System and method for processing fundus images

Abstract

A system and method for determining one or more recommendations regarding an individual's health management are disclosed. An individual's relative cardiovascular aging assessment is determined at least in part on the individual's predicted risk of cardiovascular disease (CVD) determined by a deep learning model based on one or more fundus images. Recommendations regarding the individual's health management are at least in part on the results of the relative cardiovascular aging assessment.

Description

[Technical Field]

[0001] Cross-reference of related applications This application is based on Australian Patent Application No. 2023901630, filed on 24 May 2023, and Australian Patent Application No. 2023904102, filed on 18 December 2023, the entire contents of which are incorporated herein by reference.

[0002] This technology relates to a system and method for processing fundus images, more specifically to processing fundus images to determine the risk level of cardiovascular disease (CVD), and to determining relative indications of cardiovascular aging. [Background technology]

[0003] Cardiovascular disease (CVD) is the leading cause of hospitalization and premature death in the United States, and its most common complications include both unchangeable factors such as age and sex, and modifiable factors such as blood glucose control, blood pressure, cholesterol, and smoking exposure.

[0004] National CVD risk management guidelines recommend treatment decisions based on predicted CVD risk. CVD risk varies significantly across the population (from mild to severe), and identifying an individual's CVD risk using current statistical methods presents accuracy challenges. The low accuracy of current CVD risk prediction formulas (i.e., too many false positives and false negatives) is primarily due to the fact that all available predictors are indirect scales of CVD. These equations employ regression models that apply parameters such as age, sex, ethnicity, socioeconomic poverty, smoking, duration of diabetes, systolic blood pressure, total cholesterol to HDL ratio, glycated hemoglobin A1c (HbA1c), and urinary albumin to creatinine ratio (ACR). More accurate stratification of CVD risk is needed to more precisely target appropriate medications and treatment programs to the right population.

[0005] The retina is the only part of the human vascular system that can be directly observed by non-invasive means. Several recent studies have shown that CVD risk can be estimated using artificial intelligence (AI) deep learning retinal imaging algorithms. However, in all these methods, retinal images are trained against a single label. Some studies use chronological age as the training "label," and the model's results are called "retinal age." A discrepancy between the labeled (chronological) age and the estimated (retinal) age is considered to indicate a high risk of CVD events. Other studies use CVD risk calculated using conventional formulas as the "label." In this approach, the result is a single number (presumably the perceived risk), which has proven to be inaccurate. Furthermore, none of these approaches identify the main factors contributing to CVD risk (e.g., blood pressure, cholesterol, blood sugar control, and other factors).

[0006] The purpose of this disclosure is to solve at least one of the aforementioned problems, or at least to provide a useful alternative to the public.

[0007] Further aspects and benefits of this disclosure will become apparent from the following description, which is provided only as an example. [Overview of the project] [Problems that the invention aims to solve]

[0008] This technology provides a system and method for retinal image analysis using artificial intelligence (AI). Since retinal images (also known as fundus images) are routinely taken as part of medical screening procedures (such as retinal screening for diabetic retinopathy), these images can be analyzed quickly and at low cost to improve CVD risk prediction and provided immediately to patients and healthcare providers without imposing an additional burden on patients. [Means for solving the problem]

[0009] According to one aspect of this technology, a method is provided for determining an individual's relative cardiovascular aging indicator, and this method is: A step of determining the similarity between the predicted risk of cardiovascular disease (CVD) for a first individual and the risk of CVD for a set of individuals belonging to the same chronological age group as the first individual, wherein the predicted risk of CVD is determined by a deep learning model based on one or more fundus images. The steps include determining the average chronological age of the individual closest to the first individual from the perspective of predicting the CVD risk, The steps include determining the average expected chronological age of individuals with a predicted risk of CVD, The steps include determining an indication of relative cardiovascular aging for a first individual based on the determined similarity, the determined average chronological age, and the determined average predicted chronological age.

[0010] In the example, the step of determining the relative cardiovascular aging indication ("cardiac biological age") of the first individual includes the following calculation:

number

[0011] In the example, the set of individuals may include individuals of the same sex as the first individual.

[0012] In the example, Age (Y|x~Y) The implementation may include the element of similarity magnitude * cosine similarity age + (1 - similarity magnitude) * risk extrapolation age.

[0013] According to one aspect of the present technology, a method for determining at least one recommendation regarding an individual's health management is provided, the method comprising: determining an indication of an individual's relative cardiovascular aging ("cardiac biological age") based at least in part on a predicted risk of cardiovascular disease (CVD) of the individual determined by a deep learning model based on one or more fundus images; determining at least one recommendation regarding the individual's health management based at least in part on the determined cardiac biological age.

[0014] According to one aspect of the present technology, a method for determining at least one recommendation regarding an individual's health management is provided, the method comprising: determining an indication of an individual's relative cardiovascular aging ("cardiac biological age") based at least in part on a predicted risk of cardiovascular disease (CVD) of the individual determined by a deep learning model based on one or more fundus images; determining the difference ("age difference") between the individual's actual chronological age and the cardiac biological age determined for the individual; determining at least one recommendation regarding the individual's health management based at least in part on the determined age difference.

[0015] According to one aspect of the present technology, determining an indication of an individual's relative cardiovascular aging based at least in part on a predicted risk of cardiovascular disease (CVD) of the individual determined by a deep learning model based on one or more fundus images; determining the relative contribution of one or more risk contributors to the indication of relative cardiovascular aging, and a method comprising the steps is provided.

[0016] In an example, the risk factors may include two or more of blood pressure, glycated hemoglobin A1c (HbA1c), total cholesterol, and blood glucose control.

[0017] In this example, the method may include a step of determining the difference ("age difference") between an individual's actual chronological age and the relative cardiovascular aging assessment determined for that individual.

[0018] In the example, this method may include the step of comparing an individual's age difference with the age difference of a set of individuals belonging to the chronological age category to which the individual belongs. In the example, this method may include the step of determining the relative position of the individual within the set of individuals based on the age difference. In the example, this method may include the step of determining one or more recommendations for the individual's health management based at least in part on the individual's relative position within the set of individuals based on the age difference.

[0019] The example may include a step of determining the relative contribution of one or more risk contributors based on the relative position of an individual within a set of individuals based on age differences.

[0020] The example may include a step of determining one or more recommendations for an individual's health management using the relative contribution of one or more risk contributors. In the example, one or more recommendations may include initiating testing and / or investigation for a condition associated with one or more risk factors.

[0021] According to one aspect of this technology, a method is provided for predicting the risk of cardiovascular disease (CVD) from one or more fundus images, the method being performed by one or more processors. In an example, the method includes the step of processing one or more fundus images associated with an individual using a quality assurance (QA) set of one or more convolutional neural networks (CNNs) to determine whether one or more fundus images are of sufficient quality for further processing. In an example, the method further includes the step of processing one or more fundus images determined to be of sufficient quality for further processing using an eye identification set of one or more CNNs (eye ID CNNs) to identify one or more fundus images belonging to a single eye. In the example, the method further includes the step of processing one or more fundus images using one or more sets of risk contributors (RCF CNNs) of CNNs, each RCF CNN configured to output an index of the probability of presence of different risk contributors in each of the one or more fundus images, at least one RCF CNN is comprised within a jury system model comprising multiple juror CNNs, each juror CNN configured to output the probability of different features in one or more fundus images, and the outputs of the multiple juror CNNs are processed to determine the index of the probability of presence of the risk contributors output by the RCF CNNs. In the example, the method further includes the step of generating individual feature vectors based on individual metadata and the outputs of one or more sets of RCFs of CNNs. In the example, the method further includes the step of processing individual feature vectors using a CVD risk prediction neural network model to output a prediction of an individual's overall CVD risk, the CVD risk prediction neural network model configured to determine the relative contribution of each risk contributor to the prediction of the overall CVD risk. In this example, the method further includes a step of reporting the overall CVD risk, which includes a step of reporting the relative contribution of each risk contributor to the overall CVD risk.

[0022] According to one aspect of this technology, a method is provided for predicting cardiovascular disease (CVD) from one or more fundus images, which is performed by one or more processors and includes the step of processing one or more fundus images associated with an individual using one or more sets of one or more convolutional neural networks (CNNs). In an example, the one or more sets of CNNs may include two or more of the following: one or more quality assurance (QA) sets of CNNs, one or more eye identification (eye ID) sets of CNNs, one or more local modification sets of CNNs, one or more global modification sets of CNNs, and one or more meta-representation sets of CNNs.

[0023] According to one aspect of this technology, a method is provided for predicting cardiovascular disease (CVD) from one or more fundus images, the method being performed by one or more processors and comprising the step of processing one or more fundus images associated with an individual using one or more sets of risk contributors (RCFs) of one or more CNNs, each set of RCFs of one or more CNNs configured to output an index of the probability of presence of a different risk contributor for each of the one or more fundus images. In an example, the method comprises the step of generating an individual feature vector based on the individual's metadata and the outputs of one or more sets of RCFs of CNNs. In an example, the method comprises the step of processing the individual's feature vector using a CVD risk prediction neural network model and outputting a prediction of the individual's CVD risk.

[0024] In the example, one or more fundus images may be processed to predict the risk of cardiovascular disease (CVD) from one or more fundus images. In the example, this method may include the step of processing one or more fundus images associated with an individual using one or more quality assurance (QA) sets of CNNs to determine whether one or more fundus images are of sufficient quality for further processing. In the example, this method may include the step of processing one or more fundus images determined to be of sufficient quality for further processing using one or more eye identification (eye ID) sets of CNNs to identify one or more fundus images belonging to a single eye. In the example, this method may include the step of processing one or more fundus images using multiple risk contributor (RCF) sets of one or more CNNs, each RCF set of one or more CNNs configured to output an indication of the probability of presence of a different risk contributor for each of one or more fundus images. In the example, this method may include the step of generating an individual feature vector based on the individual's metadata and the outputs of multiple RCF sets of one or more CNNs. In the example, this method may include the step of processing the individual's feature vector using a CVD risk prediction neural network model to output a prediction of the individual's CVD risk.

[0025] In the example, one or more fundus images may be processed using one or more convolutional neural network quality assurance (QA) sets to determine whether the fundus images are of sufficient quality for further processing.

[0026] In the example, classifying an image as unsuitable may include determining that the image is not directed towards the relevant area of ​​the individual's eye. In the example, determining that an image is unsuitable may include determining that at least one characteristic of the image is unsuitable. For example, an image may be determined to be oversaturated, underexposed, out of focus, or blurry.

[0027] In the example, a notification may be issued warning the user that one or more of the provided fundus images are inappropriate. This allows the user to provide one or more replacement images.

[0028] In the example, one or more fundus images may be adjusted before processing. In the example, image adjustment may be image normalization, such as spatial normalization or intensity normalization. In the example, spatial normalization may include one or more of the following: cropping, scaling, and rotation of one or more fundus images.

[0029] In the example, a color balance adjustment process may be performed on one or more fundus images. In one example, a Gaussian filter may be applied to one or more fundus images to perform color balance adjustment. Color image quality can vary greatly depending on the technology and model of the fundus camera. Color balancing reduces image inconsistencies and facilitates subsequent processing. In the example, one or more fundus images may be converted from color images to grayscale or mobanochrome images.

[0030] In this example, brightness adjustment may be performed on one or more fundus images. Image brightness can vary significantly depending on environmental conditions (e.g., lighting in the clinic) and the size of the patient's pupil. Brightness adjustment normalizes these variations, making subsequent processing easier.

[0031] In examples where one or more fundus images include multiple fundus images, the multiple fundus images may be processed using a set of eye identification (eye IDs) of one or more convolutional neural networks configured to group the fundus images as belonging to a single eye, for example, for the aggregation of future clinical outcomes. In this example, the eye ID CNN works by identifying eyes as left or right, understanding the "similarity" of multiple images, and understanding one or more parameters, including but not limited to image timestamps and patient-specific IDs. A group of images is sometimes called an image set.

[0032] In the example, one or more CNNs may be configured to identify the relative position of one or more fundus images on the retina. For example, one or more CNNs may be configured to determine whether one or more fundus images are centered on the macula or on the disk. Two main landmarks on the retina are the macula, where the densest concentration of photoreceptors is located and which is responsible for central vision, and the optic disc, where the optic nerve enters the eye. In the example, an eye ID CNN may be configured to determine whether one or more fundus images are centered on the fovea. In the example, an eye ID CNN may be configured to identify the relative position of one or more fundus images on the retina.

[0033] In the example, one or more CNNs may be configured to determine the device used to capture fundus images, or the characteristics of the device. In the example, one or more CNNs may be configured to determine whether the device utilizes flash photography or white LED confocal photography. In the example, the processing of fundus images may be at least partially based on the device determination or the characteristics of the device. In the example, the adjustment of one or more fundus images before processing may be at least partially based on the device determination or the characteristics of the device.

[0034] In the example, one or more fundus images are processed by one or more sets of risk contributors (RCFs) of CNNs, and each set of RCFs of one or more CNNs is configured to output indications of the probability of presence of different risk contributors. In the example, the risk factors may include two or more of the following: blood glucose control, blood pressure, cholesterol, and exposure to smoking. In the example, each CNN may generate probabilities of indications for these risk contributors. For example, a CNN may look for “local” indications of biological or physiological changes (e.g., angiotensins, edema, etc.) and / or “global” changes in the image (e.g., pigment changes in the periapillary region, deformation of arterial / venous crossings, changes in vascular curvature, changes in vascular caliber, etc.) that may indicate the presence of blood glucose control, blood pressure, cholesterol, or exposure to smoking. Examples of indications include, but are not limited to, the appearance, clustering, and / or location of drusen, changes in the density and / or location of pigmentation, arteriovenous crossings, changes in the diameter and / or thickness of arteriovenous crossings, arteriovenous curvature, the size and / or pattern of retinal edema, and / or concentration of arteriole aneurysms.

[0035] In the example, one or more sets of risk-contributing factors (RCFs) in a CNN may be configured to target multiple labels selected from the following groups: retinopathy, maculopathy, HbA1c, systolic blood pressure, drusen, age-related macular degeneration (AMD), smoking status, total cholesterol, and macular pigment abnormalities.

[0036] In the example, at least one RCF CNN consists of a jury system model including multiple juror CNNs, each juror CNN configured to output the probability of one or more different features in fundus images, and the outputs of the multiple juror CNNs are processed to determine the probability indication of the presence of risk contributors output by the RCF CNN.

[0037] For example, investigating each risk factor (e.g., blood glucose control, blood pressure, cholesterol, exposure to smoking) may involve multiple jurors (e.g., at least five). Each juror may be configured to output a probability. The jury system model may generate a final probability based on the results of each juror. In the example, the outputs of multiple juror CNNs may be processed to determine the probability indications of the risk contributors output by the RCF CNN, based on the expected population baseline of the population to which each individual belongs.

[0038] In the example, the outputs from one or more sets of risk contributors (RCFs) of a CNN are aggregated using minimum, maximum, mean, and median values ​​at both the model and image levels to generate an individual-level fundus image feature vector. In the example, the raw output of each model can be multiple floating-point values, and the length of the output varies by model. Aggregation of the outputs is first performed at the model level. For example, for an input fundus image, five jury models give probabilities from 0 to 1, i.e., a minimum value of 0 and a maximum value of 1 (e.g., a decimal value such as 0.01454), and the probabilities for each grade level of the five models are also aggregated. In the example, the model outputs are floating-point numbers, and after aggregation using mathematical operations (including, but not limited to, weighted average, minimum, maximum, etc.), the final output remains in floating-point form. In the example, these floating-point numbers are concatenated to form a one-dimensional array (i.e., an individual-level fundus image feature vector). In the example, individual metadata associated with one or more fundus images is concatenated with individual-level fundus image feature vectors to generate an individual feature vector. In the example, a metadata vector is generated from the metadata. In the example, the metadata is preprocessed using one or more of the following: standardization and one-shot encoding. For example, numerical features such as age may be standardized to mean 0 and standardized variance 1. For example, categorical features (e.g., gender or ethnicity) may be converted from string data to numerical vectors using one-shot encoding. In the example, the individual-level fundus image feature vector and the metadata vector may be concatenated to generate an individual feature vector. This provides a meta-representation that a neural network can understand.

[0039] In the example, the CVD risk prediction neural network model utilizes a fully connected neural network (FCNN). In the example, the FCNN may have at least five layers. In the example, the relative contribution of each modifiable factor (e.g., blood glucose control, blood pressure, cholesterol, smoking exposure) to the overall CVD risk score is determined. This combination is an algorithmic rather than equational approach, where patient biometrics are appropriately combined and weighted with retinal images within deeper layers of the overall FCNN design.

[0040] For example, two or more functions of each set of one or more convolutional neural networks disclosed herein may be provided by a single set of one or more convolutional neural networks.

[0041] In the example, the system may be configured to report CVD risk at one or more of the individual and population levels. At the individual level, the system may report an individual's overall CVD risk, i.e., the overall risk of CVD for an individual related to the processed fundus image. In the example, the system may be configured to report contributors to an individual's overall CVD risk, including immutable contributors (e.g., based on patient meta-information such as age, sex, and / or ethnicity) and modifiable contributors (e.g., based on glycemic control, blood pressure, cholesterol, and exposure to smoking). In the example, the system may be configured to identify the relative contribution of each modifiable contributor. In the example, the system may be configured to rank the modifiable contributors according to their relative contribution to the individual's overall CVD risk.

[0042] At the population level, the system may be configured to present a report analysis that generates an overall cohort cardiovascular risk profile and its contributing factors. For example, the cohort could be a population at the local, regional, or national level, a population of healthcare providers, an organizational population, or a subset thereof (e.g., risk level within the overall population). Similar to individual overall CVD risks, the system may be configured to report the relative contributions of each contributing factor that are moduloable at the population level.

[0043] In the example, the system may be configured to provide recommendations for managing an individual's condition based on the determined risk. For example, a risk level scale may be provided, with associated recommendations for each risk level. In the example, at least one recommendation may be provided based on the relative contribution of each modifiable contributing factor. Such recommendations may relate to one or more decisions regarding lifestyle (e.g., diet and exercise), further clinical evaluation (e.g., consultation with a cardiologist), or medication (e.g., medication adherence).

[0044] For example, the results can be sent to an institution (such as a health insurance company that conducts population health analysis) for further analysis.

[0045] In the example, the system may be configured to compare an individual's overall CVD risk and at least one of the relative contributions of each risk contributor to the overall CVD risk with at least a portion of a population of individuals whose overall CVD risk is predicted by a CVD risk prediction neural network model, and to report an indication of the comparison.

[0046] In the example, the system may be configured to predict changes in overall CVD risk based on changes in one or more risk-contributing factors. In the example, the system may be configured to predict group-wide CVD risk for at least a portion of a population of individuals whose overall CVD risk is predicted by a CVD risk prediction neural network model. In the example, the system may be configured to predict group-wide CVD risk based on changes in one or more risk-contributing factors for at least a portion of a population of individuals.

[0047] According to one aspect of this technology, a system is provided comprising a memory for storing program instructions and at least one processor configured to execute the program instructions stored in the memory, wherein the program instructions cause the processor to perform the task of determining the relative cardiovascular aging of an individual described herein.

