Retinal image‐based artificial intelligence in detecting and predicting kidney diseases: Current advances and future perspectives

Artificial intelligence (AI) has reformed the healthcare system with its compelling capabilities of processing biomedical data for disease diagnosis, prediction, and individualized management. The eye, as a non‐invasive observation window for many systemic diseases, can be used to detect the signs of chronic kidney diseases, and other diseases like hypertension and type 2 diabetes mellitus, based on specific manifestations of retinal images. Recent advances using AI technology have posed a great potential of using retinal images for rapid mass screening and prognosis prediction of kidney diseases. Herein, we outlined the key applications of AI in ophthalmology and the detection of systemic diseases based on retinal imaging, especially the current progress of retinal image‐based AI models for the detection and prediction of kidney diseases. We hope to shed light on the current opportunities and future challenges in this field to provide suggestions for further improvement and applications.


INTRODUCTION
Artificial Intelligence (AI), which sprung from Alan Turing's paper 1,2 in 1950, has revolutionized numerous realworld industries, such as translation, natural language processing, and image recognition after decades of development; and medicine is no exception. Machine learning (ML) and its important subset deep learning (DL) achieved prominent performance in multiple tasks of medical image processing: classification, object detection, tracking, and semantic segmentation. 3 Powered by DL, AI is now used as a classifier and screening tool in some realtime in-vivo imaging processes, and also provides referral advice and reference diagnostics in image-centric specialties, such as ophthalmology, 4,5 radiology, 6,7 dermatology, 8 and cardiology. 9,10 In ophthalmology, several AI-based diagnostic platforms based on retinal imaging such as IDx-DR 11,12 and EyeArt 13 AI Screening System for diabetic retinopathy (DR) have shown great application advantages in clinical practice. To date, the US Food and Drug Administration (FDA) has approved more than 300 AI medical products, and ophthalmic disease screening applications are among the first to be approved. Given the rich information on pathological changes and accessibility of retinal images in primary care settings, retinal imaging-based AI models are widely used for detecting and predicting diseases such as type 2 diabetes mellitus (T2DM), 4 cardiovascular diseases (CVD), 14 hepatobiliary diseases, 15 and chronic kidney disease (CKD). 16 Kidney diseases, especially CKD, have become a major public health problem worldwide for their high prevalence and poor prognosis, with more than 10% of adults showing evidence of declined kidney function. 17 As symptoms do not usually manifest in the early stages of CKD, less than 10% of patients with CKD realize their diseases before their progression to end-stage renal disease (ESRD). 18 Based on the close tie between the eye and kidney, many DL models are built to screen and grade CKD from retinal images, making early diagnosis of CKD possible and easy to conduct in practice. Evaluation of the therapeutic effect of ESRD patients and prediction of disease progression expands the potential benefit of retinal imaging-based AI in personalized health care and large-scale management. Due to the aging of the world's population, overall shortage and an unbalanced distribution of medical resources, and low health knowledge awareness, there is a huge demand and application prospect for AI in chronic diseases.
In this review, we will introduce the applications of retinal image-based AI technologies in ophthalmology and eye-related systemic diseases first. Next, the internal link between the eye and kidney, with their common pathological mechanism and clinical manifestations, is discussed in brief. Then, recent efforts of retinal image-based AI models for the detection, diagnosis, and prognosis prediction of kidney diseases are introduced separately (Scheme 1). In the end, we will discuss the challenges and future perspectives in the field. This review aims to give a relatively comprehensive demonstration of the promising AI strategies for the individualized diagnosing, and prognostic evaluations of kidney disease patients via retinal imaging data.

S C H E M E 1
Graphical schematic illustration of retinal image-based artificial intelligence for the detection, diagnosis, and prediction of kidney diseases.

RETINAL IMAGING AND AI APPLICATIONS IN OPHTHALMOLOGY
As an image-centric specialty, the process of diagnosis and treatment of ophthalmology generates massive multimodal image data from anterior to posterior segments of the eye, well-suited for developing AI models. Trained with anterior-segment optical coherence tomography (AS-OCT) and slit-lamp images, AI achieves robust performance in the detection of keratoconus, 19,20 glaucoma, 21,22 and cataracts. 5,23 Moreover, imaging of the eye fundus as a noninvasive, cost-effective method, with more information on microvascular dysfunction and neuropathy to prevent severe vision loss and blindness, are readily available in primary clinics. Currently, AI models concerning ophthalmology focused on the diagnosis and management of posterior segment diseases based on retinal imaging.

