https://orcid.org/0000-0003-2394-9561
Figures
Abstract
Background
Retinal fundus images provide a noninvasive window into systemic health, offering opportunities for early detection of metabolic disorders such as metabolic syndrome (METS).
Objective
This study aimed to develop a deep learning model to predict METS from fundus images obtained during routine health checkups, leveraging a multi-task learning approach.
Methods
We retrospectively analyzed 5,000 fundus images from Japanese health checkup participants.
Convolutional neural network (CNN) models were trained to classify METS status, incorporating fundus-specific data augmentation strategies and auxiliary regression tasks targeting clinical parameters such as abdominal circumference (AC). Model performance was evaluated using validation accuracy, test accuracy, and the area under the receiver operating characteristic curve (AUC).
Results
Models employing fundus-specific augmentation demonstrated more stable convergence and superior validation accuracy compared to general-purpose augmentation. Incorporating AC as an auxiliary task further enhanced performance across architectures. The final ensemble model with test-time augmentation achieved a test accuracy of 0.696 and an AUC of 0.73178.
Conclusion
Combining multi-task learning, fundus-specific data augmentation, and ensemble prediction substantially improves deep learning-based METS classification from fundus images. This approach may offer a practical, noninvasive screening tool for metabolic syndrome in general health checkup settings.
Citation: Itoh T, Nishitsuka K, Fukuma Y, Wada S (2025) Multi-task deep learning for predicting metabolic syndrome from retinal fundus images in a Japanese health checkup dataset. PLoS One 20(8): e0325337. https://doi.org/10.1371/journal.pone.0325337
Editor: Pan Li, Institute for Basic Science, KOREA, REPUBLIC OF
Received: May 11, 2025; Accepted: July 9, 2025; Published: August 21, 2025
Copyright: © 2025 Itoh et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The dataset used in this study was provided by the Japanese Society of Artificial Intelligence in Ophthalmology (JSAIO) as part of the 5th Annual Ophthalmic AI Competition. These fundus images were originally collected through the Japan Ocular Imaging Registry (JOIR) and were distributed for research purposes only, under strict conditions. According to the data use agreement signed with the organizers, use of the images is limited to the applicant’s own research, and redistribution to third parties is strictly prohibited. Furthermore, the dataset includes sensitive health information and is governed by ethical and legal frameworks under the supervision of JSAIO and the National Institute of Informatics, who have imposed these access restrictions to ensure compliance with Japanese data protection regulations. Requests for data access may be directed to the corresponding author, who will consult with the appropriate data custodians. Contact: Japanese Society of Artificial Intelligence in Ophthalmology Website: https://www.joir.jp.
Funding: The author(s) received no specific funding for this work.
Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interests: K.N., K.I., and Y.W. have received consulting fees from SAI Inc. Y.F. is the President of SAI Inc., the provider of the Eye AIRT technology used in this study. These relationships are unrelated to the submitted work. This does not alter our adherence to PLOS ONE policies on sharing data and materials.
Introduction
Metabolic syndrome (METS) is a condition characterized by a combination of visceral fat accumulation and metabolic risk factors such as hypertension, dyslipidemia, and impaired glucose tolerance. It is known to significantly increase the risk of cardiovascular disease and type 2 diabetes [1,2]. As the global prevalence of METS continues to rise, early detection and risk stratification have become increasingly important in public health. Traditionally, diagnosing METS requires physical examination and blood tests, which may limit accessibility in resource-limited or non-clinical settings.
In recent years, advances in artificial intelligence (AI), particularly deep learning, have enabled the development of image-based diagnostic tools that can non-invasively estimate systemic health conditions from medical images [3,4]. Among such modalities, retinal fundus photography is especially promising, as the retinal vasculature and optic nerve can reflect systemic pathophysiological changes, positioning the retina as a “window to overall health” [5].
