DR is the most common microvascular complication of diabetes mellitus. It has been demonstrated to be a leading cause of preventable blindness in the working-age population in most countries (1).

The American Academy of Ophthalmology (AAO), in 2019, stated that the prevalence of DR among diabetic patients worldwide is about 34.6%. 10.2% (28 million) diabetic patients suffer from vision-threatening DR (2). Effective management of DR requires a deep understanding of the predisposing factors, early diagnosis, and timely therapeutic intervention. Early identification of patients at risk of developing DR is the key to effective intervention, which is significant in reducing the progression of DR and thereby reducing the risk of blindness (3). Moreover, individual patients can be stratified according to different risk levels and get optimal treatment. Due to no typical symptoms in the early stage of the disease, however, most patients with DR may not seek medical evaluation until progression to the proliferative stage, resulting in irreversible visual damage (4). Therefore, methods for accurate prediction of the risk of DR are in urgent need.

At present, several DR risk prediction models based on cross-sectional studies have been developed (5–9). Deep learning algorithms were also applied (10). Although these models can predict the occurrence of DR at the index date, they cannot predict DR occurrence and development of the same patient at designated time points in the future. This will obviously restrict their clinical application. Similarly, based on the clinical characteristics related to the occurrence or development of DR, several models have been developed for optimization of the screening interval in DR screening (11, 12). However, due to the small number of cases in the studies, the proposed models have not been fully validated so far.

On the other hand, several studies investigated the pathogenesis and risk factors of DR to provide guidance for DR management. Epidemiological studies have shown that age, course of diabetes, hemoglobin A1c(HbA1c), fasting blood glucose (FBG), blood pressure, blood lipids, body mass index (BMI), smoking, proteinuria, and several others are all risk factors for DR (13, 14). Among them, duration of diabetes and hyperglycemia were demonstrated as strong risk factors for the occurrence and development of DR (15, 16). However, patients without DR were hardly unusual among those suffering from diabetes for a long time (17). The influences of other factors in DR occurrence also need to be proven. Further studies are required to elucidate the correlation and thus construct standard procedures for the management of this disease.

In this retrospective cohort study, we collected electronic health record data from hospitalized patients with type 2 diabetes and established the prediction model for future risks of DR based on a machine learning (ML) algorithm. To our knowledge, this is the first study to predict the occurrence of DR at each follow-up time point in up to 10 years. We also explore the risk factors that may affect the occurrence of DR and hope this work can provide a basis for further studies concerning the prevention and management of DR.

In this study, we constructed five ML-based models for the risk of developing DR among patients with type 2 diabetes using EHR data, the XGboost model of which demonstrated the best performance. We further evaluated the predictive value of this model and found that it can accurately predict the occurrence of DR at each time point over 10 years. According to feature importance analysis, five risk factors, SUA, LDL-C, TC, eGFR, and TG, were recognized as important indicators of DR for the first time. Moreover, analysis of data from multiple time points revealed that predicting DR risk with the XGBoost model would obviously reduce the amount of time needed for the diagnosis of DR.

XGBoost algorithm is known for its good scalability and high running speed, which is thus commonly used and developed in recent years (28–30). Specifically, the XGBoost model provides the following advantages. First, the XGBoost algorithm reduces the order of magnitude of features while retaining all the features by introducing regularization terms into the objective function to avoid over-fitting. Secondly, the algorithm also used the random forest algorithm for reference, which can not only reduce the over-fitting problem but also reduce the computational complexity. We hypothesized that these reasons would be responsible for the XGBoost algorithm’s outstanding prediction ability for DR risk compared with the other four methods adopted in this study.

There were several studies gaining insights into DR risk prediction, some of which developed ML-based models. However, most of them were cross-sectional studies. In a study in Iran (31), 3734 diabetic patients were included to build a logistic regression model with an AUC of 0.704. The input variables included sex, age, diabetes duration, BMI, blood pressure (BP), HbA1c, FPS, cholesterol, and triglycerides. In the cross-sectional study by Ein hoh et al. (7) in South Korea, a LASSO prediction model was constructed for predicting the risk of DR, reaching an AUC of 0.82 and an accuracy of 75.2%. 490 diabetic patients were selected for the study, and several variables (age, sex, smoking history, drinking history, waist circumference, BMI, physical exercise status, medication history, blood pressure, and relevant laboratory results) were used as input. In another cross-sectional study in Taiwan Province of China (6), 536 patients with type 2 diabetes were selected and 10 predictors, including systolic blood pressure, diastolic blood pressure, BMI, age, gender, duration of diabetes, family history of diabetes, whether there was a blood glucose test, whether there was exercise, and whether there was insulin treatment, were included. Then, four prediction models (SVM, DT, artificial neural network, and LR) were constructed, respectively. The model based on SVM showed the best performance with an accuracy of 79.5% and an AUC of 0.839. Mo et al. (5). included 4170 diabetic patients in a cross-sectional study to predict the risk of DR. Seven input variables, including age, diabetes duration, postprandial blood glucose, HbA1c, urine creatinine, urine microalbumin, and systolic blood pressure were used to construct a multivariate logistic regression model. The model showed moderate predictive ability with an AUC of 0.715 in the validation set. The predictive models mentioned above were developed to identify patients with DR rather than predict the occurrence of DR in the future. Therefore, they can hardly be adapted to the scenarios that require pre-diagnosis screening and prompt intervention. To our knowledge, only one follow-up study was performed for predicting the occurrence of DR (32). The study was conducted with 5034 type 2 diabetes patients not affected by retinopathy at the time of the recruitment and a median follow-up time of 1.2 years. A Cox risk prediction model was constructed with seven predictors including diabetes duration, HbA1c, systolic blood pressure, diastolic blood pressure, proteinuria, creatinine clearance, and diabetes drug treatment, and showed a helpful predictive ability (C-index=0.746). However, this model cannot meet all the clinical requirements of long-term management of type 2 diabetes and associated complications considering that it can only predict the occurrence of DR in the next 1 to 4 years.

