Performance of AI-based diabetic retinopathy screening is highly dependent on evaluation setting: A five-year, multi-framework study

Performance of AI-based diabetic retinopathy screening is highly dependent on evaluation setting: A five-year, multi-framework study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Performance of AI-based diabetic retinopathy screening is highly dependent on evaluation setting: A five-year, multi-framework study Gwenolé Quellec, Mathieu Lamard, Sarah Matta, Laurent Borderie, and 10 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9427241/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Despite near-perfect performance reported on benchmark datasets, the real-world behavior of artificial intelligence (AI) systems for diabetic retinopathy (DR) screening remains insufficiently characterized. Here, we report a five-year, multi-framework evaluation of a CE-marked AI system for automated detection of referable DR and diabetic macular edema from color fundus photographs. The system was initially validated on the Messidor-2 benchmark dataset and subsequently assessed across three independent evaluation settings: a large-scale masked comparative study (US Veterans Affairs), an external validation using handheld fundus photography in Finland, and an open, large-scale comparative evaluation within the UK National Health Service. Across these evaluation frameworks, substantial variability in observed sensitivity and specificity was found. Differences in imaging devices, population characteristics, referral definitions, and handling of ungradable images contributed to these variations. In masked settings, interpretation of comparative performance was limited by the lack of system-level attribution, while open evaluations remained sensitive to protocol design. Taken together, these results show that performance in AI-based DR screening is not a fixed property of a system, but an emergent property of the evaluation framework. This work calls for context-aware validation strategies and more standardized evaluation protocols for clinical AI systems. Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Health care Physical sciences/Mathematics and computing Health sciences/Medical research Diabetic retinopathy screening artificial intelligence real-world evaluation domain shift external validation performance variability medical imaging screening programs Full Text Additional Declarations Competing interest reported. L.B., P.O.S., P.D. and A.L.G. are employees of Evolucare Technologies. P.H. is affiliated with Optomed and holds equity in the company. G.Q. and M.L. report that a software license related to the evaluated system has been granted to Evolucare Technologies. G.Q. has also served as a consultant for Evolucare Technologies. The remaining authors declare no competing interests. Supplementary Files SupplementaryMaterial.pdf Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 18 May, 2026 Reviewers agreed at journal 12 May, 2026 Reviewers agreed at journal 05 May, 2026 Reviews received at journal 04 May, 2026 Reviewers agreed at journal 04 May, 2026 Reviewers agreed at journal 04 May, 2026 Reviewers invited by journal 04 May, 2026 Editor assigned by journal 18 Apr, 2026 Submission checks completed at journal 17 Apr, 2026 First submitted to journal 15 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9427241","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":637432727,"identity":"b69c062f-da8c-4e35-a704-9a0e8bc41034","order_by":0,"name":"Gwenolé Quellec","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDklEQVRIie2QMUvEMBiGv1sylXNN6XH5CymFuhz9HY4pgWbJcY4nHNJb6iLOhfsfnT84OJe6d2wR6loXwUWNeh0ccjgK5hmSkHwP70sAHI4/y+XXOsEWYE6Od2enFf69oQCIRsXPf6uk46RVYRfy0A58AdNGIop1ou6Cm+7xSgM9tzhhk6mw5Bn4TSZQ1HJZzOoofKiAztCi7HQceHwPvNEc0wKXBdXE31ZwTS3Fwt3qxSjvRlkNmL6hIlQ9vRqF2hQWaGIU/EwBTHMUhIp4ckrhrI/9kkvPr3uO4iBDUywyxag95Xbf02GdzKf3smuHTcJYqbrnbbWwpxy/xQP8+WATTEo+ntA643A4HP+dD6xkWbOQxVfaAAAAAElFTkSuQmCC","orcid":"","institution":"Inserm","correspondingAuthor":true,"prefix":"","firstName":"Gwenolé","middleName":"","lastName":"Quellec","suffix":""},{"id":637432728,"identity":"8d941d17-5ae2-4883-ba0f-c7d6007fdb87","order_by":1,"name":"Mathieu Lamard","email":"","orcid":"","institution":"University of Western Brittany","correspondingAuthor":false,"prefix":"","firstName":"Mathieu","middleName":"","lastName":"Lamard","suffix":""},{"id":637432729,"identity":"bf3da3a9-dc43-4ea2-b663-cbb819c4b8e7","order_by":2,"name":"Sarah Matta","email":"","orcid":"","institution":"University of Western Brittany","correspondingAuthor":false,"prefix":"","firstName":"Sarah","middleName":"","lastName":"Matta","suffix":""},{"id":637432730,"identity":"91d76fcf-1a82-48a5-aa43-538f04350f80","order_by":3,"name":"Laurent Borderie","email":"","orcid":"","institution":"Evolucare