Success rate of proximal tooth-coloured direct restorations in primary teeth at 24 months: a meta-analysis

The aim was to determine the survival of tooth-coloured restorative materials in proximal restorations of primary teeth at 24 months of follow-up and the influence of the following variables: use of coating, use of cavity conditioner, use of rubber dam isolation, the cavity form, the dentist’s experience and the methodological characteristics of the studies. We conducted a search until May 2019, obtaining 16 articles from which 30 independent studies were extracted, which were considered as units of analysis. Four outcome measures were extracted from each study: retention, marginal integrity, anatomic form, and absence of recurrent caries. Separate meta-analyses were carried for each outcome and multiple meta-regression model was applied. The outcomes with the highest mean success rates were absence of recurrent caries and anatomic form. The type of material significantly influenced success rates. The best materials were resin-based material plus total-etching adhesion and resin-modified glass ionomer cement (RMGIC), and the worst high viscosity glass ionomer cement (HVGIC). Atraumatic restorative treatment (ART) had a lower success rate than the conventional cavity form. RMGIC had the best clinical performance and HVGIC the worst. The form of the cavity, blinding and the experience of the operator were the variables that influenced success rates. Proximal primary molar restorations should be performed with RMGIC as it combines good mechanical performance of the resins together with the prevention of secondary caries of glass ionomers.

inclusion and exclusion criteria. Using the components of the Participants, Interventions, Comparisons, Outcomes, and Study designs (PICOS) system 26 , the studies to be included in this meta-analysis the studies had to meet the following criteria: • Participants: primary teeth with proximal caries in children aged 2-14 years.
• Outcomes: Success rate at 24 months of follow-up for retention, marginal integrity, anatomic form, and absence of recurrent caries. • Study designs: randomized controlled trials (RCTs), non-randomized controlled trials (nRCTs), and uncontrolled trials.
In addition, published and unpublished studies were accepted. We excluded reviews, clinical cases, in vitro studies, observational studies, studies of permanent teeth, studies evaluating the survival of materials in classes I, III, IV and V, studies dealing only with amalgam or cermet restorations or stainless steel crowns, and studies with a follow up other than 24 months. Search strategy. We exhaustively searched the following electronic databases: PubMed, MEDLINE, SciELO, Embase, Scopus, WOS, LILACS and BBO. We also carried out a manual search to find studies not included in the electronic databases. We adapted the search strategy to the requirements of each database. The references of the studies recovered were also reviewed to identify studies that might fulfil the selection criteria. The search languages were English, Spanish and Portuguese, and the search covered 1985 to May 2019. We included the following search terms, making the appropriate adaptations to the language required by the different databases: Data extraction. The most relevant data were extracted from the articles and included in a database which collected the main characteristics of the intervention and the evaluation criteria of the restoration materials. The complete database is shown in Supplementary Dataset file. For the extraction of the results, within the same article, each restoration material, or the same material under different experimental conditions, was considered as an independent study.
The four evaluation categories used were: • Retention. No detachment of the material and no partial or total fracture that required repair or a new restoration. • Marginal integrity. No discoloration, filtration or defects in marginal adaptation. • Anatomic form. No alterations in the shape and texture of the surface of the restoration.
• Absence of secondary caries.
The most relevant characteristics of the intervention studied were: • Conditioning of the cavity, referring only to glass ionomer materials.
• Complete isolation with rubber dam.
• Cavity form, differentiating between cavities made according to ART principles (manual cavities made with spoon) and conventional cavities (cavities made by using rotary instruments and burrs). • Operator experience, considering non-graduate student as inexperienced and graduate dentists as experienced.
The following patient characteristics were extracted: mean and SD of age (in years) and sex (% male). The following methodological characteristics were extracted: design (RCT, nRCT, or uncontrolled trial), random assignment, blinded assessment, reporting bias, and sample size. Financial sources and year of the study were also extracted. To check the reliability of the extraction of the study characteristics, two independent coders doubly coded all studies. The results were highly satisfactory overall, with kappa coefficients ranging between 0.927 and 1.0 (mean = 0.993) for categorical variables, and intra-class correlations between 0.805 and 1.0 (mean = 0.979) for continuous variables. Inconsistencies between the coders were resolved by consensus.
