Clinical audit: a practical guide

Clinical audit is a quality improvement process that measures care against an explicit standard, identifies where practice falls short, changes something, and then measures again. It is defined by that last step: an audit that does not re-measure is a snapshot, not an audit.

The standard is what separates it from everything else. You are not asking "how are we doing?" — you are asking "are we meeting this specific, written, defensible standard, and if not, by how much?" That makes the result actionable in a way that a general review is not.

Every version of the cycle has the same shape. The number of stages varies between four and six depending on who drew it; the logic does not.

A criterion is the thing you measure. A standard is the level you expect. "Antibiotics reviewed within 72 hours" is a criterion; "in 95% of cases" is the standard.

Criteria are conventionally split three ways, after Donabedian:

Most audits should measure process. Outcomes are what everyone cares about, but they are affected by case mix, chance and much that is outside the service's control — which means a change in an outcome rarely tells you whether practice changed. Process criteria are directly actionable and need far smaller samples to interpret.

This is the question audit teams ask most often and the one most often answered with "twenty, because that is what we did last time".

The honest answer depends on what the audit is for. If you only need to know roughly where compliance sits, a modest sample is fine. If you need to show a change between two audits is real rather than noise, you need enough records that the confidence intervals do not overlap — and that is usually more than people expect.

The practical rule: the smaller the change you need to detect, the larger the sample. Detecting a move from 80% to 90% takes roughly 250 records per round. Detecting 80% to 85% takes closer to 1,000 — which is usually the point at which a team decides the audit question needs rethinking rather than the sample enlarging.

Most of the error in an audit enters here, and almost none of it is statistical.

Audit analysis is mostly proportions, and the discipline is in reporting them honestly rather than in anything advanced.

Always report the denominator. "78% compliance" means something different from 18 of 23 than from 390 of 500. Every figure in an audit report should carry the counts behind it.

Put a confidence interval on it. A single percentage implies a precision the sample does not have. An interval shows the reader how much the figure could move on a different fifty records, and it is what stops a re-audit "improvement" being celebrated when it is noise.

For repeated measurement, use a run chart. If you are auditing monthly, plotting the points over time tells you far more than comparing two of them. Statistical process control distinguishes ordinary variation from a real shift, which is exactly the question "has it improved?" is asking.

The stage that gets skipped. An audit that identifies 62% compliance and recommends "increased awareness" has not changed anything, and the re-audit will find 64%.

This decides whether you need ethics approval, and getting it wrong is the most common governance failure in NHS projects.

The practical test: if you are allocating anyone to anything, or if the answer is intended to change practice beyond your own organisation, it is probably research. The HRA decision tool is the formal route, and the answer should be recorded before data collection starts, not justified afterwards.

More on service evaluation and how it differs.

Note that every change is to a system rather than to effort, and every re-audit repeats the original design exactly. Those two things account for most of the difference between audits that improve care and audits that produce reports.