brianletort.ai
← Library

Johns Hopkins Children's Center / Johns Hopkins Medicine

Verified evidenceMoved to the visitQueue eliminationAutonomous + backstop

Autonomous AI diabetic eye exam at point of care (ACCESS RCT; LumineticsCore/IDx-DR)

Closing the diabetic eye-screening care gap requires persuading youth to attend a separate specialist appointment. Instead, the diagnostic exam itself is brought into the routine diabetes visit and read autonomously by AI in under a minute.

Healthcare screening · Baltimore, Maryland, USA (Johns Hopkins Pediatric Diabetes Center, two sites); JHM system deployment across adult primary care

Collections: Queue eliminated · Regulated autonomy · Embodied work

A family waiting for a separate eye appointment contrasts with retinal imaging performed during an existing clinic visit.

Executive brief

The operating-model shift, in one view.

The binding constraint on screening was never diagnostic capacity; it was the separate appointment. Moving an autonomous, minute-long diagnostic into a visit the patient already attends closed a care gap that referral-plus-education could not, and it did so without introducing racial, ethnic, or socioeconomic disparities.

AI value · Diabetic eye exam completion within 6 months (care-gap closure)

Verified

Eye-exam completion: 22% → 100% in the control arm

Peer-reviewed RCT under trial conditions. Outcome is process completion, not vision preserved. Control-arm denominator appears as n=83 in the abstract and 18/82 (Table 3, n=163) in the body; both round to 22%.

Before

Refers youth to an external eye-care provider with scripted education and a paper guide. → Schedule, travel to, and attend a separate eye appointment within 6 months → Perform dilated diabetic eye exam and return results

After

Captures two fundus images during the routine endocrinology visit without pharmacologic dilation. → Returns one of three results within 60 seconds: DED present, absent, or insufficient image quality. → Communicates the result and, if DED is present, gives scripted referral education.

Human boundary

The AI makes the screening diagnosis autonomously (its FDA De Novo authorization basis for adults is diagnosis without human oversight). In the youth trial, the AI output was the result communicated to the patient and drove the referral decision, with a retina-specialist overread of every image as a safety layer…

Why it matters

Diabetic eye screening can happen during the existing clinic visit.

How the work changed

Before

How the work ran before the change.

  1. Step 1 of 3

    Study coordinator / diabetes clinic staff

    Refers youth to an external eye-care provider with scripted education and a paper guide.

    ControlReferral documentation; completion tracked via EHR and phone follow-up

  2. Step 2 of 3

    Patient and family

    Schedule, travel to, and attend a separate eye appointment within 6 months

    ControlPatient initiative; 78% of the control arm never completed the exam in the window

  3. Step 3 of 3

    Eye care provider

    Perform dilated diabetic eye exam and return results

    ControlECP clinical judgment

What changed

Diabetic eye screening can happen during the existing clinic visit.

Decision rightAI diagnoses; specialists retain treatment and the safety backstop

After

How the same work runs now.

  1. Step 1 of 3

    Trained clinic operator (no ML expertise required)

    Captures two fundus images during the routine endocrinology visit without pharmacologic dilation.

    ControlAI image-quality algorithm guides acquisition and forces retakes (max 3 attempts)

  2. Step 2 of 3

    Autonomous AI diagnostic system (LumineticsCore)

    Returns one of three results within 60 seconds: DED present, absent, or insufficient image quality.

    ControlLocked deterministic medical device under FDA De Novo regulation; in this off-label youth deployment, all images were additionally overread by a board-certified retina specialist (estimated sensitivity 100%, specificity 78.9% vs level-4 reference standard)

  3. Step 3 of 3

    Clinic staff

    Communicates the result and, if DED is present, gives scripted referral education.

    ControlScripted educational intervention

Process model built from the published workflow evidence for Johns Hopkins Children's Center / Johns Hopkins Medicine. Every step, actor, and control appears in full below.
Every step, actor, and control

Exception path

Insufficient image quality after 3 attempts triggers referral for eye care; a 'DED present' output triggers scripted referral to an eye care provider for a dilated exam.

Work removed

  • Separate screening appointment scheduling and travel
  • Pharmacologic dilation (unnecessary in youth)
  • Referral coordination and chase-up for the screening itself

Decision authority

The AI makes the screening diagnosis autonomously (its FDA De Novo authorization basis for adults is diagnosis without human oversight). In the youth trial, the AI output was the result communicated to the patient and drove the referral decision, with a retina-specialist overread of every image as a safety layer because the device is not cleared for under-22s. Treatment decisions remain with eye care providers.

Before

  1. 01

    Study coordinator / diabetes clinic staff

    Refers youth to an external eye-care provider with scripted education and a paper guide.

    Control: Referral documentation; completion tracked via EHR and phone follow-up

  2. 02

    Patient and family

    Schedule, travel to, and attend a separate eye appointment within 6 months

    Control: Patient initiative; 78% of the control arm never completed the exam in the window

  3. 03

    Eye care provider

    Perform dilated diabetic eye exam and return results

    Control: ECP clinical judgment

After

  1. 01

    Trained clinic operator (no ML expertise required)

    Captures two fundus images during the routine endocrinology visit without pharmacologic dilation.

