Perioperative interventions in preventing postoperative delirium following orthopaedic surgery in the elderly: a Bayesian network meta-analysis

Abstract

Background

Postoperative delirium (POD) is an acute neuropsychiatric syndrome that occurs following surgery, characterized by inattention and broader cognitive deficits. This study aimed to synthesize randomized evidence on perioperative strategies that prevent POD in elderly patients undergoing orthopaedic surgery.

Methods

We conducted a Bayesian network meta-analysis of randomised controlled trials (RCTs) retrieved from PubMed, Embase, Web of Science, and Cochrane Library from inception to March 2026. Eligible trials enrolled adults aged 60 years or older undergoing orthopaedic surgery and reported POD incidence with validated assay assessments. Risk ratios (RR) and mean differences with 95% Credible Interval (CrI) were estimated. The risk of bias was assessed using the Cochrane tool, and the certainty of the evidence was evaluated with the GRADE (Grading of Recommendations Assessment, Development, and Evaluation).

Results

79 RCTs comprising 16,012 patients were included. Compared with placebo, dexmedetomidine (RR 0.49, 95% CrI 0.39-0.61) has the most consistent evidence for reducing the risk of POD incidence, while ketamine (RR 0.39, 95% CrI 0.23-0.64), rivastigmine (RR 0.33, 95% CrI 0.16-0.65), olanzapine (RR 0.35, 95% CrI 0.17-0.71), and lidocaine (RR 0.41, 95% CrI 0.21-0.77) showed promising but heterogeneous benefits across current literature. Evidence for analgesic and anaesthetic interventions was limited.

Conclusion

Dexmedetomidine offers the most substantial evidence for lowering POD incidence after orthopaedic surgery in elderly; ketamine, rivastigmine, olanzapine, and lidocaine warrant further evaluation. High-quality, low-bias RCTs are needed to verify the effects of specific analgesic or anaesthetic regimens on POD occurrence.

Translational potential statement

This network meta-analysis highlights the promising translational potential of dexmedetomidine in preventing POD for elderly orthopaedic patients. Dexmedetomidine emerges as the most reliable option for immediate clinical adoption, while ketamine, rivastigmine, olanzapine, and lidocaine show promise but require confirmatory trials before routine use. The findings set the stage for personalized anaesthetic protocols that minimize POD and its associated morbidity and healthcare costs.

Graphical abstract

1. Introduction

Postoperative delirium (POD) is a common and highly distressing adverse event following surgery [1]. Depending on the patient population and surgical acuity, the reported incidence of POD varies widely, ranging from 14% in elective non-cardiac procedures to as high as 64% in emergency and acute care settings [[2], [3], [4]]. Although transient, POD is associated with severe clinical consequences including major postoperative complications, accelerated long-term cognitive decline, and increased mortality [5,6]. Furthermore, POD imposes a substantial burden on both patients and healthcare systems. The economic costs are significantly escalated due to prolonged hospitalization, the need for intensive nursing care, and higher rates of subsequent institutionalization [2].

Preventing POD has become a vital target for perioperative therapy, and recent studies highlight both pharmacological interventions, such as dexmedetomidine and melatonin, and non-pharmacological options, such as refined analgesia/anaesthesia protocols [7,8]. However, evidence specific to elderly patients undergoing orthopaedic procedures, those who are at higher risk for ischemia-reperfusion injury, pain, blood loss, comorbid diabetes, and functional impairment, remains limited [9,10]. Moreover, most studies only focused on a single pharmacological agent and direct pairwise comparisons, with less attention paid to analgesic and anaesthetic techniques.

Recently, several network meta-analyses (NMAs) have evaluated various POD prevention strategies. For example, a comprehensive study demonstrated the efficacy of specific drugs like dexmedetomidine [11]. Similarly, other parallel syntheses confirmed the benefits of multi-component non-pharmacological [12,13] and refined anaesthetic protocols [14]. However, these broad reviews share critical limitations. By pooling data across highly heterogeneous surgical specialties (e.g., cardiac, general, and orthopaedic) and various age groups, the clinical applicability of those findings may be diluted for specific populations. Furthermore, their literature searches were censored earlier, missing the most contemporary trials.

