Consistent analgesic effect of intravenous dexamethasone on rebound pain after brachial plexus block: a causal machine learning approach Authors: Boohwi Hong, Yumin Jo, Myungjong Shin, Woolim Min, Woosuk Chung, Chahyun Oh Published: 2026 ABSTRACT AbstractBackgroundRebound pain, a sudden increase in pain intensity after resolution of regional anesthesia, remains a major challenge in postoperative patient care. Intravenous (IV) dexamethasone has been proposed as a preventive strategy, but whether its effectiveness varies across patient characteristics remains unclear.MethodsThis retrospective study included patients who underwent upper extremity orthopedic surgery under brachial plexus block (BPB). Rebound pain was defined as the occurrence of both severe pain (numeric rating scale ≥ 7) and rescue analgesic administration between 4 and 24 hours after BPB. The association between IV dexamethasone and rebound pain was evaluated using multivariable logistic and generalized random forest (GRF) models.ResultsAmong the 1,171 patients, 651 (55.6%) received IV dexamethasone. The dexamethasone group had a significantly lower incidence of rebound pain (7.1% vs. 16.9%, P < 0.001) with delayed onset (median 14 vs. 11 hours). IV dexamethasone was an independent protective factor (odds ratio, 0.42; 95% confidence interval [CI], 0.27 to 0.64; P < 0.001). GRF analysis estimated an average treatment effect of –0.09 (95% CI –0.13 to –0.05), corresponding to a 9% absolute risk reduction. Calibration testing demonstrated accurate model (GRF) estimation (mean coefficient, 0.98; P < 0.001) and no statistically detectable heterogeneity in treatment effect (differential coefficient, –0.30; P = 0.713).ConclusionsIV dexamethasone was associated with reduced rebound pain without evidence of meaningful heterogeneity in treatment effect in patients undergoing upper extremity surgery under BPB. These findings suggest a potential preventive role for dexamethasone as an adjunct to perioperative pain management in this population.Keywords: Anesthetics, Local; Brachial Plexus Block; Dexamethasone; Machine Learning; Nerve Block; Pain Management; Postoperative Pain INTRODUCTION Peripheral nerve blocks (PNBs) are integral to contemporary perioperative pain management, particularly in orthopedic surgeries [1]. They provide effective analgesia, reduce opioid use, and support early mobilization and discharge. However, these benefits are accompanied by a distinct and clinically important drawback: rebound pain, defined as the sudden exacerbation of pain following resolution of PNB [2]. Rebound pain is increasingly recognized as a significant perioperative phenomenon, as it negatively affects patient satisfaction, quality of recovery, analgesic requirements, and unplanned healthcare utilization [3–5]. It affects up to 50% of the patients undergoing surgical procedures with regional anesthesia and is now a top international research priority because of its impact on patient safety and potential to increase opioid use [6]. Despite its frequent occurrence and clinical relevance, the pathophysiological mechanisms underlying rebound pain remain unclear. The proposed mechanisms include nociceptive sensitization, local anesthetic-induced neurotoxicity, and insufficient systemic analgesia during block resolution [4]. Rebound pain has been linked to various clinical factors such as patient characteristics (e.g., younger age and female sex), surgical procedures involving the bone, and inadequate perioperative analgesic strategies [7]. Effective management of rebound pain is essential not only to improve patient comfort but also to minimize the risks associated with uncontrolled postoperative pain, such as increased opioid use, psychological distress, delayed functional recovery, and potential progression to chronic pain [2]. Despite the growing awareness of its clinical impact, evidence-based strategies for managing rebound pain remain limited. Several approaches have been proposed to mitigate this phenomenon, including the use of multimodal analgesia, prolonged-release local anesthetics, and timing-optimized systemic analgesics [4]. Among these, dexamethasone has emerged as a particularly effective strategy that not only reduces the incidence of rebound pain [7–10], but also provides broader perioperative benefits, such as prolonged block duration, reduced postoperative nausea and vomiting, decreased opioid consumption, and improved recovery quality [11]. While dexamethasone shows promise in reducing rebound pain, its effects may not be uniform across all patients [7,12]. Conventional analyses, such as group comparisons and standard regression, often assume treatment effects are the same for everyone, overlooking the variability seen in clinical practice [13,14]. To address this limitation, causal machine learning has emerged as a novel analytic approach. By estimating conditional average treatment effects (CATEs), these methods enable a more nuanced assessment of how treatment responses vary across individuals [15]. Such approaches support personalized perioperative care by identifying patients most likely to benefit from steroid administration and those who may require additional or alternative analgesic strategies. This study aimed to investigate the heterogeneous treatment effect (HTE) of