Pivotal trial of a deep-learning-based retinal biomarker (Reti-CVD) in the prediction of cardiovascular disease: data from CMERC-HI.

OBJECTIVE: The potential of using retinal images as a biomarker of cardiovascular disease (CVD) risk has gained significant attention, but regulatory approval of such artificial intelligence (AI) algorithms is lacking. In this regulated pivotal trial, we validated the efficacy of Reti-CVD, an AI-Software as a Medical Device (AI-SaMD), that utilizes retinal images to stratify CVD risk. MATERIALS AND METHODS: In this retrospective study, we used data from the Cardiovascular and Metabolic Diseases Etiology Research Center-High Risk (CMERC-HI) Cohort. Cox proportional hazard model was used to estimate hazard ratio (HR) trend across the 3-tier CVD risk groups (low-, moderate-, and high-risk) according to Reti-CVD in prediction of CVD events. The cardiac computed tomography-measured coronary artery calcium (CAC), carotid intima-media thickness (CIMT), and brachial-ankle pulse wave velocity (baPWV) were compared to Reti-CVD. RESULTS: A total of 1106 participants were included, with 33 (3.0%) participants experiencing CVD events over 5 years; the Reti-CVD-defined risk groups (low, moderate, and high) were significantly associated with increased CVD risk (HR trend, 2.02; 95% CI, 1.26-3.24). When all variables of Reti-CVD, CAC, CIMT, baPWV, and other traditional risk factors were incorporated into one Cox model, the Reti-CVD risk groups were only significantly associated with increased CVD risk (HR = 2.40 [0.82-7.03] in moderate risk and HR = 3.56 [1.34-9.51] in high risk using low-risk as a reference). DISCUSSION: This regulated pivotal study validated an AI-SaMD, retinal image-based, personalized CVD risk scoring system (Reti-CVD). CONCLUSION: These results led the Korean regulatory body to authorize Reti-CVD.

Efficacy of deep learning-based artificial intelligence models in screening and referring patients with diabetic retinopathy and glaucoma.

Purpose: To analyze the efficacy of a deep learning (DL)-based artificial intelligence (AI)-based algorithm in detecting the presence of diabetic retinopathy (DR) and glaucoma suspect as compared to the diagnosis by specialists secondarily to explore whether the use of this algorithm can reduce the cross-referral in three clinical settings: a diabetologist clinic, retina clinic, and glaucoma clinic. Methods: This is a prospective observational study. Patients between 35 and 65 years of age were recruited from glaucoma and retina clinics at a tertiary eye care hospital and a physician's clinic. Non-mydriatic fundus photography was performed according to the disease-specific protocols. These images were graded by the AI system and specialist graders and comparatively analyzed. Results: Out of 1085 patients, 362 were seen at glaucoma clinics, 341 were seen at retina clinics, and 382 were seen at physician clinics. The kappa agreement between AI and the glaucoma grader was 85% [95% confidence interval (CI): 77.55-92.45%], and retina grading had 91.90% (95% CI: 87.78-96.02%). The retina grader from the glaucoma clinic had 85% agreement, and the glaucoma grader from the retina clinic had 73% agreement. The sensitivity and specificity of AI glaucoma grading were 79.37% (95% CI: 67.30-88.53%) and 99.45 (95% CI: 98.03-99.93), respectively; DR grading had 83.33% (95 CI: 51.59-97.91) and 98.86 (95% CI: 97.35-99.63). The cross-referral accuracy of DR and glaucoma was 89.57% and 95.43%, respectively. Conclusion: DL-based AI systems showed high sensitivity and specificity in both patients with DR and glaucoma; also, there was a good agreement between the specialist graders and the AI system.

Cardiovascular disease risk assessment using a deep-learning-based retinal biomarker: a comparison with existing risk scores.

