中国中医眼科杂志2025，Vol.35Issue(11)：1017-1023,7.DOI:10.13444/j.cnki.zgzyykzz.2025.11.003
四诊知识图谱与深度学习的DR模型对早期DR的预测应用价值
Application Value of Four Diagnostic Knowledge Graph and Deep Learning-based DR Model in Early Prediction of Diabetic Retinopathy
摘要
Abstract
OBJECTIVE To observe the application value of an objective four diagnostic knowledge graph and a deep learning-based diabetic retinopathy(DR)prediction model in the early diagnosis of DR.METHODS A total of 360 patients with type 2 diabetes from the ophthalmology clinics of the Eye Hospital,China Academy of Chinese Medical Sciences and The First Medical Center of Chinese PLA General Hospital.Four diagnostic information and fundus images were collected.A knowledge graph containing 6,824 nodes and 10,536 relationships was constructed using natural language processing technology,and a multimodal deep learning model was designed for DR prediction.RESULTS(1)Overall model performance evaluation:Significant differences were found in the average reasoning time for DR among the three groups(F=142.863,P=0.000).The reasoning time of the single-modal model group was lower than that of the joint model group and the traditional machine learning group(tJoint=4.234,P=0.002;tTraditional=12.473,P=0.000).The reasoning time of the joint model group was shorter than that of the traditional machine learning group(t=8.954,P=0.000).The differences were all statistically significant.ROC curve analysis for DR diagnostic capability showed that the joint model group had an AUC of 0.926,95%CI(0.901,0.951),a Youden index of 0.738,a sensitivity of 0.925,and a specificity of 0.907.The single-modal model group had an AUC of 0.847,95%CI(0.823,0.871),a Youden index of 0.685,a sensitivity of 0.852,and a specificity of 0.833.The traditional machine learning group had an AUC of 0.832,95%CI(0.807,0.857),a Youden index of 0.648,a sensitivity of 0.833,and a specificity of 0.815.(2)Comparison of early DR identification ability:The accuracy of early DR identification in the joint model diagnosis group(89.53%)was higher than that in the traditional diagnosis group(76.74%),with a statistically significant difference(χ2=4.123,P=0.042).ROC curve analysis showed that the joint model group had an AUC of 0.912,95%CI(0.887,0.937),a Youden index of 0.724,a sensitivity of 0.894,and a specificity of 0.830.The traditional diagnosis group had an AUC of 0.765,95%CI(0.701,0.829),a Youden index of 0.547,a sensitivity of 0.765,and a specificity of 0.782.(3)Clinical application value assessment:The joint model group had a shorter average diagnosis time,higher expert ratings,and lower screening costs compared to the traditional diagnosis group,with statistically significant differences(tdiagnosis time=16.284,texpert rating=5.927,tscreening cost=15.836,all P=0.000).(4)Model stability analysis:Parameter sensitivity analysis showed stable model performance within the learning rate range of 0.0001-0.0100 and batch size range of 16～64.The model demonstrated good robustness to Gaussian noise(σ≤0.1),maintained over 85%accuracy with 20%random feature missing,showed less than 1%performance variation across different GPU platforms,and exhibited no significant performance degradation after continuous operation for 168 hours.CONCLUSIONS The deep integration model of the four diagnostic knowledge graph and fundus images provides a reliable auxiliary tool for early DR screening.As a supplementary method to fundus examination,this approach can offer clinicians a more comprehensive assessment reference.关键词
四诊知识图谱/深度学习/糖尿病视网膜病变/多模态融合/早期预测Key words
four diagnostic knowledge graph/deep learning/diabetic retinopathy/multi-modal fusion/early prediction分类
医药卫生引用本文复制引用
吴星,王瑛,付海英,金鑫,李佳豪,杨永升..四诊知识图谱与深度学习的DR模型对早期DR的预测应用价值[J].中国中医眼科杂志,2025,35(11):1017-1023,7.基金项目
中国中医药科技发展中心中西医协同慢病管理研究项目(CXZH2024179) （CXZH2024179）
