Prior to the start of the ESBA program, the authors DC, GD, GZ, and IC (dentists and hygienists of the Unit of Pediatric Dentistry and Special Needs Patients, Policlinico Umberto I) completed a 1-year training period with a licensed therapist, with formal training and certified experience in autism spectrum disorders (specifically in behavioral, relational, and communication strategies used in evidence-based interventions), to acquire the relational skills necessary for effective communication and interaction with the children participating in the program. Skill acquisition was ensured through supervised training, direct observation, and competency-based evaluation during the 1-year period. The dentists and hygienists involved in the program were required to demonstrate adequate proficiency in the use of communication strategies, behavior support techniques, and child–clinician interaction methods before participating in the program.

In preparation for clinical activities, intra- and inter-examiner reliabilities were assessed to ensure consistency in caries detection and oral health status evaluation. Calibration was performed on a sample of pediatric patients not involved in the study, following the World Health Organization guidelines for oral health surveys. The assessment was conducted on a tooth-by-tooth basis using Kappa statistics (22). Kappa values of 0.82 (for intraexaminer) and 0.84 (for interexaminer) were obtained.

For each of the 85 children in the final sample, a clinical history was retrospectively obtained from the patient's’ clinical charts that the dentist completed at the end of each visit during the period of study (June 2013–March 2020). In this respect, it should be noted that in the ESBA program, treatments followed a structured sequence of weekly sessions over three main phases and a follow-up session, as described in Table 1. Throughout all phases, a licensed autism therapist monitored the sessions, provided feedback to the medical team, and suggested strategies for improvement. When a lack of cooperation prevented a certain treatment to be given and the treatment required urgent implementation (e.g., trauma, pain, and rapid progression of disease or infection), the child was referred for treatment under general anesthesia.

The clinical charts, used to obtain the clinical history of the children in the study, contained information about the treatments that the dentist had completed during each visit, as well as the dentist's annotations about both positive and negative child behaviors (e.g., acceptance, reluctance to accept or refusal of the treatments, crying, willingness to comply with the dentist, negative attitudes, and fear).

Based on the clinical history of the 85 children in the study, the dentists retrospectively rated each child’s behavior and cooperation, assigning (for the first visit and for the end of each of the three phases in the study) a value corresponding to one of the four categories of the Frankl scale (definitely negative, negative, positive, and definitely positive). The Frankl scale is a widely used and reliable behavior rating system in both clinical dentistry and ASD research (18, 23, 24).

From the clinical history, an additional outcome variable, Type of treatment, was retrospectively assessed at the end of each of the three phases established in the ESBA. For a given child, the variable Type of treatment represents the type of treatments received by the child during a particular phase according to the following classifications: approach, checkups, preventive procedures, dental cleaning, conservative procedures, surgical procedures, and anesthesia. It must be noted that during a certain phase, a patient can undergo more than one treatment.

From the participants' records at the first visit, 10 baseline covariates were also retrospectively obtained. These covariates were related to demographic attributes (age at the first visit, gender, nationality), medical history (type of delivery, vaccines, medications, previous experiences of general anesthesia), and clinical data (dental health status, gingival status, observational plaque index).

Information on demographic attributes and medical history was obtained from a questionnaire, especially designed for the study, completed by parents in a face-to-face interview with the dentist at the first visit. Clinical data were obtained from the records of the initial examination performed at the Unit of Pediatric Dentistry and Special Needs Patients, Policlinico Umberto I.

The recorded values of the sum of decayed, missing, and filled teeth in the primary dentition (dmft) and in the permanent dentition (DMFT), denoted by DMFT/dmft, were used to determine the dental health status. The recorded values of plaque and gingival indexes (with four and three categories, respectively, described in Table 2) were used to assess the general prevalence of periodontal health in accordance with the World Oral Health Survey (WHO) guidelines (25, 26).