Retinal imaging used in ophthalmology
There are several imaging methods used for diagnosing and monitoring eye diseases: 1) Color fundus photography: a non-invasive method that uses special cameras to take photographs of the eye fundus (retina) and blood vessels. 2) Optical coherence tomography (OCT): a noninvasive imaging test that uses light waves to show crosssectional scans of the interlayer structure of the retina and thickness measurement of the retina and its nerve fiber layer. 3) OCT angiography (OCT-A): a novel technology using laser light reflectance to image the microvasculature of the retina and the choroid while also providing cross-sectional structural information with B-scans. 4) Fluorescein angiography (FA): injecting a fluorescent dye into the bloodstream and taking photo series of the blood vessels in the eye. 5) dynamic retinal vessel analysis (DVA): observation of retinal vessel diameter changes to physiological stimuli. Each of these imaging methods has its own strengths and is used in different clinical situations and AI models to diagnose and monitor different eye diseases ( Figure 1).

Retinal imaging used in AI models
Unprecedented new insight into the status of the retina and beyond ocular illness is being provided by significant advancements in retinal imaging technologies. Many modalities of retinal imaging are now clinically available, including, but not limited to retinal color fundus photographs, ultrasound, OCT, OCT-A, FA, and DVA. It should be noted that retinal imaging usually includes information from the underlying choroid, which is closely adjacent to the retina structurally and functionally. According to the types of retinal images, different datasets are built for retinal photographs (EyePACS), 14 OCT images, 24 or both (UK Biobank). 25 Retinal photography is the most widely used retinal imaging technique, followed by OCT and OCT-A in training AI models. 26 The advantage of retinal photography lies in its low cost and easy feasibility with widespread implementation in point-of-care testing and primary care settings. From typical 30-to 50-degree field to ultra-widefield cameras, then hand-held and smartphone-based cameras, fundus cameras have evolved over time. OCT and its advancement, OCT-A, represent non-invasive in vivo optical sections of retinal vascular and neural structure with high speed and resolution. OCT and OCT-A could provide an abundance of clinical data in the form of both images and quantitative metrics for training AI models, though the segmentation and quantification of vasculature remain big obstacles.

Retinal image-based AI applications in ophthalmology
With the growing need for processing multimodal data and screening in massive populations, DL-powered retinal AI models have been adopted in ophthalmology to solve these problems. After the first DL algorithm for DR screening by Google was published in 2016, 4 DL models begin to be used in the screening and diagnosis of retinal diseases, such as age-related macular degeneration (AMD) 27 and glaucoma. 28,29 These models were built with retinal photographs, OCT, or OCT-A, and performed as well by ophthalmologists. Other specific signs including narrowing retinal vessel calibers, papilledema, and optic disc cupping 30 can also be well-detected in different studies. Monitoring and predicting disease progression and treatment response are also problems for physicians. To address this challenge, AI is applied to the progression prediction of neovascular AMD 31 or diabetic macular edema 32 using OCT, and also in the prediction of intraocular pressure and visual field in glaucoma patients. 33 For the management of diseases, Artificial intelligence-assisted interaction, and intelligent-assisted judgment, as a collection of doctorpatient networking platforms to assist physicians and staff to complete remote visits, referral advice, followup reminders, and chronic disease management through telephone, messages, or online applications. 34 A semiautomated model was developed by Xie et al. 35 to classify retinal images using AI, and those annotated as abnormal are reviewed by human graders with teleophthalmology.
As the first explorative stage of AI development in ophthalmology, though most of these studies show promising results, it should be noted that none of these previous studies have further tested their respective algorithms in external test sets. Additionally, limited by the need for additional infrastructure and manpower to implement, it is still up for debate as to how these algorithms perform in real-world settings, like primary care clinics in rural and remote areas.

RETINAL IMAGE-BASED AI MODELS FOR SYSTEMIC DISEASES
Retina is a perfect window for systemic diseases: it could provide non-invasive, direct visualization of microvascular circulation and the central nervous system, reflecting signs and severities of systemic diseases. 36 With automated feature extraction from retinal images to form "retinal fingerprint" of diseases, 37 AI enables retinal imaging for screening and diagnosing major systemic diseases; including diabetes mellitus, CVDs, and Alzheimer's disease, it also paves the way for future applications in kidney and other diseases.

Similar bases of the eye and major systemic diseases
Embryologically, the retina originates from the neuroectoderm, being an integral component of the nervous system, 38 while the choroid derives from the mesoderm, comprised of a part of the circulatory and immune system. 39 Anatomically, the retinal vascular structure is lined by a single continuous layer of non-fenestrated endothelial cells that resemble the cerebral vessels, while the fenestrated capillaries exit in the choroid and glomeruli of the kidneys. The retinal layers with cell body, dendrites, and axons, share similar structures of neurons in the central nervous system. Physiologically, retinal circulation exits autoregulation in response to fluctuations in blood pressure 40 but fails in DR. 41 Without autoregulation, the blood flow of choroids could reflect varying systemic blood pressure levels 42 and carbon dioxide tension. 43

Application of retinal image-based AI models in diabetes
Currently, 425 million people are affected by diabetes mellitus worldwide. 44 And at least one-third of patients with diabetes suffer from DR, leading to different stages of microaneurysms, exudates, hemorrhages, and eventually irreversible blindness. A DL algorithm developed by Gulshan et al. 4 can check for signs of retinal photographs to help physicians screen more patients. A dataset of more than 128,000 images was created to train a deep neural network to detect DR, achieving 90.3% and 87.0% sensitivity and 98.1% and 98.5% specificity in 2 validation sets, performed with comparable accuracy as the ophthalmologists. Later numerous AI models detecting DR were developed and validated in multiethnic populations, 45 multi-continent levels, 46 and also in different levels of care. 47 The system was put in a primary setting to help predict the DR development with an area under the curve (AUC) of 0.79 (95% confidence interval [CI]: 0.77-0.81) in the internal validation set and 0.70 in the external validation set. And the economic analysis has proved that the AI screening of DR combined with specialists can reach the lowest cost, which is US$62 per patient per year, compared to $66 for a fully automated model and $77 for the human assessment model. 48 However, real-world studies have somehow shown that AI screening fell short in clinical practice, especially in remote rural areas with few medical resources. The project is being carried out in Thailand, where the same DL system has been installed in 11 clinics in the provinces of Pathum Thani and Chiang Mai. 49 With inexperienced operators and unmet conditions, acquired images were somewhat different from the images required by the algorithm, and connect speed impeded the future applications of algorithms.

Application of retinal image-based AI models in CVDs
Certain CVDs can be presented with specific ocular manifestations, the most typical one being hypertensive retinopathy with narrowing retinal arteriolar calibers, arteriovenous nicking, exudates, and hemorrhages for different phases. 50 Therefore, retinal images, including retinal photographs and OCT, could be used to detect and stratify the risk of CVDs in the population. At first, retinal vessel diameters were assessed through manual grading or measured by semi-automated software. 51  In recent years, the applications of AI in the diagnosis of CVD including hypertension and myocardial infarction have also been substantially developed and demonstrate its huge potential for clinical use. Researchers includ-ing Kim et al., 52 Cheung et al., 53 and Diaz-Pinto et al., 54 employed retinal fundus images to predict hypertension, CVD risk, and myocardial infarction by predicting left ventricular mass and left ventricular end-diastolic volume, respectively.

Application of retinal image-based AI models in Alzheimer's disease
Alzheimer's disease (AD) is another major hurdle that scientists pay a lot of effort to overcome. 55 Certain retinopathy signs (e.g., microaneurysms, hard exudates, and ischemic signs) are associated with stroke incidence 56 and mortality, 57 and the retinal microvasculature change shares a link with dementia, in particular AD. 58 Wisely et al. 59  respectively). However, the generalizability and accessibility of the model remained unclear given the model contains a small group of patients without known ocular diseases and requires multiple high-cost examinations with a pre-screening procedure to exclude poor-quality images.

THE EYE/KIDNEY CONNECTION: FROM MOLECULAR TO CLINICAL
The eye is a window for direct observation of many diseases, and the eye and kidney share similar development, structure, and pathogenic mechanisms. The alteration of ocular vascular structure and function is closely related to different kinds of kidney diseases, including diabetic nephropathy and hypertension nephropathy. Herein kidney diseases can be studied non-invasively by the in vivo visualization of the retinal microvasculature.

Shared physiological and molecular pathways of eye and kidney diseases
The link between kidney diseases and blindness was first reported by Richard Bright in 1836. 60 Nowadays, scientists identify that both choroids of the eye and kidney develop from mesoderm around the fourth week, and they can be both affected by the same factors during this period. Basically, nephrons and choroids share some similar structures consisting of nerve fibers, stroma, and fenestrated microvessels, which are highly permeable to proteins and result in stroma edema. They both expressed renin-angiotensin-aldosterone receptors, 61 innervated by the autonomic nervous system, and regulated by systemic perfusion pressure ( Figure 2). 39,62 Eye and kidney diseases have common embryonic, structural, pathogenic, and underlying molecular pathways. A variety of human congenital oculorenal syndromes have been described. 63 Pax genes consist of a group of gene families that control development in encoding nuclear transcription factors, such as Waardenburg syndrome (Pax3), Aniridia (Pax6), Peter's anomaly (Pax6), and renal coloboma syndrome (Pax2). 64 Bone morphogenetic protein-7 (BMP-7) 41 is from another important growth factor-secreting family, the transforming growth factor-beta (TGF-beta) family. During embryogenesis of the mammalian kidney and eye, they would be both affected by the lack of BMP-7, exhibiting renal dysplasia and anophthalmia. Wilms' tumor suppressor (WT1) 65 could also cause aniridia, cataract, corneal clouding, and nephroblastoma (Wilms' tumor) in children.

Epidemiologic and clinical association of eye and kidney diseases
Common risk factors of old age, smoking, diabetes mellitus, hyperlipidemia, and hypertension were reported shared between CKD and ocular problems, either alone or in combination. 66 Clinically, the comorbid eye diseases in patients with CKD and ESRD have some common manifestations. 67,68 Retinopathies and certain ocular fundus signs of cotton-wool spots and hemorrhages in superficial or deep capillaries of the retina were often found in CKD patients. Systemic or retinal edema is caused by hypoalbuminemia and retention of excess fluid accumulation in interstitial spaces or subretinal spaces. 69 Renal anemia may cause retinal hypoxia, which leads to infarction of the nerve fiber layer and results in cotton wool spots. And the elevated blood cholesterol is associated with increased severity and extent of hard exudates.
Epidemiologic studies reveal a high prevalence of retinopathies in patients with CKD, independent of diabetes, hypertension, and other risk factors. 70 A population-based cross-sectional study of 9670 Chinese participants in Beijing by Gao et al., 71 found that retinopa-thy is more prevalent in patients with CKD (28.5% vs. 16.3%, p < 0.001). Also, the severity of retinopathy is associated with the level of kidney function loss (estimated glomerular filtration rate, eGFR) and predicts the future progression of CKD. In the Chronic Renal Insufficiency Cohort (CRIC) study 70 of 1904 CKD patients in the United States found CKD patients with eGFR <30 ml/min/1.73m 2 are associated with three times higher risk of retinopathies. A later serial study of 1583 CRIC participants showed that with the progression of retinopathy, the odds ratio (OR) for CKD progression was 2.24 (95% CI: 1.28-3.91) compared with participants with stable retinopathy, indicating a bidirectional link between the severity of kidney and eye diseases.
Most kidney diseases and retinopathies start or worsen from vascular dysfunction and systemic diseases like diabetes mellitus and hypertension. 72 The narrowing diameters of arterioles and widening of venules are thought to reflect the severity of both diseases. 73 In the past two decades, studies have been focused on the qualitative measurement of retinal vessel calibers, including central retinal arteriolar equivalent (CRAE), central retinal venular equivalent (CRVE), arteriole-to-venule ratio (AVR), and the analysis of vessel network geometry, fractal dimension. [66][67][68][69] Bao et al. 74 aimed to examine the association between CRAE, CRVE, and AVR with CKD in rural China to provide the scientific basis for the early detection and diagnosis of CKD. After controlling for potential confounding factors, participants with the first quartile of AVR were shown to have a greater risk of albuminuria (OR = 1.261, 95% CI: 1.02-1.57) and CKD (OR = 1.240, 95% CI: 1.00-1.54) than the fourth quartile. It was investigated by Sabanayagam et al. 75 if baseline retinal vessel diameters were related to future risk of impaired kidney function or vice versa, in prospective research involving numerous assessments of serum creatinine and retinal artery diameters (CRAE and CRVE). Retinopathy was also assessed whether it could predict future CVD events in participants of the CRIC study. Worsening early treatment DR study retinopathy scores were reported in 9.8% of individuals and were related to an elevated risk of incidence of any CVD (OR = 2.56, 95% CI: 1. 25-5.22). 76 A large scale of studies show that these retinal vascular indices are associated with CKD prevalence, but their abilities to predict CKD incidence or progression are equivocal. 75,[77][78][79] The measurement of consistent retinal vascular metrics may overlook the heterogeneity of different study populations and other information on fundus signs (e.g., microaneurysms, exudates, hemorrhages, and papilledema). And it also suggests that the sensitivity of conventional analysis of fundus photography cannot meet the need for precise identification of patients at risk.

F I G U R E 2
The eye as a window to the kidney: common microcirculation organization and physiological regulation in kidney and eye. AT1R, angiotensin II type 1 receptor; CRA, central retinal artery; CRV, central retinal vein; ET, endothelin; ETAR, endothelin type A receptor; ETBR, endothelin type B receptor; pO2, partial pressure of oxygen; RAAS, renin-angiotensin-aldosterone system; RNFL, retinal nerve fiber layer. Source: Copyright 2020, Elsevier.

DETECTING AND DIAGNOSING KIDNEY DISEASES FROM RETINAL IMAGING
Kidney diseases, like glomerulonephritis and acute kidney injury, usually entail some decline in renal function, even-tually developing CKD or kidney failure over time. CKD is the most common type of kidney disease. The high prevalence of CKD has become a major public health problem worldwide, affecting almost 10%-16% of the population in Asia, Europe, Australia, and the United States. 17,80 With an increasing incidence of obesity and T2DM, combined with an aging population worldwide, a heavier burden is expected in the future decades, especially in low-and middle-income countries. 81 Despite its severity, CKD is usually asymptomatic and progresses insidiously until the very late stage or ESRD, leading to complications like CVD and higher mortality. The idea of early detection is appealing because the inexpensive intervention can ameliorate renal function at an early stage and prevent continuous deterioration. The shortcomings of traditional analysis methods make a new comprehensive assessment approach urgently needed in kidney disease patients and a larger scale of population. The use of ML and DL combined with retinal imaging is a relatively prominent frontier in this research area (Figure 3).

Screening for risk factors of CKD
The increasing prevalence of kidney diseases worldwide is closely associated with risk factors, including smoking, obesity, hypertension, and metabolic syndromes like hyperlipidemia and diabetes. 82  Other models were also developed to detect different factors related to kidney diseases, like dyslipidemia with elevated triglyceride (MAE = 0.49, R 2 = 0.03), total cholesterol (MAE = 0.75, R 2 = 0.03) and low-density lipoprotein cholesterol (LDL-c, MAE = 0.72, R 2 = -0.03). 80 The performance of the models was relatively poor in test sets and needed to be improved. Classification models of dyslipidemia 81 identification show a better ability with an accuracy of 66.7% and an AUC of 0.703.
Diabetes and hypertension are common underlying conditions associated with CKD. In patients with diabetes, the prevalence of CKD is approximately 30% to 40%, and F I G U R E 3 An illustration of the general workflow of retinal detection and prediction of kidney diseases. First, input data includes retinal information and metadata (text, electrical health record (EHR) and results from ultrasound or magnetic resonance imaging (MRI)), second: data processing (e.g., missing data imputation, image enhancement, etc.), third: feature selection, forth: model building with different machine learning (ML)/deep learning (DL) approaches, fifth: output classified into detection, diagnosis, and treatment of kidney diseases, and their applications in clinical practice and research. EHR, electrical health record; SVM, support vector machines; RF, random forest; MLP, multilayer perceptron; KNN, k-nearest neighbors; ANN, artificial neural network; RRT, renal replacement therapy. the numbers are rising with socioeconomic change and population aging. And hypertension accounts for 30% of all kidney diseases, only secondary to diabetes, leading to hypertensive nephropathy. The renal injury can be insidious and progressive, but the retinal signs can be the earliest findings to identify target organ damages, even before the presence of proteinuria and decrease of eGFR. AI models for the detection and diagnosis of diabetes and hypertension are discussed in Section 3, and early identification

Alerting signs of early kidney diseases
CKD is often insidious, with patients remaining asymptomatic (stages G1 and G2) for long periods and consequential low awareness. When progressing to stage G3, a symptomatic stage with polyuria or fatigue caused by anemia, patients are at a significantly higher risk of complications and progression to ESRD. The assessment of kidney function chiefly depends on the level of GFR, and serum creatinine concentration is generally measured to calculate the value of eGFR by specific formula (for example, the CKD-EPI formula). CKD could also be alerted by abnormal laboratory blood tests of urea nitrogen, cystatin C, and proteinuria/microalbuminuria from urine tests in routine health checks or population screening.
Patients with increased blood urea nitrogen and creatinine were associated with the incidence of posterior subcapsular cataract 84 86 with moderate performance (MAE = 0.11, R 2 = 0.12) in the Korean dataset, however, it falls far short of robustness in multi-ethnic datasets (R 2 = 0.06, 0.01 for Singapore Epidemiology of Eye Diseases and UK Biobank, respectively). Serum creatinine level may be affected by age, food intake, inflammation status, and drugs, the calculated eGFR was adopted as a more reliable indicator, whose decline could be observed at the early stage of kidney injury. Kang et al. 87 attempted to predict early renal impairment, defined as eGFR < 90 ml/min/1.73m 2 , trained and tested with 25,706 fundus images from 6212 patients ( Figure 4A). The AUC was 0.81 in the overall population, higher in subgroups of elevated serum hemoglobin A 1c (HbA 1c ) level (0.81, 0.84, 0.85, and 0.87 for HbA 1c level of ≤6.5%, >6.5%, >7.5%, and >10%, respectively), but observed poor specificity (60%) in diabetes patients. The performance of this model lacked external validation, and it should be noted that the concerned retinal-vessel features, and abnormal retina signs marked by saliency maps, like exudation and hemorrhage, may be due to pathological changes caused by other ophthalmic or systemic comorbidities. Proteinuria or microalbuminuria, which is defined as albumin/creatinine ratio (ACR) >17.0 for males or ACR>25.0 for females, is recommended as the preferred screening strategy for all patients with diabetes and hypertension. Lim et al. 88 quantitatively measured fundus photographs by a semiautomated program to predict the presence of albuminuria, and the retinal vascular parameters were considered associated with renal function from regression models, showing better discriminative ability than the traditional risk factors model (AUC 0.80 vs. 0.77).

