Authors: Shengjie Li, Jun Ren, Fenglin Wang, Jianing Wu, Yingzhu Li, Xuanxuan Wang, Mengyu Zhang, Henggui Hu, Yunxiao Song, Wenjun Cao, Xingtao Zhou, Meiyan Li
Categories: Article, Eye diseases, Predictive markers, Machine learning
Source: Nature Communications
Authors: Shengjie Li, Jun Ren, Fenglin Wang, Jianing Wu, Yingzhu Li, Xuanxuan Wang, Mengyu Zhang, Henggui Hu, Yunxiao Song, Wenjun Cao, Xingtao Zhou, Meiyan Li
High myopia can lead to cataract, glaucoma, retinal detachment, choroidal neovascularisation, and macular degeneration, causing irreversible vision loss. Imaging detects these complications, but population screening is limited by equipment, and specialist availability. Here we show that a machine learning model using routine blood test results identifies people at increased risk of complications related to high myopia during standard health examinations. We develop the model in a multicentre study of 10,661 participants and validate it in two independent cohorts. The model shows high accuracy across centres (area under the receiver operating characteristic curve=0.9010-0.9649) and flags individuals who receive a clinical diagnosis in a hospital-based prospective follow-up study of 5,067 participants. In a community screening study of 311,254 adults, the model increases the yield of detected complications among those referred for ophthalmic assessment (positive predictive value = 74%). This scalable blood-based approach supports opportunistic screening and earlier referral in primary care and community settings.
Myopia is a global epidemic, with high myopia identified as a growing concern because of its associated risk of complications^1,2^. Myopia, particularly high myopia, is not merely a refractive error but is associated with markedly increased risks of vision-threatening complications^3,4^, including cataracts (3–8-fold^5,6^), glaucoma (~7-fold^7^), retinal detachment (12–39^8^), myopic macular degeneration (~20^9,10^), and choroidal neovascularisation (~10^8^), compared with non-myopic eyes. While the risk of complications increases with increasing myopia, the population impact lower degrees are more prevalent and thus contribute greatly to the overall burden, whereas high myopia, although less common, is associated with a much higher individual risk and accounts for more than half of pathological myopia. Some complications, such as retinal detachment and glaucoma, may benefit from early detection and timely intervention (e.g., prophylactic laser for high-risk retinal lesions), whereas effective treatments for myopic macular degeneration are lacking, and this condition is the leading cause of irreversible blindness in working-age populations, particularly in urbanised regions^11^. The substantial economic burden associated with these complications and their impacts on quality of life underscore the critical need for effective early screening strategies^12^.
Current diagnostic paradigms rely heavily on specialised imaging modalities such as optical coherence tomography (OCT), fundus autofluorescence, and perimetry. While these techniques provide detailed structural information about the eyes, they present substantial limitations in real-world clinical practice. First, considerable barriers exist to accessing these technologies, with approximately 78% of low- and middle-income countries lacking basic OCT infrastructure^13^. Second, considerable variability has been documented in clinician interpretations of the imaging findings, with studies reporting up to 22% variability in glaucoma detection using OCT parameters^14^. Third, the resource-intensive nature of these examinations limits their applicability for population-level screening, resulting in delayed diagnoses, a critical concern given that nearly 40% of myopic macular lesions are asymptomatic in the early stages^15^. Recent advances in portable fundus cameras and smartphone-based imaging offer cost-effective, accessible, and point-of-care alternatives for retinal assessment^16,17^. These technologies facilitate the early detection and monitoring of retinal pathologies, including diabetic retinopathy, age-related macular degeneration, and retinal vascular disorders, thereby improving accessibility to essential diagnostic services. Nevertheless, challenges remain regarding image quality, standardisation, regulatory approval, and medicolegal considerations, highlighting the continued need for complementary screening approaches such as blood-based risk stratification models.
Emerging evidence has demonstrated the underutilised potential of routine clinical blood tests in ocular risk assessments. Haematological markers such as the lymphocyte count, basophil count, mean platelet volume, and the neutrophil-to-lymphocyte ratio can predict the risk of retinal detachment^18,19^, whereas metabolic parameters such as the level of haemoglobin A1c (HbA1c) are correlated with maculopathy progression^20^, and higher levels of inflammatory markers such as IL-6 are clearly associated with an increased glaucoma risk^21^. These findings suggest that routinely available clinical laboratory data may provide valuable insights into the risks of myopia-related complications; nevertheless, they remain systematically underused in current screening protocols. From a cost perspective, blood-based screening is also highly efficient. In health examination centres and primary care settings, the required haematological parameters are already determined through routine blood tests, enabling opportunistic screening for high myopia-related complications at no additional expense. Even if blood tests were performed specifically for ocular screening, the median cost is approximately 300 RMB per person, which is lower than that of fundus-based imaging (400–600 RMB). Additionally, blood tests provide concurrent systemic health information, offering dual-purpose clinical utility. These features underscore blood-based models as a complementary, scalable, and low-cost strategy for identifying high myopia-related complications.
Routine clinical laboratory tests are among the most common diagnostic methods used in medical practice and provide valuable and timely insights into various diseases^22^. As blood cells interact with different tissues and organs, these tests can serve as powerful diagnostic tools. Their key advantages—low cost, accessibility, consistency, and widespread use in primary healthcare settings—make them indispensable in routine medical evaluations^23^. The integration of machine learning (ML) with routinely available clinical laboratory data represents a transformative approach to address the challenges mentioned above. Recent advances have shown that ML algorithms can effectively predict high myopia-related complications by leveraging routinely available blood data^18,19,24^. These technological advancements could help address the urgent need for scalable screening tools, particularly in resource-limited settings, by enabling opportunistic screening during routine health examinations, where the burden of high myopia-related complications is increasing most rapidly.
This large-scale study was conducted in three phases—a retrospective multicentre case–control study, a hospital-based prospective cohort study, and a community-based cross-sectional study—and leveraged routine blood test data combined with advanced machine learning architectures to establish a paradigm for opportunistic screening during routine health examinations for high myopia–related complications. In the retrospective multicentre case–control phase, we analysed data from a diverse cohort of 15,000 patients across five tertiary care centres in Asia to develop a robust predictive model integrating haematological, metabolic, and inflammatory markers. The performance of the model was validated through two large-sample cohort a hospital-based prospective cohort study and a community-based cross-sectional study. We hypothesise that blood-based screening during routine health examinations will be clinically advantageous over a direct ophthalmic examination in settings where specialised imaging is unavailable or impractical, particularly in health examination centres and community primary care facilities. This approach prioritises scalability and generalisability, aligns with the principles of precision medicine, and holds promise for reducing the global burden of preventable blindness associated with high myopia.
The structure and methodology of the case‒control study are outlined in Fig. 1. Between Jan. 1, 2016, and Dec. 30, 2024, 10,661 individuals who met the inclusion criteria were enrolled at one of five hospitals in China. The Eye & ENT cohort (discovery cohort) included 5102 participants. The first external validation cohort included 3125 participants from Xuhui Central Hospital of Fudan University and Putuo People’s Hospital, and the second external validation cohort included 2434 participants from Wanbei Coal–Electricity Group General Hospital and the First Affiliated Hospital of Anhui Medical University. The numbers of patients with high myopia overall and associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment in the discovery cohort, validation cohort 1, and validation cohort 2 are illustrated in Fig. 1. Across all cohorts, patients with uncomplicated high myopia were generally younger than those with complicated high myopia, with the differences reaching statistical significance in the discovery cohort, validation cohort 1, and validation cohort 2 (P < 0.001). No significant differences in the sex distribution were observed between the uncomplicated and complicated high myopia groups in any cohort (P > 0.05). Detailed information about the variables of interest in the discovery cohort, validation cohort 1, and validation cohort 2 is presented in Tables S1–S3. Approximately 70% of the routine laboratory parameters differed significantly among the six groups (the overall high myopia group and the five complication groups). Effect sizes were calculated as follows to better assess their potential clinical η² for continuous variables from one-way ANOVA (0.01 = small effect, 0.06 = medium effect, and ≥0.14 = large effect) and Cramér’s V for categorical variables (0.1 = small effect, 0.3 = medium effect, and ≥0.5 = large effect). Notably, 38.71% (discovery cohort), 31.15% (validation cohort 1), and 37.7% (validation cohort 2) of the parameters showed a medium or large effect size, suggesting potential clinical relevance beyond statistical significance.Fig. 1Overview of the blood-based machine learning pipeline for screening high myopia-related complications.This flowchart outlines the development and validation of a machine learning model using routine blood test parameters to screen for HM and its complications. After the exclusion criteria were applied, the data from 5102 patients were used for model development, and the data from two independent cohorts (n = 3125 and 2434) were used for validation. Five HM-related complications were included. Models constructed from multiple algorithms were compared through key metrics (AUC, PR-AUC, sensitivity, specificity, PPV, NPV, and F1 score). A random forest model based on nine selected features showed optimal performance and was evaluated with a ROC curve analysis and confusion matrices. A web-based application was developed to facilitate clinical use. HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve, PR-AUC area under the precision‒recall curve, RF random forest, ROC receiver operating characteristic curve, F1 score the harmonic mean of precision and recall, UA uric acid, EOS eosinophil, PCT thrombocytocrit, BASO% basophil percentage, TG triglyceride.
In the discovery cohort, all 61 variables from the routine clinical blood tests were employed to train screening models based on the 10 ML algorithms. The performance of the models developed from the different ML algorithms is shown in Fig. 2A, and the detailed results are provided in Fig. S1 and Table S4. In terms of the area under the receiver operating characteristic curve (AUC) (Fig. 2B), the performance of the random forest (RF) (macroaverage AUC = 0.9707 [95% CI = 0.9480–0.9891]) and extreme gradient boosting (XGBoost) (macroaverage AUC = 0.9601 [95% CI = 0.9452–0.9806]) models was significantly better (Table S5; P < 0.05) than that of the other models according to the DeLong test, whose macroaverage AUC values ranged from 0.6007 to 0.9561. The results of the DeLong test (Table S5) also showed no significant difference (P = 0.184) between the RF and XGBoost algorithms.Fig. 2Performance of machine learning models for screening high myopia-related complications using the results of routine laboratory tests in the discovery cohort.A Comparison of multiple machine learning algorithms using key performance metrics, including AUC, sensitivity, specificity, PPV, NPV, accuracy, and F1 score. The results obtained through fivefold cross-validation are reported as the means ± standard deviations. Source data are provided as a Source Data file. B Heatmap comparing the AUC values of different models. The DeLong test, a nonparametric statistical method, was used to assess differences between the areas under the ROC curves. C ROC curves of the RF model for screening high myopia-related complications. D Confusion matrix heatmap illustrating the accuracy, sensitivity, and specificity of the predictions from the RF model. E ROC curves of the XGBoost model for screening high myopia-related complications. F Confusion matrix heatmap illustrating the accuracy, sensitivity, and specificity of the predictions from the XGBoost model. HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve, ROC receiver operating characteristic curve, F1 score the harmonic mean of precision and recall, RF random forest, DT decision tree, GB gradient boosting, KNN k-nearest neighbours, LGBM light gradient boosting machine, LR logistic regression, MLP multilayer perceptron, NB naïve Bayes, XGBoost extreme gradient boosting.
As shown in Fig. 2C, the RF model achieved AUCs of 0.9725 (95% CI = 0.9523–0.9901), 0.9782 (95% CI = 0.9581–0.9940), 0.9752 (95% CI = 0.9539–0.9918), 0.9721 (95% CI = 0.9504–0.9888), 0.9617 (95% CI = 0.9391–0.9815), and 0.9643 (95% CI = 0.9422–0.9834); negative predictive value (NPV)s of 0.9249 (95% CI = 0.8972–0.9489), 0.9333 (95% CI = 0.9041–0.9568), 0.9631 (95% CI = 0.9397–0.9843), 0.9698 (95% CI = 0.9451–0.9895), 0.9867 (95% CI = 0.9701–0.9971), and 0.9813 (95% CI = 0.9595–0.9962); and positive predictive value (PPV)s of 0.7626 (95% CI = 0.7271–0.7942), 0.7922 (95% CI = 0.7581–0.8239), 0.9072 (95% CI = 0.8773–0.9348), 0.9516 (95% CI = 0.9278–0.9735), 0.9231 (95% CI = 0.8929–0.9499), and 0.8868 (95% CI = 0.8535–0.9173) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. A confusion matrix (Fig. 2D) was created to visualise the performance of the RF model, revealing sensitivities of 0.8133 (95% CI = 0.7784–0.8469), 0.8966 (95% CI = 0.8681–0.9248), 0.7586 (95% CI = 0.7212–0.7929), 0.7108 (95% CI = 0.6749–0.7450), 0.6857 (95% CI = 0.6491–0.7192), and 0.7581 (95% CI = 0.7228–0.7906) and specificities of 0.9008 (95% CI = 0.8722–0.9271), 0.8603 (95% CI = 0.8251–0.8919), 0.9878 (95% CI = 0.9715–0.9970), 0.9961 (95% CI = 0.9827–0.9996), 0.9976 (95% CI = 0.9849–0.9997), and 0.9924 (95% CI = 0.9739–0.9991) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively.
The XGBoost model (Fig. 2E) achieved AUCs of 0.9620 (95% CI = 0.9526–0.9713), 0.9728 (95% CI = 0.9657–0.9799), 0.9694 (95% CI = 0.9592–0.9787), 0.9588 (95% CI = 0.9393–0.9752), 0.9595 (95% CI = 0.9331–0.9829), and 0.9720 (95% CI = 0.9568–0.9856); NPVs of 0.9174 (95% CI = 0.8983–0.9356), 0.9122 (95% CI = 0.8917–0.9328), 0.9589 (95% CI = 0.9467–0.9711), 0.9646 (95% CI = 0.9536–0.9747), 0.9868 (95% CI = 0.9796–0.9928), and 0.9802 (95% CI = 0.9718–0.9880); and PPVs of 0.7300 (95% CI = 0.6868–0.7755), 0.7746 (95% CI = 0.7405–0.8091), 0.8416 (95% CI = 0.7757–0.8990), 0.8594 (95% CI = 0.7879–0.9265), 1.000 (95% CI = 1.0000–1.0000), and 0.9388 (95% CI = 0.8824–0.9864) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. The confusion matrix (Fig. 2F) revealed sensitivities of 0.7967 (95% CI = 0.7520–0.8373), 0.8621 (95% CI = 0.8312–0.8931), 0.7328 (95% CI = 0.6667–0.8000), 0.6627 (95% CI = 0.5750–0.7471), 0.6857 (95% CI = 0.5428–0.8065), and 0.7419 (95% CI = 0.6469–0.8298) and specificities of 0.8846 (95% CI = 0.8618–0.9057), 0.8510 (95% CI = 0.8254–0.8748), 0.9784 (95% CI = 0.9689–0.9866), 0.9884 (95% CI = 0.9818–0.9936), 1.000 (95% CI = 1.0000–1.0000), and 0.9962 (95% CI = 0.9925–0.9988) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively.
Additionally, the area under the precision‒recall curve (PR-AUC) values of the RF (0.8917 [95% CI = 0.8786–0.9108]) and XGBoost (0.8788 [95% CI = 0.8514–0.8906]) models (Fig. S2) were significantly higher (P < 0.05) than those of the other models (Fig. S2) according to the DeLong test, confirming the superior performance of the models developed based on these two algorithms. Given these results, the RF and XGBoost algorithms were selected for the subsequent steps in the development of the screening model.
Feature reduction was performed on the data from the discovery cohort to optimise the performance of the models and increase their clinical applicability. The number of features was sequentially reduced from 61 to 1 based on their importance rankings to determine the minimum number of features that yielded optimal predictive capability.
The all-feature RF model was significantly better than the 8-feature model (P = 0.0227) but not significantly better than the 9-feature model (P = 0.0970) (Fig. 3A, Table S6). In other words, the macroaverage changes in the AUC value indicated that the RF model incorporating nine features (EOS, PCT, BASO%, M, UA, LPB, BASO, EOS%, and TG levels; macroaverage AUC = 0.9649 [95% CI = 0.9528–0.9770]) retained nearly the highest predictive ability. For the XGBoost algorithm, the all-feature model was significantly better than the 19-feature model (P < 0.0010, macroaverage AUC = 0.9493) but not significantly better than the 20-feature model (P = 0.1860, macroaverage AUC = 0.9621) (Fig. 3A, Table S6). Based on the AUC (0.9649 vs. 0.9621) and the number of included features (9 vs. 20), the nine-feature RF model was selected as the final screening model.Fig. 3Feature characteristics and identification of the final model for screening high myopia-related complications in the discovery cohort.A AUC values for models with different numbers of features, showing that the all-feature RF model performed better than the 8-feature model did (P = 0.0227) but did not perform significantly better than the 9-feature model (P = 0.0970). With respect to the XGBoost model, the all-feature model outperformed the 19-feature model (P < 0.0010) but did not significantly outperform the 20-feature model (P = 0.1860). The DeLong test (two-sided) was used to assess differences between the areas under the ROC curves. Source data are provided as a Source Data file. B Feature importance ranking scores for all variables associated with complications of high myopia. C ROC curves of the RF model based on the top nine features for screening high myopia-related complications. D PR-AUC curves of the RF model based on the top nine features. E Confusion matrix heatmap illustrating the accuracy, sensitivity, and specificity of the predictions from the RF model based on the nine selected features. HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve, ROC receiver operating characteristic curve, PR-AUC area under the precision‒recall curve, RF random forest, XGB extreme gradient boosting, EOS eosinophils, PCT thrombocytocrit, BASO% basophil percentage, M monocytes, UA uric acid, LPB apolipoprotein b, BASO basophils, EOS% eosinophil percentage, TG triglyceride, P phosphorus, LDL low-density lipoprotein, HB haemoglobin, GLB globulin, PA prealbumin, MPV mean platelet volume, PT prothrombin time, DD d-dimer, FIB fibrinogen, GLU glucose, LDH lactic dehydrogenase, SHAP SHapley Additive exPlanations.
