Development and validation of an explainable machine learning model for differentiating diabetic nephropathy from diabetic retinopathy in patients with type 2 diabetes.
Zhang Y, Feng S, et al. • Frontiers in endocrinology • 2026
An explainable XGBoost machine learning model using five routine laboratory features achieved strong discrimination between diabetic nephropathy and diabetic retinopathy in type 2 diabetes patients, with AUC of 0.997 in held-out validation.
Key Findings
Results
XGBoost outperformed six other machine learning algorithms in differentiating diabetic nephropathy from diabetic retinopathy in type 2 diabetes patients.
Seven machine learning algorithms were developed and compared using 47 routinely available laboratory and demographic variables extracted from electronic health records.
The training/internal validation cohort included 2,309 DN cases and 855 DR cases; the held-out validation cohort included 578 DN cases and 214 DR cases.
Recursive feature elimination (RFE) was employed to identify the most informative subset of features and enhance model performance and interpretability.
XGBoost was selected as the final model architecture based on its highest predictive performance among all algorithms evaluated.
Results
The final explainable XGBoost model using five features achieved very high discrimination in the training/internal validation cohort.
AUC = 0.991 (95% CI: 0.989–0.994) in the training/internal validation cohort.
AP (area under the precision-recall curve) = 0.979 (95% CI: 0.973–0.984) in the training/internal validation cohort.
The model was built using only the top five features selected based on importance rankings from recursive feature elimination.
This cohort comprised 2,309 DN cases and 855 DR cases collected from a large tertiary hospital in China.
Results
The explainable XGBoost model demonstrated even stronger discrimination in the independent held-out validation cohort.
AUC = 0.997 (95% CI: 0.996–0.999) in the held-out validation cohort.
AP = 0.993 (95% CI: 0.988–0.997) in the held-out validation cohort.
The held-out validation cohort included 578 DN cases and 214 DR cases.
Performance in the independent cohort exceeded that in the training/internal validation cohort, suggesting strong generalizability.
Results
SHAP analysis identified five laboratory features as the most influential predictors for differentiating diabetic nephropathy from diabetic retinopathy.
The five most influential features identified were: α-hydroxybutyrate dehydrogenase, creatine kinase-MB, creatinine, urinary α1-microglobulin, and N-acetyl-β-D-glucosaminidase.
All five features are routinely available in clinical laboratory settings and were extracted from electronic health records.
SHAP (SHapley Additive exPlanations) values were used to interpret both global feature importance and individual-level predictions.
The use of SHAP analysis was described as addressing 'key concerns regarding transparency and clinical decision-making.'
Methods
The study used a large dataset from a single tertiary hospital in China with a notable class imbalance between diabetic nephropathy and diabetic retinopathy cases.
Total dataset included 2,887 DN cases (2,309 training + 578 validation) and 1,069 DR cases (855 training + 214 validation).
Data were collected from a large tertiary hospital in China and split into training/internal validation and independent held-out validation cohorts.
A total of 47 routinely available laboratory and demographic variables were extracted from electronic health records.
The study design focused on hospitalized patients with type 2 diabetes mellitus.
Conclusions
The authors concluded that an explainable machine learning model for predicting microvascular complications in T2DM patients demonstrated high feasibility and effectiveness with potential for real-world clinical implementation.
The model was described as having 'strong potential to support clinical management and improve patient outcomes.'
Incorporation of SHAP analyses was highlighted as addressing concerns about model transparency.
The model relies solely on routine laboratory data, making it potentially accessible in standard clinical settings.
The authors characterized findings as highlighting 'the model's potential for real-world clinical implementation.'
What This Means
This research suggests that a type of artificial intelligence model called XGBoost can very accurately tell the difference between two common diabetes complications — kidney disease (diabetic nephropathy) and eye disease (diabetic retinopathy) — using only standard blood and urine test results. The model was trained on data from over 3,900 hospitalized diabetes patients in China and tested on a separate group of nearly 800 patients it had never seen before. In that independent test, it was correct about 99.7% of the time (as measured by AUC), which is an unusually high level of accuracy.
The researchers also used a tool called SHAP to explain which lab values drove the model's decisions. The five most important tests were α-hydroxybutyrate dehydrogenase, creatine kinase-MB, creatinine, urinary α1-microglobulin, and N-acetyl-β-D-glucosaminidase — all of which are routinely measured in hospital laboratories. This transparency is important because it allows clinicians to understand why the model made a particular prediction, rather than simply trusting a 'black box' result.
This research suggests that such AI tools, built from standard lab data already collected during routine hospital care, could potentially help clinicians more quickly and accurately identify which type of microvascular complication a diabetes patient has, which matters because kidney disease and eye disease may require different monitoring and treatment approaches. However, since this study was conducted at a single hospital in China, further research across diverse populations and healthcare settings would be needed before such a model could be broadly adopted in clinical practice.
Zhang Y, Feng S, Xue Y, Xue L, Luo J. (2026). Development and validation of an explainable machine learning model for differentiating diabetic nephropathy from diabetic retinopathy in patients with type 2 diabetes.. Frontiers in endocrinology. https://doi.org/10.3389/fendo.2026.1902508