Efficacy of therapeutic suggestions under general anesthesia: a systematic review and meta-analysis of randomized controlled trials

Eligible studies were randomized controlled trials that investigated therapeutic suggestions presented during general anesthesia to adult patients undergoing surgery or medical procedures. If the intervention group received a combination of therapeutic suggestions and another psychological intervention or if therapeutic suggestions were not solely implemented intraoperatively, the study was excluded. Eligible control groups were “treatment as usual” (defined as the standard surgical care policy of the hospital) and “attention control” groups (defined as providing same amount of time and attention in addition to standard surgical care; e.g., blank tape, white noise). The included trials reported on at least one of the following outcomes measured via self- and/or observer reports: pain intensity, mental distress, recovery, use of medication, measured postoperatively within hospitalization. In addition, intraoperative outcomes, i.e., length of procedure and physiological parameters, were included (Additional file 1: Table S1).

Deviating from the protocol [21], we did not limit study inclusion to trials with a sample size of at least 20 participants in each trial arm, but rather tested this restriction in sensitivity analyses.

Electronic searches were carried out in the following databases (last search February 23, 2015): MEDLINE, CENTRAL, Web of Science, PsycINFO according to a search strategy that specified terms referring to the patient population (e.g., surg$.ti.ab.kw, General Surgery/, Anesthesia. General/), treatment (e.g., suggestion$.ti.ab.kw, Suggestion/), and study design (e.g., randomized controlled trial.pt). The search strategy was developed with consideration of validated search strategies for retrieving randomized controlled trials [22]. The MEDLINE search strategy is shown in Appendix. We adapted the strategy for the Cochrane Central Register of Controlled Trials (CENTRAL), Web of Science and PsycINFO.

In order to identify further trials, lists of references of relevant articles and previous reviews were also checked. Additionally, we screened ProQuest Dissertations and Theses Full Text Database to identify any unpublished material. One author (DJ) screened titles and abstracts of database records and retrieved full texts for eligibility assessment.

A pilot-tested data extraction form was used to collect the following information from eligible trials: characteristics of patients, intervention, control group, outcomes, bibliographic information, and effect size related data.

Data were independently extracted by two raters (DJ, JR). Inter-rater disagreement was resolved through consensus. In case of missing information, study authors were contacted. If information on effect sizes was missing and could not be retrieved, data had to be approximated using different estimation methods (e.g., estimating statistics from graphs without numerical data, setting an effect size to zero if non-significant results were mentioned without reporting statistical parameters).

To assess risk of bias in the included studies, common markers of internal validity from the Cochrane Risk of Bias Tool were extracted [23]. The risk of bias assessment was conducted by two independent raters (DJ, SK) who were previously trained and blinded to extracted effect size estimates. Disagreements were resolved by discussion with one author (JR). Inter-rater agreement for the risk of bias assessment using Cohen’s kappa (κ) was excellent, κ = 0.76 [24].

Corrected standardized mean differences (Hedges’ g) were calculated for each assessment time-point and measurement multiplied by a small sample bias correction factor [25]. An effect size of 0.5 thus indicates that the mean of the experimental group is half a standard deviation larger than the mean of the control group. The magnitude of Hedges’ g was interpreted within the same ranges as Cohen’s d, regarding 0.20, 0.50, and 0.80 as small, medium, and large effect sizes, respectively [26]. Since such effect sizes are generally not easy to interpret in terms of clinical significance, effect sizes Hedges’ g were transformed into numbers needed to treat (NNT) [27]. For all dichotomous outcomes, Log Odds Ratios were computed and converted to Hedges’ g [28] in order to pool across different effect size formats.

If a study comprised more than one intervention group [29‐31] the shared control group was divided out approximately evenly among the comparisons [32].

Outcome data were meta-analyzed using a random-effects approach. The generic inverse variance method was applied with heterogeneity estimated using the DerSimonian-Laird method [33]. Statistical heterogeneity between trials was assessed with χ² heterogeneity tests (Cochran’s Q) and I² statistic [34]. I² describes the percentage of the variability in effect estimates that is due to heterogeneity rather than chance, with values from 0 to 40% indicating no important heterogeneity, 30 to 60% moderate, 50 to 90% substantial, and 75 to 100% considerable heterogeneity, respectively [35].

In order to test for publication bias funnel plots were inspected visually and the Egger test was run [36]. Additionally, Duval & Tweedie’s trim and fill procedure was used to obtain an adjusted estimate of the treatment effect after the publication bias had been taken into account and to indicate how many missing trials have been imputed to correct for publication bias [37].

We conducted sensitivity analyses in order to test the robustness of findings, examining if meta-analytic results change when excluding approximated effect sizes and when excluding small samples (n ≤ 20 per group). Moderator analyses were planned to explain statistical heterogeneity [38]. However, heterogeneity was not important (I² < 40%). Therefore, we conducted stratified analyses in order to exploratory examine potential moderators.

All data analyses were performed using Comprehensive Meta-Analysis (CMA; Version 2.0; Biostat Inc.).

A total of 7427 records was screened and N = 32 randomized controlled trials were included in the meta-analysis. Figure 1 contains a flow chart of the study selection process.