[0048] According to one aspect of the present technology, a computer program product is provided, comprising a non-temporary computer-readable medium storing computer-readable program code, the computer-readable program code, when executed by a processor, includes instructions causing the processor to perform a method for determining the relative cardiovascular aging indication of an individual as described herein.

[0049] According to one aspect of this technology, a system is provided comprising a memory for storing program instructions and at least one processor configured to execute the program instructions stored in the memory, wherein the program instructions cause the processor to perform the task of determining the relative cardiovascular aging of an individual described herein.

[0050] According to one aspect of the present technology, a computer program product is provided, comprising a non-temporary computer-readable medium storing computer-readable program code, the computer-readable program code, when executed by a processor, includes instructions causing the processor to perform a method for predicting cardiovascular disease (CVD) as described herein.

[0051] The above and other features will become apparent from the following description and attached drawings. [Brief explanation of the drawing]

[0052] Further aspects of this disclosure will become apparent from the following description, which is provided for illustrative purposes only with reference to the attached drawings. [Figure 1] This is a schematic diagram of a system showing various computing elements that can be used individually or together according to embodiments of this technology. [Figure 2A] This figure shows a system design illustrating the flow for processing fundus images to predict the risk of cardiovascular disease (CVD) according to an aspect of this technology. [Figure 2B] This figure shows the jury model structure of the CNN set used in the system relating to this embodiment of the technology. [Figure 3] A diagram illustrating an exemplary architecture of a convolutional neural network (CNN) used according to the embodiments of this technology is shown. [Figure 4A] This figure shows another flow for processing fundus images to predict cardiovascular disease (CVD) risk and cardiac biological age according to an aspect of this technology. [Figure 4B] This figure shows a system design for processing fundus images to predict the risk of cardiovascular disease (CVD) and cardiac biological age, according to an aspect of this technology. [Figure 5A] The receiver operating characteristic (ROC) curve of the model used to predict the risk of CVD according to the aspects of this technology is shown. [Figure 5B]The precision-recall curve of the model used to predict the risk of CVD according to the aspects of this technology is shown. [Figure 6] The box plots shown illustrate the spread of cardiac biological age values ​​obtained for each chronological age group, as predicted according to the aspects of this technology. [Figure 7A] The distribution of cardiac biological age and chronological age in the first dataset according to an aspect of this technology is shown. [Figure 7B] The distribution of cardiac biological age and chronological age in the second dataset according to this technology is shown. [Figure 8] This is a box plot showing the relationship between the severity of diabetic retinopathy and age differences according to an embodiment of this technology. [Figure 9A] This shows the chronological age of men ranked by mean logLTL decile in studies conducted according to aspects of this technology. [Figure 9B] This shows the cardiac biological age issued by the DL model for men, ranked by mean logLTL decile. [Figure 10A] This shows the change in chronological age of women, ranked by mean logLTL decile, in a study conducted according to an aspect of this technology. [Figure 10B] This shows the cardiac biological age issued by the DL model for women, ranked by mean logLTL decile. [Figure 11A] This shows the systolic blood pressure, HbA1c, and 10-year PCE CVD risk score of male individuals who were issued cardiac biological age scores in the upper quartile (Q1) and lower quartile (Q4) of their chronological peer group, classified by mean logLTL decile. [Figure 11B] This shows the systolic blood pressure, HbA1c, and 10-year PCE CVD risk score of male individuals who were issued cardiac biological age scores in the upper quartile (Q1) and lower quartile (Q4) of their chronological peer group, classified by mean logLTL decile. [Figure 11C]This shows the systolic blood pressure, HbA1c, and 10-year PCE CVD risk score of male individuals who were issued cardiac biological age scores in the upper quartile (Q1) and lower quartile (Q4) of their chronological peer group, classified by mean logLTL decile. [Figure 12A] This shows the systolic blood pressure, HbA1c, and 10-year PCE CVD risk score of male individuals who were issued cardiac biological age scores in the upper quartile (Q1) and lower quartile (Q4) of their chronological peer group, classified by mean logLTL decile. [Figure 12B] This shows the systolic blood pressure, HbA1c, and 10-year PCE CVD risk score of male individuals who were issued cardiac biological age scores in the upper quartile (Q1) and lower quartile (Q4) of their chronological peer group, classified by mean logLTL decile. [Figure 12C] This shows the systolic blood pressure, HbA1c, and 10-year PCE CVD risk score of male individuals who were issued cardiac biological age scores in the upper quartile (Q1) and lower quartile (Q4) of their chronological peer group, classified by mean logLTL decile. [Figure 13] An example report of cardiac biological age generated according to an aspect of this technology is shown. [Modes for carrying out the invention]

[0053] This technology is generally aimed at determining a person's biological age (hereinafter referred to as BioAge) by using artificial intelligence technology with retinal photographs (also called fundus images) and demographic information (details such as age, sex, and ethnicity). These images and information can be obtained, for example, by an optometrist as part of a routine eye examination.

[0054] Biological age is considered a reliable alternative indicator to the Framingham risk score, a tool recommended by the Heart Foundation to assess an individual's 10-year risk of developing cardiovascular disease (CVD). The biological age metric is not a substitute for the Framingham risk score. However, one key finding by its inventors was that individuals whose biological age was, on average, five years older than their chronological age had twice as high a Framingham risk score as their peers. Because not everyone ages at the same rate, there is growing interest in distinguishing between an individual's chronological age and their biological age.

[0055] While chronological age is defined as the number of years a person has lived, biological age refers to the degree to which the cells in a person's body have aged and are functioning. Therefore, biological age is a better indicator of lifespan and future functional capacity. Although various theories exist to explain this process, it is recognized that the rate at which cells deteriorate depends heavily on factors such as the genes we inherit, lifestyle choices, the level of stress, and the body's response to infection. Research in this field has already shown that estimates of biological age can predict mortality and the onset of a wide range of physical and mental illnesses, from cardiovascular disease (CVD) to clinically significant depression, more accurately than chronological age.

[0056] In contrast to known tools used to assess biological age, this technology does not require data collection through clinical or laboratory assessments (such as blood tests and / or radiographs), which are relatively difficult to access. The retina has long been recognized as a unique window to an individual's health. This is because biomarkers present in retinal images can provide valuable insights into aging, inflammatory health, nervous system health, and cardiovascular health. The retinal vascular system is the only body part from which scientists and clinicians can directly assess the health of an individual's neurovascular tissue, allowing for non-invasive, direct observation of the body's microvessels. While some of these parameters can be assessed by humans, many others are invisible to the naked eye but can be detected by current technology.

[0057] Knowing one's biological age can lead to a better, healthier life, which is believed to benefit the individual, their family, and their community. For those who want to lower their biological age, there are actions they can take to contribute to a healthier lifestyle. There is strong evidence that simple health-improving interventions, such as losing weight, increasing exercise, quitting smoking, and controlling blood pressure, can have a significant impact on biological age.

[0058] Figure 1 shows a schematic diagram of a system 1000 that illustrates various computing elements that can be used individually or together according to aspects of the present technology. System 1000 includes a processing system 1002. In an example, the processing system 1002 may comprise processing functions represented by one or more processors 1004, a memory 1006, and other elements that are typically present in such a computing environment. In the illustrated exemplary embodiment, the memory 1006 stores information accessible by the processor 1004, which includes instructions 1008 that can be executed by the processor 1004 and data 1010 that can be retrieved, manipulated, or stored by the processor 1004. The memory 1006 may include any suitable means known in the art, including a computer-readable medium or other medium for storing data that can be read with the help of an electronic device, capable of storing information in a manner accessible by a processor. The processor 1004 may be any suitable device known to those skilled in the art, and although the processor 1004 and memory 1006 are illustrated as being in a single unit, this is not limiting, and it should be understood that each function described herein may be performed by multiple processors and memories, which may be separated from each other or located together.

[0059] Instruction 1008 may include any set of instructions suitable for execution by processor 1004. For example, instruction 1008 may be stored as computer code on a computer-readable medium. Instructions can be stored in any suitable computer language or format. Data 1010 may be retrieved, stored, or modified by processor 1004 according to instruction 1008. Data 1010 may also be formatted in any suitable computer-readable format. Again, although it is shown that the data is stored in a single location, this is not limited to this, and it should be understood that the data may be stored in multiple memories or locations. Data 1010 may include database 1012.

[0060] In some embodiments, one or more user devices 1020 (e.g., mobile communication-enabled devices such as a smartphone 1020-1, a tablet computer 1020-2, or a personal computer 1020-3) can communicate with the processing system 1000 via the network 1022 and access the functions and data of the processing system 1002. The network 1022 can consist of various configurations and protocols, such as the internet, an intranet, a virtual private network, a wide area network, a local network, a private network using one or more proprietary corporate communication protocols (wired or wireless), or a combination thereof. For example, fundus images acquired from one or more fundus imaging devices (hereinafter referred to as "fundus cameras" 1030) can be input to the processing system 1002 via the user device 1020.

[0061] A fundus camera typically includes an image capture device that is held close to the outside of the eye during use, illuminating and photographing the retina to provide a 2D image of a portion of the eye's interior. Many clinically important areas of the eye can be imaged, including the retina, macula, fovea, and optic nerve head. A single undilated fundus image only captures less than 45 degrees of the back of the eye. In practice, clinicians often choose to take multiple images, instructing patients to look up, down, left, and right to widen the field of view of the retina.

[0062] 1. First Exemplary Model Figure 2 shows a method / process architecture 2000 for processing fundus images according to an aspect of the present technology. For completeness, the deep learning models and frameworks disclosed herein are provided as examples, and it is understood that viable alternatives will be obvious to the experienced reader.

[0063] Method 2000 utilizes various convolutional neural networks ("CNNs"). CNNs are deep learning architectures particularly well-suited for analyzing visual images. A typical CNN architecture for image processing consists of a series of convolutional layers interspersed with pooling layers. The convolutional layers apply filters learned from training data to small regions of the input image to detect more relevant image features. The pooling layers downsample the output of the convolutional layers to reduce dimensionality. The output of a CNN can take various forms depending on the application, such as one or more probabilities or class labels.

[0064] The first dataset used as training data included measurements for both non-diabetic and diabetic patients. Since not all measurements are related to CVD risk, irrelevant columns were discarded following expert advice. This resulted in 35 columns corresponding to 21 fields, including age, sex, ethnicity, deprivation score, family history, smoking, systolic blood pressure, BMI, TC / HDL, HbA1c, diabetes status (Y / N), type of diabetes, atrial fibrillation, antihypertensive medication, antithrombotic medication, lipid-lowering medication, eGFR, metrasone in the past 6 months, lipids in the past 6 months, LLD in the past 6 months, anticoagulants in the past 6 months, anticoagulants in the past 6 months, and CVD events and dates. These columns were retained based on expert advice to avoid missing useful variables, but it is not necessary to use all of them in the modeling. Regarding total visits, each patient typically has multiple visits over time (i.e., there may be multiple sets of biometric information for a single patient). Based on expert advice, only the first consultation for each patient was maintained to maximize the study observation time. The first dataset, obtained according to the screening process described below, contained 95,992 images from 51,956 patients. The second dataset, created using the screening process described below, contained 14,280 images from 3,162 patients. This second dataset was used to tune and validate the model developed using the above training data.

[0065] One example is the modified Inception-ResNet-v2 CNN architecture shown in Figure 3. Inception-ResNet-v2 is a convolutional neural architecture based on the Inception family architecture, incorporating residual connections. It consists of 164 layers and dozens of initial residual blocks. Each initial residual block consists of multiple parallel branches to which convolutional kernels of different sizes and strides are applied. For example, one branch moves to a 1*1 convolution operation, while others move to 1*7, 7*1, 1*3, 3*1, or 3*3. The different-sized convolutional kernels are intended to capture image features from different perspectives. Residual connections are designed to build deeper networks. The idea behind residual connections is relatively simple: the input of each block is appended to the output of the block to preserve the input information. This allows the model to ignore some blocks as needed, and facilitates the propagation of gradients along the network. In this example of the technology, Inception-ResNet-v2 is used as the feature extractor, the final layer is adapted to meet the requirements of the technology, and the probability of the learned features being present is generated.

[0066] Returning to Figure 2, in the input stage 2002, one or more fundus images (e.g., a collection of individual fundus photographs) are received. Quality assurance is performed on the received images to determine if they are suitable for further processing. In this example, quality assurance is performed by a set of one or more Quality Assurance ("QA") CNNs 2004.

[0067] 1.1 QAC NN QA CNN 2004 is trained by inputting sample images pre-labeled by expert clinicians and training for a sufficient number of iterations. In one example, QA CNN is based on a modified XCEPTION design (although a modified Inception-ResNet-v2 design as described above may also be used) and trained using a dataset of 20,000 images, the dataset containing four types of images in roughly equal proportions: Type 1: Eyeballs, rooms, and other irrelevant images; Type 2: Images that are extremely oversaturated or underexposed; Type 3: Images that are not perfect but may be useful for clinicians to perform manual analysis; Type 4: High-quality images.

[0068] The experiment was run on an Intel Xeon Gold 6128 10 Professional with 16GB of RAM and an NVIDIA GeForce TiTan V VOLTA 12GB graphics card. Tensorflow 1.11.0 and Python 3.6.6 were used to run the QA CNN 3004 model.

[0069] The hyperparameters include the following: (i) Batch size: 64. The batch size refers to the number of training samples used in one step. The larger the batch size, the larger the required memory area. When the input image size is 320×320 and the GPU memory is 12GB, the batch size is set to 64. (ii) Training / validation / test split: (70 / 15 / 15). (iii) Epochs: 100. One epoch refers to one forward pass and one backward pass of all training examples. (iv) Learning algorithm: The ADAM optimizer, an advanced version of stochastic gradient descent, is used. (v) Initial learning rate: 10e-3. The learning rate controls how much the model adjusts the weights with respect to the loss gradient. Typical learning rates are on the order of [10e-1, 10e-5]. Considering the use of the ADAM optimizer and batch normalization, the initial learning rate is set to 10e-3. (vi) Loss function: Softmax cross-entropy. (vii) Dropout rate: 0.5.

[0070] The above QA CNN achieved 99% accuracy when classifying input images into categories. After training, all type 1 and type 2 images were deleted. Type 3 images are displayed to the clinician but not used for further processing. Type 4 images are used as part of subsequent processing.

[0071] 1.2 Lighting Type CNN In an example, one or more lighting type CNNs 2005 can be configured to determine the device or the characteristics of the device used to capture an input fundus image. There are mainly two imaging techniques for fundus photography. a) Flash photography and b) white LED confocal photography, which generate visually different images respectively. Depending on the camera source (and the image), subsequent processing (described below) is adjusted.

[0072] 1.3 Eye ID CNN Clinicians often take multiple images from one eye to create a wider field of view at the back of the eye. A 2006 eye recognition (eye ID) CNN set was trained to find similarities between multiple viewpoint images from the same eye and group them into a single image set. Identifying images belonging to the same eye is crucial because the final clinical outcome may be the sum of the analyses of each image in that set.

[0073] The exemplary training environment for Eye-ID CNN 2006 is similar to that of QA CNN 2004 described above. A database of 160,585 images was created from 75,469 eyes of 40,160 individuals. Each image was labeled with left / right eye, patient ID (if available), and image acquisition timestamp. Eye-ID CNN 2006 was trained on this dataset to identify image orientation (left / right) and group images based on ID / acquisition time. Eye-ID CNNs trained in 2006 achieved accuracy exceeding 99%. When implemented, Eye-ID CNNs group multiple images submitted by clinicians into eye and patient subgroups.

[0074] The ocular ID CNN from 2006 was further trained to identify the location of images on the retina, and was also able to distinguish whether that location was centered on the macula or on the disk.

[0075] 1.4 Image Preparation Fundus images, after being processed by the eye ID CNN 2006, may be adjusted before further processing in the image preparation stage 2008, for example, by adjusting brightness and color balance for normalization, or by cropping and scaling the images for standardization.

[0076] For example, a Gaussian filter may be applied to the original fundus photograph. An example of such a filter can be expressed as follows:

number

[0077] 1.5 Feature Extraction Next, high-quality images with relevant labels are passed through multiple AI models. These AI models are trained to detect risk-contributing factors (RCFs) CNN 2010 to detect instructions for blood glucose control, blood pressure, cholesterol, and exposure to smoking. a from 2010 n The set includes, but is not limited to, the appearance, clustering, and / or location of drusen, changes in the density and / or location of pigmentation, arteriovenous crossings, changes in the diameter and / or thickness of arteriovenous crossings, arteriovenous curvature, the size and / or pattern of retinal edema, and / or concentration of arterioles.

[0078] RCF CNN 2010: Searching for indications of blood glucose control, blood pressure, cholesterol, and smoking exposure in the retina. a from 2010 n Each functions as a "jury" system. Referring to Figure 2B, each RCF CNN 2010 is a multiple jury CNN 2011 (in this example, five jury CNN 2011) a ~2011 e Each CNN 2011 is composed of ) and is configured to generate probabilities of features that have been trained to focus on (e.g., the presence and concentration of drusen).

[0079] CNN 2010 for extracting features from fundus images nIn an exemplary implementation, 101 layers are stacked using residual connections and an inception block, resulting in 24,276,481 parameters being generated. A dataset of 95,992 images is created from 51,956 patients. The data points recorded for each patient are as follows: gender, date of birth, date of death (if applicable), ethnicity, socioeconomic poverty index, HbA1c, SCR, TCHDL, ACR, blood pressure lowering medication (Y / N), lipid lowering medication (Y / N), anti-thrombotic medication (Y / N), oral antidepressant (Y / N), insulin (Y / N), AF, CVD event, CVD event date, EGFR. This dataset is obtained from multiple ophthalmology clinics over 15 years using several different fundus camera models.

[0080] For each CNN 2010 n the dataset is split into training, validation, and test sets (70%, 15%, 15% respectively) or a similar ratio. The fundus images are first cropped and resized to an 800x800 (or similar) pixel size. In order to maximize GPU memory utilization during training, the batch size is set to 6. To update the parameters to minimize the loss, an Adam optimizer is employed with a learning rate of 1*10e-3. Dropout is enabled at a rate of p = 0.2 and the model is trained for at least 100 epochs. This exemplary implementation is executed by the Python programming language version 3.7. To guide the optimization of the model parameters, a cross-entropy loss function is employed. The goal of training is to minimize the loss function to obtain the most accurate probability prediction of CVD events. Typically, cross-entropy loss is utilized in the context of classification problems. However, although CVD event risk prediction is not a classification task, the applied labels are 1 and 0 (indicating whether a CVD event has occurred). Therefore, a cross-entropy loss approach is adopted and the overall loss is formalized as follows.

Number

[0081] In the post-processing 2012 of the jury system, the results of each juror CNN 2011 are weighted equally or unequally compared to the remaining jurors and considered, and a statistical representation of the likelihood of changes observed in a particular fundus image is created.

[0082] 2010 n After receiving the raw output from the CNN of 2010, the output results are aggregated in 2014. For example, in the case of an input fundus image, the five juror models give probabilities from 0 to 1, that is, the minimum value is 0 and the maximum value is 1 (for example, a decimal value such as 0.01454), and the probabilities of each grade level of the five models are also aggregated. In the example, the output of the model is a floating-point number, and after being aggregated using mathematical operations (including but not limited to weighted average, minimum value, maximum value, etc.), the final output remains in the form of a floating-point number concatenated to form a one-dimensional array in 2016 (that is, the individual-level fundus image feature vector).

[0083] Step 2018 involves receiving, for example, an individual's metadata associated with one or more fundus images. This metadata may include gender, ethnicity, HbA1c, TCHDL, etc. The metadata may also include categorical data such as gender, ethnicity, deprivation value, and medication, as well as numerical data such as age and HbA1c. The metadata is preprocessed using standardization and one-shot encoding. After loading this metadata into memory, the categorical data is converted to one-hot encoding. For example, 3 bits might be used to represent gender: [1,0,0] represents male, [0,1,0] represents female, and [0,0,1] represents other. In the case of numerical biometrics, standardization is applied to ensure the values ​​are on the same scale by subtracting the mean from each value and dividing by the standard deviation; for example, the normal HbA1c range is 30-120, while a TCHDL value is typically less than 8. In yet another example, numerical features such as age may be standardized to have a mean of 0 and a standard variance of 1. This generates an individual metadata vector.

[0084] 1.6 CVD Risk Prediction After the processing pipeline for the fundus image 2016 and metadata 2020 is complete, in step 2022, the individual-level fundus image feature vector and metadata vector are concatenated to form the individual feature vector. For example, the metadata vector may be in the form [0,1,0,0,1], and the individual-level fundus image feature vector may be in the form [0.3,0.5,0.4,0.35,0.43,…]. The concatenated vector is [0,1,0,0,1,0.3,0.5,0.4,0.35,0.43,…]. This concatenated vector provides a meta-representation that the neural network can understand.

[0085] Each feature vector is processed by the CVD risk prediction neural network model 2024, which utilizes a fully connected neural network (FCNN). In this example, the FCNN may have at least five layers. The sizes of each layer (i.e., the number of neurons) are 512, 256, 128, 64, and 1, respectively. In this exemplary embodiment, the ReLU activation function is used in each layer except the last layer. The last layer uses a sigmoid function to compress the output between [0,1], which serves as the predicted risk / probability.

[0086] The model is trained using an Adam optimizer with a backpropagation algorithm and a cross-entropy loss function to describe the target prediction. The training data includes labels indicating whether each individual experienced a CVD event (such as heart failure) after fundus images were taken and metadata was recorded. Thus, the cross-entropy loss can be used to measure the risk predicted by the AI ​​model as true. If an individual has experienced a CVD event, the label is 1. The model prediction of 0.6 represents an error of 1-0.6=0.4. If an individual has never experienced a CVD event, the label is 0, and the model prediction of 0.2 indicates an error of |0-0.2|=0.2. Next, the mean error for each batch is calculated. After obtaining the loss term, the backpropagation method is used to calculate the gradient of each trainable parameter (218,113 parameters in the exemplary model) with respect to the final loss. Then, the Adam algorithm is used to update the parameters in the direction of the negative gradient.

number

[0087] Therefore, the model learns the optimal parameter set from the data, minimizing overall loss and obtaining better predictive results. The relative contribution of each factor to the final risk can be estimated using other methods, such as controlled experiments. Another method is to "turn on / off" the CNNs of jurors responsible for one aspect (e.g., blood glucose control) and calculate the deviation from the calculated total risk.

[0088] During training, the process converts patient information files into a matrix. The matrix is ​​then converted into a streaming dataset, enabling memory- and time-efficient image loading. The streaming dataset is then augmented for training purposes. The augmented streaming dataset is optimized using gradient descent via a provided optimizer and loss function.

[0089] After the preprocessing step, all available biometric identifiers for each patient are converted into 1D vectors of length 31. Thus, stacking the feature vectors of all patients in the dataset yields a feature matrix of shape (44123,31).

[0090] In addition to metadata, fundus images also need to be preprocessed. After loading the fundus images, they are resized to a predefined fixed shape, such as 600*600. Then, various image enhancement techniques are used, such as random adjustments to brightness and saturation, random inversion, rotation, and simulation of JPEG noise.

[0091] Due to the large number of images (a total of 95,992 fundus images in the example), there are technical limitations that prevent loading all images into memory at once for training. Therefore, patient metadata, along with the corresponding fundus images, is converted to a streaming format. In one example, a data generator is created using TensorFlow, which generates mini-batches of data each time. Each mini-batch contains multiple patient consultations, including biometrics and fundus images. The streamed dataset is then sent to the model for training.

[0092] The CVD risk prediction neural network model 2024 operates on the fundamental understanding that patterns in fundus image features (e.g., arterial wall thickening) can occur in individuals or in combination due to multiple reasons (e.g., risk factors such as high blood pressure and high cholesterol). In other words, one change in the retina cannot be attributed to a single disease. Therefore, probabilities are used. A “jury-based” local-to-whole statistical model representing changes in retinal images is used to estimate the risk of cardiovascular events within a given period (e.g., 5-10 years) and to identify the relative contributions of various elements to the estimated cardiovascular risk. Here, the jury-based probabilities from RCF CNN 2010 are grouped and evaluated against the probabilities of changes due to immutable factors (e.g., age, race, sex) and modifiable factors (e.g., blood glucose control, blood pressure, cholesterol, smoking exposure). In doing so, the CVD risk prediction neural network model 2024 learns whether a change in one modifiable factor correlates with changes in other modifiable factors. For example, in a patient whose risk of cardiovascular events over the next five years was calculated to be 20%, when the jury's estimated probabilities of (i) local changes in arterial wall thickening and (ii) changes in the overall color pattern indicating thinning of the retinal layer are combined with the probabilities of changes due to age, sex, and race, it is estimated that 8% of the risk is due to unchangeable factors (age, race, sex, etc.), and of the remaining 12%, it is estimated that 6% is due to hypertension, 3% to diabetes, and the remainder to renal function and smoking.

[0093] 1.7 Presentation of CVD risk results The final output is a prediction of CVD risk that can be decomposed into contributors, including immutable contributors (e.g., based on patient meta-information such as age, sex, and ethnicity) and modifiable contributors (e.g., based on glycemic control, blood pressure, cholesterol, and exposure to smoking). For example, this can be achieved through individual or group analysis of the relative contribution of CNNs to the effects of each factor (e.g., smoking), such as inclusion / exclusion analysis, weight adjustment analysis, and sensitivity analysis.

[0094] 2. Second exemplary model Figure 4A shows a method / process architecture 4000 for processing fundus images according to an aspect of the present technology. For completeness, the deep learning models and frameworks disclosed herein are provided as examples, and it is understood that viable alternatives are obvious to the experienced reader.

[0095] Figure 4B shows an exemplary model consisting of four different levels.

[0096] 2.1 Level 1 The first level 4100 includes an image quality control CNN (QC) 4102, a left / right asymmetry (left eye / right eye) detector CNN 4104, and an image position (foveal / non-foveal) detector CNN 4106. The input to this layer is only fundus images. This process ensures that only sufficiently high-quality foveal images are accepted into the model. Identifying left / right asymmetry in images helps in aggregating all images from each eye of each individual during analysis.

[0097] 2.2 Level 2 Level 2, 4200, includes nine AI ensembles, each consisting of five CNNs (45 in total) trained on various labels from the UK BioBank. Each ensemble targets unique labels within fundus images: 1. Retinopathy 4202, 2. Maculopathy 4204, 3. HbA1c 4206, 4. Systolic blood pressure 4208, 5. Drusen 4210, 6. Age-related macular degeneration (AMD) 4212, 7. Smoking status 4214, 8. Total cholesterol 4216, and 9. Macular pigment abnormalities 4218.

[0098] These CNNs follow a modified version of the Inception-Resnet-V2 or ResNet50 structure. Taking the single retinopathy CNN model as an example, this model has a deep structure consisting of 164 layers, using a combination of inception and residual blocks. The inception block uses a combination of convolutional layers with different filter sizes, while the residual block uses skip connections to allow the model to learn from previous layers. Batch normalization and bottleneck layers are employed to improve training efficiency. Overall, the model architecture is designed to extract features at multiple scales and capture fine details within images, making it suitable for detecting levels of retinopathy and other biomarkers. For each CNN, the dataset is split into 70%, 15%, and 15% for training, validation, and test, respectively. Extraneous background is trimmed from fundus images, and the resulting images are resized to 800x800 pixels. A batch size of 8 is chosen to optimize GPU memory during training. The Adam optimizer is employed with a learning rate of 1*10e-3 to update parameters to minimize loss. Dropout is enabled at rate p=0.2, and the model is trained for at least 100 epochs. All code related to this work is written using Python 3.7.

[0099] Furthermore, a jury system (described above) is implemented to arrive at the final prediction for each biomarker. To illustrate with the retinopathy model, there are six different levels of retinopathy (R0 to R5). Five jury models are employed to evaluate each eye, resulting in 30 probability values ​​for each eye. These probabilities are then combined for both eyes to generate the final value for each patient.

[0100] 2.3 Third Level The third level, 4300, is a multilayer perceptron (MLP) that uses the output of the second level CNN along with the patient's chronological age, sex, and ethnicity to reproduce the pooled cohort equation (PCE) CVD risk score. This PCE CVD risk score is a ground truth label calculated from nine fields in the UK BioBank dataset for each participant. The model architecture consists of an input layer followed by five dense layers with progressively decreasing neuronal counts of 1024, 512, 256, 128, and 32. These layers are interspersed with batch normalization and a LeakyReLU activation function with a leaky rate of 0.1. To address concerns about overfitting, a dropout layer with a rate of 0.3 is incorporated after the third, fourth, and fifth dense layers. The final layer consists of a single neuron and a linear activation function that predicts the target value. For optimization, an Adam optimizer is used with an exponentially decreasing learning rate schedule, initialized to 3e-3 and decreasing by a coefficient of 0.95 every 1000 steps. The Huber loss function is employed to guide the updating of model parameters. Early stopping is implemented to suppress overfitting and ensure efficient training.

