Clinical article · Specialties, Medicines and results

Clinical risk scores and formulas: how to choose, calculate and interpret them

Written for
Hospital physicians, emergency doctors and residents
Reading time
7 minutes
Published
Updated

Written by the Sana Health team for clinicians. It summarises published guidance, is not a guideline and has not been peer reviewed. Check current local protocols before acting.

Key points

  • Write down the decision first; a score is useful only if it was built and validated for that decision in patients like yours.
  • Check the version: guidelines change their preferred tools, as the 2024 ESC atrial fibrillation guideline did with CHA2DS2-VA.
  • Most calculation errors come from units, item definitions and missing values that were silently assumed normal.
  • A score can rank patients well and still misjudge absolute risk in a new population, so read discrimination and calibration separately.
  • Creatinine-based kidney function estimates are imprecise in individuals and invalid when creatinine is not at steady state.
  • Confirm any result close to a threshold that would change management in a validated calculator, and record the version used.

Start with the decision, not the score

A score answers one question for one kind of patient. Diagnostic rules estimate the probability that a condition is present now, prognostic models estimate the risk of a future event, and formulas such as the estimated glomerular filtration rate or the anion gap estimate a physiological quantity [1]. Before choosing a tool, state the decision it should inform: admit or discharge, image or not, anticoagulate or not, which dose.

There are far more models than well-validated ones. A systematic review of cardiovascular risk in the general population alone found 363 prediction models and 473 external validations [2]. Clinical prediction rules also perform differently from one population to another [3], so a well-known name is not enough.

Choosing a score: five questions

  • Was it derived and externally validated in patients like this one, in a similar setting, country and severity of illness?
  • Is it the version current guidance uses? The 2024 European Society of Cardiology (ESC) atrial fibrillation guideline moved to CHA2DS2-VA, which omits the point for female sex [4].
  • Does it fit the decision? Scores built to identify low-risk patients, such as the Glasgow-Blatchford score in gastrointestinal bleeding, perform better than scores asked to pick out high-risk patients [7].
  • Are its items and outcomes clearly defined? The TRIPOD statement sets out what a prediction model report should contain so that this can be judged [1].
  • Has its performance been tested in the subgroups you see? Wells rule strategies with fixed or age-adjusted D-dimer were examined separately in inpatients, patients with cancer and people aged 75 or over [8].
DecisionTool in common usePoint to watch
Anticoagulation in atrial fibrillationCHA2DS2-VASc or CHA2DS2-VAGuidelines differ on the version; name the one used [4]
Bleeding risk on anticoagulationHAS-BLED and similar scoresUsed to find and correct modifiable risks, not on their own to withhold anticoagulation [4]
Outpatient care after upper gastrointestinal bleedingGlasgow-Blatchford scoreBest at identifying very low-risk patients [7]
Suspected pulmonary embolismWells rule or revised Geneva score with D-dimerFewer patients are ruled out in some subgroups [8]
Drug dosing by kidney functionCockcroft-Gault or eGFRFollow the method the drug's labelling and local pharmacy use [12]

A prognostic score is not a screening test. In a meta-analysis, qSOFA predicted mortality in suspected infection with poor sensitivity [9], so a low qSOFA does not exclude sepsis.

Calculating without error

  1. Confirm the units of every input: creatinine in micromol/L or mg/dL, urea or blood urea nitrogen, glucose in mmol/L or mg/dL.
  2. Use each item's definition from the source publication or guideline, not just its label; terms such as vascular disease or labile INR have specific meanings.
  3. Use values from the right time point; most acute severity scores use the values at presentation.
  4. Mark missing inputs as missing, then work out the lowest and highest possible result and whether the range crosses a decision threshold.
  5. Recalculate a result that sits near a threshold, or confirm it in a validated calculator, and record the version used.

Formulas need the same discipline. With sodium 128, chloride 98 and bicarbonate 12 mmol/L, the anion gap is 128 minus 110, or 18 mmol/L. Correcting for an albumin of 28 g/L by adding 0.25 for each g/L below 40 raises it to about 21 mmol/L. With glucose 31 and urea 14 mmol/L, calculated osmolality is twice the sodium plus glucose plus urea, about 301 mmol/kg. Corrected for glucose with the traditional factor, the sodium is about 135 mmol/L.

The Cockcroft-Gault equation estimates creatinine clearance as (140 minus age) multiplied by body weight in kg, divided by 72 times the serum creatinine in mg/dL [10], with a further factor of 0.85 in women. For a 68-year-old man weighing 82 kg with a creatinine of 1.8 mg/dL, it gives about 46 mL/min with actual weight and about 37 mL/min with an ideal body weight of 66 kg for his height of 170 cm. A gap of that size can move a drug into a different dosing band.