The global SHapley Additive exPlanations (SHAP) analysis provided insights into the overall behaviour of the RF model, with SHAP summary plots (Fig. 3B) illustrating the ranked contributions of each feature based on their average SHAP values in descending order. The top nine contributing features were the EOS, PCT, BASO%, M, UA, LPB, BASO, EOS%, and TG levels. The distributions and correlations of these nine features are presented in Fig. S3A. Strong correlations were observed between the EOS and EOS% and between the BASO and BASO%, whereas no significant correlations were identified among the other variables.
The 9-feature RF model achieved a macroaverage AUC of 0.9649 (95% CI = 0.9528–0.9770) (Fig. 3C) and an PR-AUC of 0.8812 (95% CI = 0.8045–0.9579) (Fig. 3D). At the subgroup level, the RF model achieved AUCs of 0.9736 (95% CI = 0.9480–0.9880), 0.9790 (95% CI = 0.9342–0.9885), 0.9705 (95% CI = 0.9272–0.9772), 0.9589 (95% CI = 0.9348–0.9706), 0.9475 (95% CI = 0.8976–0.9788), and 0.9599 (95% CI = 0.9214–0.9746); NPVs of 0.9387 (95% CI = 0.8915–0.9926), 0.9333 (95% CI = 0.8821–0.9692), 0.9650 (95% CI = 0.9150–0.9797), 0.9669 (95% CI = 0.9465–0.9759), 0.9867 (95% CI = 0.9478–0.9974), and 0.9824 (95% CI = 0.9523–0.9894); and PPVs of 0.8602 (95% CI = 0.8075–0.9178), 0.8237 (95% CI = 0.7660–0.8618), 0.7860 (95% CI = 0.7436–0.8160), 0.7451 (95% CI = 0.6978–0.7964), 0.7273 (95% CI = 0.6941–0.7578), and 0.7869 (95% CI = 0.7352–0.8387) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively (Fig. 3C). The confusion matrix (Fig. 3E) revealed sensitivities of 0.8423 (95% CI = 0.7987–0.8823), 0.8934 (95% CI = 0.8398–0.9297), 0.7759 (95% CI = 0.7182–0.8305), 0.6867 (95% CI = 0.6486–0.7280), 0.6857 (95% CI = 0.6513–0.7428), and 0.7742 (95% CI = 0.7265–0.8065) and specificities of 0.9463 (95% CI = 0.8967–0.9943), 0.8864 (95% CI = 0.8383–0.9175), 0.9689 (95% CI = 0.9257–0.9755), 0.9832 (95% CI = 0.9416–0.9940), 0.9915 (95% CI = 0.9546–1.0000), and 0.9849 (95% CI = 0.9517–0.9929) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. The optimal cut-off values of the final RF model for the detection of high myopia (Fig. S4A) and high myopia associated with cataracts (Fig. S4B), macular degeneration (Fig. S4C), choroidal neovascularisation (Fig. S4D), glaucoma (Fig. S4E), and retinal detachment (Fig. S4F) were 0.47, 0.62, 0.16, 0.18, 0.11, and 0.12, respectively. A detailed description of the performance, including sensitivity, specificity, PPV, NPV, accuracy, and F1 score, of the final RF model is provided in Table 1.Table 1Nine features-based RF Model Performance in the discovery cohort and validation cohortsDiseaseSensitivity (95%CI)Specificity (95%CI)Accuracy (95%CI)PPV (95%CI)NPV (95%CI)F1 (95%CI)AUC (95%CI)PR-AUC (95%CI)DiscoveryCohort 1HM0.8423(0.7987, 0.8823)0.9463(0.8967, 0.9943)0.9171(0.8799, 0.9478)0.8602(0.8075, 0.9178)0.9387(0.8915, 0.9926)0.8512(0.8117, 0.8843)0.9736(0.9480, 0.9880)0.9495(0.8912, 0.9933)HM associated with cataract0.8934(0.8398, 0.9297)0.8864(0.8383, 0.9175)0.8890(0.8530, 0.9410)0.8237(0.7660, 0.8618)0.9333(0.8821, 0.9692)0.8571(0.8070, 0.9109)0.9790(0.9342, 0.9885)0.9709(0.9427, 0.9788)HM associated with MD0.7759(0.7182, 0.8305)0.9689(0.9257, 0.9755)0.9428(0.8960, 0.9737)0.7965(0.7429, 0.8375)0.9650(0.9150, 0.9797)0.7860(0.7436, 0.8160)0.9705(0.9272, 0.9772)0.9034(0.8528, 0.9520)HM associated with CN0.6867(0.6486, 0.7280)0.9832(0.9416, 0.9940)0.9544(0.9109, 0.9617)0.8143(0.7694, 0.8728)0.9669(0.9465, 0.9759)0.7451(0.6978, 0.7964)0.9589(0.9348, 0.9706)0.8350(0.7926, 0.8748)HM associated with glaucoma0.6857(0.6513, 0.7428)0.9915(0.9546, 1.0000)0.9790(0.9439, 0.9847)0.7742(0.7180, 0.8307)0.9867(0.9478, 0.9974)0.7273(0.6941, 0.7578)0.9475(0.8976, 0.9788)0.7795(0.7436, 0.8299)HM associated with RD0.7742(0.7265, 0.8065)0.9849(0.9517, 0.9929)0.9696(0.9276, 0.9790)0.8000(0.7568, 0.8489)0.9824(0.9523, 0.9894)0.7869(0.7352, 0.8387)0.9599(0.9214, 0.9746)0.8488(0.7957, 0.8823)Mean0.7764(0.6894, 0.8633)0.9602(0.9187, 1.0000)0.9420(0.9065, 0.9775)0.8115(0.7808, 0.8422)0.9622(0.9391, 0.9853)0.7923(0.7364, 0.8482)0.9649(0.9528, 0.9770)0.8812(0.8045, 0.9579)Validationcohort 1HM0.7229(0.6743, 0.7656)0.8702(0.8292, 0.9084)0.8298(0.7907, 0.8767)0.6779(0.6468, 0.7272)0.8926(0.8541, 0.9308)0.6997(0.6500, 0.7546)0.9021(0.8543, 0.9505)0.8288(0.7767, 0.8682)HM associated with cataract0.8482(0.7973, 0.9016)0.7610(0.7012, 0.8011)0.7930(0.7330, 0.8493)0.6735(0.6156, 0.7194)0.8961(0.8361, 0.9302)0.7508(0.6975, 0.7876)0.9133(0.8626, 0.9640)0.8746(0.8337, 0.9333)HM associated with MD0.5888(0.5396, 0.6430)0.9864(0.9422, 0.9943)0.9359(0.8910, 0.9891)0.8630(0.8088, 0.9175)0.9428(0.8843, 0.9950)0.7000(0.6619, 0.7469)0.8951(0.8505, 0.9437)0.7497(0.7118, 0.7872)HM associated with CN0.5840(0.5208, 0.6374)0.9886(0.9546, 0.9958)0.9505(0.9256, 0.9616)0.8424(0.7994, 0.8731)0.9581(0.9261, 0.9640)0.6898(0.6320, 0.7355)0.8948(0.8629, 0.9390)0.7268(0.6859, 0.7643)HM associated with glaucoma0.7481(0.6910, 0.7890)0.9996(0.9757, 1.0000)0.9861(0.9379, 0.9954)0.9902(0.9665, 0.9974)0.9860(0.9415, 0.9964)0.8523(0.8130, 0.8934)0.9324(0.8739, 0.9649)0.8028(0.7624, 0.8561)HM associated with RD0.5592(0.5028, 0.6036)0.9940(0.9655, 1.0000)0.9577(0.9256, 0.9655)0.8939(0.8424, 0.9381)0.9612(0.9286, 0.9747)0.6880(0.6354, 0.7293)0.8745(0.8319, 0.9193)0.6958(0.6650, 0.7501)Mean0.6752(0.5539, 0.7965)0.9333(0.8309, 1.0000)0.9088(0.8269, 0.9908)0.8235(0.6921, 0.9549)0.9395(0.9000, 0.9789)0.7301(0.6628, 0.7974)0.9020(0.8815, 0.9225)0.7797(0.7091, 0.8504)Validationcohort 2HM0.7456(0.7152, 0.7969)0.8774(0.8188, 0.9281)0.8389(0.7816, 0.8880)0.7149(0.6720, 0.7737)0.8932(0.8408, 0.9397)0.7299(0.6953, 0.7731)0.9064(0.8604, 0.9370)0.8438(0.7902, 0.8797)HM associated with cataract0.8551(0.8099, 0.9115)0.7828(0.7458, 0.8198)0.8086(0.7669, 0.8517)0.6867(0.6497, 0.7418)0.9065(0.8732, 0.9464)0.7617(0.7264, 0.8107)0.9164(0.8678, 0.9569)0.8766(0.8318, 0.9290)HM associated with MD0.6007(0.5342, 0.6616)0.9812(0.9437, 0.9924)0.9330(0.8846, 0.9756)0.8224(0.7896, 0.8766)0.9443(0.8867, 0.9784)0.6943(0.6536, 0.7296)0.8988(0.8520, 0.9389)0.7511(0.7140, 0.8046)HM associated with CN0.5696(0.5041, 0.6110)0.9894(0.9620, 0.9975)0.9477(0.9146, 1.0000)0.8562(0.8169, 0.9096)0.9542(0.9064, 0.9600)0.6841(0.6251, 0.7152)0.8752(0.8167, 0.9216)0.7010(0.6427, 0.7384)HM associated with glaucoma0.7890(0.7427, 0.8284)0.9995(0.9667, 1.0000)0.9896(0.9679, 1.0000)0.9885(0.9388, 0.9981)0.9897(0.9671, 0.9951)0.8776(0.8404, 0.9104)0.9351(0.8876, 0.9651)0.8257(0.7719, 0.8682)HM associated with RD0.6145(0.5490, 0.6812)0.9930(0.9463, 1.0000)0.9637(0.9393, 0.9781)0.8800(0.8447, 0.9207)0.9685(0.9309, 0.9813)0.7237(0.6732, 0.7553)0.8743(0.8440, 0.9277)0.7063(0.6683, 0.7636)Mean0.6957(0.5733, 0.8182)0.9372(0.8444, 1.0000)0.9136(0.8373, 0.9899)0.8248(0.7079, 0.9416)0.9427(0.9041, 0.9813)0.7452(0.6713, 0.8192)0.9010(0.8761, 0.9259)0.7841(0.7057, 0.8625)HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, CI confidence interval, PPV positive predictive value, NPV negative predictive value; AUC area under the receiver operating characteristic curve, PR-AUC area under the precision-recall curve, RF random forest, F1 the harmonic mean of precision and recall.
