A-063 Assessing the risk of measurement variability of model inputs on the accuracy and precision of kidneyintelX.dkd, a machine learning prognostic test for progression in Diabetic Kidney Disease
Year: 2025
Type: conference-abstract
Abstract: Background Diabetic Kidney Disease (DKD) is a severe complication of diabetes that leads to kidney failure in about 5-15% of cases. The kidneyIntelX.dkd is an FDA approved prognostic blood test for assessing the risk of DKD progression, enabling a more personalized approach to treatment and care for those individuals based on risk. The test uses a non-linear random forest algorithm which combines novel prognostic biomarkers Tumor Necrosis Factor Receptors 1 and 2 (TNFR1, TNFR-2) and Kidney Injury Molecule (KIM-1) with readily available clinical laboratory values of Blood Urea Nitrogen (BUN), Hemoglobin A1c (HbA1c) and urinary Albumin Creatinine ratio (UACR). The test reports a low, moderate, or high risk for either a sustained 40% decline in eGFR or kidney failure over 5 years. Measurement of novel biomarkers is performed using an analytically validated assay on received test samples with results algorithmically combined with the aforementioned standard laboratory values obtained from the patient’s health record within the past 12 months. In this study, we sought to characterize the risk and impact of imprecision in the measurement of the inputs, including those taken from patient records, on the accuracy of final kidneyintelX.dkd result. Methods Simulation study assessed the cumulative measurement imprecision of six (6) inputs on the random forest algorithm outputs. The kidneyintelX.dkd test provides a predicted probability of the kidney event and categorizes patients as low, moderate and high based on predetermined cut-offs. Imprecision for KIM-1, TNFR-1 and TNFR-2 was established from precision studies conducted in accordance with CLSI and FDA guidelines. For HbA1c, BUN and UACR, the %CV studied was based on data derived from College of American Pathologists (CAP) and American Proficiency Institute (API) survey data, published literature on observed measurement variability and data published in decision summaries on the FDA website on cleared diagnostic devices and methods routinely used in clinical practice. Subjects (n=675) included in the clinical validation study submitted to FDA and included in product literature were assessed in these analyses. Assuming a normal distribution of measurement variability, input feature levels were randomly varied (n=100 per subject) across the range of X-Y to X+Y where X was the actual input feature level reported for subjects and Y was the value after applying the expected range of variation. Results The mean intra-individual %CV of predicted probabilities from the kidneyintelX.dkd test in 100 simulation experiments in each subject was 3.3%. In total, 65,700 simulation experiments were performed with just 3.9% resulting in a recategorization of the kidneyintelX.dkd risk level result (e.g. from high to moderate risk). Across the three risk categories, the mean %CV and the % of experiments resulting in risk level reclassification were as follows: Low: 1.2% and 2.3%, Moderate: 6.3% and 5.7%, and High 4.6% and 6.2%. Conclusion The cumulative level of variability in the kidneyintelX.dkd test result is significantly less than that observed in the individual inputs, thereby providing a comprehensive, accurate, reliable, and robust assessment of risk for progressive decline in kidney function or kidney failure.
Source: Clinical Chemistry
Authors: Andrew Peabody, Michael Donovan, Kara Moran, Katherine Edwards, Fergus Fleming
Language: English
Subfield: Nephrology
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4414741122
Cites:
Cited by:
Related to: