Types of studies

We included studies if they met all the following criteria:

• randomised controlled trials (RCTs) with an adequate method of concealment of randomisation (e.g. use of an interactive voice response system and sealed, opaque envelopes);

• double‐blind trials, in which both participants and personnel, or outcome assessors, were blinded to treatment;

• placebo‐controlled trials;

• parallel‐group or cross‐over trials. For cross‐over trials, we only included the results gained from the first treatment arm;

• trials that had a treatment duration of at least eight weeks, and recorded baseline seizure data.

Types of participants

People of any age with focal epilepsy, defined as seizures that only affect part of the brain at onset, including simple focal, complex focal, or secondary generalised tonic‐clonic seizures, who failed to respond to one or more antiepileptic drugs; or with drug‐resistant focal epilepsy, defined as people who had not achieved sustained seizure freedom during adequate trials of two tolerated and appropriately chosen antiepileptic drugs were eligible for inclusion (International League Against Epilepsy definition (Kwan 2010)).

Types of interventions

The intervention treatment group received perampanel, in addition to one or more existing antiepileptic drugs, from the time of randomisation.

The control group received a matched placebo, in addition to one or more existing antiepileptic drugs, from the time of randomisation.

Types of outcome measures

Electronic searches

We searched the following databases on 20 October 2022:

• Cochrane Register of trials (CRS Web), using the search strategy given in Appendix 1;

• MEDLINE (Ovid, 1946 to 19 October 2022), using the search strategy given in Appendix 2.

CRS Web includes randomised or quasi‐randomised, controlled trials from Specialized Registers of Cochrane Review Groups, including Epilepsy; the Cochrane Central Register of Controlled Trials (CENTRAL); PubMed; Embase; ClinicalTrials.gov; and the World Health Organization International Clinical Trials Registry Platform (ICTRP). In MEDLINE (Ovid) the coverage end date always lags a few days behind the search date.

We also conducted electronic database searches on: 24 November 2015, 26 June 2017, 25 August 2018, 19 September 2019, and 1 June 2021.

Searching other resources

We reviewed the reference lists of the retrieved trials to search for any additional reports of relevant trials. We contacted the pharmaceutical manufacturer Eisai Inc., which produces perampanel, to obtain relevant data, and the original investigators of the relevant trials to identify any additional published or unpublished data. We also handsearched abstracts published from June 2018 to October 2022 from International Epilepsy Congress meetings.

Selection of studies

Two review authors (RB and RH) independently read the titles and abstracts of all reports identified by the search strategies implemented in August 2018, September 2019, June 2021, and October 2022 and assessed their eligibility for the review. Two separate review authors (Y Xiao and J Huang) screened the titles and abstracts of all reports identified by the previous searches, conducted in November 2015 and June 2017. Once we had retrieved the full text of all the potentially relevant papers, each review author independently evaluated the full text of each paper for inclusion. We resolved any disagreements by discussion, or with a third author (JW) acting as an arbitrator.

Data extraction and management

Two review authors (RB and RH) independently extracted the following data, using a data extraction form:

• participants: total and number in each group, age, gender, seizure types, seizure frequency at the time of randomisation, exclusion criteria;

• methods: study design, study duration, randomisation method, allocation concealment method, blinding methods;

• interventions: details of perampanel treatment, such as administration method, dosage, and duration; number of background drugs;

• outcomes: proportion of participants with a 50% or greater reduction in seizure frequency; proportion of participants with seizure freedom during the whole treatment period; proportion of participants who withdrew during the course of the treatment period for any reason; proportion of participants experiencing dizziness, somnolence, fatigue, and headache, or other clinically important adverse effects reported in the included trials;

• other: country and setting, publication year, sources of funding, intention‐to‐treat (ITT) analysis.

The review authors had no disagreements about the data extraction.

Assessment of risk of bias in included studies

Two review authors (RB and RH) independently assessed the risk of bias of the included trials using the RoB 1 tool recommended by the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011). The RoB 1 tool consists of seven parameters: 1. random sequence generation, 2. allocation concealment, 3. blinding of participants and personnel, 4. blinding of outcome assessment, 5. incomplete outcome data, 6. selective outcome reporting, and 7. other bias. For each entry, we provided our judgement ('low risk' of bias, 'high risk' of bias, or 'unclear risk' of bias), followed by a description of the design, conduct, or observations that underlie the judgement (Higgins 2011). The review authors had no disagreement in assessing the risk of bias.

Measures of treatment effect

We analysed data according to the ITT principle. For dichotomous outcomes (50% or greater reduction in seizure frequency, seizure freedom, treatment withdrawal), we used risk ratios (RRs) with 95% confidence intervals (CIs) to analyse the outcomes. For individual adverse effects, we used 99% CIs to make allowance for multiple hypothesis testing. For the primary analysis, we assumed that participants who did not complete follow‐up, or for whom there were inadequate seizure data, were non‐responders.

