it, the most recent creatinine result is eight months old, an earlier value was never entered and two source systems disagree about the date of a key event. Each of these is a different kind of uncertainty. Collapsed into one sentence, they become a confident statement that no one actually verified.

This happens in rule engines, dashboards and language-model summaries alike. The fix is not a better model; it is an explicit state model that every layer respects.

Five states worth separating

State Meaning What the reviewer should see
Present A valid value from an identified source within the agreed time window The value, unit, time and source
Missing No value recorded where one was expected A clear "not recorded" label, never a zero or blank
Stale A value exists but is outside the window that makes it relevant The value, its age and the window
Conflicting Two or more sources disagree Both values and sources, unresolved until reviewed
Unlinked A value or explanation has no traceable source A warning that it cannot be verified

A recorded zero or a negative test result is information. A missing result is the absence of information. Software that shows both as "0" or as an empty cell invites the wrong conclusion.

Carry the state through every layer

  1. Data model. Store the state explicitly alongside the value, with the source identifier and timestamp. If you use FHIR R4, the Observation dataAbsentReason element and Provenance resources can carry part of this.
  2. Rules and models. Rules should declare what they do with each state: a rule that needs a recent value should return "cannot evaluate" when the value is missing or stale, not "false".
  3. Summaries. Any generated text must inherit the weakest state of its inputs. If one input is conflicting, the summary says so.
  4. Interface. Use consistent badges, and let the reviewer move from a cohort count to the case, from the case to the input and from the input to the source record in a few clicks.
  5. Audit. Record who resolved a conflict or accepted a stale value, and why.

Progressive disclosure for reviewers

Reviewers do not want to read every field. A good evidence view has three levels:

  • Summary: the finding and an overall evidence state.
  • Inputs and explanation: which inputs the finding used, their states and the rule or reasoning step that combined them.
  • Sources: the original records with identifiers and times.

Our clinical walkthrough shows this pattern with synthetic cases: a missing value stays distinct from a recorded zero, a conflict remains open and a cohort summary leads back to the individual case.

Testing it

Write acceptance scenarios for each state before building the interface: a case where a value is missing, one where it is stale, one with a conflict and one where an explanation has no source. Each scenario states what the reviewer must see. Clinical owners sign off these scenarios; engineers make them pass and keep them as regression tests.

Where this fits in a product

This is not a clinical decision in itself. It is the foundation that lets qualified people judge whether a finding is usable. It applies to clinical decision support, research cohort tools, registry software and any AI feature that summarizes patient information. If you are designing one of these and want a second pair of eyes on the evidence model, tell us about the workflow.

Technical references

Questions & answers

Why not just impute missing values?

Imputation can be appropriate for some statistical analyses, but in a review interface it hides the fact that a test was never done. Keep the original state visible and label any derived value as derived.

Does FHIR support this?

FHIR R4 Observation has a dataAbsentReason element for values that are absent, and provenance can be represented with Provenance resources. You still need to decide how your product displays and reasons over these states.

How many states do we need?

Start with five: present, missing, stale, conflicting and unlinked. Add product-specific states only when a reviewer would act differently on them.

Does this slow reviewers down?

The design aims to reduce manual searching with a clear state badge and a direct path to the source. Test review time and comprehension with your own reviewers before claiming an improvement.

LET’S START A CONVERSATION

Let’s start with your workflow.

Tell us what you want to improve and which systems you use. We will reply within one business day with questions, a view on fit and a practical next step.

Discuss a project