As a first step in the data analysis, univariate distributions and summary statistics for the variables in the study were computed separately for the two subcohorts of children defined by their uncooperative and cooperative behaviors at the initial visit. The first cohort (children with initially uncooperative behavior) comprised the 45 children in the sample with the Frankl scale rating of negative or definitely negative at the first visit. The second cohort (children with initially cooperative behavior) comprised the 40 children in the sample with the Frankl scale rating of positive at the first visit (none of the children in the sample were rated as definitely positive at the first visit).

The distinction between the two groups is important because the effectiveness of the ESBA has a different meaning for the two subcohorts. As the primary goal is successful dental treatment, cooperative behavior is required on the part of children. For initially uncooperative children, we consider the ESBA approach to be successful if, after intervention, the children's behavior changes to cooperative (the Frankl scale changes from negative or definitely negative to positive or definitely positive during the program). On the other hand, for initially cooperative children, we consider the ESBA approach to be successful if their cooperation (already present before intervention) improves during the program (the Frankl scale changes from positive to definitely positive during the program). The underlying rationale is that, for initially cooperative children, improvement in cooperation reflects an improved perception of the dental experience, which, in turn, may simplify and accelerate the administration of the required treatments and reduce distress.

In order to distinguish between the two cohorts of children and measure the success of the ESBA approach, we defined the new variable, Init.Rating, and the notion of improvement. The Init.Rating variable takes the value 0 for children with initially uncooperative behavior (with the Frankl scale rating of negative or definitely negative at the first visit) and the value 1 for children with initially cooperative behavior (with the Frankl scale rating of positive at the first visit). The notion of improvement takes into account the question whether the ESBA approach has achieved its goals by the end of a given phase: (i) A child demonstrating initially uncooperative behavior (Init.Rating = 0) is considered to have improved by the end of phase k (k = I, II, III) if their Frankl scale rating is either positive or definitely positive. (ii) A child demonstrating initially cooperative behavior (Init.Rating = 1) is considered to have improved by the end of phase k (k = I, II, III) if their Frankl scale rating is definitely positive.

As a preliminary assessment of the efficacy of the ESBA, for each of the two cohorts of initially uncooperative and cooperative children, we computed the distribution of the type of treatment performed and the percentage of children who had improved by the end of phases I, II, and III. In order to test the statistical significance of the differences between the Frankl scale ratings at baseline (i.e., at the first visit) and the corresponding rate at the end of each of the three phases, the Stuart–Maxwell test (27) was used.

A more rigorous assessment of the efficacy of the ESBA was conducted through statistical modeling, relating the response variable (Frankl scale assessment) to the set of explanatory variables in the study. A categorical predictor with m levels was represented using m−1 dummy variables. In all cases, the reference category was the first level of the variable, as reported in Table 2.

In order to exploit the ordering of the components of the response variable (Frankl scale assessment is an ordinal categorical variable with four levels), an analysis was performed by fitting a set of 3,328 cumulative logit models with proportional odds, the most widely used logistic model for ordinal response (27). Each of the 3,328 models corresponded to a different set of explanatory variables and interactions to be included in the modeling. All models were fitted using the R software (version 4.3.1) and the clm() function from the Ordinal package available from the Comprehensive R Archive Network (CRAN) at https://CRAN.R-project.org/package=ordinal (28).

Model selection was performed using the Bayesian Information Criterion (BIC), which is a penalized likelihood method that favors parsimonious models and has a Bayesian interpretation in terms of the posterior probability of the models (29). The best model is the one with the minimum BIC.

In order to account for uncertainty in BIC scores due to sampling variability, the “one standard error rule” (1se rule) for model selection was applied. The rule was first proposed in (30) for tree size selection. In our study, the rule was implemented in three steps. Step 1: 2,000 bootstrap “copies” of the original dataset were produced. Step 2: For each bootstrap sample, the 3,328 cumulative logit models were refitted, and the corresponding BIC score were computed. In this way, for each model, 2,000 bootstrap BIC scores were produced, for which the average and the standard deviation were computed. Step 3: The model with the minimum average BIC was identified. Let BIC¯ and se be the average and the standard deviation of the 2,000 BIC scores for this model, respectively. The best BIC-1se model was defined as the most parsimonious model whose average BIC fell within one standard deviation of BIC¯ [i.e., within the interval ( BIC¯ -se, BIC¯ +se)].

The best BIC-1se model was then used to compute 95% confidence intervals for the probability that an initially uncooperative child and an initially cooperative child would improve by the end of phase I, II, and III of the ESBA. The 95% confidence intervals were obtained using the bootstrap and the percentile method (31). Again, 2,000 bootstrap “copies” of the original dataset were produced. For each bootstrap sample, the selected BIC-1se model was refitted and the corresponding predicted probabilities of interest were computed. In this way, for each event of interest, 2,000 predicted bootstrap probabilities were produced. The corresponding 95% confidence interval was obtained by computing the 0.025 and 0.975 percentiles of these 2,000 predicted bootstrap probabilities.

Summary statistics for the variables in the study are presented in Tables 2–4, separately for the two cohorts of children with uncooperative and cooperative behaviors at the initial visit (45 and 40 children, respectively). For each variable, the actual sample size (n) is reported in parentheses.

The univariate distribution of the variables related to demographic attributes (age at the first visit, gender, nationality), medical history (type of delivery, vaccines, medications, previous experiences of general anesthesia), and clinical data (dental health status, gingival status, observational plaque index) is presented in Table 2. The first two groups of variables display similar distribution patterns in the two cohorts of children with uncooperative and cooperative behaviors at the initial visit. In both cohorts, approximately three-quarters of the sample are boys and the age of the children ranges between 7 and 12 years, with a mean of approximately 9.2 and a standard deviation of 1.5. The vast majority of the children (more than 88.6% in both cohorts,) are Italian and received vaccinations on schedule according to the Italian National Health Service Plan. Fewer than half of the children were born by natural childbirth. With regard to medication use, based on parental/legal guardian declarations collected through questionnaires, approximately 20% of the children in both cohorts regularly or temporarily used medications, either for anxiety and tension control or for comorbidity with pathologies of a different origin (cardiac, neurological, or psychiatric). Previous experience with anesthesia was reported in 15.4% of children with cooperative behavior at the first visit and 11.1% of children with initially uncooperative behavior. In the latter group, the large majority (91.1%) presented with bleeding and/or gingival tartar, and approximately seven out of 10 (71.1%) had visible plaque at the gingival margin or abundant accumulation of calculus beyond the margin. These percentages dropped to 80.0 and 57.5, respectively, in the cohort of children with cooperative behavior at the first visit. The average DMFT/dmft index was 3.1 in both cohorts, with a standard deviation of approximately 2.9.

As a preliminary descriptive evidence of the efficacy of the ESBA, Table 3 reports, for each level (treatment) of the variable Type of treatment, the percentage of patients in the two cohorts who had received the treatment during phase I, by the end of phase II, and by the end of phase III. In phase I, educational intervention was carried out in the vast majority (95.6%) of the 45 initially uncooperative children at the first visit and in the majority (77.5%) of the 40 initially cooperative children at the first visit. These percentages had increased to 100 and 82.5, respectively, by the end of phase III. Regular checkups were performed in approximately 60% of the initially uncooperative children (in 62.2% of them in phase I and in 64.4% and 66.7% by the end of phase II and III, respectively), whereas they were largely postponed until phase III for the initially cooperative children (10% of these children received a checkup in phase III and none of them received a checkup in the first two phases of the ESBA). A substantial percentage of the initially cooperative children received preventive procedures (such as routine dental cleaning and fluoride treatments) (47.5) and more demanding/invasive treatments such as conservative procedures (57.5), minor surgery (17.5), and dental cleaning (57.5). This, however, was not the case for the cohort of initially uncooperative children; in phase I, only 33.3% of this cohort received dental cleaning and less than 2% received preventive or conservative procedures and minor surgeries. As children progressed through phases II and III of the ESBA program, some of these differences between the two cohorts narrowed, while others largely persisted. By the end of phase III, 66.7% of the initially uncooperative children at the first visit had received dental cleaning and approximately the same percentage (64.4) preventive procedures (remarkable increases compared with phase I, where these percentages were 33.3 and 0, respectively). The corresponding percentages for the initially cooperative children—who by the end of phase III had received dental cleaning and preventive procedures—were not too different (87.5 and 77.5, respectively).

The percentages of initially uncooperative children who received conservative and surgical procedures also increased significantly during the ESBA program (from 2.2 in phase I to 15.6 by the end of phase III for conservative procedures and from 0 in phase I to 11.1 in phase III for surgical procedures). Despite this favorable evolution, these percentages remained much lower (approximately one-fifth and half) than the corresponding percentages for the cohort of children who were initially cooperative, among whom 72.5% of the children had received conservative procedures and 22.5% surgical procedures by the end of phase III.

In both cohorts, when educational intervention did not translate into collaboration and treatment deferral was not an option (due to the urgency of the case), the patients were treated under general anesthesia. Five (11.1%) of the initially uncooperative children were treated under general anesthesia in phase I and 12 (26.7%) by the end of phase III. These percentages were much lower in the cohort of initially cooperative children. None of the children in this group (0%) was treated under anesthesia in phase I and only four (10%) were treated by the end of phase III. Notably, these four children, who although cooperated at the first visit, demonstrated a decline in cooperation in phase II of the ESBA.

Additional descriptive evidence of the ESBA efficacy is provided in Table 4, which shows the distribution of the Frankl scale at the first visit and at the end of phases I, II, and III for the two cohorts of children with initial uncooperative and cooperative behaviors. Boldface is used to denote cases in which the patients improved compared with the first visit. As we explained earlier, according to our definition, an initially uncooperative child (with a Frankl scale rating of negative or definitely negative at first visit) was considered improved if their rating at the end of the phase became positive or definitely positive, while an initially cooperative child (with a Frankl scale rating of positive at first visit) was considered improved if their rating at the end of the phase became definitely positive.

As shown in the table, at the first visit, six of the 45 initially uncooperative children (13.3%) were rated in the Frankl scale as definitely negative, while the remaining 39 (86.7%) were rated as negative. However, as they completed the different phases of the ESBA program, most of these children demonstrated cooperation (improved according to our definition). By the end of phase II, 60% had displayed cooperative behaviors, i.e., their Frankl scale rating was either positive (57.8%) or definitely positive (2.2%). This percentage had increased to 84.4 by the end of phase III (in which 80% of the initially uncooperative children were rated as positive and 4% as definitely positive). None of the initially uncooperative children regressed in their behavior during the ESBA program.

In order to test whether the outlined differences between the distribution of Frankl scale ratings at the first visit and the corresponding distributions at the end of each ESBA phase were statistically significant, we used the Stuart–Maxwell test. The p-values for this test (one for each phase) were all lower than 0.001, indicating that for the cohort of 45 initially uncooperative children, these differences were statistically significant at the α = 0.05 level.

If a numerical score from 1 to 4 is assigned to the four levels of the Frankl scale (1 corresponding to definitely negative and 4 to definitely positive), the sample average rating for these 45 children at the first visit will increase from 1.9 to 2.2, 2.6, and 2.9 by the end of phases I, II, and III, respectively, indicating an upward trend from negative to positive behavior.

On the other hand, as shown in Table 4, the 40 initially cooperative children (all rated as positive at the first visit) displayed very limited change in the Frankl scale ratings from the first visit through the phases of the ESBA program. The percentage of these children who maintained the same rating as at the first visit was 97.5 at the end of phase I, 87.5 at the end of phase II, and 90 at the end of phase III. Only one child (2.5% of the total) had improved (i.e., their rating had changed to definitely positive) by the end of phase II, and this percentage had increased only to 7.5 (i.e., three children) by the end of phase III. In addition, four children (10%) with initially cooperative behaviors regressed in phase II (they changed their behavior to uncooperative and were rated as negative in the Frankl scale at the end of phase II). While three of these children returned to their original rating (positive) during the last phase of the ESBA program, one maintained the negative rating also in phase III. If, as before, a numerical score from 1 to 4 is assigned to the four levels of the Frankl scale (1 corresponding to definitely negative and 4 to definitely positive), the sample average rating for the 40 initially cooperative children at the first visit would be 3 and would stay approximately constant through the phases of the ESBA (3 at the end of phase I and III and 2.9 at the end of phase II). Not surprisingly, for this cohort of children, according to the Stuart–Maxwell test, the differences between the distribution of the Frankl scale rating at the first visit and the corresponding distribution at the end of each phase were not statistically significant at the α = 0.05 level (the p-values for this test, one for each phase, were all greater than 0.17).

A total of 3,328 proportional odds cumulative logit models that relate the response variable (Frankl scale assessment) to a set of ten explanatory variables were fitted to the data. For baseline behavior, the binary variable Init.Rating (which takes the value 0 if the Frankl scale rating at the first visit is negative or definitely negative and 1 if the Frankl scale rating at the first visit is positive) was used to distinguish between the two cohorts of children with uncooperative and cooperative behaviors at the first visit. The other explanatory variables were Phase (indicating the phase at which the value of the response variable was retrospectively assessed), Age, Gender, medical history variables (Type of delivery, Medications, and General anesthesia), and clinical variables (Dental health status, Gingival status, and Observational plaque index). The two variables Nationality and Vaccines were excluded from the modeling because of the large number of missing values and their low sample variability (as shown in Table 2, regardless of the cohort of children considered, more than 88.6% of the patients shared the same value for each variable).

Model selection was performed using the BIC criterion and the 1se rule described previously. For the model selection, all 3,328 models were fitted using the same sample of 76 children, which corresponded to the largest sample for which complete information about the 10 regressors included in the modeling was available. The selected model (and the models of interest) was then reestimated using the largest sample for which complete information about the regressors was available. Parameter estimates, confidence intervals, and p-values for the coefficients of the full model, the minimum BIC model, and the best BIC-1se model are presented in Table 5. As shown in the table, four predictors (Phase, Init.Rating, Medications, and Dental health status) were statistically significant (at the level α = 0.05) in the full model. These four predictors also defined the model with minimum BIC.

The best BIC-1se model contained only the predictors Phase and Init.Rating, which emerged as the two most important explanatory variables in our statistical analysis. As shown in Table 5, both predictors were statistically significant and had positive estimated coefficients in the full model, the selected minimum BIC model, and the best BIC-1se model. Thus, this indicated that adding any of these two covariates to the model, improved the predictions of children’s behavior and cooperation as measured by the Frankl scale. This is confirmed by the corresponding 95% confidence intervals in the table that do not include the zero.

The positive estimated coefficient for Init.Rating indicated that children who were already cooperative at the beginning of the study (Init.Rating = 1) tended to have a more favorable Frankl scale assessment at the end of each of the three phases of the proposed protocol than those who were initially uncooperative (Init.Rating = 0). Similarly, the positive estimated coefficient for Phase indicated that children’s behavior and cooperation, as measured by the Frankl scale, tended to improve as they progressed from one phase to the next in the proposed approach (for a more technical interpretation of the coefficients of Phase and Init.Rating within the cumulative logit models in Table 5, see Appendix A).

The importance of the Phase and Init.Rating variables is illustrated in the side-by-side boxplot of BIC scores in Figure 1. The figure considers a partition of the 3,328 cumulative logits models in five subgroups. For each subgroup, the number of models is reported in parenthesis below the subgroup label. The different subgroups correspond to (i) models that do not contain any of the two candidate predictors (Phase and Init.Rating); (ii) models that contain Phase but not Init.Rating; (iii) models that contain Init.Rating but not Phase; (iv) models that contain both predictors (Phase and Init.Rating) but no interaction term; and (v) models with both Phase and Init.Rating and the interaction of Phase with another predictor.¹

As shown in Figure 1, the corresponding BIC values cluster into four groups with increasing BIC range. The ordering of the four groups in terms of BIC makes apparent the relevance of Phase and Init.Rating as predictors of children’s behavior and cooperation as measured by the Frankl scale.

The best group (with BIC values in the range 300–369), represented in gray in Figure 1, comprised the models in subgroups (iv) and (v), which included both Phase and Init.Rating. The second-best group of models (with BIC values in the range 322–373) contained the Init.Rating predictor but not Phase. The third-best group (with BIC values in the range 347–408) contained Phase but not Init.Rating. The worst group of models (with BIC values in the range 362–407) consisted of models that did not contain either of the two regressors (neither Init.Rating nor Phase).

In line with this analysis, the best BIC-1se model contained only the covariates Phase and Init.Rating, both with positive and statistically significant coefficients. The estimated coefficients of Phase and Init.rating for the minimum BIC model were also positive and statistically significant. However, the presence of the interaction between Phase and Init.rating in the minimum BIC model, in practice, implied a different estimated coefficient for Phase, depending on the Frankl scale assessment at the first visit. Children with problematic behavior and limited cooperation at the beginning of the study (Init.Rating = 0) had an estimated coefficient of 1.43, which was more than 10 times larger than the corresponding coefficient for initially cooperative children at the beginning of the study (Init.Rating = 1), which was 0.11(1.43–1.32 = 0.11). Thus, in agreement with our descriptive findings, the minimum BIC model predicted that children with problematic behavior and limited cooperation at the beginning of the study tended to improve more across the different phases of the proposed approach compared with initially cooperative children. Despite the lack of an interaction term, the best BIC-1se model also captured this important feature of behavioral evolution during the ESBA program. As shown in Table 6, the best BIC-1se model predicted that an autistic patient uncooperative at the beginning of the study (Init.Rating = 0) with a probability of 0.40 will be cooperative by the end of phase I (Frankl scale at the end of phase I, positive/definitely positive). The probability of being cooperative had increased to 0.62 and 0.80 by phase II and phase III, respectively. In contrast, for an autistic patient already cooperative at the beginning of the study (Init.Rating = 1), these predicted probabilities of improvement were much lower and increased at a lower rate as a function of Phase (since there is little room for improvement). In particular, the best BIC-1se model predicted that an autistic patient cooperative at the beginning of the study with a probability of 0.02 (very small) would have improved their behavior and cooperation by the end of phase I (Frankl scale at the end of phase I, definitely positive). This probability increased only to 0.05 and 0.11 as the patients completed phase II and phase III, respectively.

The 95% confidence intervals in Table 6 provide informative bounds for the estimated probabilities of improvement. For example, with a 95% of confidence, it can be claimed that an autistic patient uncooperative at the beginning of the study with a probability of at least 0.28 will be cooperative by the end of phase I, and this probability will increase to 0.53 and 0.71 by the end of phases II and III, respectively. Similarly, with a 95% of confidence, it can be claimed that for an autistic patient cooperative at the beginning of the study, the probability of improvement by the end of phase III will be lower than 0.20.

The predicted probabilities in Table 6 can be transformed into predicted percentages multiplying each predicted probability by 100. These predicted percentages can be compared with the observed percentages obtained from Table 4 and can also be used as an indication of the fit of the best BIC-1se model to the data. This comparison is shown in Figure 2, which demonstrates that the best BIC-1se model quite effectively predicts the percentage of children who improved their behavior and cooperation across the different phases of the ESBA. The 95% bootstrap-based confidence intervals in each case contained the true proportion of children who improved their behavior. The good fit of the best BIC-1se model was confirmed by using the Pearson's chi-square test with an associated p-value of 0.148.

In order to test the proportional odds assumption underlying the best BIC-1se cumulative logit model, a likelihood ratio test (LRT) that examines the proportional odds model vs. the same model without proportional odds (i.e., vs. the model having a separate parameter for each logit) was conducted. The LRT test statistic was 7.64 with 4 degrees of freedom and a p-value of 0.11. Thus, there is no significant evidence to reject the proportional odds assumption at the α = 0.05 level.