Search

Abstract Background Sensory changes due to aging or disease can impact brain tissue. This study aims to investigate the link between glaucoma, a leading cause of blindness, and alterations in brain connections. Methods We analyzed diffusion MRI measurements of white matter tissue in a large group, consisting of 905 glaucoma patients (aged 49-80) and 5292 healthy individuals (aged 45-80) from the UK Biobank. Confounds due to group differences were mitigated by matching a sub-sample of controls to glaucoma subjects. We compared classification of glaucoma using convolutional neural networks (CNNs) focusing on the optic radiations, which are the primary visual connection to the cortex, against those analyzing non-visual brain connections. As a control, we evaluated the performance of regularized linear regression models. Results We showed that CNNs using information from the optic radiations exhibited higher accuracy in classifying subjects with glaucoma when contrasted with CNNs relying on information from non-visual brain connections. Regularized linear regression models were also tested, and showed significantly weaker classification performance. Additionally, the CNN was unable to generalize to the classification of age-group or of age-related macular degeneration. Conclusions Our findings indicate a distinct and potentially non-linear signature of glaucoma in the tissue properties of optic radiations. This study enhances our understanding of how glaucoma affects brain tissue and opens avenues for further research into how diseases that affect sensory input may also affect brain aging.

Objective: To describe visual acuity data representation in the American Academy of Ophthalmology Intelligent Research in Sight (IRIS) Registry and present a data-cleaning strategy Design: Reliability and validity study. Participants: Patients with visual acuity records from 2018 in the IRIS Registry. Methods: Visual acuity measurements and metadata were identified and characterized from 2018 IRIS Registry records. Metadata, including laterality, assessment method (distance, near, and unspecified), correction (corrected, uncorrected, and unspecified), and flags for refraction or pinhole assessment were compared between Rome (frozen April 20, 2020) and Chicago (frozen December 24, 2021) versions. We developed a data-cleaning strategy to infer patients’ corrected distance visual acuity in their better-seeing eye. Main Outcome Measures: Visual acuity data characteristics in the IRIS Registry. Results: The IRIS Registry Chicago data set contains 168 920 049 visual acuity records among 23 001 531 unique patients and 49 968 974 unique patient visit dates in 2018. Visual acuity records were associated with refraction in 5.3% of cases, and with pinhole in 11.0%. Mean (standard deviation) of all measurements was 0.26 (0.41) logarithm of the minimum angle of resolution (logMAR), with a range of − 0.3 to 4.0 A plurality of visual acuity records were labeled corrected (corrected visual acuity [CVA], 39.1%), followed by unspecified (37.6%) and uncorrected (uncorrected visual acuity [UCVA], 23.4%). Corrected visual acuity measurements were paradoxically worse than same day UCVA 15% of the time. In aggregate, mean and median values were similar for CVA and unspecified visual acuity. Most visual acuity measurements were at distance (59.8%, vs. 32.1% unspecified and 8.2% near). Rome contained more duplicate visual acuity records than Chicago (10.8% vs. 1.4%). Near visual acuity was classified with Jaeger notation and (in Chicago only) also assigned logMAR values by Verana Health. LogMAR values for hand motion and light perception visual acuity were lower in Chicago than in Rome. The impact of data entry errors or outliers on analyses may be reduced by filtering and averaging visual acuity per eye over time. Conclusions: The IRIS Registry includes similar visual acuity metadata in Rome and Chicago. Although fewer duplicate records were found in Chicago, both versions include duplicate and atypical measurements (i.e., CVA worse than UCVA on the same day). Analyses may benefit from using algorithms to filter outliers and average visual acuity measurements over time. Financial Disclosure(s): Proprietary or commercial disclosure may be found found in the Footnotes and Disclosures at the end of this article.

Objective: To report the incidence of and evaluate demographic, ocular comorbidities, and intraoperative factors for rhegmatogenous retinal detachment (RRD) and retinal tear (RT) after cataract surgery in the American Academy of Ophthalmology IRIS® Registry (Intelligent Research in Sight). Design: Retrospective cohort study. Participants: Patients aged ≥ 40 years who underwent cataract surgery between 2014 and 2017. Methods: Multivariable logistic regression was used to evaluate demographic, comorbidity, and intraoperative factors associated with RRD and RT after cataract surgery. Main Outcome Measures: Incidence and risk factors for RRD or RT within 1 year of cataract surgery. Results: Of the 3 177 195 eyes of 1 983 712 patients included, 6690 (0.21%) developed RRD and 5489 (0.17%) developed RT without RRD within 1 year after cataract surgery. Multivariable logistic regression odds ratios (ORs) showed increased risk of RRD and RT, respectively, among men (OR 3.15; 95% confidence interval [CI], 2.99–3.32; P < 0.001 and 1.79; 95% CI, 1.70–1.89; P < 0.001), and younger ages compared with patients aged > 70, peaking at age 40 to 50 for RRD (8.61; 95% CI, 7.74–9.58; P < 0.001) and age 50 to 60 for RT (2.74; 95% CI, 2.52–2.98; P < 0.001). Increased odds of RRD were observed for procedure eyes with lattice degeneration (LD) (10.53; 95% CI, 9.82–11.28; P < 0.001), hypermature cataract (1.61; 95% CI, 1.06–2.45; P = 0.03), complex cataract surgery (1.52; 95% CI, 1.4–1.66; P < 0.001), posterior vitreous detachment (PVD) (1.24; 95% CI, 1.15–1.34; P < 0.001), and high myopia (1.2; 95% CI, 1.14–1.27; P < 0.001). Lattice degeneration conferred the highest odds of RT (43.86; 95% CI, 41.39–46.49; P < 0.001). Conclusion: In the IRIS Registry, RRD occurs in approximately 1 in 500 cataract surgeries in patients aged > 40 years within 1 year of surgery. The presence of LD conferred the highest odds for RRD and RT after surgery. Additional risk factors for RRD included male gender, younger age, hypermature cataract, PVD, and high myopia. These data may be useful during the informed consent process for cataract surgery and help identify patients at a higher risk of retinal complications. Financial Disclosure(s): The author(s) have no proprietary or commercial interest in any materials discussed in this article.

Purpose: Neovascular age-related macular degeneration (nAMD) shows variable treatment response to intravitreal anti-VEGF. This analysis compared the potential of different artificial intelligence (AI)-based machine learning models using OCT and clinical variables to accurately predict at baseline the best-corrected visual acuity (BCVA) at 9 months in response to ranibizumab in patients with nAMD. Design: Retrospective analysis. Participants: Baseline and imaging data from patients with subfoveal choroidal neovascularization secondary to age-related macular dengeration. Methods: Baseline data from 502 study eyes from the HARBOR (NCT00891735) prospective clinical trial (monthly ranibizumab 0.5 and 2.0 mg arms) were pooled; 432 baseline OCT volume scans were included in the analysis. Seven models, based on baseline quantitative OCT features (Least absolute shrinkage and selection operator [Lasso] OCT minimum [min], Lasso OCT 1 standard error [SE]); on quantitative OCT features and clinical variables at baseline (Lasso min, Lasso 1SE, CatBoost, RF [random forest]); or on baseline OCT images only (deep learning [DL] model), were systematically compared with a benchmark linear model of baseline age and BCVA. Quantitative OCT features were derived by a DL segmentation model on the volume images, including retinal layer volumes and thicknesses, and retinal fluid biomarkers, including statistics on fluid volume and distribution. Main Outcome Measures: Prognostic ability of the models was evaluated using coefficient of determination (R2) and median absolute error (MAE; letters). Results: In the first cross-validation split, mean R2 (MAE) of the Lasso min, Lasso 1SE, CatBoost, and RF models was 0.46 (7.87), 0.42 (8.43), 0.45 (7.75), and 0.43 (7.60), respectively. These models ranked higher than or similar to the benchmark model (mean R2, 0.41; mean MAE, 8.20 letters) and better than OCT-only models (mean R2: Lasso OCT min, 0.20; Lasso OCT 1SE, 0.16; DL, 0.34). The Lasso min model was selected for detailed analysis; mean R2 (MAE) of the Lasso min and benchmark models for 1000 repeated cross-validation splits were 0.46 (7.7) and 0.42 (8.0), respectively. Conclusions: Machine learning models based on AI-segmented OCT features and clinical variables at baseline may predict future response to ranibizumab treatment in patients with nAMD. However, further developments will be needed to realize the clinical utility of such AI-based tools. Financial Disclosure(s): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Abstract Technological advancements are dramatically changing the landscape of therapeutic development. The convergence of advances in computing power, analytical methods, artificial intelligence, novel digital health tools, and cloud‐based platforms has the potential to power an exponential acceleration of evidence generation. For regulatory agencies responsible for evidence evaluation and oversight of medical products, these advances present both promises and challenges. Ultimately, realizing the translation and impact of these innovations that could potentially enhance therapeutic development and improve the health of individuals and the public will require a nimble and responsive regulatory approach. Supporting an adaptive policy‐making infrastructure that is poised to address novel regulatory considerations, creating a workforce to ensure relevant expertise, and fostering more diverse collaborations with a broader group of stakeholders are steps toward the goal of modernizing the regulatory ecosystem. This article outlines approaches that can help provide the flexibility and tools needed to foster innovation, while ensuring the safety and effectiveness of medical products.

