Machine-Learning Model Targets Cystatin C Testing in Adults with Diabetes

A machine-learning model built from routine blood tests can identify which adults with diabetes are most likely to show a large gap between their creatinine-based and cystatin C-based kidney-function estimates, according to a retrospective two-center study published in Frontiers in Endocrinology. When researchers restricted cystatin C testing to the highest-ranked 20 percent of patients, the model detected 61.2 percent of cases in its development cohort and 55.4 percent in an independent external validation cohort, enriching case detection nearly threefold over random testing.

The target phenotype, labeled D2, was defined as a ratio of cystatin C-based estimated glomerular filtration rate (eGFRcys) to creatinine-based eGFR (eGFRcr) below 0.70, meaning the cystatin C estimate was more than 30 percent lower than the creatinine estimate. Such large negative discordance may reveal kidney-function differences that creatinine alone does not capture. Previous meta-analyses have associated this pattern with higher risks of mortality, cardiovascular events, heart failure, and kidney failure, and cohort studies in adults with diabetes have linked eGFR differences with microvascular complications and incident albuminuria.

Cystatin C offers complementary information to creatinine because the two markers are influenced by different non-GFR factors, and clinical guidance favors combined creatinine-cystatin C estimates when cystatin C is available. However, universal cystatin C testing is often impractical due to assay availability and implementation barriers, raising the practical question of whom to test first. The study’s intended application was prioritization of cystatin C measurement using demographic and laboratory information already available in routine care.

The retrospective study included adults aged 18 or older with a recorded diabetes diagnosis and paired serum creatinine and cystatin C measurements from the same blood specimen, analyzed on identical Atellica CH 930 analyzers at two Chinese hospitals. The development cohort, assembled at Wanbei Hospital between May 2022 and November 2023, comprised 1,170 participants, of whom 214 (18.29 percent) were D2-positive. The external validation cohort, drawn from 2024 records at the First Affiliated Hospital of Anhui Medical University, included 1,619 participants with 112 D2-positive cases (6.92 percent). Diabetes type could not be reliably classified across the two centers.

Seven classification algorithms were evaluated against a fixed pool of 28 routinely available laboratory and demographic variables. XGBoost outperformed the alternatives on the prespecified primary metric, the area under the precision-recall curve, and was advanced to SHAP-guided feature reduction. The final nine-predictor model used age, sex, eGFRcr, hemoglobin, platelet count, blood urea nitrogen, albumin, alkaline phosphatase, and glucose. In development, using pooled out-of-fold predictions, it achieved an AUROC of 0.863 (95 percent CI, 0.833 to 0.889) and an AUPRC of 0.600 (95 percent CI, 0.532 to 0.668), compared with 0.791 and 0.461 for a prespecified clinical core of age, sex, and eGFRcr alone.

Applied unchanged to the external cohort, the model retained useful discrimination despite the much lower D2 prevalence there: AUROC was 0.806 (95 percent CI, 0.769 to 0.840) and AUPRC was 0.258 (95 percent CI, 0.193 to 0.337), versus 0.661 and 0.126 for the clinical core. At an external operating point achieving at least 80 percent sensitivity, specificity was 70.0 percent for XGBoost versus 35.4 percent for the clinical core. At the 20 percent testing allocation externally, the model identified 62 of 112 D2 cases, compared with 42 for the clinical core and 43 for a three-variable logistic model using eGFRcr, blood urea nitrogen, and albumin, a paired recall difference of 17.0 percentage points (95 percent CI, 8.6 to 27.2). The corresponding difference in development was smaller and uncertain, at 1.9 percentage points.

Discrimination, however, did not transport as absolute probability. The raw external predictions substantially overestimated risk, with a mean predicted probability of 29.9 percent against an observed prevalence of 6.9 percent, and the external Brier score of 0.144 was worse than the clinical core’s 0.130. A post-hoc, cross-fitted logistic recalibration performed within the external cohort improved local calibration markedly, but the authors emphasized this was a local calibration assessment rather than a second independent validation. Decision-curve analysis using the original external probabilities also did not show consistently positive net benefit across the exploratory 0 to 12 percent threshold range examined. The findings indicate the model is better supported for risk ranking and testing prioritization than for transferring uncalibrated probability thresholds between centers.

SHAP interpretation identified eGFRcr as the largest contributor to model output, which the authors noted is expected because it is mathematically embedded in the D2 definition and should not be read as an independent biological risk factor. Blood urea nitrogen, albumin, platelet count, hemoglobin, glucose, alkaline phosphatase, age, and sex contributed additional, often nonlinear, information. Among D2-negative participants, those selected by the model showed consistently higher prevalence of related cystatin C-based phenotypes in both cohorts, including lower eGFRcys, more negative eGFRcys-minus-eGFRcr differences, higher cystatin C concentrations, and more frequent absolute eGFRcys reductions below 60 mL/min/1.73 m². Because these related measures all share cystatin C, they do not constitute independent confirmation of kidney disease.

The authors detailed several limitations. The study was retrospective, and development performance estimates represent model-selection results rather than independent internal validation, since hyperparameter tuning and out-of-fold evaluation shared the same five-fold partition. D2 and related phenotypes were defined from a single paired measurement and do not establish chronicity-confirmed chronic kidney disease; measured GFR was unavailable, so the study evaluates discordance between estimates rather than determining which estimate better reflects true filtration. In small selectively retested subsets, D2 classification agreed after 30 to 365 days in 66.1 percent of development participants (n=59) and 86.8 percent of external participants (n=53). Participants had already undergone paired testing and were required to have measured blood counts, potentially limiting applicability to unselected populations, and evaluation was confined to two Chinese centers.

The testing-allocation analysis was also retrospective and did not assess downstream changes in clinical management, patient outcomes, or cost-effectiveness. The authors concluded that routine clinical laboratory data can support prioritization for large negative eGFR discordance in adults with diabetes, and that model-based ranking may help target cystatin C testing when universal measurement is impractical, but raw probabilities require local calibration before any threshold-based decisions, and prospective evaluation is needed before clinical implementation.

Subject of Research: Development and external validation of a machine-learning model to prioritize cystatin C testing for large negative eGFR discordance in adults with diabetes

Article Title: Targeted cystatin C testing for large negative eGFR discordance in adults with diabetes: development and external validation of a machine-learning model

Article References: Wang, X., Zhang, M., Zhang, K., Xu, Y., Chen, Y., & Hu, Y. (2026). Targeted cystatin C testing for large negative eGFR discordance in adults with diabetes: development and external validation of a machine-learning model. Frontiers in Endocrinology, 17, Article 2002540. https://doi.org/10.3389/fendo.2026.2002540

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.2002540

Keywords: cystatin C, eGFR discordance, diabetes, kidney function, machine learning, XGBoost, external validation, targeted testing, chronic kidney disease, clinical laboratory, risk prediction, Frontiers in Endocrinology