Methods and systems for analyzing reproductive aging using retinal age gap
The retinal age gap, calculated via AI-enhanced retinal imaging and genetic data integration, addresses the invasiveness and variability of current reproductive aging measures, providing a reliable, non-invasive biomarker for ovarian reserve and fertility prediction.
Patent Information
- Application Number
- PCT/US2026/027171
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-09
- Filing Date
- 2026-05-07
- Publication Date
- 2026-11-12
AI Technical Summary
Current methods for measuring reproductive aging are invasive and subject to variability due to hormonal fluctuations, necessitating non-invasive biomarkers for enhanced predictive accuracy.
Utilizing artificial intelligence to calculate the retinal age gap, defined as the difference between predicted retinal age and chronological age, as a biomarker for reproductive aging, through retinal fundus imaging and a deep learning framework like FLEX, which integrates genetic data to predict AMH levels and reproductive health.
The retinal age gap provides a non-invasive, accurate assessment of reproductive aging, correlating with AMH levels and reproductive events, offering insights into ovarian reserve and fertility, and potentially predicting menopause and other health indicators.
Smart Images

Figure US2026027171_12112026_PF_FP_ABST
Abstract
Description
Attorney Docket No. 71951-716.602METHODS AND SYSTEMS FOR ANALYZING REPRODUCTIVE AGING USING RETINAL AGE GAP CROSS REFERENCE TO RELATED APPLICATIONS
[0001] This application claims the benefit of International Patent Application No.PCT / CN2025 / 093808, filed on May 9, 2025, entitled “RETINAL AGE GAP AS A NON-INVASIVE BIOMARKER FOR REPRODUCTIVE AGING,” which is entirely incorporated herein by reference for all purposes.FIELD
[0002] The present disclosure relates generally to methods and systems using retinal age gap as a biomarker for reproductive aging.BACKGROUND
[0003] Reproductive aging is a complex biological process influenced by various hormonal, genetic, and environmental factors. As women age, the decline in ovarian reserve significantly impacts fertility, highlighting the need for reliable biomarkers to assess reproductive aging.However, current method for measuring reproductive aging involves invasive procedures, and its accuracy is subject to variability due to hormonal fluctuations. This underscores the need for complementary indicators using noninvasive measurement methods to enhance predictive accuracy. The present disclosure addresses these and other needs.SUMMARY
[0004] In some embodiments, provided herein are methods and systems using artificial intelligence-derived retinal age gap, which is the difference between predicted retinal age and chronological age, as a marker for reproductive aging in women.
[0005] In some embodiments, provided herein is a method comprising: obtaining retinal fundus images of a subject; using artificial intelligence to predict a retinal age of the subject based on the retinal fundus images; calculating a retinal age gap, which is the difference between predicted retinal age and chronological age; and using the retinal age gap as a biomarker for a disease or condition in the subject. In some embodiments, the disease or condition is biological aging. In some embodiments, the disease or condition is reproductive aging. In some embodiments, the disease or condition is stroke. In some embodiments, the disease or condition is a kidney disease. In some embodiments, the disease or condition is Parkinson’s disease. In some embodiments, the disease or condition is central obesity. In some embodiments, the disease or condition is mortality risk. In some embodiments, the disease or condition is an age-related disease. In any one of theAttorney Docket No. 71951-716.602embodiments disclosed herein including any of the preceding embodiments, the retinal age gap can be negatively associated with Anti -Mullerian Hormone (AMH) level in the subject.
[0006] In some embodiments, provided herein is a method, comprising: (a) obtaining retinal fundus images of a subject; (b) determining a retinal age of the subject based on the retinal fundus images, thereby determining a retinal age gap, which is the difference between the retinal age and the subject’s chronological age; and (c) determining, from the retinal age gap, a state of reproductive aging in the subject.
[0007] In any one of the embodiments herein including any preceding embodiment, the method can comprise obtaining the retinal fundus images using non-invasive fundus imaging. In any one of the embodiments herein including any preceding embodiment, one or more of the retinal fundus images can comprise microvascular features.
[0008] In any one of the embodiments herein including any preceding embodiment, the method can further comprise predicting the retinal age of the subject using a model trained on fundus image features from a healthy cohort of subjects. In some embodiments, the model is trained to learn optimal weights of the fundus image features that minimize the difference between a predicted retinal age and its corresponding chronological age.
[0009] In any one of the embodiments herein including any preceding embodiment, the model can be trained using transfer learning from the fundus image features from the healthy cohort of subjects. In any one of the embodiments herein including any preceding embodiment, the model can be trained using supervised fine-tuning based on fundus image features from a medium AMH (Anti -Mullerian Hormone) group in the healthy cohort of subjects. In some embodiments, the medium AMH group comprises the middle two quartiles of the healthy cohort which is stratified by AMH levels into quartiles.
[0010] In any one of the embodiments herein including any preceding embodiment, the model can comprise a dual-channel transfer learning architecture. In any one of the embodiments herein including any preceding embodiment, the model can comprise two parallel Swin Transformer encoders. In any one of the embodiments herein including any preceding embodiment, the model can comprise (i) a first encoder, the parameters of which are kept frozen, (ii) a second encoder, the parameters of which are allowed to learn from new data, preserving the model's feature extraction while updating it to suit the fundus images from the healthy cohort of subjects. In any one of the embodiments herein including any preceding embodiment, the model can further comprise fusing outputs of the first encoder and the second encoder. In any one of the embodiments herein including any preceding embodiment, the model can further comprise processing the fused outputs through normalization and adaptive pooling to predict the retinal age.Attorney Docket No. 71951-716.602
[0011] In any one of the embodiments herein including any preceding embodiment, the retinal age gap can be a positive reginal age gap indicating accelerated retinal aging. In any one of the embodiments herein including any preceding embodiment, the retinal age gap can be correlated with an AMH (Anti -Mullerian Hormone) level in the subject. In any one of the embodiments herein including any preceding embodiment, the method can further comprise predicting the subject has a low AMH level based on a more positive retinal age gap. In any one of the embodiments herein including any preceding embodiment, the method can further comprise predicting the subject has accelerated reproductive aging based on a more positive retinal age gap. In any one of the embodiments herein including any preceding embodiment, the method can further comprise predicting the subject has low ovarian reserve based on a more positive retinal age gap. In any one of the embodiments herein including any preceding embodiment, the method can further comprise predicting the subject has a higher level of a sexual hormone based on a more positive retinal age gap. In any one of the embodiments herein including any preceding embodiment, the sexual hormone can comprise follicle-stimulating hormone (FSH). In any one of the embodiments herein including any preceding embodiment, the sexual hormone can comprise prolactin (PRL).
[0012] In any one of the embodiments herein including any preceding embodiment, the method can further comprise predicting earlier menopause based on a more positive retinal age gap. In any one of the embodiments herein including any preceding embodiment, the subject can be a woman between 30 and 39. In any one of the embodiments herein including any preceding embodiment, the subject can be a woman between 40 and 44. In any one of the embodiments herein including any preceding embodiment, the subject can be a woman between 45 and 49. In any one of the embodiments herein including any preceding embodiment, the subject can be a woman between 50 and 54. In any one of the embodiments herein including any preceding embodiment, the subject can be a woman 55 or older.