Aims: This study aims to evaluate the ability of a deep-learning-based cardiovascular disease (CVD) retinal biomarker, Reti-CVD, to identify individuals with intermediate- and high-risk for CVD. Methods and results: We defined the intermediate- and high-risk groups according to Pooled Cohort Equation (PCE), QRISK3, and modified Framingham Risk Score (FRS). Reti-CVD's prediction was compared to the number of individuals identified as intermediate- and high-risk according to standard CVD risk assessment tools, and sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated to assess the results. In the UK Biobank, among 48 260 participants, 20 643 (42.8%) and 7192 (14.9%) were classified into the intermediate- and high-risk groups according to PCE, and QRISK3, respectively. In the Singapore Epidemiology of Eye Diseases study, among 6810 participants, 3799 (55.8%) were classified as intermediate- and high-risk group according to modified FRS. Reti-CVD identified PCE-based intermediate- and high-risk groups with a sensitivity, specificity, PPV, and NPV of 82.7%, 87.6%, 86.5%, and 84.0%, respectively. Reti-CVD identified QRISK3-based intermediate- and high-risk groups with a sensitivity, specificity, PPV, and NPV of 82.6%, 85.5%, 49.9%, and 96.6%, respectively. Reti-CVD identified intermediate- and high-risk groups according to the modified FRS with a sensitivity, specificity, PPV, and NPV of 82.1%, 80.6%, 76.4%, and 85.5%, respectively. Conclusion: The retinal photograph biomarker (Reti-CVD) was able to identify individuals with intermediate and high-risk for CVD, in accordance with existing risk assessment tools.

Non-invasive chronic kidney disease risk stratification tool derived from retina-based deep learning and clinical factors.

Despite the importance of preventing chronic kidney disease (CKD), predicting high-risk patients who require active intervention is challenging, especially in people with preserved kidney function. In this study, a predictive risk score for CKD (Reti-CKD score) was derived from a deep learning algorithm using retinal photographs. The performance of the Reti-CKD score was verified using two longitudinal cohorts of the UK Biobank and Korean Diabetic Cohort. Validation was done in people with preserved kidney function, excluding individuals with eGFR <90 mL/min/1.73 m2 or proteinuria at baseline. In the UK Biobank, 720/30,477 (2.4%) participants had CKD events during the 10.8-year follow-up period. In the Korean Diabetic Cohort, 206/5014 (4.1%) had CKD events during the 6.1-year follow-up period. When the validation cohorts were divided into quartiles of Reti-CKD score, the hazard ratios for CKD development were 3.68 (95% Confidence Interval [CI], 2.88-4.41) in the UK Biobank and 9.36 (5.26-16.67) in the Korean Diabetic Cohort in the highest quartile compared to the lowest. The Reti-CKD score, compared to eGFR based methods, showed a superior concordance index for predicting CKD incidence, with a delta of 0.020 (95% CI, 0.011-0.029) in the UK Biobank and 0.024 (95% CI, 0.002-0.046) in the Korean Diabetic Cohort. In people with preserved kidney function, the Reti-CKD score effectively stratifies future CKD risk with greater performance than conventional eGFR-based methods.

Validation of a deep-learning-based retinal biomarker (Reti-CVD) in the prediction of cardiovascular disease: data from UK Biobank.

BACKGROUND: Currently in the United Kingdom, cardiovascular disease (CVD) risk assessment is based on the QRISK3 score, in which 10% 10-year CVD risk indicates clinical intervention. However, this benchmark has limited efficacy in clinical practice and the need for a more simple, non-invasive risk stratification tool is necessary. Retinal photography is becoming increasingly acceptable as a non-invasive imaging tool for CVD. Previously, we developed a novel CVD risk stratification system based on retinal photographs predicting future CVD risk. This study aims to further validate our biomarker, Reti-CVD, (1) to detect risk group of ≥ 10% in 10-year CVD risk and (2) enhance risk assessment in individuals with QRISK3 of 7.5-10% (termed as borderline-QRISK3 group) using the UK Biobank. METHODS: Reti-CVD scores were calculated and stratified into three risk groups based on optimized cut-off values from the UK Biobank. We used Cox proportional-hazards models to evaluate the ability of Reti-CVD to predict CVD events in the general population. C-statistics was used to assess the prognostic value of adding Reti-CVD to QRISK3 in borderline-QRISK3 group and three vulnerable subgroups. RESULTS: Among 48,260 participants with no history of CVD, 6.3% had CVD events during the 11-year follow-up. Reti-CVD was associated with an increased risk of CVD (adjusted hazard ratio [HR] 1.41; 95% confidence interval [CI], 1.30-1.52) with a 13.1% (95% CI, 11.7-14.6%) 10-year CVD risk in Reti-CVD-high-risk group. The 10-year CVD risk of the borderline-QRISK3 group was greater than 10% in Reti-CVD-high-risk group (11.5% in non-statin cohort [n = 45,473], 11.5% in stage 1 hypertension cohort [n = 11,966], and 14.2% in middle-aged cohort [n = 38,941]). C statistics increased by 0.014 (0.010-0.017) in non-statin cohort, 0.013 (0.007-0.019) in stage 1 hypertension cohort, and 0.023 (0.018-0.029) in middle-aged cohort for CVD event prediction after adding Reti-CVD to QRISK3. CONCLUSIONS: Reti-CVD has the potential to identify individuals with ≥ 10% 10-year CVD risk who are likely to benefit from earlier preventative CVD interventions. For borderline-QRISK3 individuals with 10-year CVD risk between 7.5 and 10%, Reti-CVD could be used as a risk enhancer tool to help improve discernment accuracy, especially in adult groups that may be pre-disposed to CVD.