Unit of analysis issues

We did not include any cross‐over studies in this review, therefore, did not encounter any unit of analysis issues. If we were to identify a cross‐over study that was eligible for inclusion in a future review update, we would extract data from the first treatment period only, and then treat these data as though they had been derived from a parallel‐group study. This would prevent data from the same participant contributing to both the intervention and control treatment groups, and avoid any unit of analysis issues.

Dealing with missing data

We attempted to obtain additional information and missing data from the study authors and the sponsoring company through personal communication. Some unpublished data for subgroup analysis were not available from the sponsor at present. Once the data become available, we will add them to an update of this review. We used ITT analysis to account for missing data.

Assessment of heterogeneity

We evaluated clinical and methodological heterogeneity of included trials by comparing the characteristics of participants (age, baseline seizure frequency, epilepsy duration, and geographical origin of included participants), interventions (administration dose and duration, co‐interventions), and study design (randomisation, allocation concealment, and blinding methods) between studies. We evaluated statistical heterogeneity using the Chi² test and the I² statistic. A P value equal to, or greater than 0.10 from the Chi² test indicated no significant statistical heterogeneity. When the P value was less than 0.10 from the Chi² test, we interpreted the heterogeneity according to the percentage ranges of the I² statistic, as follows (Higgins 2011):

• 0% to 40%: might not be important;

• 30% to 60%: may represent moderate heterogeneity^a;

• 50% to 90%: may represent substantial heterogeneity^a;

• 75% to 100%: represents considerable heterogeneity^a.

^aThe importance of the observed value of the I² statistic depends on the magnitude and direction of effects and the strength of the evidence for heterogeneity (e.g. P value from the Chi² test or a CI for the I² statistic).

Assessment of reporting biases

We requested trial protocols in order to examine whether the outcomes of interest, stated in the trial protocol, were reported and analysed in the respective study reports. When the protocol was not provided, we compared the methods section of the study reports against the results section and stated that this was the case in the risk of bias table.

We included a limited number of studies so we were unable to investigate potential publication biases using funnel plots and visually inspect them for asymmetry to assess reporting bias, according to the approach outlined in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2011).

Data synthesis

We used Review Manager 5 (RevMan 5) to synthesise the available data (Review Manager 2014). We used a fixed‐effect or a random‐effects model dependent mainly on the results of the Chi² test and the I² statistic for heterogeneity (Higgins 2011). If a P value was greater than 0.10, indicating no significant statistical heterogeneity, we used a fixed‐effect model. If a P value was less than 0.10 but the I² statistic indicated no important or moderate heterogeneity, we used a fixed‐effect model. If a P value was less than 0.10 and the I² statistic indicated substantial heterogeneity, we first explored factors contributing to clinical heterogeneity to determine whether a subgroup analysis according to clinical subgroups was needed. If the substantial heterogeneity could not readily be explained, we adopted a random‐effects model.

Subgroup analysis and investigation of heterogeneity

We planned to conduct subgroup analysis according to the different ages of participants (children younger than 17 years versus adults), and the different doses of perampanel (e.g. 2 mg/day, 4 mg/day, 8 mg/day, or 12 mg/day). However, because of the limited data gained from the included studies, we only conducted the subgroup analysis according to different doses of perampanel, not according to different ages of the participants.

Sensitivity analysis

We intended to conduct sensitivity analyses, whereby we would exclude studies that had missing data for the primary outcome (i.e. data were not available in the original reports, or were not supplied by the study authors following correspondence) from the meta‐analysis. All seven studies provided data for 50% or greater reduction in seizure frequency therefore it was not necessary to conduct the planned sensitivity analyses.

Summary of findings and assessment of the certainty of the evidence

We used the GRADE approach to summarise findings, as detailed in the GRADE handbook (Schünemann 2013). We exported data from Review Manager 5 to GRADEpro GDT software to create a summary of findings table (GRADEpro GDT; Review Manager 2014). We assessed the certainty of the evidence for outcomes across five domains (risk of bias; inconsistency; indirectness; imprecision; publication bias), according to the GRADE guidelines (Balshem 2011).

We selected the following outcomes as the most important, to include in the summary of findings table: 50% or greater reduction in seizure frequency, seizure freedom, treatment withdrawal due to any reason, treatment withdrawal due to adverse effects, and the proportion of participants who experienced at least one adverse effect. We also included the two adverse effects that demonstrated the largest effect size for all doses of perampanel versus placebo (i.e. the proportion of participants who experienced ataxia, and the proportion of participants who experienced dizziness). This enabled us to assess whether the estimated effect size was likely to be an accurate reflection of the true treatment effect.