The retinal age gap: an affordable and highly accessible biomarker for population-wide disease screening across the globe.

Traditional biomarkers, such as those obtained from blood tests, are essential for early disease detection, improving health outcomes and reducing healthcare costs. However, they often involve invasive procedures, specialized laboratory equipment or special handling of biospecimens. The retinal age gap (RAG) has emerged as a promising new biomarker that can overcome these limitations, making it particularly suitable for disease screening in low- and middle-income countries. This study aimed to evaluate the potential of the RAG as a biomarker for broad disease screening across a vast spectrum of diseases. Fundus images were collected from 86 522 UK Biobank participants aged 40–83 (mean age: 56.2 ± 8.3 years). A deep learning model was trained to predict retinal age using 17 791 images from healthy participants. The remaining images were categorized into disease/injury groups based on clinical codes. Additionally, 8524 participants from the Brazilian Multilabel Ophthalmological Dataset (BRSET) were used for external validation. Among the 159 disease/injury groups from the 2019 Global Burden of Disease Study, 56 groups (35.2%) exhibited RAG distributions significantly different from healthy controls. Notable examples included chronic kidney disease, cardiovascular disease, blindness, vision loss and diabetes. Overall, the RAG shows great promise as a cost-effective, non-invasive biomarker for early disease screening.

retinal age gap, machine learning, retinal age prediction, deep learning 1. Introduction In the evolving landscape of global healthcare, timely disease detection and intervention are vital for improving population health outcomes and reducing costs [1]. Historically, traditional disease biomarkers, notably those procured from blood tests, have been instrumental in this pursuit [2]. The efficacy of these biomarkers stems from their ability to concurrently provide insights into multiple disease pathways, thereby providing a comprehensive snapshot of an individual’s health status. However, the analysis and quantification of most disease biomarkers available today requires invasive procedures, sophisticated laboratory equipment or sensitive handling of the biospecimens, making them challenging to deploy within global healthcare systems [3,4]. The retina, a unique and complex structure located at the back of the eye, can give a snapshot of the body’s overall health and has emerged as a potential biomarker for detecting a wide range of diseases [5–7]. Due to its direct connection with the central nervous system and the systemic circulation, the retina can show changes and anomalies that may signify the onset or progression of various health conditions, not just those confined to the eye. Retinal imaging technologies allow for the non-invasive visualization of the retina’s microvasculature and neural tissue, enabling the detection of subtle changes that might indicate systemic diseases such as diabetes, chronic kidney disease, and cardiovascular diseases, as well as neurodegenerative disorders like Alzheimer’s disease [8–10]. This capability is particularly valuable because it offers a painless and easily accessible means of early disease detection, potentially before clinical symptoms manifest. The exploration of the retina as a source of disease biomarkers is a testament to the ongoing convergence of technology and healthcare, offering promising avenues for precision medicine [11]. One specifically promising emerging retinal biomarker is known as the retinal age gap (RAG), which measures the difference between the biological retinal age and the chronological age of an individual [12]. In this context, the biological retinal age is typically estimated using a machine learning model trained on retinal fundus images of healthy subjects [13]. Recently, the RAG has been found to be a biomarker that is associated with a variety of diseases, including stroke [14], kidney failure [15] and Parkinson’s disease [16]. Importantly, RAG estimation can be performed using low-cost and easy-to-operate standard fundus imaging equipment, widely available in ophthalmic practices worldwide or even through smartphone-based solutions that enable at-home and rural data acquisition [17]. Furthermore, the RAG does not require advanced laboratory equipment and special handling of biospecimens, making it an ideal candidate for disease screening, particularly in low- and middle-income countries, but also in high-income countries with cost-strained healthcare systems. Despite