Statistical analysis. From each sample, the proportions of success at 24 months were extracted for the four outcome measures: retention, marginal integrity, anatomic form, and absence of recurrent caries. Separate meta-analyses were made for each outcome. To normalize the distribution of the success rates, they were transformed using the logit event rate: Lp = Ln [p/(1 − p)], with p being the success rate, Ln the natural logarithm, and Lp the logit event rate. In each meta-analysis, a random-effects model was assumed 38 and, consequently, the logit event rates were weighted by the inverse variance, defined as the sum of the within-study and between-studies variance. The latter was estimated using the DerSimonian and Laird method 39 . The sampling variance of each logit event rate, V (Lp), was calculated as: V (Lp) = 1/(np) + 1/[n (1 − p)], with n being the sample size. Subsequently, the results were back-transformed to success rates to facilitate interpretation by means of: p = e Lp /(e Lp + 1), with e being the base of the natural logarithm 40 . In each meta-analysis, the 95% confidence limits around the mean success rate were computed using the method proposed by Hartung 41 . The heterogeneity of the success rates was assessed by constructing a forest plot and calculating the Q statistic and the I 2 index. I 2 values of about 25%, 50%, and 75% can be considered as reflecting low, moderate, and high heterogeneity 42 . To test whether publication bias was a threat to the validity of the meta-analytic results, funnel plots were constructed applying the trim-and-fill method 43 .
If studies were found to have heterogeneity, moderator analyses were carried out using meta-regressions and weighted ANOVAs for continuous and categorical variables, respectively, assuming a mixed-effects model 44 www.nature.com/scientificreports www.nature.com/scientificreports/ moderators, respectively. The proportion of variance accounted for by the moderator variables was estimated using R 2 , an index that takes into account the total and residual between-studies variance 47 . Finally, to identify the study characteristics that best explained variability of the success rates, a multiple meta-regression model was applied. The moderator variables included in the model were selected taking into account both statistical and practical significance achieved in the previous analyses. All statistical analyses were carried out with the metafor package for R 48 .

Results
Results of the systematic search. Figure 1 shows a flowchart of the selection process. The search strategy yielded 920 articles: 220 from PubMed, 68 from Embase, 212 from WOS, 134 from MEDLINE, 2 from SciELO, 203 from Scopus, 33 from LILACS and 48 from BBO. After excluding 795 duplicate articles, 125 articles remained.
Seventy-one articles were accepted due to the title and, after reading the corresponding abstracts, 36 were excluded, due to: • Study in permanent teeth: 4.
• Comment on an article: 1.
• Language other than English, Spanish, Portuguese: 1. (Summary in English, full article in Korean) 64 .
• Identical to a previously selected item: 1. One article was found with a different title but with the same authors and an identical sample. We excluded one of the two articles 66 .
In nine studies, the dentists who evaluated the restorations were unaware of the material used, as they were triple blinded 11,12,16,31,32,34,36,37 . In 14 articles there was no reporting bias 11,12,16,18,20,27,29,[31][32][33][34][35][36][37] , and 15 articles did not declare whether there were conflicts of interest or not 11,12,16,18,20,[27][28][29][30][31][33][34][35][36][37] . The source of funding was public in six of the articles 11,16,28,33,34,37 and private or mixed in the remaining 10 12,18,20,27,29-32,35,36 . Thirty independent studies were extracted from the 16 articles selected. The most important characteristics of each study are shown in Table 1. Table 2 shows the mean success rates, the 95% confidence intervals, and the heterogeneity statistics (Q and I 2 ) for each outcome. The absence of recurrent caries and anatomic form were the outcomes with the highest mean success rates (p + = 0.909 and p + = 0.901, respectively). The mean success rates for marginal integrity and retention were 0.898 and 0.879, respectively. As retention was the main outcome, Fig. 2 shows a forest plot of the 27 retention success rates, and Supplementary File 3 shows the forest plots of the success rates for marginal integrity, anatomic form and absence of recurrent caries. The Q statistic was significant (p < 0.0001) for all four outcomes and the I 2 indices were >75% in all cases ( Table 2). The wide heterogeneity of the success rates was investigated by analysing the influence of moderator variables.