Poplin et al. demonstrated that deep learning could predict systemic risk factors such as blood pressure, gender, and smoking status from retinal images with high accuracy [6]. Furthermore, Lee et al. showed that METS could be predicted from fundus images in over 13,000 subjects using a vision transformer-based model, achieving an AUC of 0.7752 [7]. However, their model was developed using data from a Korean population and applied the international diagnostic criteria for METS, in which abdominal circumference (AC) is one of five optional components [1]. In contrast, the Japanese definition requires AC (≥85 cm for men) as a mandatory component, along with additional metabolic risk factors [8].
The Japan Ocular Imaging Registry (JOIR), a real-world ophthalmic image database, has provided access to high-quality retinal images from health checkup participants in Japan [9]. In this study, we aimed to develop a deep learning model for METS prediction using these fundus images and a limited set of associated clinical data. We employed a multi-task learning approach incorporating AC as an auxiliary loss, implemented fundus-specific data augmentation strategies, and evaluated model robustness using ensemble learning and test-time augmentation (TTA). This approach demonstrated promising results and was recognized with the first-place award at an AI competition organized by the Japanese Society of Ophthalmic AI.
Methods
Dataset and participants
The dataset used in this study was provided by the Japanese Society of Ophthalmic AI as part of an AI model development competition. It is based on the Japan Ocular Imaging Registry (JOIR) [9] and includes 5,000 anonymized retinal fundus photographs from Japanese male health checkup participants. All subjects were male, a standardization applied by the organizers to reduce variation during model training. An additional set of 500 cases was provided as an independent test dataset. Clinical variables accompanying the images included age, AC, systolic blood pressure (SBP), diastolic blood pressure (DBP), triglycerides (TG), high-density lipoprotein cholesterol (HDL), and fasting plasma glucose (FPG). In this study, METS was defined according to the Japanese criteria established by the Committee to Evaluate Diagnostic Standards for Metabolic Syndrome, which require AC (≥85 cm for men and ≥90 cm for women) as a mandatory component, plus two or more of the following: elevated TG (≥150 mg/dL), low HDL (<40 mg/dL), high blood pressure (≥130/85 mmHg), or elevated FPG (≥110 mg/dL) [8]. Fig 1 illustrates the distributions and interrelationships of key clinical variables used to define METS in the JOI dataset. Among these, abdominal circumference (AC) demonstrated the strongest correlation with the presence of METS (Pearson correlation coefficient = 0.578). This statistical relationship informed our choice of AC as the auxiliary task in our multi-task learning framework.
Abdominal circumference (AC) showed the strongest positive correlation with the presence of metabolic syndrome (METS), followed by systolic blood pressure (SBP) and diastolic blood pressure (DBP). The results guided the selection of auxiliary variables for the multi-task learning framework. Abbreviations: METS, metabolic syndrome; AC, abdominal circumference; SBP, systolic blood pressure; DBP, diastolic blood pressure; TG, triglycerides; HDL, high-density lipoprotein cholesterol; FPG, fasting plasma glucose.
The presence or absence of METS was annotated based on Japanese diagnostic criteria. This retrospective study used fully anonymized data obtained from routine health checkups conducted as part of a public health program. No identifying personal information was accessible to the authors. According to the Ethical Guidelines for Life Science and Medical Research Involving Human Subjects in Japan, studies using such anonymized, non-interventional data are exempt from requiring ethics committee review. Therefore, no institutional review board approval or informed consent was required for this study.
Image preprocessing and quality control
Images with excessive blur, poor contrast, or pathological findings (e.g., diabetic retinopathy, macular scars, central retinal artery occlusion) were excluded. Additionally, cases with abdominal circumference values falling outside ±3 standard deviations were removed to mitigate the impact of outliers. After exclusion, 4,785 training cases and 500 test cases remained. Original fundus images were acquired at a resolution of 1920 × 1280 pixels. Each image was cropped without preserving the original aspect ratio and then resized according to the input requirements of each model architecture: 288 × 288 pixels for ConvNeXt-Base, and 256 × 256 pixels for SE-ResNeXt-50 and Swin Transformer V2 Base. All images were subsequently normalized prior to model input.