Retinal photograph-based deep learning predicts biological age, and stratifies morbidity and mortality risk.

BACKGROUND: ageing is an important risk factor for a variety of human pathologies. Biological age (BA) may better capture ageing-related physiological changes compared with chronological age (CA). OBJECTIVE: we developed a deep learning (DL) algorithm to predict BA based on retinal photographs and evaluated the performance of our new ageing marker in the risk stratification of mortality and major morbidity in general populations. METHODS: we first trained a DL algorithm using 129,236 retinal photographs from 40,480 participants in the Korean Health Screening study to predict the probability of age being ≥65 years ('RetiAGE') and then evaluated the ability of RetiAGE to stratify the risk of mortality and major morbidity among 56,301 participants in the UK Biobank. Cox proportional hazards model was used to estimate the hazard ratios (HRs). RESULTS: in the UK Biobank, over a 10-year follow up, 2,236 (4.0%) died; of them, 636 (28.4%) were due to cardiovascular diseases (CVDs) and 1,276 (57.1%) due to cancers. Compared with the participants in the RetiAGE first quartile, those in the RetiAGE fourth quartile had a 67% higher risk of 10-year all-cause mortality (HR = 1.67 [1.42-1.95]), a 142% higher risk of CVD mortality (HR = 2.42 [1.69-3.48]) and a 60% higher risk of cancer mortality (HR = 1.60 [1.31-1.96]), independent of CA and established ageing phenotypic biomarkers. Likewise, compared with the first quartile group, the risk of CVD and cancer events in the fourth quartile group increased by 39% (HR = 1.39 [1.14-1.69]) and 18% (HR = 1.18 [1.10-1.26]), respectively. The best discrimination ability for RetiAGE alone was found for CVD mortality (c-index = 0.70, sensitivity = 0.76, specificity = 0.55). Furthermore, adding RetiAGE increased the discrimination ability of the model beyond CA and phenotypic biomarkers (increment in c-index between 1 and 2%). CONCLUSIONS: the DL-derived RetiAGE provides a novel, alternative approach to measure ageing.

Computer-aided detection and abnormality score for the outer retinal layer in optical coherence tomography.

BACKGROUND: To develop computer-aided detection (CADe) of ORL abnormalities in the retinal pigmented epithelium, interdigitation zone and ellipsoid zone via optical coherence tomography (OCT). METHODS: In this retrospective study, healthy participants with normal ORL, and patients with abnormality of ORL including choroidal neovascularisation (CNV) or retinitis pigmentosa (RP) were included. First, an automatic segmentation deep learning (DL) algorithm, CADe, was developed for the three outer retinal layers using 120 handcraft masks of ORL. This automatic segmentation algorithm generated 4000 segmentations, which included 2000 images with normal ORL and 2000 (1000 CNV and 1000 RP) images with focal or wide defects in ORL. Second, based on the automatically generated segmentation images, a binary classifier (normal vs abnormal) was developed. Results were evaluated by area under the receiver operating characteristic curve (AUC). RESULTS: The DL algorithm achieved an AUC of 0.984 (95% CI 0.976 to 0.993) for individual image evaluation in the internal test set of 797 images. In addition, performance analysis of a publicly available external test set (n=968) had an AUC of 0.957 (95% CI 0.944 to 0.970) and a second clinical external test set (n=1124) had an AUC of 0.978 (95% CI 0.970 to 0.986). Moreover, the CADe highlighted well normal parts of ORL and omitted highlights in abnormal ORLs of CNV and RP. CONCLUSION: The CADe can use OCT images to segment ORL and differentiate between normal ORL and abnormal ORL. The CADe classifier also performs visualisation and may aid future physician diagnosis and clinical applications.

