Protocol and registration

The review protocol was registered by the International Prospective Register of Systematic Reviews (PROSPERO, protocol number CRD42017073532) and was previously published [27].

Differences between protocol and review

Some adjustments to the protocol version were made such as: i) eligibility criteria was altered to include RCT in which patients were diagnosed with rheumatoid arthritis regardless of diagnostic criteria, ii) trials of cross-over design were excluded due to the difficulty of using data to perform comparisons of results; iii) some adjustments were necessary in the search strategy; iv) we performed the indirect meta-analyzes due to the heterogeneity presented by RCT included in this review; and v) subgroup analyzes were not performed due to heterogeneity between studies.

Eligibility criteria

Information sources

We searched the following electronic databases with no restrictions regarding publication status or language: COCHRANE (CENTRAL), MEDLINE; EMBASE; CINAHL; Web of Science and Virtual Health Library. References for all included studies, other reviews, guidelines and related articles were searched examining reference lists. Ongoing studies were searched in the trial registry ClinicalTrials.gov and World Health Organization International Clinical Trials Registry Platform. The searching was carried out in order to identify all relevant publications up to December 14 of 2019.

Search

The search strategy was created using terms of the Medical Subject Headings (MeSH) and keywords. The associated keywords: i) intervention (anti-inflammatory agents); ii) condition (rheumatoid arthritis), and iii) methodological filters were applied to limit retrieval to RCT.

The search strategy was adapted for each database and designed with the assistance of a trained librarian (S1 File).

Study selection

Four reviewers (MDGP, SB-F, LGM, FCA), working in pairs and independently, screened titles and abstracts. The same reviewers, in pairs and independently, assessed eligibility of each full-text article. In case of duplicate publications, we would just include the article with most complete data, however this situation did not occur. Disagreements were resolved by consensus or by a third review author (CCB or LCL) if necessary.

Data collection process

All reviewers, in pairs and independently, extracted the data using standardized and pretested forms with instructions and contacted study authors to clarify any uncertainties. Studies where important data were incomplete or missing, we contacted the authors, however, we have not received reply from any authors. Whenever possible, we computed missing standard deviation (SD) from other statistics, such as standard error (SE) [24].^.For the studies that did not provide enough data, we verified whether these values could be extracted from graphs using web based tools (https://automeris.io/WebPlotDigitizer/).

The information extracted were: year and country of the publication, register protocol, study design, characteristics of the population (diagnostic criteria, pain relief medications, number of patients, mean and standard deviation age, percentage of women); interventions and comparators (drug, dose diary, via of administration, duration of the treatment in weeks), risk of bias and outcomes.

Geometry of the network

The data were summarized in a network meta-analysis. The model was proposed by Bucher et al. [29] and draws on both direct evidence (treatments compared in the same trial) and indirect evidence (different treatments studied in separate trials, but compared when they use a common comparator), with the benefit of randomization in each study retained.

Network meta-analysis using mixed treatment comparisons technique was carried out to unite in a single analysis direct and indirect evidences, the main objective being increasing precision of the estimation. The network diagram is made up of lines and nodes. In the diagram, the notes represent every intervention, and the size of the nodes means the number of participants. The lines indicate direct comparisons between different interventions and the thickness of the line means the amount of studies [30].

Risk of bias within individual studies

Using a modified version of the Cochrane collaboration risk of bias tool [24], the same reviewers assessed the risk of bias for each trial, in pairs and independently, according to the following criteria: random sequence; allocation concealment; blinding of the patients, care provider and outcome assessor for each outcome measure; incomplete outcome data; selective outcome reporting; and other biases.

To determine the risk of bias of a study, each criterion was rated as ‘definitely yes’, ‘probably yes’, being assigned a low risk of bias and ‘probably no’ and ‘definitely no’, assigned a high risk of bias [31]. Disagreements were resolved by consensus.

Incomplete outcome data, lost follow-up less than 10% and a difference of less than 5% in missing data in intervention and control groups were considered low risk of bias. In order to determine whether there was reporting bias or not, we first determined whether the protocol for the assessed RCT was published before recruitment of patients had started. For studies published after July 1st, 2005, we screened the Clinical Trial Register at the International Clinical Trials Registry Platform of the World Health Organization (http://apps.who.int/trialsearch) [32].

In cases where study protocol registration reports and safety results were not found, we used the classification for high risk of bias. The absence of a criterion for diagnosis of rheumatoid arthritis was identified as a possible source of bias and classified as high risk of bias in the criterion "other risks of bias".

The bias classification was done using the Review Manager 5 software and a third review author (CCB or MTS) carried out any final decisions when necessary.

Summary measures and methods of analyses

Analyses were carried out for each anti-inflammatory drug and for each outcome of interest. Estimates of comparative effectiveness were measured using standardized mean differences (SMD) with associated 95% confidence intervals (95% CI); and estimates of comparative safety were measured using odds ratio (OR) with 95% CI. Subgroup analysis could not be performed due to heterogeneity of the studies. The graphical representation of meta-analyses was not performed since there were few direct comparisons.

We are provided summary tables when the meta-analysis was not appropriate (S2 File). The analyses were carried out using Stata software (version 14.2). We adopted a comparison of mixed treatment with mixed generalized linear models to analyze the indirect and direct comparisons between the networks. The comparisons presented were derived from indirect and direct comparison, if available. The maximum restricted likelihood method was used to estimate the random effect model.

The inconsistency was evaluated using the loop-specific approach to evaluate the agreement between direct and indirect estimates, and it was regarded as a better consistency when the inconsistency factors (IF) was equal to zero [33]. We assess the possibility of intransitivity comparing trials participants between direct comparisons that contributed to an indirect comparisons [34]. Publication bias was graphically accessed via a comparison-adjusted funnel plot, and Egger’s test were applied to measure the asymmetry; the results of Egger test (P > .05) were defined as non-significant publication bias among included studies [35].

We calculated the relative ranking of agents for induction of clinical remission as their surface under the cumulative ranking (SUCRA), which represents the percentage of efficacy or safety achieved by a drug compared to other that is always the best without uncertainty (SUCRA = 100%) [36]. This parameter was used to estimate the average treatment ranking to explore potential orderings of treatments. The trial nodes that were not connected to the network were excluded.

Quality of evidence

We followed the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach to appraise the confidence in estimates derived from network meta-analysis of outcomes [34, 37]. RCT start at high confidence and can be rated down based on risk of bias, indirectness, imprecision, inconsistency and publication bias; they can then be graded at levels of moderate, low and very low confidence [38].

The risk of bias was evaluated for each outcome as low, moderate or high and is represented in figures of network meta-analysis in colors green, yellow and red, respectively. If direct and indirect estimates were coherent, then the higher of their ratings was assigned to the network meta-analysis estimates (S3 File).

Study selection

Of a total of 10,498 publications (reasons for exclusion are in S4 File), 26 studies met the inclusion criteria (NSAIDs = 21 and corticoids = 5) (Fig 1, S5 File).

Fig 1

Study flowchart.

VHL: Virtual Health Library; WHO: World Health Organization; NSAIDs: Non-steroidal anti-inflammatory drugs; corticoids drugs.

Non-steroidal anti-inflammatories for rheumatoid arthritis

Description of studies

Twenty-one trials (10,503 patients) were included in this review. Most studies reported the functional class of the disease (n = 17). Clinical trials comprised ten NSAIDs (aceclofenac, aspirin, celecoxib, diclofenac, etodolac, etoricoxib, indomethacin, ketoprofen, meloxicam, nabumetone, naproxen, piroxicam and tenoxicam). One study evaluated the use of a patch formulation, the others described oral administration. Follow-up time ranged from 14 to 182 days. Most studies reported the use of rescue medication (n = 16) and 17 studies described concomitant use of DMARD therapy, but did not describe the medicine used. The mean age of the patients varied between 46.9 and 58.7 years (Table 1).

Table 1

Characteristics of the included studies on non-steroidal anti-inflammatories for rheumatoid arthritis (n = 21).

