Model Quantifies Uncertainty in Kidney Function Decline for Diabetic Patients

Researchers have developed and internally evaluated a statistical model designed to predict the decline of kidney function in adults with type 2 diabetes mellitus, specifically focusing on the uncertainty surrounding these long-term forecasts. The study, a secondary analysis of data from the multicenter Action to Control Cardiovascular Risk in Diabetes (ACCORD) trial, addresses a common limitation in current medical informatics where long-horizon kidney trajectory models typically report only point estimates or event risks. By incorporating conformal intervals, the new approach provides a clearer picture of the predicted change in estimated glomerular filtration rate (eGFR) over a 48-month period, making the inherent uncertainty of the prediction explicit rather than leaving it implicit.

The research team analyzed data from 6,853 adults with type 2 diabetes and high cardiovascular risk who participated in the ACCORD trial. The primary model utilized a comprehensive set of baseline clinical variables, including demographics, cardiometabolic measures, medication history, and prior medical history.

In their methodological approach, the researchers compared several penalized linear, random forest, histogram gradient boosting, and XGBoost models. These advanced machine learning techniques were benchmarked against mean-outcome and clinical linear models to assess their predictive capabilities. The study employed a rigorous evaluation strategy, using five-fold training-set cross-validation to select the optimal model. Separate calibration and test sets were then used to evaluate point accuracy, calibration, and the discrimination of rapid decline. Additionally, the team assessed the coverage and width of both split and locally adaptive 90% prediction intervals to ensure the model’s reliability in quantifying uncertainty.

The results of the study, based on a test set of 1,371 participants, indicated that the random forest model achieved a mean absolute error of 3.370 mL/min/1.73 m² per year, with a 95% confidence interval ranging from 3.221 to 3.527. The model’s explanatory power was reflected in an R-squared value of 0.359, with a 95% confidence interval from 0.233 to 0.466.

A key finding of the study was the performance of the locally adaptive intervals, which achieved a coverage of 0.909, with a 95% confidence interval from 0.894 to 0.923. This coverage is close to the nominal 90% target, indicating that the model’s uncertainty quantification is statistically sound. However, the mean width of these intervals was 14.706 mL/min/1.73 m² per year, which the authors note remains broad.

The researchers also examined the prevalence of rapid kidney decline, defined as an eGFR drop of at least 5 mL/min/1.73 m² per year. In the test set, this rapid decline occurred in 467 participants, representing 34.1% of the group. The model demonstrated an area under the curve (AUC) of 0.732 for discriminating rapid decline, with a 95% confidence interval from 0.708 to 0.758.

Further analysis revealed that the uncertainty in the model’s predictions was not uniform across all patients. Notably, participants in Q5 had higher levels of HbA1c (8.6% vs. 8.2%), higher triglycerides (253.3 vs. 177.5 mg/dL), and a higher proportion of females (49.6% vs. 30.7%) compared to those in Q1, despite having similar baseline eGFR values. These findings indicate that factors beyond baseline kidney function, such as glycemic control and lipid levels, contribute significantly to the modeled uncertainty. This variation in conditioned width suggests that the model captures distinct subgroups of patients with different levels of predictability.

The study concludes that conformal intervals can achieve near-nominal marginal coverage, providing a statistically valid framework for quantifying prediction uncertainty in kidney function decline. However, the intervals remain broad, reflecting the inherent difficulty of long-term forecasting in a complex disease like type 2 diabetes. The conditioned width differences across HbA1c, triglycerides, and sex reveal that modeled uncertainty varies based on specific clinical characteristics, offering insights into which patient profiles are more or less predictable.

Subject of Research: Uncertainty-aware prediction of 48-month eGFR decline in type 2 diabetes mellitus: a secondary analysis of ACCORD

Article Title: Uncertainty-aware prediction of 48-month eGFR decline in type 2 diabetes mellitus: a secondary analysis of ACCORD

Article References: Olshvang, D., Harris, C., Chellappa, R., Parikh, C., & Santhanam, P. (2026). Uncertainty-aware prediction of 48-month eGFR decline in type 2 diabetes mellitus: a secondary analysis of ACCORD. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03921-3

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03921-3

Keywords: Uncertainty-aware, prediction, month, eGFR, decline, type, diabetes, mellitus, secondary, analysis, ACCORD, scientific research