We subsequently performed a subgroup analysis after excluding patients with active systemic inflammatory conditions (n = 126), recent infections (n = 240), immunosuppressive medication use (n = 64), diabetes (n = 277), or hypertension (n = 455), conditions that could affect haematological biomarkers, to assess the performance of the final RF model.
The final RF model achieved a macroaverage AUC of 0.9560 (95% CI: 0.9236–0.9884), which was not significantly different from that of the model including all patients (DeLong test, P = 0.056). As shown in Fig. S5A, the RF model achieved AUCs of 0.9701 (95% CI = 0.9377–1.0000), 0.9751 (95% CI = 0.9427–1.0000), 0.9685 (95% CI = 0.9361–1.0000), 0.9536 (95% CI = 0.9212–0.9860), 0.9236 (95% CI = 0.8912–0.9560), and 0.9423 (95% CI = 0.9099–0.9747) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. Other performance metrics, including sensitivity, specificity, PPV, NPV, accuracy, and F1 score, are summarised in Table S7, with macroaverage values of 0.7639, 0.9557, 0.9229, 0.9625, 0.9250, and 0.8250, respectively.
Models developed in the discovery cohort require validation with independent validation cohorts to ensure robustness. Therefore, two external validation cohorts were used to evaluate the performance of the 9-feature RF model. In validation cohort 1, the RF model achieved a macroaverage AUC of 0.9020 (95% CI = 0.8815–0.9225; Fig. 4A) and an PR-AUC of 0.7797 (95% CI = 0.7091–0.8504; Fig. 4B). As shown in Fig. 4A, the RF model achieved AUCs of 0.9021 (95% CI = 0.8543–0.9505), 0.9133 (95% CI = 0.8626–0.9640), 0.8951 (95% CI = 0.8505–0.9437), 0.8948 (95% CI = 0.8629–0.9390), 0.9324 (95% CI = 0.8739–0.9649), and 0.8745 (95% CI = 0.8319–0.9193) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. The confusion matrix (Fig. 4C) further illustrates the model’s performance, revealing sensitivities of 0.7229 (95% CI = 0.6743–0.7656), 0.8482 (95% CI = 0.7973–0.9016), 0.5888 (95% CI = 0.5396–0.6430), 0.5840 (95% CI = 0.5208–0.6374), 0.7481 (95% CI = 0.6910–0.7890), and 0.5592 (95% CI = 0.5028–0.6036) and specificities of 0.8702 (95% CI = 0.8292–0.9084), 0.7610 (95% CI = 0.7012–0.8011), 0.9864 (95% CI = 0.9422–0.9943), 0.9886 (95% CI = 0.9546–0.9958), 0.9996 (95% CI = 0.9757–1.0000), and 0.9940 (95% CI = 0.9655–1.0000) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively.Fig. 4External validation of the final RF model in validation cohorts 1 and 2.A ROC curves of the RF model based on the top nine features for screening high myopia-related complications in validation cohort 1. B PR-AUC curves of the RF model based on the top nine features in validation cohort 1. C Confusion matrix heatmap illustrating the accuracy, sensitivity, and specificity of the predictions from the RF model based on the nine selected features in validation cohort 1. D ROC curves of the RF model based on the top nine features for screening high myopia-related complications in validation cohort 2. E PR-AUC curves of the RF model based on the top nine features in validation cohort 2. F Confusion matrix heatmap illustrating the accuracy, sensitivity, and specificity of the predictions from the RF model based on the nine selected features in validation cohort 2. G A comparison of multiple machine learning algorithms was conducted in validation cohort 1 using key performance metrics, including the AUC, sensitivity, specificity, PPV, NPV, accuracy, and F1 score. Except for the final RF model, the performance of all the other models was analysed using the complete set of blood test variables as input features. Source data are provided as a Source Data file. H A comparison of multiple machine learning algorithms was conducted in validation cohort 2 using key performance metrics, including the AUC, sensitivity, specificity, PPV, NPV, accuracy, and F1 score. Except for the final RF model, the performance of all the other models was analysed using the complete set of blood test variables as input features. Source data are provided as a Source Data file. HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, PPV positive predictive value, NPV negative predictive value, AUC area under the receiver operating characteristic curve, ROC receiver operating characteristic curve, PR-AUC area under the precision‒recall curve, RF random forest, DT decision tree, GB gradient boosting, KNN k-nearest neighbours, LGBM light gradient boosting machine, LR logistic regression, MLP multilayer perceptron, NB naïve Bayes, XGBoost extreme gradient boosting.
The distributions and correlations of the nine blood test variables in validation cohort 1 are presented in Fig. S3B. Strong correlations were observed between the EOS and EOS% and between BASO and BASO%, whereas no significant correlations were found among the other variables, suggesting limited multicollinearity and supporting the independent contributions of these features to model performance.
Similar results were observed in validation cohort 2, where the RF model achieved a macroaverage AUC of 0.9010 (95% CI = 0.8761–0.9259; Fig. 4D) and an PR-AUC of 0.7841 (95% CI = 0.7057–0.8625; Fig. 4E). As shown in Fig. 4D, the RF model achieved AUCs of 0.9064 (95% CI = 0.8604–0.9370), 0.9164 (95% CI = 0.8678–0.9569), 0.8988 (95% CI = 0.8520–0.9389), 0.8752 (95% CI = 0.8167–0.9216), 0.9351 (95% CI = 0.8876–0.9651), and 0.8743 (95% CI = 0.8440–0.9277) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. The confusion matrix (Fig. 4F) confirmed these findings, with sensitivities of 0.7456 (95% CI = 0.7152–0.7969), 0.8551 (95% CI = 0.8099–0.9115), 0.6007 (95% CI = 0.5342–0.6616), 0.5696 (95% CI = 0.5041–0.6110), 0.7890 (95% CI = 0.7427–0.8284), and 0.6145 (95% CI = 0.5490–0.6812) and specificities of 0.8774 (95% CI = 0.8188–0.9281), 0.7828 (95% CI = 0.7458–0.8198), 0.9812 (95% CI = 0.9437–0.9924), 0.9894 (95% CI = 0.9620–0.9975), 0.9995 (95% CI = 0.9667–1.0000), and 0.9930 (95% CI = 0.9463–1.0000) for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively.
The distributions and correlations of these nine blood test variables in validation cohort 2 are presented in Fig. S3C. Strong correlations were observed between the EOS and EOS% and between the BASO and BASO%, whereas no significant correlations were found among the other variables. A detailed description of the performance, including sensitivity, specificity, PPV, NPV, accuracy, and F1 score, of the final RF model in the validation cohort is provided in Table 1.
A subgroup analysis of validation cohort 1 was performed after excluding patients with active systemic inflammatory conditions (n = 81), recent infections (n = 144), immunosuppressive medication use (n = 25), diabetes (n = 222), or hypertension (n = 367) to assess the performance of the final RF model. The final RF model achieved a macroaverage AUC of 0.9010, which was not significantly different from that of the model including all patients (DeLong test, P = 0.868). As shown in Fig. S5B, the RF model achieved AUCs of 0.9064, 0.9164, 0.8988, 0.8752, 0.9351, and 0.8743 for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. Other performance metrics, including sensitivity, specificity, PPV, NPV, accuracy, and F1 score, are summarised in Table S8, with macroaverage values of 0.6758, 0.9327, 0.8407, 0.9398, 0.7279, and 0.7359, respectively.
Furthermore, a subgroup analysis of validation cohort 2 was performed after excluding patients with active systemic inflammatory conditions (n = 33), recent infections (n = 76), immunosuppressive medication use (n = 18), diabetes (n = 109), or hypertension (n = 138) to assess the performance of the final RF model. The final RF model achieved a macroaverage AUC of 0.9008, which was not significantly different from that of the model including all patients (DeLong test, P = 0.976). As shown in Fig. S5C, the RF model achieved AUCs of 0.9042, 0.9135, 0.8890, 0.8833, and 0.9456 for identifying high myopia overall and high myopia associated with cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment, respectively. Other performance metrics, including sensitivity, specificity, PPV, NPV, accuracy, and F1 score, are summarised in Table S9, with macroaverage values of 0.6834, 0.9332, 0.8486, 0.9398, 0.7286, and 0.7429, respectively.
In the sensitivity analysis excluding UA, the only kidney function indicator included in the model, the final RF model maintained robust performance (Table S10), achieving macroaverage AUCs of 0.8589 and 0.8613 in validation cohorts 1 (Fig. S6A) and 2 (Fig. S6B), respectively.
Furthermore, a comparison between the final nine-feature RF model and other machine learning algorithms—constructed using the complete set of blood test variables as input features—was conducted in validation cohorts 1 (Table S11, Fig. S7) and 2 (Table S12, Fig. S8) using key performance metrics. As shown in Fig. 4G, in validation cohort 1, the final RF model outperformed all the other models, including those utilising the full set of blood test variables. Notably, its performance was comparable to that of the conventional RF model trained on the complete feature set. Similar results were observed in validation cohort 2 (Fig. 4H), further confirming the robustness of the final RF model.
Taken together, these findings indicate that the nine-feature RF model demonstrated strong screening performance across all the validation cohorts, consistent with the findings in the discovery cohort.
DCA of the final RF model indicated that it consistently offered net benefits for the discovery cohort and both validation cohorts over a range of threshold probabilities. In the discovery cohort (Fig. S9A), validation cohort 1 (Fig. S9B), and validation cohort 2 (Fig. S9C), the RF model outperformed the “treat none” strategy, indicating its practical utility in clinical decision-making.
The calibration curves further confirmed the model’s robust screening performance. In the discovery cohort (Fig. 5A), the RF model showed good overall calibration, particularly for high myopia and high myopia associated with macular degeneration. High myopia associated with cataracts was slightly overestimated in the mid-to-high probability range, whereas the model tended to underestimate the likelihood of high myopia associated with choroidal neovascularisation, glaucoma, and retinal detachment, as indicated by calibration curves consistently falling below the ideal diagonal line, suggesting a conservative prediction tendency for these categories.Fig. 5Calibration curves for the RF model in the discovery cohort and validation cohorts 1 and 2.A Calibration curve of the RF model for the discovery cohort, showing good calibration for HM and HM associated with MD at higher predicted probabilities (>0.6). The model slightly overestimated HM with cataracts and underestimated HM with choroidal neovascularisation, glaucoma, and RD, as indicated by the curves below the ideal line. B Calibration curve of the RF model for validation cohort 1, with performance similar to that in the discovery cohort. C Calibration curve of the RF model for validation cohort 2, confirming the reliability and generalisability of the model. HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, RF random forest.
Similar calibration performance was observed in validation cohort 1 (Fig. 5B) and validation cohort 2 (Fig. 5C), further supporting the reliability and generalisability of the model across different populations.
Given the challenges clinicians face in adopting diagnostic models with limited interpretability, we employed SHAP to elucidate the decision-making process of our final model by quantifying the contribution of each feature to the model predictions. The SHAP summary plot (Fig. 6A) presents the relative importance values of the nine selected features based on their mean absolute SHAP values, providing a global interpretation of model behaviour. SHAP dependence plots were generated to illustrate how individual features influence the model output and to further improve the interpretability. Each plot depicts the relationship between the actual values of the features and their corresponding SHAP values.Fig. 6Analysis of the clinical utility of the model.A SHAP summary plot depicting the relative importance of the nine selected features based on their mean absolute SHAP values, providing a global understanding of the model’s behaviour. B SHAP dependence plots showing the influence of individual features on the model’s predictions. We developed a freely accessible online web application to support clinical integration, which is available at https://high-myopia.streamlit.app/. This interactive platform allows clinicians to enter the nine required feature values and then automatically calculates and displays the predicted risk based on the random forest model. HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, SHAP SHapley Additive exPlanations, UA uric acid, EOS% eosinophil percentage, TG triglycerides, EOS eosinophils, LPB apolipoprotein B, M monocytes, PCT plateletcrit, BASO% basophil percentage, BASO basophils.
We developed a freely accessible online web application to facilitate seamless integration into clinical practice (Fig. 6B), which is available at https://high-myopia.streamlit.app/. This interactive platform allows clinicians to input the nine required feature values, after which the application automatically computes and displays the predicted risk based on the RF model.
A total of 5067 participants with high myopia were enrolled in the hospital-based prospective cohort study to screen for high myopia-related complications (Fig. 7A). Of these participants, 4572 met the eligibility criteria and were assessed in real time with the online web application developed for assessing the risk of high myopia-related complications. According to the screening results of the model, 1523 individuals were classified as high risk. Among them, 127 declined to cooperate, and 84 withdrew voluntarily, resulting in 211 individuals not being referred for further ophthalmic evaluation.Fig. 7External validation of the RF screening model in hospital-based and community-based cohorts.A Screening workflow of the hospital-based prospective cohort study. B Confusion matrix heatmap depicting the prediction performance (accuracy, sensitivity, and specificity) of the RF model in the hospital-based cohort. C Summary of the prediction metrics (accuracy, sensitivity, specificity, PPV, and NPV) of the RF model in the hospital-based cohort. Source data are provided as a Source Data file. D Screening workflow of the community-based cross-sectional study. HM high myopia, RD retinal detachment, MD macular degeneration, CN choroidal neovascularisation, PPV positive predictive value, NPV negative predictive value, RF random forest.
Among them, 1312 high-risk participants were referred to the Department of Ophthalmology for further evaluation, resulting in the confirmation of 938 cases of high myopia-related complications. An additional 2270 individuals identified as being at low risk of complications were also referred for further ophthalmic evaluation, and 347 cases of high myopia-related complications were identified at EENT Hospital.
Figure 7B presents the confusion matrix of the final RF model for detecting the five high myopia-related complications. The model achieved a mean sensitivity of 70.0% (95% CI = 66.43%–74.22%), a specificity of 94% (95% CI = 90.18%–98.22%), a PPV of 86% (95% CI = 82.33%–89.87%), and an NPV of 94% (95% CI = 91.60%–97.04%), as detailed in Fig. 7C.
The final RF model yielded false-positive rates of 14.22% for cataracts, 0.23% for macular degeneration, 0.52% for choroidal neovascularisation, 0% for glaucoma, and 0.18% for retinal detachment, indicating a low proportion of false alarms. In terms of downstream costs, the higher false-positive rate for cataracts may lead to some additional routine ophthalmologic visits, while the very low false-positive rates for macular degeneration, choroidal neovascularisation, and retinal detachment minimise unnecessary procedures. Overall, the model helps allocate ophthalmic resources efficiently, focusing on confirmatory evaluations of high-risk individuals and limiting unnecessary healthcare expenditure.
A total of 311,254 participants were included in the community-based cross-sectional study, among whom 1905 were identified as having high myopia (Fig. 7D). Of these, 1894 met the inclusion criteria and were assessed in real time with the online web application. According to the predictions of the RF model, 267 individuals were classified as high risk.
A total of 210 high-risk individuals were subsequently referred to the Department of Ophthalmology for further clinical evaluation, which confirmed 148 cases of high myopia-related complications. Owing to the cross-sectional nature of the study, follow-up validation was not performed for participants in the low-risk group.
The final RF model had favourable PPVs (Fig. 7D) in detecting the five high myopia-related complications, with the highest PPV observed for the prediction of macular degeneration (0.795), followed by glaucoma (0.75), retinal detachment (0.75), choroidal neovascularisation (0.733), and cataracts (0.65), indicating strong performance in identifying true positive cases within the high myopia population.
In this study, we developed and validated a machine learning framework leveraging routine blood test data to enable opportunistic screening for vision-threatening complications of high myopia during routine health examinations, thereby addressing critical limitations of current imaging-based diagnostic paradigms. Our key innovation lies in showing that a readily obtainable set of nine routine blood biomarkers—EOS, PCT, BASO%, M, UA, LPB, BASO, EOS%, and TG levels—can predict five major complications with robust accuracy and the validation of the model through a hospital-based prospective cohort study and a community-based cross-sectional study. These findings advance ocular risk stratification by establishing haematological features as accessible proxies for microstructural pathology, particularly in resource-limited settings.