[0013] In some embodiments, provided herein is a system comprising: at least one hardware processor; and one or more software modules configured to, when executed by the at least one hardware processor, perform the method of any one of the embodiments herein including any preceding embodiment.
[0014] In some embodiments, provided herein is a non-transitory computer-readable medium having instructions stored thereon, wherein the instructions, when executed by a processor, cause the processor to perform the method of any one of the embodiments herein including any preceding embodiment.Attorney Docket No. 71951-716.602
[0015] In some embodiments, provided herein is a system comprising: at least one hardware processor; a non-transitory computer-readable medium coupled to at least one hardware processor, optionally wherein the coupling is over a network; and instructions stored in the non-transitory computer-readable medium, wherein the instructions when implemented by the processor, configure the system to perform the method of any one of the embodiments herein including any preceding embodiment.
[0016] Additional aspects and advantages of the present disclosure will become readily apparent to those skilled in this art from the following detailed description, wherein only illustrative embodiments of the present disclosure are shown and described. As will be realized, the present disclosure is capable of other and different embodiments, and its several details are capable of modifications in various obvious respects, all without departing from the disclosure. Accordingly, the drawings and description are to be regarded as illustrative in nature, and not as restrictive.BRIEF DESCRIPTION OF THE DRAWINGS
[0025] The following drawings illustrate certain embodiments of the features and advantages of this disclosure. These embodiments are not intended to limit the scope of the appended claims in any manner. Like reference symbols in the drawings indicate like elements.
[0026] FIG. 1 provides an overview of the workflow of the proposed FLEX framework, outlining the development and application of retinal age models in healthy women to explore the relationship between retinal aging and reproductive aging.
[0027] FIGS.2A-2F provide data distribution and retinal age prediction in the Healthy Cohort and the Healthy Female Cohort.
[0028] FIGS. 3A-3F provide differences in retinal age gap between different AMH levels across age groups and the correlation with log AMH. The association between retinal age gap and AMH levels is analyzed through group comparisons and correlation analyses across different age groups.FIG. 3A shows a comparison of retinal age gap between the high AMH group (red) and the low AMH group (blue) overall and across different age groups. * - value < 0.05. FIGS.3B-3F show the correlation between retinal age gap and log AMH levels across different age groups. FIG. 3B shows 25-29 years old. FIG. 3C shows 30-34 years old. FIG.3D shows 35-39 years old. FIG. 3E shows 40-44 years old. FIG. 3F shows 45-50 years old.
[0029] FIGS. 4A-4C provide logistic regression analysis of AMH levels based on retinal age gap and reproductive characteristics. Univariate and multivariate logistic regression analyses evaluating the association between AMH levels, retinal age gap, and reproductive factors. FIG. 4A shows univariate logistic regression analysis of AMH Levels based on retinal age gap across different ageAttorney Docket No. 71951-716.602groups. FIG. 4B shows univariate logistic regression analysis of AMH levels based on retinal age gap, childbirth number, pregnancy number, First childbirth age, and First pregnancy age. FIG. 4C shows multivariate logistic regression analysis of AMH levels based on retinal age gap, childbirth number, pregnancy number, first childbirth age, and first pregnancy age, adjusted for chronological age, eye laterality, smoking status, drinking status, and marriage status. Error bars represent 95% confidence intervals (CI) for odds ratios (OR). OR: Odds ratio, CI: Confidence interval. The high AMH group serves as the reference group for logistic regression analysis against the low AMH group, and - values < 0.05 are considered statistically significant.
[0030] FIG. 5 provides heat maps for different age groups. Heat maps showing age estimation using the FLEX model based on fundus images from different age groups within the Healthy Female Cohort.
[0031] FIGS. 6A-6C provide a flowchart of gene analysis and enrichment results. Overview of the gene analysis workflow and corresponding enrichment analysis results, detailing the key steps and methodologies involved. FIG. 6A shows flowchart illustrating the gene analysis process. Starting with 16,987,370 Single Nucleotide Polymorphisms (SNPs), quality control (QC) filtering resulted in 8,089,940 SNPs for subsequent gene feature selection. Gene features selection involved a Genome wide association study (GWAS) for retinal age gap, identifying 7,7'0 SNPs with P < 0.001, followed by LassoNet to select the top 4,000 ranked SNPs based on importance. The selected genes then underwent gene annotation, as well as Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. FIG. 6B shows results of the GO enrichment analysis. FIG. 6C shows results of the KEGG enrichment analysis. SNPs: Single Nucleotide Polymorphisms, QC: quality control, GWAS: Genome wide association study, GO: Gene Ontology, KEGG: Kyoto Encyclopedia of Genes and Genomes.DETAILED DESCRIPTION
[0032] All publications, patents, and patent applications mentioned in this specification are herein incorporated by reference in their entirety to the same extent as if each individual publication, patent, or patent application was specifically and individually indicated to be incorporated by reference in its entirety. In the event of a conflict between a term herein and a term in an incorporated reference, the term herein controls.
[0033] While various embodiments of the present disclosure have been shown and described herein, it will be obvious to those skilled in the art that such embodiments are provided by way of example only. Numerous variations, changes, and substitutions may occur to those skilled in theAttorney Docket No. 71951-716.602art without departing from the present disclosure. It should be understood that various alternatives to the embodiments of the present disclosure described herein may be employed.
[0034] The section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described.
[0035] Reproductive aging is a complex biological process influenced by various hormonal, genetic, and environmental factors. As women age, the decline in ovarian reserve significantly impacts fertility, highlighting the need for reliable biomarkers to assess reproductive aging. Anti-Mullerian Hormone (AMH) is widely recognized as a robust biomarker for ovarian reserve, offering valuable insights into reproductive potential. However, current method for measuring AMH levels involves invasive procedures, and its accuracy is subject to variability due to hormonal fluctuations. This underscores the need for complementary indicators using noninvasive measurement methods to enhance predictive accuracy.
[0036] Notably, AMH levels have also been linked to vascular aging, with studies showing correlations between AMH and markers of vascular health, including intima-media thickness, coronary artery calcification, and chronic kidney disease. These findings suggest a bidirectional relationship, where vascular aging could serve as a proxy for ovarian aging.
[0037] Noninvasive retinal imaging has emerged as a novel tool to leverage this connection. Retinal fundus imaging captures microvascular features in the eye that reflect systemic aging. Advances in artificial intelligence have further enabled the calculation of the retinal age gap, defined as the difference between predicted retinal age and chronological age. This measure has shown potential as a biomarker for both biological and reproductive aging, including stroke, kidney diseases, Parkinson's disease, central obesity, as well as mortality risk and age-related diseases. Given the vascular underpinnings of both retinal and ovarian aging, retinal imaging may represent a groundbreaking, non-invasive approach to predicting AMH levels and, by extension, reproductive health.
[0038] In some embodiments, the present disclosure provides Frozen and Learning Ensemble Crossover (FLEX), a deep learning framework designed to predict retinal age from fundus images with high precision. In some aspects, by leveraging fundus images and incorporating healthy model weights derived from healthy cohorts (FIG. 1, panel a), the FLEX model is developed using data from the Healthy Female Cohort and the association between the retinal age gap and AMH levels is examined. In some aspects, the link between retinal aging and reproductive health is explored, integrating genetic data to uncover the interplay between vascular aging and ovarian reserve, while also enhancing reproductive health assessment through the fusion of genetic and imaging data (FIG. IB, panel b). By combining advanced Al techniques with retinal imaging,Attorney Docket No. 71951-716.602provided herein is a multimodal approach to understanding reproductive aging, providing valuable insights into the genetic and physiological mechanisms underlying ovarian reserve, and paving the way for improved diagnostic tools and targeted interventions in reproductive medicine.