Diagnosing and grading CKD
For the early diagnosis of CKD in community and primary care clinics, two important studies have pointed out a new direction for renal disease screening mode based on retinal fundus images. In 2020, Sabanayagam et al. 89 developed and validated a DL algorithm to detect CKDs in three population-based, multi-ethnic datasets (12,970 participants in total). In this study, multi-modal data were added into three models: a model of only fundus images, a model of CKD risk factors (age, sex, ethnicity, diabetes, and hypertension), and a hybrid model combining images and risk factors ( Figure 4B). They all showed similar good performance in the internal validation set (AUC = 0.911, 0.916, and 0.914, respectively) and moderate performance in the external validation cohort (AUC = 0.733-0.858). In subgroups of diabetes and hypertension patients, AUC was estimated similar to the whole population. Given that both the image-only model and the risk-factor model could predict CKD with high AUCs, and their combination only slightly improved the performance, fundus images alone could be employed as an auxiliary or opportunistic screening tool for CKDs in community populations, even without the collection of patient information. While the positive predictive value was high (54%) in the internal validation set, it was only 14% and 9% in external validation sets due to the high prevalence of CKD in the internal validation set. To improve the positive predictive value, the image-only model is recommended only to be applied to high-risk groups, such as diabetes and hypertension patients, however, this would limit its clinical utility in a larger population. Following this study, Zhang et al. 16 built a DL model using ResNet-50 to identify CKD and early CKD patients only with fundus images or combined with clinical metadata (age, sex, height, weight, BMI, and BP). The models were trained and validated with 115,344 images from 57,672 patients. In the diagnosis model of CKD, the AUCs were 0.861 for the metadata model, and 0.918 for the model with only retinal images ( Figure 4C). If the two models are combined, the accuracy rate will further increase (AUC = 0.930), providing better solutions for capturing features of CKDs. Then patient's eGFR, the key index of renal function and grading of CKD, was predicted using retinal fundus images with R 2 of 0.327-0.507, and MAE of 11.1-13.4 ml/min/1.73m 2 , showing consistency with the measured eGFR (intraclass correlation coefficient [ICC] = 0.65, 95% CI: 0.63-0.66). For the grading of CKD, the thresholds of predicted eGFR were used to train regression models of differentiating advanced (stage G3) or severe+ CKD (stages G4 and G5) from early CKD (stages G1 and G2), which achieved good performance with AUCs of 0.853 and 0.825.
Images captured by smartphone-based devices were introduced as an external test set to further evaluate the generalizability of the AI model and assess the feasibility of a massive population in a longitudinal cohort, achieving comparably good CKD detection performance (AUC = 0.897, 95% CI: 0.85-0.89, for image-only model) and noninferior performance of eGFR prediction (ICC = 0.53, 95% CI: 0.50-0.55). The deployment of smartphone-and cloudbased AI diagnosis systems could build a bidirectional relationship between patients and clinics to improve the feasibility and broaden healthcare access by encouraging patients to self-monitor and allowing doctors to diagnose and follow up with patients remotely.
However, current studies remain at the first stage of CKD diagnosis, only concerned about the qualitative value of eGFR, and did not include other important biomarkers for kidney function, like microalbuminuria or electrolytes. For CKD patients with different causes, such as glomerulonephritis or diabetic nephropathy, their treatment, and prognosis may vary. That's why the etiologic diagnosis (the so-called etiologic/GFR/albuminuria classification system) is emphasized. This helps to prioritize the diagnosis and treatment of the underlying disease to slow down the progression of CKD. (Table 1)

GUIDING TREATMENT AND PREDICTING PROGNOSIS FROM RETINAL IMAGING
The use of routinely collected laboratory values can help with clinical decisions: guiding individual treatment, evaluating the need for more intensive interdisciplinary clinic care, and determining the timing of renal replacement treatment (RRT). The development of new models that predict the treatment response, survival, and quality of life of patients, success on dialysis, and risk of CVD and other complications in patients with CKD is needed. Given that prevalence of ophthalmic diseases in CKD patients is as high as 32.0%, a regular ophthalmic examination was recommended, and this information could be an addition to the traditional prediction method.

Individual management for common complications of CKD
The kidney is critical in maintaining volume, bone health, acid-base balance, electrolyte stability, and blood pressure. Decreased kidney function can lead to several serious complications such as anemia, hyperkalemia, mineral bone disease, hyperphosphatemia, hypertension, and CVDs. Renal anemia plays an important factor in affecting the quality of life and prognosis of the patients. Chen et al. 90 applied retinal vessel images from OCT to directly and non-evasively observe the hemoglobin concentration in the retina for the first time. Later, Mitani et al. 91 and Zhao et al. 92 leverage retinal fundus images to predict hemoglobin concentration and detect the status of anemia accurately (MAE: 0.67, AUC: 0.87; MAE: 0.83, AUC: 0.93, respectively). For the predictions of electrolyte disorders, sodium was predicted with an R 2 of 0.12 (95% CI: 0.10-0.13), while the performance was poor for potassium, calcium, and phosphorus (R 2 = 0.07, 0.07, and 0.02, respectively). 86 These studies indicate that retinal imaging has clinical potential to screen for and monitor renal anemia in patients with CKD, which could aid in deciding on treatment plans and improving patient outcomes.

Evaluating the treatment effect and predicting the complications of dialysis
When CKD patients progress to the end stage, their renal function fails to meet the basic need for body operation, and RRT is essentially needed, which consists of hemodialysis, peritoneal dialysis, and kidney transplant. The most common treatment for ESRD patients is hemodialysis, however, the occurrence of complications (intradialytic hypotension [IDH], adverse cardiovascular events, renal anemia, etc.) leads to dialysis failure and poor patient survival. To evaluate the treatment effect and prevent complications, many indicators were used to monitor and evaluate patients' fluid and metabolic status, which includes laboratory blood tests of creatine and bioimpedance for volume assessment. Retinal images and vessel measurements also represent novel and noninvasive methods to facilitate easy assessments of dialysis patients. It has been proved that the blood flow in the retina and choroid was significantly changed after a single session of hemodialysis 93,94 and the decrease in SBP was associated with the subfoveal choroidal thickness. 95 Thus, Coppolino et al. 96 performed OCT-A before the dialysis session and predict the risk of IDH within 30 days of follow-up. IDH was shown to be associated with baseline foveal vessel density of superficial and deep capillary plexus in logistic regression analysis but inversely associated with the choroid. These metrics predict the occurrence of IDH with AUCs of 0.674-0.783, with limited small and single-center data without external validation and robustness tests. But a significantly high proportion of patients reported experiencing IDH emphasizes the importance of prediction and prevention of complications with rapid and non-invasive retinal examinations.
In another cohort of hemodialysis patients, Werfel et al. 97 adopted flicker-induced dynamic retinal vessel signals to assess the microvascular function and improve the predictions of individual cardiovascular risk. For patients with ESRD and other chronic diseases, these algorithms provide a new potential for future applications of these clinically relevant outcome-specific impairments of microvascular function and will fill the gap in predictive tools for CKDs. (Table 2) 6.3 Predicting the prognosis and mortality of CKD Prediction of CKD prognosis or mortality at the onset and primary assessment could help physicians identify patients at risk of ESRD progression and deliver intensive care and timely interventions. Retinal arteriolar narrowing reflects CKD and other vascular processes and predicts renal disease progression or renal endpoints (50% renal function loss or start of RRT). The relative risk for renal endpoints of narrow arterioles was 3.7 (95% CI: 1.7-8.4) and the tertile of the narrowest arterioles was associated with more renal endpoints (log-rank p < 0.001). 79 The incidence and severity of DR 98,99 also reveal that retinal vessel changes were associated with poor renal outcomes (hazard ratio (HR) = 1.69, 95% CI: 1.16-2.45).
Zhang et al. 16 implemented a DL-based AI model to predict the risk of progression to advanced (stage 3) and severe CKD (stages 4 and 5) based on a 6-year longitudinal cohort. Combined with clinical data, the model performed well with a C-index of 0.845 (95% CI: 0.789-0.910) on the internal test set and 0.719 (95% CI: 0.627-0.807) on the external test set. Applications also extend to healthy individuals, the Kaplan-Meier method was used to stratify participants into low, medium, and high-risk groups with high degrees of separation (p < 0.001). And the incidence of CKD or advanced CKD in high-risk groups was significantly different from medium-risk groups, while no differences were observed in low-risk groups of CKD (p = 0.212 and 0.689 in the internal and external longitudinal test sets, respectively). A DL model was developed by Zhang et al. 100   increase in the risk of incident ESRD with each one-year increase in retinal age gap (HR = 1.10, 95% CI: 1.03-1.17).
As the retinal images are highly amendable for early prediction and longitudinal evaluations, they could not only help estimate the progression of ESRD but also assist doctors to be a predictor of mortality, 101 After multivariable adjustments, the retinal age gap was associated with a 2% increase in the risk of all-cause mortality (HR = 1.02, 95% CI: 1.00-1.03, p = 0.020) and a 3% increase in the risk of cause-specific mortality attributable to non-cardiovascular and non-cancer disease (HR = 1.03, 95% CI: 1.00-1.05, p = 0.041). This research suggests that retinal images have the potential to be used as a screening tool for risk assessment and the delivery of personalized treatment (See Figure 5).

FUTURE PERSPECTIVES AND CHALLENGES
As ML is exploding into many areas of the world, healthcare is no exception. The utility of big data in medicine, such as Gene Expression Omnibus Datasets, National Health and Nutrition Examination Survey, and the UK Biobank, promotes the development of DL in medicine. These large-scale databases contain in-depth genetic and health information with enormous images that we can use to screen pathogenic pathways as well as explore the relationship between different diseases. Moreover, the prediction could be conducted in the absence/early stages of the disease, as for populations at high risk of CKD like hypertension patients: high ACR within normal values could foresee CKD incidence. Not only the diagnosis of CKD could be made, but also the etiology and pathological diagnosis, early prevention of progression in high-risk groups, and guidance for cause-specific treatment could be made by AI, involving the full life-cycle of renal diseases. In addition to the only image-based models, more involvements of the electronic health record, risk factors, and clinical laboratory data have set the stage for multimodal training. In 2021, OpenAI stands out as an important representative of multimodal learning, such as CLIP 102 matching images and text, and DeepMind's Perceiver IO classifying text, images, and videos. Multimodal learning has become an important trend in medical AI. An early stage of DL was associated with unimodal AI, where the results were mapped to a single modality of data. Multimodal AI, on the other hand, is the ultimate fusion of computer vision and interactive AI intelligence models to provide computers with providing comprehensive considerations of diseases that resemble physicians in realistic diagnostic and treatment processes with broader application prospects.
Despite the high accuracy of AI models for many diseases, there are still a series of unavoidable challenges for future applications. (1) The problem of heterogeneity and varied quality comes with the larger size of datasets. Retinal images from the largest dataset, UK Biobank, were assessed by MacGillivray et al., 103 and a certain proportion (16%) of retinal images are short of quality for automated analysis. Thus, image quality management and pre-processed procedure should be added to the workflow. (2) Currently most of the algorithms are designed for the single task of distinguishing patients from healthy individuals, resulting in the lack of specificity in disease differentiation because the same pathologic signs are caused by different systemic diseases. To determine whether deep-learning models can consistently perform across the spectrum of multiple diseases, further research is required. (3) Though some of the AI screening algo-rithms were approved by FDA, the real-world utility is uncertain for now, whether it can be generalized to diverse populations in clinical practice and win the trust of patients. A new report released by Google Health shows that a retinal AI system it began testing in Thailand in 2019 has shown a strong "unconvincing" effect. Reports show that more than one-fifth of the images were rejected by the system because of clarity issues, and poor network caused delays in uploading and processing images, which need the assistance of local and timely reminder models. (4) Some questions remain in the paucity of understandability and interpretability in these AI algorithms, for which the trade-off between accurate of explainable models should be re-evaluated. (5) At this stage, given the difficulty for AI models to explain the diagnostic process, there is still an ethical risk of a "black box" and the flawed algorithms have the potential to cause significant harm to patients and lead to medical errors. Therefore, systematic training, extensive validations and testing, and prospective reviews are needed when using AI algorithms in medical practice. (6) The possibility of identifying individuals from largescale databases is increasing. Eye images including iris and fundus can be leaked as "fingerprint" biological information, which in turn may further impede privacy protection. (7) As the disproportionate distribution of ethnic groups in the datasets could increase the existing inequality in model training, the potential for bias must be respected, which calls for further collaboration and cooperation of the government and industries to ensure access to safe, equitable, and affordable medical services.

CONCLUSION
As AI is transforming almost every aspect of our lives, researchers felt an urgent need to incorporate AI in medicine, including ophthalmology and other diseases. Many studies proved the capability of retinal image-based AI models for predicting systemic diseases, such as diabetes, CVDs, and CKDs. Attempts have been made to examine the performance in screening and early diagnosis of kidney diseases and the prediction of patients' prognoses. Future study is essential for improving the interpretability of AI models and assessing the costeffectiveness and clinical deployment in real-world practice. With collaborative efforts, hope lies in that these retinal image-based AI models could be implemented and revolutionized healthcare systems in real word practice.