We, therefore, conducted a Bayesian network meta-analysis (NMA) to integrate both direct and indirect evidence across pharmacological, analgesic, and anaesthetic interventions, strictly isolating elderly patients scheduled for orthopaedic surgery. This specific age restriction is crucial; Indeed, biologically and methodologically, age profoundly modifies the neuro-pharmacological response to prophylactic agents, and excluding younger patients prevents the severe dilution of absolute treatment effects seen in mixed-age cohorts [11,15,16]. This analysis aims to estimate comparative effects, rank interventions probabilistically, and appraise the certainty of evidence with the grading of recommendations assessment, development, and evaluation (GRADE) adapted for NMA.

2. Methods

2.1. Information sources and search strategy

This study followed PRISMA guidelines [[17], [18], [19]], and we searched PubMed, Embase, Web of Science, and Cochrane Library database from inception to March 2026. The search strategy combined terms for “(pharmacological) and (analgesia) and (anaesthesia)”, along with other related interventions (Supplementary Table S1), and was limited to full-text human studies in English to avoid misinterpretation during translation from other languages [20]. In addition, closely relevant references to the current research topic were manually searched and screened. The protocol of this study was registered in PROSPERO (CRD420251178454).

2.2. Eligibility criteria

The analysis included randomized controlled trials (RCTs) enrolling older adults ≥60 years old undergoing orthopaedic surgery that evaluated perioperative interventions, including pharmacological, anaesthetic, and analgesic techniques, aimed at preventing POD. POD was assessed with validated diagnostic tools, such as the Confusion Assessment Method (CAM) and the Delirium Rating Scale-Revised-98 (DRS-R-98). No restrictions were applied by sex or race. Exclusion criteria included non-randomized designs (including observational, retrospective, or crossover designs), non-orthopaedic surgical studies, studies where the predefined lower age limit for inclusion was below 60 years (unless separate extractable data for the ≥60 years subgroup were explicitly provided), interventions initiated outside the immediate perioperative period (i.e., to be eligible as prophylaxis, interventions had to be initiated preoperatively, intraoperatively, or within the first 24 h postoperatively; interventions beginning on or after postoperative day 2 were excluded), and studies not reporting POD incidence.

Regarding the study selection process, after automated deduplication in EndNote (version 20), the residual duplicates were manually screened by comparing the first author, year, title, and page number. Two researchers independently screened titles/abstracts and full texts, resolving disagreements through discussion or consultation with a third reviewer.

2.3. Data extraction and outcomes

All data were carefully extracted from eligible studies by the two researchers independently, with predesigned data collection forms. Extracted information included authors, year of publication, trial design, sample size, perioperative interventions, surgery types, comparators, anaesthetic method, delirium assessment tool, follow-up window, and the incidence of POD per arm (Supplementary Table S2). The primary outcome was POD incidence, defined as the first positive CAM or DRS-R-98 assessment recorded within the in-hospital period or within the first 7 postoperative days.

2.4. Risk-of-bias assessment

Two investigators independently evaluated the risk of bias for each included trial using the Cochrane Risk of Bias tool (RoB 2.0). The assessment was structured across five mandatory domains: (1) bias arising from the randomization process; (2) bias due to deviations from intended interventions; (3) bias due to missing outcome data; (4) bias in measurement of the outcome; and (5) bias in selection of the reported result. Using the signaling questions provided by the tool, each domain and the overall risk of bias for each study were adjudicated as “low risk," “some concerns," or “high risk”. Any discrepancies between the reviewers were resolved through consensus or via adjudication by a third senior investigator.

2.5. Statistical analysis

For each clinical outcome, an NMA was carried out with the overall evidence that formed a connected network containing a minimum of two studies in total [21]. This ensures the minimum network structure required to evaluate at least three interventions and generate indirect evidence, without necessitating closed loops. For outcomes reported by only a single study, a standard pairwise meta-analysis was conducted. We further adopted a Bayesian framework for the NMA due to its flexibility in handling complex network geometries and underlying assumptions [22]. Specifically, the analysis was implemented as a Bayesian hierarchical random-effects model utilizing a binomial likelihood and logit link. Following established methodology for mixed treatment comparisons, this hierarchical structure distinguishes between basic and functional parameters to coherently synthesize direct and indirect evidence [23,24]. To ensure data-driven posterior estimates, non-informative priors were applied to treatment effects, alongside a standard uniform prior for between-study variance. Finally, Markov chain Monte Carlo (MCMC) simulations were employed to estimate these models and integrate disparate evidence sources into consistently relative treatment effects. Four parallel chains were run with 25,000 iterations each, including an initial 5000 burn-in for convergence, yielding 20,000 posterior samples per chain [25]. Convergence of the chains was assessed visually with trace plots and quantitatively with the potential scale reduction factor (PSRF, Gelman-Rubin statistic) [26]. If PSRF <1.05, the model is deemed to have converged.