    Control: AI image-quality algorithm guides acquisition and forces retakes (max 3 attempts)

  2. 02

    Autonomous AI diagnostic system (LumineticsCore)

    Returns one of three results within 60 seconds: DED present, absent, or insufficient image quality.

    Control: Locked deterministic medical device under FDA De Novo regulation; in this off-label youth deployment, all images were additionally overread by a board-certified retina specialist (estimated sensitivity 100%, specificity 78.9% vs level-4 reference standard)

  3. 03

    Clinic staff

    Communicates the result and, if DED is present, gives scripted referral education.

    Control: Scripted educational intervention

Work that left the path

  • Separate screening appointment scheduling and travel
  • Pharmacologic dilation (unnecessary in youth)
  • Referral coordination and chase-up for the screening itself
  • Specialist time spent on screening-negative patients

Human role before

Clinic staff refer and educate; the diagnostic determination depends entirely on the patient completing an external specialist visit.

Human role after

Clinic staff operate the fundus camera at point of care and deliver results immediately; a retina specialist overreads all images as a safety backstop in the off-label pediatric use; eye care providers see only referred DED-positive patients for management and treatment.

AI roleAutonomous diagnostic determination of referable diabetic eye disease from fundus images at point of care, with no physician oversight at the moment of diagnosis; image-quality control and operator guidance included.

Outcomes

Diabetic eye exam completion within 6 months (care-gap closure)

Verified

22% (18/82) in the control arm (scripted ECP referral plus education)→100% (81/81) in the intervention arm; difference 78 percentage points (95% CI 69-87), p<0.001; no significant differences by race, ethnicity, SES, or education

6-month post-randomization window; enrolled 2021-11-24 to 2022-06-06, follow-up completed 2022-12-06 · 164 randomized youth aged 8-21 with type 1 or type 2 diabetes at a single academic pediatric diabetes center (two sites); 41% minority groups; 47% Medicaid

Peer-reviewed RCT under trial conditions. Outcome is process completion, not vision preserved. Control-arm denominator appears as n=83 in the abstract and 18/82 (Table 3, n=163) in the body; both round to 22%.

Follow-through with an eye care provider when indicated

Verified

22% in the control arm→64% (16/25) in the intervention arm (difference 42 points, 95% CI 21-63), p<0.001

6-month window · 25 intervention participants with a 'DED present' result

Denominators differ by design (intervention: DED-positive participants completing ECP follow-up; control: participants completing the ECP exam), so the comparison is not like-for-like; the paper defines the secondary outcome this way explicitly.

DED screening adherence in the routine pediatric deployment (pre-trial)

Verified

49% baseline adherence→95% after autonomous AI implementation; 310 exams in the first year; 85.7% sensitivity and 79.3% specificity vs level-2 reference

First year of pediatric deployment (2018 onward) · 310 youth with diabetes at the JHM multidisciplinary pediatric diabetes center

Pre/post observational, single center, reported in a peer-reviewed implementation review by the operating team.

What leaders can reuse

Anti-pattern

Citing the 100% vs 22% trial result as proof of operational performance; the trial proves efficacy under RCT conditions, and the separate routine-care deployment evidence is what proves durability.

Questions

  1. 01Which of our screening or compliance gaps are actually attendance problems that point-of-care automation could eliminate?
  2. 02Where would we accept an off-label AI deployment with a specialist overread backstop, and who signs that protocol?
  3. 03Screening completion is a process metric; what is our path to the outcome metric (here, vision preserved)?

Portability conditions

  • A recurring visit the population already attends (here, pediatric endocrinology)
  • An FDA-authorized autonomous diagnostic for the target population, or a managed off-label protocol with specialist overread
  • Trained non-physician operators and EMR integration (JHM reports roughly 6 months of start-up: contracting, Epic integration, workflow planning)
  • A clear referral protocol for positive results; follow-through, not screening, becomes the next bottleneck

Reputation risk

low

Evidence and authority

What the public record supports.

Current · updated

3 peer reviewed, 2 primary; publication outcomes are verified.

Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 0bbf6e47dd114b1f

Related transformations

More in Healthcare screening →

German organised mammography screening programme (PRAIM implementation study; 12 sites) · Autonomous + backstop

AI-supported double reading with normal triaging and a post-read safety net (Vara MG)

Two unaided human readers plus consensus are not the only safe model for a national screening programme.

MIT randomized professional-writing study · Autonomous + backstop

ChatGPT-assisted drafting with human submission

The redesign is a shorter drafting loop with human revision and accountability, not autonomous publication.

Capital Region of Denmark mammography screening programme · Exceptions only

AI-triaged single versus double reading of screening mammograms (Transpara)

That every screening mammogram must be independently double-read by two radiologists to keep cancer detection and recall rates acceptable.

The Permanente Medical Group · Creator to judge

Ambient AI draft-note generation during patient encounters

Ambient tools can draft clinical notes, while physicians remain the editors and signers.

Sources

Read the evidence, freshness, caveat, and version policy.