Identification of Patients with CKD in Medical Databases: A Comparison of Different Algorithms

Abstract

Background and objectives: Despite CKD consensus definitions, epidemiologic studies use multiple different algorithms to identify CKD. We aimed to elucidate if this affects the patient characteristics and the estimated prevalence and prognosis of CKD by applying six different algorithms to identify CKD in population-based medical databases and compare the cohorts.

Design, setting, participants, & measurements: Patients with CKD in Northern Denmark (2009-2016) were identified using six different algorithms: five were laboratory based defined by (1) one measured outpatient eGFR <60 ml/min per 1.73 m² (single test, n=103,435), (2) two such findings ≥90 days apart (Kidney Disease Improving Global Outcomes, n=84,688), (3) two such findings ≥90 days apart with no eGFR >60 ml/min per 1.73 m² observed in-between (Kidney Disease Improving Global Outcomes, persistent, n=68,994), (4) two such findings ≥90 and <365 days apart (Kidney Disease Improving Global Outcomes, time limited, n=75,031), and (5) two eGFRs <60 ml/min per 1.73 m² or two urine albumin-creatinine ratios >30 mg/g ≥90 days apart (Kidney Disease Improving Global Outcomes, eGFR/albuminuria, n=100,957). The sixth included patients identified by reported in- and outpatient hospital International Classification of Diseases diagnoses of CKD (hospital-diagnosed, n=27,947). For each cohort, we estimated baseline eGFR, CKD prevalence, and 1-year mortality using the Kaplan-Meier method.

Results: The five different laboratory-based algorithms resulted in large differences in the estimated prevalence of CKD from 4637-8327 per 100,000 population. In contrast, 1-year mortality varied only slightly (7%-9%). Baseline eGFR levels at diagnosis were comparable (53-56 ml/min per 1.73 m²), whereas median time since first recorded eGFR <60 ml/min per 1.73 m² varied from 0 months (single-test) to 17 months (Kidney Disease Improving Global Outcomes, persistent). The hospital-diagnosed algorithm yielded markedly lower CKD prevalence (775 per 100,000 population), a lower baseline eGFR (47 ml/min per 1.73 m²), longer time since first eGFR <60 ml/min per 1.73 m² (median 70 months), and much higher 1-year mortality (22%).

Conclusions: Population prevalence of CKD identified in medical databases greatly depends on the applied algorithm to define CKD. Despite these differences, laboratory-based algorithms produce cohorts with similar prognosis.

Podcast: This article contains a podcast at https://www.asn-online.org/media/podcast/CJASN/2021_03_11_CJN15691020_final.mp3.

Keywords: algorithms; chronic kidney disease; epidemiology; glomerular filtration rate; hospital records; incidence; mortality; prevalence; prognosis; registries.

Graphical abstract

Figure 1.

Distribution of eGFR on the index date in patients with CKD in six cohorts, sampled using different algorithms during 2009–2016 in Northern Denmark. IQR, interquartile range; KDIGO, Kidney Disease Improving Global Outcome.

Figure 2.

Time since first test of eGFR <60 ml/min per 1.73 m² before index date (month zero) in six cohorts, sampled using different CKD-defining algorithms during 2009–2016 in Northern Denmark.

Figure 3.

Mortality in patients with CKD in six cohorts, sampled using different algorithms during 2009–2016 in Northern Denmark.

Figure 4.

Proportion of patients with CKD receiving maintenance dialysis during <8 years follow-up in six cohorts, sampled using different CKD-defining algorithms during 2009–2016. The intercept of the graph reflects how many were already in maintenance dialysis on the index date.