Interventions were grouped into arms based on their action mechanism, clinical similarity, and pharmacological class, to ensure validity and clinical relevance across studies. When multiple doses or routes of administration existed for the same intervention, they were combined into a single arm if clinical judgment and prior evidence indicated comparable effects. Each distinct intervention or combination of interventions is treated as an independent arm within the network. Common control or standard care is considered as a placebo to standardise the entire network. The resulting evidence network comprises all interventions connected directly or indirectly through at least one common control group.

The fundamental assumption of transitivity was evaluated conceptually before conducting the network meta-analysis. Because transitivity cannot be assessed via direct statistical analysis, two independent reviewers rigorously compared the distribution of potential clinical and methodological effect modifiers across the different pairwise comparisons. These modifiers included patient demographics, orthopaedic surgery site, delirium assessment method, and follow-up times. If substantial clinical violations of the transitivity assumption were identified, generating indirect evidence would be deemed invalid.

A standard Bayesian consistency model with a binomial likelihood was used for the NMA [23]. This model assumes global consistency by mathematically constraining indirect comparisons as exact liner combinations of direct comparisons. Results are presented as a pooled risk ratio (RR) with 95% credible intervals (Crls). Model fit was evaluated using the deviance information criterion (DIC), with smaller values indicating a better fit [27]. The between-study heterogeneity was estimated as the standard deviation of the random effects, and heterogeneity was further quantified using the I² statistic. When the mean residual deviation of the model approaches the number of data points, it indicates the model fits well. Consistency was rigorously evaluated at three levels: network level, loop level (using inconsistency factors), and at the comparison level via node-splitting.

2.6. Assessment of certainty of evidence

The certainty of evidence for the network estimates was evaluated with the GRADE framework adapted for network meta-analysis [28,29]. Starting from a baseline of “high” certainty for randomized controlled trials, we then systematically assessed evidence across four domains: study limitations (RoB 2.0), indirectness/intransitivity, inconsistency (via node-splitting models), and imprecision (assessed through Optimal Information Size and clinical decision thresholds of RR). The overall certainty for each comparison was ultimately categorized as high, moderate, low, or very low. Detailed operational definitions, quantitative thresholds, and rationales for each grading decision are provided in the Supplemental Methods.

Publication bias and small-study effects were evaluated using comparison-adjusted funnel plots and Egger's test (p-value <0.05). Given the limited statistical power of these methods and following Cochrane Handbook recommendations [30], formal asymmetry testing was strictly restricted to outcomes comprising at least 10 studies; outcomes with fewer studies were not assessed.

Methodological heterogeneity was assessed by examining the similarity of point estimates, the overlap of confidence intervals, and statistical tests. All analyses were performed using the R statistical software (version 4.4.1, http://www.r-project.org).

3. Results

Our search identified 3432 published articles in total, of which 1965 were removed due to irrelevant topics (4) and duplications (1,961); 1294 were excluded at title/abstract screening and 94 were excluded due to different/wrong outcomes, interventions, patient populations, or others at full text screening, leaving 79 RCTs that met inclusion criteria and proceeded to subsequent analysis [[31], [32], [33], [34], [35], [36], [37], [38], [39], [40], [41], [42], [43], [44], [45], [46], [47], [48], [49], [50], [51], [52], [53], [54], [55]], [[56], [57], [58], [59], [60], [61], [62], [63], [64], [65], [66], [67], [68], [69], [70], [71], [72], [73], [74], [75], [76], [77], [78], [79], [80], [81], [82], [83], [84], [85], [86], [87], [88], [89], [90]], [[91], [92], [93], [94], [95], [96], [97], [98], [99], [100], [101], [102], [103], [104], [105], [106], [107], [108], [109]] (Fig. 1). In these 79 included trials, 4 trials were three-arm, 2 were four-arm, and the rest were two-arm design. Orthopaedic surgeries were performed primarily on the hip and knee, and less frequently on the thoracic, lumbar, femur, pelvis, spine, and shoulder. POD was most commonly assessed with the CAM. All the included trials reported the incidence of POD, typically until discharge or 3-7 days postoperatively, and a small number reported a shorter follow-up of 1-2 days (Supplementary Table S2). In addition, regarding the transitivity assumption, a comprehensive assessment revealed that the distributions of clinical and methodological effect modifiers were sufficiently similar across the network. No substantial clinical violations were identified, confirming the validity of the indirect comparisons.

Fig. 1.

Flowchart of the study selection.

The overall analysis included 16,012 patients, covering 31 pharmacological agents, 9 analgesic techniques, and 2 anaesthetic techniques. For the network meta-analysis, the primary pharmacological network was constructed with 30 specific agents and placebo as the common comparator (Fig. 2). The random-effects model performed in NMA demonstrated excellent fit (DIC = 239.54, I² = 3%), with a total residual deviance (132.3) closely approximating the 130 data points (ratio 1.018), indicating no significant lack of fit. Across the network, dexmedetomidine was the most frequently evaluated drug with 2746 participants in 24 trials and propofol was the second drug with 1420 participants in 12 trials. Total 12 drugs significantly reduced POD risk compared to placebo; 5 were supported by high or moderate certainty evidence, including dexmedetomidine (RR 0.49, 95% CrI 0.39-0.61; moderate certainty), ketamine (RR 0.39, 95% CrI 0.23-0.64; high certainty), rivastigmine (RR 0.33, 95% CrI 0.16-0.65; moderate certainty), olanzapine (RR 0.35, 95% CrI 0.17-0.71; moderate certainty), and lidocaine (RR 0.41, 95% CrI 0.21-0.77; moderate certainty). Other significant agents, such as insulin and melatonin, were associated with lower certainty (Table 1). Additional comparative results are detailed in Supplementary Table S3.

Fig. 2.

Bayesian network results for pharmacological interventions. (a) Network plot; (b) Forest plot for different pharmacological interventions compared with placebo. Dot sizes correspond to number of participants. CrI, credible intervals.

Table 1.

Summary of findings and grading of recommendations assessment, development and evaluation.

Intervention vs Placebo	RR (95% CrI)	Study Limitations	Indirectness	Inconsistency	Imprecision	Overall Certainty
Significant Protection (95% CrI does not cross 1.0)
Dexmedetomidine	0.49 (0.39 to 0.61)	0	0	−1d	0	Moderate
Ketamine	0.39 (0.23 to 0.64)	0	0	0	0	High
Carbohydrate	0.21 (0.04 to 0.73)	−1a	0	0	−1e	Low
Hypertonic saline	0.29 (0.10 to 0.74)	−1b	0	0	−1e	Low
Rivastigmine	0.33 (0.16 to 0.65)	0	0	0	−1e	Moderate
Insulin	0.34 (0.19 to 0.59)	0	0	−1d	−1e	Low
Olanzapine	0.35 (0.17 to 0.71)	0	0	0	−1e	Moderate
Methylene blue	0.40 (0.15 to 0.99)	−1a	0	0	−1e	Low
Lidocaine	0.41 (0.21 to 0.77)	0	0	0	−1e	Moderate
Parecoxib	0.49 (0.25 to 0.92)	−1a	0	0	−1e	Low
Dexamethasone	0.50 (0.24 to 1.00)	−1a	0	0	−1e	Low
Melatonin	0.58 (0.36 to 0.84)	0	0	−1d	−1e	Low
No Significant Difference (95% CrI crosses 1.0)
Ciprofol	0.29 (0.06 to 1.13)	−1a	−1c	0	−1e	Low
Morphine	0.36 (0.03 to 3.49)	0	−1c	0	−2f	Very Low
Methylprednisolone	0.51 (0.20 to 1.23)	0	0	0	−2f	Low
Xenon	0.59 (0.17 to 2.07)	0	−1c	0	−2f	Very Low
Remimazolam	0.73 (0.34 to 1.46)	0	−1c	0	−2f	Very Low
Esketamine	0.77 (0.30 to 2.01)	0	0	0	−2f	Low
Desflurane	0.82 (0.15 to 4.79)	0	−1c	0	−2f	Very Low
Sevoflurane	0.84 (0.36 to 1.99)	0	−1c	0	−2f	Very Low
Haloperidol	0.91 (0.44 to 1.89)	0	0	0	−2f	Low
Iron	0.94 (0.40 to 2.27)	0	0	0	−2f	Low
Propofol	1.19 (0.78 to 1.84)	0	0	0	−1e	Moderate
Clonidine	1.23 (0.61 to 2.45)	−1a	0	0	−2f	Very Low
Midazolam	1.25 (0.80 to 1.94)	−1a	−1c	0	−1e	Very Low
Sufentanil	1.32 (0.28 to 6.54)	0	−1c	0	−2f	Very Low
Fospropofol	1.42 (0.51 to 4.18)	0	−1c	0	−2f	Very Low
Sodium oxybate	1.53 (0.61 to 3.98)	0	0	0	−2f	Low
Minocycline	1.78 (0.39 to 9.73)	0	0	0	−2f	Low
Ramelteon	1.90 (0.27 to 16.26)	0	0	0	−2f	Low

