Foundational Considerations for Artificial Intelligence Using Ophthalmic Images.

IMPORTANCE: The development of artificial intelligence (AI) and other machine diagnostic systems, also known as software as a medical device, and its recent introduction into clinical practice requires a deeply rooted foundation in bioethics for consideration by regulatory agencies and other stakeholders around the globe. OBJECTIVES: To initiate a dialogue on the issues to consider when developing a bioethically sound foundation for AI in medicine, based on images of eye structures, for discussion with all stakeholders. EVIDENCE REVIEW: The scope of the issues and summaries of the discussions under consideration by the Foundational Principles of Ophthalmic Imaging and Algorithmic Interpretation Working Group, as first presented during the Collaborative Community on Ophthalmic Imaging inaugural meeting on September 7, 2020, and afterward in the working group. FINDINGS: Artificial intelligence has the potential to improve health care access and patient outcome fundamentally while decreasing disparities, lowering cost, and enhancing the care team. Nevertheless, substantial concerns exist. Bioethicists, AI algorithm experts, as well as the Food and Drug Administration and other regulatory agencies, industry, patient advocacy groups, clinicians and their professional societies, other provider groups, and payors (i.e., stakeholders) working together in collaborative communities to resolve the fundamental ethical issues of nonmaleficence, autonomy, and equity are essential to attain this potential. Resolution impacts all levels of the design, validation, and implementation of AI in medicine. Design, validation, and implementation of AI warrant meticulous attention. CONCLUSIONS AND RELEVANCE: The development of a bioethically sound foundation may be possible if it is based in the fundamental ethical principles of nonmaleficence, autonomy, and equity for considerations for the design, validation, and implementation for AI systems. Achieving such a foundation will be helpful for continuing successful introduction into medicine before consideration by regulatory agencies. Important improvements in accessibility and quality of health care, decrease in health disparities, and lower cost thereby can be achieved. These considerations should be discussed with all stakeholders and expanded on as a useful initiation of this dialogue.

Choriocapillaris Degeneration in Geographic Atrophy.

Early age-related macular degeneration (AMD) is characterized by degeneration of the choriocapillaris, the vascular supply of retinal photoreceptor cells. We assessed vascular loss during disease progression in the choriocapillaris and larger vessels in the deeper choroid. Human donor maculae from controls (n = 99), early AMD (n = 35), or clinically diagnosed with geographic atrophy (GA; n = 9, collected from outside the zone of retinal pigment epithelium degeneration) were evaluated using Ulex europaeus agglutinin-I labeling to discriminate between vessels with intact endothelial cells and ghost vessels. Morphometric analyses of choriocapillaris density (cross-sectional area of capillary lumens divided by length) and of vascular lumen/stroma ratio in the outer choroid were performed. Choriocapillaris loss was observed in early AMD (Bonferroni-corrected P = 0.024) with greater loss in GA (Bonferroni-corrected P < 10-9), even in areas of intact retinal pigment epithelium. In contrast, changes in lumen/stroma ratio in the outer choroid were not found to differ between controls and AMD or GA eyes (P > 0.05), suggesting choriocapillaris changes are more prevalent in AMD than those in the outer choroid. In addition, vascular endothelial growth factor-A levels were negatively correlated with choriocapillaris vascular density. These findings support the concept that choroidal vascular degeneration, predominantly in the microvasculature, contributes to dry AMD progression. Addressing capillary loss in AMD remains an important translational target.

Identifying Ethical Considerations for Machine Learning Healthcare Applications.

Along with potential benefits to healthcare delivery, machine learning healthcare applications (ML-HCAs) raise a number of ethical concerns. Ethical evaluations of ML-HCAs will need to structure the overall problem of evaluating these technologies, especially for a diverse group of stakeholders. This paper outlines a systematic approach to identifying ML-HCA ethical concerns, starting with a conceptual model of the pipeline of the conception, development, implementation of ML-HCAs, and the parallel pipeline of evaluation and oversight tasks at each stage. Over this model, we layer key questions that raise value-based issues, along with ethical considerations identified in large part by a literature review, but also identifying some ethical considerations that have yet to receive attention. This pipeline model framework will be useful for systematic ethical appraisals of ML-HCA from development through implementation, and for interdisciplinary collaboration of diverse stakeholders that will be required to understand and subsequently manage the ethical implications of ML-HCAs.

Automated and Computer-Assisted Detection, Classification, and Diagnosis of Diabetic Retinopathy.