Routine blood indices, biochemical indices, and coagulation indices can reflect a wide range of physiological and pathological states in the body^23,25^, providing information on aspects such as inflammation, blood circulation, metabolic status, and tissue injury. In this study, the selected biomarkers were associated with known mechanisms of myopic complications. The inclusion of eosinophil and basophil counts and percentages aligns with recent evidence showing that localised inflammation drives retinal detachment and choroidal neovascularisation through IL-4/IL-13-mediated vascular leakage^26–28^. IL-4 and IL-13, key Th2-type cytokines, play critical roles in mediating localised inflammation in the eye. They promote the recruitment and activation of eosinophils and basophils, increasing vascular permeability, which may contribute to the development of retinal detachment and choroidal neovascularisation. The evidence suggests that these cytokines induce the endothelial expression of adhesion molecules, facilitating leucocyte adhesion and migration and thereby amplifying inflammatory responses in the retina and choroid. Importantly, targeting IL-4/IL-13 signalling with anti-inflammatory therapies, such as monoclonal antibodies against IL-4 or IL-13, could reduce local inflammation and vascular leakage, decreasing the risk of disease progression^29,30^. While preclinical studies support this approach, the safety and efficacy of such interventions in clinical settings remain to be established, and future prospective studies are warranted to evaluate their therapeutic potential.
UA, a potent antioxidant, paradoxically showed a protective association in our model, although this finding is consistent with its dual role in neutralising oxidative stress in early macular degeneration while exacerbating inflammation at higher concentrations^31^. In this study, we also revealed that abnormalities in triglyceride and lipoprotein levels were associated with complications of high myopia. Systemic dyslipidaemia-induced blood lipid deposition plays a critical role in the pathogenesis of macular degeneration, where lipid dysregulation promotes lipid accumulation, oxidative stress, and chronic inflammation in Bruch’s membrane, ultimately leading to retinal pigment epithelial dysfunction and disease progression^32,33^. The prominence of the thrombocytocrit and the level of lipoprotein B corroborates the roles of thrombotic and dysregulated lipid pathways in the progression of myopic glaucoma^34^. This biological agreement between the identified biomarkers and known mechanisms strengthens the clinical plausibility of the model, suggesting that systemic shifts in haematometabolic parameters may precede detectable fundus changes by months to years.
The utility of routine blood biomarkers for screening high myopia-related complications is further supported by their inherent stability and established systemic‒ocular interplay. First, routine blood parameters exhibit high intraindividual consistency over time, as evidenced by longitudinal cohort studies showing a <15% coefficient of variation for haematological indices such as the eosinophil count and thrombocytocrit across intervals of years^23^. This stability allows reliable risk stratification, in contrast to the transient imaging artefacts detectable in OCT scans. This systemic–ocular axis operates bidirectionally—not only do blood markers reflect ocular pathology but retinal imaging itself can also predict certain systemic conditions^35–40^. Babenko et al.^36^ reported strong correlations between ocular imaging features and peripheral blood parameters; specifically, conjunctival vascular density, scleral hue, and subtle haemorrhagic manifestations, including fluctuations in haemoglobin levels, platelet counts, and inflammatory markers, were shown to reflect systemic physiological changes. Cheung et al.^41^ developed fully automated artificial intelligence and deep learning-based software for analysing retinal vessels, revealing that a narrower retinal arteriolar calibre was significantly associated with elevated blood pressure, whereas a wider venular calibre was associated with a higher body mass index, elevated haemoglobin A1c levels, and current smoking status. Moreover, algorithms based on digital colour fundus photography have shown good performance in predicting hypertension, hyperglycaemia, and dyslipidaemia, with AUCs of 0.766, 0.880 and 0.703, respectively^42^. These ocular features may thus serve as noninvasive surrogates for underlying haematological or inflammatory conditions, underscoring the potential of integrated systemic–ocular diagnostic approaches. Taken together, this bidirectional crosstalk suggests that blood-based screening could serve not as a replacement for retinal exams but as a scalable first-line triage tool, particularly in regions lacking imaging infrastructure.
Our blood-based opportunistic screening during routine health examinations framework offers notable advantages over AI-driven fundus imaging for detecting high myopia-related complications, particularly in resource-limited settings. Although deep learning models based on retinal photographs have achieved strong diagnostic performance (AUCs of 0.89–0.95 for myopic maculopathy), their widespread implementation is constrained by the need for high-resolution imaging equipment, which is unavailable in more than 90% of rural clinics in low-income countries, and the requirement for trained personnel. These challenges are further compounded by the frequent need for pharmacologic pupil dilation in highly myopic eyes with peripheral retinal pathology. In contrast, routine blood testing is almost universally accessible, with 98% of primary care facilities worldwide equipped to perform basic haematological assessments. Moreover, blood collection requires no ophthalmic expertise and remains feasible in conditions that compromise fundus imaging, such as media opacities (e.g., cataracts) or severe fundus distortions associated with high myopia.
Nevertheless, our model had a nonnegligible false-positive rate, suggesting that its optimal role is as a prescreening tool in community clinics, primary care facilities, or health examination centres to identify individuals at high risk who should undergo a confirmatory ophthalmic evaluation. In particular, within health examination centres, individuals with high myopia could be screened at no additional cost or resource burden, since the required haematological parameters are already included in standard blood test panels. By leveraging routinely collected data, our model enables efficient risk stratification for multiple high myopia-related complications, thereby reducing missed diagnoses and supporting the more rational allocation of limited ophthalmic resources.
Cost considerations further support the practical utility of our model. In health examination centres, the haematological parameters required are already incorporated into standard blood panels, which means that our approach can identify individuals at risk of multiple high myopia-related complications without introducing any additional expense. Similarly, in community and primary care settings, patients often undergo routine blood testing for nonophthalmic conditions, and incorporating these results into our model allows opportunistic case-finding at no extra cost. When equipment amortisation and personnel expenditures are accounted for, the median cost per blood-based screening is approximately ¥300 (≈$43 or €37), which is considerably lower than that of most ophthalmic imaging modalities. Beyond ophthalmic risk prediction, blood tests also provide concurrent systemic health information (e.g., uric acid levels for gout and triglyceride levels for dyslipidaemia), thus offering dual-purpose clinical benefits that cannot be achieved through fundus imaging alone.
Moreover, we acknowledge the rapid development of point-of-care ophthalmic imaging technologies. Smartphone-based fundus photography and portable fundus cameras, for example, have emerged as cost-effective, portable, and accessible methods for the early detection of retinal diseases such as diabetic retinopathy, age-related macular degeneration, and retinal vascular disorders, significantly broadening access to diagnostic services^16,17,43^. Rather than competing with these imaging-based modalities, our model is best positioned as a complementary by leveraging routinely collected blood test data, it enables opportunistic, low-cost risk stratification for high myopia-related complications, thereby facilitating targeted referral for a confirmatory ophthalmic evaluation and promoting the more efficient allocation of limited healthcare resources.
This blood-based screening model is specifically designed for opportunistic case-finding during routine health examinations in community and primary care settings where specialised ophthalmic imaging is unavailable. The target population includes asymptomatic adults with high myopia (≤–6.0 D) who do not routinely access ophthalmologic care, estimated to represent approximately 60–70% of individuals with high myopia in Chinese populations^44,45^. By integrating this approach into standard health check-up workflows, early identification of high-risk individuals can be achieved without the need for specialised ophthalmic equipment, thereby facilitating timely referral for confirmatory imaging and intervention.
Although the RF models displayed high overall discriminative performance, the sensitivities for detecting macular degeneration, choroidal neovascularisation, and retinal detachment were relatively low (<0.6). Clinically, missed cases in these categories could delay diagnosis and timely intervention for vision-threatening complications. This discrepancy can be explained by several factors. First, decision thresholds were determined using the Youden index, which aims to balance sensitivity and specificity rather than maximising sensitivity alone. Second, class imbalance existed in the dataset, with some conditions being less frequent than others, which likely reduced the sensitivity for these less common categories. Collectively, these factors explain why the sensitivity did not fully reflect the exceptionally high AUC values. Furthermore, in the community-based cross-sectional study, the detection of 347 complication cases among low-risk individuals indicated a substantial false-negative rate. Recognising this limitation is important for interpreting model performance and for guiding clinical application, particularly when considering potential threshold adjustments or future model refinements to optimise sensitivity without compromising overall accuracy.
