Artificial Intelligence Methods for Screening Multiple Systemic Diseases Based on Retinal Imaging
By constructing the Reti-pioneer framework and combining retinal images and clinical metadata, the problem of insufficient generalization ability in the screening of endocrine and metabolic diseases in existing technologies has been solved. It enables unified screening and high-accuracy prediction of multiple diseases and is suitable for early screening of multi-ethnic populations.
Patent Information
- Application Number
- CN202611310266.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2026-04-10
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing endocrine and metabolic disease screening technologies based on retinal imaging suffer from insufficient generalization ability of single-disease models, lack of representativeness of multi-ethnic populations and multi-disease risk stratification capabilities, making it difficult to construct a unified framework and failing to meet the comprehensive, accurate and scalable early assessment needs of high-risk populations.
The Reti-pioneer framework is constructed, employing a multi-task learning approach. It combines retinal images and clinical metadata, utilizing the Swin Transformer, Vision Mamba, and RET Found modules for feature extraction and fusion. A quality-aware module processes images of varying quality, integrating multimodal data for screening and prediction of various endocrine and metabolic diseases.
The model's generalization ability under different resource environments has been improved, enabling unified screening of various endocrine and metabolic diseases. It has enhanced prediction accuracy and interpretability, demonstrating effectiveness and cross-regional applicability in multi-ethnic populations, and significantly improving the detection accuracy of T2DM, hypertension, and hyperlipidemia.
Smart Images

Figure CN123118519A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image analysis, and more specifically, to an artificial intelligence method for screening various systemic diseases based on retinal images. Background Technology
[0002] Against the backdrop of a global aging population, the incidence of endocrine and metabolic diseases, such as diabetes and thyroid diseases, is rising significantly, placing a heavy burden on healthcare systems worldwide. Early screening and intervention are crucial to curbing disease progression. Current mainstream screening methods heavily rely on blood-based biomarker testing. While these methods are highly accurate, they involve invasive blood collection, cold chain logistics, high testing costs, and poor patient compliance, severely limiting their feasibility for high-frequency, longitudinal monitoring in large populations, especially in remote areas with limited medical resources, where this supply-demand imbalance is even more pronounced. In this context, oculomics, an interdisciplinary technology that uses artificial intelligence (AI) to analyze ophthalmic images to monitor overall health, has emerged, offering a non-invasive solution to these challenges. Studies have shown that microscopic changes in the retinal blood vessels and nerve fiber layers often precede the clinical manifestations of endocrine diseases. By analyzing color fundus photographs (CFP) using deep learning, early predictions of metabolic diseases such as diabetes can be achieved. This is expected to break down the geographical and economic barriers of traditional screening, alleviate medical inequality, and enable early warning of endocrine and metabolic abnormalities through the analysis of retinal images.
[0003] While ocular omics and basic medical models offer novel approaches to disease screening, current technologies still face significant challenges when applied to endocrine and metabolic disease screening. First, most existing ocular omics studies operate within a "single disease, single model" framework, typically employing a train-from-scratch approach. This not only demands massive amounts of high-quality labeled data but also struggles to capture the complex multimorbidity patterns prevalent in reality, resulting in insufficient generalization ability when models face patients with multiple coexisting diseases. Second, while existing basic medical models aim to reduce data dependence and improve generalization, their application in the endocrine and metabolic disease field is limited, lacking in-depth optimization for this disease spectrum. Furthermore, the training data for existing models lack representativeness across ethnic groups and sufficient external validation in multi-disease risk stratification. These deficiencies prevent current systems from constructing a unified framework capable of simultaneously identifying and characterizing multiple endocrine and metabolic diseases, failing to meet the urgent clinical need for comprehensive, accurate, and scalable early assessment of high-risk populations. Summary of the Invention
[0004] The present invention aims to overcome at least one deficiency in the prior art and provide an artificial intelligence method for screening multiple systemic diseases based on retinal images, in order to solve the problems mentioned in the background art above.
[0005] The technical solution adopted in this invention is an artificial intelligence method for screening multiple systemic diseases based on retinal imaging, comprising the following steps: S1. Obtain macular CFP data from the crowd cohort and hospital cohort, construct a multimodal dataset, and divide the multimodal dataset into a training set, a validation set, and an external test set; S2. Construct the Reti-pioneer framework, a retina-based disease prediction pre-training model, for the simultaneous screening and prediction of six endocrine and metabolic diseases. S3. Validate and evaluate the generalization ability of the Reti-pioneer framework using external ethnic data; S4. Enhance the interpretability of the Reti-pioneer framework through saliency images; S5. Obtain potential retinal features and disease-specific plasma proteins, and verify the biological rationale of the Reti-pioneer framework by analyzing the relationship between potential retinal features and disease-specific plasma proteomics maps. S6. The ecological validity and workflow efficiency of the model were measured in a primary healthcare setting using a prospective silent trial.
[0006] By constructing the multi-task framework Reti-Pioneer and conducting biologically relevant, stepwise, multi-center clinical validation studies across diverse resource environments, it provides a unified screening for a range of endocrine and metabolic diseases.
[0007] To improve prediction accuracy, the Reti-pioneer framework includes a quality-aware module and three frozen pre-trained visual foundation modules. The quality-aware module is used to evaluate image quality; the three frozen pre-trained visual foundation modules are Swin Transformer, Vision Mamba, and RET Found, which are used to extract features from CFP images.
[0008] At the image level, the quality-aware module enables the preservation and effective use of low-quality CFPs, which may still reflect the overall endocrine state. The Reti-pioneer framework does not exclude these images but instead uses image quality metrics to guide model decisions in the early processing stages.
[0009] To achieve multimodal integration, the Reti-pioneer framework also includes a feature fusion module and a scaling exponential linear unit. The feature fusion module is used to fuse the extracted features, and the scaling exponential linear unit is used to scale the fused features.
[0010] Feature fusion is performed through a feature fusion module, integrating image quality and clinical features, including age, sex, race, and weight.
[0011] Preferably, the construction of step S2 based on the Reti-pioneer framework, a pre-trained model for retinal disease prediction, specifically involves: Image features are extracted using the Swin Transformer, Vision Mamba, and RET Found modules, extracting clinical features. A feature fusion module then fuses image quality and clinical features, including age, gender, ethnicity, and weight. A quality-aware module, in conjunction with the feature fusion module, utilizes bilinear operations to combine quality indicators with F-dimensional deep features, generating fused features of the same dimension. Quality indicators are categorized as "low," "medium," and "high," and their probability scores are calculated. A scaling exponential linear unit scales the fused features. The fused features are then combined into a linear layer for outputting disease probabilities. A SELU activation function is applied between intermediate layers to enhance training stability and maintain self-normalization properties. Fine-tuning is applied to various downstream tasks, including screening and prediction.
[0012] The quality-aware module includes trained modules that score image quality and categorize them into low, medium, and high quality labels. SELU activation and linear transformation maintain this dimension, ensuring consistency in feature representation. The generalization capabilities of the Swin Transformer, Vision Mamba, and RET Found models were systematically evaluated in independent external tests by combining bilateral CFP and clinical features.
[0013] Step S2, which uses the Reti-pioneer framework to screen for and predict six endocrine and metabolic diseases, specifically involves: The fundus image is input into the Reti-pioneer framework, where the Swing Transformer, Vision Mamba, and RETFound modules are integrated through weighted soft voting and fused by the model architecture to generate a single, unified prediction.
[0014] Since all model development and evaluation analysis units are individual participants, the evaluation is conducted on a per-subject basis, using fundus images from both eyes as input, and integrating the trained prediction head to generate a single, uniform prediction.
[0015] Preferably, the external ethnic data in step S300 includes data from a first dataset and a second dataset, wherein the first dataset was collected under dilated pupil conditions and the second dataset was collected under non-dilated pupil conditions.
