For the search, two researchers searched a total of seven electronic databases or electronic journals (PubMed, Cochrane, EMBASE-OVID, CINAHL, World of Science, SCOPUS and RISS) for papers published in English and Korean until December 31, 2023. The search was conducted from January 18, 2024 to February 22, 2024, and the search strategy based on PICO-SD (Population, Intervention, Comparison, Outcome, Study Design) was shown in Table 1.

The results of this study were reported with reference to the PRISMA 2020 checklist (accessed at https://www.prisma-statement.org/ on April 1, 2024). The selection criteria were as follows. The study population (P) was diabetes aged 19 years and older, the intervention (I) was aromatherapy, the control (C) was conventional care or no intervention, and the outcome (O) was pain. When two or more interventions were used in an article, each intervention was separated into its own study and the effect size was calculated. Only studies that presented the number of subjects, mean, and standard deviation were selected for accurate effect size calculation. The study design (SD) included randomized controlled trials (RCTs) and a quasi-experimental study. The exclusions were as follows: studies that the subjects were not adult diabetics; studies that the intervention was not aromatherapy; studies that pain was not reported as an outcome variable; studies that the mean, standard deviation, or number of subjects was not reported; studies not reported in English or Korean; studies that the full text could not be found; and single group studies.

Two researchers independently searched the data according to the inclusion and exclusion criteria to select studies for analysis. The selected studies were extracted by author, publication year, country, Institutional Review Board (IRB), fund, number of subjects, study design, intervention characteristics (oil type, application method, application area, control group intervention method, intervention duration, intervention session, intervention time/session, number of measurement time), and outcome variables and then recorded in a coding book created in Microsoft Excel Spreadsheet Software. In case of coding discrepancies, the original articles were reviewed again and final coding values were determined (Table 2).

The quality assessment of the selected literatures was independently performed by two researchers using the Joanna Briggs Institute (JBI) checklist. The five articles included in the analysis were all RCTs, with a mean value of 9.40±2.30 using the 13-item JBI Checklist for RCTs tool [10]. Only one article clearly reported " Q7. Were outcome assessors blind to treatment assignment?", and only two articles clearly reported " Q2. Was allocation to treatment groups concealed?", "Q4. Were participants blind to treatment assignment?", and "Q5. Were those delivering the treatment blind to treatment assignment?"

MIX 2.0 Pro (Ver. 2.0.1.6, BiostatXL, 2017) was used to calculate and merge effect sizes for pain, the primary outcome of the studies. For the pooled overall effect of pain, Hedge's g was used as the effect size due to the small number of studies, and a synthesis forest plot was used using a random effects model with reweighting to account for variations in subject characteristics and study-specific heterogeneity. The significance of the effect was determined using 95% confidence interval (CI), p value less than .05, and the weight of each effect size was derived using the inverse of variance [11]. The heterogeneity of the included studies was assessed using Q statistics and Higgin's I2 values [12], and 95% CIs were reported to account for the bias of the heterogeneity I2, which is a point estimate in small meta-analysis studies; an I2 greater than 50% was interpreted as heterogeneity [13]. Meta-regression and exclusion sensitivity analysis were performed to identify factors that determine the heterogeneity of studies on aromatherapy for pain in diabetes. Publication bias was also tested using funnel plot, trim and fill plot, Begg's test, and trim and fill method, correcting for pooled overall effect [14].

A total of 1,788 articles were retrieved from seven databases according to the search strategy, and 1,061 articles were extracted after excluding duplicates. Based on the inclusion and exclusion criteria, we selected the final five articles (Figure 1). As shown in Table 2, two studies were published after 2021, three were conducted in West Asia, four were reviewed on IRB and one was funded. The study designs were all RCTs, and one studies had fewer than 60 subjects. The characteristics of the interventions were: oil type was lavender oil in two studies and blended oil in two studies; aromatherapy was applied to the feet in two studies, hands and feet in two studies, and other oils were applied in the control group in two studies; intervention duration was 28 days (4 weeks) or more in three studies; intervention sessions were 12 or more in three studies; and intervention time per session was less than 30 minutes in three studies. Three studies measured the dependent variable pre- and post-intervention, two studies measured it mid-intervention, and four studies had a quality score of 9 or higher (Table 2, Appendix 1).

The pooled overall effect of pain in diabetes on aromatherapy was Hedge's g = −1.83 (95% CI: −2.76 to −0.91), which was found to be a large effect based on the criteria for interpreting effect sizes outlined by Brydges [15] (Figure 2). The studies were highly heterogeneous, with a Q = 166.31 (Q-df = 158.31, p < .001) and a Higgins' I2 of 96.4% in the heterogeneity test. Univariate meta-regression was performed to identify potential influences on the effectiveness of aromatherapy on pain in diabetes and found that West Asian (Z = −6.46, p < .001), being reviewed on IRB (Z = −4.92, p < .001), and having a fund (Z = −8.75, p < .001) were associated with a reduction in pain. In terms of study intervention characteristics, pain reduction was more likely to occur when the participants were younger than 60 (Z = 5.00, p < .001), when the type of aromatic oil used was lavender oil (Z = −7.16, p < .001), and when the intervention was massage therapy rather than topical application (Z = −5.03, p < .001). Fewer hours per session (Z = 4.02, p < .001) and more repeated measurements (Z = −5.03, p < .001) than two pre- and post-test measurements (Z = −5.03, p < .001) were associated with pain reduction. However, the duration of the intervention (Z = −1.88, p = .060), number of sessions (Z = 1.11, p = .268), and quality of life scores (Z = −1.54, p = .123) did not have a significant effect on pain reduction (Table 3). When one study was excluded using an exclusion sensitivity test [16], Hedge's g was −0.99 to −2.36, a large effect size, and the 95% CI (−1.71 to −3.58, −0.27 to −1.17) did not include zero, all of which were statistically significant.

To check for publication bias in the study, we performed funnel plot and trim and fill plot analyses and found that the individual effect sizes of the seven studies included in the study (blue circles) were skewed slightly to the left, indicating some publication bias (Figure 3A), and the trim and fill plot showed that one additional study (white circle) should be added (Figure 3B). For further analysis of publication bias, Begg's test was performed, and Tau b = −0.57, ties = 0.00 (Z = −1.80, p = .072), confirming that there was no publication bias. Furthermore, the number of articles that should have been added to the study using the trim and fill method [17] was found to be 1, and the effect size of the 8 corrected studies was −0.49 (95%CI: −0.65, −0.34). The effect size of pain reduction after correction was somewhat smaller than before correction, but it was still statistically significant after correction. In conclusion, the study was free of publication bias (Table 4).