Background: The introduction of artificial intelligence (AI) in medicine has raised significant ethical, economic, and scientific controversies. Introduction: Because an explicit goal of AI is to perform processes previously reserved for human clinicians and other health care personnel, there is justified concern about the impact on patient safety, efficacy, equity, and liability. Discussion: Systems for computer-assisted and fully automated detection, triage, and diagnosis of diabetic retinopathy (DR) from retinal images show great variation in design, level of autonomy, and intended use. Moreover, the degree to which these systems have been evaluated and validated is heterogeneous. We use the term DR AI system as a general term for any system that interprets retinal images with at least some degree of autonomy from a human grader. We put forth these standardized descriptors to form a means to categorize systems for computer-assisted and fully automated detection, triage, and diagnosis of DR. The components of the categorization system include level of device autonomy, intended use, level of evidence for diagnostic accuracy, and system design. Conclusion: There is currently minimal empirical basis to assert that certain combinations of autonomy, accuracy, or intended use are better or more appropriate than any other. Therefore, at the current stage of development of this document, we have been descriptive rather than prescriptive, and we treat the different categorizations as independent and organized along multiple axes.

DYSFUNCTIONAL AUTONOMIC REGULATION OF THE CHOROID IN CENTRAL SEROUS CHORIORETINOPATHY.

PURPOSE: To study the effect of changing perfusion pressures on retinal and choroidal structure in central serous chorioretinopathy (CSC). METHODS: This prospective observational case series included seven healthy volunteers (14 eyes) and seven patients (14 eyes) with CSC. Each patient underwent spectral domain optical coherence tomography with enhanced depth imaging in the upright (sitting) and supine positions. Image segmentation focused on central macular thickness, subretinal fluid, total macular volume, choroidal thickness, and choriocapillaris thickness. Blood pressure and heart rate were measured in the upright and supine positions. RESULTS: Choriocapillaris thickness was thicker in CSC participants (34.23 µm; range, 30.9-36.5 µm) compared with healthy controls (13.96 µm; range, 7.15-23.87 µm) (P ≤ 0.001). The choroid was similarly thicker in CSC participants (371.4 µm; range, 200.2-459.4 µm) compared with healthy controls (231.4 µm; range 161.8-287.5 µm) (P ≤ 0.001). Choroidal thickness increased in patients with CSC when transitioning from upright (371.4 µm) to supine (377.8 µm) (P ≤ 0.01). By contrast, there was an 11.97% decrease in choroid thickness in normal controls when transitioning from upright (231.4 µm) to supine (203.9 µm). There were no significant hemodynamic changes. CONCLUSION: We demonstrated that choroidal thickness increased in response to increased perfusion pressures in patients with CSC and not in normal controls. These findings likely represent an autonomic dysregulation of choroidal blood flow in patients with CSC.

RetFM-J, an ImageJ-based module for automated counting and quantifying features of nuclei in retinal whole-mounts.

The present article introduces RetFM-J, a semi-automated ImageJ-based module that detects, counts, and collects quantitative data on nuclei of the inner retina from H&E-stained whole-mounted retinas. To illustrate performance, computer-derived outputs were analyzed in inbred C57BL/6J mice. Automated characterization yielded computer-derived outputs that closely matched manual counts. As a method using open-source software that is freely available, inexpensive staining reagents that are robust, and imaging equipment that is routine to most laboratories, RetFM-J could be utilized in a wide variety of experiments benefiting from high-throughput, quantitative, uniform analyses of total cellularity in the inner retina.

Quantitative measurement of retinal ganglion cell populations via histology-based random forest classification.

The inner surface of the retina contains a complex mixture of neurons, glia, and vasculature, including retinal ganglion cells (RGCs), the final output neurons of the retina and primary neurons that are damaged in several blinding diseases. The goal of the current work was two-fold: to assess the feasibility of using computer-assisted detection of nuclei and random forest classification to automate the quantification of RGCs in hematoxylin/eosin (H&E)-stained retinal whole-mounts; and if possible, to use the approach to examine how nuclear size influences disease susceptibility among RGC populations. To achieve this, data from RetFM-J, a semi-automated ImageJ-based module that detects, counts, and collects quantitative data on nuclei of H&E-stained whole-mounted retinas, were used in conjunction with a manually curated set of images to train a random forest classifier. To test performance, computer-derived outputs were compared to previously published features of several well-characterized mouse models of ophthalmic disease and their controls: normal C57BL/6J mice; Jun-sufficient and Jun-deficient mice subjected to controlled optic nerve crush (CONC); and DBA/2J mice with naturally occurring glaucoma. The result of these efforts was development of RetFM-Class, a command-line-based tool that uses data output from RetFM-J to perform random forest classification of cell type. Comparative testing revealed that manual and automated classifications by RetFM-Class correlated well, with 83.2% classification accuracy for RGCs. Automated characterization of C57BL/6J retinas predicted 54,642 RGCs per normal retina, and identified a 48.3% Jun-dependent loss of cells at 35 days post CONC and a 71.2% loss of RGCs among 16-month-old DBA/2J mice with glaucoma. Output from automated analyses was used to compare nuclear area among large numbers of RGCs from DBA/2J mice (n = 127,361). In aged DBA/2J mice with glaucoma, RetFM-Class detected a decrease in median and mean nucleus size of cells classified into the RGC category, as did an independent confirmation study using manual measurements of nuclear area demarcated by BRN3A-immunoreactivity. In conclusion, we have demonstrated that histology-based random forest classification is feasible and can be utilized to study RGCs in a high-throughput fashion. Despite having some limitations, this approach demonstrated a significant association between the size of the RGC nucleus and the DBA/2J form of glaucoma.