Data augmentation strategies
To improve generalization under limited data conditions, we compared two types of image augmentation strategies during model training. The first strategy, referred to as Case 1, was specifically designed for retinal fundus images. It included anatomically conservative transformations such as small-angle rotation, brightness and contrast adjustment, color saturation modulation, and local contrast enhancement using CLAHE. Horizontal flipping was deliberately excluded in Case 1 to preserve the orientation of anatomical landmarks such as the optic disc and macula. This strategy aimed to maintain biological plausibility while introducing sufficient variability to prevent overfitting.
The second strategy, Case 2, employed general-purpose augmentations commonly used in standard image classification tasks. This included transformations such as horizontal flipping, affine transformation, Gaussian noise injection, motion blur, and channel-wise color shifting. Unlike Case 1, Case 2 did not account for the unique anatomical features of retinal images, and therefore served as a baseline for comparison. The specific augmentation techniques included in each strategy are summarized in Table 1. The impact of each augmentation strategy on model performance was evaluated using five-fold cross-validation on the training dataset. Performance metrics reported in this study, however, are based solely on the independent 500-case test dataset.
Model architecture and training
We developed a deep learning model to predict the presence of METS based on fundus images. As shown in Fig 2, fundus images were input into a shared convolutional backbone for feature extraction. We employed a multi-task learning strategy, where the main task was binary classification of METS status, and the auxiliary task was regression of a related clinical parameter such as AC. During training, both tasks were optimized simultaneously, but only the binary classification output for METS was used for the final prediction. Loss functions included binary cross-entropy (BCE) for classification and mean squared error (MSE) for regression, with a loss weighting of 0.8:0.2.
Fundus images were used as input to a convolutional neural network (shared backbone) for feature extraction. The model employed a multi-task learning strategy, with the main task being binary classification of metabolic syndrome (METS) status, and the auxiliary task being regression of clinical parameters such as abdominal circumference (AC) during training. Final output was the binary classification of METS presence (positive/negative).
Initially, ResNet-50 was used as the backbone, and for final performance evaluation, ConvNeXt-Base, SE-ResNeXt-50, and Swin Transformer V2 Base architectures were also employed. Swin Transformer V2 Base used global average pooling. In contrast, ConvNeXt-Base and SE-ResNeXt-50 employed Generalized Mean (GeM) pooling in place of conventional global average pooling to improve spatial feature aggregation and enhance generalization. The dropout rate was increased to 0.5 before the final classification layers in all models.
Model training and evaluation
Each model was trained using stratified five-fold cross-validation. To prevent data leakage, cross-validation was performed at the patient level; that is, all fundus images from a single subject were assigned to the same fold to ensure that no images from the same individual appeared in both training and validation sets. Final model performance was evaluated using the independent 500-case test dataset provided by the competition organizers. At the time of model submission, ground truth labels for this test dataset were not disclosed to participants. After the conclusion of the competition, these labels were publicly released by the organizers. All test accuracy and AUC values reported in this study were recalculated by the authors using the released labels.
Three advanced architectures—ConvNeXt-Base, SE-ResNeXt-50, and Swin Transformer V2 Base —were trained using the optimal configuration (Case 1 augmentation + AC auxiliary loss). Their predictions were ensembled by averaging outputs from five cross-validated models for each architecture, followed by average fusion across the three architectures.
TTA was applied only to the final ensemble model. Each test image was rodomly rotated evaluated and vertically flipped three times. Final predictions were obtained by averaging these three outputs. Performance metrics reported in the Results section are based solely on the 500-case independent test dataset. To ensure reproducibility, all models were retrained in a standardized software environment with consistent preprocessing and fixed library versions.
Model interpretation and error analysis
To assess model interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to the ConvNeXt-Base CNN model. We utilized the open-source `pytorch-grad-cam` library (https://github.com/jacobgil/pytorch-grad-cam) to generate heatmaps for visualizing the spatial regions in fundus photographs that contributed most to the model’s prediction of METS presence or absence.