Study	Functional capacity	Interventions (dose in mg/day)	Outcomes reported	Sample size (N)	Lost follow-up (%)	DMARDs Use	Rescue medication	Duration (weeks)	Mean age (years)	Woman (%)
Bernhard et al., 1987[56]	II or III	nabumetone 1,000, aspirin 900	5,6,7,8, 10	234	49.1	not allowed	acetaminophen	24	50.7	75
Collantes et al., 2002[40]	I, II or III	placebo, etoricoxib 90 naproxen 1,000	1,3,4,7,8,9,10	687	29.7	allowed	aspirin	12	52.3	NR
Emery et al., 1992[41]	I, II or III	nabumetone 2,000 naproxen 1,000	1,5,10	284	4.9	allowed	acetaminophen	12	53.2	NR
Emery et al., 1999[59]	I, II or III	celecoxib 400 diclofenac 150	1,2,3,4,5,7,8,10	497	31.8	allowed	NR	24	55.2	96.7
Furst et al., 2002[39]	NR	placebo, meloxicam 7.5, 15, 22.5 diclofenac 150	1,2,3,4,7,8, 10	888	0.7	allowed	acetaminophen	12	55.4	NR
Geusens et al., 2002[42]	I, II or III	Placebo naproxen 1,000	3,4,5,7,8,10	1023	NR	allowed	acetaminophen	12	53.6	82.8
Geusens et al., 2004[43]	I, II or III	naproxen 500 placebo	1,3,4,7,8,9,10	726	54.9	allowed	acetaminophen	26	53.5	88
Gibofsky et al., 2007[44]	II or III	naproxen 1,000 placebo	1,2,3,4,5,7,8,9,10	340	49.4	allowed	acetaminophen	12	55.9	68.5
Jacob et al., 1986[45]	I, II or III	placebo, etodolac 50, 100, 200, aspirin 3,900	1,3,4,5,6,10	264	42.4	allowed	acetaminophen	6	52.9	60.2
Kawai et al., 2010[46]	I, II, III or IV	placebo, ketoprofen 20	1,10	652	3.7	allowed	NR	2	58.7	85.8
Kornasoff et al., 1996[47]	I, II or III	aceclofenac 200 indomethacin 100	3,4,5,6,7,8,10	219	17.8	NR	acetaminophen	12	56.0	70.7
Krug et al., 2000[48]	I, II or III	nabumetone 2,000 naproxen 1,000	3,4,7,8,10	344	0.6	allowed	acetaminophen	12	54.0	70.9
Lightfoot, 1997[49]	NR	etodolac 400, 600 piroxicam 20	3,4,5,10	361	37.3	allowed	acetaminophen	12	57.0	84.2
Matsumoto et al., 2002[50]	I, II or III	placebo, etoricoxib 90 naproxen 1,000	1,3,4,7,8,9,10	448	68.7		aspirin	12	55.6	NR
Pasero et al., 1995[51]	NR	aceclofenac 200 diclofenac 150	1,5,6,10	327	7.6	NR	NR	24	50.7	81.3
Perez Ruiz; Alonso Ruiz; Ansoleaga, 1996[52]	NR	aceclofenac 200 tenoxicam 20	1,5,6,10	237	13	not allowed	acetaminophen	12	56.6	98.7
Shi et al., 2004[53]	I, II, III or IV	diclofenac 100, meloxicam 15 nabumetone 1,000, celecoxib 200	10	407	31.2	allowed	NR	24	46.9	76.9
Vasey et al., 1987[57]	II or III	nabumetone 1,000 naproxen 500	5,6,7,8,10	318	54.4	allowed	acetaminophen	24	55.0	NR
Williams et al., 2006[58]	I, II or III	placebo naproxen 500	7,8,10	1093	59.5	NR	NR	12	56.2	76.4
Wojtulewski et al., 1996[54]	I, II or III	meloxicam 7.5 naproxen 750	1,3,4,5,6,7,8,10	306	23.8	allowed	acetaminophen	26	NR	NR
Zhao et al., 2000[55]	I, II or III	Placebo celecoxib 100, 200, 400 naproxen 1,000	2,9,10	688	67	allowed	acetaminophen	12	54.5	NR

Notes. Outcomes reported: 1: pain; 2: functional disability score; 3: swollen joint count; 4: tender joint count; 5: morning stiffness; 6: grip strength; 7:physician assessment; 8:patient assessment; 9:quality of life scale; 10:adverse events. NR: not reported. Capacity Functional (Class I–Completely able to perform usual activities of daily living; Class II–Able to perform usual self-care and vocational activities but limited in avocational activities; Class III–Able to perform usual self-care activities but limited in vocational and avocational activities; and Class IV–Limited in ability to perform usual self-care, vocational, and avocational activities).