[0101] Here, we use Shapley's Additive Explanation (SHAP) algorithm to reveal the contribution of each variable to the final prediction of CVD risk. This methodology provides a unified approach to interpretability in machine learning and facilitates a comprehensive and integrated understanding of the importance of features in predictive models. Inspired by cooperative game theory, SHAP values ​​reveal a fair distribution of contributions across features in each prediction, thereby facilitating the attribution of each feature's impact on the predictive outcome. In contrast to other locally interpretable, model-independent explanations, SHAP values ​​attribute contributions in a consistent manner according to the principles of local precision, missing values, and consistency. This ensures that cumulative attributions align with the overall effect.

[0102] 2.4 Level 4 After the PCE CVD risk score is generated at the third level, at the fourth level (4400), a parameter (cardiac biological age) is derived that represents the deviation of the CVD risk score from the standard as a concept of acceleration or deceleration of cardiovascular aging. This cardiac biological age score is calculated for each individual using the following method. ● Determine the similarity between an individual's predicted CVD risk and CVD risk scores obtained from other people of the same age. ●The average (chronological) age of the individual's nearest neighbor is determined based on the state of the retina, which is represented by multidimensional features extracted from the DL model. ● Determine the average expected chronological age of individuals with a predicted CVD risk. ●These data are combined probabilistically to calculate the final value of cardiac biological age.

[0103] The calculation of cardiac biological age can be summarized by the following formula:

number

[0104] This equation has three main components. 1. P(risk(x)|x∈X) is the conditional probability that patient x has a risk(x) of being at risk of CVD, assuming that the patients belong to a set of individuals X of similar age and sex. a. This is estimated using a normal distribution in the following steps: 1. Calculate the sample mean and standard deviation of the predicted CVD risk for set X, and construct a standard normal distribution. 2. Calculation of the Z-value for patient x:

number

[0105] 2.5 Model Evaluation To evaluate the accuracy of the AI ​​ensemble predicting an individual's CVD score, the scores issued by the AI ​​model are compared to the scores generated by the PCE equation. If the score exceeds 7.5%, the individual is determined to be at high risk, and the accuracy of the AI ​​model's CVD risk prediction ability is calculated using the confusion matrix shown in Table 1 below. The output of the PCE equation is assumed to be true. [Table 1] Establishment of labels for estimating the accuracy of CLAiR when predicting individuals with high PCE CVD scores. [Table 1]

[0106] This confusion matrix is ​​used to calculate the overall model's accuracy, sensitivity, and specificity.

[0107] Table 2 shows the model's ability to identify high-risk individuals (PCE-generated CVD score > 7.5%) compared to scores generated by the PCE equation. By identifying high-risk individuals, our model achieves ROCAUC of 88.6%, sensitivity of 83.7%, and specificity of 90.2% on the test set. ROC and its associated accuracy-recall F1 curve plots are shown in Figures 5A and 5B. [Table 2] Confusion matrix comparing PCE scores generated by CLAiR and CVD scores calculated using formula PCE. [Table 2]

[0108] To test the validity of the cardiac biological age generated by the Level 4 output, the UK BioBank dataset is divided into several chronological age groups (40-43, 43-46, 46-49, 49-52, 52-55, 55-59, 59-62, 62-65, 65-68, 68-72, 72-75). The oldest subgroup is deliberately selected with a wide range (10-year intervals) to ensure sufficient participants in that age group. The AI ​​model generates a range of cardiac biological age values ​​for each chronological age group. A key belief in the concept of cardiac biological age is that the biometrics of an individual segment of a particular chronological age are typically distributed around the chronological age. Therefore, individuals with a cardiac biological age higher than their chronological age represent a cohort of people whose cardiovascular system is aging more rapidly than expected. Conversely, individuals with a cardiac biological age younger than expected represent a group of people whose aging is progressing more slowly than expected. Within each age group, individuals whose cardiac biological age is 5 years or more higher than their chronological age (i.e., cardiac biological age - chronological age > 5) were arbitrarily defined as individuals exhibiting "accelerated aging" within their chronological age category.

[0109] To test the validity of the heart biological age generated by the model, compare the biomarkers (CVD risk, HbA1c, systolic blood pressure) known to be associated with heart aging between the individual cohort of "accelerated aging" and the calendar peers. The hypothesis is that the group in the process of accelerated aging should have significantly worse values of these biomarkers compared to their peers.

[0110] Figure 6 shows the distribution of the heart biological age values obtained for each calendar age group.

[0111] To examine the effectiveness of identifying individuals showing "accelerated heart aging" of the heart biological age, calculate the mean values of PCE 10-year CVD risk, HbA1c, systolic blood pressure, and total cholesterol for each age group. Next, the mean values of the same dataset are calculated for individuals within each age group for a) individuals with "accelerated aging" (i.e., heart biological age - age > 5), and b) all others (i.e., heart biological age ≤ 5).

[0112] After establishing the mean of the data points for the accelerated aging group and the remaining group, calculate the difference from the overall mean of the group for each age category (e.g., mean 血圧40-43YO - mean 血圧40-43YO 加速老化 vs mean 血圧40-43YO - mean 血圧40-43YO そのほか ). Next, use the paired t-test to compare these deviations from the mean values of the combined ages of the "accelerated aging" patients and the other patients. As shown in Table 3, for all age groups and both genders, the mean scores of PCE 10-year CVD risk, HbA1c, and systolic blood pressure obtained from the accelerated elderly were statistically significantly higher than the mean scores obtained from the other elderly. The exception was total cholesterol, where for all age categories and both genders, the scores obtained from the accelerated aging cohort were statistically significantly lower than those from the other cohort. [Table 3] Differences in PCE 10-year CVD risk, HbA1c, systolic blood pressure, and total cholesterol between accelerated aging patients (heart biological age - age 5 years or more) and other people

Table 3

[0113] The results demonstrate that this technique, using only retinal images, can reliably detect individuals with moderate and high PCE CVD risk, along with their age, ethnicity, and sex, with an ROC AUC of 88.6%, sensitivity of 83.7%, and specificity of 90.2%. To date, developers of DL algorithms designed to predict CVD risk have reported model performance as an AUC statistic, with successful models being accepted as those achieving AUC > 0.70. While this approach has merits, the value is limited simply knowing that a model can predict CVD risk with an AUC > 0.70, as biometric data tends to be normally distributed. When biometric data is clustered around the mean, a model can achieve high accuracy by simply learning the mean data points. Therefore, high accuracy does not necessarily mean that the algorithm has learned what was expected, and for an algorithm to function reliably, its output needs to be biologically valid and clinically meaningful. Current techniques utilize multilayer perceptrons (MLPs) to derive the final predicted PCE CVD risk score. Therefore, by deriving the Shaply value, the magnitude of the attribution coefficient assigned to each component variable that constitutes an individual's PCE score can be determined. The advantage of this approach is that it allows us to display the relative contribution of the constituent risk factors on the PCE risk score generated for each individual. This not only allows physicians to gain a deeper understanding of the relative importance of these constituent risk factors for individual patients, but also provides a mechanism to open the "black box" and gain a deeper understanding of how the model arrived at its results and to test its validity.

[0114] While baseline CVD risk increases with age, the actual likelihood of a CVD event varies even among individuals of the same chronological age. Individuals age at different rates, which affects both their risk of developing chronic diseases and the severity of their effects. This difference in individual aging rates is not captured by conventional regression-based risk equations, which assume that all individuals in a population age at the same rate. As a result, these equations are relatively insensitive to variations in individual risk due to biological aging.

[0115] A key feature of this technology makes it possible to identify individuals whose CVD risk profile is accelerating compared to their peers. After generating PCE cardiac risk scores for all individuals, the mean expected chronological age of individuals with a specific individual's predicted CVD risk is determined using a clustering method (this indication is referred to here as the individual's cardiac biological age). Once the individual's cardiac biological age is determined, this value can be compared to their actual chronological age. Thus, an "age difference" arises from the difference between chronological age and cardiac biological age. Compared to their peers, individuals with an age difference of more than 5 years, both male and female, have significantly higher CVD risk scores, systolic blood pressure, and HbA1c. Total cholesterol is significantly lower. These data suggest that, due to aspects of current technology, it is possible to extract data from the retina that can not only accurately predict CVD within a population but also identify individuals whose cardiac risk profile is significantly "older" compared to their peers.

[0116] 3. Cardiac biological age and age difference Further model training and evaluation will be carried out as described herein. The UK Biobank will be used for training and internal validation. The validation subset will represent 20% of the data randomly selected before development. UK Biobank data is accessible through a direct request to the UK Biobank (IRBUOA-86299) and will be obtained using approved data management and data transfer protocols. This study will use 89,894 fundus images collected from 44,176 participants in the UK Biobank. While UK Biobank participants were recruited from the general UK population, only about 5% of UK Biobank participants identified themselves as having "physician-diagnosed" diabetes.

[0117] A 10K dataset from the US-based iPAX study, which includes several differences from the UK Biobank, will be used for external validation (IRB UCB 2017-09-10340). The dataset used in this analysis (iPAX 10K) consists of a subset of 8,969 individuals with sufficient clinical data to calculate the conventional PCE risk score. Of these, 978 were excluded because they had developed CVD before the date of retinal imaging. Therefore, the external validation dataset consists of 18,900 retinal images from 7,861 individuals. The composition of the dataset used in this study is shown in Table 4. [Table 4] Demographic and risk factor composition of the training and internal test datasets from the UK Biobank and the iPAX 10K external validation dataset used in this study. Significance tests were performed between training and testing in the UK Biobank, and between training in the biobank and external validation in the iPAX 10K (*P<0.01 z-test, **P<0.001 z-test, †P<0.01 chi-squared). In the UK Biobank, Hispanic ethnicity was not included as an option, and white participants were mainly from the UK and Ireland. Since the iPAX 10K included Hispanic, Black, and White options, it was not possible to distinguish between Hispanic Black participants and Hispanic White participants. [Table 4-1] [Table 4-2]

[0118] Furthermore, levels of diabetic retinopathy were extracted from the iPAX 10K dataset, and the most severe retinopathy items observed in both eyes were recorded (see Table 5). [Table 5] Distribution of diabetic retinopathy grades in the iPAX 10K dataset: R0 = no DR, R1 = mild nonproliferative diabetic retinopathy in one eye only, R1 + = mild NPDR in both eyes, R2 = moderate NPDR, R3 = severe NPDR, R4 = proliferative DR [Table 5]

[0119] In this example, inclusion criteria included individuals with data in the dataset that included demographic information such as age, sex, and ethnicity, at least one good-quality image from each eye, and information on CVD risk factors (such as systolic blood pressure, HbA1c, total cholesterol, and HDL cholesterol).

[0120] In this example, the inclusion criteria included individuals who were under 40 years old or over 75 years old at the time of retinal imaging, those who were pregnant at the time of retinal imaging, those with a known history of cardiac events such as stroke or heart attack prior to retinal imaging, and those whose datasets recorded persistent visual impairment, congenital eye disease, or severe eye injury (in one or both eyes at the time of retinal imaging).

[0121] 3.1 Model Evaluation Individuals within each dataset are divided into age groups based on 10-year lifespans, and cardiac biological age is generated. To validate the output of the deep learning model, the age difference (cardiac biological age - chronological age) of individuals is calculated. Individuals within each age cohort are ranked in the highest to lowest quartile of the age difference, and profiles of CVD risk biomarkers (systolic blood pressure (SBP), HbA1c, and total cholesterol / HDL ratio (TChol / HDL) in this example) are compared between individuals in the upper and lower quartiles of the age difference. Finally, similar calculations are performed for individuals with diabetes to compare the presence and severity of diabetic retinopathy (DR). The following statistical analyses are performed and reported: ● Linear regression between biological age and chronological age Report: R2, slope, intercept Plot: Scatter plot of biological age and chronological age ● Comparison of biomarker profiles between individuals with age differences in the upper quartile of an age cohort (the top 25% of individuals with the highest age differences) and individuals with age differences in the lower quartile of an age cohort (the bottom 25% of individuals with the lowest age differences): o Systolic blood pressure oHbA1c o Total cholesterol / HDL ratio o Whether diabetic retinopathy is present in the iPAX 10K dataset (if diabetic retinopathy is asymmetric, the highest grade from both eyes will be used). o Decade CVD risk defined by the pooled cohort equation (PCE).

[0122] The correlation between an individual's predicted cardiac biological age and chronological age in both the UK Biobank and iPAX 10K datasets is shown in Figures 7A and 7B, respectively. Figure 7A shows the distribution of cardiac biological age and chronological age in the UK Biobank dataset (R-squared: 0.89, slope: 0.93, intercept: 3.38). Figure 7B shows the distribution of cardiac biological age and chronological age in the iPAX 10K dataset (R-squared: 0.51, slope: 0.54, intercept: 29).

[0123] In the internally validated UK Biobank dataset, mean cardiac biological age was very close to recorded chronological age across all age groups (R-squared 0.89). In contrast, the mean cardiac biological age in the externally validated iPAX 10k dataset was significantly higher compared to chronological age across all age groups under 62 years. Subsequently, mean cardiac biological age became comparable to chronological age and even lower than chronological age in the older age group (R-squared 0.51). The difference between cardiac biological age and chronological age was greatest in younger patients and gradually decreased until the early 60s.

[0124] In both the UK Biobank internal validation dataset and the iPAX 10K external validation dataset, individuals were categorized into three age groups: 40-50, 50-60, and 60-70 years, based on their life decade. Within each age cohort, the age difference for each individual was calculated, and the cohorts were arranged in ascending quartiles, with the highest age difference grouped in the highest quartile and the lowest age difference grouped in the lowest quartile. Next, the mean (and standard deviation) of the following cardiac biomarkers were evaluated: SBP, HbA1c, TChol / HDL ratio, and 10-year CVD.

[0125] In both the internal and external validation datasets, mean HbA1c and mean SBP were significantly higher for individuals in the upper quartile of the age distribution compared to individuals in the lower quartile of the age distribution across almost all age groups (see Tables 6-9). A similar trend was observed in the UK Biobank total TChol / HDL ratio results, but there were no significant differences in TChol / HDL ratios across any age cohort in the iPAX 10K dataset (see Tables 10 and 11). [Table 6] Mean and standard deviation of HbA1c (%) measurements for the top 25% and bottom 25% of age-differential participants in each age decile in the UK Biobank [Mean (Standard Deviation)] [Table 6] [Table 7] Mean and standard deviation of HbA1c (%) measurements for the top 25% and bottom 25% of age-differential participants in each age decile in the iPAX 10K [Mean (Standard Deviation)] [Table 7] [Table 8] Mean and standard deviation (mean (SD)) of systolic blood pressure (mmHg) measurements for the highest 25% and lowest 25% of age-differential participants in each age quintile at the UK Biobank. [Table 8] [Table 9] Mean and standard deviation (mean (SD)) of systolic blood pressure (mmHg) measurements for the systolic 25% and diastolic 25% of age-differential participants in each age decile in the iPax 10K. [Table 9] [Table 10] Mean and standard deviation [mean (SD)] of TChol / HDL measurements for the top 25% and bottom 25% of age-differential participants in 10 age-group categories in the UK Biobank. [Table 10] [Table 11] Mean and standard deviation [mean (SD)] of TChol / HDL ratio measurements for the top 25% and bottom 25% of age-differential participants in each age decile in the iPax 10K. [Table 11]

[0126] The iPAX 10K dataset also provides information on diabetic retinopathy grades. Examining the proportion of individuals with diabetic retinopathy in two quartiles, individuals with age differences in the upper quartile of the age cohort distribution had a significantly higher prevalence of diabetic retinopathy compared to individuals in the lower quartile (see Table 12). As the severity of diabetic retinopathy increased, the observed age difference gradually increased in all categories except proliferative diabetic retinopathy (see R4, Figure 8: Relationship between diabetic retinopathy severity and age difference in the iPAX dataset (F-value = 168, p-value = < 0.001)). [Table 12] Presence of any level of diabetic retinopathy (shown as the percentage with DR) in the top 25% and bottom 25% of age-gap participants at each age 10% of the iPax 10K dataset. [Table 12]

[0127] Once an individual's cardiac biological age is determined, this value is compared to their actual chronological age. The difference between an individual's chronological age and cardiac biological age in this age group is called the "age difference." When grouped by life decade compared to their peers, individuals with age differences in the upper quartiles of each age cohort had significantly higher SBP and HbA1c than individuals with age differences in the lower quartiles of most age cohorts, in both internal and external validation datasets. The TChol / HDL data is less clear. In both the upper and lower quartiles of age differences in the iPax dataset, over 70% of individuals aged 50 and older were taking statins. In comparison, only 20% of high-risk elderly patients in the UK Biobank were taking medication.

[0128] The inventors found a correlation between chronological age and cardiac biological age in both datasets, but the strength of this correlation was stronger in the UK Biobank compared to the iPAX 10K dataset (R-squared 0.89 vs 0.51, slope 0.93 vs 0.54: UK Biobank vs iPAX 10K). The DL model can identify a trend that, at least at the population level, diabetic individuals age more rapidly than non-diabetic individuals. The cardiac biological age of individuals in the iPAX 10K dataset was significantly higher than that of their non-diabetic peers in all age groups compared to individuals in the UK Biobank, and therefore the age difference was also greater. The DL model can recognize that the age difference in younger individuals with diabetes in the iPAX 10K dataset is significantly worse than that of older individuals with diabetes.

[0129] One of the components of the DL model suite that makes up the cardiac biological age DL model is a DR function detector, which is uniquely trained to determine the degree of diabetic retinopathy. Reviewing the iPAX 10K data revealed that the DR function detector was clearly activated. In the upper quartile of the age difference, less than 47% had R0, and the remainder had R2 or lower. In comparison, in the lower quartile of the age difference, more than 88% had R0, while only 5% had R2 or lower (see Table 13). [Table 13] Prevalence of diabetic retinopathy in the iPAX 10K dataset (upper and lower quartiles for each age group and cardiac biological age) [Table 13]

[0130] The inventors found that in all age groups, the presence of DR was associated with a significantly higher age difference compared to individuals without DR (see Table 12). Furthermore, apart from proliferative diabetic retinopathy, where the cardiac biological age difference was similar to that of severe non-proliferative DR, the age difference increased continuously with the severity of the individual's retinopathy when diabetic retinopathy was present (see Figure 8).

[0131] This study externally validates a cardiac biological age model designed to detect individuals with higher cardiovascular risk factors compared to their peers, based solely on retinal images and limited demographic data. Furthermore, this study demonstrates that the DL model can further stratify risk based on the presence or absence of DR within a high-risk population of individuals with diabetes.

