For this retrospective study, data from the Samsung Medical Center-Non Cardiac operation (SMC-NoCop) registry (KCT 0006363) were used. This cohort consisted of 203 787 consecutive adult patients (≥ 18 years of age) who underwent non-cardiac surgery between January 2011 and June 2019 at Samsung Medical Center, Seoul, Korea.

The Institutional Review Board of Samsung Medical Center approved this retrospective study (SMC 2021-06-078). The need for written informed consent from individual patients was waived because the data for this study was initially collected in de-identified form. From this registry, patients were excluded if they were aged < 60 years, if they presented with preoperative delirium, preoperative dementia, or were undergoing outpatient surgery, if they had American Society of Anesthesiologists physical status (ASA-PS) ≥ Ⅳ, if they required perioperative dialysis or continuous renal replacement therapy, or if they were undergoing neurosurgery. This manuscript was written following the Strengthening the Reporting of Observational Studies in Epidemiology reporting guideline [18].

The raw data for this registry were collected by reviewing the electronic medical charts using the Clinical Data Warehouse Darwin-C from the Samsung Medical Center. This is an electronic retrieval system that was built to allow researchers to extract and retrieve data from electronic medical records in a de-identified form. The system provides access to hospital records for over 4 million patients, including more than 1.3 billion laboratory findings and 400 million prescriptions. Mortality data outside our institution are regularly updated using the National Population Registry of the Korea National Statistical Office with a personally designated identification number.

Potential confounding variables were obtained from a preoperative evaluation sheet by investigators who were independent from this study. The baseline characteristics of patients including demographics (sex, age, body mass index, and ASA-PS), pre-existing mental or behavioral disorders (mood disorder, schizophrenia, alcohol disorder, substance abuse, sleep disorder, and personality disorder), social and medical history (alcohol, smoking, hypertension, diabetes, chronic kidney disease, dialysis, stroke, coronary artery disease, heart failure, arrhythmia, peripheral artery disease, and chronic obstructive pulmonary disease), preoperative laboratory results (creatinine, serum sodium, serum potassium, serum phosphorus, and serum chloride), and preoperative medications (β blockers, calcium channel blockers, angiotensin-converting enzyme inhibitors or angiotensin II receptor blockers, antiplatelet agents, and statins) were collected. The Charlson Comorbidity Index score was also calculated to assess the overall severity of comorbidities and general health status [19]. In addition, operative variables including risk of surgery, emergency, general anesthesia, operation duration, use of intraoperative inotropics, and intraoperative red blood cell (RBC) transfusion were collected. According to the non-cardiac surgery guidelines from the Society of Cardiology/European Society of Anesthesiology, the surgical risk was divided into three groups: low, intermediate, and high [20].

Preoperative hemoglobin level was measured in all patients as a routine laboratory blood test. For the diagnosis of preoperative anemia, the hemoglobin level obtained most recently prior to surgery was referenced and anemia was determined according to the World Health Organization (WHO) definition: hemoglobin < 13 g/dl for males and < 12 g/dl for females [21]. The patients with anemia were further classified as mild (hemoglobin > 11 g/dl and < 12 g/dl for females and > 11 g/dl and < 13 g/dl for males) or moderate-to-severe (hemoglobin < 11 g/dl for both males and females) anemia [22].

The primary endpoint was the risk of POD within 7 days after surgery. The retrospective diagnosis of POD was determined based on the medical chart [23] using the following methods to maximize sensitivity: (1) International Classification of Diseases, Ninth Revision and Tenth Revision (ICD-9/10) codes for delirium newly diagnosed after surgery; (2) search the entire medical chart for keywords or derivatives to diagnose POD; and (3) identification of medications used to treat delirium after surgery. The examples of ICD-9/10 codes and terms used to screen for POD are presented in the Supplementary Table 1. Next, our independent researchers reviewed the medical charts (progress notes, nursing chart, consciousness evaluation sheet, and psychiatry consultation notes) of suspected POD cases, and only those that were definitively diagnosed with delirium were classified as POD.

For secondary outcomes, mortality during one- and three-year follow-up after surgery was investigated. The survival time was calculated from the day of surgery to the date of death, or the day when the patient was censored (the last visit to the hospital). The mortality data were finally extracted on September 30, 2023.

Continuous variables were reported as either mean ± standard deviation or median with interquartile range as appropriate. To determine normality, we visually assessed the distribution of data using histograms. Categorical variables were presented as number with percentage.

To reduce the effect of confounding resulting from nonrandomization of this retrospective study, the inverse probability of treatment weighting (IPTW) method was used in the analysis. The probability of treatment was estimated using a logistic regression model that included preoperative anemia as the outcome variable and patient demographic and clinical characteristics presented in Table 1 as the predictors (except for preoperative hemoglobin). For each patient, the weight was calculated as the inverse of the probability of exposure to preoperative anemia. Furthermore, we adjusted the weights by trimming those below the first percentile and above the 99th percentile to mitigate the impact of extreme values [24,25]. To assess the balance between the two groups, the absolute standardized difference (ASD) was evaluated before and after weighting. The two groups were considered to be well balanced if the ASD was less than 0.10.