[0039] In some embodiments, provided herein is a method, comprising: (a) obtaining retinal fundus images of a subject; (b) determining a retinal age of the subject based on the retinal fundus images, thereby determining a retinal age gap, which is the difference between the retinal age and the subject’s chronological age; and (c) determining, from the retinal age gap, a state of reproductive aging in the subject. In some embodiments, the method comprises predicting the retinal age of the subject using a model trained on fundus image features from a healthy cohort of subjects. In some embodiments, the method comprises training the model to learn optimal weights of the fundus image features that minimize the difference between a predicted retinal age and its corresponding chronological age. In some embodiments, the model is trained using supervised fine-tuning based on fundus image features from a medium AMH group in the healthy cohort of subjects. In some embodiments, the medium AMH group comprises the middle two quartiles of the healthy cohort which is stratified by AMH levels into quartiles.
[0040] In some embodiments, the model comprises a dual-channel transfer learning architecture. In some embodiments, the model comprises two parallel Swin Transformer encoders. In some embodiments, the model comprises (i) a first encoder, the parameters of which are kept frozen, (ii) a second encoder, the parameters of which are allowed to learn from new data, preserving the model's feature extraction while updating it to suit the fundus images from the healthy cohort of subjects. In some embodiments, the model comprises fusing outputs of the first encoder and the second encoder. In some embodiments, the model comprises processing the fused outputs through normalization and adaptive pooling to predict the retinal age. An example of the workflow is shown in FIG. 1.
[0041] In FIG. 1, panel a shows Healthy Cohort Retinal Age Model Weights were trained using transfer learning based on the Healthy Cohort, and these weights were subsequently utilized for fundus feature extraction. FIG. 1, panel b shows the study involved ophthalmic examinations, questionnaires, and AMH testing for the Healthy Female Cohort (middle panel). The cohort was stratified by AMH levels into quartiles, with the middle two quartiles (medium AMH group) being used to train a Female Retinal Age Model using the Healthy Cohort Retinal Age Model Weights (upper left panel). The trained model was then applied to predict retinal age for the highest quartile (high AMH group) and the lowest quartile (low AMH group). The retinal age gap served as a measure of an individual's reproductive aging level. Further analyses were conducted to explore the relationship between the retinal age gap and AMH levels, as well as between reproductiveAttorney Docket No. 71951-716.602characteristics and AMH levels (lower left panel). Whole-genome sequencing identified Single Nucleotide Polymorphisms (SNPs), which were analyzed in relation to the predicted retinal age gap (top right panel). Additionally, retinal images were integrated with genetic data to predict AMH levels (bottom right panel).
[0042] FIGS.2A-2F provide a comprehensive overview of participant demographics and the performance assessment of retinal age prediction in the Healthy Cohort and Healthy Female Cohort. FIG. 2A shows the distribution of male and female participants across different age groups in the Healthy Cohort, with females represented in red and males in teal. FIG.2B shows a scatter plot illustrating the relationship between age and AMH levels within the Healthy Female Cohort, stratified into three groups based on AMH levels: high (cyan), medium (purple), and low (yellow).FIG. 2C shows scatterplots showing the correlation between retinal age and chronological age in the Healthy Cohort, where data points are color-coded based on the retinal age gap, with larger gaps shown in red and smaller gaps in blue. FIG. 2D shows Bland-Altman plots illustrating the agreement between retinal age and chronological age in the Healthy Cohort, with similar colorcoding based on retinal age gap. FIG. 2E shows scatterplots displaying the correlation between retinal age and chronological age in the Healthy Female Cohort, with data points color-coded according to the retinal age gap (red for larger gaps and blue for smaller gaps). FIG.2F shows Bland- Altman plots depicting the agreement between retinal age and chronological age in the Healthy Female Cohort, with the same color scheme based on retinal age gap.
[0043] In some embodiments, the retinal age gap is a positive reginal age gap indicating accelerated retinal aging. In some embodiments, the retinal age gap is correlated with an AMH level in the subject. In some embodiments, the method comprises predicting the subject has a low AMH level based on a more positive retinal age gap. In some embodiments, the method comprises predicting the subject has accelerated reproductive aging based on a more positive retinal age gap. In some embodiments, the method comprises predicting the subject has low ovarian reserve based on a more positive retinal age gap. In some embodiments, the method comprises predicting the subject has an higher level of a sexual hormone based on a more positive retinal age gap. In some embodiments, the sexual hormone comprises follicle-stimulating hormone (FSH). In some embodiments, the sexual hormone comprises prolactin (PRL). In some embodiments, the method comprises predicting earlier menopause based on a more positive retinal age gap. In some embodiments, the subject is a woman between 30 and 39. In some embodiments, the subject is a woman between 40 and 44. In some embodiments, the subject is a woman between 45 and 49. In some embodiments, the subject is a woman between 50 and 54. In some embodiments, the subject is a woman 55 or older.Attorney Docket No. 71951-716.602
[0044] In some embodiments, a method disclosed herein comprising retinal image-based age prediction, for instance, using a model that forecast retinal age based on fundus images. In some embodiments, the model produces consistent retinal age predictions across both temporally and spatially different sources. In some embodiments, a method disclosed herein comprising using the retinal age gap, which is associated with AMH levels, to predict AMH levels (e.g., using the retinal age gap alone or optionally in combination with one or more reproductive events). In some embodiments, age-specific associations between retinal age gap and AMH levels are used to predict reproductive aging of the subject, e.g., a woman between 40 and 44. In some embodiments, sexual hormones and / or time of menopause are associated with retinal age gap, which is used to predict reproductive aging of the subject.
[0045] Unless otherwise defined, all of the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art in the field to which this disclosure belongs.
[0046] As used in this specification and the appended claims, the singular forms "a", "an", and "the" include plural references unless the context clearly dictates otherwise. Any reference to "or" herein is intended to encompass "and / or" unless otherwise stated.
[0047] " About" and "approximately" shall generally mean an acceptable degree of error for the quantity measured given the nature or precision of the measurements. Exemplary degrees of error are within 20 percent (%), typically, within 10%, and more typically, within 5% of a given value or range of values.
[0048] As used herein, the terms "comprising" (and any form or variant of comprising, such as "comprise" and "comprises"), "having" (and any form or variant of having, such as "have" and "has"), "including" (and any form or variant of including, such as "includes" and "include"), or "containing" (and any form or variant of containing, such as "contains" and "contain"), are inclusive or open-ended and do not exclude additional, un-recited additives, components, integers, elements, or method steps.
