2.1

Search

Three electronic databases (Medline via PubMed, Embase via Ovid, and Cochrane Central) were searched independently by two reviewers (SR, FS) to identify potentially eligible studies. A three-pronged search strategies using the Boolean operator AND was used, combining: disease centered terms (caries, decay, carious), intervention centered terms (sealing, infiltration) and trial centered terms (randomized, random, trial, participants, clinical studies, lesion progression). No controlled vocabularies (MeSH etc.) were employed. Reviewers screened the identified titles and abstracts of records against the inclusion criteria. If either author found a record potentially eligible, full-texts were assessed, again independently and in duplicate. Inclusion was decided by the two reviewers in consensus, or in consultation with a third reviewer (GG). Further articles were identified by examining the references of retrieved full-text studies. The search was not restricted by language. Neither authors nor journals were blinded to reviewers.

2.2

Selection

Study design and interventions: Parallel group and split-mouth randomized controlled clinical trials (RCTs) comparing micro-invasive strategies against each other, against NI or placebo were included. We excluded cross-over trials from this review as the condition, dental caries, cannot return to baseline level following the initial intervention, neither can the treatment (i.e. sealing or infiltration etc. cannot be reversed).

Participants: Participants with permanent or primary teeth, and proximal carious lesions with intact surface status (non-cavitated), presumed clinically (visually-tactile if possible) or by radiographic diagnosis (with a lesion extension to which cavitation is unlikely) on posterior teeth. A sensitivity analysis to evaluate the impact of different lesion depth (into enamel or into dentin radiographically) was planned, but not conducted given insufficient data reporting depth. Treatment of buccal/vestibular lesions using micro-invasive means was not considered.

Outcome measures: Our primary outcome was lesion progression, assessed using digital subtraction radiography (DSR) or, if not available, pairwise reading or, if not available, lesion staging (for example according to radiographic enamel depth – E1, E2 outer/inner enamel half – or dentin depth – D1-3 in dentin thirds). We chose to prioritize outcome measures this way as this implies a ranking of sensitivity (DSR being most sensitive). Our secondary outcomes were:

• –

  subjective evaluation of the treatments by participants and dentists

• –

  efficiency (time needed for the intervention), costs or cost-effectiveness (regardless of how effectiveness was defined); and

• –

  any safety issues (e.g., allergies) that are noted related to the interventions.

2.3

Data extraction and management

Duplicative data extraction was performed independently by two calibrated reviewers (GG, FS). Disagreements were resolved through discussion. We did not contact study authors. Only data from the most recent publication of a study (longest follow-up) was extracted. Data was recorded according to guidelines outlined by the Cochrane Collaboration . For the present evaluation, the following items were extracted:

• •

  Study details – design, year of publication, number randomized/analysed, study setting (e.g. school, practice);

• •

  Population – age, gender and number of participants, baseline caries experience or caries risk;

• •

  Potentially important effect modifiers (dentition; lesion depth if given);

• •

  Interventions – detailed description of the interventions, including number of teeth treated per participant;

• •

  Outcome data – details of the outcomes reported and the outcomes measures, including method of assessment and timepoint(s) assessed;

• •

  Funding sources, declarations/conflicts of interest.

2.4

Risks of bias, heterogeneity and transitivity

Two review authors (SR, FS) independently assessed the risks of bias of each included study using the domain-based tool of the Cochrane Handbook for Systematic Reviews of Interventions. The presence of heterogeneity was assessed by evaluating the trial and study population characteristics across trials. The assumption of transitivity was assessed by evaluating the distribution of potential effect modifiers across the trials and trial arms forming the network for NMA.

2.5

Treatment effects

Our primary outcome (lesion progression) was analysed by calculating the odds ratio (OR). No analysis of continuous outcomes was performed. Results from NMA were presented as summary relative effect sizes (OR) for each possible pair of interventions. Mean ranks and the cumulative ranking curve (SUCRA) were used to rank strategies according to their efficacy for arresting proximal lesions.

Where multiple lesions within a person were evaluated, the individual was considered to be the statistical unit and the lesions to be clustered within an individual. To account for clustering we computed the design effect . Therefore the size of each trial was reduced to its effective sample size , sometimes referred to as Donner or C adjustment . The design effect is given by 1 + ( M − 1 ) × I C C , where M is the average cluster size and ICC is the intracluster correlation coefficient. As all included studies were split-mouth studies, we computed the average cluster size M by l e s i o n s × 0.5 p a t i e n t s .

