Construction Of Diabetic Retinopathy Database And Research On Predictive Model Based On Machine Learning

Aims: Diabetic retinopathy(DR)is the main cause of irreversible blindness in working-age adults worldwide.Based on machine learning algorithm and large sample database,this topic carried out a series of studies from four aspects of influencing factor analysis,disease prediction,patient follow-up and screening.The first purpose was to determine whether the occurrence of diabetic retinopathy and its related factors are affected by diabetes type(Latent Autoimmune Diabetes in Adults [LADA],type 1 diabetes mellitus [T1DM],and type 2 diabetes mellitus[T2DM]).And then,the main aim was to investigate diabetic retinopathy risk factors and predictive models by machine learning using a large sample dataset,and explore the clinical application of the diabetic retinopathy follow-up management system based on mobile terminals in ophthalmology.The final terminal of the study was to develop a scientific and useful DR clinical screening tool that can be easily popularized for patients with T2 DM.Methods:(1)LADA patients were matched for age(±2 years)and sex to T1DM(1:1)and T2DM(1:2)patients.Retrieved variables included demographic characteristics,diabetes history,laboratory test findings,and history of DR screening,etc.Multiple Logistic regression analysis was applied to identify influencing factors of DR.A decision tree was used to explore interactions between diabetes type and other influencing factors of DR.(2)Information on 32 452 in-patients with type-2 diabetes mellitus were retrieved.The period is from January 1,2013 to December 31,2017.Sixty variables(including demography information,physical and laboratory measurements,system diseases and insulin treatments)were retained for baseline analysis.The optimal 17 variables were selected by recursive feature elimination(RFE).The prediction model was built based on XGBoost algorithm,and it was compared with three other popular machine learning techniques: Logistic Regression,Support Vector Machine,and Random Forest.In order to explain the results of XGBoost model more visually,the SHapley Additive ex Planation(SHAP)method was used.(3)Ophthalmologists designed the database structure according to the characteristics of disease diagnosis and treatment,and cooperated with software engineers to develop a followup management system for diabetic retinopathy.Diabetes patients use the follow-up software to complete the follow-up plan with the help of the doctor.(4)A DR prediction model based on the Logistic regression algorithm was established on the development dataset containing 778 samples.The development dataset was randomly assigned to the internal validation dataset and the training dataset at a ratio of 3:7.The generalization capability of the model was assessed using an external validation dataset containing 128 samples.The DR risk calculator was developed through We Chat Developer Tools using Java Script,which was embedded in the We Chat Mini Program.Results:(1)We included 110 LADA,101 T1 DM,and 220 T2 DM patients.DR prevalence was 26.4%in LADA patients,lower than that in T1DM(50.5%)and T2DM(47.7%)patients(P<0.001).Logistic regression analysis demonstrated that diabetic nephropathy(DN)(OR=42.39,95% CI:10.88–165.11,P<0.001)and diabetes duration(OR=1.15,95% CI: 1.1–1.26,P<0.001)were independent risk factors for DR,and regular DR screening(OR=0.33,95% CI :0.16–0.69,P=0.003)was an independent protective factor.Decision tree analysis showed that in patients without DN with a diabetes duration of at least 10.5 years,T1 DM and LADA patients had a higher incidence of DR than T2 DM patients(72.7% vs.55.1%).(2)DR occurred in 2038(6.28%)T2DM patients.The XGBoost model was identified as the best prediction model with the highest AUC value(0.90)and showed that an Hb A1 c value greater than 8%,nephropathy,a serum creatinine value greater than 100 μmol/L,insulin treatment,and diabetic lower extremity arterial disease(DLEAD)were associated with an increased risk of DR.A patient’s age over 65 was associated with a decreased risk of DR.The force plot of SHAP visualized the risk of DR in each patient.While outputting the prediction results,personalized analysis was made on the effect of the control level of each clinical indicator on the patient’s concurrent DR.(3)Diabetic retinopathy follow-up management system includes six data modules:medical history,personal information,system physical examination,ophthalmology examination,ophthalmology treatment,and system treatment.It has functions such as mobile office,doctor-patient communication,medical record management,follow-up reminder,scientific research management,and security protection.It has been put into clinical application and runs stably.(4)The model revealed risk factors(duration of diabetes,diabetic nephropathy,and creatinine level)and protective factors(annual DR screening and hyperlipidemia)for DR.In the internal and external validation,the recall ratios of the model were 0.92 and 0.89,respectively,and the area under the curve values were 0.82 and 0.70,respectively.Conclusions:(1)The prevalence of DR in diabetes patients was affected by diabetes duration,DN occurrence,and regular DR screening.Diabetes type indirectly affects DR occurrence through its interaction with diabetes duration and DN.Correct LADA diagnosis is necessary,and DR screening needs to be well-implemented.(2)With better comprehensive performance,XGBoost model had high reliability to assess risk indicators of DR.The most critical risk factors of DR and the cutoff of risk factors can be found by SHAP method to render the output of the XGBoost model clinically interpretable.(3)Using mobile terminal-based follow-up system to track and manage diabetic patients can provide reference for clinical diagnosis and treatment decisions of diabetic retinopathy and help patients to identify the disease and effectively prevent it.On the other hand,it can systematically collect clinical data providing high-quality data for scientific research of diabetic retinopathy.Its development has broad application prospects and social benefits.(4)The DR screening tool integrated education,risk prediction,and medical advice function,which could help clinicians in conducting DR risk assessments and providing recommendations for ophthalmic referral to increase the DR screening rate among patients with T2 DM.