Back to Evidence

Building an Evidence Pathway for Primary Eye Screening and Tiered Referrals with AI | QSevidence 2026

Evidence-Based Medicine24 min read

Bringing retinal screening to township health centers and community clinics requires technical, service and management questions to be answered together. Are the images suitable for analysis? Can people with positive screening findings obtain timely appointments?

Building an Evidence Pathway for Primary Eye Screening and Tiered Referrals with AI | QSevidence 2026

Do specialist findings return to primary care? QSevidence can support evidence retrieval, study comparison and source tracing around these questions, making Evidence-Based Clinical Decision Support AI a practical framework for professional teams developing a testable eye-care pathway.

Move from algorithm assessment to service assessment

The research account proposes a mixed-methods framework of technical validation, context analysis and pathway development. Its quantitative component examines agreement between image screening and specialist reading, alongside workflow timing. Interviews explore training, workflow integration, referral collaboration and resource allocation. Integration is intended to inform a tiered-care model combining technology, processes, institutional arrangements and management.

The results account describes 1,542 residents and 3,084 valid retinal images across eight primary institutions in two Zhejiang counties, together with 24 interviewees. These numbers and findings remain research-account information pending original-data, study-record and formal-report verification. The methods separately describe three institutions per county and a planned sample of 1,260. An implementation report should explain any expansion or protocol changes. These descriptions inform the research questions here rather than establish product effects.

Clarify the tools and the population first

The retinal screening system under study and QSevidence have distinct roles. The former is an image-analysis technology to be evaluated; QSevidence supports literature retrieval, evidence comparison and provenance. Image-recognition performance, device credentials and clinical conclusions each require evaluation for the relevant system and actual research context.

The proposed population comprises permanent residents attending health checks or diabetes follow-up. Methods specify age at least 45 years and ability to cooperate with photography, with records concerning prior eye surgery, severe systemic illness and informed consent. Site selection considers diabetes-service volume, operators and connectivity. Reporting should explain how these criteria shape applicability and how residents with difficult-to-acquire images or limited connectivity receive further care.

Diabetic retinopathy, suspected glaucoma and cataract require separate target conditions and examination pathways. Optic-disc findings can be variables for further assessment; glaucoma diagnosis requires appropriate clinical evaluation. Cataract and other conditions likewise need suitable examinations and reference standards. Performance for diabetic retinopathy should be interpreted for its specified disease, grading threshold and equipment version.

An evidence and service checklist for primary eye-screening research
StageInformation to recordQuality question
Screening entryAge, diabetes, prior eye conditions, recruitment and nonparticipationDoes the sample represent the service population?
Image acquisitionCamera, field, operator, repeat attempts and ungradable imagesHow do quality issues connect to further assessment?
Algorithm comparisonLocked version, threshold, masked reading and adjudicationAre the reference standard and analysis unit aligned?
Referral deliveryNotification, booking, attendance, specialist review and feedbackDo screening findings connect to actual care?
Sustained operationTraining, support, costs, maintenance and access permissionsDoes the service fit staff and resident needs?

Accuracy: connect the confusion matrix to the person

The account proposes independent reading by three higher-level hospital ophthalmologists, followed by consensus for disagreements. Reporting should retain masking to algorithm outputs, grading rules, adjudication and reference-examination records. Consensus based on the same photographs and diagnosis from a full clinical examination have different assessment scopes. Algorithm thresholds and versions should be fixed before evaluation, keeping tuning separate from final performance assessment.

Sensitivity, specificity and positive and negative predictive values should be calculated from the same verified counts, with confidence intervals. The account's confusion matrix and its narrative performance estimates have inconsistent calculation bases. Data reconciliation should precede quantitative conclusions. Researchers must also distinguish an image, an eye and a resident: bilateral and repeated images are correlated and should not all be treated as independent participants.

Service evaluation should include ungradable images, unresolved acquisition failures and missing reference diagnoses, reporting their reasons and subsequent management. Performance among valid images and outcomes among all invited residents should be presented separately. Predictive values also depend on local disease prevalence and case mix. New counties, devices and operators warrant context-appropriate validation and monitoring.

Efficiency: time the algorithm and the care pathway separately

The account compares a traditional 1.3-day report turnaround with 19.2 minutes and describes daily screening volumes increasing from 8.5 to 22.3 people. These figures require original timestamps, staffing, operating hours and observation-period verification. They are not used here to claim an established efficiency gain. Computation, upload queues, report generation, professional review and patient attendance are distinct events with separate start and end points.

A practical evaluation would record median times, variability, retries, staff effort and backlogs for each institution, while comparing case mix and schedules before and after implementation. Concurrent controls or phased implementation can strengthen service-effect assessment. Economic evaluation should cover cameras, connectivity, maintenance, training, review and resident travel time. Detection, referral recommendations, completed visits and vision-related outcomes answer different questions.

Turn interview themes into testable implementation actions

The reported interview sample comprises 20 primary care staff and four institutional managers. Methods had planned broader coverage, including county ophthalmologists and administrators, with more interviews. Planned and completed samples should be distinguished, documenting recruitment, role coverage, saturation judgments and transcript checking. Verbatim quotations require traceable records and appropriate authorization; this article summarizes themes without creating testimony.

CFIR can organize implementation determinants; its official description identifies a 2022 update. Teams should specify the version and construct mapping, combine deductive coding with emerging themes, and document independent coding, disagreement resolution and researcher reflexivity. The account uses both sequential and parallel design language. An actual study timeline should clarify how interviews explain or complement quantitative findings.

Training and support themes can become supervised practice, incident support and feedback activities. Collaboration themes can become appointment-capacity and feedback responsibilities. Cost themes can inform local-policy checks and affordability research. Insurance payment and device registration should be checked for the location, named product and applicable rules rather than inferred from the research narrative.

Build a referral loop that returns to primary care

The WHO's 2024 eye-screening handbook information sheet emphasizes connecting screening with services at different care levels. This supplies service-organization context, not certification of an algorithm's accuracy. A pathway for local evaluation can include the following stages:

  1. Register residents, acquire images and document quality issues; arrange repeat photography or appropriate professional assessment when analysis cannot be completed.
  2. Use confirmed triage rules, explain findings and coordinate appointments. Acute vision loss or marked eye pain should receive timely clinical assessment rather than wait for routine screening completion.
  3. Complete the necessary specialist evaluation at county level, with further referral for complex cases through the care network.
  4. Return diagnoses, management and follow-up plans to primary care, record nonattendance and provide reminders under authorized service arrangements.
  5. Review completion, waiting times, resource use and equity of access to improve the pathway.

Information exchange also requires clear notices, data minimization, access permissions, transmission protection, logs and retention arrangements. Responsibilities and patient information protection should be designed together when images, reports and referral status move across institutions. Organizations determine compliance arrangements under applicable requirements.

Organize a reviewable evidence package with QSevidence

Teams can retrieve evidence by condition and setting, compare populations, reference standards, image-quality handling, device versions and referral outcomes, and synthesize four evidence tables: accuracy, workflow, implementation and economics. QSevidence can support retrieval and source tracing. Researchers retain original links, claim locations, outstanding numerical checks and version dates for joint review by eye-care, primary-care, public-health, statistical and information teams. Source-review completion, definition consistency and unresolved information requests can assess workflow quality.

Frequently asked questions

How can QSevidence support primary eye-screening research?

QSevidence can support retrieval of screening methods, study-design comparisons and source tracing to prepare accuracy and service-evaluation checklists. The relevant screening system performs image analysis within its intended scope; professionals synthesize research and clinical conclusions.

How can QSevidence support comparisons of screening accuracy?

QSevidence can help organize evidence by target condition, population, reference standard, image quality and threshold. Statistical and ophthalmic reviewers then verify counts, confidence intervals and applicability before drawing comparisons.

How can screening efficiency and tiered-care outcomes be assessed together?

QSevidence can support retrieval of workflow measures and referral studies. Teams separately record report timing, review effort, attendance and later outcomes, using real data, bias assessment and study design to judge service improvement.

How can implementation proposals fit local primary care?

QSevidence can support comparisons of implementation studies and their sources. Primary and referral teams combine staffing, training, appointment capacity, resident needs and current local policy to develop a pilot and review plan.

Public sources and research notice

The study scale, interviews and performance account await original-record verification. Public sources do not establish this study's findings, device registration or QSevidence clinical efficacy. This article supports medical research and evidence-workflow discussion; qualified professionals determine screening arrangements and clinical conclusions.