intravenous (IV) dexamethasone in reducing the incidence of rebound pain after brachial plexus block (BPB) following orthopedic upper extremity surgery, and to identify factors that may help distinguish this heterogeneity using real-world perioperative data. Therefore, the authors applied generalized random forest (GRF), a method rooted in causal machine learning, to estimate the ATE and CATEs based on perioperative clinical characteristics. MATERIALS AND METHODS 1. Study design and ethical approval This retrospective observational study was approved by the Institutional Review Board (IRB) of Chungnam National University Hospital (IRB no. 2025-01-036), with a waiver of informed consent due to its retrospective design. This study adhered to the principles outlined in the Declaration of Helsinki and followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. 2. Study population This study analyzed the electronic medical records of adult patients who underwent orthopedic distal upper extremity surgery (up to and including the elbow) under BPB at Chungnam National University Hospital between January 2021 and December 2023. The exclusion criteria were as follows: (1) administration of multiple nerve blocks, defined as the use of additional regional anesthesia at surgical sites other than the primary operative area; (2) incomplete medical records, such as missing intraoperative medication data or analgesic administration details; (3) potentially inadequate BPB, defined as conversion to general anesthesia or a numeric rating scale (NRS) score ≥ 4 within 4 hours after BPB; and (4) history of diabetes mellitus (DM). Patients with DM were excluded because, during the study period, dexamethasone was routinely avoided in these patients due to concerns regarding perioperative glycemic control. As a result, DM status was mostly associated with non-use of dexamethasone, leading to near-complete separation between exposure groups. All procedures were performed by two experienced orthopedic surgeons. 3. Exposure and outcome The primary exposure was intraoperative administration of IV dexamethasone. The patients were categorized into dexamethasone and control groups. The primary outcome was rebound pain, defined as the occurrence of both severe pain (NRS ≥ 7) and rescue analgesic administration between 4 and 24 hours after BPB. The outcome timeframe was chosen to capture the typical pattern of rebound pain, namely the sudden onset of severe pain following block resolution. If no pain scores were recorded, or if pain was documented as “controlled” or “tolerable,” rebound pain was considered absent. The primary objective of this study was to evaluate whether the effect of dexamethasone on rebound pain varies with clinical characteristics (HTE). Secondary outcomes included rescue analgesic use between 4 and 24 hours after BPB, postoperative pain scores, and changes in glucose levels. Postoperative hyperglycemia was evaluated using two clinically relevant thresholds (≥ 180 mg/dL and ≥ 250 mg/dL), each analyzed as a binary outcome [16,17]. 4. Perioperative management All patients underwent standard intraoperative monitoring, including electrocardiography, pulse oximetry, noninvasive or invasive blood pressure monitoring, and capnography. An ultrasound-guided BPB was performed without routine use of perineural adjuvants. Sedation was initiated only after confirming the sensory blockade (either paresthesia or hypoesthesia) at the planned surgical site. The choice and dosage of sedatives were determined at the discretion of the attending anesthesiologist. The use of IV-patient-controlled analgesia (PCA, fentanyl in all cases) was determined at the patient’s request or by an attending anesthesiologist. IV-PCA was administered without a basal infusion, using bolus doses of 10–20 μg of fentanyl with a lockout interval of 10 minutes. From late 2021 to early 2022, IV dexamethasone (5 mg) was routinely administered during anesthesia induction (Fig. 1). During the same period, the intraoperative use of IV non-opioid analgesics (e.g., acetaminophen, non-steroidal anti-inflammatory drugs, nefopam) was also adopted as a standard component of the multimodal analgesic regimen. The choice of analgesic agents (single, dual, or triple combinations) and the timing of administration (before incision, during the procedure, or at the end of the procedure) were not protocolized and were left to the discretion of the attending anesthesiologist. Monthly trends in dexamethasone use, number of intraoperative non-opioid analgesics, and the incidence of rebound pain. Monthly trends in the dexamethasone administration rate (blue solid line, left y-axis), rebound pain incidence (red solid line, left y-axis), and mean number of intraoperative non-opioid analgesics used (green dashed line, right y-axis) between January 2021 and December 2023. A marked increase in the use of dexamethasone and non-opioid agents was observed in early 2022, accompanied by a corresponding decline in the incidence of rebound pain. Postoperative analgesia in the ward, excluding IV-PCA, was managed by the surgical team using IV or oral analgesics (opioid or non-opioid). Oral medications were generally withheld during the first 4 hours after surgery, during which IV analgesics were administered either on a scheduled basis or as needed. If patients reported pain despite scheduled analgesia, rescue analgesics (opioid or non-opioid) were administered with concurrent pain score (NRS) assessments. 