Zeboulon 2020a.

Study characteristics
Patient Sampling	Retrospective, machine‐learning, experimental study. The Orbscan (Bausch & Lomb, Bridgewater, NJ) database was exported using the batch export functionality both as image files and the underlying numeric data matrixes represented by each colour map. They selected 3000 examinations in total, 1000 per class (normal, keratoconus, history of refractive surgery). All 3000 examinations were obtained from different people, and only 1 eye per person was selected. They balanced the examinations to have exactly 500 left eyes and 500 right eyes in each class. The selection process was as follows: consecutive examinations were preselected by a resident and reviewed by a corneal tomography expert.
Patient characteristics and setting	Keratoconus group: an anterior curvature map showing 1 of the classic keratoconus patterns described by Rabinowitz et al. associated with corneal thinning. Refractive surgery group: examinations were an oblate anterior surface (flat in its centre), a prolate posterior surface (steep in its centre), central corneal thinning, and lower central curvature values compared with the periphery (cases of myopic laser surgery). Normal group: examinations were preselected if no corneal condition could be detected.
Index tests	The possibility of using numeric data matrixes instead of colour maps to train a convolutional neural network (CNN) for a classification task. Specifically, the investigators used 4 maps that are frequently used in clinical practice, stacked together as if they were 4 colour channels of a single image to classify examinations into 3 categories: normal, keratoconus, and history of refractive surgery. The training set was trained during 15 epochs with a learning rate of 0.0001 and a batch size of 2.
Target condition and reference standard(s)	The diagnosis was made by a resident and corneal tomography specialist with at least 5 years of experience. The diagnosis was made before the convolutional neural network analysis.
Flow and timing	All participants were diagnosed by 2 cornea specialists. All cases were included in the analysis.
Comparative	Not applicable
Notes	The study authors received no funds or support for the study.
Methodological quality
Item	Authors' judgement	Risk of bias	Applicability concerns
DOMAIN 1: Patient selection
Was a consecutive or random sample of patients enrolled?	Yes
Was a case‐control design avoided?	No
Did the study avoid inappropriate exclusions?	No
Could the selection of patients have introduced bias?		High risk
Are there concerns that the included patients and setting do not match the review question?			High
DOMAIN 2: Index test (All tests)
Were the index test results interpreted without knowledge of the results of the reference standard?	Yes
If a threshold was used, was it pre‐specified?	Unclear
Was the model designed in an appropriate manner?	Unclear
Could the conduct or interpretation of the index test have introduced bias?		Low risk
Are there concerns that the index test, its conduct, or interpretation differ from the review question?			Low concern
DOMAIN 3: Reference standard
Is the reference standard likely to correctly classify the target condition?	Yes
Were the reference standard results interpreted without knowledge of the results of the index tests?	Yes
Could the reference standard, its conduct, or its interpretation have introduced bias?		Low risk
Are there concerns that the target condition as defined by the reference standard does not match the question?			Low concern
DOMAIN 4: Flow and timing
Did all patients receive the same reference standard?	Yes
Were all patients included in the analysis?	Yes
Could the patient flow have introduced bias?		Low risk
DOMAIN 5: Comparative