This retrospective cohort study analyzed data from an extensive surgical database containing 376,933 patients who underwent non-cardiac surgery between January 2008 and August 2019. After applying systematic exclusions, 223,415 patients remained eligible for screening. Perioperative ischemic stroke (PIS) was identified in 525 patients based on clinical presentation, imaging findings, temporal association with surgery, and neurologist confirmation. Patients younger than 65 years were excluded, leaving 178 elderly patients with PIS for the final analysis. The screening process is illustrated in Figure 1.

The study was approved by the Ethics Committee of the Chinese PLA General Hospital (S2025-123-01) with a waiver of informed consent and was conducted by STROBE guidelines.

The primary exposure was the type of intraoperative opioid, fentanyl or sufentanil, administered during general anesthesia. Opioid selection was determined by attending anesthesiologist preference and drug availability, without a standardized institutional protocol. The primary outcome was postoperative delirium (POD) within 7 days of surgery.

PIS was defined as any neurological impairment affecting movement, sensation, or cognition caused by a localized blockage of blood supply in the brain, spinal cord, or retina within 30 days post-surgery (17). Stroke diagnoses were confirmed by at least one ICD-9-CM or ICD-10-CM code (Supplementary Table S1) documented in discharge records.

Secondary outcomes included overall survival, measured from the date of surgery to the date of death from any cause during the follow-up period.

POD was diagnosed by a neurologist based on descriptors such as altered mental status, confusion, disorientation, agitation, inappropriate behavior, inattention, hallucinations, and combative behavior (18). Diagnoses were supported by documentation in the anesthesiology department’s database, as described previously (18)^.

Medical records were also reviewed for administration of quetiapine, olanzapine, haloperidol, or risperidone postoperatively. The diagnostic criteria followed the DSM-IV guidelines (19). Patients who exhibited delirium symptoms or received these medications in their preoperative records were excluded.

Preoperative variables included age, sex, and body mass index (BMI), and comorbidities such as hypertension, diabetes, coronary heart disease, previous ischemic stroke, atrial fibrillation, peripheral vascular disease, and renal dysfunction.

Surgical variables included American Society of Anesthesiologists (ASA) classification, intraoperative opioid type and dose, morphine and remifentanil dosage, estimated blood loss, transfusion requirements, and intraoperative fluid administration (colloid and crystalloid volumes). Use of dexmedetomidine and non-steroidal anti-inflammatory drugs (NSAIDs) during the perioperative period was also recorded.

Laboratory data included hemoglobin, total bilirubin, and glucose levels from the most recent test conducted within 3 days before surgery. Inflammatory markers, including neutrophil-to-lymphocyte ratio (NLR) and platelet-to-lymphocyte ratio (PLR), were calculated from complete blood counts. BMI was computed from measured height and weight.

POD assessments were performed by trained healthcare professionals using standardized screening tools for the first five postoperative days. Length of hospital stay was calculated from the date of surgery to discharge.

Continuous variables with normal distribution were expressed as mean ± standard deviation (SD) and compared using Student’s t-test. Non-normally distributed variables were reported as median ± interquartile range (IQR) and compared using the Mann–Whitney U test. Categorical variables were presented as counts (percentages) and compared using a chi-squared test or Fisher’s exact test, as appropriate.

Kaplan–Meier survival analysis with log-rank testing was used to compare overall survival between patients with and without POD.

The association between intraoperative opioid type and POD risk was evaluated using univariate and multivariate logistic regression models. Opioid type (fentanyl vs. sufentanil) was treated as a binary exposure variable. Covariates were selected based on clinical relevance and prior literature identifying established risk factors for POD. Model 2 (patient-related confounders) included: sex, age, BMI, ASA physical status, hypertension, diabetes, coronary heart disease, previous ischemic stroke, atrial fibrillation, peripheral vascular diseases, renal dysfunction, hemoglobin, total bilirubin, NLR, and PLR. Model 3 (surgery-related confounders) included: blood loss, colloid, crystalloid, NSAIDs, blood transfusion, morphine, total remifentanil dosage, glucose, and dexmedetomidine. Model 4 (fully adjusted) included all of the above confounders. Results from univariate analysis for all covariates in Model 4 are presented in Supplementary Table S3, and univariate results in the PSM cohort are provided in Supplementary Table S4.

To reduce confounding, 1:1 propensity score matching (PSM) was performed using logistic regression, incorporating all covariates described above. Covariate balance was assessed using standardized mean difference (SMD), with SMD < 0.2 considered acceptable (20). Logistic regression was then repeated in the matched cohort to confirm associations.

Results were reported as odds ratios (ORs) with 95% confidence intervals (CIs). A p < 0.05 was considered statistically significant. All analyses were performed using R software (version 4.0.3, R Foundation for Statistical Computing, Vienna, Austria) with relevant packages including tableone, MatchIt, pROC, Matching, Cobalt, and rms.

Among 178 patients with perioperative ischemic stroke, the overall incidence of postoperative delirium (POD) was 40.4% (72/178), underscoring the substantial burden in this high-risk population. During follow-up, 87 patients (52.7%) died.

Kaplan–Meier survival analysis showed no statistically significant difference in overall survival between patients with and without POD (p = 0.156, Figure 2). Median survival was 365 days (IQR: 30–365) for patients without delirium and 67 days (IQR: 18–75) for those with delirium. Although delirium was associated with a trend toward shorter survival, this did not reach statistical significance. Cox regression analysis confirmed that POD was not significantly associated with overall survival (HR: 1.068; 95% CI: 0.697–1.637; p = 0.762; Supplementary Table S2). Baseline characteristics of the study population are presented in Table 1.

Seventy-two patients developed POD. In unadjusted logistic regression, fentanyl use was significantly associated with a lower risk of POD compared to sufentanil (OR: 0.042; 95% CI: 0.014–0.103; p = 0.005; Table 2). This association of lower POD risk with fentanyl compared to sufentanil persisted after adjusting for patient-related factors (OR: 0.025; 95% CI: 0.006–0.079; p < 0.001), surgery-related confounders (OR: 0.039; 95% CI: 0.011–0.112; p < 0.001), and all predefined confounders (OR: 0.017; 95% CI: 0.003–0.070; p = 0.017; Table 2).

After propensity score matching (PSM), fentanyl remained associated with significantly lower POD risk compared to sufentanil (OR:0.005; 95% CI: 0–0.057; p < 0.001). Detailed results are provided in Supplementary Table S3.

PSM generated matched cohorts of 44 patients each in the fentanyl and sufentanil groups, with standardized mean differences < 0.2 for most covariates (Figures 3a,b). Logistic regression in the matched cohort confirmed a significantly lower POD risk with fentanyl (OR: 0.005; 95% CI: 0–0.057; p < 0.001; Table 2). The association was consistent across both unadjusted and adjusted analyses, reinforcing the robustness of the findings. Additional PSM details are presented in Supplementary Table S4.

Subgroup analyses by sex, age (≤70 vs. >70 years), and diabetes status showed that fentanyl consistently reduced POD risk across all categories: females (OR: 0.001, 95% CI: 0–0.061, p = 0.00934), males (OR: 0, 95% CI: 0–0.003, p = 0.00146), age ≤70 years (OR: 0.006, 95% CI: 0–0.177, p = 0.0168), age >70 years (OR: 0.012, 95% CI: 0.001–0.088, p < 0.001), non-diabetic patients (OR: 0.001, 95% CI: 0–0.017, p < 0.001), and diabetic patients (OR: 0.007, 95% CI: 0–0.124, p = 0.0104).

No significant interaction was observed between any subgroup variable and the lower POD risk associated with fentanyl versus sufentanil (p_-interaction > 0.05), indicating that the lower POD risk with fentanyl compared to sufentanil was consistent across patient subgroups (Figure 4).

This study has several notable strengths. It specifically targeted perioperative ischemic stroke patients—a group particularly high-risk for delirium—using a large, detailed dataset that included comprehensive patient characteristics and anesthetic management variables. This granularity enabled a thorough analysis of delirium risk factors. The observed association of lower delirium risk with fentanyl compared to sufentanil was confirmed through multiple statistical approaches, including propensity score matching, which enhances the robustness of the findings. Importantly, the study centered on opioid choice, an aspect within the anesthesiologist’s direct control.

However, there are important limitations. Delirium is a multifactorial condition, making prediction based on opioid selection alone inherently challenging. The retrospective, single-center design limits the generalizability of our findings to other healthcare settings with different patient populations, surgical practices, or anesthetic protocols. Our results may remain susceptible to unmeasured confounding despite extensive adjustment. Key factors such as anesthesia depth, postoperative pain control strategies, and detailed stroke characteristics were not fully captured but may have influenced outcomes. Moreover, while propensity score matching was employed to minimize confounding, the specific types of surgical procedures were not included as a covariate in our models; given that delirium incidence varies substantially with surgical stress and tissue trauma extent, this represents a potential source of residual confounding. Similarly, detailed information regarding anesthesia type (regional, general, combined, or multimodal approaches) was not fully transparent in the study population, and more granular inclusion and exclusion criteria related to these factors would strengthen future investigations. Critically, we did not assess long-term cognitive outcomes beyond the acute postoperative period, precluding conclusions about persistent cognitive impairment or functional recovery trajectories. Additionally, although dexmedetomidine and NSAIDs were included as covariates in our models, residual confounding from these agents cannot be entirely excluded, as their dosing, timing, and indications may have varied and potentially influenced delirium outcomes independently. Beyond these factors, other potential sources of bias warrant consideration: selection bias may have arisen from the retrospective identification of perioperative stroke cases, as milder or undiagnosed strokes could have been missed; information bias related to delirium ascertainment may exist given the reliance on clinical documentation rather than standardized prospective screening; temporal changes in clinical practice over the 11-year study period (2008–2019) could introduce era effects; indication bias cannot be excluded as opioid selection may have been influenced by unmeasured patient characteristics or physician preferences correlated with delirium risk; and survivor bias may affect our findings if patients who died early had different opioid exposures. Future multi-center prospective trials are needed to validate these results and to define optimal opioid protocols for perioperative ischemic stroke patients.