Additionally, predicted probability scores (pred_score) for the METS-positive class were extracted from the softmax output. Based on the ground truth and predicted labels, the test set images were categorized into four groups: true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN). To investigate how prediction confidence varied among these groups, we plotted histograms of the predicted scores using Python (seaborn and matplotlib libraries).
Implementation environment
All models were implemented in Python 3.12 using PyTorch 2.4.1 and trained on a workstation equipped with an NVIDIA RTX 3090 GPU. Image preprocessing and augmentation were conducted using the Albumentations 1.4.7 library. Model prototyping and configuration management were supported using EyeAIRT (Sensor to AI, Inc.).
Results
Comparison of auxiliary loss variables and validation strategy
We investigated the effect of different auxiliary loss configurations on validation performance in multi-task learning models using ConvNeXt-Base, SE-ResNeXt-50 and Swin Transformer V2 Base architectures. Each model was trained to classify METS status either without auxiliary loss (Main only), or with various clinical parameters as auxiliary regression targets. The evaluated configurations included: AC only; AC combined with SBP and DBP; and all four variables—AC, SBP, DBP, TG, and FPG. As shown in Table 2, the inclusion of AC consistently improved validation accuracy and AUC in both architectures. While adding further variables occasionally led to marginal gains, the overall improvement compared to the AC-only setting was limited and inconsistent.
Based on these findings, the AC-only configuration was selected for all subsequent training and evaluation. This approach balances predictive performance with model simplicity and interpretability, and was consistently effective across all evaluated architectures.
Comparison of data augmentation strategies
We compared the effects of two different data augmentation strategies: Case 1 (fundus-specific augmentation) and Case 2 (general-purpose augmentation). As shown in Fig 3, the learning curves for training and validation accuracy over 30 epochs revealed that Case 1 achieved more stable convergence and higher validation performance compared to Case 2. While Fig 3 illustrates the training dynamics, a quantitative comparison of validation performance between the two strategies is provided below. As summarized in Table 3, the validation accuracy of the Case 1 models was consistently higher than that of the Case 2 models across different architectures. Specifically, Case 1 achieved a higher average validation accuracy (0.66290) compared to Case 2 (0.63762), supporting its effectiveness in fundus-specific model development.
(A) Learning curves for Case 1 using fundus-specific data augmentation. Training accuracy and validation accuracy both increase steadily over the course of 25 epochs, with only a small gap between them. This indicates that the model trained with fundus-specific augmentation achieved stable convergence and generalization without signs of overfitting. (B) Learning curves for Case 2 using general-purpose data augmentation. While training accuracy increases continuously, validation accuracy plateaus early and then gradually declines. This pattern suggests that the model is overfitting the training data and fails to generalize effectively to unseen data.
Final model performance with ensemble and TTA
To evaluate the final classification performance of our system, we tested three advanced architectures—ConvNeXt-Base, SE-ResNeXt-50 and Swin Transformer V2 Base —on the independent test dataset comprising 500 cases. All models were trained under the optimal configuration using fundus-specific augmentation (Case 1) and AC as the auxiliary task. As shown in Table 4, the individual models achieved test accuracies of 0.66400 (ConvNeXt-Base), 0.66000 (SE-ResNeXt-50), and 0.66600 (Swin Transformer V2 Base), with corresponding AUCs of 0.72136, 0.70139, and 0.71712, respectively.
Subsequently, we constructed an ensemble model by averaging the predictions from all three architectures and applying TTA. Each test image was evaluated in three forms (original, rotated, vertically flipped), and the final prediction was determined by averaging the outputs. The ensemble model with TTA achieved a test accuracy of 0.69600 and an AUC of 0.73178, outperforming all individual models. While this configuration yielded the highest performance, the improvement may partly reflect the complementary nature of the integrated architectures rather than TTA alone. These findings support the value of ensemble learning in fundus image-based classification tasks.