^aDowngraded due to ≥ 50% “Some concerns”.

^bDowngraded due to ≥50% “High risk”.

^cDowngraded due to indirect evidence contributed to ≥50%.

^dDowngraded due to substantial unexplained heterogeneity (dexmedetomidine:79.2%; insulin 91.3%; melatonin 84%) was observed within the comparisons.

^eDowngraded due to total sample size <400.

^fDowngraded due to total sample size <400 and the 95% CrI was simultaneously crossing clinical thresholds for both appreciable benefit and harm.

The NMA results comparing different analgesic techniques are presented in Fig. 3. Fascia iliac compartment block (FICB) was the most frequently studied analgesic technique (Fig. 3a). Several techniques showed point estimates potentially reducing POD incidence, such as patient-controlled intravenous analgesia (PCIA) (RR 0.13, 95% CrI 0.01-1.17), lumbosacral plexus block (LPB) (RR 0.34, 95% CrI 0.07-1.38), and FICB (RR 0.50, 95% CrI 0.19-1.57), but none showed a statistically significant reduction compared to placebo medication for analgesia (PMA) (Fig. 3b). In the standard pairwise meta-analysis, no significant differences were found between general and regional anaesthetic techniques (pooled RR 0.94, 95%CI 0.61-1.44) (Fig. 4).

Fig. 3.

Bayesian network results for analgesic techniques. (a) Network plot; (b) Forest plot for interventions compared with PMA. PMA, placebo medication for analgesic; LPBdeep, lumbosacral plexus block (deep sedation, BIS: 40 to 60); LPBlight, lumbosacral plexus block (light sedation, BIS: 60 to 80); FICB, fascia iliac compartment block; PENG, pericapsular nerve group block; FNB, femoral nerve block; ESPB, erector spine plane block; PCIA, patient-controlled intravenous analgesia. Dot sizes correspond to number of participants. CrI, credible intervals.

Fig. 4.

Pairwise results for anaesthesia techniques by forest plot.

Overall, among the 79 included RCTs, 50 studies (63.29%) were classified as having a low or possibly low risk of bias. The remaining 29 studies (36.71%) were judged to have some concerns (n = 22) or a high risk of bias (n = 7), predominantly due to selective reporting and poor trial registration (Fig. 5). Regarding specific domains, 61 trials (77.22%) generated adequate random sequences and implemented sufficient allocation concealment. For deviations from intended interventions, 70 studies (88.61%) were judged as low risk, indicating successful blinding of participants and researchers. No risk for missing outcome data was identified in any trial (100%), and 69 trials (87.34%) achieved low risk in the measurement of outcomes. Additionally, 56 trials (70.89%) demonstrated no evidence of bias in the selection of reported results (Supplementary Table S4).

Fig. 5.

Summary plot of risk of bias assessments for included randomized controlled trials, using Cochrane Risk of Bias tool (version 2).

In the sensitivity analysis excluding comparisons with blank control or standard of care (DIC = 218.94, I² = 4%), melatonin no longer demonstrated a significant protective effect. Carbohydrate was removed because its evidence base relied on comparisons with the excluded blank control groups (Supplementary Fig. S1). To account for potential bias from small study effects, analysis excluding intervention evaluated in only one trial and trials of fewer than 100 participants (DIC = 171.83, I² = 2%) confirmed the sustained efficacy of rivastigmine, insulin, ketamine, lidocaine, parecoxib, dexmedetomidine, and melatonin (Supplementary Fig. S2). Besides, after excluding studies at high risk of bias (DIC = 229.33, I² = 3%), the protective effect of dexamethasone failed to reach significance, while hypertonic saline was excluded as its evidence base relied entirely on high-risk trial (Supplementary Fig. S3).

The Gelman-Rubin convergence test showed that all parameters converged and the trace plots were stable (Supplementary Fig. S4 and Table S5). Potential publication bias was examined using a comparison-adjusted funnel plot for all pairwise intervention comparisons (Supplementary Figs. S5 and S6). Most effects lay within the funnel triangle with generally symmetrical distributions, particularly among comparisons with low standard errors. Egger's test detected no evidence of small-study effects for any of the intervention comparisons, suggesting an overall low risk of publication bias across the analysis. Furthermore, the node-splitting model revealed no significant inconsistency between direct and indirect evidence (all p-values>0.05), supporting the validity of the consistency assumption across the entire network (Supplementary Table S6).

4. Discussion

In this systematic review and Bayesian network meta-analysis of 79 RCTs enrolling 16,012 elderly patients undergoing orthopaedic surgery, we identified 12 pharmacological agents that reduced POD incidence compared with placebo/standard care. Crucially, only five of these were supported by high-to-moderate certainty evidence (dexmedetomidine, ketamine, rivastigmine, olanzapine, and lidocaine), underscoring their relative clinical reliability. In contrast, while most analgesic interventions yielded point estimates favourable benefit, none of them reached the threshold for statistical credibility. Similarly, our findings suggest that the choice between general and regional anaesthesia does not significantly influence POD risk.

POD is an acute, fluctuating disturbance of consciousness and cognition arising from a multifactorial pathophysiology. Surgery-induced neuroinflammation, accompanied by microglial and astrocytic activation, endothelial dysfunction, and peripheral immune activation with release of pro-inflammatory cytokines, was suggested to serve as a potential key driver of POD [110,111]. Neurotransmitter imbalances, particularly cholinergic deficiency, relative dopaminergic hyperactivity, and alterations in serotonin and noradrenaline signalling, may further contribute to attentional deficits and cognitive impairment [111]. Additionally, sleep-wake cycle disruption can exacerbate the risk of POD by disturbing circadian rhythm regulation and affecting neurotransmitters [112]. Additional contributors include metabolic and endocrine dysregulation, blood-brain barrier compromise, impaired cerebral oxygenation, and adverse drug effects [[113], [114], [115]].

Recent guidance from the European Society of Anaesthesiology and Intensive Care (2024) and a previous study suggest that pharmacology, such as dexmedetomidine, may reduce the incidence of POD and shorten intensive care unit length of stay [28,116]. Dexmedetomidine is a highly selective α₂-adrenergic receptor agonist with sedative and anxiolytic effects. Its preventive effect on POD may include improvements in hypoxemia, analgesia, sleep quality, and inflammatory control [8,85,117]. Its efficacy has been reported across cardiac, non-cardiac, and neurosurgical settings [8,118,119], and our analysis strengthens the evidence specifically in older orthopaedic patients. Our findings demonstrate that perioperative ketamine reduced the incidence of POD in elderly patients undergoing orthopaedic surgery. This observation concurs with the latest meta-analytical evidence [120], which reported a robust protective effect in elderly populations. Although our results diverge from the 2023 review by Fellous et al. [121], which found no significant benefit, the discrepancy likely stems from substantial heterogeneity in ketamine regimens, administration timing, and baseline risk stratification across earlier trials.