According to Masood et al. we set the ICC to 0.2 The number of lesions and the number experiencing the event were then divided by the design effect.

Multi-arm studies were treated as multiple independent two-arm studies in pairwise meta-analyses. A continuous correction of +1 was used for trial arms with zero events. Missing data were not imputed.

2.6

Data synthesis

Pairwise random effects meta-analyses used for direct treatment comparisons was implemented using the metafor package in R (R Foundation for Statistical Computing, 2017). As described, all included studies used a split-mouth design, but reported data only in marginal form. To properly compute ORs we applied the Becker-Balagtas method , again setting the ICC at 0.2 .

Indirect and mixed comparisons were performed using Bayesian random-effects modelling and Markov Chain Monte Carlo simulations using JAGS implemented in the R package gemtc 0.8-2 . Networks of interventions were constructed by plotting different treatments (as nodes) and comparisons (as edges) . Binomial likelihood was used to model the data . To fit the model, we used non-informative priors, for the basic parameters from a normal distribution N(0,1000), and a uniform prior U(0,4) for the random-effects standard deviation. The first 20,000 iterations were discarded as ‘burn-in’ and then further 80,000 iterations were undertaken for 4 chains with a thinning of 1. The convergence was assessed based on the Brooks-Gelman-Rubin criteria and inspection of trace plots.

Median OR and their 95% credible intervals (95% CrI) were reported. Credible intervals are the range of estimated parameters after exclusion of extreme values . Different strategies were ranked according to their probability of having the lowest versus the highest odds of arresting lesions , and the average rank calculated. The surface under the cumulative ranking (SUCRA) line was plotted and the area under the plot (SUCRA value) calculated.

2.7

Statistical heterogeneity

In pairwise meta-analyses we estimated different heterogeneity variances for each pairwise comparison. In NMA we assumed a common estimate for the heterogeneity variance across the different comparisons. For pairwise meta-analysis, Chi ² test was used to assess heterogeneity, with p < 0.1 indicating significant heterogeneity. I ² statistic and its 95% confidence interval was used to quantify heterogeneity. For NMA, a total I-squared value for heterogeneity in the network was calculated . In order to evaluate the level of (in)consistency, we applied node-splitting, which evaluates one comparison at a time by separating the direct evidence on that comparison from the network of indirect evidence . Comparison adjusted funnel plot were assessed to check for the existence of publication bias .

2.8

Trial sequential analysis (TSA)

Conventional meta-analysis uses Z-values to compare two interventions, with Z = 0.0 indicating no difference between intervention groups. If Z exceeds ±1.96, a difference is traditionally assumed to be statistically significant (p ≤ 0.05, two-sided test). TSA plots the series of Z-values against the number of event, patients or information accumulated with time . This ‘cumulative Z-curve’ is assessed regarding its relation to the conventional significance boundaries (Z = ±1.96), the required information size (RIS), and the trial sequential monitoring boundaries (TSMB) for benefit, harm, or futility.

The RIS was calculated based on type I error risk of α = 0.05, a type II error risk of β = 0.20 (equivalent to a power of 0.80) and the control event proportion. The relative risk reduction (RRR) was based on an a priori defined worthwhile interventional effect of 20%. It should be noted that smaller intervention effects might well be relevant. This, however, would increase the RIS even further . The required RIS was further adjusted for the diversity in the meta-analysis (diversity-adjusted RIS: DARIS). The Lan-DeMets version of the O’Brien–Fleming function was used for calculating the TSMBs. Results of cumulative Z-value crossing the conventional boundary of significance (Z = ±1.96) but not the TSMBs for benefit or harm were defined as spuriously significant. Firm evidence was assumed to be reached when the Z-curve crossed the TSMB for benefit or harm before the DARIS was reached. Firm evidence of futility was confirmed by the Z-curve crossing the TSBM for futility. TSA 0.9 (Copenhagen Trial Unit, Copenhagen, Denmark) was used .

2.9

Grading of evidence

The certainty of the accrued evidence was graded according to the GRADE working group of evidence using Grade Profiler 3.6 GRADE following guidelines for rating the certainty in estimates from network meta-analyses .

2.1

Search

Three electronic databases (Medline via PubMed, Embase via Ovid, and Cochrane Central) were searched independently by two reviewers (SR, FS) to identify potentially eligible studies. A three-pronged search strategies using the Boolean operator AND was used, combining: disease centered terms (caries, decay, carious), intervention centered terms (sealing, infiltration) and trial centered terms (randomized, random, trial, participants, clinical studies, lesion progression). No controlled vocabularies (MeSH etc.) were employed. Reviewers screened the identified titles and abstracts of records against the inclusion criteria. If either author found a record potentially eligible, full-texts were assessed, again independently and in duplicate. Inclusion was decided by the two reviewers in consensus, or in consultation with a third reviewer (GG). Further articles were identified by examining the references of retrieved full-text studies. The search was not restricted by language. Neither authors nor journals were blinded to reviewers.

2.2

Selection

Study design and interventions: Parallel group and split-mouth randomized controlled clinical trials (RCTs) comparing micro-invasive strategies against each other, against NI or placebo were included. We excluded cross-over trials from this review as the condition, dental caries, cannot return to baseline level following the initial intervention, neither can the treatment (i.e. sealing or infiltration etc. cannot be reversed).

Participants: Participants with permanent or primary teeth, and proximal carious lesions with intact surface status (non-cavitated), presumed clinically (visually-tactile if possible) or by radiographic diagnosis (with a lesion extension to which cavitation is unlikely) on posterior teeth. A sensitivity analysis to evaluate the impact of different lesion depth (into enamel or into dentin radiographically) was planned, but not conducted given insufficient data reporting depth. Treatment of buccal/vestibular lesions using micro-invasive means was not considered.

Outcome measures: Our primary outcome was lesion progression, assessed using digital subtraction radiography (DSR) or, if not available, pairwise reading or, if not available, lesion staging (for example according to radiographic enamel depth – E1, E2 outer/inner enamel half – or dentin depth – D1-3 in dentin thirds). We chose to prioritize outcome measures this way as this implies a ranking of sensitivity (DSR being most sensitive). Our secondary outcomes were:

• –

  subjective evaluation of the treatments by participants and dentists

• –

  efficiency (time needed for the intervention), costs or cost-effectiveness (regardless of how effectiveness was defined); and

• –

  any safety issues (e.g., allergies) that are noted related to the interventions.

2.3

Data extraction and management

Duplicative data extraction was performed independently by two calibrated reviewers (GG, FS). Disagreements were resolved through discussion. We did not contact study authors. Only data from the most recent publication of a study (longest follow-up) was extracted. Data was recorded according to guidelines outlined by the Cochrane Collaboration . For the present evaluation, the following items were extracted:

• •

  Study details – design, year of publication, number randomized/analysed, study setting (e.g. school, practice);

• •

  Population – age, gender and number of participants, baseline caries experience or caries risk;

• •

  Potentially important effect modifiers (dentition; lesion depth if given);

• •

  Interventions – detailed description of the interventions, including number of teeth treated per participant;

• •

  Outcome data – details of the outcomes reported and the outcomes measures, including method of assessment and timepoint(s) assessed;

• •

  Funding sources, declarations/conflicts of interest.

2.4

Risks of bias, heterogeneity and transitivity

Two review authors (SR, FS) independently assessed the risks of bias of each included study using the domain-based tool of the Cochrane Handbook for Systematic Reviews of Interventions. The presence of heterogeneity was assessed by evaluating the trial and study population characteristics across trials. The assumption of transitivity was assessed by evaluating the distribution of potential effect modifiers across the trials and trial arms forming the network for NMA.

2.5

Treatment effects

Our primary outcome (lesion progression) was analysed by calculating the odds ratio (OR). No analysis of continuous outcomes was performed. Results from NMA were presented as summary relative effect sizes (OR) for each possible pair of interventions. Mean ranks and the cumulative ranking curve (SUCRA) were used to rank strategies according to their efficacy for arresting proximal lesions.

Where multiple lesions within a person were evaluated, the individual was considered to be the statistical unit and the lesions to be clustered within an individual. To account for clustering we computed the design effect . Therefore the size of each trial was reduced to its effective sample size , sometimes referred to as Donner or C adjustment . The design effect is given by 1 + ( M − 1 ) × I C C , where M is the average cluster size and ICC is the intracluster correlation coefficient. As all included studies were split-mouth studies, we computed the average cluster size M by l e s i o n s × 0.5 p a t i e n t s .