Purpose: To identify clinical factors associated with the need for future surgical intervention following closed globe ocular trauma. Design: Retrospective cohort study. Subjects, Participants, and/or Controls: Patients in the American Academy of Ophthalmology IRIS® Registry (Intelligent Research in Sight) with a diagnosis of closed globe ocular trauma occurring between 2013 and 2019, identified using International Classification of Disease, 10th Revision and Systematized Nomenclature of Medicine codes. Methods: Diagnosis codes were used to identify multiple concomitant diagnoses present on the date of closed globe ocular trauma. Survival analyses were performed for each outcome of interest, and linear regression was used to identify clinical factors associated with the risk of surgical intervention. Main Outcome Measures: Outcomes included retinal break treatment, retinal detachment (RD) repair, retinal break treatment or RD repair, glaucoma surgery, and cataract surgery. Results: Of the 206 807 patients with closed globe ocular trauma, 9648 underwent surgical intervention during the follow-up period (mean, 444 days): 1697 (0.8%) had RD repair, 1658 (0.8%) had retinal break treatment, 600 (0.3%) had glaucoma surgery, and 5693 (2.8%) had cataract surgery. Traumatic cataract was the strongest risk factor for cataract surgery (hazard ratio, 13.0; 95% confidence interval, 10.8–15.6), traumatic hyphema showed highest risk for glaucoma surgery (7.24; 4.60–11.4), and vitreous hemorrhage was the strongest risk factor for retinal break treatment and detachment repair (11.01; 9.18–13.2 and 14.2; 11.5–17.6, respectively) during the first 60 days after trauma date. Vitreous hemorrhage was a risk factor for cataract surgery at > 60 days after trauma date only. Iris–angle injury was the strongest risk factor for glaucoma surgery > 60 days after trauma, while vitreous hemorrhage remained the strongest factor for retinal break treatment and detachment repair at > 60 days. Traumatic hyphema was a risk factor for all surgical outcomes during all follow-up intervals. Conclusions: Diagnosis of concomitant traumatic cataract, vitreous hemorrhage, traumatic hyphema, and other risk factors may increase the likelihood of requiring surgical intervention after closed globe ocular trauma.

Objective: To obtain complete DNA sequences of adenoviral (AdV) D8 genome from patients with conjunctivitis and determine the relation of sequence variation to clinical outcomes. Design: This study is a post hoc analysis of banked conjunctival swab samples from the BAYnovation Study, a previously conducted, randomized controlled clinical trial for AdV conjunctivitis. Participants: Ninety-six patients with AdV D8-positive conjunctivitis who received placebo treatment in the BAYnovation Study were included in the study. Methods: DNA from conjunctival swabs was purified and subjected to whole-genome viral DNA sequencing. Adenovirus D8 variants were identified and correlated with clinical outcomes, including 2 machine learning methods. Main Outcome Measures: Viral DNA sequence and development of subepithelial infiltrates (SEIs) were the main outcome measures. Results: From initial sequencing of 80 AdV D8-positive samples, full adenoviral genome reconstructions were obtained for 71. A total of 630 single-nucleotide variants were identified, including 156 missense mutations. Sequence clustering revealed 3 previously unappreciated viral clades within the AdV D8 type. The likelihood of SEI development differed significantly between clades, ranging from 83% for Clade 1 to 46% for Clade 3. Genome-wide analysis of viral single-nucleotide polymorphisms failed to identify single-gene determinants of outcome. Two machine learning models were independently trained to predict clinical outcome using polymorphic sequences. Both machine learning models correctly predicted development of SEI outcomes in a newly sequenced validation set of 16 cases (P = 1.5 × 10−5). Prediction was dependent on ensemble groups of polymorphisms across multiple genes. Conclusions: Adenovirus D8 has ≥ 3 prevalent molecular substrains, which differ in propensity to result in SEIs. Development of SEIs can be accurately predicted from knowledge of full viral sequence. These results suggest that development of SEIs in AdV D8 conjunctivitis is largely attributable to pathologic viral sequence variants within the D8 type and establishes machine learning paradigms as a powerful technique for understanding viral pathogenicity.

