Communicating Model Uncertainty to Clinicians

Why Uncertainty Needs Its Own Message

A dose recommendation is often shown as a single value, but the model’s estimate is not exact. A clinician needs to know whether uncertainty concerns an estimated model quantity, a future patient outcome, or the reliability of the data supplied to the model. These are different issues and should not be collapsed into one vague label such as “low confidence.”

For example, a model may estimate a patient’s expected concentration with uncertainty because its parameters were estimated from finite data. Even if that expected concentration were known exactly, an individual future concentration could still vary. Separately, a sample time entered incorrectly could make the model’s output unreliable for reasons not represented by either statistical interval. Clear communication identifies which source is being described and what it means for interpretation.

State What an Interval Represents

An interval is interpretable only when its target and construction are clear. An approximate 95% confidence interval for a model-based mean describes uncertainty in estimating that mean. In repeated use under the stated model and sampling assumptions, intervals constructed this way would contain the fixed mean about 95% of the time. It does not mean there is a 95% probability that the fixed mean lies in this particular interval.

A prediction interval instead concerns a future patient outcome or observation. It is generally wider because it includes outcome variability in addition to uncertainty in the estimated mean. In a Bayesian analysis, a credible interval has a different interpretation: conditional on the specified model, data, and prior, it contains the parameter with the stated posterior probability. The display should name the interval type; a bare “95% interval” leaves an important ambiguity.

A common approximate calculation for a normally distributed estimate is estimate plus or minus 1.96 standard errors. This gives a two-sided 95% confidence interval when the normal approximation is appropriate and the standard error estimates uncertainty in the target estimate. It is not automatically a prediction interval. The assumptions, target, and source of the standard error must be checked before using the formula.

Translate Prediction Uncertainty Into Clinical Quantities

A prediction interval can be useful, but often a probability tied to a clinically relevant range is easier to interpret. If a future concentration \(X\) is assumed to follow a normal predictive distribution with mean \(\mu\) and standard deviation \(\sigma\), then a probability for a target range can be calculated by standardizing its limits. This calculation depends on the predictive distribution being suitable for the patient and outcome; the number is not reliable merely because the arithmetic is correct.

Show the Sources of Predictive Uncertainty

A useful explanation distinguishes uncertainty in the estimated mean from variability among individual outcomes. In a simple setting where the two components are independent and approximately normal, their variances add. This gives a direct justification for why a prediction interval is wider than an interval for the mean. Real models can require more complex calculations, but the communication principle remains: specify which sources are included and avoid describing a parameter interval as if it covered individual outcomes.

The variance addition above requires independence of the components in this simplified example. If they are dependent, a covariance term is needed; if the predictive distribution is skewed or otherwise non-normal, a normal interval may also be inappropriate. The actual model’s predictive distribution, rather than a convenient formula, should determine the reported interval or probability.

Choose Graphics That Preserve the Meaning

A good graphic makes both the result and its uncertainty legible without implying more certainty than the model supports. The graphic should label the quantity being shown, its units, the interval type and level, and any clinical target range. A point estimate can be shown with an interval whisker, but the caption or label must say whether the whisker is a confidence interval, credible interval, or prediction interval.

• For an estimated mean: show the estimate and its confidence or credible interval, labelled as uncertainty in the mean.
• For an individual prediction: show a predictive distribution or prediction interval. If a target range matters, mark that range distinctly and, when justified, report the probability within it.
• For multiple dose options: use a consistent scale and show the same type of uncertainty for each option. Do not mix interval types without explicit labels.
• For warnings: use a separate visual element and plain-language text for input concerns or model applicability. Do not encode these as wider statistical intervals unless they are actually represented in the model.

A shaded predictive density can show where probability is concentrated and can highlight the portion inside a target band. However, the shaded area should not obscure the target limits, and a density plot should not be presented as a guarantee that the realized concentration will follow that shape. A compact point-and-interval display may be clearer when clinicians need to compare a small number of quantities.

Keep Statistical Uncertainty Separate From Warnings

A model’s interval quantifies uncertainty under a specified statistical model. It does not automatically account for every concern relevant to a clinical decision. A sample with an uncertain collection time, a missing covariate, or a value that fails a data-quality check may not be represented by the interval calculation. Likewise, a patient outside the population or setting in which the model is supported may face model-applicability concerns that the interval does not capture.

Communicate these issues separately and specifically. For example: “The predicted trough is 8.0 mg/L; the 95% prediction interval is 3.4 to 12.6 mg/L. The recorded sample time is unverified, so the input may not represent a trough.” The interval reports model-based outcome uncertainty; the second sentence identifies an input concern. If the model does not represent a relevant clinical circumstance, say that as a limitation rather than suggesting the interval has incorporated it.

An interval can be narrow and still be misleading if the model is misspecified or applied beyond its supported context. Conversely, a wide interval can be an honest reflection of limited information rather than evidence of bad data. Clinically useful communication explains the source and consequence of the concern, not just a confidence label.

Diagnostics, Alternatives, and Common Pitfalls

Before presenting an interval or probability, check that the model output is finite, on the intended scale, and compatible with the stated units. Examine whether the assumed distribution is reasonable, whether the interval is implausibly asymmetric or extends into impossible values, and whether the patient’s inputs fall within the model’s supported range. Check that uncertainty calculations correspond to the intended target: a model parameter, a conditional mean, or a future individual outcome. If a calculation relies on a normal approximation, inspect whether that approximation is defensible; a normal interval for a nonnegative concentration can assign probability to negative values.

The main alternatives serve different purposes. A single point estimate is concise but hides uncertainty. A confidence interval communicates uncertainty about an estimated mean, while a prediction interval communicates the spread of future outcomes. A predictive probability within a target range can be more directly relevant to a clinical question, but depends on the predictive distribution and chosen target boundaries. A Bayesian credible interval can express posterior uncertainty conditional on the model, data, and prior; it should be named as such rather than treated as interchangeable with a frequentist confidence interval.

• Wrong: “There is a 95% chance the fixed mean is in this frequentist confidence interval.” Correction: describe the repeated-sampling coverage of the method, or use a Bayesian credible interval when that posterior probability statement is intended.
• Wrong: “The 95% confidence interval contains 95% of future patient concentrations.” Correction: use a prediction interval for individual outcomes.
• Wrong: “The interval is wide, so the data must be poor.” Correction: width can reflect parameter uncertainty, individual variability, or both. Assess data quality directly.
• Wrong: “The model’s interval covers all uncertainty.” Correction: state its included sources and separately disclose data concerns and model limitations that are not represented.