The final model yielded a false-positive rate of 14.22% for cataracts, which could translate into a substantial number of unnecessary ophthalmic referrals in large-scale applications. For instance, in a screening population of 10,000 individuals with high myopia, approximately 1422 false-positive referrals would require confirmatory ophthalmic evaluation, potentially increasing healthcare burden and costs. To mitigate this, several optimisation strategies may be considered. First, adjusting decision thresholds for specific complications could improve specificity without markedly compromising sensitivity. Second, implementing a two-stage screening workflow—a blood-based ML pre-screen (Stage 1) followed by a clinical risk score (Stage 2) incorporating established cataract risk factors such as diabetes, prolonged ultraviolet exposure, smoking, and obesity—may further reduce false positives. Individuals exceeding stage-specific thresholds would be referred for confirmatory ophthalmic examination, while those with intermediate risk could undergo short-interval follow-up. Third, restricting model application to older high-myopia populations (≥50 years), in whom cataracts are most prevalent and established clinical referral pathways already exist, could enhance the practical value and cost-effectiveness of opportunistic screening. These refinements should be evaluated in future prospective implementation studies to balance diagnostic sensitivity with healthcare resource utilisation.
Several limitations of this study should be acknowledged. First, although the multicentre case–control study provided a large and diverse dataset for model development, its retrospective and noninterventional nature may introduce selection bias. Second, the hospital-based prospective cohort and community-based cross-sectional studies were conducted exclusively in Asian populations, which may limit the generalisability of our findings to other ethnic groups and geographic regions. External validation is warranted in African, European, Latin American, and other ancestry groups, as haematological parameters and the prevalence of myopia-related complications may vary across populations. Third, the cross-sectional design of the community-based cohort precludes the assessment of temporal or causal relationships between routine blood test markers and high myopia-related complications. Furthermore, the fact that only participants who screened positive underwent confirmatory ophthalmic examinations, while model-negative individuals were not systematically assessed, introduces a potential verification bias that may lead to an overestimation of the sensitivity, specificity, and predictive values. In particular, the lack of systematic evaluation among model-negative participants prevents an accurate estimation of the false negative and negative predictive values, which are essential for assessing the clinical utility of the model. Fourth, the proportion of patients with high myopia-related complications was higher than that of patients with uncomplicated high myopia in both the discovery and validation cohorts, reflecting the tertiary referral nature of the participating hospitals; although a class–SMOTE strategy was applied during model training and the model performed well in both hospital- and community-based settings, the overrepresentation of complication cases may still limit the generalisability to populations with lower prevalence. Fifth, while our model displayed promising predictive performance, we did not conduct a formal cost-effectiveness or health economic analysis to quantify potential clinical or economic benefits. Sixth, although we incorporated multiple haematological, metabolic, and inflammatory markers, other potentially relevant biomarkers were not included, and future studies integrating multiomics data may further improve the predictive accuracy. Finally, a key limitation is the inability to determine whether the model preferentially detects early-stage or advanced disease, which has implications for clinical utility and screening timing. Theoretically, preferential detection of early-stage disease would provide greater clinical benefit by enabling timely intervention to prevent irreversible visual loss, whereas uniform performance across all stages would enhance population-level case identification. Future prospective studies with detailed staging and longitudinal follow-up are needed to define the stage-specific performance and optimal application window of this blood-based model.
Our framework provides a promising strategy for opportunistic screening of high myopia-related complications during routine health examinations, serving as a practical first-line approach in health examination centres and community primary care settings. With performance approaching that of specialised imaging models, this strategy may be particularly valuable in low-resource settings, where the burden of preventable blindness is disproportionately high. Future integration with telemedicine platforms and portable diagnostic tools could further support the development of more accessible and scalable presymptomatic care models, contributing greatly to global efforts to mitigate myopia-related visual impairment.
Future studies should explore the integration of blood-based biomarkers with smartphone-assisted fundus imaging, potentially establishing a cost-effective, two-tiered screening strategy suitable for wide-scale deployment. Moreover, mechanistic studies are needed to determine whether altering the levels of the identified biomarkers—such as the use of fibrates to reduce triglyceride levels—can mitigate the risk of myopic complications, thereby opening new therapeutic avenues.
This study received ethical approval from the institutional review boards of the participating Xuhui Central Hospital (Approval No. 2025-017), the Eye & ENT Hospital of Fudan University (Approval No. 2025124), Wanbei Coal–Electricity Group General Hospital (Approval No. WBZY-LLWYH-2025-020), and the First Affiliated Hospital of Anhui Medical University (Approval No. PJ2025-06-34). This study was conducted in accordance with the Declaration of Helsinki and the Ethical Guidelines for Medical and Health Research Involving Human Subjects. This study was conducted in three phases. First, a multicentre retrospective case–control study was performed to develop and internally and externally validate the machine learning-based model using routine blood test data from patients with high myopia, both with and without complications. Second, the clinical performance of the model was prospectively validated in a hospital-based cohort. Finally, the model was further evaluated in a community-based cross-sectional study to examine its utility for population-level screening. Given that the multicentre case–control study and community-based cross-sectional study were noninterventional and retrospective in nature, the institutional review boards waived the requirement for informed consent. In the prospective, hospital-based cohort, all participants provided written informed consent prior to sample collection and study enrolment.
A multicentre, case‒control study was conducted across five hospitals in China between January 1, 2016, and December 30, 2024. Patients from the Eye & ENT Hospital of Fudan University composed the discovery cohort. The first validation cohort consisted of patients from Xuhui Central Hospital of Fudan University and Putuo District Central Hospital, and the second validation cohort included patients from Wanbei Coal–Electricity Group General Hospital and The First Affiliated Hospital of Anhui Medical University. The overall study design is illustrated in Fig. 1. Sex of participants was recorded as male or female according to hospital registration data, which were based on self-reported information at the time of admission. As sex was not directly relevant to the study objectives, no sex-based analyses were performed.
Individuals were excluded if they met any of the following criteria to minimise the effects of potential confounding a history of haematologic diseases or cancers; systemic illnesses such as acute infections, metabolic syndrome, or autoimmune diseases; pregnancy or intraocular surgery within the past six months; aged under 18 years; or a lack of available laboratory test results. In the subgroup analyses, patients with active systemic inflammatory conditions, recent infections, immunosuppressive medication use, diabetes, or hypertension were excluded to further validate the robustness of the model. This information was collected through a review of medical records and laboratory data, and individuals meeting one or more criteria were removed from the analysis prior to model development.
During the multicentre, case–control study phase, all patients with high myopia who visited one of the five participating hospitals between January 1, 2016, and December 30, 2024, were included, without sampling. The discovery cohort comprised 1464 patients with uncomplicated high myopia and patients with high myopia who were divided into the following complication-specific 1914 patients with cataracts, 698 with macular degeneration, 478 with choroidal neovascularisation, 185 with glaucoma, and 363 with retinal detachment. External validation cohort 1 included 937 patients with uncomplicated high myopia, 986 patients with high myopia complicated with cataracts, 463 patients with high myopia complicated with macular degeneration, 335 patients with high myopia complicated with choroidal neovascularisation, 114 patients with high myopia complicated with glaucoma, and 290 patients with high myopia complicated with retinal detachment. External validation cohort 2 included 787 patients with uncomplicated high myopia and with high myopia presenting with the following 771 patients with cataracts, 255 with macular degeneration, 192 with choroidal neovascularisation, 130 with glaucoma, and 299 with retinal detachment. No participants had two or more of these five high myopia complications. Venous blood samples were collected from patients newly diagnosed with high myopia or high myopia-related complications within 24 h of presentation.
The performance of the screening model was validated through two large-sample cohort a hospital-based prospective cohort study and a community-based cross-sectional study. A prospective screening programme was initiated at the Eye & ENT Hospital of Fudan University and Xuhui Central Hospital of Fudan University. Briefly, all patients with high myopia who presented to the hospital from January 2025 to May 2025 were sequentially enrolled, with approximately 5000 patients recruited during the study period. Routine blood tests were performed for all eligible participants. Individuals who were identified as positive for high myopia-related complications according to the established screening model were further evaluated with ophthalmic examinations to confirm the presence of these complications. Furthermore, individuals identified as negative for high myopia-related complications were further assessed through ophthalmic examinations.
A community-based cross-sectional screening programme was launched in Xuhui District, Shanghai, in July 2024^46^. Residents of Xuhui District with high myopia were invited to participate through two local community health service centres, the Changqiao Community Health Service Centre and the Caohejing Community Health Service Centre, with whom we have maintained strong collaborative relationships over the past decade. Interested individuals were screened for eligibility according to predefined inclusion and exclusion criteria. Eligible participants were provided with detailed information about the study objectives, procedures, and potential risks. Approximately 300,000 residents of Xuhui District were included. Routine blood tests were performed for all eligible participants as part of their routine physical examinations; those who were identified as positive for high myopia-related complications according to the model were further evaluated and confirmed to be positive through ophthalmic examinations conducted at the Eye & ENT Hospital of Fudan University or Xuhui Central Hospital of Fudan University. Furthermore, individuals who were identified as negative for high myopia-related complications were not further evaluated with ophthalmic examinations.
The inclusion criteria were as no prior medical history of high myopia-related complications; ability to understand, sign, and date a written informed consent form; good physical condition; and clarity of consciousness.