the potential of RAG to be a low-cost and widely available imaging-biomarker for many diseases, its utility has only been shown for a few selected diseases while its systematic utility as a disease biomarker has not been shown so far. Thus, the aim of this work was to perform a large-scale machine learning-based analysis using fundus imaging data from more than 80 000 subjects included in the UK Biobank [18] to determine the RAG distribution for a wide spectrum of diseases and injuries with the selection motivated and guided by the 2019 Global Burden of Disease and Injury Study [19]. Furthermore, this work also aimed to externally validate the developed model using data from the Brazilian Multilabel Ophthalmological Dataset (BRSET) [20]. The workflow developed to train the machine learning models for retinal age prediction and to further analyse RAG predictions across disease and injury groups is summarized in figure 1 and described in more detail in the following sections. 2. Material and methods (a). Dataset The UK Biobank, initiated in 2006, is a large-scale population-based study examining genetic, lifestyle and environmental interactions in various diseases such as cardiovascular diseases, cancer and neurological disorders [18]. With data from over 500 000 participants aged between 40 and 69 years old at recruitment, the UK Biobank is one of the most extensive biomedical research resources available worldwide. It offers retinal fundus images from over 80 000 participants. Briefly described, retinal fundus photographs were obtained using the Topcon 3D OCT-1000 Mark II system, manufactured by Topcon Corp in Tokyo, Japan. Each eye was captured as a 45° non-mydriatic and non-stereo image, ensuring that both, the optic disc and the macula, were centred within the frame. The BRSET dataset was utilized in this work for external validation to evaluate the generalization, and therefore clinical potential, of the trained model [20]. This dataset comprises 16 266 images from 8524 participants, collected over the period from 2010 to 2020, with each participant contributing macula-centred fundus images from both eyes. Additionally, the BRSET dataset includes information about a few diseases for each participant. Retinal images in this dataset were captured using a Nikon NF505 and a Canon CR-2 fundus camera. (b). Data preprocessing and demographic information From the UK Biobank database, 177 354 fundus images were obtained from 86 522 participants for this work. Originally made available with an image size of 2048 × 1536 pixels, the images were preprocessed to centre the retina and standardize the optic disc orientation. For computational efficiency, these images were resized to 800 × 800 pixels. After image quality screening using the deep learning model described by Fu [21], a total of 97 611 images from 55 024 participants were included for experimental analysis in this work. This image quality screening model identifies key indicators of poor image quality, including blurring, uneven illumination and low contrast. Only images classified as having sufficient quality by this method were included for further analysis. The same preprocessing procedures applied to the UK Biobank images were also used for preprocessing of the BRSET dataset, yielding a total of 9650 images from 5313 participants. The demographic details of the UK Biobank cohort included in the experimental analysis are provided in table 1. Electronic supplementary material, table S1 presents the demographics of UK Biobank participants excluded due to poor image quality. Table 2 summarizes the demographic information for the BRSET participants, whose clinical data are relatively sparse compared to the UK Biobank cohort. Table 1. | N | mean | std | min | Q1 | Q2 | Q3 | max | | |---|---|---|---|---|---|---|---|---| | age (years) | 55 024 | 56.46 | 8.27 | 40 | 50 | 58 | 63 | 83 | | diastolic blood pressure (mm Hg) | 54 629 | 81.45 | 9.99 | 44.50 | 74.50 | 81.00 | 88.00 | 134.00 | | systolic blood pressure (mm Hg) | 54 629 | 136.50 | 18.27 | 76.50 | 123.50 | 135.00 | 148.00 | 245.00 | | HDL cholesterol (mmol L−1) | 47 036 | 1.49 | 0.39 | 0.36 | 1.21 | 1.44 | 1.72 | 4.13 | | triglycerides (mmol L−1) | 49 903 | 1.66 | 0.95 | 0.25 | 1.01 | 1.42 | 2.03 | 11.19 | | glucose (mmol L−1) | 46 997 | 5.12 | 0.99 | 1.25 | 4.69 | 4.99 | 5.32 | 25.49 | | weight (kg) | 54 801 | 77.70 | 15.75 | 33.60 | 66.20 | 76.10 | 87.10 | 188.10 | | body mass index (kg m−2) | 54 782 | 27.13 | 4.67 | 12.65 | 23.93 | 26.48 | 29.52 | 65.98 | | arterial stiffness index | 54 387 | 9.51 | 3.62 | 0.97 | 7.23 | 9.21 | 11.24 | 350.00 | | left eye logMAR | 54 477 | 0.02 | 0.21 | −1.06 | −0.12 | −0.04 | 0.10 | 1.35 | | right eye logMAR | 54 516 | 0.02 | 0.20 | −0.66 | −0.12 | −0.04 | 0.10 | 1.35 | | alcohol intake | 54 926 | never: 4140; special occasions only: 6258; one to three times a month: 6374; once or twice a week: 14 021; three or four times a week: 13 018; daily or almost daily: 11 063; prefer not to answer: 52 | |||||| | smoking status | 54 926 | never smoked: 31 421; previous smoker: 18 661; current smoker: 4672; prefer not to answer: 172 | |||||| | sex | 55 024 | female: 30 063; male: 24 961 | |||||| | ethnicity | 54 914 | white: 51 501; mixed background: 392; black: 1042; Asian: 1037; other: 942 | Table 2. | N | mean | std | min | Q1 | Q2 | Q3 | max | | |---|---|---|---|---|---|---|---|---| | age (years) | 5313 | 57.39 | 18.26 | 5 | 47 | 60 | 71 | 97 | | sex | 5313 | female: 3351; male: 1962 | (c). Machine learning model training An EfficientNetB3 convolutional neural network with 12 million parameters and 395 layers was trained for retinal age prediction, initialized with weights pretrained on ImageNet [22]. Subsequently, the model was fine-tuned on fundus images from the UK Biobank using the Adam optimizer with a learning rate of and a weight decay of to minimize the L2 loss between the known chronological age and the predicted value. To reduce overfitting, data augmentation techniques were applied during training, including random rotations up to 10°, translations up to 20% of the image size, horizontal and vertical flips with a 0.5 probability, and colour jitter adjustments of up to 20% for brightness, contrast, saturation and hue. Training occurred over 50 epochs with a batch size of 8. After performance assessment using fivefold cross validation, the final model underwent training with UK Biobank data using an 80/10/10 split for training, validation and testing, respectively. Model selection was performed to minimize overfitting by selecting the model with the best performance on the validation set. PyTorch 1.13.1 was used to implement the EfficientNetB3 model architectures, and training was performed using an NVIDIA RTX 3090 GPU. (d). Dataset construction and computation of RAG To identify healthy UK Biobank participants for training the age prediction models, the stringent filtering criteria proposed by Zhu et al. was used [12], where a participant was considered healthy if they had no self-reported medical conditions prior to imaging acquisition. Following the methodology of Zhang et al. [15], images from both eyes were used for machine learning training to maximize the volume of training data. To prevent statistical bias in the evaluation of the retinal age prediction model’s performance, the partitioning of fundus images into training, validation and testing datasets was carefully managed. This ensured that individual participants were allocated exclusively to either the training, validation or testing datasets. Furthermore, to avoid statistical bias while investigating the relationship between RAG and disease/injury groups, only a single image per participant was analysed for testing, prioritizing the right eye and using the left if the right was unavailable. A total of 19 053 fundus images from 10 828 participants were identified as healthy and used for training the age prediction models. In the BRSET dataset, a participant was deemed healthy if they exhibited no previously reported diseases. In total, 1247 images were obtained from 796 healthy participants in the BRSET dataset. To investigate the utility of the RAG as a disease biomarker, the RAG was computed for 321 disease/injury groups included in the 2019 Global Burden of Disease and Injury Study. In total, fundus images from 44 196 participants were organized into disease/injury groups based on the presence of specific International Classification of Diseases 10th Revision (ICD-10) codes attained prior to image acquisition as shown in electronic supplementary material, table S2. Out of the total 321 different disease/injury groups identified, 159 groups contained at least 50 participants and were used to investigate statistical differences in RAG value distributions between healthy controls and each disease/injury group. An overview of the data used for training the age prediction models and to analyze the different disease/injury groups is shown in figure 2. Once the machine learning models for estimating biological retinal age were trained and validated using data from healthy UK Biobank participants, RAG values were calculated for each participant in the 159 identified disease/injury groups from the UK Biobank. Practically, the