Clinical Prediction Rules
Tools that combine several clinical variables to estimate the probability of a diagnosis or outcome and guide a decision.
A clinical prediction rule (CPR) combines findings from history, examination and simple tests into a score that estimates the probability of a diagnosis or outcome — e.g. Wells, the Ottawa rules, HEART, PERC, CURB-65. A rule must be derived, then externally validated, and ideally impact-analysed, before clinical use. Appraise its discrimination (AUC) and its calibration.
A good rule discriminates → AUC 0.85; the gold dot is the score cut-off you actually triage on
How well a rule separates patients who will have the outcome from those who won’t, across every possible score cut-off. The further the curve bows to the top-left, the better the discrimination (a higher AUC); the dashed diagonal is a coin toss. The gold dot is the single cut-off the rule recommends for action — for example the HEART ≤3 “low-risk” threshold.
Read the whole curve (AUC) for overall discrimination, then read the chosen cut-off for the sensitivity and specificity you’d actually get when you apply the rule. But discrimination is only half the story: also ask whether the rule was externally validated in patients like yours, and whether its predicted risks are calibrated — a curve drawn from derivation data flatters the rule.
CPRs standardise decisions and can safely cut investigation and admission — but only within the population and setting where they were validated. A rule is a decision aid, not a replacement for judgement: it estimates probability, it does not examine the patient in front of you.
- Lifecycle:
derive → validate (external) → impact study - Appraise both discrimination (AUC) and calibration
- A derivation-only rule overfits and overperforms
HEART score — the model often held up as doing it properly. After derivation (Six et al., Neth Heart J 2008) it was prospectively externally validated by Backus et al. in 2,440 ED chest-pain patients across 10 hospitals (Int J Cardiol 2013): c-statistic 0.83 for 6-week MACE, beating TIMI (0.75) and GRACE (0.70), with a HEART ≤3 “low-risk” group running ~1.7% MACE. It was then taken further with impact studies (e.g. the HEART Pathway / HEART-Impact RCTs). A rule earns ED use through this whole chain — derive, externally validate, then impact-analyse — not on a single impressive derivation AUC.
- Adopting a rule after derivation only, with no external validation.
- Applying a rule to a different population or spectrum than it was validated in.
- Treating the rule as a replacement for clinical judgement rather than a decision aid.
Quick check
A newly-derived rule reports excellent performance in its development sample — can you adopt it straight into practice?
Answer: No — it needs external validation in a separate population (and ideally an impact study) first. Derivation performance is optimistic because the rule was fitted to that exact data.
Ready to build your plan? EMF Premium gives you all 40,000+ questions, 20 mocks and 1,215 OSCE stations from £29/month — or a one-off 3- or 6-month pass.