[0016] The first dataset consists of medical institutions in regions such as Tibet, Xinjiang, and Guangxi in China, while the second dataset consists of physical examination centers and large tertiary hospitals in Guangdong, China.
[0017] To quantify the contribution of the core components in the Reti-pioneer framework, step S3, which involves evaluating the generalization ability of the Reti-pioneer framework using external ethnic data, specifically involves: Compare the complete Reti-Pioneer framework with the following: (1) An ablation version without a quality-aware module, trained on all images; (2) The same ablation version with a quality-aware module, specifically trained on high-quality images; (3) Using only fundus images or clinical metadata as a single-modal baseline to isolate performance gains attributable to the model’s full multimodal architecture; All experimental results were categorized according to disease type and visualized using calibration curves and decision analysis curves.
[0018] All images include those from the internal training and test sets. High-quality images are obtained by the quality-aware module scoring the images, with the scoring results including low-quality, medium-quality, and high-quality images.
[0019] The purpose of the above experiments was to isolate the performance gains attributed to the model-complete multimodal architecture.
[0020] Preferably, step S4, which enhances the interpretability of the Reti-pioneer framework through saliency images, specifically involves: Saliency mapping is performed to identify and visualize the analytical regions of interest that contribute most to the model's diagnostic and predictive decisions in the Gaussian-filtered fundus color image; an integral gradient algorithm is used to generate a visual explanation of the model's behavior, attributing predictive relevance to individual pixels; the saliency image shows the relative contribution of each pixel to the final prediction.
[0021] The visual representation of model behavior is the salient regions of the heatmap, which are generated at the pixel level to identify fundus regions that are crucial to model prediction.
[0022] To understand retinal features associated with endocrine and metabolic disorders, step S5 specifically involves obtaining potential retinal features and disease-specific plasma proteins: Based on plasma proteomic analysis data from a multimodal dataset, proteins were quantified and captured using the Olink Explore 3072 platform. Protein expression levels were reported as normalized protein expression on a $log_2$ scale. After quality control, proteins with missing data >30% and missing data >50% were excluded, and the remaining proteins were retained as disease-specific plasma proteins for analysis. Orthogonal partial least squares discriminant analysis was used to screen for potential retinal features. Elastic network regularization analysis was used to select potential retinal features to achieve dimensionality reduction and identify key predictors from the Reti-Pioneer framework. The priority of candidate proteins was determined based on their coefficients, and biologically relevant biomarkers were extracted for selection.
[0023] The 256-dimensional fundus color image features from the penultimate layer of the model were used to extract latent retinal features, which were then selected using orthogonal partial least squares discriminant analysis (OPLS-DA). Elastic network regularized regression was used to select the top 5 core proteins associated with each disease from all protein data, and the coefficients of the candidate proteins were obtained during the OPLS-DA operation. By linking retinal features with the biological basis of endocrine and metabolic pathophysiology, orthogonal partial least squares discriminant analysis (OPLS-DA) was used to derive predictive components for each disease. After adjusting for age and sex, except for osteoporosis, the predictive components showed a significant correlation with the diseases.
[0024] Step S5, which verifies the biological rationale of the Reti-pioneer framework by examining the relationship between latent retinal features and disease-specific plasma proteomics maps, specifically involves: Using retinal latent features derived from variational autoencoders, orthogonal partial least squares discriminant analysis (OPLS-DA) was performed on a model to assess the separation between the control and case groups. Predictive components were extracted, and systematic variances directly related to inter-group differences were captured. Logistic regression was used to test the impact of predictive components on the prevalence of corresponding diseases. Pearson correlation analysis was used to quantify the linear relationship between individual protein levels and corresponding predictive components. Logistic regression was used to assess the association between retinal latent features derived from OPLS-DA and the top-ranking plasma proteins in six endocrine and metabolic diseases, resulting in both unadjusted and adjusted models for age, sex, race, body mass index, image quality, and assessment center.
[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Improve the performance and generalization ability of the Reti-pioneer framework on both resource-constrained and resource-rich datasets. The Reti-Pioneer framework maintained robust performance on a combination of datasets from resource-limited regions (Tibet, Xinjiang, and Guangxi) and resource-rich environments. Its performance for T2DM, hyperlipidemia, and osteoporosis was significantly higher in mixed-resource environments compared to resource-limited environments alone. The model was validated in the Singapore Eye Disease Epidemiology Study (SEED), demonstrating its effectiveness across diverse ethnic populations and its superior performance in hypertension detection. The integration of the quality-aware module and multiple pre-trained base models further enhanced Reti-Pioneer's performance. Ablation studies showed that combining the quality-aware module significantly improved the AUROC for T2DM compared to using only high-quality images (P=0.011) or using full-quality images without the module (P=0.021). It also significantly improved the accuracy for hypertension (P<0.001) and hyperlipidemia (P=0.037) compared to using only high-quality images. Furthermore, the multimodal combination of retinal images and clinical metadata yielded superior predictive accuracy compared to any single modality input.
[0026] 2. Reti-Pioneer enables longitudinal risk stratification of endocrine and metabolic diseases over 5 and 10 years. The Reti-Pioneer framework was evaluated in a fully preserved longitudinal subset of the UK Biobank dataset to assess its performance in predicting six endocrine disorders at 5-year and 10-year intervals. Reti-Pioneer achieved an AUROC of 0.755 for 5-year onset of type 2 diabetes mellitus (T2DM), decreasing to 0.736 at 10-year prediction. For hypertension, the model achieved AUROCs of 0.755 and 0.719 at 5-year and 10-year intervals, respectively. Corresponding values for hyperlipidemia were 0.748 at 5 years and 0.735 at 10 years. In a sensitivity analysis excluding individuals diagnosed within the first year after fundus imaging, the model maintained comparable performance across both 5-year and 10-year prediction intervals. Attached Figure Description
[0027] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention. To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0028] Figure 1 This diagram provides an overview of the Reti-pioneer framework for multi-disease screening based on retinal imaging, as presented in this invention.
[0029] Figure 2This is a schematic diagram illustrating the external testing of the Reti-pioneer of the present invention under different medical resource environments.
[0030] Figure 3 This is a schematic diagram of a prospective silent trial evaluating the operational efficiency of Reti-Pioneer in primary healthcare according to the present invention.
[0031] Figure 4 This is a schematic diagram illustrating the multi-disease screening performance and acceptability of the Reti-Pioneer of the present invention in a preliminary study.
[0032] Figure 5 This is a schematic diagram of the overall architecture of the Reti-Pioneer framework of the present invention.
[0033] Figure 6 This diagram illustrates the performance of the Reti-Pioneer in different external validation datasets, along with decision curve analysis.
[0034] Figure 7 This diagram illustrates the performance of the Reti-Pioneer in different external validation datasets, along with calibration curve analysis.
[0035] Figure 8 This is a saliency map of patients with endocrine and metabolic diseases according to the present invention.
[0036] Figure 9 This is a schematic diagram showing the detailed performance indicators of the Reti-pioneer in the silent test of this invention.
[0037] Figure 10 The flowchart shows the population included in the construction and validation of the Reti-pioneer model of this invention.
[0038] Figure 11 This is a saliency map of patients with endocrine and metabolic diseases according to the present invention.
[0039] Figure 12 This is a schematic diagram illustrating the biological interpretability and cross-omics integration analysis of the Reti-Pioneer model of the present invention. Detailed Implementation
[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.
[0042] Example 1 This embodiment provides an AI-based method for screening multiple systemic diseases based on retinal imaging. It incorporates the multi-task framework Reti-Pioneer and conducts biologically relevant, stepwise, multi-center clinical validation studies to provide unified screening for a range of endocrine and metabolic diseases across diverse resource environments. An overview of the Reti-Pioneer workflow is as follows: Figure 1 As shown, where Figure 1 a represents the data acquisition process: obtaining color fundus images with relevant clinical metadata from population and hospital cohorts; Figure 1 b represents the model architecture: The quality-aware module is combined with three frozen pre-trained visual base models, Swin Transformer, Vision Mamba, and RET Found, for the simultaneous screening of six endocrine and metabolic diseases; Figure 1 c represents the model validation and interpretation process: assessing generalization ability through external multi-ethnic validation; examining biological interpretability by correlating latent retinal features with plasma proteomics and multi-gene risk score data; Figure 1 d represents a prospective clinical assessment: a silent trial measured ecological validity and workflow efficiency in primary care; a subsequent preliminary study evaluated diagnostic accuracy, clinical applicability, and acceptance by clinicians and patients.
[0043] The first step involved constructing a multimodal dataset containing CFP images of varying quality and pairing them with structured clinical metadata. A total of 107,730 CFP images from 53,865 individuals were used to construct a heterogeneous dataset covering different imaging conditions and clinical contexts. This data was sourced from the community-based UK Biobank and the Chinese Tertiary Hospital Registry.
[0044] The second step involved developing a multimodal learning model that integrates CFP images of varying quality with structured clinical metadata. This architecture integrates large-scale visual foundational models, including Swin Transformer, Vision Mamba, and RETFound, to leverage their complementary capabilities. Subsequently, the model underwent comprehensive external testing on multicenter datasets spanning diverse geographic regions.
[0045] The third step systematically evaluated the framework's generalization ability in disease prediction and assessed its interpretability by correlating latent retinal features with proteomics and genetic markers. Furthermore, its ability to improve ophthalmologist diagnostic performance was evaluated in a human-AI collaborative image reading study.
[0046] Finally, the model was deployed in a prospective silent trial, comparing its efficiency and time cost in routine primary care compared to traditional screening methods. Following successful validation, a preliminary clinical study was conducted to assess the feasibility and practical impact of the framework in clinical workflows. By demonstrating a scalable and adaptable AI framework, the findings fill a key gap in current screening paradigms and provide groundbreaking insights for the practical deployment of ocular omics-driven systemic health assessments.
[0047] Results 1. Reti-Pioneer's performance and generalization ability on datasets with both limited and abundant resources. The demographic and clinical characteristics of the participants in the training dataset are summarized in Table 1. The framework was fine-tuned on an internal dataset containing 107,730 CFPs from 53,865 participants and further validated on independent datasets from resource-limited and resource-rich regions in China and a multi-ethnic cohort in Singapore. These external validation sets contain a total of 23,232 CFPs from 11,616 participants.
[0048] At the image level, image quality is evaluated by a quality-aware module, with results categorized as low, medium, and high quality. This module uses image quality metrics to guide model decisions during processing, combining three frozen pre-trained models with three learnable prediction heads. At the individual level, these three trained heads are integrated through a weighted soft voting ensemble, such as... Figure 5 As shown.
[0049] Overall, Reti-Pioneer achieved the following AUROC values on the internal test dataset: T2DM 0.833 (95% CI, 0.810–0.856); gout 0.832 (0.799–0.866); osteoporosis 0.787 (0.742–0.833); hypertension 0.740 (0.726–0.755); hyperlipidemia 0.736 (0.721–0.751); and thyroid disease 0.699 (0.667–0.730), as shown in Table 2. Stratified by sex yielded highly consistent performance across male and female subgroups, as shown in Table 3.
[0050] Reti-Pioneer was evaluated on two independent external test sets representing vastly different healthcare resource environments in China.
[0051] like Figure 2 As shown, based on aggregated data from medical institutions in low-income and remote areas, namely Tibet, Xinjiang, and Guangxi, the model demonstrated strong discriminative performance for six chronic diseases, with AUROC values as follows: T2DM 0.821 (0.792-0.851); hypertension 0.805 (0.776-0.833); hyperlipidemia 0.628 (0.560-0.676); gout 0.731 (0.662-0.799); osteoporosis 0.904 (0.876-0.932); and thyroid disease 0.821 (0.785-0.858). Figure 2 This presentation showcases the receiver operating characteristic (ROC) curves of Reti-pioneer tests on six endocrine and metabolic diseases using independent external datasets under different healthcare resource environments. The Chinese dataset is divided into three healthcare environments: resource-limited environment, resource-rich environment A, and resource-rich environment B. Resource-rich environment A represents a large tertiary hospital, while resource-rich environment B represents a health checkup center. The dataset from Singapore is derived from the SEED cohort.
[0052] Reti-Pionee maintained robust performance on a combined dataset containing both low-income and high-income regions of China, demonstrating significantly better performance for T2DM, hyperlipidemia, and osteoporosis in this mixed resource environment compared to low-income and remote regions only. The breakdown results for each external population are shown in Table 4.
[0053] To assess its cross-ethnic generalization ability, this protocol was validated in the multi-ethnic Singapore Eye Disease Epidemiology Study (SEED). The overall AUROCs were: T2DM 0.686 (0.680–0.700); hypertension 0.749 (0.740–0.760); hyperlipidemia 0.615 (0.600–0.620). Stratified analysis by ethnicity showed consistent performance across groups. For hypertension, the corresponding AUROCs were 0.731, 0.746, and 0.769. For T2DM, the AUROCs were 0.646 for Indian participants, 0.674 for Chinese participants, and 0.692 for Malay participants. These results demonstrate that Reti-Pioneer is effective across diverse ethnic populations, particularly excelling in hypertension detection across all groups.
[0054] In external validation, calibration plots and Brier scores indicate that Reti-Pioneer has moderate to good calibration accuracy and overall performance, such as... Figure 6 As shown in figure a, Reti-Pioneer's decision curve analysis on external test datasets demonstrates that the model provides net clinical benefit in these diseases. This finding is consistent across multiple external validation environments, including both resource-constrained and resource-rich settings, and applicable SEED cohorts, such as... Figure 6 As shown in b.
[0055] The integration of the quality-aware module and the strategic integration of multiple pre-trained base models are crucial to the performance of Reti-Pioneer. Ablation studies show that combining the quality-aware module significantly improves the AUROC for T2DM compared to using only high-quality images (P=0.011) or using full-quality images without the module (P=0.021). The module also significantly improves accuracy for hypertension (P<0.001) and hyperlipidemia (P=0.037) compared to using only high-quality images, as shown in Table 5. Furthermore, the multimodal combination of retinal images and clinical metadata yields superior prediction accuracy compared to any single modality input, as shown in Table 5.
[0056] 2. Reti-Pioneer enables longitudinal risk stratification of endocrine and metabolic diseases over 5 and 10 years. Given that retinal manifestations of endocrine and metabolic diseases may precede clinical diagnosis, this protocol further evaluated the Reti-Pioneer framework in a fully preserved longitudinal subset of images collected from the UK Biobank between January 2006 and December 2021 to assess its performance in predicting six endocrine disorders at 5-year and 10-year intervals. The analysis included 15,704 participants and 31,748 images, 57.4% of whom were female, with a mean age of 56.6 ± 8.0 years. The self-reported racial distribution was as follows: 83.7% White, 6.7% Black, 4.3% Asian, and 5.4% other. Participants with pre-existing conditions at baseline were excluded from the analysis.
[0057] As expected, predicting disease onset over longer time intervals proved more challenging than cross-sectional screening, as shown in Table 6. The Reti-Pioneer model achieved an AUROC of 0.755 (95% CI: 0.702–0.809) for T2DM development within 5 years, decreasing to 0.736 (0.694–0.779) for 10 years. Similarly, for hypertension, the model achieved AUROCs of 0.755 (0.704–0.805) and 0.719 (0.679–0.759) at 5-year and 10-year intervals, respectively. The corresponding values for hyperlipidemia were 0.748 (0.695–0.801) at 5 years and 0.735 (0.694–0.775) at 10 years. The decision curves for the Reti-Pioneer framework, derived from the prospective validation test, are presented in Table 6. Figure 7 In a sensitivity analysis excluding individuals diagnosed within the first year after fundus image acquisition, the model maintained comparable performance across both the 5-year and 10-year prediction intervals, as shown in Table 7.
