Topic Development

This topic was proposed by Patrick Pun, MD, MHS, and the Nephrology Field Advisory Committee.

Conceptual Model

We developed a conceptual model to clarify the relationship of the KQs to the overall pathway of patients who undergo MRI studies with GBCAs. As depicted in Figure 1, patients who undergo an MRI or MRA imaging study may or may not receive gadolinium exposure to obtain the clinically required diagnostic information. KQ 1 addresses the rate of nephrogenic systemic fibrosis (first box) in all patients who receive GBCA exposure during the course of an MRI/MRA study. Of particular interest are certain subpopulations (KQ 1A-C) identified in the purple box (eg, patients with different types of kidney-related disease). Similarly, KQ 2 addresses the relative risk of NSF among patients who receive newer versus older GBCAs during the course of an MRI/MRA study and examines the risk in the same key subpopulations (KQ 2 A-C). We have also identified other important concepts such as individual patient factors that may increase or modify the risk of NSF and other types of adverse effects among patients who are exposed to GBCAs.

Figure 1

Conceptual Model. Abbreviations: AKI=acute kidney injury; CKD=chronic kidney disease; ESRD=end-stage renal disease; HD=hemodialysis; KQ=key question; MRA=magnetic resonance angiography; MRI=magnetic resonance imaging; PD=peritoneal dialysis; PTH=parathyroid (more...)

Search Strategy

In collaboration with an expert medical librarian, we conducted a primary literature search from inception to January 7, 2019 of MEDLINE^® (via PubMed^®), Embase, Cochrane Register of Controlled Trials, and Web of Science. We used a combination of database-specific subject headings and keywords (eg, gadolinium, contrast media, nephrogenic fibrosis) and searched in the titles and abstracts (Appendix C). We also conducted hand searches of key references^7,9,11–27 for relevant citations that may not have been captured in the database search.

Study Selection

We used the artificial intelligence (AI) technology developed as part of the DistillerSR software (Evidence Partners Inc., Manotick, ON, Canada), called DistillerAI, to assist with screening abstracts.²⁸ Using prespecified inclusion/exclusion criteria (Table 2), the titles and abstracts of a subset of articles (approximately n=100) identified through our primary search were classified independently by 2 senior investigators (KMG, JL) for relevance to the KQs. After resolving disagreements between the 2 investigators, this set of included and excluded articles was used to train the Distiller AI program.

The Distiller AI program screened the remaining titles and abstracts and assigned a prediction score of relevance to the study questions. All citations classified with a prediction score ≤0.5 underwent screening by a single investigator. Potentially relevant studies included by the investigator or with an AI prediction score >0.5 underwent full-text screening. At the full-text screening stage, 2 independent investigators agreed on a final inclusion/exclusion decision (see Appendix D for justification of excluded studies). All articles meeting eligibility criteria were included for data abstraction. All results were tracked in an electronic database (for referencing, EndNote^®, Clarivate Analytics, Philadelphia, PA; for data abstraction, DistillerSR; Evidence Partners Inc., Manotick, ON, Canada).

Table 2 describes the study eligibility criteria organized by PICOTS elements (population, intervention, comparator, outcome, timing, setting) and other criteria such as study design, language, and publication type. We included a broad range of study designs ranging from randomized trials to case reports in order to capture any study type quantitatively reporting NSF in association with GBCA exposure. Studies were excluded if they did not report the number of patients exposed by specific GBCA. Similarly, studies were excluded if they only identified the specific GBCA exposure for those patients ultimately diagnosed with NSF but not the rest of the study population. We also included case reports and case series for patients with NSF that clearly described exposure to an ACR group II and/or III GBCA.

In order to align our KQs with existing guidelines pertaining to the use of GBCAs and their associated risk of NSF, we adopted the groupings for GBCAs given by the American College of Radiology (ACR) in their 2018 guidelines.⁹ Thus, “newer gadolinium-based contrast agents” are referred to throughout the report as ACR group II/III agents and “older gadolinium-based contrast agents” are referred to as ACR group I agents.

Table 2

Study Eligibility Criteria.

Data Abstraction

Data from published reports were abstracted into a customized DistillerSR database by 1 reviewer and over-read by a second reviewer. Disagreements were resolved by consensus or by obtaining a third reviewer’s opinion when consensus was not reached. Data elements included descriptors to assess applicability, quality elements, intervention/exposure details, and outcomes.

Key characteristics abstracted included patient descriptors, specifics of gadolinium agent exposure (eg, specific agent, dose, number of doses received), comparator (if any), outcomes (confirmed or suspected diagnosis of NSF), and source of study funding. Note that if a study included a non-contrast comparator arm, we did not abstract data from the non-contrast arm as the comparison between GBCA exposure and non-GBCA exposure was not the focus of this review. Multiple reports from a single study were treated as a single data point, prioritizing results based on the most complete and appropriately analyzed data. Key features relevant to applicability included the match between the sample and target populations (eg, age, Veteran status).

We defined cases of NSF as “confounded” when there was clear evidence that the patient had been exposed to multiple GBCAs prior to the development of NSF; conversely, “unconfounded” refers to cases in which a patient was noted specifically to have been exposed only to a single GBCA (even if multiple doses of the same GBCA) prior to disease development. When it was not clearly stated whether or not a patient had received exposures to multiple GBCAs, we considered them conservatively as confounded.

Quality Assessment

Quality assessment was done by the investigator abstracting or evaluating the included article and was over-read by a second, highly experienced investigator. Disagreements were resolved by consensus between the 2 investigators or, when needed, by arbitration by a third investigator.

For randomized, nonrandomized, and controlled before-after studies, we used criteria from the Cochrane EPOC risk of bias (ROB) tool.²⁹ These criteria are adequacy of randomization and allocation concealment; comparability of groups at baseline; blinding; completeness of follow-up and differential loss to follow-up; whether incomplete data were addressed appropriately; validity of outcome measures; protection against contamination; selective outcomes reporting; and conflict of interest. We assigned a summary ROB score (low, unclear, high) to individual studies, defined as follows:

• Low ROB: Bias, if present, is unlikely to alter the results seriously.
• Unclear ROB: Information required to determine risk of bias was not clearly specified in the peer-reviewed paper or unable to be obtained to make a judgment.
• High ROB: Bias may alter the results seriously.

For observational cohort and case-control studies, we adapted the Newcastle-Ottawa scale (from the version modified by Guyatt et al).³⁰ This scale includes quality assessment criteria for selection of cases and controls, comparability of cases and controls, and ascertainment of exposure (or outcome as relevant). For questions relevant to cohort studies with an exposed and unexposed group, we consider “exposed” to mean patients who received any ACR Group II or III agent of interest and “nonexposed” to mean patients who received an agent not of primary interest (eg, ACR Group I agents). For cohorts that only report an exposed group, we included a “not applicable” response option for questions specific to exposed and nonexposed groups. Similarly, we modified a question about matching for confounding variables to include adequate statistical adjustment or stratification for confounders if matching was not applicable. See Appendix E for our modified ROB form. Given the number of eligible cohort and case-control studies, we did not evaluate the ROB for case reports or case series studies.

Data Synthesis

We summarized the primary literature using data abstracted from the eligible studies. Summary tables describe the key characteristics of the primary studies overall and by specific gadolinium agent. Next, we determined the feasibility of completing a quantitative synthesis (ie, meta-analysis) to estimate summary effects. The feasibility of conducting a meta-analyses depended on the volume of relevant literature, conceptual homogeneity of the included studies, and completeness of results reported in those included studies. Due to heterogeneity of study methodology, patient population, and follow-up time points across studies, we elected not to conduct meta-analysis.

While we did not calculate summary estimates across studies, we do present forest plots of the point estimates from individual studies grouped by category of kidney function (all patients, patients with risk factors for CKD, and patients with CKD of any stage) within each KQ. To create these categories, we identified the stages of CKD that were included by a given study. For studies that only reported eGFR ranges, we converted them to standard CKD stages (note: some studies did not report eGFR but only CKD stages). We did not include studies in the forest plots that were not designed to identify cases of NSF as a primary outcome, although the findings of these studies are reported narratively in each result section. Also, there was inconsistency in the reporting of whether or not cases of NSF were confounded across included studies. Thus, in order to facilitate comparisons in the forest plots, the number of cases of NSF reported for each study is the total found in a given study and may include confounded cases.

Because of the variability in methods across included studies and the low numbers of NSF cases found, we report the occurrence of NSF cases per index GBCA exposure as opposed to a relative risk, prevalence, or incidence. This allows for accurate reporting of the phenomena of interest and for comparisons across studies that use both the term incidence and prevalence. We refer to “index GBCA exposure” as the contrast agent identified in each study as the primary exposure in questions related to NSF occurrence, acknowledging that some patients were exposed to multiple agents potentially both before and after the index exposure. Finally, we calculated an exact upper 95% confidence interval (CI) for each individual study, which is also displayed in the forest plots. Analyses were performed with the R statistical package version 3.5.3 (R Foundation; https://www.R-project.org/). Exact 95% confidence intervals³¹ were obtained with the binom.test function.

Because quantitative synthesis was not indicated, we narratively analyzed outcomes for both KQs. For narrative analyses, we gave more weight to evidence from higher quality studies (ie, low ROB) when possible. Our narrative synthesis focused on documenting and identifying patterns of NSF development across categories of kidney function and types of GBCA exposure. For KQ 2, we did not calculate risk ratios or odd ratios for the following reasons: the included studies were not designed for this type of comparison originally, it was unclear if the populations receiving different GBCAs were directly comparable, and there is reason to suspect confounding by indication (eg, certain GBCAs are preferred for MRIs of different organs). We also analyzed potential reasons for inconsistency in treatment effects across studies by evaluating differences in the study population, intervention, comparator, and outcome definitions.

Rating the Body of Evidence

The certainty of evidence (COE) for each key question was assessed using the approach described by the Grading of Recommendations Assessment, Development, and Evaluation (GRADE) working group.³² We limited GRADE ratings to those outcomes identified by the stakeholders and technical expert panel as critical to decision-making (ie, development of NSF). Additionally, we limited COE assessment to the highest order study designs (ie, EPOC criteria studies, prospective and retrospective cohorts). In brief, this approach requires assessment of four domains: risk of bias, consistency, directness, and precision. Additional domains to be used when appropriate are coherence, dose-response association, impact of plausible residual confounders, strength of association (magnitude of effect), and publication bias. These domains were considered qualitatively, and a summary rating was assigned after discussion by 2 investigators (KMG, AMG) as high, moderate, low, or very low COE. COE was not assessed for studies that only enrolled patients with chronic liver disease.

Peer Review

A draft version of this report was reviewed by technical experts and clinical leadership. A transcript of their comments and our responses is in Appendix F.