STARD Checklist

What STARD is

STARD is the reporting guideline for diagnostic accuracy studies: studies that compare the results of one or more index tests against a reference standard in the same participants. The current version, STARD 2015, has a 30-item checklist and a participant flow diagram, plus a separate checklist for abstracts. Journals that publish diagnostic research generally expect it, and many ask for the completed checklist at submission.

When it applies

Use STARD whenever your main outcome is how well a test identifies a condition, reported as sensitivity, specificity, predictive values, likelihood ratios or area under the ROC curve. That covers imaging, laboratory and clinical tests, and increasingly the evaluation of AI and machine learning models against a reference standard (a dedicated extension for AI-based tests is in development).

What the checklist covers

• Title and abstract: identified as a study of diagnostic accuracy, with at least one accuracy measure named.
• Introduction: the scientific and clinical background, including the intended use and clinical role of the index test, and the study hypotheses.
• Methods: whether data were collected prospectively or retrospectively, eligibility criteria and how participants were identified, the setting and dates, the index test and reference standard in enough detail to replicate, the rationale for the reference standard, whether readers were masked to other results, the analysis methods, how indeterminate results and missing data were handled, and the intended sample size.
• Results: the flow of participants (the STARD diagram), baseline characteristics, the distribution of disease severity, the time between index test and reference standard, a cross tabulation of index test against reference standard results, accuracy estimates with confidence intervals, and any adverse events from testing.
• Discussion: limitations, including sources of bias and uncertainty, and implications for practice.

What authors most often miss

• The 2 x 2 table (or its equivalent) cross tabulating index test and reference standard results.
• Masking: no statement of whether the people reading the index test knew the reference standard result, or the reverse.
• Indeterminate or missing results quietly excluded, which inflates accuracy.
• Spectrum of participants: a case-control sampling of clear-cut cases and healthy controls presented as if it reflected clinical use.
• Thresholds chosen after looking at the data without saying so.

Ophthalmology: imaging, AI and fellow eyes

Diagnostic studies in ophthalmology often evaluate imaging devices or AI models on fundus photographs and OCT scans, frequently including both eyes per patient. Report whether accuracy is per eye or per patient, how you accounted for clustering, and where the training and test images came from when evaluating an algorithm.