The variables with ASD ≥ 0.10 after IPTW were considered insufficiently balanced and were further included as covariates in the outcome analysis to account for residual confounding. Specifically, we incorporated the variable into the weighted logistic regression model for POD and the weighted Cox proportional hazards regression model for mortality as additional covariates. The proportional hazard assumption was confirmed using Schoenfeld residuals. The results were reported as odds ratio (OR) or hazard ratio (HR) with 95% CI. To further compare the outcomes between the non-anemic group and the mild or moderate-to-severe anemic group, the same statistical methods were used as described above.

In addition, we investigated whether any variations in the effects of preoperative anemia on POD existed across the different subgroups, including by age (< 75 years or ≥ 75 years), sex (female or male), Charlson Comorbidity Index (< 2 or ≥ 2), mental or behavioral disorder (no or yes), type of anesthesia (no general anesthesia or general anesthesia), risk of surgery (low to moderate or high), and intraoperative RBC transfusion (no transfusion or transfusion), by using IPTW cohort. To estimate the stratum-specified HRs, we used a single logistic regression model that included treatment-covariate interactions.

For sensitivity analysis, we restricted the inclusion criteria to a subgroup of patients who were admitted to ICU after surgery, considering the detection bias associated with the diagnosis of POD. We used POD within 24 hours after surgery as an additional endpoint to investigate the impact of preoperative anemia on postoperative early delirium. All statistical analyses were performed using R 4.2.0 (http://www.R-project.org/) and P value < 0.05 was considered statistically significant. We utilized ‘ipw’ R package for estimating inverse probability weights.

After excluding 141 187 patients who met the exclusion criteria, the remaining 62 600 patients were included for final analysis (Fig. 1). Based on preoperative hemoglobin level, 45 199 (72.2%) patients were included in the normal group and 17 401 (27.8%) in the anemia group. Overall, the mean preoperative hemoglobin level was 13.9 ± 1.1 g/dl in the normal group and 11.1 ± 1.2 g/dl in the anemia group. The baseline characteristics of patients and perioperative data are summarized in Table 1. Patients in the anemia group were older and tended to have a higher incidence of comorbidities and abnormal laboratory results than subjects in the normal group. In addition, patients in the anemia group underwent high-risk surgery, emergency operations, and intraoperative transfusions more frequently than subjects in the normal group. After IPTW adjustment, satisfactory balance (ASD < 0.10) was attained between the two groups for all variables.

The overall incidence of POD within 7 days after surgery was 3.9% (2447/62 600). The incidence of POD was higher in the anemia group (7.2%) than in the normal group (2.6%). In adjusted analysis, significant association was found between preoperative anemia and POD, with an adjusted OR of 1.42 (95% CI [1.30−1.55], P < 0.001). One-year and three-year mortalities were also significantly higher in the anemia group (adjusted HR: 2.68, 95% CI [2.45−2.93], P < 0.001 for one-year mortality; and adjusted HR: 2.19, 95% CI [2.07−2.33], P < 0.001 for three-year mortality; Table 2) (Supplementary Fig. 1). Long-term follow-up data, including the number of censored events at selected time points, are presented in Supplementary Table 2.

The patients with anemia were further divided into mild anemia and moderate-to-severe anemia groups; there were 10 664 (17.0%) patients in the mild anemia group (Supplementary Table 3) and 6737 (10.8%) patients in the moderate-to-severe anemia group (Supplementary Table 4). The incidence of POD was 6.1% (652/10 664) in the mild anemia group and 9.0% (604/6737) in the moderate-to-severe anemia group. Pairwise comparison was conducted with the normal group using IPTW and the risk of POD increased with the severity of anemia (adjusted OR: 1.32, 95% CI [1.18−1.47], P < 0.001 in the mild anemia group; and adjusted OR: 1.70, 95% CI [1.50−1.93], P < 0.001 in the moderate-to-severe anemia group) (Table 3). A similar significant association was observed for both one-year mortality (adjusted HR: 2.24, 95% CI [2.02−2.49], P < 0.001 in the mild anemia group; and adjusted HR: 4.13, 95% CI [3.67−4.65], P < 0.001 in the moderate-to-severe anemia group) and three-year mortality (adjusted HR: 1.96, 95% CI [1.83−2.10], P < 0.001 in the mild anemia group; and adjusted HR: 3.03, 95% CI [2.79−3.30], P < 0.001 in the moderate-to-severe anemia group).

Subgroup analysis revealed that the observed association between preoperative anemia and POD significantly interacted with variables, namely sex (P for interaction = 0.048) and Charlson Comorbidity Index (P for interaction = 0.035). The significant association between preoperative anemia and POD was more evident in males (OR: 1.52, 95% CI [1.37−1.69], P < 0.001) and was only applicable to patients who had a low Charlson Comorbidity Index (OR: 1.45, 95% CI [1.33−1.59], P < 0.001) (Fig. 2).

Patients who were admitted to the ICU after surgery were included in this subgroup, comprising a total of 8 007 patients (Supplementary Table 5). Similar to the result of primary analysis, we found a significant association between preoperative anemia and early delirium in patients admitted to the ICU after surgery using IPTW (adjusted OR: 1.24, 95% CI [1.11−1.38], P < 0.001) (Supplementary Table 6).