According to Masood et al. we set the ICC to 0.2 The number of lesions and the number experiencing the event were then divided by the design effect.

Multi-arm studies were treated as multiple independent two-arm studies in pairwise meta-analyses. A continuous correction of +1 was used for trial arms with zero events. Missing data were not imputed.

2.6

Data synthesis

Pairwise random effects meta-analyses used for direct treatment comparisons was implemented using the metafor package in R (R Foundation for Statistical Computing, 2017). As described, all included studies used a split-mouth design, but reported data only in marginal form. To properly compute ORs we applied the Becker-Balagtas method , again setting the ICC at 0.2 .

Indirect and mixed comparisons were performed using Bayesian random-effects modelling and Markov Chain Monte Carlo simulations using JAGS implemented in the R package gemtc 0.8-2 . Networks of interventions were constructed by plotting different treatments (as nodes) and comparisons (as edges) . Binomial likelihood was used to model the data . To fit the model, we used non-informative priors, for the basic parameters from a normal distribution N(0,1000), and a uniform prior U(0,4) for the random-effects standard deviation. The first 20,000 iterations were discarded as ‘burn-in’ and then further 80,000 iterations were undertaken for 4 chains with a thinning of 1. The convergence was assessed based on the Brooks-Gelman-Rubin criteria and inspection of trace plots.

Median OR and their 95% credible intervals (95% CrI) were reported. Credible intervals are the range of estimated parameters after exclusion of extreme values . Different strategies were ranked according to their probability of having the lowest versus the highest odds of arresting lesions , and the average rank calculated. The surface under the cumulative ranking (SUCRA) line was plotted and the area under the plot (SUCRA value) calculated.

2.7

Statistical heterogeneity

In pairwise meta-analyses we estimated different heterogeneity variances for each pairwise comparison. In NMA we assumed a common estimate for the heterogeneity variance across the different comparisons. For pairwise meta-analysis, Chi ² test was used to assess heterogeneity, with p < 0.1 indicating significant heterogeneity. I ² statistic and its 95% confidence interval was used to quantify heterogeneity. For NMA, a total I-squared value for heterogeneity in the network was calculated . In order to evaluate the level of (in)consistency, we applied node-splitting, which evaluates one comparison at a time by separating the direct evidence on that comparison from the network of indirect evidence . Comparison adjusted funnel plot were assessed to check for the existence of publication bias .

2.8

Trial sequential analysis (TSA)

Conventional meta-analysis uses Z-values to compare two interventions, with Z = 0.0 indicating no difference between intervention groups. If Z exceeds ±1.96, a difference is traditionally assumed to be statistically significant (p ≤ 0.05, two-sided test). TSA plots the series of Z-values against the number of event, patients or information accumulated with time . This ‘cumulative Z-curve’ is assessed regarding its relation to the conventional significance boundaries (Z = ±1.96), the required information size (RIS), and the trial sequential monitoring boundaries (TSMB) for benefit, harm, or futility.

The RIS was calculated based on type I error risk of α = 0.05, a type II error risk of β = 0.20 (equivalent to a power of 0.80) and the control event proportion. The relative risk reduction (RRR) was based on an a priori defined worthwhile interventional effect of 20%. It should be noted that smaller intervention effects might well be relevant. This, however, would increase the RIS even further . The required RIS was further adjusted for the diversity in the meta-analysis (diversity-adjusted RIS: DARIS). The Lan-DeMets version of the O’Brien–Fleming function was used for calculating the TSMBs. Results of cumulative Z-value crossing the conventional boundary of significance (Z = ±1.96) but not the TSMBs for benefit or harm were defined as spuriously significant. Firm evidence was assumed to be reached when the Z-curve crossed the TSMB for benefit or harm before the DARIS was reached. Firm evidence of futility was confirmed by the Z-curve crossing the TSBM for futility. TSA 0.9 (Copenhagen Trial Unit, Copenhagen, Denmark) was used .

2.9

Grading of evidence

The certainty of the accrued evidence was graded according to the GRADE working group of evidence using Grade Profiler 3.6 GRADE following guidelines for rating the certainty in estimates from network meta-analyses .