Similarly, evidence for the neuromodulator agents, rivastigmine and olanzapine, to prevent POD remains conflicting across the literature [[122], [123], [124]]. The proposed mechanism for rivastigmine involves enhancing cholinergic transmission by increasing acetylcholine availability, while olanzapine is considered to modulate dopaminergic and serotonergic imbalance via receptor blockade. Our results showed signals of benefit for both agents despite prior inconsistencies, underscoring the need for larger, well-designed trials with clear stratification. Notably, our findings indicate a significant protective effect of lidocaine against POD, a result that diverges from the 2026 network meta-analysis [11]. In their synthesis, lidocaine did not show a statistically significant reduction in delirium incidence across broader surgical populations. This discrepancy may be attributed to the specific high-risk profile and inflammatory burden of our orthopaedic elderly cohort compared to the more diverse settings examined in larger-scale analyses. Given the conflicting evidence and the potential for population-specific benefits, further high-precision research is essential to definitively elucidate the relationship between perioperative lidocaine and delirium, as well as to identify the optimal dosing and timing for this intervention.

Pain is closely associated with POD, and adequate analgesia is, therefore, expected to reduce POD risk through pain relief [125]. A Cochrane review reported a lower pooled incidence of POD of 10.5% in non-pharmacological interventions compared to 18.4% in standard care [126]. In addition, a previous trial reported that regional nerve block reduced POD in patients without cognitive impairment [127]. Our analysis did not replicate this effect, possibly because our synthesis grouped block types broadly and did not stratify by anatomical site or techniques. Meanwhile, most trials included in this study employed optimised and refined intraoperative analgesia programmes encompassing preoperative, intraoperative, and postoperative analgesia and sedation, which likely reduced additional benefits from regional blocks on POD. Current meta-analyses also did not show a lower incidence of POD with general anaesthesia compared to regional anaesthesia [128,129], and our findings align with this literature and strengthen the overall evidence base.

This study has several strengths. First, to the best of our knowledge, it is the first comprehensive Bayesian network meta-analysis to examine pharmacological, analgesic, and anaesthetic interventions for preventing POD in elderly orthopaedic patients. We synthesized both direct and indirect evidence across a broad set of interventions and assessed certainty using the GRADE framework, with appropriate downgrading for risks, such as lack of blinding and other bias. In addition, to maximize generalizability, we imposed no restrictions on sample size, intervention type, delirium assessment tools, or follow-up duration.

However, some limitations should be acknowledged. First, although diverse interventions were included in our study, some comparisons cannot be made into a single evidence network, which prevented direct cross-class comparisons between non-pharmacological and pharmacological interventions. Second, evidence and sample sizes for specific agents, such as rivastigmine, insulin, and olanzapine, are relatively limited, which reduces the quality of the evidence. Third, our search was restricted to English-language publications to ensure accurate data extraction and risk-of-bias assessment; however, methodological evidence suggests that excluding non-English trials rarely biases the overall conclusions of clinical meta-analyses [20]. Finally, due to the limited power of statistical tests when study numbers are low, publication bias and small-study effects were only formally evaluated for outcomes involving 10 or more trials. While this approach aligns with current methodological standards to prevent false-negative results, it limits our ability to definitively assess the reporting bias for interventions supported by a smaller evidence base. Therefore, these results should be interpreted with caution. Further evidence would be strengthened by well-designed multicentre RCTs with standardized medication protocols, clearly defined outcomes, and sufficient statistical power.

5. Conclusions

Among elderly patients undergoing orthopaedic surgery, dexmedetomidine has the most consistent evidence for reducing the risk of POD, while ketamine, rivastigmine, olanzapine, and lidocaine showed promising but heterogeneous benefits across current literature. We found no credibly better analgesic technique and no differences between general and regional anaesthesia. Additional low bias RCTs across orthopaedic surgical procedures are warranted to validate these findings and to identify patients who are most likely to benefit.

Author contributions

DM, QC, and XC contributed to the conceptualization and design of the study. XC and KS designed and wrote the entire manuscript. XC, HX, and KS collected the references for the manuscript and prepared the figures. DM, QC and XC completed the subsequent revisions of the manuscript. All authors contributed to the manuscript revision and approved the final version of the manuscript.

Data statement

All data generated or analysed during this study are included in this published article (and its supplementary information files). Raw data are available from the corresponding author on reasonable request.

Declaration of generative AI and AI-assisted technologies in the writing process

The authors declare that they didn't use any generative AI and AI-assisted technologies in the writing process.

Declaration of competing interest

The authors declare that they have no conflicts of interest.

Appendix A. Supplementary data

The following is the Supplementary data to this article: