Long-term risk of depression after sevoflurane versus propofol anesthesia: a global propensity score–matched cohort study

Abstract

Preclinical studies indicate that volatile anesthetics such as sevoflurane induce neuroinflammation and oxidative stress—mechanisms implicated in the pathophysiology of depression—whereas propofol exhibits neuroprotective properties. However, whether these pharmacological differences translate into distinct long-term psychiatric outcomes in humans remains unclear. We conducted a multinational propensity score–matched cohort study using the TriNetX Global Federated Network (2005–2024) to determine whether sevoflurane anesthesia is associated with a higher risk of long-term depression compared with propofol. To isolate the effect of anesthetic choice, we applied a new-user design with a one-year latency washout, excluding patients with preexisting depression or antidepressant use. The primary outcome was new-onset depression diagnosed more than one year after anesthesia. The matched cohort included 37,936 patients balanced across more than 100 covariates. Sevoflurane exposure was associated with a higher risk of new-onset depression compared with propofol (adjusted hazard ratio [aHR], 1.51; 95% CI, 1.41–1.61). The association remained consistent across severity strata (aHR, 1.51 for moderate-to-severe depression) and surgical subgroups. Notably, we observed a graded dose-response relationship, where repeated sevoflurane exposures conferred progressively higher risk (aHR, 1.75 for ≥2 exposures), consistent with a biologic gradient. In this large multinational cohort, sevoflurane use was associated with a higher long-term risk of incident depression compared with propofol. These findings align with preclinical evidence of anesthetic-related neuroinflammatory mechanisms and indicate that anesthetic selection may have implications for long-term neuropsychiatric risk. Further mechanistic and prospective studies are warranted.

Introduction

Depression is a leading contributor to global disability and premature mortality, affecting more than 300 million people worldwide and imposing a profound burden on individuals, families, and health systems [1]. The World Health Organization estimates that depression accounts for more than US$1 trillion annually in lost productivity and healthcare expenditures [2–4]. Beyond its socioeconomic toll, major depressive disorder (MDD) is associated with increased risks of cardiovascular disease, metabolic dysfunction, suicide, and all-cause mortality [5–7]. The rising prevalence of depression in aging, post-pandemic societies underscores an urgent need to identify modifiable medical and iatrogenic contributors to this highly disabling condition [8, 9]. While genetic and environmental determinants are well studied, potential iatrogenic triggers within routine medical care—such as exposure to general anesthesia—remain largely overlooked. Importantly, the majority of clinical investigations in this field focus on outcomes occurring within days to weeks after surgery, and even studies extending follow-up typically do not exceed 3 to 6 months [10].

Although postoperative neurocognitive disorders—including delirium, cognitive decline, and affective disturbances—are now recognized as significant long-term complications of anesthesia and surgery, perioperative anesthetic safety has historically been evaluated primarily through a short-term recovery lens, with emphasis on outcomes such as postoperative delirium (POD) and postoperative cognitive dysfunction (POCD) [11–16]. Systematic research examining de novo mood disorders emerging beyond one year after anesthesia remains strikingly limited [17, 18].

Evidence concerning long-term psychiatric sequelae, particularly depression developing beyond one year after surgery, remains remarkably scarce. This temporal distinction is critical: whereas POD and early POCD likely reflect transient inflammatory or metabolic perturbations, late-onset depression may represent enduring neurobiologic alterations or cumulative neurotoxicity. Anesthetic agents themselves may exert enduring neurobiologic effects through mechanisms involving neurotransmission, inflammation, oxidative stress, and synaptic plasticity [11, 19, 20]. Preclinical studies have shown that volatile anesthetics such as sevoflurane induce microglial activation, IL-6 and TNF-α elevation, and oxidative stress–mediated neuronal injury, pathways also implicated in the pathophysiology of depression [20–23]. In contrast, propofol demonstrates antioxidant, anti-inflammatory, and neuroprotective properties, mitigating NF-κB–mediated neuroinflammation [24–28]. Notably, although certain preclinical models have reported transient antidepressant-like effects of sevoflurane under acute stress conditions [29, 30], whether these short-term pharmacologic observations translate into durable mental health benefits—or risks—in humans remains uncertain, underscoring a substantial translational gap between experimental findings and long-term clinical outcomes [11]. These divergent mechanistic profiles raise a biologically plausible question: could anesthetic choice influence the long-term risk of postoperative depression?