[0058] 3. Visualization and biological rationale of Reti-Pioneer To explain the diagnostic decisions within the Reti-Pioneer framework, this approach generates pixel-level saliency maps to identify fundus regions crucial for model predictions, such as... Figure 11 As shown. Furthermore, this protocol elucidates the biological explanation for the connection between underlying retinal features and endocrine pathophysiology.
[0059] This study investigated the biological basis linking retinal features with endocrine and metabolic pathophysiology by analyzing 256-dimensional latent embeddings extracted from the penultimate layer of the model. Using orthogonal partial least squares discriminant analysis (OPLS-DA), the study derived the principal components predicting retinal diseases from these embeddings. After adjusting for age and sex, these predictive components showed significant correlations with diseases except for osteoporosis (all P < 0.001), as shown in Table 8.
[0060] Subsequently, this approach integrated 2,920 plasma proteomics maps and applied Elasticnet regression to identify the top five disease-specific protein features for each of the six diseases, such as... Figure 12 As shown. Several key plasma proteins, including SCARA5 (T2DM), PLA2G7, PTPRF, and APOM (hyperlipidemia), have shown significant associations with OPLS-DA-derived predictive components of potential retinal features, such as... Figure 8 As shown, this is true even after adjusting for age, sex, race, body mass index, image quality, and assessment center, and after applying false detection rate (FDR) correction for multiple tests, as shown in Table 9.
[0061] In contrast, the polygenic risk score (PRS) for related features showed limited associations: only the PRS for LDL-C was significantly associated with the predictive components of potential retinal features, and this was observed only in the unadjusted model, as shown in Table 9. In summary, these findings provide biological justification for Reti-Pioneer by demonstrating that the retinal features identified by the model are associated with disease-related protein features and, to a lesser extent, with genetic risk factors.
[0062] 4. Reti-Pioneer Enhances Clinical Decision Making (AI Copilot) To benchmark the diagnostic performance of Reti-Pioneer against clinical experts, this protocol included an image interpretation study with participants being retinal specialists from independent institutions with 3-10 years of experience. Each specialist interpreted bilateral CFPs from 1,200 cases across 200 different patients with each of the six target diseases from the internally retained test set.
[0063] In subsequent stages, conducted 1–2 weeks apart to mitigate recall bias, the same ophthalmologist re-evaluated the images after reviewing the Reti-Pioneer diagnostic predictions and saliency heatmaps. With Reti-Pioneer assistance (Copilot), retinal specialists achieved an average accuracy of 88.0% for T2DM, 79.0% for gout, and 70.0% for thyroid diseases, compared to 71.0%, 51.0%, and 63.0% respectively without assistance when applied to the same images (as shown in Table 10). This trend of improved accuracy from AICopilot is consistent across other disease screening tasks. As expected, the performance of human specialists varied with their level of clinical experience; nevertheless, the diagnostic accuracy of Reti-Pioneer Copilot surpassed that of any single retinal specialist.
[0064] 5. The silent trial demonstrated the operational efficiency of Reti-Pioneer. To validate the deployment feasibility of the Reti-Pioneer framework, a web platform, http: / / retipioneer.cn, was developed, and a prospective silent trial was conducted in a primary healthcare setting. This trial evaluated the framework's performance in screening for six systemic diseases and benchmarked it against comprehensive physical and blood tests, such as... Figure 3 As shown in a. Figure 3 This is a prospective, silent trial evaluating the operational efficiency of Reti-pioneer in primary healthcare. Figure 3 This is a clinical workflow that integrates Reti-Pioneer screening with routine physical and laboratory tests.
[0065] In the clinical workflow, bilateral fundus images of 1,017 participants were acquired at the point of care and transmitted in real time to a central GPU computing infrastructure for immediate Reti-Pioneer analysis without disrupting routine physical examination procedures. Baseline characteristics of the 1,017 participants are summarized in Table 11.
[0066] Silent testing confirmed the high robustness of the integrated system, with an image acquisition success rate of 98.7% and an AI inference success rate of 100%. At the operational level, the Reti-Pioneer workflow demonstrated exceptional throughput efficiency. The median total time from image acquisition to report generation was significantly shorter than that required for laboratory or physical examination reports (30.6 ± 6.0 seconds vs. approximately 7.97–8.11 hours; P < 0.001), and also significantly shorter than the Finnish Diabetes Risk Score (FINDRISC) for T2DM (30.6 ± 6.0 seconds vs. 126.6 ± 48.1 seconds; P < 0.001). Figure 3 As shown in b and 3c. Figure 3 b is an example of a screening report generated by AI. Figure 3 c represents the median time to diagnosis among Reti-pioneer, laboratory tests, and the FINDRISC questionnaire (n=1,017). Bars represent the mean time to diagnosis, and error bars represent the standard deviation (SD). Paired two-tailed t-tests were used to determine statistical significance (P<0.001).
[0067] Furthermore, Reti-Pioneer demonstrates significant computational advantages compared to models trained from scratch, such as... Figure 3 As shown in d. Figure 3 d represents a comparison of the computational efficiency of Reti-Pioneer and the de novo trained model, expressed as the number of training parameters and total training time. Detailed performance metrics of Reti-Pioneer in the silent experiment are available in [link to silent experiment]. Figure 9 Provided by China.
[0068] 6. A prospective preliminary study evaluated the clinical feasibility of the Reti-Pioneer framework. To evaluate the integration of the Reti-Pioneer system into routine multi-disease screening care, a prospective real-world study was conducted at a community health service center and a physical examination center. A total of 606 participants were enrolled and assessed by primary care providers (PCPs) using the Reti-Pioneer Copilot. At baseline visits, retinal fundus images and clinical metadata were collected from all participants, and standard biochemical measurements were performed. Patients undergoing routine physical examinations were given Reti-Pioneer AI tests by their PCPs and were followed up two weeks later to confirm the biochemical results. Baseline characteristics of the cohort are summarized in Table 11.
[0069] Reti-Pioneer analyzed retinal images and metadata from all participants to screen for six endocrine and metabolic disorders. To minimize missed diagnoses, screening was conducted at a pre-defined decision threshold corresponding to 85% sensitivity, based on results determined in internal validation. Under a composite diagnostic criterion combining laboratory tests, physical examination, and self-reported medical history, Reti-Pioneer demonstrated superior discrimination for T2DM compared to the FINDRISC questionnaire (AUROC 0.776, 95% CI 0.710–0.842 vs. 0.565, 0.459–0.670; P < 0.001). Reti-Pioneer also exhibited a balanced sensitivity and specificity with a high negative predictive value, as shown in Table 12. The corresponding AUROC values for the other five diseases are: hypertension 0.843 (0.804-0.881); hyperlipidemia 0.699 (0.651-0.748); gout 0.804 (0.752-0.855); osteoporosis 0.877 (0.845-0.910); thyroid disease 0.646 (0.592-0.700). Figure 4 As shown in a and 4b. Figure 4 To assess the multi-disease screening performance and acceptability of Reti-Pioneer in preliminary studies, Figure 4 'a' represents the probability density plot of the six disease risks predicted by Reti-Pioneer, stratified according to disease status based on composite diagnostic criteria. Figure 4 b represents the area under the receiver operating characteristic (AUROC) curve for the six diseases.
[0070] Compared to all image models that use only high-quality images or do not have quality guidance, the integration of the quality-aware module produced improved or comparable AUROCs in all six diseases, as shown in Table 13.
