A Multi-center Study to Evaluate Performance of an Automated Device for the Detection of Diabetic Retinopathy

From January 2017 to July 2017, 900 participants were prospectively enrolled in this observational study at 10 primary care practice sites throughout the United States. The study was approved by the institutional review board for each site, and all participants provided written informed consent. The study, which was funded by IDx LLC, was designed by the authors with input from the U.S. Food and Drug Administration (FDA) on the endpoints, statistical testing, and study design. Emmes Corp, a contract research organization (CRO), provided overall project management, including data management and independent monitoring and auditing services for all sites. CCR, Inc., an Algorithm Integrity Provider (AIP), was contracted to lock the AI system, hold any intermediate and final results and images in escrow, and interdict access to these by the Sponsor, from prior to the start of the study until final data lock. Because the Sponsor was thus interdicted from access to the AI system, the AIP performed all necessary maintenance and servicing activities during the study as well as throughout closeout.

The target population was asymptomatic persons, ages of 22 and older, who had been diagnosed with diabetes and had not been previously diagnosed with DR. A diagnosis of diabetes was defined as meeting the criteria established by either the World Health Organization (WHO) or the American Diabetes Association (ADA); Hemoglobin A1c (HbA1c) ≥ 6.5% based on repeated assessments; Fasting Plasma Glucose (FPG) ≥ 126 mg/dL (7.0 mmol/L) based on repeated assessments; Oral Glucose Tolerance Test (OGTT) with two-hour plasma glucose (2-hr PG) ≥ 200 mg/dL (11.1 mmol/L) using the equivalent of an oral 75 g anhydrous glucose dose dissolved in water; or symptoms of hyperglycemia or hyperglycemic crisis with a random plasma glucose (RPG) ≥ 200 mg/dL (11.1 mmol/L).

To help enroll a sufficient number of mtmDR participants for the evaluation of sensitivity, a stepwise enrichment strategy, as indicated in the prespecified protocol, was utilized mid-study to recruit sufficient numbers of mtmDR participants. The enrichment strategy sought higher risk participants with elevated HbA1c (>9.0%) levels or elevated Fasting Plasma Glucose; this enrichment was independently activated by the statistician while always remaining masked to the AI system outputs and the ETDRS disease levels. To account for any unintentional spectrum bias in the no/mild population, the study pre-defined a specificity outcome parameter to correct for any potential spectrum bias resulting from this enrichment strategy as co-primary.

All primary care sites in the study identified one or more inhouse operator trainees to perform the AI system protocol. After installation of the equipment by the Sponsor at the site, but before any participant was recruited, AI system operator trainees had to attest that they had not previously performed ocular imaging. Also, before start of study recruitment at each site, AI system operator trainees underwent a one-time standardized 4 h training program. They were trained how to acquire images, how to improve image quality if the AI system gave an insufficient quality output, and how to put images for analysis into the AI system. No additional training was provided to any of the AI system operators for the duration of the study. Independently, FPRC certified expert photographers were identified in geographic locations close to each site by the CRO, and documented 4W-D FPRC certification was required before any participant was imaged. The CRO independently completed site initiation visits at each site to ensure each site met all the good clinical practice requirements prior to start of enrollment.

Study protocol. All participants gave written informed consent to participate in both the AI system protocol, as well as the FPRC imaging protocol, using two different cameras:

The AI system protocol consisted of the following steps:

1. Operator takes images with a nonmydriatic retinal camera (NW400, Topcon Medical Systems, Oakland, NJ) according to a standardized imaging protocol with one disc and one fovea centered 45° image per eye;
2. Operator submits images to the AI system for automated image quality and protocol adherence evaluation;
3. If the AI system outputs insufficient quality, steps 1-2 are repeated until sufficient quality is output or 3 attempts were made. If the AI system still indicates that images are of insufficient quality, the participant's pupils are dilated with tropicamide 1.0% eyedrops, (provided by the Sponsor at each site), until the pupil diameter is at least 5mm in each eye or 30 minutes have passed, and steps 1-2 are repeated until sufficient quality is output or 3 attempts were made. If the AI system still outputs that images are of insufficient quality, the AI system output of insufficient quality is automatically provided to the CRO via secure data transfer;
4. Whenever the AI system indicates sufficient quality, the AI system disease level output (either mtmDR detected or mtmDR not detected) is automatically provided to the CRO via secure data transfer; the final AI system output provided to the CRO after this protocol was mtmDR detected, mtmDR not detected or insufficient quality

The FPRC imaging protocol was then conducted, and consisted of the following steps, all performed by an FPRC certified photographer:

1. If participant is not already dilated, dilating eye drops of tropicamide 1.0% are administered;
2. Digital widefield stereoscopic fundus photography is performed, using a camera capable of widefield photography, (Maestro, Topcon Medical Systems, Oakland, NJ) according to the FPRC 4W-D stereo protocol, by an FPRC certified photographer;
3. Anterior segment photography for media opacity assessment is performed according to the Age Related Eye Disease Study, by an FPRC certified photographer;
4. OCT of the macula is performed using a standard OCT system capable of producing a cube scan containing at least 121 B scans, (Maestro, Topcon Medical Systems, Oakland, NJ) according to the FPRC OCT protocol, by an FPRC certified photographer.

The FPRC certified photographers were masked to the AI system outputs at all times.

Reference standards. The FPRC grading protocol consisted of determination of ETDRS Severity Scale (SS) levels for fundus photographs and standardized OCT grading, as follows: the 4W-D images were read by three experienced and validated readers at the FPRC according to the well established ETDRS SS, using a majority voting paradigm. The macular OCT images were evaluated for the presence of center-involved DME by experienced readers at the FPRC according to the DRCR grading paradigm. For each participant, the ETDRS levels were mapped to mtmDR+(ETDRS level 35 or higher and /or DME present), or mtmDR- (ETDRS level 10-20 and DME absent), taking the worst of two eyes to correspond to the outputs of the AI system at the participant level.16 To measure sensitivity for the cases requiring immediate followup, called vision threatening DR, we defined vtDR + as ETDRS level 53 or higher, and/ or DME present, See Supplemental Table 2 for the mapping from ETDRS and DME levels to dichotomous mtmDR- and mtmDR +and vtDR+. Because DME can be identified both on the basis of retinal thickening on stereoscopic fundus photographs, as well as on the basis of retinal thickening on OCT, we separately analyzed both. Stereoscopic fundus based Clinically Significant DME (CSDME) was identified if there was either retinal thickening or adjacent hard exudates < 600 μm from the foveal center, or a zone of retinal thickening > 1 disc area, part of which is less than 1 disc diameter from the foveal center, according to the FPRC, in any eye. OCT based center-involved DME was identified if a participant had central subfield (a 1.0mm circle centered on the fovea) thickness that was >300 μm according to the FPRC, in any eye.20 Accordingly, we further specify the definition of mtmDR where relevant:

fundus mtmDR+ is defined as

• ETDRS level ≥ 35 (determined from fundus photographs)and/or
• CSDME (determined from fundus photographs)

and multimodal mtmDR+is defined as:

• ETDRS level ≥ 35 (determined from fundus photographs), and / or
• CSDME (determined from fundus photographs) and / or
• center-involved DME (determined from OCT).

and similarly for vtDR + . FPRC readers were masked to the AI system outputs at all times, masked to the fundus photograph reading when evaluating the OCT images, and masked to OCT readings when evaluating fundus photographs.