Non-steroidal anti-inflammatory drugs: what is the actual risk of chronic kidney disease? A systematic review and meta-analysis

An initial limited search of PUBMED was undertaken to identify articles on the topic [9]. The text words contained in the titles and abstracts of relevant articles, and the index terms used to describe the articles were used to develop a full search strategy (Additional file 1). The search strategy, including all identified keywords and index terms, was adapted for each included database and/or information source. Studies published in either English or other languages (with English abstract) were included. There were no date limitations on the search. We searched the following electronic databases from inception to May 2023: MEDLINE, Cochrane Library, Web of Science and Science direct. We also hand-searched the reference lists of relevant articles and contacted experts in the field for additional studies or retrieval of full text of studies.

Data extraction was performed by one reviewer using a standardized data extraction form and checked by another reviewer. The extracted data included: study information; (name, journal name, year, authors, title, country), study design, aim, setting, participants description (inclusion, exclusion, total number and number per group), exposure description and measurements, outcome (CKD description and measurements), results summary, number and percentage of NSIADs users with CKD and non NSAIDs users with CKD and reported main effect size measure.

Another form was used to assess the risk of bias of included studies using JBI’s critical appraisal tools for observational [10]. The risk of bias assessment was conducted independently by two reviewers. Both reviewers evaluated the studies separately and were blinded to each other’s assessments. Each potential risk of bias was judged as high, low and no information. Any discrepancies between the two reviewers were resolved through discussion, and when necessary, a third reviewer was consulted to reach a consensus. We used Robvis online tool for data visualization to summarize the risk of bias for each included study [11].

We assessed included studies for clinical and methodological heterogeneity by inspecting all studies for setting or participants that are clearly different from other studies and methodological outliers were discussed, possible sources were considered in the analysis.

Out of the 40 included studies, 39 were meta-analyzable. We used the generic inverse variance method to pool log Odds ratios and log Hazard ratios and their standard error for binary, and time-to-event effect measures, respectively. Methodological heterogeneity in measurement of exposures and outcomes necessitated the use of a random-effects model. We assessed the statistical heterogeneity among the studies (heterogeneity in effect size variance) by visual inspection of the results to examine the statistical heterogeneity and by using the I-squared statistic and the Cochran’s Q test. We explored the sources of heterogeneity using subgroup analysis and meta-regression based on the study characteristics, such as study design, setting (hospital, outpatient, community and secondary data), quality, population (comorbidities, age groups, gender), exposure definition (Duration and type of NSAIDs use) and outcome definition. We assessed studies included in meta-analysis for publication bias using funnel plots for studies reporting Odds ratio and HR separately (Additional file 2). Stata software (version 18) native suite of commands, “meta”, was used to perform meta-analysis. The suite is broad and simple. “meta set” command was used for meta-analysis setup, followed by “meta summarize” and “meta forest” commands. These commands allow for subgroup analyses. We used the “leave one out” command to test the effect of deleting each study on the pooled effect size.

We identified 4108 publications from electronic and hand-searches from inception to May 2023 for screening. After excluding studies failing to meet our inclusion criteria and duplicates, 260 full-text studies were assessed for eligibility. Of these, 40 studies met the inclusion criteria were included. The PRISMA flow chart is provided in Figure (1).

Figure 1.

PRISMA flowchart of study selection.

This systematic review included 40 studies, 19 cohort, 11 cross sectional, and 10 case control studies reviewing a total of 1,757,118 participants. All the included 40 studies were available in a full-text format and were in English or with English abstract except for Yarger et al study [12] which was published as abstract only, but we found enough data to include it in our systematic review and meta-analysis.

Although the included studies were between 1986 and 2023, more than half of them (56%) were recent (23 studies out of the 40 were in the last 10 years between 2013–2023). Main characteristics of included 40 studies, settings and populations’ characteristics are shown in Table (1).

Thirty-nine studies out of 40 studies were meta-analyzable (18 cohort, 11 cross sectionals, and 10 case control) studies reviewing a total of 1,756,934 participants. A study on CKD patients found that COX-2 inhibitors significantly reduced eGFR after three and six months, with no notable change post withdrawal at one and two years [13]. This study was excluded from our meta-analysis as it was the only one using means and standard deviations for continuous variables to estimate effect size.

Over all 14 studies were from United States and 2 from Canada, 14 from Asia, 9 from Europe and 1 from Africa. The map in Fig (2) represents the geographical distribution of studied populations.