[0071] Satisfaction surveys conducted with participants and clinicians demonstrated the system's high acceptance and usability. Over 80% of participants reported "very satisfied" across all evaluation dimensions, including ease of use, clarity of information, and efficiency in managing multiple diseases. Regarding willingness to pay, 52.0% of participants indicated they were willing, while 4.9% were not, primarily due to cost concerns, such as... Figure 4 As shown in c. Figure 4 c represents the participant's reported satisfaction after the Reti-Pioneer screening, expressed as a 5-point Likert scale and a willingness-to-pay response. Clinicians rated key deployment factors, identifying system integration, patient acceptance, and regulatory issues as the most important considerations, while giving high scores (mean score > 4.4 / 5) across all decision support dimensions, including workflow compatibility and accuracy trust. Figure 4As shown in d. Figure 4 d represents potential barriers to the deployment of Reti-Pioneer multi-disease screening as assessed by clinicians using a structured questionnaire, with n=15 independent clinicians. Bars represent mean scores, and error bars represent standard deviations (SD).
[0072] discuss This project aims to develop a low-cost, clinically applicable AI framework for multi-disease screening using accessible, practical, and cost-effective retinal imaging technology, with the potential to reduce disease burden, disability, and mortality at the population level, particularly in remote and underserved areas. To address this pressing public health need, this project developed Reti-Pioneer, a retinal imaging-based framework designed to detect and predict six major endocrine and metabolic diseases: type 2 diabetes, hypertension, hyperlipidemia, gout, osteoporosis, and thyroid disease. The framework was trained, validated, and externally tested using multinational and multi-ethnic datasets from China, the UK, and Singapore. Its real-world applicability was further demonstrated through a prospective silent trial and a pilot study conducted in a routine care setting.
[0073] The main findings of this study demonstrate that Reti-Pioneer can accurately detect six endocrine and metabolic diseases, with AUROCs of 0.833 for type 2 diabetes, 0.832 for gout, 0.787 for osteoporosis, 0.740 for hypertension, 0.736 for hyperlipidemia, and 0.699 for thyroid disease. Its performance remained consistent across external test datasets comprising diverse ethnic and geographical populations from Singapore, as well as various regions of China, including resource-rich and resource-limited environments. In these cohorts, Reti-Pioneer also effectively identified multiple morbidities, outperforming models based solely on clinical characteristics. Notably, the framework exhibited strong predictive power for future disease onset, with AUROCs predicting 10-year onset for: type 2 diabetes 0.736, hypertension 0.719, hyperlipidemia 0.735, gout 0.753, osteoporosis 0.813, and thyroid disease 0.662. Despite performance gaps observed between some cohorts—which may be attributed to racial differences, differences in image acquisition protocols such as pupil dilation versus non-dilation, label noise from electronic health records or self-reports, and systemic missed diagnoses of diseases in routine care—the model maintained strong generalization ability across all test populations and settings.
[0074] In this study, our approach integrates a quality-aware module into the framework because fundus image quality has been shown to be associated with overall health status. Reti-Pioneer achieves superior AUROC performance across the entire image quality range compared to models trained only on high-quality images or on all-quality images without explicit quality integration, making it particularly suitable for deployment in primary care and resource-constrained environments where image quality can vary significantly. While clinicians often attempt to re-capture suboptimal fundus images, this is often ineffective in patients with systemic comorbidities. This approach suggests that, in real-world clinical settings, particularly with automated imaging systems, low-quality images, while limiting ophthalmic diagnosis, may still be valuable for developing and improving base models. Embedding quality-aware assessment as an auxiliary module for multimodal data fusion enhances the model's adaptability to real-world clinical conditions. Future research should investigate whether specific imaging features, such as blurred patterns or artifacts, show independent associations with systemic physiological decline, such as Parkinson's disease or diabetes.
[0075] Reti-Pioneer further enhances clinical interpretability by generating saliency maps that highlight the areas in the fundus most relevant to model predictions. These visualizations help demonstrate the model's ability to capture disease-specific heterogeneity in diagnostic and prognostic tasks. For example, mounting evidence suggests that diabetes primarily affects retinal microvessels and peripheral retinal structures through endothelial dysfunction, oxidative stress, and thickening of the capillary basement membrane. When clinicians are provided with Reti-Pioneer's heatmaps and predictive outputs, they can make more informed decisions, effectively enhancing the clinical decision-making process. To further elucidate the underlying biological mechanisms, this protocol analyzes the relationship between latent retinal features derived from dimensionality reduction of retinal imaging parameters and disease-specific plasma proteomics maps. These proteins show significant associations with retinal features, providing insights into the molecular pathways linking protein expression to retinal pathophysiology. By combining oculomics with plasma proteomics, this protocol establishes a multimodal biomarker framework, enabling the AI system to decode spatially resolved disease mechanisms while maintaining clinical interpretability. Although several disease-related proteins failed to retain independent associations with potential retinal features after adjusting for confounding factors, the associations observed in the unadjusted model still support the biological interpretability of the model predictions. These findings suggest that retinal features identified by Reti-Pioneer can serve as surrogate indicators of potential protein signatures, even if they lack strict independence after multivariate adjustment.
[0076] To date, health economics analyses increasingly suggest that AI-driven retinal screening can generate substantial cost savings by reducing reliance on expert diagnosis, particularly in resource-constrained settings. However, prospective real-world deployment studies remain scarce. Our prospective evaluation of this protocol demonstrates that Reti-Pioneer significantly improves operational efficiency, with a median time from image acquisition to report generation considerably shorter than laboratory tests or the FINDRISC questionnaire for type 2 diabetes. Compared to conventional models trained from scratch, the Reti-Pioneer framework exhibits significant computational efficiency, and its deployment in primary care settings confirms clinical applicability, demonstrating consistent real-world performance and maintaining diagnostic accuracy. While our prospective silent trials and preliminary clinical studies provide crucial initial evidence for the operational stability of this framework, broader clinical adoption will require overcoming several key hurdles. First, the regulatory approval pathway for AI-based medical devices remains complex and geographically heterogeneous. Qualification will require well-designed, large-scale, multicenter trials that meet regional safety and efficacy criteria. Second, while current research establishes predictive accuracy, generating robust evidence of impact on individual patients remains essential. Future work should include long-term randomized controlled trials to evaluate whether deploying Reti-Pioneer can shorten the time to diagnosis of endocrine and metabolic diseases, optimize the allocation of expert resources, and ultimately strengthen long-term risk stratification and improve clinical outcomes.
[0077] While Reti-Pioneer's general capabilities have been well validated, several limitations warrant further consideration. First, although the training dataset incorporates demographic diversity, performance discrepancies observed in external multi-ethnic SEED cohorts highlight the need for further improvements in dataset balance and representativeness to enhance the model's generalization ability across all populations. Second, while Reti-Pioneer has shown potential in identifying endocrine and metabolic diseases from CFPs, particularly in individuals not routinely undergoing blood-based screening, its current diagnostic and predictive accuracy remains below the threshold required for widespread clinical adoption. Third, the longitudinal assessments in this study were conducted as a binary classification task within a fixed timeframe, rather than as a formal event occurrence time analysis. Future research should employ survival analysis approaches to more rigorously model risk. Furthermore, predicting chronic disease incidence is inherently limited by the inaccurate definition of disease onset in electronic health records, and potential misclassification between cases and controls could affect the effectiveness of longitudinal predictions. Fourth, although Reti-Pioneer is designed for multi-task analysis of common endocrine and metabolic diseases, its current scope does not include rare or severe, life-threatening conditions such as coronary artery disease, stroke, or death. However, Reti-Pioneer's flexible architecture allows for the inclusion of additional disease targets through the introduction of specialized encoders and efficient sample fine-tuning for task-specific decoders. Fifth, while this protocol employs analytical methods such as resilient network regularization and OPLS-DA to mitigate confounding factors and focus on disease-specific signals, residual confounding factors cannot be completely eliminated. The results of this protocol primarily support its biological interpretability by demonstrating that the AI model's predictions can serve as surrogate indicators of protein characteristics, rather than establishing independent associations. Finally, although this prospective preliminary study suggests that Reti-Pioneer has the potential to improve risk stratification, the effectiveness of "Reti-Pioneer-guided prevention strategies" as an alternative to traditional whole-body screening still needs to be validated in larger-scale randomized trials.