Analysis of publication bias.
To determine whether publication bias was a threat to the conclusions of the meta-analysis, funnel plots were constructed applying Duval and Tweedie's trim-and-fill method. Figure 3 shows the funnel plot of the retention success rates, with a slightly higher concentration of data on the right side of the mean success rate. By applying the trim-and-fill method, seven additional success rate estimates were imputed to achieve symmetry in the funnel plot. Adding the seven success rates led to a slight decrease in the mean success rate from the original 0.879 to 0.839 (95% CI: 0.764-0.893), implying a 4.8% decrease, which may be considered negligible. Therefore, publication bias did not threaten the overall success rate for retention outcome.
Similar analyses were carried out for the other three outcomes. Supplementary File 3 presents the funnel plots for marginal integrity, anatomic form, and absence of recurrent caries. For both marginal integrity and anatomic form, five new success rates were imputed to adjust the funnel plots to symmetry (see Supplementary Figs 4 and 5, respectively). The mean success rates for marginal integrity obtained with the 27 original success rates and after imputing data, were 0.898 and 0.881 (95% CI: 0.832-0.916), respectively, with a negligible decrease of 1.9% when data imputation was applied. For anatomic form, the mean success rates for the original and once imputed data were 0.901 and 0.876 (95% CI: 0.817-0.919), respectively, implying a negligible decrease of 2.9%. With respect to the absence of recurrent caries, the trim-and-fill method added eight success rates on the left side of the funnel plot to achieve symmetry (see Supplementary Fig. 6 in File 3), leading to a decrease in the mean success rate from 0.909 (with the 24 original success rates) to 0.878 (95% CI: 0.828-0.915) once data were imputed. In this case, the decrease of 3.5% can also be considered negligible. Therefore, these results enabled publication bias to be ruled out as a threat to the validity of the meta-analytic results.
Analysis of moderator variables. The analysis of moderator variables was carried out separately for the four outcomes. We only present here the results of the retention outcome, since the pattern of results was very similar for the remaining outcomes. However, the results of applying ANOVAs and simple meta-regressions for the marginal integrity, anatomic form and absence of recurrent caries are shown in Supplementary File 4. Table 3 presents the simple meta-regressions applied for each continuous moderator variable on the estimated retention success rates. Of the moderators analysed, only sample size was significantly negatively associated with the success rates (p = 0.001), accounting for 37% of variance. A marginally significant result was found for anatomic form (p = 0.052, accounting for 17% of variance; see Table 3 in Supplementary File 4).
Regarding categorical moderators, Table 4 shows the results of the weighted ANOVAs applied to the estimated retention success rates. The type of material significantly influenced the success rates (p = 0.011), explaining 36% of variance. Specifically, better results were found for resin-based material plus total-etching adhesion (p + = (2020) 10  The form of the cavity was also associated with the success rates (p < 0.001), accounting for 46% of variance: the mean retention success rate was lower for ART (p + = 0.649) than for conventional cavity design (p + = 0.936). Similar results were found for marginal integrity (p = 0.040, R 2 = 0.21; Table 2 in Supplementary File 4), anatomic form (p = 0.001, R 2 = 0.63; Table 4 in Supplementary File 4), and absence of recurrent caries (p = 0.046, R 2 = 0.16; Supplementary Table 6 in File 4). The dentist's experience was significantly associated with retention success rates (p = 0.048; Table 4), although it only explained 7% of variance. More experienced dentists had better success rates than non-experienced ones (p + = 0.895 vs.0.554, respectively). Of the methodological variables analysed, assessor blinding was significantly associated with retention success rates (p = 0.028, R 2 = 0.32; Table 4), with lower retention success rates when the assessor was blinded. These results were repeated for marginal integrity (p = 0.023, R 2 = 21; Table 2 in Supplementary File 4), anatomic form (p = 0.032, R 2 = 0.33; Table 4 in Supplementary File 4), and absence of recurrent caries (p = 0.002, R 2 = 0.28; Supplementary Table 6 in File 4). explanatory models. Although some of the moderators were significantly associated with retention success rates, none showed non-significant results in the model misspecification tests (Q E and Q W for meta-regressions and ANOVAs, respectively), suggesting residual heterogeneity among the success rates after including the moderator. Similar results were found for marginal integrity, anatomic form, and recurrent caries success rates (see Supplementary File 4). Therefore, multiple meta-regression models including the most relevant characteristics of the studies were applied to explain the variability in the different success rates.
The predictors included in the meta-regression models for explaining success rates were selected as a function of both statistical and practical significance achieved in the previous results of the ANOVAs and simple meta-regressions. A moderator variable was included in the model when the F statistic was significant and the R 2 index was >30%. With respect to retention success rates, four predictors were included: the sample size, the material (dichotomized as: 1 for resin-based material plus total-etching adhesion, RMGIC, and open sandwich restoration; and 0 for resin-based material plus self-etching adhesion and HVGIC), the form of the cavity (0: ART, and 1: conventional cavity design), and assessor blinding (0: no blinding, and 1: blinding).
Due to missing data in some variables, the number of studies included in the meta-regression was k = 27. The results are shown in Table 5. The full model showed a significant association with the retention success rates (p = 0.001), accounting for 45% of variance. However, once the other predictors were controlled for, none showed www.nature.com/scientificreports www.nature.com/scientificreports/ a significant association with the success rates. This may be due to collinearity among the predictors. Inspection of the bivariate correlations between the predictors showed a significant association between the form of the cavity and the type of material (r = 0.66, p < 0.001), as ART usually uses HVGIC material. When the form of the cavity was removed from the meta-regression model, the type of material (dichotomized) was significant (p = 0.029) as was the full model, F (3, 23) = 7.34, p = 0.001, accounting for 49% of the variance. Sample size and assessor blinding were not significant in the multiple meta-regression model.
The results of the multiple meta-regressions for marginal integrity, anatomic form, and recurrent caries success rates are shown in Supplementary Tables 7-9 in File 4. With respect to marginal integrity, two moderators were included in the model: cavity form and assessor blinding. There was a trend to significance for the full model  www.nature.com/scientificreports www.nature.com/scientificreports/ (p = 0.051), with 22% of variance accounted for, while neither predictors were significant once the influence of each on the other was controlled for. Cavity form and assessor blinding were included in the meta-regression model for anatomic form success rates. A significant association was found for the full model (p = 0.005) and 61% of variance was accounted for. In addition, the cavity form was significantly associated with anatomic form success rates (p = 0.017) after controlling for assessor blinding. With respect to recurrent caries, three moderators were included in the meta-regression model: use of coat, cavity form, and assessor blinding. The full model was significant (p = 0.007), with 83% of variance accounted for. Of the three moderators, only the use of coat was negatively associated with recurrent caries success rates (p = 0.022).

Discussion
Since the FDA advised, in July 2010, that dental amalgam should not be used in children aged <6 years, due to its greater sensitivity to the potential toxic effects of mercury 67 , and the EU banned its use, from July 1, 2018, for the restoration of primary teeth, in children aged <15 years, and pregnant or breastfeeding women 68 , paediatric dentists need to know what the best alternative material for the restoration of primary teeth is. Therefore, this meta-analysis tried to answer the question: which tooth-coloured restoration material has the best clinical behaviour in proximal restorations of primary teeth at 24 months? We chose proximal restorations because they have the highest failure rate, especially when functional, due to the presence of antagonistic teeth 52 .
We studied the success rate at 24 months ought to the dropout rates of study subjects increase over time because there is a marked increase in the rates of physiological exfoliation, characteristic of childhood growth 53 . Moreover, after that time many studies have reported a high level of failure, depending on clinical variables and patient related factors. In general, the annual failure rate was 17% for restorations in primary molars 69 , although some studies recorded losses of around 50% at 24 months of follow up 23,57 . The success rate for class II in primary teeth was 68% at 18 months 70 and 52´58% at 36 months 71 . This way, a study found that after a 7-year follow-up, only 1% of initial restorations completed the study 72 .
