The First Deployed Machine Learning Algorithm From Routine Labs

Disease Detection Series • Chronic Kidney Disease

Chronic kidney disease (CKD) kills nearly 58,000 Americans annually and affects 35.5 million more, most of them unaware. Nine in 10 adults with CKD have not been diagnosed.¹ The disease writes itself into the comprehensive metabolic panel (CMP) for years before kidney function falls to a level that triggers clinical concern. Machine learning that reads the creatinine trajectory, not just the current value, alongside concurrent shifts in blood urea nitrogen (BUN), albumin, and hemoglobin, identifies patients on the path to kidney failure two to five years before conventional thresholds are crossed. One algorithm is now CE-marked and deployed.

Klinrisk, the first CE-marked machine learning algorithm for CKD from routine labs, achieves AUC 0.88 and was validated on 4.8 million US adults across commercial, Medicare, and Medicaid populations.

The Treatment Effect That Makes Early Detection Urgent

The treatment window for CKD is far more effective early than late. SGLT2 inhibitors reduce CKD progression by 30 to 40 percent in patients with preserved estimated glomerular filtration rate (eGFR). ACE inhibitors and angiotensin II receptor blockers (ARBs) most effectively slow proteinuria-driven progression in Stages 1 to 3. Finerenone reduces kidney failure risk by 20 percent in CKD patients with diabetes.² Detecting CKD at Stage 2 rather than Stage 4 delivers 5 to 10 additional years of nephroprotective therapy before dialysis becomes necessary—a difference of approximately $450,000 to $900,000 in lifetime dialysis costs per patient, and the difference between a treatable condition and a terminal one.

The Signature in the Blood

CKD progression records itself in routine labs well before clinical thresholds. Serum creatinine rises insidiously, from 0.9 to 1.0 to 1.1 to 1.3 mg/dL, while remaining in the reference range until up to 50 percent of nephron mass has been lost. The eGFR slope—the annual rate of decline—is more informative than any single value: accelerating decline of >3 mL/min/1.73m²/year signals imminent Stage 3.³

BUN rises with declining filtration. Albumin falls due to inflammation, malnutrition, and proteinuria. Hemoglobin declines as erythropoietin production falls, typically beginning in Stage 3. Potassium rises, calcium falls, phosphorus rises, and bicarbonate declines as tubular regulation fails. This multi-parameter trajectory forms an interconnected pattern that no single-value alert can identify, but that machine learning can read across simultaneous dimensions.

The Machine Learning Case

The Klinrisk random survival forest algorithm was developed on 77,196 Manitoba patients and externally validated on 107,097 Alberta patients, achieving an area under the curve (AUC) of 0.88 at two years and 0.84 at five years. It was subsequently validated in the CANVAS Program and CREDENCE trial, achieving an AUC of 0.88 and outperforming the KDIGO heatmap at every time interval (p<0.01).⁴⁵

Most definitively, Tangri and colleagues validated Klinrisk on 4.8 million United States adults across commercial, Medicare, and Medicaid populations, achieving an AUC of 0.80 to 0.87—the largest and most diverse validation of any CKD machine learning algorithm.⁶ Roche received a CE-mark for the Kidney Klinrisk Algorithm on its navify® platform in October 2025, making Klinrisk the first regulatory-approved machine learning algorithm for CKD from routine laboratory data. LifeLabs, Canada’s largest diagnostic laboratory, holds the exclusive North American commercial partnership.

What This Proves

CKD is the furthest advanced of all 11 diseases in this series: a deployed, CE-marked machine learning algorithm validated on 4.8 million patients. Klinrisk demonstrates that the creatinine trajectory, eGFR slope, and concurrent metabolic shifts carry enough longitudinal information to predict kidney failure years before it occurs. The question for CKD is no longer whether machine learning works; it is how quickly health systems will integrate it.

Endnotes

1.U.S. Centers for Disease Control and Prevention. “Chronic Kidney Disease in the United States, 2023.” 2024.

Comprehensive epidemiological update documenting CKD in 35.5 million Americans, 9 in 10 undiagnosed. Establishes CKD as a leading cause of death with a massive undiagnosed burden—the clinical context that makes machine learning-based early detection compelling.

2.Perkovic V, Jardine MJ, Neal B, et al. “Canagliflozin and Renal Outcomes in Type 2 Diabetes and Nephropathy.” N Engl J Med. 2019;380(24):2295–2306.

CREDENCE trial demonstrates that canagliflozin reduces kidney failure risk by 30% and cardiovascular events by 20% in CKD patients with diabetes. Key evidence that nephroprotective treatment is available and effective when delivered early, establishing why early CKD detection from routine labs directly saves lives.

3.Tangri N, Grams ME, Levey AS, et al. “Multinational Assessment of Accuracy of Equations for Predicting Risk of Kidney Failure: A Meta-Analysis.” JAMA. 2016;315(2):164–174.

Meta-analysis validating the Kidney Failure Risk Equation across 31 cohorts, establishing eGFR slope and albuminuria as primary predictors of kidney failure. Establishes the predictive signal in routine CMP values and is the clinical foundation on which Klinrisk builds its machine learning enhancement.

4.Ferguson T, Ravani P, Sood MM, et al. “Development and External Validation of a Machine Learning Model for Progression of CKD.” Kidney Int Rep. 2022;7(8):1772–1781.

Development and external validation of Klinrisk RSF on 77,196 plus 107,097 patients, achieving AUC 0.88 at two years. Primary publication establishing Klinrisk performance, the first large-scale demonstration that machine learning on routine CMP values predicts kidney failure better than existing clinical tools.

5.Tangri N, Ferguson TW, Bamforth RJ, et al. “Machine Learning for Prediction of Chronic Kidney Disease Progression: Validation of the Klinrisk model in the CANVAS Program and CREDENCE Trial.” Diabetes ObesMetab. 2024;26(8):3371–3380.

Clinical trial validation of Klinrisk in CANVAS and CREDENCE, achieving AUC 0.88 and outperforming KDIGO heatmap at every time interval. Establishes regulatory-quality evidence for Klinrisk performance in high-risk CKD populations with diabetes.

6.Tangri N, Ferguson TW, Teng CC, et al. “Validation of the Klinrisk Machine Learning Model for CKD Progression in a Large Representative US Population.” J Am Soc Nephrol. 2026;37(2):326–337.

Validation of Klinrisk on 4.8 million US adults across commercial, Medicare, and Medicaid populations, achieving AUC 0.80–0.87. The largest and most diverse validation of any CKD machine learning algorithm, establishing generalizability across US insurance types, ages, and ethnic backgrounds.