Point-of-care screening
AI devices that check for diabetic retinopathy at the point of care, or help primary care physicians decide whether to refer a skin lesion.
LumineticsCore and EyeArt autonomously check retinal images for diabetic retinopathy, and DermaSensor gives primary care physicians a result on a skin lesion to inform referral. The American Medical Association's AI taxonomy cites retinal imaging, Current Procedural Terminology code 92229, as its example of autonomous AI, which generates clinically meaningful conclusions without concurrent physician involvement. Medicare set payment for the code at $45 to $64 per exam from Jan. 1, 2022, according to a 2022 paper. DermaSensor was the first AI-enabled skin cancer device FDA authorized for non-specialists, a commentary noted.
A 2024 meta-analysis of 34 studies on real-world images found pooled sensitivity of 94% and specificity of 89% for AI diabetic retinopathy screening, among gradable images. In a test of seven anonymized algorithms from five companies, sensitivities ranged from 50.98% to 85.90%, most did no better than human graders, and the authors called for real-world testing of every algorithm before clinical use. Open questions include ungradable images in routine care and unnecessary referrals from low specificity.
DermaSensor
The handheld device sends brief pulses of light through a probe tip to the surface of a skin lesion and collects the light scattered back from beneath the surface, a technique called elastic scattering spectroscopy, then displays "Investigate Further" or "Monitor", according to the company.
Why this grade. DERM-SUCCESS, a blinded prospective accuracy study at 22 primary care sites in the U.S. and Australia (1,005 participants, 1,579 lesions) in which the device's readings did not guide care, found 95.5% sensitivity and 20.7% specificity for skin cancer; the evidence on referral decisions comes from reader studies.
The evidence
In DERM-SUCCESS, 30 primary care physicians scanned lesions they suspected of cancer, the device's readings did not guide care, and all lesions were biopsied; in patients 40 and older, sensitivity was 96.3% overall and 90.2% for melanoma, and across the study the negative predictive value was 96.6% and the positive predictive value 16.6%. Among patients 40 and older, the device returned "Monitor" for four melanomas and four keratinocyte carcinomas. In a prospective blinded study of 440 lesions at eight U.S. and two Australian sites, selected and biopsied by dermatologists, sensitivity for melanoma was 95.5% and specificity 32.5%. In a company-sponsored reader study of 108 primary care physicians and 100 lesion cases, the device's output raised referral sensitivity from 82.0% to 91.4% and lowered referral specificity from 44.2% to 32.4%. In a reader study of 118 internal medicine and family medicine physicians and 100 lesions, the area under the receiver operating characteristic curve for management decisions was 0.671 with the device versus 0.630 without, a significant increase.
- Used by
- The company said hundreds of physicians at health systems, and hundreds more in private practices, were using the device; it said more than 40,000 lesions had been scanned, MassDevice reported in September 2026.
- FDA
- FDA granted De Novo authorization DEN230008 on Jan. 12, 2024, creating a new device type: software-aided adjunctive diagnostic devices for use by physicians on lesions suspicious for skin cancer. The decision summary indicates it for lesions suggestive of melanoma, basal cell carcinoma or squamous cell carcinoma in patients aged 40 and older, to help decide on referral to a dermatologist, and not as the sole diagnostic criterion.
- Limits
- The company sponsored DERM-SUCCESS and one reader study and paid the other's first author; a commentary noted that 97.1% of DERM-SUCCESS patients were White and called on FDA to monitor real-world specificity and the risk of unnecessary care from false positives.
- Cost
- A company spokesperson told MDedge in January 2024 that pricing was by subscription: $199 a month for five patients or $399 a month for unlimited use.
EyeArt
EyeArt is autonomous AI that detects more-than-mild and vision-threatening diabetic retinopathy at the same time from retinal images, according to Eyenuk. The company says it is also certified in the European Union for autonomous detection of age-related macular degeneration and glaucomatous optic nerve damage.
Why this grade. EyeArt's studies measured accuracy, led by two prospective studies: a pivotal study at 15 U.S. sites (893 patients; 95.5% sensitivity and 85.0% specificity for more-than-mild retinopathy) and an evaluation of 30,405 English NHS screening episodes graded by EyeArt alongside human graders; none has measured its effect on care, and the San Ysidro Health randomized trial has no results yet.
The evidence
The San Ysidro Health trial (NCT06721351), which randomizes patients to point-of-care EyeArt screening or usual care, is listed as in progress with no results. In the pivotal study, sponsored by Eyenuk, EyeArt's results were compared with reading-center grades; sensitivity and specificity for vision-threatening retinopathy were 95.1% and 89.0% without dilation, and the system could grade 87.4% of eyes without dilation and 97.4% after protocol dilation. In the NHS evaluation, which ran EyeArt alongside human grading in three English screening programs, sensitivity for referable retinopathy was 95.7%, and specificity was 68% against a human grade of no retinopathy and 54.0% when non-referable retinopathy was included. An earlier study of 20,258 patients in English NHS screening, funded by the National Institute for Health Research, found 93.8% sensitivity for referable retinopathy and concluded that EyeArt and Retmarker were cost-effective alternatives to human grading. A retrospective study of 101,710 visits at 404 primary care clinics, with Eyenuk staff among the authors, found 91.3% sensitivity and 91.1% specificity.
- Used by
- Eyenuk said more than 500,000 patients had been screened in primary care, diabetes centers and specialty clinics worldwide. San Ysidro Health sponsors a randomized trial of point-of-care EyeArt screening.
- FDA
- FDA cleared EyeArt through 510(k) K200667 on Aug. 3, 2020, and version 2.2.0 through K223357 on June 16, 2023; that clearance added the Topcon NW400 camera to the Canon cameras already cleared, the company said. The macular degeneration and glaucoma claims rest on European certification, not FDA clearance.
- Limits
- The studies measured accuracy against human or reading-center grades, not what happened to patients when EyeArt's results were used in care, and Eyenuk sponsored the pivotal study.
LumineticsCore
LumineticsCore analyzes retinal images taken at the point of care and reports whether more-than-mild diabetic retinopathy is present, generating results without physician evaluation, according to Digital Diagnostics. It is intended for use by health care providers.
Why this grade. In the ACCESS randomized trial of 164 youth with diabetes at a Johns Hopkins pediatric diabetes center, diabetic eye exam completion within six months was 100% with point-of-care autonomous AI exams versus 22% with referral to an eye care provider; a cluster randomized trial at an eye clinic in Bangladesh found 40% higher clinic productivity with AI.
Caution. A report from a family medicine residency practice in Columbus, Ohio, whose authors reported no conflicts of interest, found 36% of exams were insufficient as of May 2023, when the practice began offering dilation to patients older than 64 or whose earlier undilated exams had failed, and 22% over the next eight months (Yun et al., Annals of Family Medicine, 2024); the pivotal trial reported imageability of 96.1%.
The evidence
ACCESS randomized 81 youth aged 8 to 21 to the AI exam and 83 to referral. The Bangladesh trial randomized 105 clinic days (993 patients), and specialists completed 1.59 care encounters an hour with AI versus 1.14 without.
In the company-funded pivotal trial at 10 U.S. primary care offices (900 enrolled), sensitivity for more-than-mild diabetic retinopathy was 87.2% and specificity 90.7%, and 23.6% of participants needed pupil dilation to get a result. A prospective study of 310 youth at a pediatric diabetes center (the SEE study) found screening adherence rose from 49% to 95% after autonomous AI was introduced; its abstract does not name the device, and Johns Hopkins Medicine reported the same rise after Johns Hopkins began using IDx-DR in clinics that treat children. In an external validation of IDx-DR 2.0 in the Hoorn Diabetes Care System in the Netherlands (1,415 people with type 2 diabetes), 63.5% had images of sufficient quality, and sensitivity for referable retinopathy was 91% or 68%, depending on the grading criteria.
- Used by
- Johns Hopkins uses the system in clinics that treat children with diabetes, Johns Hopkins Medicine reported in December 2023. Digital Diagnostics announced that it was available in nine Labcorp patient service centers in Alabama.
- FDA
- FDA granted De Novo authorization DEN180001 on April 11, 2018, to the system, then named IDx-DR, for detecting more-than-mild diabetic retinopathy and diabetic macular edema; the pivotal trial report called it the first FDA-authorized autonomous AI diagnostic system in any field of medicine. A later 510(k), K203629, was cleared on June 10, 2021.
- Limits
- The company, then IDx, funded the pivotal trial, and a company director co-authored both randomized trials, which measured exam completion at one U.S. pediatric center and clinic throughput at one eye clinic in Bangladesh, not vision outcomes. ACCESS used the system in youth aged 8 to 21, although the paper says it is not labeled for youth under 22.
Grades: A, randomized evidence of benefit; B, evidence from clinical use; C, accuracy studies only; D, little or no independent evidence. How the grades work. Reviews of the published evidence, not medical advice or an endorsement of any product.