[0049] As used herein, ordinal terms such as "first", "second", "third", etc., in the claims to modify a claim element does not by itself connote any priority, precedence, or order of one claim element over another or the temporal order in which acts of a method are performed, but are used merely as labels to distinguish one claim element having a certain name from another element having a same name (but for use of the ordinal term) to distinguish the claim elements. Similarly, use of a), b), etc., or i), ii), etc. does not by itself connote any priority, precedence, or order of steps in the claims. Similarly, the use of these terms in the specification does not by itself connote any required priority, precedence, or order.Attorney Docket No. 71951-716.602EXAMPLESExample 1 — Deep learning framework designed to predict retinal age from fundus images with high precision
[0050] Reproductive aging impacts women's health through fertility decline, disease susceptibility, and systemic aging. This example explores the retinal age gap-the difference between predicted retinal age and chronological age-as a biomarker for reproductive aging. By developing a Swin-Transformer-based dual -channel transfer learning model with data from healthy women, associations between the retinal age gap and Anti-Mullerian Hormone (AMH), a key marker of ovarian reserve, were examined. Findings revealed a negative association between the retinal age gap and AMH levels, particularly among women aged 40-50. Lower AMH levels correlated with earlier reproductive aging milestones, emphasizing the predictive value of retinal aging. Genetic data from genome-wide association studies further supported these associations and enhanced AMH prediction through multimodal modeling. These findings highlight the retinal age gap as a promising, non-invasive biomarker for reproductive aging and its potential role in disease prediction and personalized health interventions in women.
[0051] Cohorts and Subjects Characteristics
[0052] Two cohorts were utilized to develop and validate retinal age prediction models and investigate the relationship between retinal age gap and reproductive aging. The Healthy Cohort comprised 20,139 fundus images from 10,013 subjects (55.8% female), with ages ranging from 19.11 to 85.30 years (mean ± SD, 38.00 ± 4.76 years; see FIG.2A). This cohort was used to construct a generalized retinal age prediction model, capturing the normal aging process across a broad population.
[0053] The Healthy Female Cohort, consisting of 2,560 fundus images from 1,294 females aged between 25.01 and 47.14 years, was specifically designed to explore the link between retinal aging and reproductive aging. Patients were categorized into 3 groups based on AMH levels: high (n = 320, median = 4.66 ng / mL, IQR: 3.38-6.06), medium (n = 654, median = 1.96 ng / mL, IQR: 1.22-2.88), and low (n = 320, median = 0.58 ng / mL, IQR: 0.20-1.02), with no significant age differences between groups ( = 0.912) (Table 1 and FIG.2B). The retinal age model was then trained on the medium AMH group to capture typical female aging patterns, avoiding the extremes represented by high or low AMH levels. The low AMH group exhibited significantly lower mean age at first childbirth (26.79 ± 2.87) and first pregnancy (26.36 ± 2.71) compared to the high and medium AMH groups (P = 0.003) (Table 1).Attorney Docket No. 71951 -716.602Attorney Docket No. 71951-716.602
[0054] Anti -Mullerian Hormone (AMH), which is not normally distributed, is described by the median and interquartile range (IQR). Body Mass Index (BMI), Systolic Blood Pressure (SBP), Diastolic Blood Pressure (DBP), Fasting Plasma Glucose (FPG) and other normally distributed variables are expressed as mean (SD). Smoking, drinking, and marriage statuses are presented as counts and percentages (n, %). P- value from ANOVA or X2 is used to compare females of different AMH levels, with statistical significance indicated by P < 0.05.
[0055] Retinal Image-Based Age Prediction Using the FLEX Model
[0056] To achieve precise retinal age prediction, fundus image features from the Healthy Cohort were used by the model to learn optimal weights that minimize the difference between predicted retinal age and true chronological age. The model demonstrated strong predictive performance, with a Pearson's Correlation Coefficient (PCC) of 0.96, a Mean Absolute Error (MAE) of 2.69 years, a Coefficient of Determination (R2) of 0.91, and an Intraclass Correlation Coefficient (ICC) of 0.96 (FIG. 2C, FIG. 2D).
[0057] The final FLEX model was established by fine-tuning the initial weights using the medium AMH group data from the Healthy Female Cohort, allowing it to capture the unique characteristics of female aging. The FLEX model achieved robust predictive performance, with a PCC of 0.83, an MAE of 2.21 years, an R2of 0.68, and an ICC of 0.83, and revealed that the predicted retinal age often deviates from the true chronological age, highlighting individual variations (FIG. 2E, FIG. 2F).
[0058] To further validate the robustness and generalizability of the FLEX model, its predictive performance in cross-eye and retrospective predictions was evaluated. In the spatial consistency validation experiment, each image from one eye was paired with the retinal age predicted by the FLEX model from the contralateral eye of the same individual, using retinal images from 1,243 women. The model achieved a PCC of 0.90 and an MAE of 1.60 years when predicting the right eye based on the left, and a PCC of 0.88 and an MAE of 1.63 years when predicting the left eye based on the right. Additionally, the model's ability to forecast retinal age based on fundus images taken two years prior (623 fundus images from 317 women taken in 2019) was confirmed, achieving a PCC of 0.86 and an MAE of 1.83 years. These results demonstrate FLEX model's ability to produce consistent retinal age predictions across both temporally and spatially different sources.
[0059] The retinal age gap is associated with AMH levels
[0060] The FLEX model was used to predict retinal age in women from high and low AMH groups to explore the relationship between retinal age gap and AMH levels. While mean retinal age was similar across groups, the low AMH group showed a significantly higher retinal age gap (meanAttorney Docket No. 71951-716.602= 0.71 ± 2.50 years) compared to the high AMH group (mean = 0.36 ± 2.77 years, P 0.021) (FIG. 3A).
[0061] Age subgroup analysis revealed significant differences in the 40-50 years range, with the high AMH group showing a significantly smaller retinal age gap compared to the low AMH group of the same age group. A significant negative correlation emerged in women aged 40-50, particularly in the 40-44 years (R = -0.098, P = 0.019) and 45-50 years (R = -0.250, P = 0.002) age brackets (FIGS. 3B-3F), suggesting an inverse relationship between retinal age gap and AMH levels in women aged 40-50. However, no significant correlation between retinal age gap and AMH levels was found in the 25-39 age group. (P = 0.599 to P = 0.969) (FIGS. 3A-3D).
[0062] Retinal age gap and reproductive events predict AMH levels
[0063] Age-specific associations between retinal age gap and AMH levels were observed in women aged 40-44, where each additional year in the retinal age gap was linked to a 12% increase in the odds of low AMH levels (OR= 1.12, 95% CI: 1.02-1.22, P = 0.018) (FIG. 4A). This association was even more pronounced in women aged 45-50, with a 20% higher likelihood of low AMH levels for each additional year in the retinal age gap (OR = 1.20, 95% CI: 1.02-1.44, P = 0.038). No significant associations were found in women aged 25-39 (P > 0.05).
[0064] Retinal age gap and reproductive events were examined independently. The analysis showed that accelerated retinal aging, as indicated by a more positive retinal age gap, was associated with a 5% increased risk of having low AMH levels (OR = 1.05, 95% CI: 1.01-1.10, P = 0.022; FIG. 4B). This association remained significant after adjusting for potential confounders (FIG. 4C). Additionally, having a greater number of childbirths and being at younger ages during the first pregnancy and childbirth is also positively associated with having lower AMH level than average (FIG. 4B), and these trends were still observed while adjusting for other confounding factors (FIG. 4C). Notably, the number of pregnancies did not show any significant correlation with AMH levels whether alone or considering other variables.