Purpose: A deep learning model was developed to detect nonexudative macular neovascularization (neMNV) using OCT B-scans. Design: Retrospective review of a prospective, observational study. Participants: Normal control eyes and patients with age-related macular degeneration (AMD) with and without neMNV. Methods: Swept-source OCT angiography (SS-OCTA) imaging (PLEX Elite 9000, Carl Zeiss Meditec, Inc) was performed using the 6 × 6-mm scan pattern. Individual B-scans were annotated to distinguish between drusen and the double-layer sign (DLS) associated with the neMNV. The machine learning model was tested on a dataset graded by humans, and model performance was compared with the human graders. Main Outcome Measures: Intersection over Union (IoU) score was measured to evaluate segmentation network performance. Area under the receiver operating characteristic curve values, sensitivity, specificity, and positive predictive value (PPV) and negative predictive value (NPV) were measured to assess the performance of the final classification performance. Chance-corrected agreement between the algorithm and the human grader determinations was measured with Cohen’s kappa. Results: A total of 251 eyes from 210 patients, including 182 eyes with DLS and 115 eyes with drusen, were used for model training. Of 125 500 B-scans, 6879 B-scans were manually annotated. A vision transformer segmentation model was built to extract DLS and drusen from B-scans. The extracted prediction masks from all B-scans in a volume were projected to an en face image, and an eye-level projection map was obtained for each eye. A binary classification algorithm was established to identify eyes with neMNV from the projection map. The algorithm achieved 82%, 90%, 79%, and 91% sensitivity, specificity, PPV, and NPV, respectively, on a separate test set of 100 eyes that were evaluated by human graders in a previous study. The area under the curve value was calculated as 0.91 (95% confidence interval, 0.85–0.98). The results of the algorithm showed excellent agreement with the senior human grader (kappa = 0.83, P < 0.001) and moderate agreement with the junior grader consensus (kappa = 0.54, P < 0.001). Conclusions: Our network (code is available at https://github.com/uw-biomedical-ml/double_layer_vit) was able to detect the presence of neMNV from structural B-scans alone by applying a purely transformer-based model.

Purpose: Investigate associations of natural environmental exposures with exudative and nonexudative age-related macular degeneration (AMD) across the United States. Design: Database study. Participants: Patients aged ≥ 55 years who were active in the IRIS Registry from 2016 to 2018 were analyzed. Patients were categorized as nonexudative, inactive exudative, and active exudative AMD by International Classification of Diseases 10th Revision and Current Procedural Terminology (CPT) codes. Patients without provider-level ZIP codes matching any ZIP code tabulation area were excluded. Methods: Environmental data were obtained from public sources including the US Geological Survey, National Renewable Energy Laboratory, National Oceanic and Atmospheric Administration, and Environmental Protection Agency. Multiple variable, mixed effects logistic regression models with random intercepts per ZIP code tabulation area quantified the association of each environmental variable with any AMD versus non-AMD patients, any exudative AMD versus nonexudative AMD, and active exudative AMD versus inactive exudative and nonexudative AMD using 3 separate models, while adjusting for age, sex, race, insurance type, smoking history, and phakic status. Main Outcome Measure: Odds ratios for environmental factors. Results: A total of 9 884 527 patients were included. Elevation, latitude, solar irradiance measured in global horizontal irradiance (GHI) and direct normal irradiance (DNI), temperature and precipitation variables, and pollution variables were included in our models. Statistically significant associations with active exudative AMD were GHI (odds ratio [OR], 3.848; 95% confidence interval [CI] with Bonferroni correction, 1.316–11.250), DNI (OR, 0.581; 95% CI, 0.370–0.913), latitude (OR, 1.110; 95% CI, 1.046–1.178), ozone (OR, 1.014; 95% CI, 1.004–1.025), and nitrogen dioxide (OR, 1.005; 95% CI, 1.000–1.010). The only significant environmental associations with any AMD were inches of snow in the winter (OR, 1.005; 95% CI, 1.001–1.009) and ozone (OR, 1.011; 95% CI, 1.003–1.019). Conclusions: The strongest environmental associations differed between AMD subgroups. The solar variables GHI, DNI, and latitude were significantly associated with active exudative AMD. Two pollutant variables, ozone and nitrogen dioxide, also showed positive associations with AMD. Further studies are warranted to investigate the clinical relevance of these associations. Our curated environmental dataset has been made publicly available at https://github.com/uw-biomedical-ml/AMD_environmental_dataset.