[0132] 4. Comparison of cardiac biological age and leukocyte telomere length Studies have reported an inverse correlation between CVD events, risk, and hypertension and the biological marker leukocyte telomere length (LTL). We compared cardiac biological age metrics using a subset of the UK Biobank dataset.

[0133] Individuals were divided into males and females, and each cohort was ranked by LTL, grouped into 10 stages from shortest to longest. Next, for each LTL decile, the individual's retinal image was presented to a cardiac biological age DL model to determine cardiac biological age. Then, individuals within each LTL decile were ranked by cardiac biological age, and the upper and lower quartiles of each decile were determined. For each LTL 10-minute, the means of the following variables obtained from individuals in the upper and lower quartiles were compared: systolic blood pressure, HbA1c, and PCE 10-year CVD risk score.

[0134] Chronic age and cardiac biological age issued by the DL model for men, ranked by mean logLTL decile, are shown in Figures 9A and 9B, respectively, and for women in Figures 10A and 10B. Systolic blood pressure, HbA1c, and 10-year PCE CVD risk scores for individuals belonging to the upper (Q1) and lower (Q4) quartiles of their chronological peer group, for which cardiac biological age scores were issued, classified by mean logLTL decile, are shown in Figures 11A-11C (men), Figures 12A-12C (women), and Table 14. [Table 14] Mean biomarker values ​​for men and women in the upper quartile (Q1) of cardiac biological age compared to the lower quartile (Q4) of the same age group. [Table 14]

[0135] In both men and women, mean systolic blood pressure, mean HbA1c, and mean 10-year PCE CVD risk score were significantly higher in individuals in the highest quartile of cardiac biological age compared to individuals in the lowest quartile (P<0.001 - see Table 14). Furthermore, LTL was significantly shorter in men compared to older women, and in both men and women, increasing chronological age and increasing cardiac biological age were inversely correlated with LTL (see Figures 9A and 9B, 10A and 10B).

[0136] The inventors found that worsening biomarker profiles, increased chronological age, and increased cardiac biological age were associated with shortened LTL. Furthermore, they found that the DL cardiac biological age model could accurately classify patients into those at high risk of CVD and those at low risk of CVD disease based on relevant CVD biomarkers and CVD risk scores derived from 10-year PCE. Consistent with conventional CVD models, both men and women whose cardiac biological age was issued in the upper quartile of their peer group had significantly higher mean SBP, HbA1c, and 10-year PCE CVD scores compared to individuals in the lower quartile (e.g., PCE).

[0137] Novel biomarkers such as LTL have been suggested to potentially capture cardiac health, and studies using Mendelian randomization in particular have provided compelling evidence linking LTL shortening to an increased risk of atherosclerotic cardiovascular events and hypertension. Lifetime telomere shortening is thought to be determined by both intrinsic (genetic) and extrinsic (non-genetic) factors, both of which appear to contribute to the association between LTL and CVD risk. At the time of application, the precise mechanisms underlying telomere shortening and CVD are not yet fully understood, but LTL is thought to reflect both an individual's cumulative inflammatory exposure and oxidative stress, as well as their genetically determined ability to repair vascular damage. Regardless of the mechanism by which LTL shortens, the results support the hypothesis that a well-trained DL model may not only be able to predict an individual's CVD risk as assessed by conventional biomarkers, but may also be able to assess risk using the novel biomarker LTL.

[0138] 5. Announcement of Cardiac Biological Age Results Referring to Figure 13, an individual's cardiac biological age results may be presented in Report 5000, which includes a cardiac biological age infographic 5002 showing the individual's chronological age 5004 and cardiac biological age 5006 on a relative scale. In the example shown, the age difference is positive, meaning that the cardiac biological age is higher than the chronological age. A textual explanation of this is provided in the Age Difference section 5008.

[0139] Explanation Section 5010 summarizes key individual recommendations for improving cardiac biological age. In the illustrated example, where the age difference is relatively high, lifestyle changes known to be associated with health and well-being are strongly recommended. These recommendations may be adjusted based on an individual's age difference; for example, individuals with a negative or neutral age difference may be advised to continuously monitor their biological age (e.g., annually), while individuals with a positive but not high age difference may be encouraged to consider lifestyle changes.

[0140] Current technologies have the potential to significantly improve access to personalized health recommendations, particularly CVD risk prevention strategies, by generating risk predictions that do not require clinical or laboratory assessments to generate individual risks. Since retinal images are routinely taken during optometrial examinations, this technology can be implemented without significant additional investment in primary care. This characteristic makes these technologies particularly suitable for resource-constrained environments. Finally, AI-based predictive tools that assess risk at the individual level can help in treatment decisions based on an individual's specific needs, thereby increasing the likelihood of favorable health outcomes.

[0141] All references, including patents or patent applications, cited herein are incorporated herein by reference. No reference is considered to constitute prior art. Discussions regarding references are those of their authors, and applicants reserve the right to object to the accuracy and validity of the cited references. While numerous prior art documents are referenced herein, it is clear that these references do not constitute an admission that any of these documents form part of the common general knowledge of the art in any country of the world.

[0142] Unless otherwise clearly indicated in the context, throughout the description and claims, words such as “includes” and “equipment” shall be interpreted in a comprehensive sense, i.e., “includes, but not limited to,” and not in an exclusive or exhaustive sense.

[0143] Broadly speaking, the present invention consists of the parts, elements, and features referred to or shown herein, individually or as a set, or of any combination of two or more or all of the said parts, elements, or features. Where, in the foregoing description, an integer or its equivalent is referred to as a known component, those integers are incorporated herein as if they were described individually.

[0144] Various changes and modifications to the currently preferred embodiments described herein will be apparent to those skilled in the art. Such changes and modifications can be made without departing from the spirit and scope of this disclosure and without impairing the advantages associated with this disclosure. Accordingly, such changes and modifications are intended to be included within the scope of this disclosure as defined by the appended claims.

Claims

1. A method for determining at least one recommendation regarding an individual's health management, A step of determining an individual's relative cardiovascular aging indicator ("cardiac biological age") based at least in part on the individual's predicted risk of cardiovascular disease (CVD) determined by a deep learning model based on one or more fundus images, A method comprising the step of determining at least one recommendation for the health management of an individual, based at least in part on the determined cardiac biological age.

2. This includes a step of determining the difference ("age difference") between an individual's actual chronological age and the cardiac biological age determined for that individual. The method according to claim 1, wherein the step of determining at least one of the aforementioned recommendations relating to the health management of an individual is based at least in part on the determined age difference.

3. The method according to claim 2, further comprising the step of comparing the age difference of the individual with the age difference of a set of individuals belonging to the chronological age category to which the individual belongs.

4. The method according to claim 3, further comprising the step of determining the relative position of individuals within a set of individuals based on the age difference.

5. The method according to claim 4, wherein the step of determining at least one recommendation regarding the health management of the said individual is at least in part based on the individual's relative position within the set of individuals based on age difference.

6. The method according to any one of claims 1 to 5, comprising the step of determining the relative contribution of one or more risk contributors to an indication of relative cardiovascular aging.

7. The method according to claim 6, wherein the risk contributing factors include two or more of the following: blood pressure, glycated hemoglobin A1c (HbA1c), total cholesterol, and blood glucose control.

8. The method according to claim 6 or 7, wherein the relative contribution of the one or more risk-contributing factors is used to determine at least one recommendation regarding the health management of the individual.

9. The step of determining the cardiobiological age is, A step of determining the similarity between the predicted risk of cardiovascular disease (CVD) for a first individual and the risk of CVD for a set of individuals belonging to the same chronological age group as the first individual, wherein the predicted risk of CVD is determined by a deep learning model based on one or more fundus images. The steps include determining the average chronological age of the person closest to the first individual from the perspective of predicting the risk of the aforementioned CVD, The steps include determining the average expected chronological age of individuals with a predicted risk of CVD, The method according to any one of claims 1 to 8, comprising the step of determining an indication of relative cardiovascular aging for the first individual based on the determined similarity, the determined average chronological age, and the determined average predicted chronological age.

10. The step of determining the cardiac biological age of the first individual includes the following calculation: [Math 1] Here, x refers to the first individual, and X is a set of individuals in the same age group as the first individual. Here, P(risk(x) | x∈X) is the conditional probability that the first individual x has a risk(x) that is a CVD risk, assuming that the first individual belongs to the set of individuals X of the same age. Here, age (X) is the average age of set X. The method according to claim 9, wherein age (Y | x ~ Y) is the average age of patient points close to the first individual x.

11. The method according to claim 10, wherein the set of individuals includes individuals of the same sex as the first individual.

12. The method according to claim 10 or claim 11, wherein the implementation of age (Y | x ~ Y) includes the magnitude of similarity * cosine similarity age + (1 - magnitude of similarity) * risk extrapolation age.

13. A system for determining at least one recommendation relating to personal health management, the system comprising one or more processors and one or more storage devices that, when executed by the one or more processors, store instructions causing the one or more processors to perform the following operations, the operations being: A step of determining an individual's relative cardiovascular aging indicator ("cardiac biological age") based at least in part on the individual's predicted risk of cardiovascular disease (CVD) determined by a deep learning model based on one or more fundus images, A system comprising the step of determining at least one recommendation for the health management of the individual, based at least in part on the determined cardiac biological age.

14. A computer program product for determining at least one recommendation regarding an individual's health management, Based at least in part on an individual's predicted risk of cardiovascular disease (CVD) determined by a deep learning model based on one or more fundus images, an individual's relative cardiovascular aging indicator ("cardiac biological age") is determined. A computer program product comprising a non-temporary computer-readable storage medium containing computer program code for determining at least one recommendation regarding the health management of the individual, at least in part based on the determined cardiac biological age.

Images