The XGBoost model we constructed performed well in different follow-up periods (AUC=0.834-0.966), which indicates that it could predict whether and when a patient with type 2 diabetes will develop DR in the next 1 to 8 years or even over 8 years. Results also show the proposed model is effective to demonstrate in patients who had DR of different severity levels during follow-up. For the individuals with moderate NPDR, the model exhibited the highest performance. As a transitional stage, moderate NPDR can progress and advance to vision-threatening retinopathies. Intervention methods such as intravitreous anti-VEGF treatment can be applied to secure moderate NPDR patients. As the patient progresses to severe NPDR and PDR, it is hard to get the same treatment effect. Moreover, the method requires no extra laboratory tests since the model input was drawn from demographic and clinical characteristics and routine test results. Therefore, the model will facilitate clinicians risk-stratifying patients with type 2 diabetes. For patients at risk of DR, clinicians can provide adequate follow-up management and effective therapeutic intervention. While for the low-risk population, visit frequency will be appropriately reduced, and thus personal and societal healthcare burdens will be reduced accordingly. We further developed nomograms to depict the association between the clinical variables and the probabilities of DR. The nomogram integrating the XGBoost-based machine learning output achieved a higher predictive performance, which provides an intuitive way to interpret the model and shows its potential to be a clinical decision support tool.

To identify the key features contributing to the pre-diagnosis of DR, the importance scores of the risk factors were calculated by XGBoost. According to the feature importance analysis, HbA1c, diabetes duration, age, and FBG were highly ranked, which corresponds to the current clinical perceptions. In addition, we found that SUA, LDL-C, TC, eGFR, and TG may have the potential to be strong predictors for DR. Moreover, we further calculated the importance scores of the risk factors by RF, and found that the results were similar to and complementary to those from XGBoost. In previous studies, Krizova et al. found that an elevated level of vitreous uric acid was closely related to the development of DR, and uric acid concentration was closely related to the degree of DR progression (33). Zhu et al. found that SUA in the vitreous activates the expression of retinal inflammatory factors through the Notch signaling pathway, which will induce oxidative stress and inflammatory response and thus promote the occurrence and development of DR (34). Obesity is a metabolic disease that is associated with insulin resistance and diabetes mellitus (35). In the Wisconsin Epidemiologic Study of Diabetic Retinopathy (WESDR), it was found that high TC levels could increase the risk of retinal hard exudation in patients with type 2 diabetes (36). High LDL-C and high TC/LDL-C levels were associated with retinal macular edema in SN-DREAMS studies (37). Besides, statin lipid-lowering drugs could significantly reduce the risk of DR (38). Since DR is a microvascular complication of diabetes, eGFR can be used as an important biochemical indicator reflecting DR. Chen et al. found that higher eGFR was positively correlated with the risk of DR (39). Therefore, the results presented in our study indicated that these risk factors could be used as early and effective predictors of DR for type 2 diabetes patients before recognizable symptoms appear.

There are still several limitations to this study. First, the prediction models were constructed with the risk factors extracted from EHR, except for those with excessive missing data including the ones that might be related to the occurrence of DR (e.g., hip circumference and neck circumference). Secondly, individualized management approaches for type 2 diabetes and associated complications were not fully described. The monitoring and tracking features were also not collected during the follow-up period. The above information will be recorded and used to optimize the XGBoost model in the future. In addition, this is a single-center retrospective study. We are planning to collect cases from several institutions and conduct a prospective study to validate the generalization ability of the proposed model.

In summary, we developed and evaluated an ML-based model for predicting the risk of DR. The results suggest that it is possible to pre-diagnose DR without fundus images. We also demonstrated the potential of applying XGBoost models to facilitate clinicians in accurately identifying the high-risk population for DR and formulating patient-specific management strategies, thereby reducing the occurrence and development of DR.

The original contributions presented in the study are included in the article/ Supplementary Material . Further inquiries can be directed to the corresponding authors.

The studies involving human participants were reviewed and approved by Ethics Committee of Dalian Central Hospital. The ethics committee waived the requirement of written informed consent for participation.

XHL, ZG, and XYL conceived of and designed the study. YZ contributed to data collection and wrote the manuscript. MD contributed to the conceptualization of the project and edited the manuscript. SL contributed to discussion and revision of the manuscript. HY contributed to model building and wrote the corresponding manuscript. WC contributed to the modeling scheme design and technical support. MZ, PL, and QY contributed to part of manuscript proofreading. XHL, ZG, and XYL are the guarantor of this project and had full access to all of the data in the study and accuracy of the data analysis and reviewed the manuscript. All authors contributed to the article and approved the submitted version.

This work was funded by the National key R&D Program of China (2016YFC0901200), the Dalian Medical Science Research Project Grant Support (2011007).

MD and WC are employed by Infervision.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.