[0065] Reproductive factors significantly influence AMH levels, particularly in women aged 30-39. For women aged 35-39, each additional childbirth increases the risk of having a low AMH level by 51%. Delaying first childbirth and first pregnancy by one year was associated with a 17% and 16% reduction, respectively, in the risk of low AMH levels among women aged 30-34. Among women aged 35-39, the corresponding reductions were 10% and 9%, respectively. The number of pregnancies did not show a significant association with AMH levels among any age groups (P > 0.05). Notably, the impact of reproductive factors on AMH level is most significant at the 30-39 range, indicating a critical period where reproductive choices may significantly affect ovarian reserve.Attorney Docket No. 71951-716.602
[0066] Sexual Hormones and Time of Menopause are Associated with Retinal Age Gap
[0067] Sex hormone levels in 71 females (140 eyes) were assessed, focusing on Follicle-Stimulating Hormone (FSH), Luteinizing Hormone (LH), progesterone (P), estradiol (E2), testosterone (T), and prolactin (PRL). For each hormone, the data were stratified into quartiles, with the top and bottom 25% classified as high and low hormone levels, respectively. Higher levels of FSH and PRL were significantly associated with higher retinal age gaps, particularly in women aged 40-50. However, there were no significant correlations between retinal age gap and E2, LH, P, and T levels.
[0068] Additionally, among 403 postmenopausal women (773 eyes), with a mean menopause age of 49.99 years (SD = 3.00), age of menopause was found to be negatively correlated with retinal age gap (R = -0.17, P < 0.001), where women who experienced menopause later had smaller retinal age gaps (P < 0.001). Each additional year in the retinal age gap increases the risk of experiencing menopause before age 45 by 36% (P < 0.001) and increases the risk of experiencing menopause between 45-49 by 12% (P = 0.004), relative to those experiencing menopause at age 55 or older. When considering other variables with retinal age gap, it was found that each additional year of the retinal age gap increased the risk of experiencing menopause before the age of 45 by 22% (OR = 1.22, P = 0.025).
[0069] Model explainability
[0070] By applying a Visualizer to identify salient features influencing the model's output, vascular regions in the fundus images were found to be most prominent in age estimation. As predicted age increased, the model's attention shifted from the nasal vasculature around the optic disc to the temporal vasculature, with a more refined focus concentrated on vascular regions, a pattern most noticeable in the 40-50 age group (FIG. 5).
[0071] Genome-wide Association Study for reproductive aging related retinal age gap
[0072] There are likely shared genetic factors between retinal and reproductive aging, knowing that genetic factors contribute to reproductive aging. The genome-wide association study (GW AS) on retinal age gap has provided several important insights into the genetic and biological factors that influence retinal aging. A total of 4,000 single nucleotide polymorphisms (SNPs) were identified as significantly associated with retinal age variation, highlighting the genetic basis of individual differences in retinal aging. Gene Ontology (GO) analysis pinpointed 38 key GOs, mostly of the cellular components (CC) domain, with enriched GO terms such as synaptic membrane, postsynaptic membrane, ion channel complex, cation channel complex, and extrinsic component of the plasma membrane, among others (Adjusted / 5< 0.05) (FIGS. 6A-6C). These findings,Attorney Docket No. 71951-716.602particularly enrichment of functions in neuronal signaling, suggest a crucial role of neuronal signaling in regulating retinal aging (FIG. 6B).
[0073] Among genes of interest in retinal aging, 14 pathways were significantly enriched.Notably, pathways related to reproductive and hormonal functions, such as progesterone-mediated oocyte maturation and oxytocin signaling, were identified, suggesting shared genetic factors among reproductive and retinal aging. Additionally, several cardiovascular pathways were found, including Arrhythmogenic right ventricular cardiomyopathy, Adrenergic signaling in cardiomyocytes, Renin secretion, and Platelet activation. Neuronal signaling pathways such as circadian entrainment, glutamatergic synapse, and dopaminergic synapse were also identified. These results collectively emphasize the intricate relationships between retinal aging, reproductive aging, cardiovascular health, and neuronal signaling, suggesting that genetic factors influencing retinal aging may have broader implications for overall health and aging processes (FIG. 6C).
[0074] Enhanced Prediction of AMH Levels Through Integration of Retinal and Genomic Data
[0075] Given that multimodal approaches can improve predictive accuracy, both fundus images and genetic data were leveraged to predict AMH levels as a proxy for ovarian reserve. The focus was on predictions in women aged 40-50 due to this period being a critical transition phase in reproductive health, often leading up to menopause, making it an ideal group to explore the relationship between retinal aging, AMH levels, and reproductive health. Using data from a cohort of 176 women (352 fundus images), two regression models were developed, one based solely on fundus images and the other integrating fundus images with SNP data (fusion model). While both achieved satisfactory performance in predicting AMH level, the fusion model did show a modest improvement (PCC = 0.78, MAE = 0.27 ng / mL, R2= 0.52). These results highlight the potential of retinal images, especially when combined with genetic data, in providing reliable predictions of AMH levels and offering clinically meaningful fertility guidance for women in this age group.
[0076] In this example, the potential of retinal age, as quantified by the retinal age gap, was investigated as a non-invasive marker of reproductive aging. By analyzing fundus images of healthy women, the example demonstrates that the retinal age gap is closely associated with key indicators of reproductive health, including AMH levels and hormonal factors. A significant association was found between lower AMH levels and an increased retinal age gap, indicating accelerated retinal aging, particularly in women aged 40-50. This age range is crucial for reproductive health, as AMH levels typically begin to decline in the years leading up to menopause. With the progress of reproductive aging, AMH levels decline, leading to a reduction in sex hormones like estrogen and progesterone, which are known to exert protection over theAttorney Docket No. 71951-716.602retina. This decline in protective hormones may accelerate retinal aging, creating the link between reproductive aging and retinal health. Additionally, women with lower AMH levels also showed a higher frequency of earlier reproductive events, such as childbirth. These results suggest that the retinal age gap may serve as a valuable non-invasive marker for reproductive aging, offering insight into the physiological processes that underpin ovarian reserve and fertility.
[0077] Further analysis revealed that the retinal age gap is not only linked to AMH but also to other hormonal markers, including FSEI and PRL. Elevated FSEI and PRL levels, which have been previously described to be linked to ovarian aging, were also found to accelerate retinal aging, increasing the retinal age gap, particularly in women aged 40-50 years. While links between retinal aging and other hormones related to ovarian aging, such as E2 and LH, were not observed, this absence can be attributed to the influences of the menstrual cycle on these hormones’ levels, which could be an important variable to consider in future research. In contrast, AMH proves to be a more stable marker for assessing reproductive aging, as its levels remain almost consistent throughout the menstrual cycle.
[0078] Notably, the smaller retinal age gap correlated with later menopause, indicative of prolonged reproductive function, further supporting its potential role as an indicator of reproductive aging. These findings align with previous studies that have shown that retinal aging could be an external reflection of internal, reproductive processes. The connection between retinal age and hormonal factors underscores the broader potential of retinal imaging in assessing age-related physiological changes.
[0079] Reproductive timing and the number of childbirths were also associated with AMH levels, with younger reproductive ages and higher childbirth frequency linked to lower AMH. However, prior studies on this relationship have been controversial. In younger women, follicle loss occurs through atresia, but the depletion becomes nonlinear around ages 37.5-38, making reproductive events more impactful in this age range due to accelerated follicle loss. Additionally, earlier and more frequent childbearing could have a cumulative effect on ovarian reserve, potentially leading to lower AMH levels.
[0080] The genetic underpinnings of retinal aging were also explored, identifying significant associations between retinal age variation and genetic factors involved in reproductive senescence. Genetic pathways related to progesterone-mediated oocyte maturation and oxytocin signaling were enriched in the analysis. Both pathways are critical components of reproductive function, highlighting the potential genetic link between retinal and reproductive aging. Furthermore, pathways associated with cardiovascular health, such as adrenergic signaling and renin secretion, were also implicated. Given that cardiovascular, retinal, and ovarian health are all interconnectedAttorney Docket No. 71951-716.602through vascular structures, the decline in cardiovascular health can have a systemic impact, impairing blood supply to the retina and ovary and subsequently accelerating their aging.
[0081] Salient mapping further showed that the model focuses on blood vessels around the optic disc, which are areas sensitive to systemic and female aging. Genetic investigations also identified genes related to reproductive conditions, such as premature ovarian insufficiency (POI) and polycystic ovary syndrome (PCOS), to be associated with retinal age gap. While out of the current study's scope, these findings may offer insights for future therapeutic interventions aimed at delaying reproductive decline.
[0082] The potential clinical applications of this study are significant. The retinal age gap could serve as a useful, non-invasive biomarker for reproductive aging, offering insights into a woman's ovarian reserve and overall reproductive health without the need for invasive procedures. Such a tool could be valuable for fertility counseling, deci si on -making regarding reproductive treatments, and monitoring during the menopausal transition.
[0083] The model to predict retinal age gap, as measured through non-invasive fundus imaging, can be used as a proxy for AMH levels and ovarian reserve. The findings highlight the potential of retinal imaging as a tool for assessing reproductive aging, offering a promising direction for early intervention and personalized care in clinical and community settings. Ultimately, the integration of Al, retinal imaging, and genetic data holds great promise for advancing the understanding of reproductive aging and improving women's health care outcomes.
[0084] Study Design and Participants
[0085] A cross-sectional study was conducted using data from two cohorts within the China Bioage Investigation Consortium, the Healthy Cohort and the Healthy Female Cohort (FIG. 1).The Healthy Cohort, derived from the China Optimal Health Cohort Study (COACS), included participants from northern and southern China for retinal age model development (FIG. 1, panel a). The Healthy Female Cohort comprised premenopausal women aged 25-50 years for reproductive health assessments (FIG. 1, panel b).
[0086] Inclusion criteria for the Healthy Cohort were general good health and absence of serious systemic diseases, nor gynecological diseases, surgical menopause, or hormone therapy. The Healthy Female Cohort excluded participants with gynecological diseases, surgical menopause, hormone therapy, or major systemic conditions like diabetes. More specifically, individuals with fasting blood glucose levels (FPG) > 7.0 mmol / L, systolic blood pressure (SBP) > 140 mmHg, or diastolic blood pressure (DBP) > 90 mmHg, were excluded. These thresholds were applied to minimize the confounding effects of systemic vascular conditions on retinal aging. Participants were randomly assigned to training (70%), internal validation (20%), and testing (10%) sets.Attorney Docket No. 71951-716.602
[0087] Ophthalmic Assessments
[0088] At least two ophthalmologists performed comprehensive eye examinations, including best-corrected visual acuity testing and refractive status assessment using an autorefractometer (KR800; Topcon; Tokyo, Japan) for all participants, excluding patients with ocular disease or vision loss. 45° nonmydriatic fundus images (CR2AF; Canon; Tokyo, Japan) were then captured and manually reviewed, excluding low-quality images from analysis. Chronological age was calculated as the difference between birth date and image capture date.
[0089] Retinal Age Gap Definition
[0090] A deep learning model was developed to predict retinal age from fundus images. The retinal age gap was defined as the difference between chronological age and predicted retinal age, calculated on a per-image basis. A higher gap indicates greater physiological aging.
[0091] Reproductive health assessments and AMH stratification
[0092] In the Healthy Female Cohort, AMH levels were measured and gynecological evaluations were conducted, including uterine ultrasound and cervical cytology. A subset of participants underwent sex hormone tests (FSH, LH, P, E2, T, PRL) and whole-genome sequencing for SNP analysis. A group of healthy postmenopausal women, defined as those reporting cessation of menstruation for over one year (excluding those with menopause onset before age 35 or after age 60), was also included.
[0093] Participants were stratified into age-specific AMH quartiles: the lowest quartile as the low AMH group (high reproductive aging), the middle two quartiles as the medium AMH group, and the highest quartile as the high AMH group (low reproductive aging). This quartile-based stratification strategy aligns with previous studies on reproductive aging, in which AMH levels were commonly divided into quartiles to capture population-level variations in ovarian reserve and facilitate comparisons between groups with distinct reproductive profiles. To capture typical aging patterns and reduce bias from extreme values, the FLEX model was fine-tuned using data from the medium AMH group. The low and high AMH groups were then compared to examine the relationship between retinal age gap and reproductive aging.
[0094] Deep Learning Model for Age Prediction in Healthy Cohort
[0095] To extract age-related features from retinal images, the Healthy Cohort Retinal Age Model Weights were developed by training a deep learning model on data from the Healthy Cohort. These model weights were then applied to analyze fundus images from the Healthy Female Cohort. The model was built using a Swin Transformer architecture for age regression, specifically with the Swin-Base configuration, which includes a patch size of 4, a window size of 12, and an inputAttorney Docket No. 71951-716.602resolution of 384x384 (swin_base_patch4_windowl2_384). Pretrained weights from the Japan Ocular Imaging Registry's JOI model (joir.jp / data / en_index) were used to initialize the network.
[0096] Training occurred over 100 epochs, starting with a learning rate of 0.75e-4, which was dynamically adjusted using StepLR. The AdamW optimizer was used with a weight decay of 1 ,5e-3, and a batch size of 8. To avoid overfitting, early stopping was implemented if the validation loss did not improve after 10 epochs. Stratified sampling was employed to ensure the dataset reflected the overall age distribution, dividing participants into 10-year age intervals.
[0097] Modified Model FLEX for Age Prediction in Healthy Female Cohort
[0098] The FLEX model were trained on retinal images from the medium AMH level group, indicating nonextreme reproductive aging, to estimate retinal age and analyze its trends in different AMH levels. The FLEX network features a Dual -Channel Transfer Learning architecture, using two parallel Swin Transformer encoders. One encoder's parameters were kept frozen, while the other was allowed to leam from the new data, preserving the original model's feature extraction while updating it to suit the Healthy Female Cohort's images. The outputs of the two encoders were fused and processed through normalization and adaptive pooling to predict retinal age.
[0099] To mitigate overfitting, training was limited to 5 epochs, with data augmentation techniques, such as flipping and random rotation. The training process also included age-stratified sampling at 5-year intervals. The final model weights, based on the lowest validation loss, were used to predict retinal age for participants in the low and high AMH groups.
