Developing and validating a clinlabomics-based machine-learning model for early detection of occult diabetic kidney disease: implications for primary care screening

Background Occult diabetic kidney disease (DKD) is a subtle yet high-risk microvascular complication of type 2 diabetes mellitus (T2DM). Early-stage DKD often goes undetected because traditional screening markers remain within the normal range. This study aimed to develop and validate an explainable machine learning (ML) model using routine clinical and laboratory data for the early detection of occult DKD. Its potential value for primary care screening was also evaluated. Methods This multicenter retrospective study included 1,916 hospitalized patients with T2DM. The derivation cohort consisted of 1,066 patients from Wanbei Coal-Electricity Group General Hospital and was used to train the model. An independent cohort of 850 patients from the First Affiliated Hospital of Anhui Medical University served for external validation. Thirty-two routine clinical variables were initially considered. Eight ML algorithms were compared to identify the optimal model. SHapley Additive exPlanations (SHAP) was employed to rank feature importance, reduce variables, and interpret the model. Finally, a quartile-based risk stratification system and a web-based tool were developed. Results Among the eight algorithms, logistic regression (LR) showed the best performance. Using SHAP rankings, a simplified LR model was built with eight features: HGB, HbA1c, HTN, UA, sex, MicroVCs, CVD, and A/G. The model performed well in both the training cohort (AUC = 0.824) and the external validation cohort (AUC = 0.786). SHAP analysis identified HbA1c, uric acid (UA), and hemoglobin (HGB) as the top contributors. The risk stratification system demonstrated clear separation, with the incidence of occult DKD rising from 1.5% in the lowest-risk quartile (Q1) to 55.8% in the highest-risk quartile (Q4). Additionally, decision curve analysis demonstrated that the model provides substantial clinical net benefit, and the final model was implemented as an interactive web-based calculator for real-time risk assessment. Conclusion An explainable ML model was successfully developed to accurately predict occult DKD using routine clinical data. The model combines good performance with clear interpretation. It may serve as a practical tool for large-scale screening and early intervention in primary care.