Risk of bias (Fig 2)

One study had minimum risk of bias [39]. Allocation concealment was insufficiently described in the majority them [40–55].

Fig 2

Risk of bias for studies on non-steroidal anti-inflammatories (n = 21).

Shi et al. (2004) [53] did not describe the blinding of patients or healthcare professionals due to the fact the study was an open trial in which both researchers and participants were aware of which treatments were administered.

Most of the RCT had attrition bias, since they did not report their results with the intention to analyse reported withdrawals >10% of sample and/or did not discuss the implications of patients lost follow-up, except for five studies [39, 41, 46, 48, 51].

Gibofsky et al. (2007) [44] and Zhao et al. (2000) [55], even though reported adverse events, did not evaluate other important outcomes (such as improvement of pain, swelling and duration of morning stiffness), and thus showed reporting bias. Also, two other clinical trials which described disease diagnostic did not cite the criteria used [56, 57].

Assessed outcomes and network meta-analysis

Nine NSAIDs were compared at 13 different dosages and placebo groups. Of the 12 trials, 8 (66.6%) were two-arm studies, whereas 4 (33.3%) were multiple-arm studies (Fig 3). Overall, 4,016 patients were included for “improvement of pain” (Fig 3A); 2,447 patients for “improvement of physical function” (Fig 3B); 4,962 patients for “number of tender/painful joints” (Fig 3C); 4,962 for “number of swollen joints” (Fig 3D), 4,152 for “patient’s global assessment” (Fig 3E); and 4,152 for “physician’s global assessment” (Fig 3F).

Fig 3

Evidence structure of eligible comparisons for network meta-analysis: Effectiveness outcomes.

A—Pain (9 studies, 10 non-steroidal anti-inflammatories—NSAIDs, 22 arms, 4,016 patients); B—Physical function (4 studies, 5 NSAIDs, 11 arms, 2,447 patients); C- Number of tender/painful joints (8 studies, 6 NSAIDs, 21 arms, 4,962 patients); D—Number of swollen joints (8 studies, 6 NSAIDs, 21 arms, 4,962 patients); E—Patient’s Global Assessment (6 studies, 6 NSAIDs, 17 arms, 4,152 patients); F—Physician’s Global Assessment (6 studies, 6 NSAIDs, 17 arms, 4,152 patients).

The outcomes of studies that could not be included in the network meta-analysis are described in S2 File.

Pain

This outcome included 13 studies [39–46, 50–52, 54, 59]. Naproxen 1,000 reduced pain compared to placebo (SMD: -10.28, 95% CI: -20.39; -0.17) (evidence of very low quality) (Fig 4, S3 File). Heterogeneity cannot be estimated due to insufficient number of studies. According to the Egger´s test (p = 0.042), the results may be influenced by small-size effects (S6 File).

Fig 4

Comparative effectiveness outcomes between drugs using network meta-analysis.

Comparisons should be read from left to right. Standardized mean difference (SMD) for comparisons are located in the common cell between the column-defining and row-defining treatment. Numbers on highlighted background are statistically significant. A. Pain; B. Physical function; C. Number of tender/painful joints (upper right quarter) and number of swollen joints (lower left quarter); D. Patient’s global assessment (upper right quarter) and physician’s global assessment (lower left quarter).

Safety of the interventions

All studies were included in the meta-analysis. Overall, 10,072 patients reported a number of adverse events for 12 different NSAIDs (56 arms). Of the 21 trials, 10 (41.6%) were two-arm studies and 11 (58.3%) were multiple-arm studies (Fig 5). Although adverse events were reported for most drugs, only etoricoxib 90 was associated with more adverse events compared to placebo (RR: 4.43, 95% CI: 1.22; 16.08) (evidence of low quality) (Fig 6). Heterogeneity cannot be estimated due to insufficient number of studies. According to the Egger´s test (p = 0.718), the results indicate no obvious publication bias (S6 File).

Fig 5

Evidence structure of eligible comparisons for network meta-analysis: Adverse events.

Lines connect the interventions that have been studied in head-to-head (direct) comparisons in the eligible randomized controlled trials. The width of the lines represents the cumulative number of randomized controlled trials for each pairwise comparison and the size of every node is proportional to the number of randomized participants (sample size).

Fig 6

Comparative adverse events between drugs using network meta-analysis.

Comparisons should be read from left to right. Odds ratio for comparisons are located in the common cell between the column-defining and row-defining treatment. Numbers on highlighted background are statistically significant.

In most of the studies, gastrointestinal adverse events were the most commonly reported by patients using NSAIDs. Abdominal pain, diarrhea, dyspepsia and nausea were the most frequent events reported in 18 studies [39–45, 47–50, 52–54, 56–59]. NSAIDs responsible for the highest incidence of these events were diclofenac [39, 53, 59] and naproxen [40–44, 48, 50, 58] at any dose.

Hypertension [40, 42, 44] and headache [40, 44, 49, 57–59] were commonly reported adverse events. No study reported serious adverse events leading to death or hospitalization. Early discontinuation due to treatment failure or to adverse events did not differ statistically between the groups and was not associated with any specific NSAID.

Ranking of treatments and outcomes (Table 2)

Values of SUCRA provided the hierarchy of 24 treatments on the outcomes assessed based on absolute rank probabilities. Tenoxicam 20, nabumetone 2,000 and aceclofenac 200 were most effective at reducing pain (4.3%, 4.6% and 4.8%, respectively).

Table 2

Bold numbers on highlighted background are first in ranking. For the safety outcome, the highlight represents the safety. For effectiveness outcomes, the highlight represents the most effective.

Ranking of treatments for the outcomes assessed.

Interventions	Safety		Improvement of pain		Improvement of physical function		Number of tender/painful joints		Number of swollen joints		Patient’s Global Assessment		Physician’s Global Assessment
	SUCRA	Mean Rank (SE)	SUCRA	Mean Rank (SE)	SUCRA	Mean Rank (SE)	SUCRA	Mean Rank (SE)	SUCRA	Mean Rank (SE)	SUCRA	Mean Rank (SE)	SUCRA	Mean Rank (SE)
Celecoxib 200 mg	78.1	6.0 (0.3)	-	-	-	-	-	-	-	-	-	-	-	-
Placebo	70.4	7.8 (0.3)	10.7	9.9 (0.2)	10.7	6.4 (0.1)	12.9	8.0 (0.2)	13.0	8.0 (0.2)	4.0	7.7 (0.1)	2.3	7.8 (0.1)
Nabumetone 2.0g	69.9	7.9 (0.3)	63.9	4.6 (0.2)	-	-	-	-	-	-	-	-	-	-
Ketoprofen 20 mg	66.9	8.6 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Celecoxib 800 mg	66.1	8.8 (0.3)	-	-	-	-	-	-	-	-	-	-	-	-
Naproxen 1.0g	64.3	9.2 (0.3)	-	-	73.2	2.6 (0.1)	65.9	3.7 (0.1)	54.8	4.6 (0.1)	63.9	3.5 (0.1)	52.8	4.3 (0.1)
Meloxicam 7.5 mg	60.1	10.2 (0.4)	35.9	7.4 (0.2)	37.0	4.8 (0.1)	36.2	6.1 (0.2)	55.2	4.6 (0.1)	31.0	5.8 (0.1)	44.3	4.9 (0.1)
Celecoxib 400 mg	59.3	10.4 (0.4)	45.2	6.5 (0.2)	45.6	4.3 (0.1)	55.1	4.6 (0.2)	42.1	5.6 (0.2)	57.7	4.0 (0.1)	30.0	5.9 (0.1)
Etodolac 200 mg	52.9	11.8 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Aceclofenac 200g	52.3	12.0 (0.4)	62.4	4.8 (0.2)	-	-	-	-	-	-	-	-	-	-
Naproxen 750 mg	51.2	12.2 (0.4)	55.2	5.5 (0.2)	-	-	79.6	2.6 (0.2)	99.2	1.1 (0.1)	-	-	-	-
Diclofenac 150 mg	51.0	12.3 (0.4)	55.0	5.5 (0.2)	43.9	4.4 (0.1)	50.0	5.0 (0.1)	32.4	6.4 (0.1)	58.3	3.9 (0.1)	57.9	3.9 (0.1)
Meloxicam 22.5 mg	50.4	12.4 (0.4)	49.5	6.0 (0.2)	72.2	2.7 (0.1)	37.8	6.0 (0.2)	52.8	4.8 (0.1)	52.4	4.3 (0.1)	58.7	3.9 (0.1)
Etodolac 100 mg	49.0	12.7 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Meloxicam 15 mg	49.3	12.7 (0.4)	53.9	5.6 (0.2)	67.4	3.0 (0.1)	26.6	6.9 (0.2)	18.7	7.5 (0.1)	47.7	4.7 (0.1)	57.4	4.0 (0.1)
Etodolac 50 mg	48.7	12.8 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Naproxen 500 mg	43.8	13.9 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Nabumetone 1.0g	40.8	14.6 (0.4)	51.4	5.9 (0.2)	-	-	-	-	-	-	-	-	-	-
Indomethacin 100mg	38.6	15.1 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Tenoxicam 20 mg	38.9	15.1 (0.4)	67.1	4.3 (0.2)	-	-	-	-	-	-	-	-	-	-
Diclofenac 100 mg	26.8	17.8 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Etoricoxib 90 mg	25.8	18.1 (0.4)	-	-	-	-	86.1	2.1 (0.1)	81.9	2.4	85.0	2.0 (0.1)	96.6	1.2 (0.1)
Aspirin 3.600 mg	24.1	18.5 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-
Aspirin 3.900 mg	21.5	19.1 (0.4)	-	-	-	-	-	-	-	-	-	-	-	-

Notes. SUCRA: surface under the cumulative ranking curve. Mean Rank: average ranking of treatments according to their relative effectiveness. The first ranked treatment is most likely to be the most effective treatment regarding a particular outcome compared to other treatments in the network. Numbers on highlighted background are statistically significant. SE: standard error.

Best results for improvement of physical function were observed for naproxen 1,000 (2.6%), meloxicam 22.5 (2.7%) and meloxicam 15 (3.0%).

Regarding number of tender/painful joints and swollen joints, etoricoxib 90 (2.1%) and naproxen 750 (1.1%) had the highest improvement rates, respectively.

As for patient’s global assessment and physician’s global assessment, etoricoxib 90 was considered the best intervention for both variables (2.0% and 1.2%, respectively). Celecoxib 200, placebo and nabumetone 2,000 were associated with a smaller number of adverse events and had the best safety profile (6.0%, 7.8% and 7.9%, respectively).

Corticoids for rheumatoid arthritis

Description of studies

Five trials (1,544 patients) were included. No study reported the functional class of the disease. These studies investigated the drugs prednisone 5, 7.5, 10 and 15 and prednisolone 7.5, administered orally. Patient follow-up ranged from 12 to 104 weeks. No studies reported the use of rescue medication and four studies described concomitant use of DMARD therapy. The mean age of the participants ranged from 39.9 to 58 years (Table 3).

Table 3

Characteristics of the studies on corticoids anti-inflammatories for rheumatoid arthritis (n = 5).

Study	Functional capacity	Interventions	Outcomes reported	Sample size (N)	Lost follow-up (%)	DMARD Use	Rescue medication	Duration (weeks)	Mean age (years)	Woman (%)
Bakker et al., 2012[12]	NR	methotrexate* and prednisone 10, methotrexate* and placebo	1,2,5	236	27.9	allowed	NR	52	53.5	60.1
Buttgereit et al., 2013[60]	NR	prednisone 5, placebo	2,5	350	7.7	allowed	NR	12	57.3	84
Choy et al., 2008[11]	NR	methotrexate alone** methotrexate** and ciclosporin*** methotrexate** and prednisolone methotrexate**, ciclosporin*** and prednisolone$	1,2,3,4,5	467	18.8	allowed	NR	52	54	69.5
Ding et al., 2012[61]	NR	prednisone 7.5, prednisone 15 placebo$ $	5	266	5.6	not allowed	NR	12	43	85.3
Hafstrom et al., 2014[62]	NR	prednisolone 7.5 placebo	1	225	46.2	allowed	NR	104	54.5	64

Notes.

*dose increased by 5 mg/week until remission

**starting at 7.5 mg/week, increasing incrementally up to target dose of 15 mg/week

***ciclosporin started 3 months after methotrexate (initial dose 100 mg/day, increased gradually up to target dose of 3 mg/kg daily)

^$60 mg/day initially, reduced to 7.5 mg daily from 6 to 28 weeks, stopped by week 34

^$ $all groups received leflunomide 20 mg/day and methotrexate 10 mg/day; Outcomes reported: 1. Progression of the disease assessed by radiological imaging of joints; 2. Disease activity; 3. Function; 4. Quality of life; 5. Adverse events. NR: not reported. Capacity Functional (Class I–Completely able to perform usual activities of daily living; Class II–Able to perform usual self-care and vocational activities but limited in avocational activities; Class III–Able to perform usual self-care activities but limited in vocational and avocational activities; and Class IV–Limited in ability to perform usual self-care, vocational, and avocational activities).

Risk of bias (Fig 7)

All of the assessed trials considered eligible were at high risk of bias, except for Choy et al. (2008) [11] which was at minimum risk of bias. Allocation concealment was insufficiently described in four studies [12, 60–62]. One study did not describe the blinding of patients and healthcare professionals and failed to report whether there were patients lost at follow-up [60].

Fig 7

Risk of bias for studies on corticoids (n = 5).

Bakker et al. (2012) [12], Buttegereit et al. (2013) [60], Ding et al. (2012) [61] and Hafstrom et al. (2014) [62] did not describe other important outcomes, such as reduction of pain, swelling and duration of morning stiffness or the main adverse events reported by patients, leading to reporting bias. Two clinical trials did not cite diagnostic criteria used [12, 62].

Summary of evidence and comparison of findings with previous studies

This systematic review assessed the available evidence on effectiveness and safety of NSAIDs and corticoids for the treatment of rheumatoid arthritis. RCT assessed the use of NSAIDs, mainly via oral administration. The main methodological flaws were absence of allocation concealment and high dropout rates; it was also often unclear how analyses were performed or how the studies dealt with missing data.

In general, it was observed that naproxen 1,000 improved physical function and reduced overall pain and number of painful and swollen joints, providing benefits according to the patient’s and physician’s global assessment compared to placebo. Also, naproxen 750 was better than most of the NSAIDs at reducing the number of swollen joints (including naproxen 1,000), except in comparation to etoricoxib 90. However, the quality of evidence was very low overall. It can be observed that naproxen did not exhibit a dose-dependent behaviour as the 750 dose was more effective than the 1,000 dose for the number of swollen joints outcome.

Etoricoxib 90 was also reported to reduce the number of painful and swollen joints and provided better results according to the patient’s global assessment compared to placebo. In addition, it was better than both celecoxib 400 and naproxen 1,000, according to physician’s global assessment, but the quality of evidence was very low.

No study using NSAIDs reported “progression of disease as assessed by radiological imaging of joints”. Also, it was not possible to perform meta-analysis for the outcomes morning stiffness, average grip strength and improvement of quality of life.

In another review [63], celecoxib 200 showed significant pain reduction compared to placebo (11% absolute improvement; 95% CI 8% to 14%), which was not evidenced in our review. However, this study also related low quality of the evidence of your findings.

Etoricoxib 90 improved overall outcomes when compared to placebo, but did not have a good safety profile according to the ranking hierarchy of 24 treatments. Celecoxib 200, nabumetone 2,000 and placebo were associated with a smaller number of adverse events. The most common side effects reported by patients using NSAIDs were gastrointestinal events and no SAE were documented.

Findings from safety data of another systematic review showed that piroxicam or etodolac did not lead to clinically significant adverse events and that mild adverse events were evidenced by the use of celecoxib or etoricoxib such as nausea, vomiting and headaches. However, the quality of the evidence did not allow confirmation as to which NSAID is safest [6].