Figure 2.

Geographical distribution of studied populations.

Although all studies included population aged 18 years or older, six studies were limited to populations older than 60 years [12,14,15,16,17,18]. Two studies included only female population [19,20] and one study include only male population [21].

Seven studies dealt with patients diagnosed with musculoskeletal diseases [22,23,24,25,26,27,28], two on people with diabetes [29,30] and 1 for hypertensive patients [31]. Despite cancer patients are considered as frequent and chronic users of NSAIDs but they are considered as exclusion criteria in most of studies enrolled in our meta-analysis.

At baseline, eleven studies enrolled CKD patients (of those, seven studies compared the CKD patients to matched non-CKD [13,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36] controls while four studies compared CKD patients within different stages [12,37,38,39].

CKD occurrence or progression was assessed by eGFR in most of studies and serum creatinine, ACR, spot proteinuria. Different definitions used for eGFR decline are shown in table (1). Out of 40 studies, 17 studies included an eGFR< 60 mL/min/1.73 m², 4 studies included any change in eGFR. 1 study for each of the following outcome: decrease of 29 mL, 15 mL/min/1.732, >5mL/min/1.73 m²/year, ≥ 10%, 25%, 50% from baseline to measure the occurrence or progression of CKD. Eleven studies estimated the GFR using the four-variable MDRD equation, 6 studies used Cockroft–Gault formula, 5 studies used CKD-EPI, 6 studies By using ICD9-CM codes whereas eGFR method was unclear in 6 studies [12,20,29,32,33, 36].

Creatinine, BUN used in 2 studies [15,40] to assess the outcome. The other 3 studies used secondary claim data for their outcome assessment [22,23,31]. However, these diverse outcome measurements collectively contribute to defining the occurrence or progression of chronic kidney disease (CKD).

In 22 studies, information regarding chronic NSAIDs exposure was obtained through interviews (either direct or telephone) or questionnaires. In the remaining studies, assessment relied on healthcare databases and medical records (Table 1).

Estimation of the effect size was done by HR in 11 studies, by OR in 35 studies and by mean eGFR difference in one study. We extracted the adjusted measure if available otherwise we used the crude (univariate measure).

Among 35 studies that measure the OR as effect size, [15,16,20,22,23,24,29,35,41,42,43] reported hazardous effect, [14,26,44,45] reported protective effect while [12,16,17,18,19,21,25,27, 32,33,34,36,40,46,47,48] reported no effect. Some of these studies provided inconsistent effect in different subsets of their populations [16,33,45,47].

Among 11 studies that measure the HR as effect size [37] reported protective effect, [39,49] reported no effect and [28,30,31,38,50,51] reported hazardous effect. One study provided different effect in subsets of its population [38].

Odds ratios were extracted or calculated from 35 studies or subpopulations within studies. Figure 3 forest plot showed significant association between chronic NSAIDs use and CKD (occurrence and progression); pooled odds ratio was 1.24 (95% CI: 1.11–1.39, p <0.001, I² = 91.21%), (May represent substantial between studies statistical heterogeneity).

Figure 3.

Forest plot of pooled ORs.

Hazard ratios were extracted or calculated from 11 studies or subpopulations within studies. Figure 4 forest plot showed significant association between chronic NSAIDs use and CKD (occurrence and progression); pooled Hazard ratio=1.50 (95thCI:1.31:1.7), P <0.001, I2=90.77% (considerable statistical heterogeneity).

Figure 4.

Forest plot of pooled HRs.

The review authors’ judgement about each risk of bias item for each included study is summarized in figure 11 (by study and by domain) and additional file 3 (by domain).

Figure 11.

Risk of bias (by study and by domain).

In most of the cohort studies included, the risk of bias was generally low regarding participant recruitment, confounder identification, follow-up duration, and statistical methods used. Specifically, the risk of bias for the measurement of exposure was low in 10 studies. Concerning incomplete follow-up and the strategies used to address it, the risk of bias was low in 7 and 6 out of the 18 studies, respectively.

In most of the case-control studies, the risk of bias was low across all domains, except for strategies to deal with confounders, where only 5 out of 10 studies had a low risk of bias. Similarly, in the majority of cross-sectional studies, the risk of bias was low for all domains except for the measurement of exposure, with only 5 out of 11 studies demonstrating a low risk of bias in this area.