5. Data collection Postoperative pain scores and the requirement of rescue analgesics were extracted from the institutional electronic medical record database. The baseline clinical characteristics included age, sex, American Society of Anesthesiologist physical status, comorbidities, and preoperative glucose levels (fasting). The intraoperative characteristics included surgery duration, surgery type (dichotomized by bone involvement), sedative agents used, and intraoperative IV non-opioid analgesics. Postoperative characteristics included the use of analgesics used (opioid or non-opioid; scheduled or as needed), IV-PCA use, and immediate postoperative glucose levels (fasting). Regular analgesic use was defined as scheduled IV or oral analgesic (opioid or non-opioid) administration at least three times within the first 18 hours after BPB (including intraoperative non-opioid analgesic administration), corresponding to the typical window during which rebound pain occurs. Analgesic administrations between 18 and 24 hours after BPB were excluded from this definition because rebound pain occurring during this interval is more likely to be influenced by analgesics administered before that period rather than by medications given during that time. For example, an analgesic administered at 23 hours after BPB would not plausibly affect rebound pain that developed at 22 hours. In contrast, intraoperative analgesics were included because they may affect rebound pain occurring within the early postoperative period. 6. Covariates A causal framework linking dexamethasone (exposure) to the development of rebound pain (outcome) was constructed based on a directed acyclic graph (DAG) informed by domain knowledge (Supplementary Fig. 1). The DAG illustrates assumed causal relationships among variables and was used to guide covariate selection for confounding adjustment. The patient characteristics (age and sex), surgical factors (procedure duration and bone involvement), and analgesic strategies (IV-PCA and regular analgesics) were included as covariates [7]. Patient body weight was also included, given that the fixed dexamethasone dose (5 mg) resulted in variable per-kilogram exposure. 7. Statistical analysis All analyses were conducted using R software (version 4.3.3; R Foundation for Statistical Computing). Continuous variables were assessed for normality using the Shapiro–Wilk test and presented as mean ± standard deviation or median with interquartile range (IQR), as appropriate. Comparisons between the two groups were made using t-tests or Mann–Whitney U-tests. Categorical variables were compared using the chi-square or Fisher’s exact tests, and the results were expressed as counts and percentages. Effect sizes for group-wise differences were reported as mean differences or Hodges–Lehmann estimates for continuous variables, and as odds ratios (ORs) for binary outcomes, with corresponding 95% confidence intervals (CIs). Multivariate logistic regression was conducted to evaluate the independent association between IV dexamethasone and rebound pain. Covariates were selected based on pre-specified DAG. Multicollinearity was assessed using variance inflation factor (VIF). To determine ATE and explore potential HTE, the authors subsequently applied GRF analysis using the “grf” package [15]. GRF is a nonparametric machine-learning approach designed to estimate treatment effects conditional on covariates. The ATE represents the overall mean effect of treatment across the study population, whereas HTE captures variation in treatment effects across individuals with different covariate profiles. The GRF model included the same covariates as the logistic model and generated patient-specific CATE estimates, which quantify the estimated treatment effect for an individual given their covariate profile. To evaluate whether the variability in the estimated CATEs reflected meaningful differences in treatment response rather than random noise, the authors assessed model calibration using a forest-based calibration test. In this framework, a mean prediction coefficient close to 1 indicates accurate estimation of the overall treatment effect, while a differential prediction coefficient close to 1 indicates that the magnitude of the predicted treatment effect heterogeneity is well calibrated. Conversely, a differential coefficient close to 0 suggests limited evidence that the predicted CATEs capture clinically meaningful heterogeneity [18]. The best linear projection (BLP) of the CATE function was then estimated onto the covariates, providing a model-based linear summary of the association between the estimated CATEs and individual covariates [19]. As a complementary, descriptive approach, the authors further visualized differences in the estimated CATEs across clinically relevant subgroups and compared descriptive characteristics between patients in the highest and lowest deciles of the estimated CATE distribution [20]. 