Novel method using 3-dimensional segmentation in spectral domain-optical coherence tomography imaging in the chick reveals defocus-induced regional and time-sensitive asymmetries in the choroidal thickness.

Studies into the mechanisms underlying the active emmetropization process by which neonatal refractive errors are corrected, have described rapid, compensatory changes in the thickness of the choroidal layer in response to imposed optical defocus. While high frequency A-scan ultrasonography, as traditionally used to characterize such changes, offers good resolution of central (on-axis) changes, evidence of local retinal control mechanisms make it imperative that more peripheral, off-axis changes also be tracked. In this study, we used in vivo high resolution spectral domain-optical coherence tomography (SD-OCT) imaging in combination with the Iowa Reference Algorithms for 3-dimensional segmentation, to more fully characterize these changes, both spatially and temporally, in young, 7-day old chicks (n = 15), which were fitted with monocular +15 D defocusing lenses to induce choroidal thickening. With these tools, we were also able to localize the retinal area centralis, which was used as a landmark along with the ocular pectin in standardizing the location of scans and aligning them for subsequent analyses of choroidal thickness (CT) changes across time and between eyes. Values were derived for each of four quadrants, centered on the area centralis, and global CT values were also derived for all eyes. Data were compared with on-axis changes measured using ultrasonography. There were significant on-axis choroidal thickening that was detected after just one day of lens wear (∼190 µm), and regional (quadrant-related) differences in choroidal responses were also found, as well as global thickness changes 1 day after treatment. The ratio of global to on-axis choroidal thicknesses, used as an index of regional variability in responses, was also found to change significantly, reflecting the significant central changes. In summary, we demonstrated in vivo high resolution SD-OCT imaging, used in combination with segmentation algorithms, to be a viable and informative approach for characterizing regional (spatial), time-sensitive changes in CT in small animals such as the chick.

A New Approach to Staging Diabetic Eye Disease: Staging of Diabetic Retinal Neurodegeneration and Diabetic Macular Edema.

Topic: The goal of this review was to summarize the current level of evidence on biomarkers to quantify diabetic retinal neurodegeneration (DRN) and diabetic macular edema (DME). Clinical relevance: With advances in retinal diagnostics, we have more data on patients with diabetes than ever before. However, the staging system for diabetic retinal disease is still based only on color fundus photographs and we do not have clear guidelines on how to incorporate data from the relatively newer modalities into clinical practice. Methods: In this review, we use a Delphi process with experts to identify the most promising modalities to identify DRN and DME. These included microperimetry, full-field flash electroretinogram, spectral-domain OCT, adaptive optics, and OCT angiography. We then used a previously published method of determining the evidence level to complete detailed evidence grids for each modality. Results: Our results showed that among the modalities evaluated, the level of evidence to quantify DRN and DME was highest for OCT (level 1) and lowest for adaptive optics (level 4). Conclusion: For most of the modalities evaluated, prospective studies are needed to elucidate their role in the management and outcomes of diabetic retinal diseases. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Risk Factors for Nondiagnostic Imaging in a Real-World Deployment of Artificial Intelligence Diabetic Retinal Examinations in an Integrated Healthcare System: Maximizing Workflow Efficiency Through Predictive Dilation.