Interpreting the number

Two properties matter. Discrimination is how well a score ranks patients who have the outcome above those who do not; calibration is how closely predicted risks match observed risks. A score can rank well yet overestimate or underestimate absolute risk in a population unlike its derivation cohort. CHADS2 and CHA2DS2-VASc, for example, offer only modest discrimination for individual patients [5], and decisions on thromboprophylaxis should rest on the absolute risks of stroke and bleeding and the net clinical benefit for the patient [6].

  • Features the score does not include, such as active bleeding, frailty, falls or pregnancy
  • A patient outside the derivation population, including children, extremes of body weight and unusual case mixes
  • The patient's own priorities when the benefit is small and the burden is real
  • A clinical picture that is evolving faster than a single score at one time point can capture

Kidney function: an estimate, not a measurement

Creatinine-based eGFR is imprecise in individual patients because creatinine also depends on muscle mass and other factors unrelated to filtration [11]. Combining creatinine with cystatin C improves the estimate, and the race-free CKD-EPI equation has performed less well than its predecessor in some European populations [11]. A US National Kidney Foundation workgroup has recommended race-free eGFR, adjusted to the patient's body surface area, in place of Cockcroft-Gault for medication decisions [12], while much drug labelling still refers to creatinine clearance.

All creatinine-based estimates assume a steady state. In acute kidney injury they overestimate function while creatinine is rising, and dosing decisions need pharmacy advice and repeated measurement. Local protocols take precedence over any single equation.

How Sana Health helps

Clinical Scores and Calculators restates the decision, chooses the score or formula validated for it and names the version, then sets out each input beside the patient's value with the points or substituted terms. It flags inputs that are missing instead of assuming them, gives the range of possible results, checks the arithmetic a second way, and interprets the band against the guideline that uses it. When a result sits near a threshold that would change management, the report says to confirm it in a validated calculator. Derivation and validation studies are listed as verified references.

Questions

What is the difference between discrimination and calibration?

Discrimination, usually reported as a c-statistic, describes how well a score separates patients who have the outcome from those who do not. Calibration describes whether the predicted risks match the risks actually observed. A score needs both, and calibration is the property most often lost when a score is used in a new population.

Should eGFR or Cockcroft-Gault be used for drug dosing?

Practice is changing. Many drug labels were written using Cockcroft-Gault creatinine clearance, while a US National Kidney Foundation workgroup now supports race-free eGFR adjusted to body surface area. Use the method the drug's labelling and the local pharmacy specify, and be cautious at extremes of body size and in acute kidney injury.

What should I do when a score falls right on a cut-off?

Check each input and its units, recalculate in a validated calculator, and consider which way the clinical features not captured by the score point. If the decision still turns on a single point, treat the result as uncertain and decide with the patient and, where appropriate, a senior colleague.

References

  1. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): the TRIPOD statement (opens in a new tab)Ann Intern Med, 2015
  2. Prediction models for cardiovascular disease risk in the general population: systematic review (opens in a new tab)BMJ, 2016
  3. Perforated peptic ulcer (opens in a new tab)Lancet, 2015
  4. 2024 ESC Guidelines for the management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS) (opens in a new tab)Eur Heart J, 2024
  5. Biomarkers in atrial fibrillation: a clinical review (opens in a new tab)Eur Heart J, 2013
  6. Stroke and bleeding risk assessment in atrial fibrillation: when, how, and why? (opens in a new tab)Eur Heart J, 2013
  7. Scoring systems for risk stratification in upper and lower gastrointestinal bleeding (opens in a new tab)Best Pract Res Clin Gastroenterol, 2023
  8. Wells Rule and d-Dimer Testing to Rule Out Pulmonary Embolism: A Systematic Review and Individual-Patient Data Meta-analysis (opens in a new tab)Ann Intern Med, 2016
  9. Prognostic Accuracy of the Quick Sequential Organ Failure Assessment for Mortality in Patients With Suspected Infection: A Systematic Review and Meta-analysis (opens in a new tab)Ann Intern Med, 2018
  10. Prediction of creatinine clearance from serum creatinine in patients with rheumatoid arthritis: comparison of six formulae and one nomogram (opens in a new tab)Clin Rheumatol, 2000
  11. Diagnostic standard: assessing glomerular filtration rate (opens in a new tab)Nephrol Dial Transplant, 2024
  12. Moving forward from Cockcroft-Gault creatinine clearance to race-free estimated glomerular filtration rate to improve medication-related decision-making in adults across healthcare settings: A consensus of the National Kidney Foundation Workgroup for Implementation of Race-Free eGFR-Based Medication-Related Decisions (opens in a new tab)Am J Health Syst Pharm, 2025