Background. Overview of ocular conditions and global statistics Corneal disease is the fifth leading cause of blindness in the world, and approximately 4.5 million individuals have moderate to severe vision impairment secondary to loss of corneal clarity. Compared to other leading causes of blindness, corneal disease primarily affects a younger population and therefore has a greater disability-adjusted life years. Only 1 in 70 individuals with corneal blindness ultimately undergoes corneal transplantation due to a number of issues including socioeconomic and political factors. As a result, the number of keratoplasty procedures completed in the U.S. per annum is about 50,000.
Tracking Long-Term Outcomes After Corneal Transplantation In the United States, there is currently no registry or database tracking donor or recipient longitudinal outcomes after corneal transplantation. Other organ transplants including kidney, liver, heart, lung, and pancreas have an established registry , despite corneal transplants being one of the most common transplantations in the US. Australia is one of the few countries that has an established corneal graft registry since 1985, which has provided invaluable insight to determine positive and negative prognostic factors affecting corneal graft survival. In order to obtain best subject outcomes, clinical practice should ideally be tailored to selecting the best type of surgery (i.e. penetrating keratoplasty \[PKP\], endothelial keratoplasty \[EK\], anterior lamellar keratoplasty \[ALK\], or artificial cornea) for each individual patient, based on real world outcomes data.
Developing and utilizing artificial intelligence for corneal disease Machine learning, which plays an ever-growing role in developing artificial intelligence systems for medical applications, is a powerful means of handling very large data sets. A variety of algorithms can incorporate many values more efficiently and accurately than humans. Imaging studies are particularly rich, making them well-suited for machine learning.
An accurate AI/ML-enabled algorithm assessment of various imaging studies could improve precision over physical exams, improving patient outcomes by earlier and more accurate detection of abnormalities and better prediction of future outcomes. Additionally, AI/ML-enabled remote collection of patient data presents substantial potential benefits for patients, providers, and the broader health system to monitor disease, outcomes of surgery or treatment. With home- or community-based monitoring, healthy patients can save time and money traveling frequently to the clinic. For those where issues are detected, potential ocular conditions or post-surgical complications can be identified earlier before they become more severe and require intervention or surgery, which improves both patient outcomes and saves health system resources.
Objectives. Primary: To establish the first nationwide corneal registry in the United States to include information related to the disease state, information on donor tissue, recipient data, surgical procedure, and long-term clinical outcomes. Ultimately, this prospective data collection will allow us to determine prognostic factors for successful corneal transplantation and create an algorithm to guide clinical practice based on real world outcomes.
Secondary: To collect and create a database of de-identified imaging studies (including but not limited to optical coherence topography (OCT), in vivo confocal biomicroscopy, specular biomicroscopy, and corneal topography) to ultimately develop artificial intelligence (AI) based diagnostic and prognostic algorithms for corneal disease prevalence, progression and surgery outcomes.
Study Design. Design Prospective and observational.
Study Size The initial study subject recruitment will be piloted at a variety of US centers. All eligible subjects will be recruited and consented subjects will be enrolled during the initial phase of the study.
Data Collection US based corneal surgeons will obtain corneal images pre- and post- corneal transplantation. These de-identified images, along with the clinical information (donor and recipient characteristics, surgical information, and longitudinal outcomes afterwards) will be entered into the registry.
Data Elements for Corneal Graft Registry For the subjects undergoing corneal transplantation, the following elements will be collected and entered into a secure, electronic database. The imaging data source for this study are copies of corneal topography OCT, specular biomicroscopy and in vivo confocal biomicroscopy images produced during routine clinical care. The registry will receive copies of images in any format, including electronic data transfers and CDs. OCT images from different providers and care sites may vary in quality and detail. The abstraction process will map data to a single cohesive data schema.
Data sources All OCT and corneal topography images are de-identified with no subject health information. Only the raw images will be collected for analysis, and OCT images will be compiled with an aim to create an online registry.
Data collection and storage OCT images will be submitted by healthcare providers through a secure, encrypted, imaging request platform with personnel follow-up as needed. Imaging documentation is uploaded to the study's servers and de-identified of all subject data and protected health information (PHI).
Data abstraction Study staff with expertise in assessing OCT images will review all images submitted to detect patterns. These patterns will eventually be used to train AI/ML algorithms for the collection of measurement data.