Technologies","correspondingAuthor":false,"prefix":"","firstName":"Laurent","middleName":"","lastName":"Borderie","suffix":""},{"id":637432731,"identity":"e45804df-4c7e-4bd5-8b04-b3d068c975bb","order_by":4,"name":"Philippe Zhang","email":"","orcid":"","institution":"University of Western Brittany","correspondingAuthor":false,"prefix":"","firstName":"Philippe","middleName":"","lastName":"Zhang","suffix":""},{"id":637432732,"identity":"faf45f37-a1cd-43fb-a060-fd9f7416c8f5","order_by":5,"name":"Paola Ortega Saborio","email":"","orcid":"","institution":"Evolucare Technologies","correspondingAuthor":false,"prefix":"","firstName":"Paola","middleName":"Ortega","lastName":"Saborio","suffix":""},{"id":637432733,"identity":"c6a53fd0-d1fe-41af-8637-3b40492aff74","order_by":6,"name":"Pierre Deman","email":"","orcid":"","institution":"Evolucare Technologies","correspondingAuthor":false,"prefix":"","firstName":"Pierre","middleName":"","lastName":"Deman","suffix":""},{"id":637432734,"identity":"21789aac-f85a-4b67-acc0-bcc8ec5ba14f","order_by":7,"name":"Alexandre Le Guilcher","email":"","orcid":"","institution":"Evolucare Technologies","correspondingAuthor":false,"prefix":"","firstName":"Alexandre","middleName":"Le","lastName":"Guilcher","suffix":""},{"id":637432735,"identity":"25bf97e0-2e61-4530-81ad-e5aa44f74b22","order_by":8,"name":"Ali Erginay","email":"","orcid":"","institution":"Assistance Publique - Hôpitaux de Paris","correspondingAuthor":false,"prefix":"","firstName":"Ali","middleName":"","lastName":"Erginay","suffix":""},{"id":637432736,"identity":"4aca21a3-7b2a-4104-8a17-962c170d28f5","order_by":9,"name":"Anna-Maria Kubin","email":"","orcid":"","institution":"Oulu University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Anna-Maria","middleName":"","lastName":"Kubin","suffix":""},{"id":637432738,"identity":"6e49241c-b8f1-4986-8b02-98ac25119b98","order_by":10,"name":"Nina Hautala","email":"","orcid":"","institution":"Oulu University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Nina","middleName":"","lastName":"Hautala","suffix":""},{"id":637432740,"identity":"887fbb5c-13b7-4cba-8ea2-14c0e931352d","order_by":11,"name":"Petri Huhtinen","email":"","orcid":"","institution":"Optomed","correspondingAuthor":false,"prefix":"","firstName":"Petri","middleName":"","lastName":"Huhtinen","suffix":""},{"id":637432741,"identity":"20ffc590-a940-4432-9a04-c74f22606a93","order_by":12,"name":"Béatrice Cochener","email":"","orcid":"","institution":"Centre Hospitalier Régional Universitaire de Brest","correspondingAuthor":false,"prefix":"","firstName":"Béatrice","middleName":"","lastName":"Cochener","suffix":""},{"id":637432742,"identity":"fec428e9-f3a3-4e72-87ec-31c6d43531db","order_by":13,"name":"Pascale Massin","email":"","orcid":"","institution":"Assistance Publique - Hôpitaux de Paris","correspondingAuthor":false,"prefix":"","firstName":"Pascale","middleName":"","lastName":"Massin","suffix":""}],"badges":[],"createdAt":"2026-04-15 12:53:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9427241/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9427241/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":109068278,"identity":"f583dd77-c632-451b-aecc-94289013974f","added_by":"auto","created_at":"2026-05-12 10:05:14","extension":"pdf","order_by":1,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":300165,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9427241/v1_covered_243e36ca-0953-4b89-bbdb-a7cecd45a562.pdf"},{"id":109046916,"identity":"68e84ff2-e599-4d12-a5ac-516094c953b5","added_by":"auto","created_at":"2026-05-12 05:48:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":156442,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterial.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9427241/v1/1119d3e3b4f4ff1935cbedf7.pdf"}],"financialInterests":"Competing interest reported. L.B., P.O.S., P.D. and A.L.G. are employees of Evolucare Technologies.\nP.H. is affiliated with Optomed and holds equity in the company.\nG.Q. and M.L. report that a software license related to the evaluated system has been granted to Evolucare Technologies.\nG.Q. has also served as a consultant for Evolucare Technologies.