OBJECTIVE: In the pivotal clinical trial that led to Food and Drug Administration De Novo "approval" of the first fully autonomous artificial intelligence (AI) diabetic retinal disease diagnostic system, a reflexive dilation protocol was used. Using real-world deployment data before implementation of reflexive dilation, we identified factors associated with nondiagnostic results. These factors allow a novel predictive dilation workflow, where patients most likely to benefit from pharmacologic dilation are dilated a priori to maximize efficiency and patient satisfaction. METHODS: Retrospective review of patients who were assessed with autonomous AI at Johns Hopkins Medicine (8/2020 to 5/2021). We constructed a multivariable logistic regression model for nondiagnostic results to compare characteristics of patients with and without diagnostic results, using adjusted odds ratio (aOR). P < .05 was considered statistically significant. RESULTS: Of 241 patients (59% female; median age = 59), 123 (51%) had nondiagnostic results. In multivariable analysis, type 1 diabetes (T1D, aOR = 5.82, 95% confidence interval [CI]: 1.45-23.40, P = .01), smoking (aOR = 2.86, 95% CI: 1.36-5.99, P = .005), and age (every 10-year increase, aOR = 2.12, 95% CI: 1.62-2.77, P < .001) were associated with nondiagnostic results. Following feature elimination, a predictive model was created using T1D, smoking, age, race, sex, and hypertension as inputs. The model showed an area under the receiver-operator characteristics curve of 0.76 in five-fold cross-validation. CONCLUSIONS: We used factors associated with nondiagnostic results to design a novel, predictive dilation workflow, where patients most likely to benefit from pharmacologic dilation are dilated a priori. This new workflow has the potential to be more efficient than reflexive dilation, thus maximizing the number of at-risk patients receiving their diabetic retinal examinations.

Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: the ACCESS randomized control trial.

Diabetic retinopathy can be prevented with screening and early detection. We hypothesized that autonomous artificial intelligence (AI) diabetic eye exams at the point-of-care would increase diabetic eye exam completion rates in a racially and ethnically diverse youth population. AI for Children's diabetiC Eye ExamS (NCT05131451) is a parallel randomized controlled trial that randomized youth (ages 8-21 years) with type 1 and type 2 diabetes to intervention (autonomous artificial intelligence diabetic eye exam at the point of care), or control (scripted eye care provider referral and education) in an academic pediatric diabetes center. The primary outcome was diabetic eye exam completion rate within 6 months. The secondary outcome was the proportion of participants who completed follow-through with an eye care provider if deemed appropriate. Diabetic eye exam completion rate was significantly higher (100%, 95%CI: 95.5%, 100%) in the intervention group (n = 81) than the control group (n = 83) (22%, 95%CI: 14.2%, 32.4%)(p < 0.001). In the intervention arm, 25/81 participants had an abnormal result, of whom 64% (16/25) completed follow-through with an eye care provider, compared to 22% in the control arm (p < 0.001). Autonomous AI increases diabetic eye exam completion rates in youth with diabetes.

Effectiveness of artificial intelligence screening in preventing vision loss from diabetes: a policy model.

The effectiveness of using artificial intelligence (AI) systems to perform diabetic retinal exams ('screening') on preventing vision loss is not known. We designed the Care Process for Preventing Vision Loss from Diabetes (CAREVL), as a Markov model to compare the effectiveness of point-of-care autonomous AI-based screening with in-office clinical exam by an eye care provider (ECP), on preventing vision loss among patients with diabetes. The estimated incidence of vision loss at 5 years was 1535 per 100,000 in the AI-screened group compared to 1625 per 100,000 in the ECP group, leading to a modelled risk difference of 90 per 100,000. The base-case CAREVL model estimated that an autonomous AI-based screening strategy would result in 27,000 fewer Americans with vision loss at 5 years compared with ECP. Vision loss at 5 years remained lower in the AI-screened group compared to the ECP group, in a wide range of parameters including optimistic estimates biased toward ECP. Real-world modifiable factors associated with processes of care could further increase its effectiveness. Of these factors, increased adherence with treatment was estimated to have the greatest impact.

Autonomous AI systems in the face of liability, regulations and costs.

Autonomous AI systems in medicine promise improved outcomes but raise concerns about liability, regulation, and costs. With the advent of large-language models, which can understand and generate medical text, the urgency for addressing these concerns increases as they create opportunities for more sophisticated autonomous AI systems. This perspective explores the liability implications for physicians, hospitals, and creators of AI technology, as well as the evolving regulatory landscape and payment models. Physicians may be favored in malpractice cases if they follow rigorously validated AI recommendations. However, AI developers may face liability for failing to adhere to industry-standard best practices during development and implementation. The evolving regulatory landscape, led by the FDA, seeks to ensure transparency, evaluation, and real-world monitoring of AI systems, while payment models such as MPFS, NTAP, and commercial payers adapt to accommodate them. The widespread adoption of autonomous AI systems can potentially streamline workflows and allow doctors to concentrate on the human aspects of healthcare.

Considerations for addressing bias in artificial intelligence for health equity.

Health equity is a primary goal of healthcare stakeholders: patients and their advocacy groups, clinicians, other providers and their professional societies, bioethicists, payors and value based care organizations, regulatory agencies, legislators, and creators of artificial intelligence/machine learning (AI/ML)-enabled medical devices. Lack of equitable access to diagnosis and treatment may be improved through new digital health technologies, especially AI/ML, but these may also exacerbate disparities, depending on how bias is addressed. We propose an expanded Total Product Lifecycle (TPLC) framework for healthcare AI/ML, describing the sources and impacts of undesirable bias in AI/ML systems in each phase, how these can be analyzed using appropriate metrics, and how they can be potentially mitigated. The goal of these "Considerations" is to educate stakeholders on how potential AI/ML bias may impact healthcare outcomes and how to identify and mitigate inequities; to initiate a discussion between stakeholders on these issues, in order to ensure health equity along the expanded AI/ML TPLC framework, and ultimately, better health outcomes for all.

Autonomous artificial intelligence increases real-world specialist clinic productivity in a cluster-randomized trial.

Autonomous artificial intelligence (AI) promises to increase healthcare productivity, but real-world evidence is lacking. We developed a clinic productivity model to generate testable hypotheses and study design for a preregistered cluster-randomized clinical trial, in which we tested the hypothesis that a previously validated US FDA-authorized AI for diabetic eye exams increases clinic productivity (number of completed care encounters per hour per specialist physician) among patients with diabetes. Here we report that 105 clinic days are cluster randomized to either intervention (using AI diagnosis; 51 days; 494 patients) or control (not using AI diagnosis; 54 days; 499 patients). The prespecified primary endpoint is met: AI leads to 40% higher productivity (1.59 encounters/hour, 95% confidence interval [CI]: 1.37-1.80) than control (1.14 encounters/hour, 95% CI: 1.02-1.25), p < 0.00; the secondary endpoint (productivity in all patients) is also met. Autonomous AI increases healthcare system productivity, which could potentially increase access and reduce health disparities. ClinicalTrials.gov NCT05182580.

Generative Artificial Intelligence Through ChatGPT and Other Large Language Models in Ophthalmology: Clinical Applications and Challenges.

The rapid progress of large language models (LLMs) driving generative artificial intelligence applications heralds the potential of opportunities in health care. We conducted a review up to April 2023 on Google Scholar, Embase, MEDLINE, and Scopus using the following terms: "large language models," "generative artificial intelligence," "ophthalmology," "ChatGPT," and "eye," based on relevance to this review. From a clinical viewpoint specific to ophthalmologists, we explore from the different stakeholders' perspectives-including patients, physicians, and policymakers-the potential LLM applications in education, research, and clinical domains specific to ophthalmology. We also highlight the foreseeable challenges of LLM implementation into clinical practice, including the concerns of accuracy, interpretability, perpetuating bias, and data security. As LLMs continue to mature, it is essential for stakeholders to jointly establish standards for best practices to safeguard patient safety. Financial Disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Validation of tablet-based evaluation of color fundus images.

PURPOSE: To compare diabetic retinopathy (DR) referral recommendations made by viewing fundus images using a tablet computer with those made using a standard desktop display. METHODS: A tablet computer (iPad) and a desktop computer with a high-definition color display were compared. For each platform, 2 retinal specialists independently rated 1,200 color fundus images from patients at risk for DR using an annotation program Truthseeker. The specialists determined whether each image had referable DR and also how urgently each patient should be referred for medical examination. Graders viewed and rated the randomly presented images independently and were masked to their ratings on the alternative platform. Tablet-based and desktop display-based referral ratings were compared using cross-platform intraobserver kappa as the primary outcome measure. Additionally, interobserver kappa, sensitivity, specificity, and area under the receiver operating characteristic were determined. RESULTS: A high level of cross-platform intraobserver agreement was found for the DR referral ratings between the platforms (κ = 0.778) and for the 2 graders (κ = 0.812). Interobserver agreement was similar for the 2 platforms (κ = 0.544 and κ = 0.625 for tablet and desktop, respectively). The tablet-based ratings achieved a sensitivity of 0.848, a specificity of 0.987, and an area under the receiver operating characteristic of 0.950 compared with desktop display-based ratings. CONCLUSION: In this pilot study, tablet-based rating of color fundus images for subjects at risk for DR was consistent with desktop display-based rating. These results indicate that tablet computers can be reliably used for clinical evaluation of fundus images for DR.