Automatic segmentation of corneal deposits from corneal stromal dystrophy images via deep learning.

BACKGROUND: Granular dystrophy is the most common stromal dystrophy. To perform automated segmentation of corneal stromal deposits, we trained and tested a deep learning (DL) algorithm from patients with corneal stromal dystrophy and compared its performance with human segmentation. METHODS: In this retrospective cross-sectional study, we included slit-lamp photographs by sclerotic scatter from patients with corneal stromal dystrophy and real-world slit-lamp photographs via various techniques (diffuse illumination, tangential illumination, and sclerotic scatter). Our data set included 1007 slit-lamp photographs of semi-automatically generated handcraft masks on granular and linear lesions from corneal stromal dystrophy patients (806 for the training set and 201 for test set). For external test (140 photographs), we applied the DL algorithm and compared between automated and human segmentation. For performance, we estimated the intersection of union (IoU), global accuracy, and boundary F1 (BF) score for segmentation. RESULTS: In 201 internal test set, IoU, global accuracy, and BF score with 95 % confidence Interval were 0.81 (0.79-0.82), 0.99 (0.98-0.99), and 0.93 (0.92-0.95), respectively. In 140 heterogenous external test set as a real-world data, those were 0.64 (0.61-0.67), 0.95 (0.94-0.96), and 0.70 (0.64-0.76) via DL algorithm and 0.56 (0.51-0.61), 0.95 (0.94-0.96), and 0.70 (0.65-0.74) via human rater, respectively. CONCLUSIONS: We developed an automated segmentation DL algorithm for corneal stromal deposits in patients with corneal stromal dystrophy. Segmentation on corneal deposits was accurate via the DL algorithm in the well-controlled dataset and showed reasonable performance in a real-world setting. We suggest this automatic segmentation of corneal deposits helps to monitor the disease and can evaluate possible new treatments.

Gender Prediction for a Multiethnic Population via Deep Learning Across Different Retinal Fundus Photograph Fields: Retrospective Cross-sectional Study.

BACKGROUND: Deep learning algorithms have been built for the detection of systemic and eye diseases based on fundus photographs. The retina possesses features that can be affected by gender differences, and the extent to which these features are captured via photography differs depending on the retinal image field. OBJECTIVE: We aimed to compare deep learning algorithms' performance in predicting gender based on different fields of fundus photographs (optic disc-centered, macula-centered, and peripheral fields). METHODS: This retrospective cross-sectional study included 172,170 fundus photographs of 9956 adults aged ≥40 years from the Singapore Epidemiology of Eye Diseases Study. Optic disc-centered, macula-centered, and peripheral field fundus images were included in this study as input data for a deep learning model for gender prediction. Performance was estimated at the individual level and image level. Receiver operating characteristic curves for binary classification were calculated. RESULTS: The deep learning algorithms predicted gender with an area under the receiver operating characteristic curve (AUC) of 0.94 at the individual level and an AUC of 0.87 at the image level. Across the three image field types, the best performance was seen when using optic disc-centered field images (younger subgroups: AUC=0.91; older subgroups: AUC=0.86), and algorithms that used peripheral field images had the lowest performance (younger subgroups: AUC=0.85; older subgroups: AUC=0.76). Across the three ethnic subgroups, algorithm performance was lowest in the Indian subgroup (AUC=0.88) compared to that in the Malay (AUC=0.91) and Chinese (AUC=0.91) subgroups when the algorithms were tested on optic disc-centered images. Algorithms' performance in gender prediction at the image level was better in younger subgroups (aged <65 years; AUC=0.89) than in older subgroups (aged ≥65 years; AUC=0.82). CONCLUSIONS: We confirmed that gender among the Asian population can be predicted with fundus photographs by using deep learning, and our algorithms' performance in terms of gender prediction differed according to the field of fundus photographs, age subgroups, and ethnic groups. Our work provides a further understanding of using deep learning models for the prediction of gender-related diseases. Further validation of our findings is still needed.

Development of a Web-Based Ensemble Machine Learning Application to Select the Optimal Size of Posterior Chamber Phakic Intraocular Lens.

Purpose: Selecting the optimal lens size by predicting the postoperative vault can reduce complications after implantation of an implantable collamer lens with a central-hole (ICL with KS-aquaport). We built a web-based machine learning application that incorporated clinical measurements to predict the postoperative ICL vault and select the optimal ICL size. Methods: We applied the stacking ensemble technique based on eXtreme Gradient Boosting (XGBoost) and a light gradient boosting machine to pre-operative ocular data from two eye centers to predict the postoperative vault. We assigned the Korean patient data to a training (N = 2756 eyes) and internal validation (N = 693 eyes) datasets (prospective validation). Japanese patient data (N = 290 eyes) were used as an independent external dataset from different centers to validate the model. Results: We developed an ensemble model that showed statistically better performance with a lower mean absolute error for ICL vault prediction (106.88 µm and 143.69 µm in the internal and external validation, respectively) than the other machine learning techniques and the classic ICL sizing methods did when applied to both validation datasets. Considering the lens size selection accuracy, our proposed method showed the best performance for both reference datasets (75.9% and 67.4% in the internal and external validation, respectively). Conclusions: Applying the ensemble approach to a large dataset of patients who underwent ICL implantation resulted in a more accurate prediction of vault size and selection of the optimal ICL size. Translational Relevance: We developed a web-based application for ICL sizing to facilitate the use of machine learning calculators for clinicians.

Deep-learning-based cardiovascular risk stratification using coronary artery calcium scores predicted from retinal photographs.

BACKGROUND: Coronary artery calcium (CAC) score is a clinically validated marker of cardiovascular disease risk. We developed and validated a novel cardiovascular risk stratification system based on deep-learning-predicted CAC from retinal photographs. METHODS: We used 216 152 retinal photographs from five datasets from South Korea, Singapore, and the UK to train and validate the algorithms. First, using one dataset from a South Korean health-screening centre, we trained a deep-learning algorithm to predict the probability of the presence of CAC (ie, deep-learning retinal CAC score, RetiCAC). We stratified RetiCAC scores into tertiles and used Cox proportional hazards models to evaluate the ability of RetiCAC to predict cardiovascular events based on external test sets from South Korea, Singapore, and the UK Biobank. We evaluated the incremental values of RetiCAC when added to the Pooled Cohort Equation (PCE) for participants in the UK Biobank. FINDINGS: RetiCAC outperformed all single clinical parameter models in predicting the presence of CAC (area under the receiver operating characteristic curve of 0·742, 95% CI 0·732-0·753). Among the 527 participants in the South Korean clinical cohort, 33 (6·3%) had cardiovascular events during the 5-year follow-up. When compared with the current CAC risk stratification (0, >0-100, and >100), the three-strata RetiCAC showed comparable prognostic performance with a concordance index of 0·71. In the Singapore population-based cohort (n=8551), 310 (3·6%) participants had fatal cardiovascular events over 10 years, and the three-strata RetiCAC was significantly associated with increased risk of fatal cardiovascular events (hazard ratio [HR] trend 1·33, 95% CI 1·04-1·71). In the UK Biobank (n=47 679), 337 (0·7%) participants had fatal cardiovascular events over 10 years. When added to the PCE, the three-strata RetiCAC improved cardiovascular risk stratification in the intermediate-risk group (HR trend 1·28, 95% CI 1·07-1·54) and borderline-risk group (1·62, 1·04-2·54), and the continuous net reclassification index was 0·261 (95% CI 0·124-0·364). INTERPRETATION: A deep learning and retinal photograph-derived CAC score is comparable to CT scan-measured CAC in predicting cardiovascular events, and improves on current risk stratification approaches for cardiovascular disease events. These data suggest retinal photograph-based deep learning has the potential to be used as an alternative measure of CAC, especially in low-resource settings. FUNDING: Yonsei University College of Medicine; Ministry of Health and Welfare, Korea Institute for Advancement of Technology, South Korea; Agency for Science, Technology, and Research; and National Medical Research Council, Singapore.

Detection of features associated with neovascular age-related macular degeneration in ethnically distinct data sets by an optical coherence tomography: trained deep learning algorithm.

BACKGROUND: The ability of deep learning (DL) algorithms to identify eyes with neovascular age-related macular degeneration (nAMD) from optical coherence tomography (OCT) scans has been previously established. We herewith evaluate the ability of a DL model, showing excellent performance on a Korean data set, to generalse onto an American data set despite ethnic differences. In addition, expert graders were surveyed to verify if the DL model was appropriately identifying lesions indicative of nAMD on the OCT scans. METHODS: Model development data set-12 247 OCT scans from South Korea; external validation data set-91 509 OCT scans from Washington, USA. In both data sets, normal eyes or eyes with nAMD were included. After internal testing, the algorithm was sent to the University of Washington, USA, for external validation. Area under the receiver operating characteristic curve (AUC) and precision-recall curve (AUPRC) were calculated. For model explanation, saliency maps were generated using Guided GradCAM. RESULTS: On external validation, AUC and AUPRC remained high at 0.952 (95% CI 0.942 to 0.962) and 0.891 (95% CI 0.875 to 0.908) at the individual level. Saliency maps showed that in normal OCT scans, the fovea was the main area of interest; in nAMD OCT scans, the appropriate pathological features were areas of model interest. Survey of 10 retina specialists confirmed this. CONCLUSION: Our DL algorithm exhibited high performance for nAMD identification in a Korean population, and generalised well to an ethnically distinct, American population. The model correctly focused on the differences within the macular area to extract features associated with nAMD.

Prediction of systemic biomarkers from retinal photographs: development and validation of deep-learning algorithms.

BACKGROUND: The application of deep learning to retinal photographs has yielded promising results in predicting age, sex, blood pressure, and haematological parameters. However, the broader applicability of retinal photograph-based deep learning for predicting other systemic biomarkers and the generalisability of this approach to various populations remains unexplored. METHODS: With use of 236 257 retinal photographs from seven diverse Asian and European cohorts (two health screening centres in South Korea, the Beijing Eye Study, three cohorts in the Singapore Epidemiology of Eye Diseases study, and the UK Biobank), we evaluated the capacities of 47 deep-learning algorithms to predict 47 systemic biomarkers as outcome variables, including demographic factors (age and sex); body composition measurements; blood pressure; haematological parameters; lipid profiles; biochemical measures; biomarkers related to liver function, thyroid function, kidney function, and inflammation; and diabetes. The standard neural network architecture of VGG16 was adopted for model development. FINDINGS: In addition to previously reported systemic biomarkers, we showed quantification of body composition indices (muscle mass, height, and bodyweight) and creatinine from retinal photographs. Body muscle mass could be predicted with an R2 of 0·52 (95% CI 0·51-0·53) in the internal test set, and of 0·33 (0·30-0·35) in one external test set with muscle mass measurement available. The R2 value for the prediction of height was 0·42 (0·40-0·43), of bodyweight was 0·36 (0·34-0·37), and of creatinine was 0·38 (0·37-0·40) in the internal test set. However, the performances were poorer in external test sets (with the lowest performance in the European cohort), with R2 values ranging between 0·08 and 0·28 for height, 0·04 and 0·19 for bodyweight, and 0·01 and 0·26 for creatinine. Of the 47 systemic biomarkers, 37 could not be predicted well from retinal photographs via deep learning (R2≤0·14 across all external test sets). INTERPRETATION: Our work provides new insights into the potential use of retinal photographs to predict systemic biomarkers, including body composition indices and serum creatinine, using deep learning in populations with a similar ethnic background. Further evaluations are warranted to validate these findings and evaluate the clinical utility of these algorithms. FUNDING: Agency for Science, Technology, and Research and National Medical Research Council, Singapore; Korea Institute for Advancement of Technology.

Explainable Machine Learning Approach as a Tool to Understand Factors Used to Select the Refractive Surgery Technique on the Expert Level.

Purpose: Recently, laser refractive surgery options, including laser epithelial keratomileusis, laser in situ keratomileusis, and small incision lenticule extraction, successfully improved patients' quality of life. Evidence-based recommendation for an optimal surgery technique is valuable in increasing patient satisfaction. We developed an interpretable multiclass machine learning model that selects the laser surgery option on the expert level. Methods: A multiclass XGBoost model was constructed to classify patients into four categories including laser epithelial keratomileusis, laser in situ keratomileusis, small incision lenticule extraction, and contraindication groups. The analysis included 18,480 subjects who intended to undergo refractive surgery at the B&VIIT Eye center. Training (n = 10,561) and internal validation (n = 2640) were performed using subjects who visited between 2016 and 2017. The model was trained based on clinical decisions of highly experienced experts and ophthalmic measurements. External validation (n = 5279) was conducted using subjects who visited in 2018. The SHapley Additive ex-Planations technique was adopted to explain the output of the XGBoost model. Results: The multiclass XGBoost model exhibited an accuracy of 81.0% and 78.9% when tested on the internal and external validation datasets, respectively. The SHapley Additive ex-Planations explanations for the results were consistent with prior knowledge from ophthalmologists. The explanation from one-versus-one and one-versus-rest XGBoost classifiers was effective for easily understanding users in the multicategorical classification problem. Conclusions: This study suggests an expert-level multiclass machine learning model for selecting the refractive surgery for patients. It also provided a clinical understanding in a multiclass problem based on an explainable artificial intelligence technique. Translational Relevance: Explainable machine learning exhibits a promising future for increasing the practical use of artificial intelligence in ophthalmic clinics.

Deep Learning for Automated Sorting of Retinal Photographs.

PURPOSE: Though the domain of big data and artificial intelligence in health care continues to evolve, there is a lack of systemic methods to improve data quality and streamline the preparation process. To address this, we aimed to develop an automated sorting system (RetiSort) that accurately labels the type and laterality of retinal photographs. DESIGN: Cross-sectional study. PARTICIPANTS: RetiSort was developed with retinal photographs from the Singapore Epidemiology of Eye Diseases (SEED) study. METHODS: The development of RetiSort was composed of 3 steps: 2 deep-learning (DL) algorithms and 1 rule-based classifier. For step 1, a DL algorithm was developed to locate the optic disc, the "landmark feature." For step 2, based on the location of the optic disc derived from step 1, a rule-based classifier was developed to sort retinal photographs into 3 types: macular-centered, optic disc-centered, or related to other fields. Step 2 concurrently distinguished laterality (i.e., the left or right eye) of macular-centered photographs. For step 3, an additional DL algorithm was developed to differentiate the laterality of disc-centered photographs. Via the 3 steps, RetiSort sorted and labeled retinal images into (1) right macular-centered, (2) left macular-centered, (3) right optic disc-centered, (4) left optic disc-centered, and (5) images relating to other fields. Subsequently, the accuracy of RetiSort was evaluated on 5000 randomly selected retinal images from SEED as well as on 3 publicly available image databases (DIARETDB0, HEI-MED, and Drishti-GS). The main outcome measure was the accuracy for sorting of retinal photographs. RESULTS: RetiSort mislabeled 48 out of 5000 retinal images from SEED, representing an overall accuracy of 99.0% (95% confidence interval [CI], 98.7-99.3). In external tests, RetiSort mislabeled 1, 0, and 2 images, respectively, from DIARETDB0, HEI-MED, and Drishti-GS, representing an accuracy of 99.2% (95% CI, 95.8-99.9), 100%, and 98.0% (95% CI, 93.1-99.8), respectively. Saliency maps consistently showed that the DL algorithm in step 3 required pixels in the central left lateral border and optic disc of optic disc-centered retinal photographs to differentiate the laterality. CONCLUSIONS: RetiSort is a highly accurate automated sorting system. It can aid in data preparation and has practical applications in DL research that uses retinal photographs.

Adopting machine learning to automatically identify candidate patients for corneal refractive surgery.

Recently, it has become more important to screen candidates that undergo corneal refractive surgery to prevent complications. Until now, there is still no definitive screening method to confront the possibility of a misdiagnosis. We evaluate the possibilities of machine learning as a clinical decision support to determine the suitability to corneal refractive surgery. A machine learning architecture was built with the aim of identifying candidates combining the large multi-instrument data from patients and clinical decisions of highly experienced experts. Five heterogeneous algorithms were used to predict candidates for surgery. Subsequently, an ensemble classifier was developed to improve the performance. Training (10,561 subjects) and internal validation (2640 subjects) were conducted using subjects who had visited between 2016 and 2017. External validation (5279 subjects) was performed using subjects who had visited in 2018. The best model, i.e., the ensemble classifier, had a high prediction performance with the area under the receiver operating characteristic curves of 0.983 (95% CI, 0.977-0.987) and 0.972 (95% CI, 0.967-0.976) when tested in the internal and external validation set, respectively. The machine learning models were statistically superior to classic methods including the percentage of tissue ablated and the Randleman ectatic score. Our model was able to correctly reclassify a patient with postoperative ectasia as an ectasia-risk group. Machine learning algorithms using a wide range of preoperative information achieved a comparable performance to screen candidates for corneal refractive surgery. An automated machine learning analysis of preoperative data can provide a safe and reliable clinical decision for refractive surgery.