\nThe remaining authors declare no competing interests.","formattedTitle":"Performance of AI-based diabetic retinopathy screening is highly dependent on evaluation setting: A five-year, multi-framework study","fulltext":[],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":false,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":true,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":true,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Diabetic retinopathy screening, artificial intelligence, real-world evaluation, domain shift, external validation, performance variability, medical imaging, screening programs","lastPublishedDoi":"10.21203/rs.3.rs-9427241/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9427241/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDespite near-perfect performance reported on benchmark datasets, the real-world behavior of artificial intelligence (AI) systems for diabetic retinopathy (DR) screening remains insufficiently characterized. Here, we report a five-year, multi-framework evaluation of a CE-marked AI system for automated detection of referable DR and diabetic macular edema from color fundus photographs. The system was initially validated on the Messidor-2 benchmark dataset and subsequently assessed across three independent evaluation settings: a large-scale masked comparative study (US Veterans Affairs), an external validation using handheld fundus photography in Finland, and an open, large-scale comparative evaluation within the UK National Health Service.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAcross these evaluation frameworks, substantial variability in observed sensitivity and specificity was found. Differences in imaging devices, population characteristics, referral definitions, and handling of ungradable images contributed to these variations. In masked settings, interpretation of comparative performance was limited by the lack of system-level attribution, while open evaluations remained sensitive to protocol design.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTaken together, these results show that performance in AI-based DR screening is not a fixed property of a system, but an emergent property of the evaluation framework. This work calls for context-aware validation strategies and more standardized evaluation protocols for clinical AI systems.\u003c/p\u003e","manuscriptTitle":"Performance of AI-based diabetic retinopathy screening is highly dependent on evaluation setting: A five-year, multi-framework study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-12 05:48:38","doi":"10.21203/rs.3.rs-9427241/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-18T17:57:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"47714240657042478628263821900122543726","date":"2026-05-12T15:09:24+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"263577038218226643544918215357914884499","date":"2026-05-05T16:08:14+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-05T00:44:56+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"290287316574097829352557461676094207648","date":"2026-05-04T23:48:59+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"84087269219545168369350996556236468158","date":"2026-05-04T23:31:24+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-05-04T21:57:53+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-18T15:44:44+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-18T02:43:59+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Digital Medicine","date":"2026-04-15T12:37:03+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-digital-medicine","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"npjdigitalmed","sideBox":"Learn more about [npj Digital Medicine](http://www.nature.com/npjdigitalmed/)","snPcode":"41746","submissionUrl":"https://submission.springernature.com/new-submission/41746/3","title":"npj Digital Medicine","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b3f072c2-f9c2-4874-9f43-de834645ef86","owner":[],"postedDate":"May 12th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-18T17:57:31+00:00","index":37,"fulltext":""},{"type":"reviewerAgreed","content":"47714240657042478628263821900122543726","date":"2026-05-12T15:09:24+00:00","index":36,"fulltext":""},{"type":"reviewerAgreed","content":"263577038218226643544918215357914884499","date":"2026-05-05T16:08:14+00:00","index":34,"fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-05-05T00:44:56+00:00","index":33,"fulltext":""},{"type":"reviewerAgreed","content":"290287316574097829352557461676094207648","date":"2026-05-04T23:48:59+00:00","index":32,"fulltext":""},{"type":"reviewerAgreed","content":"84087269219545168369350996556236468158","date":"2026-05-04T23:31:24+00:00","index":31,"fulltext":""},{"type":"reviewersInvited","content":"25","date":"2026-05-04T21:57:53+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67848753,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":67848754,"name":"Health sciences/Diseases"},{"id":67848755,"name":"Health sciences/Health care"},{"id":67848756,"name":"Physical sciences/Mathematics and computing"},{"id":67848757,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-05-12T05:48:38+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-12 05:48:38","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9427241","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9427241","identity":"rs-9427241","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}