[0100] Two validation experiments were conducted to assess the consistency of retinal age predictions: (1) Spatial consistency was evaluated by training the model on paired data, where each retinal image from one eye was matched with the predicted retinal age of the contralateral eye (using the left eye image paired with the right eye's predicted age, and vice versa); (2) Temporal consistency was assessed by pairing retinal images acquired in 2019 with the retinal ages predicted by the FLEX model from the corresponding images collected in 2021 for the same individuals. This design allowed us to evaluate the stability of retinal age predictions over time. For both experiments, the datasets were partitioned into training, validation, and test sets (7:2: 1), following the same procedure as the main analysis, with no subject-level overlap.
[0101] Explainability of model predictions
[0102] To enhance the explainability of the retinal age predictions, the Visualizer tool (github.com / luo3300612 / Vi sualizer) was utilized, which visualizes the attention modules of deep learning models. The tool provided local variable insights from the model without altering its internal code. Specifically, attention feature maps from the Swin Transformer were intercepted atAttorney Docket No. 71951-716.602each layer using forward hooks, and visualization was performed on the final attention block of both encoder networks in the FLEX model.
[0103] Preprocessing SNP Data in the Healthy Female Cohort
[0104] Genetic analysis was conducted on 928 female participants from the Healthy Female Cohort who completed whole-genome sequencing. This reduced the sample size from an initial 1,294 participants due to exclusions for incomplete genetic data or poor sample quality. PLINK 1.90b6.24 (June 6, 2021) was used for SNP quality control, removing variants with missing genotype rates over 2%, minor allele frequencies under 1%, or those that did not pass the Hardy -Weinberg equilibrium test (p-value < le-5). After filtering, 8,089,940 variants were retained for further analysis.
[0105] Extraction of Gene Features in the Healthy Female Cohort
[0106] Given the large number of SNPs called, traditional false discovery rate and Bonferroni correction methods were considered too conservative, potentially leading to a low true positive rate. Therefore, a combination of GWAS and deep learning was used for better identification of genes related to the retinal age gap.
[0107] For extraction of gene features, GWAS was conducted for preliminary SNP screening. Considering that genetic influences affect the body systemically and are shared across both eyes, the association analysis using the average retinal age gap derived from both eyes was performed. Linear regression models were applied using PLINK to analyze the retinal age gap. The study focused on Han Chinese women from the same geographic region with similar environmental backgrounds, assuming minimal population structure differences among samples, thus not performing principal component analysis (PCA) to adjust for population structure. SNPs with p-values less than le-3 were selected for further feature extraction. The LassoNet model was employed for further screening of these SNPs, which integrates deep learning with Lasso regression (LI regularization) to facilitate automatic feature selection within a neural network. In this model, letR* represent a dense feed-forward neural network parameterized by weights W. For any input ®and parameter 8 e the LassoNet model can be expressed as equation (1):FFNW(x) Q)
[0108] This formulation can be seen as a residual neural network with an additional skip connection. To optimize the model, the following problem is solved using projected proximal gradient descent, as shown in equation (2):Attorney Docket No. 71951-716.602mini (0, IE) isub “ject ioH " IV 3;1' h "oc < MH ii0JL for all} J ~ 1..2.... ,d (2)
[0109] Here,denotes the LI norm of 0, enforcing sparsity. The optimization process involves incrementally increasing A by a factor of 1 + c until the first-layer weightsreduce to zero. This creates a path of solutions, enabling the selection of an optimal feature set based on the desired sparsity level. The significance of each feature is determined by the point at which its corresponding weight Hj(1)becomes zero as increases. Features are chosen by identifying those that are the last to be zeroed out. To enhance model performance, hidden layers with sizes 1200, 800, and 200 were configured, increasing the complexity of the neural network. And using default parameters M = 10 and e = 0.02. Missing data was filled with the SNPs median, and data standardization was applied to ensure consistent scaling of features. The dataset was divided into 742 samples for training and 186 samples for testing. By default, 10% of the training data is automatically allocated for validation to monitor performance and prevent overfitting. The top 4000 SNPs were selected based on their importance and better model fit (lower and faster-changing MAE) for further analysis.
[0110] Gene Annotation and Enrichment Analyses[OHl] The selected SNPs underwent gene annotation, which was performed using the online Variant Effect Predictor (VEP). To explore the biological significance of the gene modules, GO and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were conducted using the clusterProfiler package in R. P-values were adjusted using the Benjamini-Hochberg algorithm, with a threshold of P < 0.05 considered statistically significant. The results were visualized using bar plots and dot plots.
[0112] AMH Concentration Prediction Using Integrated Retinal and Genetic Data
[0113] Based on the observed relationship between retinal images and AMH levels in women aged 40-50, the SNPs-Attention Swin Transformer (SNPs-AST) model was further developed to predict AMH concentrations by integrating retinal images and genetic data from this population. This model was compared to a Swin Transformer-based model that used only image data for AMH prediction.
[0114] The model for image-based prediction relied on the Swin Transformer architecture, initialized with pre-trained weights from the Healthy Cohort Retinal Age Model weights. These weights had already learned image features related to eye aging, which could potentially enhance the prediction of AMH concentrations. The final layer is configured as a fully connected (FC) layer without additional weights. All model parameters were randomly initialized andAttorney Docket No. 71951-716.602standardized to ensure uniform training conditions. Other training parameters remain consistent with previous configurations.
[0115] For the gene-image fusion model, an improved version model (SNPs-AST) was developed based on the co-attention technique. The framework consists of three primary components: an image feature embedding encoder, an SNP-specific encoder, and a cross-modal fusion mechanism utilizing a multi-head attention mechanism for integrating images and SNPs.
[0116] The image feature embedding encoder employs the Swin Transformer encoder, initialized with the pre-trained weights from the Healthy Cohort Retinal Age Model. For the SNP-specific encoder, a dedicated network was designed which comprises a simple multi-layer perceptron (two FC layers with ReLU activation functions) to embed the input SNPs into the same dimensional space as the image features. The input SNP features include SNP variation (0, 1, 2), chromosome location (1, 2, 3, etc.), and mutation type, with mutation types recoded according to their frequency (e.g., C / T is 1, G / A is 2). These three gene features are combined as the gene information input for the model.
[0117] Finally, a multi-head attention mechanism was custom-designed that leverages image features as query vectors and SNP embedded features as keys and values. This mechanism allows the model to focus on relevant parts of the input sequence (SNP features) when processing the image features. The attention function can be described by the following equation (3):Attention ( Q, K, V) = softmax j — - — | V\ v<4 / Q)
[0118] In this context, Q represents the image feature vectors, K and V represent the SNPs embedded feature vectors, with dkdenoting the dimensionality of the key vectors. The dot product QK computes the similarity between image and SNPs features, producing the attention scores, which are then scaled by to mitigate the effects of large dot product values. The softmax function is applied to obtain the attention weights, which are subsequently used to compute a weighted sum of the SNPs feature vectors, yielding the final attention output.
[0119] The attention scores calculated from this mechanism dynamically filter SNPs potential features relevant to AMH. A residual connection mechanism was then employed to add the image features to the SNP potential features, followed by an FC layer acting as a regressor to predict AMH concentrations.