Table 1 presents selected study characteristics. The majority of primary studies were published between 1986 and 2001; only one study [39] was published much earlier. Among the primary studies, there were three unpublished dissertations. One study was reported in German [40], all others were written in English. Altogether, n = 32 randomized controlled trials provided k = 37 comparisons between an intervention and a control group, incorporating a total of n = 1111 patients in intervention groups (M = 30.0, SD = 18.2) and n = 991 patients in control groups (M = 31.0, SD = 18.1). The mean age of patients in the intervention groups was 47.7 years (SD = 8.2), similarly in the control groups 47.2 years (SD = 9.4). The mean percentage of male patients was 17% (SD = 28.1) in intervention groups, and 17% (SD = 29.3) in control groups as well. This low percentage of male patients can be ascribed to a high proportion of studies including patients undergoing gynecological surgery; 16 primary studies investigated female patients only. In the majority of primary studies anesthesia was performed as “balanced anesthesia” with an opioid and an inhalational anesthetic (Table 1). In six studies, neuroleptanesthesia was used and in two studies total intravenous anesthesia (TIVA) with propofol or midazolam, respectively. Nitrous oxide was included in all except one study. In seven studies a benzodiazepine was applied for premedication. Therapeutic suggestions were presented via tape in all studies, played throughout the surgery in almost every study. Suggestion were judged as affirmative (e.g., “You will feel fine after the operation.”) in 12 intervention groups (32%), as non-affirmative (e.g., “After the operation you will not feel any nausea.”) in one (3%), and both affirmative and non-affirmative in 14 intervention groups (38%; no information reported for 10 intervention groups). In 19 intervention groups (51%), suggestion were accompanied by or alternated with soothing music or sounds. In all studies, the effects of therapeutic suggestions were compared against attention control. 18 studies (56%) used blank tapes/white noise, 7 studies (22%) offered sounds or music, and another 7 studies used spoken text (history of hospital, story of Peter Pan, parts of a cookery book) as control condition.

Additional file 2: Table S2 contains information on the risk of bias in included studies. Overall, the risk of bias in the included studies was mainly judged as low; no study indicated a high risk of bias in any quality item. However, due to missing information in the studies a high percentage of items was judged as unclear.

Across all included postoperative outcomes, there was a small, but statistically significant and homogeneous effect of therapeutic suggestions compared to attention control (g = 0.13, 95% CI [0.04; 0.23], k = 37, p = .005; I² = 0%).

When outcomes were analyzed separately, we found effects of therapeutic suggestions on pain intensity (g = 0.04, CI 95% [−0.04; 0.12], NNT = 44.3) and mental distress (g = 0.03, CI 95% [−0.11; 0.16], NNT = 68.2) to be close to zero and non-significant. However, small significant effects in favor of therapeutic suggestions appeared on medication use (g = 0.19, CI 95% [0.09; 0.29], NNT = 9.2) and on recovery (g = 0.14, CI 95% [0.03; 0.25], NNT = 13.0). Stratifying analyses on medication use and recovery with respect to outcomes, we found small, significant effects for therapeutic suggestions on PONV (g = 0.21, CI 95% [0.07; 0.36], NNT = 8.3) and analgesic use (g = 0.16, CI 95% [0.06; 0.26], NNT = 11.0). Therapeutic suggestions also revealed a small effect on antiemetic use (g = 0.22, CI 95% [−0.003; 0.45], NNT = 7.9) and on all other recovery outcomes (g = 0.11, CI 95% [−0.01; 0.24], NNT = 15.6), even though these effects were marginally significant only (Figs. 2, 3, and 4). Heterogeneity for all outcomes was not important (I² < 40%).

Regarding intraoperative outcomes, therapeutic suggestions revealed a small effect on physiological parameters, even though this effect was not significant (g = 0.13, CI 95% [−0.16; 0.42], k = 12, p = .389; I² = 62%). Effects of therapeutic suggestions on length of surgical procedure (g = −0.04, CI 95% [−0.14; 0.07], k = 28, p = .499; I² = 0%) were close to zero and non-significant.

A visual inspection of the funnel plot (see Additional file 3: Figure S1) gave no indication of publication bias as trials are distributed symmetrically around the pooled effect size. Egger’s test of funnel plot asymmetry did not indicate publication bias (t(35) = 0.18; p = .428), and Duval & Tweedie’s trim and fill procedure resulted in no trimmed studies. Hence, publication bias does not pose a threat to the accuracy of our meta-analytic results.

We tested the robustness of effects for primary outcomes. After excluding approximated effect sizes for all outcome categories the meta-analytic result patterns (size of effect estimates and significance) did not change considerably though effect sizes were slightly larger and reached significance for recovery. Furthermore, effects were robust against the exclusion of small samples (n ≤ 20 per group) yielding effect sizes comparable in size and (non-)significance (Additional file 4: Table S3).

Since heterogeneity was not important at all (I² < 40%), we did not run our pre-specified subgroup analyses. However, in order to get some ideas about potential moderators we exploratory conducted stratified analyses for PONV and antiemetic use since for all other postoperative outcomes results were homogeneous (I² = 0%). Studies applying suggestions related to the absence of PONV (e.g., “no sickness”) yielded larger effects than studies without such suggestions, but this difference was not significant for both outcomes. There was no indication of an association between treatment effects and affirmativity of suggestions. Furthermore, studies using neuroleptanesthesia did not differ from those with intravenous or inhalation anesthesia.

Stratifying the analyses according to risk of bias, we only found differences with respect to handling of incomplete outcome data which were significant by trend for PONV (p = .061) with studies evaluated as low risk bias yielding smaller effects than studies judged as unclear risk of bias. Random sequence generation had no influence of treatment effects (Additional file 5: Table S4).