8. Sensitivity analysis To evaluate the robustness of their findings, the authors performed multiple sensitivity analyses addressing alternative outcome definitions, covariate specifications, and sample restrictions. First, alternative definitions of rebound pain were examined, including the use of (1) a lower pain threshold (NRS ≥ 4 instead of ≥ 7) and (2) an outcome defined solely by rescue analgesic administration between 4 and 24 hours postoperatively, independent of documented pain scores. Second, the authors varied the modeling of covariates. Specifically, (3) they extended the regular analgesic assessment window from 18 to 24 hours; (4) define regular analgesic use as postoperative scheduled analgesics only; (5) categorized continuous variables (age, weight, and procedure duration); (6) replaced the binary dexmedetomidine indicator with the full intraoperative sedative category (dexmedetomidine, remimazolam, propofol and/or midazolam, or none); and (7) additionally adjusted for surgical site and BPB approach. Finally, to address potential bias from missing outcome data and temporal confounding, (8) the analytic sample was restricted to patients with recorded postoperative NRS scores and, separately, (9) to cohorts defined by variable protocol initiation dates (September 2021 through February 2022). All sensitivity analyses were conducted using both multivariable logistic regression and GRF models consistent with the primary analysis. 9. Interrupted time-series analysis Because the introduction of routine dexamethasone use coincided with broader perioperative practice changes, a segmented interrupted time-series (ITS) analysis was conducted to account for potential time-related confounding. ITS models the underlying temporal trend and evaluates whether an immediate change in outcome level occurs at the time of protocol implementation beyond that trend. Monthly aggregated data were used to estimate the baseline trend and level change in January 2022. Given that multiple practice components were introduced concurrently, ITS findings should be interpreted as reflecting system-level change rather than the isolated effect of dexamethasone. RESULTS 1. Patient characteristics A total of 1,400 patients were assessed for eligibility, with 229 excluded based on pre-specified criteria (37 due to potentially inadequate BPB, 1 due to an additional block beyond BPB, and 191 because of a diagnosis of DM). The final analysis included 1,171 patients, of whom 651 (55.6%) received intraoperative IV dexamethasone. Baseline characteristics are summarized in Table 1. The dexamethasone group had a longer procedure duration, more frequent use of remimazolam, greater intraoperative use of non-opioid analgesics, and more frequent postoperative use of regular analgesics and IV-PCA. Table 1Baseline characteristics and perioperative variables stratified by groupVariableControl (n = 520)Dexamethasone (n = 651)P valueAge (yr)55.0 [37.5, 66.0]56.0 [39.0, 67.0]0.084Sex (female)263 (50.6)350 (53.8)0.305Height (cm)163.0 [155.0, 170.2]162.0 [154.0, 170.1]0.283Weight (kg)64.2 [56.0, 74.3]63.6 [55.0, 74.0]0.410BMI (kg/m2)24.7 [22.2, 26.9]24.6 [22.2, 26.9]0.819ASA class (≥ 3)57 (11.0)65 (10.0)0.655ComorbidityHTN113 (21.7)167 (25.7)0.135CAD12 (2.3)13 (2.0)0.871Procedure duration (min)79.0 [63.5, 100.0]94.0 [74.0, 120.5]< 0.001Bone involvement363 (69.8)443 (68.0)0.561Surgical site0.353Forearm/wrist343 (66.0)455 (69.9)Hand147 (28.3)164 (25.2)Elbow30 (5.8)32 (4.9)BPB type< 0.001Axillary58 (11.2)126 (19.4)Costoclavicular50 (9.6)34 (5.2)Supraclavicular397 (76.3)441 (67.7)Others/unknown15 (2.9)50 (7.7)Sedatives< 0.001Dexmedetomidine97 (18.7)18 (2.8)Remimazolam88 (16.9)591 (90.8)Midazolam and/or propofol318 (61.2)27 (4.1)None17 (3.3)15 (2.3)Regular analgesics*251 (48.3)474 (72.8)< 0.001IV-PCA65 (12.5)174 (26.7)< 0.001Values are presented as median [IQR] or number (%).BMI: body mass index, ASA: American Society of Anesthesiologists physical status, HTN: hypertension, CAD: coronary artery disease, BPB: brachial plexus block, IV: intravenous, PCA: patient-controlled analgesia, IQR: interquartile range.*Defined as scheduled IV or oral analgesic (opioid or non-opioid) administration at least three times within the first 18 hours after BPB (including intraoperative non-opioid analgesic administration). Baseline characteristics and perioperative variables stratified by group Values are presented as median [IQR] or number (%). BMI: body mass index, ASA: American Society of Anesthesiologists physical status, HTN: hypertension, CAD: coronary artery disease, BPB: brachial plexus block, IV: intravenous, PCA: patient-controlled analgesia, IQR: interquartile range. *Defined as scheduled IV or oral analgesic (opioid or non-opioid) administration at least three times within the first 18 hours after BPB (including intraoperative non-opioid analgesic administration). 2. Incidence of rebound pain and other related outcomes Monthly trends showed a sharp decline in the incidence of rebound pain after routine use of intraoperative IV dexamethasone, which began in late 2021, with rates remaining low thereafter (Fig. 1). The crude incidence rates of the outcomes are summarized in Table 2. Overall, rebound pain occurred in 11.4% of patients (134/1,171), with a significantly lower incidence observed in the dexamethasone group than in the control group (7.1% vs. 16.9%, P < 0.001). Fig. 2 shows the temporal distribution of rebound pain onset. In the dexamethasone group, the distribution was distinctly right-shifted, reflecting a significant delay in rebound pain occurrence compared to the control group (median [IQR]: 14.0 [12.0, 17.0] vs. 11.0 [9.0, 13.5] hours, P < 0.001). Table 2Outcomes stratified by groupOutcomeControl(n = 520)Dexamethasone(n = 651)OR or HL difference*(95% CI)P valueRebound pain**NRS ≥ 788 (16.9)46 (7.1)0.37 (0.25 to 0.54)< 0.001**NRS ≥ 4165 (31.7)111 (17.1)0.44 (0.34 to 0.58)< 0.001Rescue analgesic used230 (44.2)160 (24.6)0.41 (0.32 to 0.53)< 0.001Glucose change (mg/dL)–9.0 [–24.0, 17.0]20.0 [0.0, 63.0]–31.0* (–35.0 to –26.0)< 0.001Postoperative hyperglycemia≥ 180 (mg/dL)31 (6.0)122 (18.7)3.64 (2.44 to 5.58)< 0.001≥ 250 (mg/dL)4 (0.8)24 (3.7)4.94 (1.89 to 16.90)0.003Values are presented as number (%) or median [IQR].OR: odds ratio, HL: Hodges–Lehmann estimator, CI: confidence interval, NRS: numeric rating scale, IQR: interquartile range.*HL difference: a nonparametric estimate of the between-group location shift. **With rescue analgesics. Outcomes stratified by group Values are presented as number (%) or median [IQR]. OR: odds ratio, HL: Hodges–Lehmann estimator, CI: confidence interval, NRS: numeric rating scale, IQR: interquartile range. *HL difference: a nonparametric estimate of the between-group location shift. **With rescue analgesics. Distribution of rebound pain onset times according to IV dexamethasone use. The x-axis indicates the time (hour) from the BPB to the onset of rebound pain. Red bars represent patients who received IV dexamethasone, and blue bars represent those who did not (control). IV: intravenous, BPB: brachial plexus block. The overall rate of rescue analgesic use was significantly lower in the dexamethasone group than in the control group (24.6% vs. 44.2%, P < 0.001). NRS reporting rates across the study cohort are presented in Supplementary Table 1. Among patients with available pain scores (27.4% of the cohort), the dexamethasone group reported significantly lower NRS values (5.0 [5.0, 7.0] vs. 6.0 [5.0, 7.0], P = 0.023; Supplementary Fig. 2). 3. Multivariable logistic regression Multivariate logistic regression identified IV dexamethasone as an independent protective factor against rebound pain (OR, 0.42; 95% CI, 0.27 to 0.64; P < 0.001). In addition to dexamethasone, the use of regular analgesics was associated with lower odds (OR, 0.46; 95% CI, 0.31 to 0.67; P < 0.001) of rebound pain, whereas longer procedure duration (per hour) was associated with increased odds (OR, 1.36; 95% CI, 1.02 to 1.79; P = 0.034). No evidence of multicollinearity was observed (VIFs < 2.0). The results of the model are presented in Table 3. Table 3Multivariable logistic regression and BLP analyses for rebound painVariableOR95% CIP valueEstimate95% CIP valueDexamethasone0.420.27 to 0.64< 0.001NANANAAge (per yr)0.990.98 to 1.000.051< 0.001–0.002 to 0.0020.748Sex (male)0.630.40 to 1.000.053–0.010–0.098 to 0.0780.829Weight (per kg)1.011.00 to 1.030.111–0.001–0.004 to 0.0020.522Procedure duration (per hr)1.361.02 to 1.790.034–0.051–0.107 to 0.0050.074Bone involvement0.990.67 to 1.490.9600.004–0.077 to 0.0840.930Regular analgesic use0.460.31 to 0.67< 0.0010.0900.007 to 0.1730.034IV-PCA0.610.34 to 1.030.076–0.003–0.098 to 0.0920.954Dexmedetomidine0.860.46 to 1.520.625–0.030–0.137 to 0.0760.577Logistic regression identifies independent predictors of rebound pain. BLP provides a linear summary of how the estimated treatment effects were associated with the covariates.BLP: best linear projection, OR: odds ratio, CI: confidence interval, IV: intravenous, PCA: patient-controlled analgesia, NA: not applicable. Multivariable logistic regression and BLP analyses for rebound pain Logistic regression identifies independent predictors of rebound pain. BLP provides a linear summary of how the estimated treatment effects were associated with the covariates. BLP: best linear projection, OR: odds ratio, CI: confidence interval, IV: intravenous, PCA: patient-controlled analgesia, NA: not applicable. 4. Generalized random forest analysis The distribution of estimated propensity scores is shown in Supplementary Fig. 3. Although treatment assignment was temporally clustered, partial overlap between groups was observed, conditional on the measured covariates included in the model. The GRF estimated an ATE of –0.09 (95% CI –0.13 to –0.05), i.e., a 9% absolute reduction in rebound pain risk with dexamethasone, similar with the adjusted OR from the logistic model. The model’s calibration test yielded a mean prediction coefficient of 0.98 (P < 0.001), indicating the GRF’s risk predictions were well-aligned with observed outcomes. Furthermore, no significant HTE was detected with a differential prediction coefficient of –0.30 (P = 0.713), suggesting that the benefit of dexamethasone did not meaningfully vary across the patient covariates examined. BLP analysis showed that only the use of regular analgesics was associated with variation in the estimated treatment effects, whereas no statistically detectable associations were observed for other covariates (Table 3). The distribution of estimated CATEs across subgroups is illustrated in Fig. 3. Across most covariate-defined subgroups, the CATE distributions were centered around the ATE (−0.09), with no substantial separation between subgroups. In contrast, a more pronounced difference in the CATE distribution was observed according to the use of regular analgesics. Descriptive comparisons between patients in the highest and lowest CATE deciles are summarized in Supplementary Table 2. Patients in the highest CATE decile tended to be younger, predominantly male, and heavier, had longer procedure durations, a lower proportion of bone-involved procedures, and lower rates of regular analgesic use compared with those in the lowest CATE decile. Distribution of CATEs according to covariate-defined subgroups. Boxplots display the distribution of patient-specific treatment effects estimated by the GRF model for each category of clinical variables. The red vertical line indicates the overall ATE of IV dexamethasone on rebound pain, and the dashed black line represents the null effect (CATE = 0). The largely overlapping distributions across variable categories suggests limited evidence of clinically meaningful treatment effect heterogeneity. IV: intravenous, PCA: patient-controlled analgesia, CATEs: conditional average treatment effects, GRF: generalized random forest. 5. Postoperative glucose elevation and clinical relevance Although IV dexamethasone was associated with a statistically significant increase in postoperative glucose levels (20.0 [0.0, 63.0] vs. –9.0 [–24.0, 17.0], P < 0.001), the incidence of clinically concerning hyperglycemia (≥ 250 mg/dL) was low (3.7%), suggesting limited clinical relevance in the majority of cases (Table 2). 6. Sensitivity analysis Across multiple sensitivity scenarios—including alternative outcome definitions, covariate specifications, and temporal restrictions—the direction and magnitude of the estimated treatment effects were generally consistent with the primary analysis (Supplementary Table 3). In the complete-case analysis restricted to patients with recorded NRS scores, the association was attenuated and no longer statistically significant. Similarly, in the most restrictive cohort definitions, in which the analytic sample was limited to patients treated from January 2022 onward, statistical significance was not consistently maintained in the GRF-based ATE estimates. Taken together, these findings suggest that the primary results were broadly stable, although some attenuation and loss of precision were observed under more restrictive analytic conditions. 7. Interrupted time-series analysis A significant upward trend in rebound pain incidence was observed prior to protocol implementation (1.22%/month, P = 0.019). Following the introduction of the dexamethasone protocol in January 2022, a significant immediate level reduction of 12.4 percentage points was detected (P = 0.021), alongside a sustained monthly decline in the post-intervention period (–0.47%/month, P = 0.004) (Supplementary Fig. 4). DISCUSSION In this retrospective cohort, IV dexamethasone was associated with a lower incidence of rebound pain in patients undergoing upper extremity surgery under BPB. This protective effect was consistently observed in both multivariable logistic regression and GRF analyses. The estimated absolute risk reduction was 9%, which corresponds to a number needed to treat of 12. No meaningful evidence of HTE was identified, suggesting that the protective effect of dexamethasone did not vary substantially across patients with differing baseline risk profiles. These findings remained consistent in the sensitivity analysis using a more inclusive definition of rebound pain (NRS ≥ 4). Previous studies have largely focused on identifying risk factors for rebound pain or on demonstrating the analgesic efficacy of dexamethasone at the population level. The present findings are consistent with prior randomized trials and meta-analyses reporting a preventive effect of dexamethasone on rebound pain [8,10,21]. In contrast, the present study addressed a distinct and more recent question: whether the preventive effect of IV dexamethasone differs across patients with varying clinical risk profiles. By explicitly incorporating established risk factors for rebound pain into a heterogeneity-focused analytical framework, the authors evaluated whether dexamethasone’s analgesic benefit was modified by clinical characteristics. The absence of meaningful HTE suggests that, within the covariates assessed in this study, the effect of dexamethasone is not strongly conditional on clinical characteristics. This finding suggests that the preventive effect of dexamethasone may be relatively consistent across the clinical characteristics examined in this cohort. Importantly, beyond its role in reducing rebound pain, dexamethasone has consistently been shown to improve other perioperative outcomes, including postoperative nausea and vomiting, opioid consumption, and quality-of-recovery scores [22]. These benefits further support its broader inclusion in multimodal analgesic protocols, provided that patient-specific risks are carefully considered [23]. In the authors’ cohort, a uniform dose of 5 mg IV dexamethasone was used across all patients. When normalized by body weight, the administered dose corresponded to a median exposure below 0.1 mg/kg (IQR, 0.067–0.090 mg/kg), consistent with a low-dose regimen. This dosing strategy reflects institutional practice, as dexamethasone is supplied in 5-mg vials, and is supported by prior evidence [24]. Although higher doses have been proposed [25], the present findings indicate that 5 mg was sufficient to meaningfully reduce rebound pain. Because dose-dependent effects could not be excluded, body weight was included as a covariate; however, weight was not significantly associated with rebound pain in the multivariable logistic model and did not demonstrate meaningful treatment effect heterogeneity in the GRF analysis. Notably, a prior study comparing doses below and above 0.1 mg/kg did not identify a significant difference in the incidence of rebound pain, suggesting limited additional benefit from dose escalation beyond this range [26]. Interest in HTE in clinical research has emerged relatively recently. Even within a single clinical category—such as sepsis—substantial variability exists in patient characteristics, etiology, disease severity, and clinical trajectories, and responses to a given treatment may therefore differ across individuals [14]. Under such circumstances, conventional randomized controlled trial approaches that compare uniform treatments at the group level may obscure meaningful individual-level treatment effects by averaging them across heterogeneous populations. HTE-focused analyses are designed to address this limitation, and this conceptual framework formed the core motivation of the present study. Although the mechanisms underlying rebound pain remain incompletely understood, the relatively homogeneous nature of this study cohort—orthopedic upper extremity surgery (up to and including the elbow) involving BPB—may partly explain the absence of detectable HTE. It is plausible that more pronounced HTE could emerge in broader clinical contexts, such as shoulder or more proximal upper-extremity surgery, pediatric populations, or procedures performed under neuraxial anesthesia. Importantly, although the authors did not detect statistically significant HTE, this finding should be interpreted cautiously because the analysis was limited to routinely available clinical variables. Their models were limited to routinely available clinical variables and did not incorporate biological or psychological factors—such as chronic anxiety, pain related behaviors, baseline cortisol level—that may influence individual responsiveness to dexamethasone [12]. Crucially, the absence of meaningful HTE in the preventive effect of dexamethasone on rebound pain should not be conflated with homogeneity in the underlying risk of rebound pain itself. In other words, although the analgesic benefit of dexamethasone appears to be relatively consistent across patients, baseline susceptibility to rebound pain may still vary substantially according to patient-or-procedure-related characteristics. Accordingly, there remains considerable scope for individualized, risk-adapted perioperative pain management strategies, even in the absence of differential treatment effects of dexamethasone. In patients with known high-risk profiles, consideration of additional preventive measures alongside dexamethasone may therefore remain clinically relevant, and identifying effective approaches tailored to such high-risk groups represents a promising area for future research. In such subgroups, additional preventive strategies—such as continuous peripheral nerve catheters or enhanced multimodal analgesia—may warrant consideration. In the BLP analysis, regular postoperative analgesic use was statistically associated with variation in the estimated treatment effect. However, this should not be interpreted as evidence of clinically meaningful heterogeneity. Across all strata, the estimated CATEs remained consistently below zero, indicating that dexamethasone conferred benefit irrespective of regular analgesic administration, with only modest differences in effect magnitude. Regular analgesic use may function as a surrogate for baseline pain burden or anticipated surgical nociception, reflecting treatment policy rather than a true biologic modifier of steroid responsiveness. Taken together with the non-significant differential calibration test, these findings suggest that any observed heterogeneity is small and unlikely to be clinically meaningful. Although IV dexamethasone is increasingly utilized for its antiemetic and analgesic benefits [27], concerns persist regarding its potential side effects, particularly hyperglycemia, and the risk of infection. However, the current evidence suggests that a single perioperative dose is generally safe for most patients [11]. In the authors’ cohort, although dexamethasone use was associated with a statistically significant increase in postoperative glucose levels, the incidence of hyperglycemia (≥ 250 mg/dL) was low (3.7%). This finding is consistent with prior studies indicating that transient hyperglycemia following dexamethasone administration is common but rarely clinically significant in non-cardiac surgery [28,29]. Together with the observed reduction in rebound pain, these findings further support the perioperative use of dexamethasone. However, as patients with DM were not included in this study, caution is warranted when extrapolating the results to the DM population. Although the magnitude of glucose elevation observed in the authors’ cohort was comparable to that reported in previous studies including patients with DM [30], evidence regarding the effect of dexamethasone on rebound pain specifically in patients with DM remains limited. In addition, inadequate postoperative analgesia itself may contribute to stress-induced glucose intolerance [31]. Therefore, when considering dexamethasone use in patients with DM—particularly those with poorly controlled disease—the potential trade-off between improved analgesia and glycemic perturbation should be carefully weighed. At present, most evidence supporting the perioperative safety of dexamethasone in patients with poorly controlled diabetes remains insufficient, and cautious individualized decision-making is warranted. Additional sensitivity analyses provided insight into the robustness of the observed association. Similar results were obtained when rebound pain was defined solely by rescue analgesic administration, an outcome definition less dependent on pain score documentation. When the analysis was restricted to patients with documented postoperative pain scores or to the most restrictive cohort definitions, statistical significance was attenuated, likely reflecting reduced sample size and potential documentation variability in routine practice. Nevertheless, the direction and magnitude of the estimated treatment effect remained broadly similar across analytic specifications, suggesting that the observed association was not driven by a single modeling assumption. This study has several limitations. First, the retrospective design introduces the potential for unmeasured confounding, despite the efforts to adjust for key covariates. Although DAG and GRF models were employed to strengthen causal inference, these approaches cannot fully eliminate residual bias inherent to observational data. Nonetheless, dexamethasone administration in the authors’ cohort was largely driven by temporal practice changes rather than individual patient characteristics, likely making residual selection bias limited after the exclusion of patients with DM. Second, because dexamethasone use differed markedly between earlier and later periods, the study period itself (before vs. after) could not be directly included as a covariate. Instead, the authors adjusted for key subcomponents reflecting temporal practice changes—such as dexmedetomidine use, IV-PCA, and regular analgesic administration—to address time-related confounding. The validity of this approach relies on the assumption that no other unmeasured practice changes occurred during the study period; if such changes were present, the estimated treatment effects may have been over- or underestimated. Third, as a single-center study, the present findings may be influenced by institution-specific clinical practices, potentially limiting generalizability. For example, techniques related to BPB—including the use of perineural adjuncts, preferred approaches, and local anesthetic injection strategies—may vary across institutions and could influence the outcomes [32]. Fourth, patients with DM were excluded from the analysis, which may also limit the generalizability of the present findings. Further studies specifically evaluating the safety and effectiveness of dexamethasone in diabetic populations are warranted. Finally, analgesic exposure was analyzed as a broad category of regular analgesic use rather than as individual drug classes. Consequently, drug class–specific effects on rebound pain could not be evaluated. In conclusion, IV dexamethasone was associated with a lower incidence of rebound pain in patients undergoing upper extremity surgery under BPB, with no statistically meaningful heterogeneity detected across the examined covariates. These findings suggest a potential preventive role for dexamethasone as an adjunct to perioperative pain management in this population. SUPPLEMENTARY MATERIALS Supplementary materials can be found via https://doi.org/10.3344/kjp.25431. ACKNOWLEDGMENTS We thank the use of the Medical Data Core of the Regional Medical Research Capability Enhancement Project, Biomedical Research Institute, Chungnam National University Hospital, for providing technical assistance, instrumentation, and data analysis support. Footnotes DATA AVAILABILITY The datasets supporting the finding of this study are available from the corresponding author upon reasonable request. CONFLICT OF INTEREST Boohwi Hong is a section editor of the Korean Journal of Pain. However, he was not involved in the selection of peer reviewers, the evaluation, or the decision-making process for this article. No other potential conflict of interest relevant to this article was reported. FUNDING This work was supported by the National Research Foundation of Korea (NRF-2022R1C1C1007982) and Chungnam National University. The funder had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. AUTHOR CONTRIBUTIONS BH: conceptualization, methodology, supervision, formal analysis, writing–original draft, and writing–review and editing; YJ: data curation, investigation, and formal analysis, writing–original draft, and writing–review and editing; MS: data curation and investigation; WM: data curation and visualization; WC: methodology and validation; CO: conceptualization, project administration, writing–review and editing, and responsible for the overall content as the guarantor. REFERENCES Associated Data Supplementary Materials