BackgroundMany artificial intelligence (AI) studies have focused on development of AI models, novel techniques, and reporting guidelines. However, little is understood about clinicians' perspectives of AI applications in medical fields including ophthalmology, particularly in light of recent regulatory guidelines. The aim for this study was to evaluate the perspectives of ophthalmologists regarding AI in 4 major eye conditions: diabetic retinopathy (DR), glaucoma, age-related macular degeneration (AMD) and cataract.MethodsThis was a multi-national survey of ophthalmologists between March 1st, 2020 to February 29th, 2021 disseminated via the major global ophthalmology societies. The survey was designed based on microsystem, mesosystem and macrosystem questions, and the software as a medical device (SaMD) regulatory framework chaired by the Food and Drug Administration (FDA). Factors associated with AI adoption for ophthalmology analyzed with multivariable logistic regression random forest machine learning.ResultsOne thousand one hundred seventy-six ophthalmologists from 70 countries participated with a response rate ranging from 78.8 to 85.8% per question. Ophthalmologists were more willing to use AI as clinical assistive tools (88.1%, n = 890/1,010) especially those with over 20 years' experience (OR 3.70, 95% CI: 1.10–12.5, p = 0.035), as compared to clinical decision support tools (78.8%, n = 796/1,010) or diagnostic tools (64.5%, n = 651). A majority of Ophthalmologists felt that AI is most relevant to DR (78.2%), followed by glaucoma (70.7%), AMD (66.8%), and cataract (51.4%) detection. Many participants were confident their roles will not be replaced (68.2%, n = 632/927), and felt COVID-19 catalyzed willingness to adopt AI (80.9%, n = 750/927). Common barriers to implementation include medical liability from errors (72.5%, n = 672/927) whereas enablers include improving access (94.5%, n = 876/927). Machine learning modeling predicted acceptance from participant demographics with moderate to high accuracy, and area under the receiver operating curves of 0.63–0.83.ConclusionOphthalmologists are receptive to adopting AI as assistive tools for DR, glaucoma, and AMD. Furthermore, ML is a useful method that can be applied to evaluate predictive factors on clinical qualitative questionnaires. This study outlines actionable insights for future research and facilitation interventions to drive adoption and operationalization of AI tools for Ophthalmology.

Abstract One‐in‐four ophthalmology trials are single‐armed, which poses challenges to their interpretation. We demonstrate how real‐world cohorts used as external/synthetic control arms can contextualize such trials. We herein emulated a target trial on the intention‐to‐treat efficacy of off‐label bevacizumab (q6w) pro re nata relative to fixed‐interval aflibercept (q8w) for improving week 54 visual acuity of eyes affected by neovascular age‐related macular degeneration. The bevacizumab arm (n = 65) was taken from the ABC randomized controlled trial. A total of 4,471 aflibercept‐treated eyes aligning with the ABC trial eligibility were identified from electronic health records and synthetic control arms were created by emulating randomization conditional on age, sex, and baseline visual read via exact matching and propensity score methods. We undertook an inferiority analysis on mean difference at 54 weeks; outcomes regression on achieving a change in visual acuity of greater than or equal to 15, greater than or equal to 10, and less than or equal to −15 Early Treatment Diabetic Retinopathy (ETDRS) letters at week 54; and a time‐to‐event analysis on achieving a change in visual acuity of greater than or equal to 15, greater than or equal to 10, and less than or equal to −15 ETDRS letters by week 54. The findings suggest off‐label bevacizumab to be neither inferior nor superior to licensed aflibercept. Our study highlights how real‐world cohorts representing the counterfactual intervention could aid the interpretation of single‐armed trials when analyzed in accord to the target trial framework. Study Highlights WHAT IS THE CURRENT KNOWLEDGE ON THE TOPIC? One‐in‐four randomized controlled trials in ophthalmology are single‐armed, which poses challenges for interpreting their efficacy relative to standard of care. Recent conceptual advances in the methods of causal inference and in the emulation of target trials suggests that the standard‐of‐care arms representing the counterfactual intervention can be approximated with observational data. WHAT QUESTION DID THIS STUDY ADDRESS? How real‐world cohorts representing the counterfactual intervention can aid the interpretation of single‐armed ophthalmological trials. WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE? Our study highlights how real‐world cohorts representing the counterfactual intervention could aid the interpretation of single‐armed ophthalmological trials when undertaken in accord with the target trial framework. HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE? External counterfactual arms could reduce the time and cost to reach potential regulatory approval.

Purpose: To investigate the incidence, seasonal variation, and differences among age, sex, and race for rhegmatogenous retinal detachment (RRD) repair, retinal break (RB) treatment, and posterior vitreous detachment (PVD) in the Intelligent Research in Sight (IRIS) Registry. Design: Retrospective database study. Participants: Patients in the IRIS Registry who underwent RRD repair, RB treatment, or cataract surgery (CS) based on Current Procedural Terminology codes and PVD diagnosis based on International Classification of Diseases, Ninth and Tenth Revision, codes. Methods: Daily incidence rates were defined as the ratio of patients who underwent RRD repair or RB treatment and patients with a diagnosis of PVD to the total number of patients followed on a given day within the IRIS Registry. The CS group was included as a comparison for seasonal variation. Rates were stratified by decade of life, sex, and race. Main Outcome Measures: Time series trends for incidence rates of RRD, RB, and PVD. Results: A total of 7 115 774 patients received a diagnosis of incident PVD, 237 646 patients underwent RRD repair, and 359 022 patients underwent RB treatment. Also included were 5 940 448 patients who underwent CS. The mean daily incidence for RRD repair, RB treatment, PVD diagnosis, and CS were 0.46 per 100 000 patients, 0.70 per 100 000 patients, 13.90 per 100 000 patients, and 11.80 per 100 000 patients, respectively. Men showed higher incidence of RRD repair and RB treatment than women, whereas women showed higher incidence of PVD diagnosis. Rhegmatogenous retinal detachment incidence was higher in White people compared with other races. Seasonal decreases in PVD, RB treatment, RRD repair, and CS corresponded to national holidays, with larger decreases in winter months. Kaplan-Meier estimates showed that RRD repair and RB treatment typically occurred within 60 days of PVD diagnosis. Conclusions: Within the IRIS Registry, the highest incidence of RRD was in the 6th and 7th decade of life. There was a higher incidence of RRD repair and RB treatment in men compared with women. The seasonal variation associated with national holidays was less pronounced for RRD repair and RB treatment.

Maloca et al. implement convolutional neural network (CNN) to automatically segment OCT images obtained from cynomolgus monkeys. The results are compared to annotations generated by human graders. The ambiguity in ground truth had noteworthy impact on machine learning results, which could be visualized.

In response to the COVID-19 global pandemic, recent research has proposed creating deep learning based models that use chest radiographs (CXRs) in a variety of clinical tasks to help manage the crisis. However, the size of existing datasets of CXRs from COVID-19+ patients are relatively small, and researchers often pool CXR data from multiple sources, for example, using different x-ray machines in various patient populations under different clinical scenarios. Deep learning models trained on such datasets have been shown to overfit to erroneous features instead of learning pulmonary characteristics in a phenomenon known as shortcut learning. We propose adding feature disentanglement to the training process. This technique forces the models to identify pulmonary features from the images and penalizes them for learning features that can discriminate between the original datasets that the images come from. We find that models trained in this way indeed have better generalization performance on unseen data; in the best case we found that it improved AUC by 0.13 on held out data. We further find that this outperforms masking out non-lung parts of the CXRs and performing histogram equalization, both of which are recently proposed methods for removing biases in CXR datasets.

Background: In many low and middle-income countries (LMICs), timely access to emergency healthcare services is limited. In urban settings, traffic can have a significant impact on travel time, leading to life-threatening delays for time-sensitive injuries and medical emergencies. In this study, we examined travel times to hospitals in Nairobi, Kenya, one of the largest and most congested cities in the developing world. Methods: We used a network approach to estimate average minimum travel times to different types of hospitals (e.g. ownership and level of care) in Nairobi under both congested and uncongested traffic conditions. We also examined the correlation between travel time and socioeconomic status. Results: We estimate the average minimum travel time during uncongested traffic conditions to any level 4 health facility (primary hospitals) or above in Nairobi to be 4.5 min (IQR 2.5–6.1). Traffic added an average of 9.0 min (a 200% increase). In uncongested conditions, we estimate an average travel time of 7.9 min (IQR 5.1–10.4) to level 5 facilities (secondary hospitals) and 11.6 min (IQR 8.5–14.2) to Kenyatta National Hospital, the only level 6 facility (tertiary hospital) in the country. Traffic congestion added an average of 13.1 and 16.0 min (166% and 138% increase) to travel times to level 5 and level 6 facilities, respectively. For individuals living below the poverty line, we estimate that preferential use of public or faith-based facilities could increase travel time by as much as 65%. Conclusion: Average travel times to health facilities capable of providing emergency care in Nairobi are quite low, but traffic congestion double or triple estimated travel times. Furthermore, we estimate significant disparities in timely access to care for those individuals living under the poverty line who preferentially seek care in public or faith-based facilities. Keywords: Emergency care, Kenya, Traffic, Poverty

Purpose: To evaluate the performance of a federated learning framework for deep neural network-based retinal microvasculature segmentation and referable diabetic retinopathy (RDR) classification using OCT and OCT angiography (OCTA). Design: Retrospective analysis of clinical OCT and OCTA scans of control participants and patients with diabetes. Participants: The 153 OCTA en face images used for microvasculature segmentation were acquired from 4 OCT instruments with fields of view ranging from 2 × 2-mm to 6 × 6-mm. The 700 eyes used for RDR classification consisted of OCTA en face images and structural OCT projections acquired from 2 commercial OCT systems. Methods: OCT angiography images used for microvasculature segmentation were delineated manually and verified by retina experts. Diabetic retinopathy (DR) severity was evaluated by retinal specialists and was condensed into 2 classes: non-RDR and RDR. The federated learning configuration was demonstrated via simulation using 4 clients for microvasculature segmentation and was compared with other collaborative training methods. Subsequently, federated learning was applied over multiple institutions for RDR classification and was compared with models trained and tested on data from the same institution (internal models) and different institutions (external models). Main Outcome Measures: For microvasculature segmentation, we measured the accuracy and Dice similarity coefficient (DSC). For severity classification, we measured accuracy, area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve, balanced accuracy, F1 score, sensitivity, and specificity. Results: For both applications, federated learning achieved similar performance as internal models. Specifically, for microvasculature segmentation, the federated learning model achieved similar performance (mean DSC across all test sets, 0.793) as models trained on a fully centralized dataset (mean DSC, 0.807). For RDR classification, federated learning achieved a mean AUROC of 0.954 and 0.960; the internal models attained a mean AUROC of 0.956 and 0.973. Similar results are reflected in the other calculated evaluation metrics. Conclusions: Federated learning showed similar results to traditional deep learning in both applications of segmentation and classification, while maintaining data privacy. Evaluation metrics highlight the potential of collaborative learning for increasing domain diversity and the generalizability of models used for the classification of OCT data.

Purpose: To evaluate whether cataract surgery is associated with decreased risks of central retinal vein occlusion (CRVO) or branch retinal vein occlusion (BRVO) development using the American Academy of Ophthalmology Intelligent Research in Sight (IRIS®) Registry. Design: Retrospective database study of the IRIS Registry data. Participants: Patients in the IRIS Registry who underwent cataract surgery and 1:1 matched control participants from the IRIS Registry using a decision tree classifier as a propensity model. Methods: Control and treatment groups initially were selected using Current Procedural Terminology codes for uncomplicated cataract surgery and other straightforward criteria. To accomplish treatment–control matching, a decision tree classifier was trained to classify patients as treatment versus control based on a set of chosen predictors for treatment, where best-corrected visual acuity and age were the most important predictors. Treatment and control participants subsequently were matched using the classifier, the visit dates, and the identifications of the practice. Cox regression was performed on the matched groups to measure the hazard ratio (HR) of retinal vein occlusion development adjusted for age, sex, race, primary insurance type, and previous diagnosis of diabetic retinopathy (DR), glaucoma, and narrow angles. Main Outcome Measure: The HR of retinal vein occlusion developing in patients who underwent cataract surgery compared with matched control participants. Results: The HRs for CRVO and BRVO developing in patients who underwent cataract surgery compared with matched control participants who did not during the first year after either cataract surgery or baseline visit were 1.26 [95% confidence interval [CI], 1.16–1.38; P < 0.001] and 1.27 [95% CI, 1.19–1.36; P < 0.001], respectively, after controlling for age, sex, race, insurance, and history of DR, glaucoma, and narrow angles. Diabetic retinopathy was the strongest predictor associated with CRVO (2.79 [95% CI, 2.43–3.20; P < 0.001]) and BRVO (2.35 [95% CI, 2.09–2.64; P < 0.001]) development after cataract surgery. Conclusions: Cataract surgery is associated with a small increase in risk of retinal vein occlusions within the first year; however, the incidence is low and likely not clinically significant.

Objective This study aimed to investigate emerging trends and increasing costs in the National Residency Matching Program (NRMP) and San Francisco Residency and Fellowship Match Services (SF Match) associated with the current applicant/program Gale–Shapley-type matching algorithms. Design A longitudinal observational study of behavioral trends in national residency matching systems with modeling of match results with alternative parameters. Patients and Methods We analyzed publicly available data from the SF Match and NRMP websites from 1985 to 2020 for trends in the total number of applicants and available positions, as well the average number of applications and interviews per applicant for multiple specialties. To understand these trends and the algorithms' effect on the residency programs and applicants, we analyzed anonymized rank list and match data for ophthalmology from the SF Match between 2011 and 2019. Match results using current match parameters, as well as under conditions in which applicant and/or program rank lists were truncated with finalized rank lists, were analyzed. Results Both the number of applications and length of programs' rank lists have increased steadily throughout residency programs, particularly those with competitive specialties. Capping student rank lists at seven programs, or less than 80% of the average 8.9 programs currently ranked, results in a 0.71% decrease in the total number of positions filled. Similarly, capping program rank lists at seven applicants per spot, or less than 60% of the average 11.5 applicants ranked per spot, results in a 5% decrease in the total number of positions filled. Conclusion While the number of ophthalmology positions in the United States has increased only modestly, the number of applications under consideration has increased substantially over the past two decades. The current study suggests that both programs and applicants rank more choices than are required for a nearly complete and stable match, creating excess cost and work for both applicants and programs. “Stable-marriage” type algorithms induce applicants and programs to rank as many counterparties as possible to maximize individual chances of optimizing the match.

Purpose To quantify differences in the age, gender, race, and clinical complexity of Medicare beneficiaries treated by ophthalmologists and optometrists in each of the United States. Design Cross-sectional study based on publicly accessible Medicare payment and utilization data from 2012 through 2017. Methods For each ophthalmic and optometric provider, demographic information of treated Medicare beneficiaries was obtained from the Medicare Provider Utilization and Payment Data from the Centers for Medicare and Medicaid Services (CMS) for the years 2012 through 2017. Clinical complexity was defined using CMS Hierarchical Condition Category (HCC) coding. Results From 2012 through 2017, ophthalmologists in every state treated statistically significantly older beneficiaries, with the greatest difference (4.99 years in 2014) between provider groups seen in Rhode Island. In most states there was no gender difference among patients treated by the providers but in 46 states ophthalmologists saw a more racially diverse group of beneficiaries. HCC risk score analysis demonstrated that ophthalmologists in all 50 states saw more medically complex beneficiaries and the differences were statistically significant in 47 states throughout all six years. Conclusions Although there are regional variations in the characteristics of patients treated by ophthalmologists and optometrists, ophthalmologists throughout the United States manage older, more racially diverse, and more medically complex Medicare beneficiaries.

Purpose. CAPTCHA (completely automated public turing test to tell computers and humans apart) was designed as a spam prevention test. In patients with visual impairment, completion of this task has been assumed to be difficult; but to date, no study has proven this to be true. As visual function is not well measured by Snellen visual acuity (VA) alone, we theorized that CAPTCHA performance may provide additional information on macular disease-related visual dysfunction. Methods. This was designed as a pilot study. Active disease was defined as the presence of either intraretinal fluid (IRF) or subretinal fluid (SRF) on spectral-domain optical coherence tomography. CAPTCHA performance was tested using 10 prompts. In addition, near and distance VA, contrast sensitivity, and reading speed were measured. Visual acuity matched pseudophakic patients were used as controls. Primary outcome measures were average edit distance and percent of correct responses. Results. 70 patients were recruited: 33 with active macular disease and 37 control subjects. Contrast sensitivity was found to be significantly different in both the IRF (p

Purpose. To evaluate the feasibility of using Mechanical Turk as a massively parallel platform to perform manual segmentations of macular spectral domain optical coherence tomography (SD-OCT) images using a MapReduce framework. Methods. A macular SD-OCT volume of 61 slice images was map-distributed to Amazon Mechanical Turk. Each Human Intelligence Task was set to $0.01 and required the user to draw five lines to outline the sublayers of the retinal OCT image after being shown example images. Each image was submitted twice for segmentation, and interrater reliability was calculated. The interface was created using custom HTML5 and JavaScript code, and data analysis was performed using R. An automated pipeline was developed to handle the map and reduce steps of the framework. Results. More than 93,500 data points were collected using this framework for the 61 images submitted. Pearson’s correlation of interrater reliability was 0.995 (p