Evidence of moderate to low quality reported fewer withdrawals among patients which had received celecoxib compared to those receiving placebo. Celecoxib 200 was associated with a smaller number of adverse events, but it was observed that patients developed more ulcers and severe short-term symptoms [63].

Prednisolone 7.5 and prednisone 5.0, 7.5 and 10 showed benefits in reduction of joint erosion [11, 12, 62] and disease activity [12, 60]. Due to different interventions employed in these clinical trials, it was not possible to perform meta-analysis. Besides that, issues regarding risk of bias decreased reliability of such findings.

Regarding safety of corticoids, gastrointestinal adverse events were the issues most often reported by patients using these drugs. Corticoids are usually used in association with methotrexate, and thus attention should be focused on the adverse effects of combined therapy.

The effects of prednisone or similar corticoid preparations administered alongside standard therapy (doses ranging from 270mg to 5,800mg, over the first year) reduced progression of joint erosion in rheumatoid arthritis, according to a systematic review published in 2007 [22]. Only one clinical trial included in this review [11] was also selected for our study due to differences in eligibility criteria. In our findings, two studies used prednisone at doses of 5–10 demonstrating benefit of this drug for this outcome [61, 62].

Systematic review showed that prednisolone administrated orally at low doses (maximum of 15 mg/day) was more effective than placebo and NSAIDs at decreasing joint tenderness and pain, but the quality of the evidence was not assessed by the authors [23]. According to older evidences, after six months, the effects of prednisolone (maximum 15 mg/day) were significantly better compared to placebo for number of painful joints, number of swollen joints, pain and physical function [21]. However, small sample size, high risk of bias and the fact that quality of evidence was not assessed compromise such findings.

Strengths and limitations of this study

The methodology employed in this review includes explicit eligibility criteria, comprehensive and extensive database searches and independent and paired evaluation to select studies. Meticulous search and selection processes were carried out, and we are confident that trials meeting the inclusion criteria were included in the review. Robust statistical techniques were used to assess risk of bias of the included studies.

The decision to provide network analysis is relevant as it provides information in situations where primary evidence is scarce or nonexistent and allows for more accurate estimates of effects [64]. Even though a broad search strategy was carried out and that we did not exclude any studies due to language barriers or to date of publication, some bias cannot be considered inexistent.

For some studies that did not provide variance measures necessary for meta-analysis, we estimated missing data with approximate values ​​derived graphically from the studies themselves. This could have created some bias, but the overall impact on the estimation of statistically significant differences between groups is probably small.

Some studies did not record the concomitant use of other analgesic agents and/or the dose used, which can mislead outcomes for pain measurement, for example.

It’s important to highlight that anti-inflammatories drugs are still used as adjuncts to control symptoms of disease. However, the impact of use mainly of NSAIDs has decreased during the recent years, since more potent DMARDs allow for better disease control as well as glucocorticoids are mainly used for bridging while the effects of DMARDs take place.

The quality of the primary studies included in this review was a limiting factor for proper analysis to be carried out. Besides that, the diversity of drugs and doses used and absence of report of some outcomes could have decreased the quality of our findings.

Implications for clinical practice and research

NSAIDs and corticoids are used as adjuncts to treat symptons of rheumatoid arthritis. Therefore, these anti-inflammatory drugs should not be used for long-term. According to our findings, even though etoricoxib 90 and naproxen 1,000 appear to be effective and celecoxib 200 seems to be the safest option, the low quality of the evidence suggests that futures RCT can demonstrate different results.

Prednisone and prednisolone seem to be effective and generate only mild adverse events in patients with rheumatoid arthritis, when used at lower doses. However, due to the high risk of bias of RCTs and not being possible to perform the meta-analysis of the studies involving these drugs, the evidence of such findings could not be confirmed.

We observed that the clinical trials had methodological limitations, differences regarding the drugs studied and their doses, concurrent use with other drugs and differents outcomes measured contributed to limiting the conclusions on our findings. Future trials should consider these limitations and therefore obtain long-term data. Also, adequate follow-up and larger sample sizes are required for future research.