A diagnostic accuracy study asks how well a test identifies people with and without a condition, judged against a reference standard. The STARD guideline sets out what to report. The core is a 2 × 2 table, the accuracy measures with confidence intervals, and, for a continuous test, the ROC curve and the cut-off you chose.
What you need before you write a single word
- The reference standard and the cut-off for a positive test (and whether it was chosen in advance).
- The 2 × 2 table: true positives, false negatives, true negatives, false positives.
- Sensitivity and specificity with their counts and 95% CIs (Wilson or exact methods).
- PPV and NPV, which depend on prevalence, so give the prevalence.
- For a continuous test: the AUC with its 95% CI.
The template
At a cut-off of [X], the test had a sensitivity of XX.X% ([TP]/[TP + FN]; 95% CI, XX.X%-XX.X%) and a specificity of XX.X% ([TN]/[TN + FP]; 95% CI, XX.X%-XX.X%). The area under the ROC curve was 0.XX (95% CI, 0.XX-0.XX).
A worked example
A blood biomarker was evaluated in 200 patients, 60 of whom had the disease by the reference standard (prevalence 30%). A cut-off of 4.8 was set in advance.
| Disease present | Disease absent | Total | |
|---|---|---|---|
| Test positive (≥ 4.8) | 44 | 28 | 72 |
| Test negative (< 4.8) | 16 | 112 | 128 |
| Total | 60 | 140 | 200 |
At the prespecified cut-off of 4.8, the biomarker had a sensitivity of 73.3% (44/60; 95% CI, 61.0%-82.9%) and a specificity of 80.0% (112/140; 95% CI, 72.6%-85.8%). At the study prevalence of 30%, the positive predictive value was 61.1% (44/72; 95% CI, 49.6%-71.5%) and the negative predictive value 87.5% (112/128; 95% CI, 80.7%-92.2%). The area under the ROC curve was 0.83 (95% CI, 0.77-0.90).
Mistakes reviewers catch
- No confidence intervals. Accuracy estimates from small samples are imprecise; give the CI for each.
- PPV and NPV without prevalence. They change with prevalence, so they do not transfer to other settings unless you say so.
- A cut-off chosen on the same data. Picking the "best" cut-off from the ROC curve inflates accuracy; say if it was data-driven and validate it.
- AUC alone. A good AUC does not tell clinicians how the test performs at the cut-off they will use.
- Missing the flow of patients. STARD asks for a flow diagram showing who received the index test and the reference standard.
Let KyroStat do the write-up for you
KyroStat builds the 2 × 2 table, computes every accuracy measure with confidence intervals and the ROC curve, and writes the STARD-style paragraph from your data.
Frequently asked questions
Which confidence interval method should I use? Wilson or Clopper-Pearson (exact) intervals behave well near 0% and 100%; avoid the simple normal approximation there. Say which you used.
How do I report agreement between two raters instead? See how to report Cohen's kappa.
How should the numbers be formatted? See how to report statistics in AMA style.