Introduction:
Diabetic nephropathy (DN) is the leading cause of end-stage renal disease globally. Early diagnosis and prognosis prediction are crucial to preventing DN progression. Renal biopsy, though effective, is risky, costly and less accessible in remote areas. DN and diabetic retinopathy (DR) are microvascular complications of diabetes, with the retinal microvasculature being the only directly observable microvasculature in the body. This study aims to construct a non-invasive diagnostic and prognostic prediction model using the mixed effects of retinal vascular geometric parameters and clinical data.
Methods:
Fundus images, clinical characteristics, renal biopsy diagnoses, and follow-up data were collected from patients with type 2 diabetes mellitus and chronic kidney disease. Unsupervised learning and ResNet neural networks were used to segment and calculate retinal vascular geometric parameters. Weighted quantile regression (WQS), LASSO, and COX univariate regressions were employed to assess the mixed effects of retinal vascular geometric parameters and select relevant clinical characteristics. Logistic regression and COX-RF were used for model construction.
Results:
We constructed a multimodal database comprising 397 patients from multi-centres. Initially, our unsupervised learning coupled with the Resnet neural network effectively segmented the retinal microvessels. Subsequently, we utilised a mixture of retinal vascular geometric parameters (WQS-diagnosis) and seven clinical characteristics to establish a diagnostic model for DN. The Logistic model demonstrated superior diagnostic performance, achieving an AUC of 0.98, accuracy of 0.92, precision of 0.92, recall of 0.85, and an F1 score of 0.88 on the test set. Notably, the performance of our model surpassed that of diagnostic models constructed using only clinical characteristics (AUC: 0.93) and two retinal vascular geometric parameters plus clinical characteristics (AUC: 0.94). Moreover, we validated the performance of the Logistic model on a multi-centre validation set, where it attained an accuracy of 0.91, precision of 0.90, recall of 0.90, F1 score of 0.90, and an AUC of 0.95. Additionally, a prognostic mixed-effects parameter, WQS-prognosis, was constructed using retinal vascular geometric parameters. For prognostic prediction, the COX-RF model performed best, achieving an AUC of 0.88.
Conclusions:
In summary, this study firstly constructed a non-invasive intelligent diagnostic and prognostic prediction model that combines mixed-effect retinal vascular geometric parameters (WQS) with clinical characteristics, significantly enhancing the accuracy of early diagnosis and prognosis of DN. Additionally, retinal vascular mixed-effect parameters identified in this study can serve as valuable biomarkers for DN diagnosis and prognosis. The results of this research reduce dependence on renal puncture biopsy, thereby lowering patient risk and medical costs. Furthermore, the model offers a feasible diagnostic method for remote areas, significantly improving early intervention and management of patients with DN.
I have no potential conflict of interest to disclose.
I did not use generative AI and AI-assisted technologies in the writing process.