Data security This study will comply with Health Insurance Portability and Accountability Act (HIPAA) security standards. In addition, the study team has a comprehensive set of security policies, including risk management strategies, incident response protocols, access controls, encryption standards, and study staff training to safeguard all subject images submitted.
Proposed Algorithm Development. Description of proposed Machine Learning method The algorithm has the opportunity to be the most versatile of any automated OCT image classifier and data collector. With enough data, it also can be the most accurate. The algorithm will be trained and optimized using a variety of OCT data from this study.
Data Management. Retention of images Images and documents pertaining to the study will be retained for the length of time required by relevant national or local health authorities, whichever is longer. After that period of time, the documents may be destroyed, subject to local regulations.
Data quality assurance policies The study team ensures the accuracy of data abstracted from OCT images through a range of measures leveraging both technology and human expertise. The image collection platform is designed to flag irregularities and low confidence images using conservative thresholds.
The study team undergoes a training program and must pass rigorous data quality testing before assuming full imaging screening responsibilities. All images are screened by a minimum of two reviewers, and difficult scenarios that are not described in standard procedures are escalated to a senior team lead, per policy. Procedures to document, review, and learn from escalations create feedback loops that improve operational effectiveness and reduce human error.
The study team will maintain logs of all data transformations and perform regular internal data quality audits. The data quality will be continuously monitored and analyzed throughout the submission and review process.
Access to Registry. Role-Based Access Control (RBAC) RBAC will be implemented to define different levels of access based on the user's role. Roles will be well-defined and correspond to specific responsibilities and permissions.
Authentication Strong authentication mechanisms, including two-factor authentication (2FA) will be in place to ensure that only authorized users can access the imaging registry. An authorization workflow where user access requests are reviewed and approved by study personnel will be utilized before access is granted. Study personnel will regularly review and manage user accounts, ensuring that only active users with legitimate access needs have accounts in the registry.
Access Granting / Revocation Access rights and permissions to users will be shared based on roles and responsibilities and the least privilege necessary will be granted for users to perform their tasks. Revoking access rights will be streamlined if users change roles or no longer require access. The study team will implement logging mechanisms to record user activities and access attempts. The study team will review logs to detect and investigate any suspicious or unauthorized activities.
Upon request, auditors from certain regulatory institutions (i.e., CMS, FDA, etc.) or other third-party institutions may be granted temporary access to the registry for auditing purposes.
Incident Reporting Security incidents or breaches related to unauthorized access will be dealt with promptly to mitigate the impact of security incidents and prevent recurrence.
Withdrawal of Imaging Data The Principal Investigator or IRB has the right to remove and imaging data for medical, safety, or administrative reasons at any time. Appropriate procedures will be followed to ensure the safe withdrawal of each image from the study.
Image De-identification.
To ensure the secure and ethical handling of OCT data, a comprehensive image de-identification process will be implemented. This process aims to systematically remove or alter identifiable information from each image and its associated metadata while preserving clinical and research value of the images. The following steps outline the key aspects of this de-identification process:
Removing direct identifiers
* All direct identifiers used for the purpose of individual identification, such as subject names, medical record and accession numbers, and dates of birth, will be thoroughly searched for and removed from each image's pixel data.
* Concurrently, these direct identifiers will also be sought out and removed from the image metadata, except for the medical record number, which will be irreversibly transformed via a cryptographic hashing function.
Pixel-level Anonymization
● If required, specific image regions containing identifiable features, such as facial details or unique markings will undergo either masking or blurring. Such regions lacking diagnostic features will be masked, while those with diagnostic features will be subject to blurring.
Quality Control
* Rigorous quality checks will be executed to ensure that the anonymization process does not compromise the clinical value of the images.
* Trained professionals will review a subset of de-identified images to verify that critical diagnostic features are preserved accurately.
Encryption
* Both the original and de-identified images will be encrypted to ensure their security during storage and transmission.
* All data will be stored in a secure environment with controlled access, adhering to regulatory requirements and industry best practices.
Documentation
* A detailed record of the de-identification process will be maintained, including a comprehensive account of the steps undertaken, personnel involved, and any challenges encountered.
* This document serves as an essential audit trail, offering transparency and aiding in demonstrating compliance with data protection regulations.
