Predicting Progression to Kidney Failure
Routine Lab Tests Make Up Prediction Model
By Bill Malone
While labs routinely report estimated glomerular filtration rate (eGFR) based on serum creatinine, the varied way in which patients progress to kidney failure leaves physicians with a tough call about when and how to treat patients with chronic kidney disease (CKD). This issue of Strategies explores a recent paper that developed a predictive model for progression to CKD based on readily obtained lab tests.
Although chronic kidney disease (CKD) is a growing problem that affects approximately 12% of the U.S. adult population, physicians lack effective tools to determine which patients need the most aggressive therapy. Even patients with similar eGFR have significant heterogeneity in risk for progression to kidney failure. A recent study addressed this quandary by developing and validating several prediction models that depended on routine lab testing (JAMA 2011;305:1553–1559).
“It turns out that kidney disease is very common, but silent, and progresses very slowly. Therefore a lot of people have it, but the risk of adverse consequences is not the same in everyone,” said Andrew Levey, MD, an author of the paper. “So, for example, we think about 40% of people over the age of 70 have chronic kidney disease. When we look at the chance of those people going on to kidney failure or needing dialysis, it’s less than one in 100, so many people who have chronic kidney disease are going to die with it and not from it.” Levey is the Gerald J. and Dorothy R. Friedman professor of medicine and chief of the William B. Schwartz Division of Nephrology at Tufts University School of Medicine in Boston, Mass.
Levey and lead investigator Navdeep Tangri, MD, and colleagues developed and externally validated a set of accurate prediction models for progression of CKD, with an eye toward creating a model that could be easily implemented in clinical practice. They used clinical, demographic, and lab data from two independent Canadian cohorts with CKD stages 3 to 5 who were referred to nephrologists between April 2001 and December 2008. The primary outcome measured was kidney failure, defined as need for dialysis or pre-emptive kidney transplantation.
The development and validation groups included 3,449 patients (11% with kidney failure) and 4,942 patients (24% with kidney failure), respectively. The researchers developed a series of seven models of increasing complexity. The first three models used only age, sex, eGFR, and albuminuria, while the other four added others types of variables, including clinical conditions such as diabetes, physical examination variables like blood pressure, and additional lab measures including serum phosphate, bicarbonate, and calcium.
The most accurate model included eight variables: age, sex, eGFR, albuminuria, calcium, phosphate, bicarbonate, and serum albumin. In the validation cohort, this model was the most accurate at predicting need for dialysis or kidney transplantation. Among the three more basic models, a four-variable model that took into account only age, sex, eGFR, and albuminuria performed the best.
Clinicians need new tools to help predict a patient’s course of CKD because of the inherent variability in the disease, the authors emphasized. The paper presents an example based on their models. In the illustration, a hypothetical 70-year-old man with eGFR of 30 mL/min/1.73 m2, urine albumin-to-creatinine ratio (ACR) of 200 mg/g, calcium of 9.0 mg/dL, phosphate of 4.5 mg/dL, serum albumin of 3.5 g/dL, and bicarbonate of 21 mEq/L would have a 5-year risk of developing kidney failure of 19.8% based on age, sex, and estimated GFR; 16.3% based on age, sex, estimated GFR, and urine ACR; and 26% with all eight parameters in the most robust model.
“These examples suggest that both the four-variable and the eight-variable models could lead to clinically meaningful reclassification of risk for individual patients,” noted Marcello Tonelli, MD and Braden Manns, MD of the Alberta Kidney Disease Network in their accompanying editorial (JAMA 2011;305:1593-1595). They called the paper’s findings “novel and important.”
Without new tools, clinicians are left making informed guesses, according to Levey. “Right now the metrics by which we can say who is likely is to do worse than the next person are fairly crude and unsatisfactory in the eyes of most physicians,” he said. “So most physicians make their best estimate based on unsatisfactory tools, or they decide to wait and see. And in some people, we end up waiting past the point when we should have done something, or in others we intervene on people who didn’t really need it. So without some sort of instrument to help us predict better, we’re destined just to make our best guess.”
At the same time, Tonelli and Manns cautioned that these models need greater study, expressing concern that they could lead to more testing or more referrals to nephrologists without proof of improved outcomes. “Developing optimal risk prediction tools is only part of the challenge ahead,” they wrote. “Data are urgently needed to clarify how better prognostic information can be incorporated into routine care to improve patient outcomes rather than simply increasing physician workload, the costs of laboratory testing, and the complexity of risk instruments. This seems particularly important given studies showing that routine estimated GFR reporting increased the likelihood of specialist referrals but did not improve outcomes. Achieving this objective will require studies that use other metrics besides discrimination and calibration—such as assessing the acceptability of risk stratification schemes to primary care physicians and the feasibility of implementation in diverse clinical settings—and will require studies that demonstrate that using better risk prediction tools will lead to clinically meaningful benefit for patients.”
However, Levey is confident that developing the risk models was a big step in the right direction, and he and his colleagues plan to publish on research validating their models in the primary care setting soon. “I think that the whole assumption is that if we could target our therapies to people who were at greater risk, that the inevitable side effects of therapies would be worth it in that group, whereas if we don’t target our therapies and we apply them to everyone irrespective of risk, then we end up with the small risk of side effects outweighing their benefit,” he said. “It’s been proven in the most effective therapy we have to slow progression of kidney disease, which is the ace inhibitors or angiotensin receptor blockers, that these have a greater benefit in people with more protein in the urine, and that’s one of the main factors in our risk equation.” He also noted that risk scores are common in other areas of medicine, such as the Framingham risk equation in cardiology.
Levey believes that labs will be a key partner in using predictive models in CKD in the future, and suggested that labs could report risk for a CKD patient in the same way they report information on cardiovascular disease risk. “For example, there’s the arterial sclerotic heart disease risk score, which is based on age, sex, and cholesterol values,” he said. “Similarly, we could have a kidney failure risk score reported by the lab on the basis of the creatinine, urine albumin, and the other tests, as well as age, sex.”
Join us for a discussion of this article, and interact with your peers, on LinkedIn