High quality retinal images were key to this pilot study’s success and formed the basis for testing an AI algorithm.
A large global biopharmaceutical company conducted a pilot study to establish a standardized and extensible procedure for retinal image acquisition in conjunction with laboratory assessments and blood pressure measurements. Once established, the methodology was to be implemented in a larger study to support the development of a machine-learning algorithm for prediction of vital sign measurements, laboratory findings, and diabetic and cardiovascular disease states based on digital retinal images.
The study was conducted at 2 sites in Africa with a planned enrollment of 300 subjects. The methodology was developed based on images from a large-scale biomedical database in Europe. Thus, a key study goal was to confirm if the algorithm would have a comparable level of accuracy in an African population. It was also planned to test the feasibility of obtaining analyzable retinal images in a more limited healthcare facility setting.
Innovations in artificial intelligence (AI) algorithms may increase access to healthcare and improve diagnostics in settings with limited healthcare infrastructure.
Conducting a pilot study with sites that are less familiar with digital retinal imaging can bring forth many challenges for Sponsors. This study presented the following potential issues:
- Camera learning curve. While the camera sourced for the study required very little technical knowledge, some issues with camera and subject alignment adjustments were expected
- Potential variability in retinal image collection. There are many potential sources of variability in fundus photographs including small pupils, need for exposure compensation for dark or light retinas, poor subject or camera alignment, and lash artifacts
- Image capture training. Extensive training on acquiring retinal images and determining if they were adequate for reading was needed
MERIT’s EXCELSIORTM platform was employed by experienced experts to help the Sponsor flawlessly manage the retinal imaging for the study.
- Thorough study startup documents. A study-specific Imaging Procedure Manual and an Imaging and Grading Charter were developed. The Imaging Procedure Manual contained step-by-step instructions, example images, and troubleshooting guidance for sites
- Team-based training approach. MERIT’s team worked closely with the camera manufacturer, the local distributor/installer, and site staff to train on how exactly to take the fundus photographs for the study. Site staff were provided feedback and support on image acquisition, export, and upload throughout the study
- Careful use of test images. An initial number of subjects from each site were deemed test subjects by protocol and ICF, providing training images to reduce variability and ensure correct procedures
- Innovative, comprehensive data platform. Image quality control; DICOM conversion and anonymization; and masked, centralized reading of images were all streamlined through EXCELSIOR resulting in data and image files with a standardized naming schema
Results and Highlights
- Few imaging issues due to training. Very few imaging issues arose during the study due to careful planning and extensive training
- No data loss. No data lost due to image quality issues
- High quality images. High quality Images were key to the study’s success and formed the basis for testing the AI algorithm
- Quick indexing for AI training. Standard image file naming schema allows data to be quickly indexed and used as a training dataset for future AI algorithms
It’s crucial to have a team of expert readers supported by experienced project managers, data managers, and technologists to assure high-quality, consistent data review and interpretation.
With MERIT, your critical endpoints are safe in our hands.