[0078] Conclusion: Reti-Pioneer represents a scalable and clinically valuable AI framework that leverages retinal imaging and cost-effective pre-trained models for risk stratification of patients with endocrine and metabolic diseases. This framework has significant potential for integration into real-world clinical workflows, providing a pathway to more efficient resource allocation and enhanced evidence-based decision-making across diverse healthcare settings.
[0079] 1. Study Population To train the model, this approach used CFPs from the UK Biobank and three hospital-based centers in China to maximize racial and demographic heterogeneity in the training dataset, such as... Figure 10 As shown. The UK Biobank study is an ongoing prospective cohort study with extensive phenotypic data. In short, macular CFP was acquired at baseline visits between 2009 and 2010 using the Topcon 3D OCT-1000 Mark II (Topcon) system without pupil dilation. Participants were randomized 2:1 from six assessment centers to either the retrospective diagnostic modeling group or the predictive modeling group, minimizing potential center-specific bias. Participants with complete CFP data and disease outcomes and related covariate data were included in the analysis.
[0080] The hospital-based cohort was retrospectively established using data from three clinical centers: Guangdong Provincial People's Hospital, the Third Affiliated Hospital of Sun Yat-sen University, and Linyi People's Hospital. Macular CFP was acquired using various standard fundus cameras, including Topcon TRC-NW6 (Topcon), Canon CR6-45NM (Canon), and Kowa Nonmyd a-DIII (Kowa), under dilated pupil conditions.
[0081] For external testing, this protocol utilized datasets from different medical settings in China and the multi-ethnic cohort “Singapore Eye Disease Epidemiology Study (SEED)”. The limited-resource dataset was retrospectively assembled from three clinical sites in China: the First Affiliated Hospital of Xinjiang Medical University, the People's Hospital of Linzhi City, Tibet Autonomous Region, and the People's Hospital of Guangxi Zhuang Autonomous Region. This dataset contains 1,373 participants and 2,746 images, taken under dilated conditions. The more abundant dataset includes 2,003 participants and 4,006 images from a health checkup center, and 587 participants and 1,174 images from a large tertiary hospital; these were retrospectively collected data, primarily acquired without dilation.
[0082] Furthermore, this protocol evaluated the model in the SEED study, a population-based cohort of Singaporean Chinese, Malay, and Indian adults. A total of 15,306 CFPs from 7,653 Singaporean adults, including 2,800 Chinese, 2,385 Malay, and 2,468 Indian, were included in the external test set.
[0083] This study was approved by the Institutional Review Board of Guangdong Provincial People's Hospital and Guangdong Academy of Medical Sciences (KY-N-2022-134). Data from the UK Biobank portion received ethical approval from the NHS National Research Ethics Service (reference: 11 / NW / 0382; project ID: 86091). All participants provided written informed consent before participating in the study. To ensure confidentiality, all data were systematically de-identified by removing personal identifiers before analysis.
[0084] 2. Definition for endocrine and metabolic outcomes To define osteoporosis, gout, and thyroid disease at baseline, this protocol utilizes self-reported information and inpatient records using International Classification of Diseases, 10th Revision (ICD-10) codes (three-character categories specific to the ICD level), as shown in Table 1.
[0085] For the diagnosis of baseline type 2 diabetes mellitus (T2DM), hypertension, and hyperlipidemia, in addition to self-reported information and ICD-10 codes, this protocol also considers blood biomarkers and medication use. Patients diagnosed with T2DM are those diagnosed by a physician with diabetes, using hypoglycemic agents or insulin, or with a glycated hemoglobin level >48 mmol / mol. Hypertension is defined as using antihypertensive medication and having a mean systolic blood pressure of at least 130 mmHg or a mean diastolic blood pressure of at least 80 mmHg. Hyperlipidemia is defined as using lipid-lowering agents or statins, or having a total cholesterol level ≥6.22 mmol / L.
[0086] For the predictive task, incidental cases of each disease were identified using hospital records and death registers. Follow-up began from the date of participation in the assessment center and continued until the earliest recorded date of diagnosis, date of death, or the last available date provided by the hospital or general practitioner, whichever occurred first. Individuals with an epidemic at baseline were excluded. A sensitivity analysis was performed to exclude individuals who developed an incidental disease during the first year of follow-up.
[0087] 3. Reti-Pioneer Development and Evaluation The overall architecture of the proposed Reti-Pioneer framework is presented in Figure 5For feature extraction, Reti-Pioneer utilizes three large-scale pre-trained models: Swin Transformer, Vision Mamba, and RET Found, which are frozen during training. To achieve multimodal ensemble, two bilinear modules are employed for feature fusion, integrating image quality and clinical features, including age, sex, ethnicity, and weight.
[0088] Specifically, the Reti-Pioneer framework introduces an integrated quality-aware module into retinal image analysis, enabling the preservation and effective use of low-quality CFPs, such as images blurred by cataracts, which may still reflect systemic endocrine states. Instead of excluding these images, the framework utilizes image quality metrics to guide model decisions in early processing stages.
[0089] The quality-aware module and clinical feature fusion module utilize a bilinear operation, followed by a scaled exponential linear unit (SELU) and a linear layer. The bilinear module combines the quality metric with F-dimensional deep features to generate fused features of the same dimension ($F$), categorizing the quality metric as "low," "medium," and "high," and calculating their probability scores. This fusion is followed by a linear layer that outputs the disease probability. The SELU activation function is applied between intermediate layers to enhance training stability and maintain self-normalization properties. Both SELU activation and linear transformation preserve this dimension, ensuring consistency in feature representation.
[0090] Fine-tuning was applied to various downstream tasks, including screening and prediction. The generalization capabilities of the SwinTransformer, Vision Mamba, and RET Found models were systematically evaluated in independent external tests by incorporating bilateral CFP and clinical features. All analytical units for model development and evaluation were individual participants. For each subject, using bilateral fundus images as input, three trained prediction heads were integrated via a weighted soft voting ensemble and then fused by the model architecture to generate a single, unified prediction.
[0091] To quantify the contribution of its core components, this scheme underwent a series of ablation studies on an internal validation set. The complete Reti-Pioneer framework was compared to: (1) an ablation version without a quality-aware fusion module, trained on all images; (2) the same ablation version trained specifically on high-quality images; and (3) a unimodal baseline using only fundus images or clinical metadata. These comparisons aimed to separate the performance gains attributable to the model's complete multimodal architecture.
[0092] 4. Explainability and biological plausibility To understand retinal features associated with endocrine and metabolic disorders, this protocol employed two complementary approaches.
[0093] Saliency mapping was performed to identify and visualize the parsing regions of interest (ROIs) that contribute most to the model's diagnostic and predictive decisions within the Gaussian-filtered CFP. This approach employs an integrated gradient algorithm to generate a visual interpretation of the model's behavior, attributing predictive relevance to individual pixels. The saliency map highlights the relative contribution of each pixel to the final prediction, thereby enhancing the model's interpretability.
[0094] To investigate the biological basis of disease prediction, this protocol utilized plasma proteomics analysis data from participants in the UK Biobank. Proteins were quantified using the Olink Explore 3072 platform, which integrates cardiometabolic, inflammatory, neurological, and oncological data, capturing 2,923 unique proteins. Expression levels were reported as log₂-scale normalized protein expression (NPX). After quality control, proteins with missing data >30% (GLIPR1, NPM1, PCOLCE) and participants with missing data >50% were excluded, retaining 2,920 proteins for analysis.
[0095] To identify disease-specific protein signatures associated with metabolic and endocrine disorders, this protocol employed a two-step analysis. First, elastic netregularization analysis was used for retinal latent signature selection to reduce dimensionality and identify key predictors from the Reti-Pioneer framework. Second, candidate proteins were prioritized based on their coefficients to refine biologically relevant biomarkers. Orthogonal partial least squares discriminant analysis (OPLS-DA) was performed using retinal latent signatures derived from variational autoencoders to assess the separation between control and case groups. Predictive components were extracted, capturing the systematic variance directly correlated with inter-group differences. Logistic regression was used to test the impact of predictive components on the prevalence of corresponding diseases. Pearson correlation analysis was used to quantify the linear relationship between individual protein levels and corresponding predictive components, with individual protein levels being continuous. Logistic regression was further used to evaluate the association between OPLS-DA-derived retinal latent signatures and the top-ranking plasma proteins in six endocrine and metabolic diseases. Both the unadjusted model and the model adjusted for age, gender, race, body mass index, image quality, and evaluation center are presented.
[0096] Similarly, this protocol used logistic regression to assess the association between OPLS-DA-derived retinal latent features and disease-associated polygenic risk scores (PRS). Standard PRS, obtained from the UK Biobank, included scores for T2DM, HbA1c, hypertension, high-density and low-density lipoprotein cholesterol, total cholesterol, triglycerides, osteoporosis, and estimated bone mineral density T-scores. Similarly, both unadjusted and adjusted models for age, sex, ethnicity, body mass index, image quality, assessment center, and the top four ancestral principal components are presented. False discovery rate (FDR) correction was applied for multiple comparisons between retinal latent features and top-ranking plasma proteins or polygenic risk scores.
[0097] 5. Clinical pipeline of silent trial The prospective silent trial was conducted in a primary care setting to evaluate the technical robustness of the Reti-Pioneer framework for seamless integration into real-world clinical workflows under blinded conditions. Figure 3 a showcases a deployed clinical workflow for real-time detection of multiple comorbidities.
[0098] During the trial, the system routinely processed 30–50 cases daily from patients referred to primary care settings for endocrine, metabolic, and hematological testing. The performance of the Reti-Pioneer screening system for six endocrine and metabolic disorders was evaluated, compared to a reference standard incorporating laboratory measurements, physical examination results, and self-reported medical history, and further compared to the FINDRISC questionnaire for type 2 diabetes mellitus (T2DM).
[0099] In the clinical workflow, fundus images of all eligible participants were systematically collected at the point of care. Regardless of differences in image quality due to operator or environmental factors, the images were transmitted in real time to a central GPU computing infrastructure for immediate analysis without interfering with routine clinical procedures. The AI-generated predictions were not disclosed to clinicians or patients and did not influence clinical decisions, thus maintaining the blinded nature of the trial.
[0100] Key operational metrics were automatically recorded during the trial, including image acquisition success rate, AI model inference success rate, and system throughput efficiency (defined as the total time from image acquisition to report generation). This data allows for direct comparison of the Reti-Pioneer assisted screening workflow with current standard workflows based on laboratory testing.
[0101] 6. Prospective pilot study This study conducted a prospective preliminary study in community health service centers and physical examination centers to evaluate the impact of the Reti-Pioneer AI system on clinical practice, patient behavior, and clinician / patient cognition in an open-label environment. Participants undergoing routine physical examinations were recruited and given the Reti-Pioneer AI test by their primary care provider (PCP). CFP and clinical metadata, as well as comprehensive laboratory-based physical examination results, were collected from all participants. AI-generated predictions were provided to clinicians and patients as part of the clinical information.
[0102] The consistency of diagnoses is assessed by using computational models to output the sensitivity, specificity, and accuracy of composite diagnostic criteria that include laboratory results and self-reported medical history.
[0103] The satisfaction and acceptance of participants and clinicians were assessed using a structured questionnaire. Participants rated their experience across multiple areas using a 5-point Likert scale, including ease of use, report comprehensibility, clarity of management recommendations, overall satisfaction, willingness to reuse and recommend the service, perceived efficiency of multi-disease screening, information load, perceived convenience compared to traditional testing, and willingness to pay (Table 14).
[0104] Clinicians completed an independent acceptance questionnaire, also based on a 5-point Likert scale, to assess overall satisfaction, willingness to integrate the system into practice, motivation to conduct multi-disease screening, workflow compatibility, perceived decision support value, trust in AI predictions, and applicability to resource-constrained environments. The complete questionnaire items are detailed in Supplementary Table 10.
[0105] 7. Statistical analysis Descriptive statistics report "mean ± standard deviation" for continuous variables and "frequency (%)" for categorical variables.
[0106] Discriminative performance was assessed using the area under the receiver operating characteristic (AUROC) curve. AUROCs were compared using the DeLong test. Calibration was assessed using a calibration plot (observed risk vs. predicted risk at the decimal point). Overall performance was assessed using the Brier score. Sensitivity, specificity, positive predictive value, and negative predictive value for the classification were all accompanied by 95% confidence intervals. Confidence intervals were obtained using the percentile bootstrap method on a bootstrap sample of 1000. Decision curve analysis was considered to assess clinical utility across a range of risk thresholds.
[0107] Data analysis and visualization were performed in Python (v.3.12) and R (v.4.2.0). Reti-Pioneer was implemented in PyTorch (version 2.8.0) and optimized by the AdamW optimizer with a warm-up strategy and a cosine annealing learning rate scheduler. The build, fine-tuning, and testing processes were performed on an RTX 3080 GPU (10GB dedicated VRAM, CUDA version 12.8).
[0108] 8. Data and Code Availability Patient-level data is available with informed consent from the institutional data management committee, but is not publicly disclosed. Requests for access to de-identified individual-level data from China can be submitted via email. SEED data may be offered to researchers who meet confidential data access standards and have received approval from the institutional review committee. Researchers who comply with the relevant agreements will be granted access. Data from UK Biobank is available through its portal under controlled access.
[0109] The code used to develop the algorithm in this research is available on GitHub: https: / / github.com / lyhyl / Reti-Pioneer.
[0110] Table 1: Baseline features of the training dataset
[0111] Data for continuous variables are expressed as mean ± standard deviation, while data for categorical variables are expressed as a percentage (%). The UK Biobank is a community-based prospective cohort, while the Chinese cohort includes hospital-based registries. T2DM stands for type 2 diabetes; SD stands for standard deviation.
[0112] Table 2: Performance Overview of Reti-Pioneer on the Internal Validation Set
[0113] Table 3: Reti-pioneer's gender segmentation performance on the internal validation set
[0114] Table 4: Performance Overview of Reti-Pioneer on External Test Datasets in Different Scenarios
[0115]
[0116]
[0117]
[0118] Performance metrics were reported as area under the receiver operating characteristic (AUROC), sensitivity, specificity, 95% confidence intervals (CIs), and Brill scores. Data sets were grouped by geographic region and healthcare resource setting. Data from China were divided into three resource settings: (1) Resource-scarce setting (aggregated data from Tibet, Xinjiang and Guangxi, China); (2) High-resource type setting A (data from Guangdong Provincial Health Examination Center, China); (3) High-resource type setting B (data from large tertiary hospitals in Guangdong Province, China).
[0119] Data from Singapore comes from the Singapore Eye Disease Epidemiology (SEED) cohort study, a population-based, multi-ethnic study. Bold text indicates aggregated indicators across all sites. A dash (-) indicates that the corresponding outcome indicator is unavailable.
[0120] Table 5: Performance of the Reti-pioneer framework on the internal validation set
[0121] Performance comparisons were performed using the full Reti-Pioneer framework as the reference model. High-quality images only: Frameworks trained and evaluated solely on quality-controlled bilateral retinal images, excluding low-quality images. No quality-aware module: Frameworks trained on all available bilateral retinal images but without an integrated quality assessment module. P-values are derived from the DeLong test for AUROC comparison. Bold text indicates statistical significance (P < 0.05).
[0122] Table 6: Reti-Pioneer longitudinal risk stratification of endocrine and metabolic diseases at 5-year and 10-year intervals
[0123] Performance metrics were reported as area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), 95% confidence interval (CI), and Brier score. T2DM refers to type 2 diabetes mellitus.
[0124] Table 7: Performance overview of Reti-Pioneer on external validation datasets in different scenarios, calibration curve and decision curve analysis
[0125] Table 8: Association analysis of retinal latent features extracted based on OPLS-DA with endocrine and metabolic diseases
[0126] Table 9: Association analysis between retinal latent features and core plasma proteins and PRS
[0127] Proteins were screened using elastic network regression. Polygenic risk scores were obtained from the UK Biobank. A logistic regression model was fitted after binary classification using retinal latent features (OPLS-DA predicted principal components) as the dependent variable and the optimal cutoff value of the predicted principal components based on retinal features. Model 1: Uncorrected. Model 2: Corrected for age and sex.
[0128] Table 10: Diagnostic accuracy of six endocrine disorders by retinal specialists with the assistance of Reti-Pioneer
[0129] Table 11: Baseline characteristics of 1,017 participants
[0130] Table 12: Reti-Pioneer Prospective Assessment Combined with Quality Perception Module
[0131] Table 13: Prospective efficacy comparison of Reti-pioneer and traditional FINDRISC scales in clinical pilot studies
[0132] Performance comparison with the full Reti-Pioneer framework (reference model). High-quality images only: The framework is trained and evaluated only on bilateral retinal images that have passed quality control, while low-quality images are excluded. No quality-aware module: Training is performed on all available bilateral retinal images, but a quality evaluation module is not integrated. Two-sided p-values for AUC comparisons are obtained using the DeLonghi test, with Reti-Pioneer as the reference model. P-values for all comparisons are provided (bold indicates P < 0.05). The Finnish Type 2 Diabetes Risk Score (FINDRISC) was used for type 2 diabetes risk assessment. AUROC, Area Under the Receiver Operating Characteristic (ROC) curve; PPV, Positive Predictive Value; NPV, Negative Predictive Value; T2DM, Type 2 Diabetes.
[0133] Table 14: Questionnaire Survey of Participants in the Reti-Pioneer Prospective Clinical Trial
[0134] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementation methods. Clearly, many modifications and variations can be made based on the content of this specification. The selection and specific description of these embodiments in this specification are intended to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. An artificial intelligence method for screening multiple systemic diseases based on retinal imaging, characterized in that, Includes the following steps: S1. Obtain macular CFP data from the crowd cohort and hospital cohort, construct a multimodal dataset, and divide the multimodal dataset into a training set, a validation set, and an external test set; S2. Construct the Reti-pioneer framework, a retina-based disease prediction pre-training model, for the simultaneous screening and prediction of six endocrine and metabolic diseases. S3. Validate and evaluate the generalization ability of the Reti-pioneer framework using external ethnic data; S4. Enhance the interpretability of the Reti-pioneer framework through saliency images; S5. Obtain potential retinal features and disease-specific plasma proteins, and verify the biological rationale of the Reti-pioneer framework by analyzing the relationship between potential retinal features and disease-specific plasma proteomics maps. S6. The ecological validity and workflow efficiency of the model were measured in a primary healthcare setting using a prospective silent trial.
2. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 1, characterized in that, The Reti-pioneer framework includes a quality-aware module and three frozen pre-trained visual foundation modules. The quality-aware module is used to evaluate image quality. The three frozen pre-trained visual foundation modules are SwinTransformer, Vision Mamba, and RET Found. The SwinTransformer, Vision Mamba, and RET Found modules are used to extract features from CFP images.
3. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 2, characterized in that, The Reti-pioneer framework also includes a feature fusion module and a scaling exponential linear unit. The feature fusion module is used to fuse the extracted features, and the scaling exponential linear unit is used to scale the fused features.
4. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 3, characterized in that, The construction of step S2, based on the Reti-pioneer framework for predicting retinal diseases, specifically involves: Image features are extracted using the Swin Transformer, Vision Mamba, and RET Found modules to extract clinical features. The feature fusion module then fuses image quality and clinical features, including age, gender, ethnicity, and weight. The quality perception module and the feature fusion module utilize bilinear operations to combine quality indicators with F-dimensional deep features to generate fused features of the same dimension. The quality indicators are categorized as "low," "medium," and "high," and their probability scores are calculated. The scaling exponential linear unit scales the fused features; After feature fusion, a linear layer is formed to output the disease probability. The SELU activation function is applied between intermediate layers to enhance training stability and maintain self-normalization properties. Fine-tuning is applied to various downstream tasks, including screening and prediction tasks.
5. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 2, characterized in that, Step S2, which uses the Reti-pioneer framework to screen for and predict six endocrine and metabolic diseases, specifically involves: The fundus image is input into the Reti-pioneer framework, where the Swing Transformer, Vision Mamba, and RET Found modules are integrated through weighted soft voting and fused by the model architecture to generate a single, unified prediction.
6. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 1, characterized in that, The external ethnic data in step S300 includes data from a first dataset and a second dataset. The first dataset was collected under dilated pupil conditions, and the second dataset was collected under non-dilated pupil conditions.
7. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 6, characterized in that, The specific steps of evaluating the generalization ability of the Reti-pioneer framework through external ethnic data in step S3 are as follows: Compare the complete Reti-Pioneer framework with the following: (1) An ablation version without a quality-aware module, trained on all images; (2) The same ablation version with a quality-aware module, specifically trained on high-quality images; (3) Using only fundus images or clinical metadata as a single-modal baseline to isolate performance gains attributable to the model’s full multimodal architecture; All experimental results were categorized according to disease type and visualized using calibration curves and decision analysis curves.
8. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 1, characterized in that, The specific steps of enhancing the interpretability of the Reti-pioneer framework through saliency images in step S4 are as follows: Saliency mapping is performed to identify and visualize the analytical regions of interest that contribute most to the model's diagnostic and predictive decisions in the Gaussian-filtered fundus color image; an integral gradient algorithm is used to generate a visual explanation of the model's behavior, attributing predictive relevance to individual pixels; The saliency image shows the relative contribution of each pixel to the final prediction.
9. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 1, characterized in that, The specific steps of obtaining potential retinal features and disease-specific plasma proteins in step S5 are as follows: Based on plasma proteomic analysis data from a multimodal dataset, proteins were quantified and captured using the Olink Explore 3072 platform. Protein expression levels were reported as normalized protein expression on a $log_2$ scale. After quality control, proteins with missing data >30% and missing data >50% were excluded, and the remaining proteins were retained as disease-specific plasma proteins for analysis. Orthogonal partial least squares discriminant analysis was used to screen for potential retinal features. Elastic network regularization analysis was used to select potential retinal features to achieve dimensionality reduction and identify key predictors from the Reti-Pioneer framework. The priority of candidate proteins was determined based on their coefficients, and biologically relevant biomarkers were extracted for selection.
10. The artificial intelligence method for screening multiple systemic diseases based on retinal imaging according to claim 9, characterized in that, Step S5, which verifies the biological rationale of the Reti-pioneer framework by examining the relationship between latent retinal features and disease-specific plasma proteomics maps, specifically involves: Using retinal latent features derived from variational autoencoders, orthogonal partial least squares discriminant analysis (OPLS-DA) was performed on a model to assess the separation between the control and case groups. Predictive components were extracted, and systematic variances directly related to inter-group differences were captured. Logistic regression was used to test the impact of predictive components on the prevalence of corresponding diseases. Pearson correlation analysis was used to quantify the linear relationship between individual protein levels and corresponding predictive components. Logistic regression was used to assess the association between retinal latent features derived from OPLS-DA and the top-ranking plasma proteins in six endocrine and metabolic diseases, resulting in both unadjusted and adjusted models for age, sex, race, body mass index, image quality, and assessment center.
