Database searches were conducted according to the PRISMA guidelines. Eligible studies were selected from several databases such as MEDLINE using several search managers, such as PubMed, Scopus, ScienceDirect, and ProQuest. The search term used was formed using the PICO framework (P: patient or problems; I: intervention being considered; C: comparison intervention; O: outcome measurements). The focus of our search was studies implementing carbohydrate counting as an intervention in pediatric (<18 years old) type 1 diabetes mellitus patients (<18 years old). The PICO criteria can be found in Supplementary Material 1.

The available PICO criteria were then used for search queries and keywords, which can be seen in Supplementary Material 2. The exploded keywords were included, as were MESH terms for MEDLINE and modified truncation according to the different search platforms.

The literature search was followed by a screening process based on the predetermined inclusion and exclusion criteria. Original studies on carbohydrate counting in T1DM pediatric patients that were published within the last 10 years were included. This includes randomized or nonrandomized controlled trials, and cohort or cross-sectional studies. Studies were excluded if they were not published in English, had a nonretrievable full text, or were commentaries, reviews, or letters/correspondences.

A quality assessment of the included studies was undertaken based on the study designs. We used the Cochrane risk of bias (RoB) 2.0 for trial studies and the JBI Critical Appraisal Checklist for observational and cross-sectional studies [9]. In RoB 2.0, the quality of included studies was assessed according to the following 5 appraisal elements: randomization process, deviations from the intended interventions, missing outcome data, measurements of the outcome, and selection of the reported results. Studies were then scored as low risk, some concerns, and high risk. Meanwhile, in the JBI Critical Appraisal Checklist, the quality assessment items are population similarity, the validity of exposure, confounding factors, measurements of the outcome, follow-up, and statistical analysis [10]. The studies then are appraised as included, excluded, and requiring further information.

All reviewers independently screened all of the titles from the search and excluded studies that were irrelevant. Following this, all the reviewers independently screened the titles and abstracts using an eligibility checklist. Potentially eligible texts were retrieved for final selection. We included studies that met our inclusion criteria. Any discrepancies in the extracted data were discussed by all reviewers.

The following data were extracted: (1) general information (author, title, year of publication); (2) study characteristics (study design, number of samples); (3) intervention and setting (methods of carbohydrate counting, follow-up period, study location); (4) outcome data (baseline and follow-up measure). The following outcome data of the glycemic index control were collected: HbA1c changes, nutritional outcomes (low-density lipoprotein [LDL] and body mass index [BMI] changes), and QoL analysis.

To summarize the effect of carbohydrate counting on type 1 diabetes, we measured the outcome by HbA1c concentration, nutritional outcomes (changes in BMI and LDL), and QoL.

The software Revman 5.4.1 was used to analyze the data. The Generic Inverse Variance with Random Effects (DerSimonian-Laird) method was used for the analysis using the mean difference as the effect measure, with 95% confidence intervals (ICs). The analysis was then extrapolated into a forest plot.

The flowchart for the study selection can be seen in Fig. 1. The literature search yielded 1,693 articles on the initial hit. Four reviewers independently screened titles and abstracts. Of the 36 articles screened, 34 articles were found to be eligible. Of those, 25 studies were excluded because their outcomes and study designs were not suitable. After the screening process, we included 9 studies based on the eligibility criteria.

The included studies (Table 1), were comprised of 4 randomized controlled trials, one prospective controlled trial, one prepost trial, and 3 observational studies (prospective crosssectional and retrospective longitudinal study), with a total of 606 participants [11-19]. All of the studies were conducted at various locations within Europe and the United States. The follow-up or trial periods for each study are different, ranging from 3 months to 2 years. Most studies used basic carbohydrate counting, which involves basic calculations on the consumed carbohydrate daily. In contrast, one study used advanced carbohydrate counting, while another study used a carbohydrate counting application, which calculates the carbohydrate consumption automatically. Advanced carbohydrate counting not only adjusts the amount and time of carbohydrate consumption (like basic carbohydrate counting), but also regulates the number of grams of carbohydrates adjusted to the insulin dose based on the ICR. Out of 606 participants, there were 440 participants in the carbohydrate counting group and 166 in the control group. Three studies did not use a control group due to their observational method, while one study used its whole population on a pre-post trial. Therefore, the first period of the trial was used as the control outcome, while the second period of trial was used as the interventional outcome. Most studies classified the sample into 2 different groups (carbohydrate counting and control). However, Enander et al. [17] used 3 groups, with the interventional group divided into manual carbohydrate counting and calculated carbohydrate counting (using software or applications).

The RoB assessment was conducted with Cochrane RoB 2.0 for randomized studies and JBI Critical Appraisal Tool for nonrandomized studies [9,10]. The results can be seen in both Supplementary Materials 3 and 4.

With regard to the RoB for randomized studies, 4 of 5 studies that were assessed were considered to have low RoB. However, the study by Marigliano et al. [18] was assessed to have some concern on the overall risk assessment. In the study by Marigliano et al. [18], the randomization process was deemed unclear due to a lack of relevant information. Their methods only mentioned that there was a random assignment using 2-group randomized block design on both groups. Furthermore, the outcome measurement domain was deemed concerning due to the unblinded nature of the data analysis. In the study by Gökşen et al. [11], the outcome measurement and data selection domains were deemed concerning due to multiple eligible outcome measurement methods. However, all studies were deemed eligible to be included in this review.

The RoB assessment for the nonrandomized studies was conducted using the JBI Critical Appraisal Tool with an assessment of the research population, recruitment process, exposure measurement, outcome measurement, follow-up period, and appropriate statistical analysis. In the 4 studies assessed, all aspects were deemed to be clear. Therefore, the 4 studies were eligible for inclusion.

The data synthesis of the included studies is summarized in Table 2. The recorded outcomes include the HbA1c changes and nutritional outcomes (LDL and BMI). The effect was measured using the mean difference. The authors also recorded the QoL assessment, if available. Notable outcomes of each study were also recorded if the authors deemed it necessary to be included. All the studies comparing the carbohydrate counting group and a control group showed a greater reduction in the HbA1c in the carbohydrate counting group than in the control group. Most studies reported an increase in HbA1c at the end of the followup period in the control group. Enander et al. [17] also reported the difference between the manual and calculated carbohydrate counting methods; they found that manual carbohydrate counting led to less increase in the HbA1c than did calculated carbohydrate counting. (0.1±0.39 vs. 0.4±0.34). Meanwhile, the studies that did not have a control group reported a 0.5% to 1% reduction in the HbA1c at the end of the follow-up period.

The quantitative analysis of the glycemic index control included 6 studies, comparing the HbA1c changes between the carbohydrate counting group and its control counterpart [11-14,16,17]. The meta-analysis showed the cumulative effect on the mean difference of -0.55 (95% CI, -0.81 to -0.28; P-value <0.001). While all studies similarly found that the mean difference reduction favored the interventional group, the heterogeneity analysis showed an I2 value of 88%, which implies high heterogeneity between the studies (Fig. 2).

The nutritional outcomes extracted from the included studies were the LDL cholesterol changes and BMI changes in both groups. Only 5 studies reported the changes in LDL values [11,13,14,18,19], while only 4 studies reported BMI changes [11,16-18]. Gökşen et al. reported a lower increasing level of LDL concentration in the intervention group as compared to that in the control group (mean differences [MD], 15.26±5.81 vs. 18.22±6.17) after 2 years of follow-up. Bayram et al. reported a higher reduction of LDL in the carbohydrate counting group than in the control group (MD, -3.68±6.49 vs. -2.13±5.98) [11,14]. However, Dalsgaard et al. [13] reported the opposite result with a higher reduction of LDL in the control group than in the intervention group (MD, -1.39±5.31 vs. -7.20±5.29) [13]. The quantitative analysis of LDL changes reported no significant difference between the interventional and control groups with a cumulative mean difference of 0.53 (95% CI, -5.62 to -0.30), and high heterogeneity of 95% (Fig. 3). This outcome was most certainly caused by the different results in Dalsgaard et al.’s study [13].

The BMI change outcomes were quite similar across the studies that reported it, with a lower increase or decrease in BMI in the control group than in the intervention group. However, Gökşen et al. [11] reported a higher BMI increase in the control group in the first year of the study, while the opposite happened in the second year of observation.

Donzeau et al. [16] was the only group to assess QoL. Donzeau et al. [16] reported higher scores in the carbohydrate counting group than in the control group based on the DISABKIDS, KIDSCREEN, and DSQOL scoring. The scoring assesses the children's physical fitness, ability to study and make social connections, eagerness to play, and high aspiration toward the future. Carbohydrate counting is beneficial in the maintenance of QoL in diabetic children.