[0120] Statistical Analysis
[0121] Descriptive statistical analysis was conducted using %2tests or Fisher's exact tests for categorical variables, and Student's t-tests or Mann-Whitney U-tests for continuous variables, depending on data distribution. Age then was segmented into 5-year intervals, with t-testsAttorney Docket No. 71951-716.602performed to compare retinal age gap differences across AMH groups within each interval.Additionally, Pearson's correlation was used to assess the relationship between log-transformed AMH (to account for its non-normal distribution) and the retinal age gap. Univariate and multivariate logistic regression analyses were conducted using a generalized linear model (GLM) to examine the influence of the retinal age gap and other fertility -related factors on outcomes within different AMH categories. The high AMH group was used as a reference to evaluate the odds ratios (OR), 95% confidence intervals, and / ?-values for the low AMH group. In all multivariate logistic regression analyses, the covariates adjusted in the model included age, eye laterality, BMI, smoking status, alcohol consumption, and marital status. To maximize statistical power, both eyes were included in the analysis. Given the potential for inter-eye differences, eye laterality was explicitly adjusted for in the regression models to account for intra-individual correlation. The inclusion of BMI, smoking, alcohol consumption, and marital status was justified, as these clinical factors may influence either AMH levels or retinal health, making them relevant to the study hypothesis. Additionally, the parameters were also analyzed through the univariate GLM segmented by different age groups, without adjustments due to the small amount of data for each age group.
[0122] Participants from the Healthy Female Cohort who underwent additional sex hormone testing were also assessed. These participants were divided into quartiles for each hormone based on 10-year age intervals. T-tests were then used to compare the retinal age gap between the highest and lowest quartile groups, offering further insights into the relationship between the retinal age gap and reproductive aging.
[0123] The menopausal timing of postmenopausal participants was analyzed. Initially, the association between menopausal timing and the retinal age gap was examined. The menopausal age was then divided into groups above and below the average and compared the retinal age gap of each group using t-tests. Additionally, the menopausal age was categorized into intervals of <45, 45-49, 50-54, and >55 years. Using the >55 years group as a reference, the association between the retinal age gap and menopausal timing was examined using univariate and multivariate logistic regression analyses. Since all individuals in the study have been married, age, eye laterality, BMI, smoking status, and alcohol consumption were adjusted for as covariates.
[0124] Model performance was assessed using PCC, MAE, R2, Bland-Altman plots, and ICC to evaluate agreement between predicted and actual values. All analyses were performed using R Studio (version 4.1.2), with model training and validation in Python (version 3.7.15). A / ?-value < 0.05 was considered statistically significant.Attorney Docket No. 71951-716.602
[0125] It should be understood from the foregoing that, while particular implementations of the disclosed methods and systems have been illustrated and described, various modifications can be made thereto and are contemplated herein. It is also not intended that the present disclosure be limited by the specific examples provided within the specification. While the present disclosure has been described with reference to the aforementioned specification, the descriptions and illustrations of the preferable embodiments herein are not meant to be construed in a limiting sense. Furthermore, it shall be understood that all aspects of the present disclosure are not limited to the specific depictions, configurations or relative proportions set forth herein which depend upon a variety of conditions and variables. Various modifications in form and detail of the embodiments of the present disclosure will be apparent to a person skilled in the art. It is therefore contemplated that the present disclosure shall also cover any such modifications, variations and equivalents.
[0126] Although the present disclosure has been described with reference to the presently preferred embodiment, it should be understood that various modifications can be made without departing from the spirit of the present disclosure.
Claims
Attorney Docket No. 71951-716.602CLAIMSWhat Is Claimed Is:
1. A method, comprising:(a) obtaining retinal fundus images of a subject;(b) determining a retinal age of the subject based on the retinal fundus images, thereby determining a retinal age gap, which is the difference between the retinal age and the subject’s chronological age; and(c) determining, from the retinal age gap, a state of reproductive aging in the subject.
2. The method of claim 1, comprising obtaining the retinal fundus images using non-invasive fundus imaging.
3. The method of claim 1 or claim 2, wherein one or more of the retinal fundus images comprise microvascular features.
4. The method of any one of claims 1-3, further comprising predicting the retinal age of the subject using a model trained on fundus image features from a healthy cohort of subjects.
5. The method of claim 4, wherein the model is trained to learn optimal weights of the fundus image features that minimize the difference between a predicted retinal age and its corresponding chronological age.
6. The method of claim 4 or claim 5, wherein the model is trained using transfer learning from the fundus image features from the healthy cohort of subjects.
7. The method of any one of claims 4-6, wherein the model is trained using supervised finetuning based on fundus image features from a medium AMH (Anti-Mullerian Hormone) group in the healthy cohort of subjects.
8. The method of claim 7, wherein the medium AMH group comprises the middle two quartiles of the healthy cohort which is stratified by AMH levels into quartiles.
9. The method of any one of claims 4-8, wherein the model comprises a dual-channel transfer learning architecture.
10. The method of any one of claims 4-9, wherein the model comprises two parallel Swin Transformer encoders.
11. The method of any one of claims 4-10, wherein the model comprises (i) a first encoder, the parameters of which are kept frozen, (ii) a second encoder, the parameters of which are allowed to learn from new data, preserving the model's feature extraction while updating it to suit the fundus images from the healthy cohort of subjects.Attorney Docket No. 71951-716.60212. The method of claim 11, further comprising fusing outputs of the first encoder and the second encoder.
13. The method of claim 12, further comprising processing the fused outputs through normalization and adaptive pooling to predict the retinal age.
14. The method of any one of claims 1-13, wherein the retinal age gap is a positive reginal age gap indicating accelerated retinal aging.
15. The method of any one of claims 1-14, wherein the retinal age gap is correlated with an AMH (Anti -Mullerian Hormone) level in the subject.
16. The method of any one of claims 1-15, further comprising predicting the subject has a low AMH (Anti -Mullerian Hormone) level based on a more positive retinal age gap.
17. The method of any one of claims 1-16, further comprising predicting the subject has accelerated reproductive aging based on a more positive retinal age gap.
18. The method of any one of claims 1-17, further comprising predicting the subject has low ovarian reserve based on a more positive retinal age gap.
19. The method of any one of claims 1-18, further comprising predicting the subject has a higher level of a sexual hormone based on a more positive retinal age gap.
20. The method of claim 19, wherein the sexual hormone comprises follicle-stimulating hormone (FSH).
21. The method of claim 19 or claim 20, wherein the sexual hormone comprises prolactin (PRL).
22. The method of any one of claims 1-21, further comprising predicting earlier menopause based on a more positive retinal age gap.
23. The method of any one of claims 1-22, wherein the subject is a woman between 30 and 39.
24. The method of any one of claims 1-22, wherein the subject is a woman between 40 and 44.
25. The method of any one of claims 1-22, wherein the subject is a woman between 45 and 49.
26. The method of any one of claims 1-22, wherein the subject is a woman between 50 and 54.
27. The method of any one of claims 1-22, wherein the subject is a woman 55 or older.
28. A system comprising:at least one hardware processor; andone or more software modules configured to, when executed by the at least one hardware processor, perform the method of any one of claims 1-27.
29. A non-transitory computer-readable medium having instructions stored thereon, wherein the instructions, when executed by a processor, cause the processor to perform the method of any one of claims 1-27.Attorney Docket No. 71951-716.60230. A system comprising:at least one hardware processor;a non-transitory computer-readable medium coupled to at least one hardware processor, optionally wherein the coupling is over a network; andinstructions stored in the non-transitory computer-readable medium, wherein the instructions when implemented by the processor, configure the system to perform the method of any one of claims 1-27.