For better understanding of the results, the materials used in the studies analysed were divided into five groups: materials containing resin bonded with total etching (composite, giomers, compomer and fluid composite); resin-containing materials bonded with a self-etching adhesive (composite, compomer and fluid composite), RMGLC, HVGIC, and open sandwich technique (RMGIC as a base and composite as a restorative material). Regardless of the material used, retention of proximal restorations was the most affected, with 12.5% of restorations lost within 24 months of placement. Marginal integrity, conservation of the anatomical shape and the absence of secondary caries, in descending order, were affected to a lesser extent.
The success of a restoration depends on factors such as the material used, the state of the tooth, the experience of the operator and, the patient's collaboration. This last aspect is of paramount importance in paediatric dentistry, since children's behaviour largely determines the selection of the material and the technique to be used to restore a tooth, conditioning, finally, the wide variations in the success rate between the different materials and studies 21 . Unifying all the materials included in the study in materials that contain resin and those that do not, the meta-analysis showed that those containing resin had a higher success rate in the four clinical categories studied. RMGIC had the highest success rate followed by resin-based materials used with total etching and self-etch adhesives. The material with the lowest success rate was HVGIC. Although only retention was significant, the trend in all clinical categories studied (marginal integrity, anatomic form and recurrent caries) was the same.
A meta-analysis showed that RMGIC performed better than conventional GIC for class II restorations in primary teeth 9 . Another study also observed a better performance of RMGICs compared with conventional GICs and composites for class II primary teeth, arguing that RMGICs combine the best properties of both materials: on the one hand they have the good mechanical properties of composites and, on the other, the self-adhesive properties of GICs 73 . Vitremer (3 M ESPE, St. Paul, MN, USA) was the RMGIC used in the largest number of studies included in our meta-analysis. Of the five studies, in four it was used together with Vitremer Primer (3 M ESPE, St. Paul, MN, USA), a light cure adhesive that contains, among other things, 2-hydroxyethyl methacrylate monomer (45-55%) and the copolymer of acrylic and itaconic acids (10-30%). The joint use of the Primer gives it a greater adhesive capacity and reduced sensitivity to the exchange of water with the surrounding environment by rapid photopolymerization 72 .
HVGIC showed, in our study, the worst retention rate (0.671), similar to other studies that found a 65% survival rate in multi-surface ART restorations using HVGIC 74 or failures in 30% of class II ART restorations during the first month using HVGIC 52 . A study that compared the use of HVGIC using ART versus a conventional cavity  Table 3. Results of the simple meta-regressions of continuous moderator variables on the retention success rate estimates. k = number of studies. bj = regression coefficient of each predictor. F = Knapp-Hartung's statistic for testing the significance of the predictor (the degrees of freedom for this statistic are 1 for the numerator and k -2 for the denominator). p = probability level for the F statistic. QE = statistic for testing the model misspecification. R 2 = proportion of variance accounted for by the predictor. ***p < 0.0001. www.nature.com/scientificreports www.nature.com/scientificreports/ technique concluded that, for HVGIC, there is a greater risk with ART cavities than with conventional cavities in primary tooth decay treatments, and success rates in classes II are worse than in class I 75 . This association between the poor results of ART and the poor clinical performance of HVGIC is strongly supported by our results, in which the bivariate correlations between predictors revealed a strong relationship (r = 0.66, p < 0.001) between the cavity form and the type of material. This may be due to the peculiarities of the ART technique, which uses HVGIC as a restorative material and does not allow complete isolation, meaning saliva contamination of the operative field is very frequent 76 and the survival of multi-surface restorations could be more dependent on the material, operator and control of the operative field than single-surface restorations 57 . However, in a randomized controlled study 10 year follow-up to evaluate the durability and clinical performance of a HVGIC (processed with a resinous coating) compared with a micro filled composite resin in conventional class I y II cavities, in permanent teeth isolated with cotton rolls and suction devices, no significant differences were found for both restorative materials in terms of marginal adaptation, anatomical form, secondary caries, postoperative sensitivity, surface texture, and retention. The HVGIC could be also considered a good alternative to amalgam 77 .