Interpretability and error analysis
To enhance interpretability, Grad-CAM visualizations were generated using the best-performing ConvNeXt-Base model. Representative examples are shown in Fig 4. The example from a correctly predicted METS-positive case (true positive, TP; Fig 4A) exhibits strong activation around the optic disc and major retinal vessels, indicating that the model leverages these anatomical regions to predict the presence of METS. In contrast, the correctly predicted METS-negative case (true negative, TN; Fig 4B) shows more localized and less prominent activation, suggesting reduced reliance on these features when METS is absent.
Grad-CAM visualizations were generated using the ConvNeXt-Base model. (A) A true positive (TP) case with a high predicted probability for METS (pred = 0.881), demonstrating strong activation over the optic disc and vascular arcades. (B) A true negative (TN) case with a low predicted probability (pred = 0.096), with activation localized to a smaller region.
To further evaluate model performance, predicted probabilities (pred_score) for the METS-positive class were grouped into four categories—TP, FP, FN, and TN—based on the ground truth and model predictions from the 500-case test dataset. Fig 5 presents histogram plots of the predicted scores for each group. The TP and TN groups demonstrated sharply peaked distributions near 1.0 and 0.0, respectively, indicating high model confidence in correct classifications. In contrast, the FP and FN distributions were more broadly distributed around the decision threshold, reflecting increased uncertainty in misclassified cases.
Histogram plots show the distribution of predicted scores (pred_score) for the METS-positive class in the 500-case test dataset. The four categories—true positives (TP), false positives (FP), false negatives (FN), and true negatives (TN)—are plotted separately. TP and TN predictions exhibited sharper peaks near 1.0 and 0.0, respectively, while FP and FN distributions were broader, suggesting lower confidence and borderline misclassifications.
Discussion
In this study, we developed a multi-task deep learning model to predict METS from retinal fundus images obtained from Japanese health checkup participants. The final model achieved a test accuracy of 0.696 through the integration of AC as an auxiliary task, fundus-specific data augmentation, and ensemble learning with TTA. These results demonstrate the feasibility of non-invasive, image-based METS screening using limited but well-curated retinal datasets.
The use of AC as an auxiliary variable in a multi-task learning framework was found to be the most effective among the clinical indicators tested. This finding is based on validation accuracy observed during model development, where the configuration using AC alone as an auxiliary loss yielded the best performance. This highlights the importance of selecting auxiliary tasks that are semantically and visually aligned with the primary classification target, thereby minimizing gradient interference during training [4]. Since AC reflects visceral fat accumulation—a central factor in METS—it may be indirectly manifested in the structure and appearance of the retinal vasculature. Epidemiological studies have also reported significant associations between retinal microvascular signs and METS components, including AC [3].
Our results also emphasize the benefits of fundus-specific data augmentation strategies. Based on validation accuracy observed during cross-validation, Case 1 augmentation—which included anatomically conservative transformations tailored to retinal fundus images—consistently outperformed the general-purpose Case 2 strategy. Case 1 led to more stable training and higher mean accuracy, while Case 2 resulted in early convergence and reduced performance. This suggests that domain-aware preprocessing plays a critical role in improving deep learning performance, particularly in medical image analysis where anatomical orientation is clinically meaningful. While clinical variables can directly determine METS diagnosis based on established international criteria, our model is intended as a pre-diagnostic screening tool that can identify at-risk individuals using non-invasive fundus imaging, particularly in settings where full clinical data may not yet be available.
The final model leveraged three advanced neural network architectures—ConvNeXt-Base, SE-ResNeXt-50, and Swin Transformer V2 Base —and demonstrated improved predictive performance through ensemble learning and TTA. Such strategies are known to enhance robustness and consistency across medical imaging tasks [5], and this was validated in our study using a 500-case independent test set.
In addition to accuracy, the AUC was also evaluated. The final ensemble model achieved an AUC of 0.73178, while individual models recorded AUCs of 0.72136 (ConvNeXt-Base), 0.70139 (SE-ResNeXt-50), and 0.71712 (Swin Transformer V2 Base), respectively. These values indicate strong discriminatory performance. Notably, Lee et al. reported an AUC of 0.7752 using a vision transformer model trained on over 13,000 images from a Korean population based on international METS criteria [7]. While our AUC is slightly lower, our model achieved this performance using a significantly smaller dataset (5,000 cases) and the Japanese diagnostic standard, which requires abdominal circumference as a mandatory component. This highlights the effectiveness of our task-specific optimization and augmentation strategies even in a more constrained dataset setting.
In the early phase of model development, we considered incorporating multiple clinical parameters related to metabolic syndrome, such as obesity, blood pressure, HbA1c, and lipid profiles, into the auxiliary tasks of multi-task learning. However, based on preliminary correlation analysis within the training dataset, we found that AC exhibited the strongest association with METS status among the candidates. Given the time constraints during the competition phase, we initially adopted a focused strategy by selecting AC as the sole auxiliary task. Subsequent analyses conducted for the purpose of formal publication reaffirmed that the use of AC alone provided the most stable and highest predictive performance for METS. These findings suggest that in multi-task learning, careful selection of auxiliary tasks that are strongly and directly linked to the main prediction target is critical for optimizing model performance. Rather than indiscriminately including multiple sub-tasks, strategic focusing based on clinical and statistical relevance can significantly contribute to achieving robust results. Furthermore, the strategic selection of AC as an auxiliary task based on preliminary correlation analysis is a novel aspect of this study. While multi-task learning has been explored in other medical imaging domains, its application to fundus-based metabolic risk prediction with systematic auxiliary variable selection has been scarcely reported. Our findings highlight the potential of auxiliary loss design in enhancing the predictive performance of fundus image-based AI models.
Recent advances in “oculomics”—the study of systemic health markers through ocular imaging—have further underscored the potential of the retina as a biomarker-rich window into overall physiology. Poplin et al. [6] reported that deep learning could predict cardiovascular risk factors such as age, blood pressure, and smoking status from fundus images with AUCs ranging from 0.70 to 0.82, depending on the variable. Similarly, Rim et al. demonstrated that AI models could infer kidney function markers such as estimated glomerular filtration rate (eGFR) from fundus photographs [10]. Other studies have shown that fundus-based AI can accurately estimate hemoglobin levels [11], biological age [12], and even cognitive decline [13], highlighting the retina’s unique role in reflecting multi-organ health. Our study contributes to this expanding body of oculomic research by focusing on METS prediction and incorporating multi-task learning to enhance clinical interpretability. In contrast to previous single-task approaches, our model’s auxiliary loss design guided by clinical correlation strengthens its connection to physiologically meaningful features. This study aligns with the broader vision of oculomics, which seeks to establish the eye as a non-invasive, accessible window to systemic health. Recent advances in this field include efforts toward integrative multi-disease screening and the incorporation of omics-level data such as genomics and proteomics. By demonstrating a multi-task learning framework for METS prediction, our study contributes to this evolving direction and lays the groundwork for future applications in preventive medicine.
In addition to improving predictive accuracy, we sought to enhance the interpretability of the model’s decision-making process. Grad-CAM visualizations (Fig 4) revealed that predictions of METS presence were consistently associated with activation around the optic disc and major vessels, suggesting that the model captured relevant anatomical features. Furthermore, histograms of predicted scores (Fig 5) indicated that correct predictions (TP, TN) were associated with higher confidence, while misclassified cases (FP, FN) tended to cluster near the decision threshold. These findings demonstrate that our model’s behavior is partially explainable and could help guide future refinement or clinical integration.
This study has several limitations. First, the dataset provided for training consisted solely of male participants. While this reduced variability and improved model stability, it limits the generalizability of the results to female populations. Future work should explore sex-stratified model development and validation. Second, the auxiliary loss was designed using only a limited number of clinical variables. Incorporating additional predictors—such as HbA1c, insulin resistance markers, or lifestyle factors like smoking and physical activity—may further improve performance. Third, all fundus images used in this study were captured using Canon retinal cameras, as confirmed via image metadata. While this ensured consistency in image quality, the generalizability of the model to images captured by different devices or under varied conditions remains to be evaluated through external validation.
Conclusion
This study presents a reproducible and accurate deep learning framework for predicting METS from retinal fundus images in a Japanese health checkup cohort. Through the integration of multi-task learning, fundus-specific data augmentation, and ensemble prediction with TTA, our model achieved clinically meaningful performance while using relatively limited training data. These findings support the potential for scalable, non-invasive AI-based screening tools for systemic diseases using retinal imaging, particularly in preventive medicine and public health settings. In addition, the strategic design of the auxiliary loss function, guided by preliminary clinical correlation analysis, represents a novel contribution to fundus-based AI research.
Acknowledgments
The clinical information and fundus photographs used in this study were provided by the Japanese Society of Artificial Intelligence in Ophthalmology and the National Institute of Informatics, and were managed by the Japan Ocular Imaging Registry website (http://joir.jp/).
References
- 1. Grundy SM, Cleeman JI, Daniels SR, Donato KA, Eckel RH, Franklin BA, et al. Diagnosis and management of the metabolic syndrome: an American Heart Association/National Heart, Lung, and Blood Institute Scientific Statement. Circulation. 2005;112(17):2735–52. pmid:16157765
- 2. Alberti KGMM, Zimmet P, Shaw J, IDF Epidemiology Task Force Consensus Group. The metabolic syndrome--a new worldwide definition. Lancet. 2005;366(9491):1059–62. pmid:16182882
- 3. Kawasaki R, Tielsch JM, Wang JJ, Wong TY, Mitchell P, Tano Y, et al. The metabolic syndrome and retinal microvascular signs in a Japanese population: the Funagata study. Br J Ophthalmol. 2008;92(2):161–6. pmid:17965107
- 4. Ruder S. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098. 2017. Available from: https://arxiv.org/abs/1706.05098
- 5. Liu Z, Luo P, Wang X, Tang X. Deep learning face attributes in the wild. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). Santiago, Chile; 2015. p. 3730–8.
- 6. Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng. 2018;2(3):158–64. pmid:31015713
- 7. Lee TK, Kim SY, Choi HJ, Choe EK, Sohn K-A. Vision transformer based interpretable metabolic syndrome classification using retinal Images. NPJ Digit Med. 2025;8(1):205. pmid:40216912
- 8. Committee to Evaluate Diagnostic Standards for Metabolic Syndrome. Definition and the diagnostic standard for metabolic syndrome. Nihon Naika Gakkai Zasshi. 2005;94(4):794–809. pmid:15865013
- 9. Miyake M, Akiyama M, Kashiwagi K, Sakamoto T, Oshika T. Japan Ocular Imaging Registry: a national ophthalmology real-world database. Jpn J Ophthalmol. 2022;66(6):499–503. pmid:36138192
- 10. Sabanayagam C, Xu D, Ting DSW, Nusinovici S, Banu R, Hamzah H, et al. A deep learning algorithm to detect chronic kidney disease from retinal photographs in community-based populations. Lancet Digit Health. 2020;2(6):e295–302. pmid:33328123
- 11. Mitani A, Huang A, Venugopalan S, Corrado GS, Peng L, Webster DR, et al. Detection of anaemia from retinal fundus images via deep learning. Nat Biomed Eng. 2020;4(1):18–27. pmid:31873211
- 12. Nagasato D, Tabuchi H, Masumoto H, Kusuyama T, Kawai Y, Ishitobi N, et al. Prediction of age and brachial-ankle pulse-wave velocity using ultra-wide-field pseudo-color images by deep learning. Sci Rep. 2020;10(1):19369. pmid:33168888
- 13. Li R, Hui Y, Zhang X, Zhang S, Lv B, Ni Y, et al. Ocular biomarkers of cognitive decline based on deep-learning retinal vessel segmentation. BMC Geriatr. 2024;24(1):28. pmid:38184539