RAG values were computed by subtracting the participant’s chronological age from the predicted biological age estimated by the machine learning models. For the BRSET dataset, participants were categorized according to the presence of specific disease labels. As for the UK Biobank data, RAG values were calculated for each subject in the disease groups that included at least 50 participants. Notably, the two disease groups that met this sample size requirement and aligned with diseases examined in the UK Biobank were diabetes mellitus, which had 939 participants and age-related macular degeneration, with 123 participants. (e). Statistical analysis Machine learning-based age prediction models are commonly known to exhibit age-related biases [23]. These biases often lead to the overestimation of ages for younger subjects and the underestimation for elderly subjects (regression towards the mean phenomenon). To correct for differences in the age distribution between groups, domain adaptation techniques were used to estimate the sampling distribution for the mean RAG value through weighted bootstrap resampling [24]. Specifically, kernel mean matching was used to estimate the weights needed to correct for the differences in age distribution during resampling [25]. Overall, 10 000 weighted bootstrap samples were employed to estimate the sampling distributions for the age corrected mean RAG for each disease/injury group. One-tailed nonparametric bootstrap hypothesis testing was performed for each disease/injury group to evaluate whether the absolute mean RAG values were larger than for healthy controls. The Bonferroni-corrected significance level for multiple comparisons testing across the 159 UK Biobank disease/injury groups was set to [26]. Similarly, for the two disease groups in the BRSET dataset, the Bonferroni corrected significance level was set to . A p-value below signified significance for each disease/injury group. 3. Results The retinal age prediction model achieved a mean absolute error (m.a.e.) of 3.11 ± 0.06 years and a Pearson correlation coefficient of 0.88 ± 0.01 for the healthy participants from the UK Biobank test set, representing new state-of-the-art performance for retinal age prediction on the UK Biobank [12]. Additionally, the MAE for the healthy participants in the external BRSET dataset was 4.03 ± 0.10 years with a Pearson correlation coefficient of 0.77 ± 0.02. Figure 3 presents a scatter plot illustrating the retinal age predictions made by the model on the healthy test sets from the UK Biobank and BRSET datasets. Figure 4 presents boxplots depicting the mean absolute RAG values from the UK Biobank for a selection of diseases and injuries, with a complete list of these values available in electronic supplementary material, table S3. The average RAG value for healthy participants is expected to be near zero, serving as a baseline against which deviations in disease or injury groups can be measured. In the UK Biobank, the average RAG value for healthy participants within the test set was calculated as 0.13 ± 0.09 years. Similarly, in the BRSET dataset, the average RAG value for healthy individuals was found to be 0.30 ± 0.23 years. Across the 159 disease/injury groups in the UK Biobank, 56 groups (35.2%) had a RAG value distribution that differed significantly from healthy controls (p-value of less than ). Notable examples included blindness and vision loss, cardiovascular diseases, diabetes mellitus and chronic kidney disease. The mean absolute RAG values across the diseases and injuries presented in figure 4 demonstrate that the average RAG value magnitude is dependent on the specific physiology of the disease or injury. Several disease/injury groups displayed relatively larger average RAG values compared to other disease/injury groups. For instance, diabetes mellitus patients exhibited notable average RAG values (RAG: 1.39 ± 0.11 years). Similarly, cardiovascular conditions like peripheral artery disease were associated with elevated average RAG values (RAG: 0.80 ± 0.12 years). Moreover, high average RAG values were noted in individuals with cataracts (RAG: 1.91 ± 0.18 years), as well as in other eye diseases such as glaucoma (RAG: 1.38 ± 0.11 years) and age-related macular degeneration (RAG: 1.79 ± 0.15). Patients suffering from neurological conditions like multiple sclerosis (RAG: 1.72 ± 0.31 years) and substance use disorders, including alcohol use disorders (RAG: 0.88 ± 0.19 years), also showed relatively elevated average RAG values. Conversely, certain disease or injury groups had relatively lower average RAG values that did not differ significantly from healthy subjects. Gynaecological conditions such as endometriosis (RAG: 0.27 ± 0.14 years), musculoskeletal disorders like lower back pain (RAG: 0.30 ± 0.10 years), and injuries from road accidents (RAG: 0.25 ± 0.14 years) were among those with smaller average RAG values. Within the BRSET dataset, the two disease groups analysed (diabetes mellitus and age-related macular degeneration) had average RAG values that were significantly larger from those of healthy controls (p-value of less than ). Furthermore, no significant differences were found when comparing average RAG values between the UK Biobank and the BRSET dataset. Specifically, for diabetes mellitus, the average RAG values were 1.39 ± 0.11 years in the UK Biobank and 1.23 ± 0.19 years in the BRSET dataset, with the difference being statistically insignificant (p = 0.78). Similarly, for age-related macular degeneration, the average RAG value was 1.79 ± 0.15 years in the UK Biobank and 1.68 ± 0.55 years in the BRSET dataset, also showing no statistically significant difference (p = 0.59). 4. Discussion The experimental results of this study reveal that the developed deep learning retinal age prediction model provides excellent performance for this task, establishing a new state-of-the-art performance on the UK Biobank. In particular, the developed model achieved an MAE of 3.11 years, outperforming several other studies on the UK Biobank that utilized the Xception convolutional neural network architecture and reported an MAE of 3.55 years [12,14–16]. Additionally, the model demonstrated robust generalizability when applied to the external validation BRSET dataset, achieving an MAE of 4.03 years. Overall, the findings of this study provide substantial new support for the use of the RAG as a potential alternative imaging-biomarker to other traditional, invasive biomarkers for many diseases. Importantly, the results of this work align well with the known physiological impact of diseases and injuries on the retina. Notably, patients with diabetes mellitus exhibited pronounced average RAG values, consistent with the retina being particularly vulnerable to hyperglycaemic damage, which results in microvascular complications, oxidative stress and retinal nerve fibre damage [27]. In addition, diabetes mellitus has been previously linked to age-related retinal degeneration [28]. Similarly, patients with cardiovascular conditions, like peripheral artery disease, exhibited larger average RAG values, probably indicative of the intertwined susceptibilities of the retina to vascular pathologies. Gynaecological ailments, s0uch as endometriosis, presented with relatively lower average RAG values. This could be attributed to such conditions predominantly impacting reproductive tissues without significant systemic vascular implications. Musculoskeletal disorders like lower back pain, or injuries from events like road accidents, also displayed low average RAG values, which may be expected. Ocular conditions like cataracts and glaucoma significantly influenced RAG, which is also an expected finding since these diseases majorly affect the eye’s anatomy and function. Participants with chronic kidney disease exhibited an elevated RAG, potentially due to similarities in microvasculature between the kidneys and the retina [15]. Many neurological disorders, such as multiple sclerosis, also manifested with an elevated RAG. This finding also aligns with prior knowledge given that these ailments can lead to inflammation of the optic nerve [29]. Additionally, substance use disorders, especially those linked to alcohol consumption, have been associated with increased risk for optic neuropathy [30]. Notably, the framework has shown to be promising for global clinical applications by demonstrating that the RAG values estimated by the model trained using UK Biobank data can effectively detect diabetes mellitus and age-related macular degeneration in a completely different cohort from a different part of this world (Brazil) based on imaging data captured by completely different devices leading to image quality differences, while yielding comparable average RAG values for these conditions. Overall, these observations reinforce the general validity of the biological age prediction framework for low-cost global disease surveillance. The results of this study highlight the enormous potential of the RAG as a sensitive, non-invasive disease biomarker, offering a non-invasive, practical alternative to traditional blood-based biomarkers of aging, such as PhenoAge [31], which relies on blood-based variables obtained through invasive procedures and laboratory analysis. Importantly, retinal images can be acquired using comparably inexpensive equipment, and even smartphones. The deep learning network developed for retinal age prediction is compatible with standard computing equipment widely available and, with some minor technical adaptations, may even be suitable for use on smartphones in the future [17]. This would allow for local calculation of the RAG on a smartphone without requiring an internet connection, offering immediate feedback, and making it an ideal assessment tool for rural and underdeveloped regions. Overall, this study shows that the RAG has potential to be deployed as an initial screening tool within clinical settings, offering a non-invasive and highly accessible method to identify individuals who may benefit from further, more specific diagnostic evaluation. Furthermore, it may be possible to build more detailed classifiers for those diseases with a significantly altered RAG. This approach leverages the RAG’s ability to provide early insights into systemic health conditions, serving as a gateway to targeted and earlier diagnostic and therapeutic interventions. While this study provides a comprehensive assessment of the RAG across various disease/injury groups, some limitations and challenges necessitate further investigation in future studies. One limitation lies in the categorization of disease and injury groups in the UK Biobank, which relies on specific ICD-10 clinical codes. These may not fully align with the disease classifications in the BRSET dataset, potentially leading to discrepancies due to differences in clinical definitions. Such variations could introduce confounding effects, especially for patients with comorbidities. Additionally, cultural activities and lifestyle factors, such as diet, sun exposure, and other behaviors, may influence retinal aging and act as confounding variables in the analysis. To address these challenges, future work should focus on investigating these relationships in greater depth. Another limitation is the significant number of low-quality fundus images in the UK Biobank, leading to the exclusion of numerous images during quality control. This critical aspect of the UK Biobank data was previously pointed out by MacGillivray et al., who observed that a large portion of the retinal images in the UK Biobank are too poor in quality for automated analysis [32]. Removing these substandard images is crucial to maintain high-quality data for training purposes. However, as shown in table 1 and electronic supplementary material, table S1, the demographic and clinical characteristics of the excluded participants are similar to those included in the analysis. This similarity supports the assertion that the exclusion of these participants due to poor image quality does not compromise the generalizability of the experimental analysis. Furthermore, in real-world clinical applications, image quality is typically assessed at the time of acquisition, allowing suboptimal images to be flagged and retaken, thereby minimizing the impact of poor image quality on downstream analyses [33]. In future work, we plan to strengthen the robustness of the proposed framework to accommodate diverse imaging conditions through utilizing generative AI techniques for image quality enhancement [34]. Lastly, the scope of diseases for which RAG value distributions could be compared between the UK Biobank and the BRSET dataset was limited by the amount of data available in the datasets. Increasing the dataset size and variability could enable comparisons across a wider range of diseases, thereby enhancing the validation of the clinical utility of the biological age prediction framework and confirming its applicability across various disease contexts although these first results show that the RAG is a promising and generalizable biomarker. 5. Conclusions This research demonstrates that the retinal age gap has the potential to support clinical practice worldwide by providing a non-invasive, highly cost-effective and accessible biomarker for population-wide screening for a multitude of diseases that put a significant burden on the healthcare infrastructure.

We are very grateful to all the individuals that took part in the UK Biobank study and the Brazilian Multilabel Ophthalmological Dataset. Contributor Information Christopher Nielsen, Email: [email protected]. Matthias Wilms, Email: [email protected]. Nils Daniel Forkert, Email: [email protected]. Ethics UK Biobank has approval from the North West Multi-centre Research Ethics Committee (MREC) as a Research Tissue Bank (RTB) approval. This approval means that researchers do not require separate ethical clearance and can operate under the RTB approval. Approval number for the UK Biobank is 11/NW/0382. Data accessibility Code and data for all parts of our analysis is provided at the Dryad repository [35]. Supplementary material is available online [36]. Declaration of AI use We have not used AI-assisted technologies in creating this article. Authors’ contributions C.N.: conceptualization, data curation, formal analysis, investigation, methodology, project administration, software, validation, visualization, writing—original draft, writing—review and editing; M.W.: funding acquisition, methodology, writing—review and editing; N.D.F.: funding acquisition, methodology, supervision, writing—review and editing. All authors gave final approval for publication and agreed to be held accountable for the work performed therein. Conflict of interest declaration We declare we have no competing interests. Funding This research was funded by the Lions Sight Centre and the Canada Research Chairs Program.

- 1. Alowais SA, et al. 2023. Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC Med. Educ. 23, 689. ( 10.1186/s12909-023-04698-z) [DOI] [PMC free article] [PubMed] [Google Scholar] - 2. Quinn JG, Tansey EA, Johnson CD, Roe SM, Montgomery LEA. 2016. Blood: tests used to assess the physiological and immunological properties of blood. Adv. Physiol. Educ. 40, 165–175. ( 10.1152/advan.00079.2015) [DOI] [PubMed] [Google Scholar] - 3. Shaw JLV. 2016. Practical challenges related to point of care testing. Pract. Lab. Med. 4, 2015. ( 10.1016/j.plabm.2015.12.002) [DOI] [PMC free article] [PubMed] [Google Scholar] - 4. Heidt B, Siqueira WF, Eersels K, Diliën H, van Grinsven B, Fujiwara RT, Cleij TJ. 2020. Point of care diagnostics in resource-limited settings: a review of the present and future of poc in its most needed environment. Biosensors(Basel) 10, 133. ( 10.3390/bios10100133) [DOI] [PMC free article] [PubMed] [Google Scholar] - 5. Yuan A, Lee CS. 2022. Retinal biomarkers for Alzheimer disease: the facts and the future. Asia Pac. J. Ophthalmol. (Phila). 11, 140–148. ( 10.1097/APO.0000000000000505) [DOI] [PMC free article] [PubMed] [Google Scholar] - 6. Rim TH, et al. 2020. Prediction of systemic biomarkers from retinal photographs: development and validation of deep-learning algorithms. Lancet Digit. Health 2, e526–e536. ( 10.1016/S2589-7500(20)30216-8) [DOI] [PubMed] [Google Scholar] - 7. Silverstein SM, Demmin DL, Schallek JB, Fradkin SI. 2020. Measures of retinal structure and function as biomarkers in neurology and psychiatry. Biomarkers Neuropsychiatry 2, 100018. ( 10.1016/j.bionps.2020.100018) [DOI] [Google Scholar] - 8. Song G, Steelman ZA, Finkelstein S, Yang Z, Martin L, Chu KK, Farsiu S, Arshavsky VY, Wax A. 2020. Multimodal coherent imaging of retinal biomarkers of Alzheimer’s disease in a mouse model. Sci. Rep. 10, 7912. ( 10.1038/s41598-020-64827-2) [DOI] [PMC free article] [PubMed] [Google Scholar] - 9. Kim YD, et al. 2020. Effects of hypertension, diabetes, and smoking on age and sex prediction from retinal fundus images. Sci. Rep. 10, 4623. ( 10.1038/s41598-020-61519-9) [DOI] [PMC free article] [PubMed] [Google Scholar] - 10. Zhang K, et al. 2021. Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images. Nat. Biomed. Eng. 5, 533–545. ( 10.1038/s41551-021-00745-6) [DOI] [PubMed] [Google Scholar] - 11. MacEachern SJ, Forkert ND. 2021. Machine learning for precision medicine. Genome 64, 416–425. ( 10.1139/gen-2020-0131) [DOI] [PubMed] [Google Scholar] - 12. Zhu Z, et al. 2023. Retinal age gap as a predictive biomarker for mortality risk. Br. J. Ophthalmol. 107, 547–554. ( 10.1136/bjophthalmol-2021-319807) [DOI] [PubMed] [Google Scholar] - 13. Chen R, Xu J, Zhang X, Zhang J, Shang X, Ge Z, He M, Wang W, Zhu Z. 2023. Glycemic status and its association with retinal age gap: insights from the UK biobank study. Diabetes Res. Clin. Pract. 202, 110817. ( 10.1016/j.diabres.2023.110817) [DOI] [PubMed] [Google Scholar] - 14. Zhu Z, et al. 2022. Retinal age gap as a predictive biomarker of stroke risk. BMC Med. 20, 466. ( 10.1186/s12916-022-02620-w) [DOI] [PMC free article] [PubMed] [Google Scholar] - 15. Zhang S, et al. 2023. Association of retinal age gap and risk of kidney failure: a UK biobank study. Am. J. Kidney Dis. 81, 2022. ( 10.1053/j.ajkd.2022.09.018) [DOI] [PubMed] [Google Scholar] - 16. Hu W, et al. 2022. Retinal age gap as a predictive biomarker of future risk of Parkinson’s disease. Age Ageing 51, c062. ( 10.1093/ageing/afac062) [DOI] [PMC free article] [PubMed] [Google Scholar] - 17. Karakaya M, Hacisoftaoglu RE. 2020. Comparison of smartphone-based retinal imaging systems for diabetic retinopathy detection using deep learning. BMC Bioinform. 21, 259. ( 10.1186/s12859-020-03587-2) [DOI] [PMC free article] [PubMed] [Google Scholar] - 18. Bycroft C, et al. 2018. The UK biobank resource with deep phenotyping and genomic data. Nature 562, 203–209. ( 10.1038/s41586-018-0579-z) [DOI] [PMC free article] [PubMed] [Google Scholar] - 19. Vos T, et al. 2020. Global burden of 369 diseases and injuries in 204 countries and territories, 1990–2019: a systematic analysis for the global burden of disease study 2019. Lancet 396, 1204–1222. ( 10.1016/s0140-6736(20)30925-9) [DOI] [PMC free article] [PubMed] [Google Scholar] - 20. Nakayama LF, Restrepo D, Matos J, Ribeiro LZ, Malerbi FK, Celi LA, Regatieri CS. 2024. BRSET: a brazilian multilabel ophthalmological dataset of retina fundus photos. MedRxiv 01.23.24301660. ( 10.1101/2024.01.23.24301660) [DOI] [PMC free article] [PubMed] [Google Scholar] - 21. Fu H, et al. 2019. Evaluation of retinal image quality assessment networks in different color-spaces. Lecture Notes Comp. Sci. 11764, 48–56. ( 10.1007/978-3-030-32239-7_6) [DOI] [Google Scholar] - 22. Tan M, Le QV. 2020. Efficientnet: rethinking model scaling for convolutional neural networks. arXiv. ( 10.48550/arXiv.1905.11946) [DOI] - 23. de Lange AMG, Cole JH. 2020. Commentary: correction procedures in brain-age prediction. NeuroImage 26, 102229. ( 10.1016/j.nicl.2020.102229) [DOI] [PMC free article] [PubMed] [Google Scholar] - 24. Walters SJ, Campbell MJ. 2004. The use of bootstrap methods for analysing health-related quality of life outcomes (particularly the SF-36). Health Qual. Life Outcomes 2, 70. ( 10.1186/1477-7525-2-70) [DOI] [PMC free article] [PubMed] [Google Scholar] - 25. Huang J, Gretton A, Borgwardt K, Schölkopf B, Smola A. 2006. Correcting sample selection bias by unlabeled data. Adv. Neural Inf. Processing Syst 19. ( 10.7551/mitpress/7503.003.0080) [DOI] [Google Scholar] - 26. Armstrong RA. 2014. When to use the Bonferroni correction. Ophthalmic Physiol. Opt. 34, 502–508. ( 10.1111/opo.12131) [DOI] [PubMed] [Google Scholar] - 27. Ebrahimi M, Thompson P, Lauer AK, Sivaprasad S, Perry G. 2023. The retina–brain axis and diabetic retinopathy. Eur. J. Ophthalmol. 33, 2079–2095. ( 10.1177/11206721231172229) [DOI] [PubMed] [Google Scholar] - 28. Li Q, Wang M, Li X, Shao Y. 2023. Aging and diabetic retinopathy: inherently intertwined pathophysiological processes. Exp. Gerontol. 175, 112138. ( 10.1016/j.exger.2023.112138) [DOI] [PubMed] [Google Scholar] - 29. Bennett JL, Costello F, Chen JJ, Petzold A, Biousse V, Newman NJ, Galetta SL. 2023. Optic neuritis and autoimmune optic neuropathies: advances in diagnosis and treatment. Lancet Neurol. 22, 89–100. ( 10.1016/s1474-4422(22)00187-9) [DOI] [PubMed] [Google Scholar] - 30. Karimi S, Arabi A, Shahraki T. Alcohol and the eye. J. Ophthalmic Vis. Res. 16, 260–270 (2021. ( 10.18502/jovr.v16i2.9089) [DOI] [PMC free article] [PubMed] [Google Scholar] - 31. Levine ME, et al. 2018. An epigenetic biomarker of aging for lifespan and healthspan. Aging 10, 573–591. ( 10.18632/aging.101414) [DOI] [PMC free article] [PubMed] [Google Scholar] - 32. MacGillivray TJ, et al. 2015. Suitability of UK Biobank retinal images for automatic analysis of morphometric properties of the vasculature. PLoS One 10, e0127914. ( 10.1371/journal.pone.0127914) [DOI] [PMC free article] [PubMed] [Google Scholar] - 33. Raj A, Tiwari AK, Martini MG. 2019. Fundus image quality assessment: survey, challenges, and future scope. IET Image Process. 13, 1211–1224. ( 10.1049/iet-ipr.2018.6212) [DOI] [Google Scholar] - 34. Fu J, Cao L, Wei S, Xu M, Song Y, Li H, You Y. 2022. A GAN-based deep enhancer for quality enhancement of retinal images photographed by a handheld fundus camera. Adv. Ophthalmol. Pract. Res. 2, 100077. ( 10.1016/j.aopr.2022.100077) [DOI] [PMC free article] [PubMed] [Google Scholar] - 35. Nielsen C, Wilms M, Forkert N. 2025. Code from: The retinal age gap: an affordable and highly accessible biomarker for population-wide disease screening across the globe. Dryad Digital Repository. ( 10.5061/dryad.f7m0cfz72) [DOI] [PubMed] - 36. Nielsen C, Wilms M, Forkert ND. 2025. Supplementary material from: The retinal age gap: an affordable and highly accessible biomarker for population-wide disease screening across the globe. Figshare. ( 10.6084/m9.figshare.c.7745636) [DOI] [PubMed] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement Code and data for all parts of our analysis is provided at the Dryad repository [35]. Supplementary material is available online [36].