Scientific RepoRtS
Although we found a higher retention success rate when the dentist was experienced, we found no influence of the type of isolation. In "in vitro" studies, adhesion to enamel and dentin of materials containing composites are very sensitive to salivary contamination, although the results of "in vivo" studies are unclear. A systematic 95% CI  www.nature.com/scientificreports www.nature.com/scientificreports/ review found that, in the longevity of direct dental restorations made with a tooth-coloured material in primary teeth, the use of a rubber dam did not influence the results compared with the use of roll of cotton together with a saliva ejector 78 . A meta-analysis concluded that there are few studies with a very low quality of evidence on the advantages of using the rubber dam compared to cotton rolls together with a saliva ejector on the survival of restorations and neither, the previous application of cavity conditioner nor the final application of coating on the glass ionomer, influenced the success of the clinical variables studied 79 .
Our results showed that, of all the methodological variables used, the type of trial, randomization, reporting bias and the source of funding did not influence the retention, marginal integrity, anatomic form and absence of recurrent caries success-rates, indicating that the type of study did not influence the results of the different materials and supports our decision not to use quality scales, where the type of study is decisive in the final score, to decide the inclusion of the works in our meta-analysis, but include all studies that met the inclusion criteria considering as an independent study each of the materials used or each of the different experimental conditions used with the same material.
Evaluator blinding was important in determining the success of the restoration in the four clinical criteria evaluated in the opposite direction as expected, since when the evaluator knew the material the degree of success measured was lower than when they did not. The influence that we have observed in the triple blind indicates that in the next studies that are carried out, the evaluator must be masked. Meta-regression of each moderator variable showed only the sample size had a significant negative association with the estimated retention success rate indicating that, as the sample size increased, the retention success rate fell. One of the limitations of the studies included in our meta-analysis was the small sample size [median 56, min-max 12-210]: studies were initiated with a small sample size with a high loss rate, which was very high at 24 months. Multiple meta-regressions to determine the predictors that explained the success of proximal restorations showed a similar behaviour for retention, marginal integrity, anatomic form, and recurrent caries success rates. All included the cavity form and assessor blinding as predictors. In addition, the sample size, the type of material and the use of coat were predictors of the index of retention.
A possible bias is the different criteria used in the studies to assess the clinical performance of the restorations. Thirteen used the various modifications of the USPHS criteria, one study used the USPHS criteria plus ART criteria, one study used the FDI criteria and one study used an own system. As there may be differences in the assessment of success depending on the criteria used, and the lack of sensitivity of the ART criteria in detecting improvements in the clinical performance of the materials currently used for dental restorations 80 , we unified the criteria used in the selected articles into four categories: retention, marginal integrity, anatomic form and absence of recurrent caries. This enabled comparison of the studies and elimination of the bias that the use of different evaluation systems could introduce.
After RMGIC, the material that presented the best results was the open-type sandwich method followed by resin-containing materials adhered with the total etching technique. The sandwich method, which uses RMGIC as a cavitary base material and composite as a surface material, seems to have the advantages of both materials. However, only one study was included in this meta-analysis, and therefore further studies are required to determine whether the good results are maintained. Likewise, studies that combine the speed and simplicity of the ART technique with RMGICs, which had the best clinical behaviour in proximal restorations at 24 months, are required.
In conclusion, the null hypothesis of our meta-analysis was disproved, as the index of success of proximal restorations in primary teeth at 24 months was found to depend on the type of coloured material used. The materials with the highest success rates were those that contained resin. Of these, RMGIC performed best. The highest failure rate was for HVGIC and with the cavity made using ART, which were significantly correlated. The shape of the cavity, triple blinding and the experience of the operator had the most influence on the success rates of proximal restorations.

Data availability
All data generated or analysed during this study are included in this published article (and its Supplementary Information files).   Table 5. Results of the multiple meta-regression model applied on the retention success rates, taking as predictors the sample size, the type of material (dichotomized), the cavity form, and the assessor blinding (k = 27). b j = regression coefficient of each predictor. t = statistic for testing the significance of the predictor (with 22 degrees of freedom). p = probability level for the t statistic. F = Knapp-Hartung's statistic for testing the significance of the full model. Q E = statistic for testing the model misspecification.