The exclusion criteria were as a history of haematologic diseases or any form of cancer; pregnancy or intraocular surgery within the past six months; and aged under 18 years.
The study exit criteria were as development of serious comorbid conditions such as significant cardiovascular, hepatic, or renal diseases or malignancies; noncooperation during follow-up diagnostic confirmation for high myopia-related complications; voluntary withdrawal for any reason; and death from any cause.
The procedures for the detailed ophthalmic examinations have been described previously^21,47,48^. All the enrolled patients underwent comprehensive eye evaluations, which included intraocular pressure measurements, slit-lamp biomicroscopy, optical coherence tomography, B-scan ultrasonography, ultrawide fundus photography, ultra-widefield autofluorescence, and fluorescein angiography. In our study, refractive error was measured using cycloplegic refraction, and high myopia was defined as a refractive error of ≤−6.0 dioptres. Uncomplicated high myopia was defined as high myopia in the absence of related complications, including cataracts, myopic macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment.
High myopia-related complications were identified based on the following standard ophthalmologic diagnostic
Cataract—diagnosed by slit-lamp biomicroscopy, defined as lens opacity graded according to the Lens Opacities Classification System III and confirmed to cause a best-corrected visual acuity reduction not attributable to other ocular diseases;
Macular degeneration—diagnosed by fundus photography and/or OCT according to the Meta-Analysis for Pathologic Myopia classification, including diffuse chorioretinal atrophy, patchy atrophy, and macular atrophy;
Choroidal neovascularisation—confirmed by OCT and fluorescein angiography or OCT angiography, showing subretinal hyperreflective material, exudation, and leakage consistent with neovascular membranes secondary to pathological myopia;
Glaucoma—diagnosed by a combination of characteristic optic nerve head changes (increased cup-to-disc ratio and rim thinning), corresponding visual field defects on standard automated perimetry, and intraocular pressure measurements, according to the European Glaucoma Society guidelines;
Retinal detachment—confirmed by a dilated fundus examination and/or ultrasonography, showing separation of the neurosensory retina from the retinal pigment epithelium, with retinal breaks identified in most cases.
All diagnoses were made by board-certified ophthalmologists at the participating centres, and patients with secondary causes unrelated to high myopia were excluded.
For this multicentre cohort study, routine clinical laboratory data were extracted from the patients’ electronic medical records. Each institution’s principal investigator collected information on 61 laboratory parameters per patient, which were classified into three major blood cell analysis parameters, biochemical analysis parameters, and coagulation analysis parameters.
At presentation, ophthalmologists first evaluated whether patients had any prior diagnosis or treatment for high myopia-related complications. Those with previously diagnosed complications or a history of treatment for such conditions were excluded to ensure that the laboratory measurements reflected the onset of newly identified disease. For patients newly diagnosed with high myopia or high myopia-related complications, venous blood samples were collected within 24 h of presentation. Laboratory tests obtained during stable or follow-up periods were not included in this analysis. Blood samples were obtained via standard venipuncture and analysed at multiple clinical laboratories, including the Eye & ENT Hospital of Fudan University, Putuo District Central Hospital, the First Affiliated Hospital of Anhui Medical University, Shanghai Xuhui Central Hospital, and Wanbei Coal–Electricity Group General Hospital. The procedures for the detailed laboratory analyses have been described previously^18^. The blood cell analyses included 24 variables, the biochemical analyses included 31 variables, and the coagulation analyses included 6 variables, as detailed in Table S13.
The selection of these haematological, biochemical, and coagulation parameters was guided by two main availability and biological plausibility supported by prior research. First, in terms of availability, all 61 parameters are routinely measured in standard clinical practice, which greatly enhances the feasibility and potential clinical utility of our model. Second, regarding biological plausibility, prior evidence supports the relevance of these parameters in ocular risk assessments. Recent studies have demonstrated that routinely available blood test data can be effective at predicting high myopia-related complications, particularly retinal detachment^18,19^. Blood markers of inflammation^49^, glucose levels^50^, and lipid levels^51^ have been linked to an increased risk of ocular diseases, including high myopia and retinal detachment. In addition, several haematological, biochemical, and coagulation parameters have been reported to be associated with the onset or severity of other eye diseases, such as glaucoma^52–54^, cataract^55–57^, and macular degeneration^58,59^. Therefore, based on the prior evidence, we developed a machine learning-based screening model using routine blood test data to identify individuals at risk for high myopia-related complications (cataracts, macular degeneration, choroidal neovascularisation, glaucoma, and retinal detachment).
Interlaboratory and intralaboratory quality control measures were implemented to ensure the reliability and consistency of laboratory results across multiple centres. All the centres participated in external proficiency testing organised by the Clinical Laboratory Centre using standardised survey samples that were analysed locally and evaluated centrally, with all the centres meeting the required criteria. In addition, standardised internal quality control (IQC) materials were measured daily before patient samples were processed to ensure assay precision, and the coefficient of variation (CV) was maintained within 5%. These procedures ensured the accuracy, precision, and comparability of the haematological data across all the study sites.
A total of 61 features (detailed in Table S13) were utilised to develop diagnostic models. For each participant, if more than 20% of the 61 laboratory variables were missing (i.e., 13 variables), the participant was excluded from the study. Conversely, participants with ≤20% missing values (i.e., 12 variables) were retained, and any remaining missing values for these variables were imputed using the medians^60^. Ten ML algorithms were employed to ensure a diverse and robust modelling approach, namely, the RF, AdaBoost, decision tree (DT), gradient boosting (GB), k-nearest neighbours (KNN), light gradient boosting machine (LightGBM, LGBM), logistic regression (LR), multilayer perceptron (MLP), naïve Bayes (NB), and XGBoost algorithms. Multiple classification methods were used to screen for different subtypes of high myopia complications. We applied a Synthetic Minority Oversampling Technique (SMOTE) approach to adjust for an unequal class distribution and mitigate the effect of class imbalance during model development. Each algorithm was selected based on its ability to balance interpretability and predictive power and capture both linear and nonlinear interactions among features. A combination of grid search and manual fine-tuning was implemented to identify the optimal hyperparameters and optimise model performance. The final hyperparameters for each of the ten machine learning models are detailed in Table S14. All modelling procedures were conducted in Python (v3.11), and the corresponding code is available at https://github.com/fudanRenjun/high-myopia/tree/master.
The models were assessed with multiple performance metrics, including precision, recall, PPV, NPV, F1 score, and accuracy. Furthermore, their classification capabilities were evaluated by calculating AUC and PR-AUC. The DeLong test^61^, a nonparametric statistical method, was used to compare AUC values across the different machine learning models. The optimal cut-off value for the models was determined as the value that maximised the Youden index and the sum of sensitivity and specificity. Based on this threshold, the sensitivity, specificity, accuracy, PPV, and NPV were calculated to evaluate model performance^62^. Decision curve analysis^63^ and calibration curves^64^ were used to further assess the performance of the model.
In the discovery cohort, feature selection was performed with the SHAP method^65^. The process refined the predictive model by narrowing the 61 input features to the most critical predictors while maintaining optimal predictive performance. The feature selection process was embedded within cross-validation to avoid information leakage and ensure a fair comparison among the models. The DeLong test^61^ was further applied to compare the AUC values of the models before and after feature selection, validating the impact of this dimensionality reduction process. The SHAP analysis also provides both global interpretability and local interpretability, offering insights into overall model behaviour and individualised predictions.
To facilitate clinical adoption, the final predictive model was deployed as a web-based application (https://high-myopia.streamlit.app/) using the Streamlit framework in Python to facilitate clinical adoption. By inputting the nine selected features (Table S15)—EOS, PCT, BASO%, M, UA, LPB, BASO, EOS%, and TG levels—which are routinely available in typical clinical settings, the clinician receives probability estimates for different high myopia-related complications along with a personalised force plot that visually explains the prediction for individual patients. This interactive tool enhances the interpretability of the model and supports clinical decision-making.
All the statistical analyses were performed using GraphPad Prism (v10) and Python (v3.11). The Kolmogorov–Smirnov test was applied to assess the normality of the data distribution. For continuous variables with a normal distribution, one-way analysis of variance (ANOVA) was used, whereas the Kruskal–Wallis test was used for non-normally distributed data. Categorical variables were analysed using the χ^2^ test, where appropriate. A false discovery rate correction was applied using the Benjamini–Hochberg method to account for multiple comparisons and control for type I error. Effect sizes were also η² (eta squared) for continuous variables from one-way ANOVA and Cramér’s V for categorical variables. Continuous data are presented as the means ± standard deviations (SDs), whereas categorical data are summarised as counts and percentages. A two-tailed P-value < 0.05 was considered to indicate statistical significance.
No statistical method was used to predetermine sample size. For each participant, if more than 20% of the 61 laboratory variables were missing (i.e., 13 variables), the participant was excluded from the study. The experiments were not randomised. The Investigators were not blinded to allocation during experiments and outcome assessment.
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
Supplementary Information Peer Review file Reporting Summary
Source Data