Despite decades of investigation into perioperative neurocognitive outcomes, whether specific anesthetic agents confer differential risk for long-term affective disorders remains unknown. Prior studies have been limited by short follow-up durations, small sample sizes, and inadequate control of key methodological biases such as confounding by indication and reverse causation, in which underlying conditions or subclinical mood disturbances may simultaneously affect anesthetic selection and subsequent psychiatric outcomes [11, 12, 14, 19, 31]. Moreover, the complex interplay of surgical stress, systemic inflammation, and postoperative pain further obscures causal attribution to anesthetic exposure alone [19, 32].

To address this unresolved question, robust, population-level, real-world evidence is needed. Federated electronic health record (EHR) platforms such as the TriNetX Global Network enable large-scale, longitudinal assessment across diverse populations, allowing temporal separation between anesthetic exposure and psychiatric outcomes while ensuring external validity. By implementing a prespecified one-year landmark (delayed-entry) design [33], our study explicitly distinguishes acute perioperative neurocognitive phenomena from incident depressive disorders emerging beyond the first postoperative year. Therefore, we conducted a multinational, propensity score–matched (PSM) cohort study using the TriNetX Global Federated Network to determine whether exposure to sevoflurane, compared with propofol, is associated with an increased long-term risk of developing depression.

Methods

Data source

This retrospective, propensity score–matched cohort study used the TriNetX Global Federated Network, a real-world electronic health record (EHR) platform integrating longitudinal, deidentified data from over 200 healthcare organizations across North America, Europe, and Asia (>250 million patients). The database includes demographics, diagnoses, procedures, prescriptions, laboratory results, and mortality, updated monthly under a standardized, privacy-preserving model. All data are deidentified in accordance with HIPAA and GDPR regulations, and institutional ISO/IEC 27001 certification ensures data security. As only aggregated, deidentified data were analyzed, institutional review board approval and informed consent were not required.

Study population

We included adults (≥18 years) who underwent surgery under general anesthesia between January 1, 2005, and December 31, 2024, within the TriNetX Global Federated Network. To ensure a depression-free baseline and prevent reverse causation, patients with depressive disorders (ICD-10 F32–F33) or antidepressant use (ATC N06A) before or within one year after anesthesia were excluded. We implemented a prespecified one-year lag-time landmark (delayed-entry) design. Only patients who survived this one-year post-anesthesia period without a diagnosis of depression or antidepressant exposure were eligible to enter follow-up. By defining eligibility at a fixed landmark exactly 365 days after anesthetic exposure, this approach explicitly addresses immortal time bias by ensuring that the latency period is not misclassified as time at risk. This landmarking approach establishes a fixed eligibility time point 365 days after exposure, thereby mitigating immortal time bias and eliminating reverse causation from transient perioperative mood disturbances, while ensuring a clear temporal sequence between anesthetic exposure and incident depressive outcomes. Because conditioning on survival to the one-year landmark may introduce selection bias, we performed mortality-censored sensitivity analyses across survivor cohorts to confirm that hazard estimates remained directionally consistent. The cohort comprised patients receiving either sevoflurane or propofol exclusively during the index operation. The main analysis allowed subsequent crossover, while a no-crossover sensitivity analysis confirmed robustness. Exposure status was verified from operative and pharmacy records within TriNetX.

Exposure and outcome definition

The primary exposure was the anesthetic agent used during the index surgery, defined as sevoflurane for volatile inhalational anesthesia and propofol for total intravenous anesthesia. Each anesthetic administration was treated as a single index event. To examine potential cumulative effects, a stratified analysis compared patients who received sevoflurane once with those who underwent two or more sevoflurane exposures within five years after the index procedure. The primary outcome was the first recorded clinical diagnosis of depression (ICD-10 codes F32.x or F33.x) occurring at least one year after anesthesia. Depressive episodes were further classified as mild or moderate-to-severe based on diagnostic coding and concurrent antidepressant prescriptions within the TriNetX dataset. Patients who developed depression or died within the first postoperative year were excluded under the predefined landmark (delayed-entry) eligibility criteria, consistent with a delayed-entry survival framework designed to prevent immortal time bias and isolate incident depressive events from acute perioperative neuroinflammatory or stress-related mood disturbances. Follow-up commenced one year after anesthesia and continued until the earliest occurrence of depression, death, loss to follow-up, or December 31, 2024. All-cause mortality was analyzed as a secondary endpoint and used for mortality-censored sensitivity analyses. All definitions conformed to standardized TriNetX data terminology to maintain analytic consistency and reproducibility across cohorts.

Index date and follow-up

In accordance with the prespecified landmark (delayed-entry) design, the index date (time zero) was set exactly 365 days after anesthetic exposure. Patients were required to survive this latency period without a depression diagnosis or antidepressant use before entering follow-up. Observation then continued until the first occurrence of depression, death, loss to follow-up, or December 31, 2024. This delayed-entry survival structure strengthens temporal ordering between exposure and outcome, mitigates reverse causation, and ensures that early postoperative psychiatric manifestations are not misclassified as long-term incident events. By redefining time zero at the one-year landmark, this delayed-entry survival structure prevents misclassification of early postoperative events as long-term outcomes and reduces bias from transient perioperative mood disturbances and conditioning on short-term survival.

Covariates

Baseline covariates included demographics (age, sex, race), lifestyle factors (smoking, alcohol use), and comorbidities such as hypertension, diabetes, dyslipidemia, coronary artery disease, chronic kidney or liver disease, thyroid disorders, chronic pain, cancer, anxiety, and sleep disorders (Table 1). Concomitant medications (opioids, corticosteroids, benzodiazepines, antipsychotics, and mood stabilizers) and laboratory markers—including hemoglobin, fasting glucose, eGFR, liver enzymes, C-reactive protein, thyroid-stimulating hormone, HbA1c, and calcidiol—were also included. Importantly, surgical category (e.g., cardiovascular, musculoskeletal, respiratory, head-and-neck, dermatologic, and other major procedure groups) was directly incorporated into the propensity score model as a baseline covariate to mitigate confounding by procedural indication. Missing data were managed through TriNetX’s automated data harmonization, with no manual imputation.

Table 1.

Baseline characteristics of patients receiving sevoflurane or propofol anesthesia before and after propensity-score matching.

Characteristic	Before: Sevoflurane (n = 18,975)	Before: Propofol (n = 1,054,534)	ASMD (Before)	After: Sevoflurane (n = 18,968)	After: Propofol (n = 18,968)	ASMD (After)
Demographic Characteristics
Age at index, mean ± SD (y)	40 ± 4.36	40 ± 4.66	0.096	40 ± 4.37	40 ± 4.46	0.003
Male, No. (%)	8162 (43.0)	451,932 (42.9)	0.003	8160 (43.0)	8013 (42.3)	0.016
Female, No. (%)	10,804 (56.9)	568,064 (53.9)	0.062	10,799 (56.9)	10,943 (57.7)	0.015
White, No. (%)	11,398 (60.1)	633,553 (60.1)	0.000	11,397 (60.1)	11,238 (59.3)	0.017
Black or African American, No. (%)	4234 (22.3)	166,780 (15.8)	0.166	4228 (22.3)	4587 (24.2)	0.045
Asian, No. (%)	609 (3.2)	59,408 (5.6)	0.118	609 (3.2)	606 (3.2)	0.001
Other or unknown race, No. (%)	2600 (13.7)	177,664 (16.9)	0.089	2600 (13.7)	2409 (12.7)	0.024
Comorbid Medical Conditions
Type 2 diabetes mellitus, No. (%)	1142 (6.0)	51,861 (4.9)	0.048	1141 (6.0)	1042 (5.5)	0.022
Hypertensive diseases, No. (%)	3250 (17.1)	136,052 (12.9)	0.119	3245 (17.1)	3032 (16.0)	0.030
Ischemic heart disease, No. (%)	419 (2.2)	15,154 (1.4)	0.058	418 (2.2)	386 (2.0)	0.012
Pulmonary heart disease / pulmonary circulation, No. (%)	258 (1.4)	7973 (0.8)	0.059	257 (1.4)	233 (1.2)	0.011
Cerebrovascular disease, No. (%)	321 (1.7)	10,294 (1.0)	0.062	321 (1.7)	303 (1.6)	0.008
Chronic lower respiratory disease, No. (%)	1683 (8.9)	60,814 (5.8)	0.119	1681 (8.9)	1545 (8.2)	0.026
Acute respiratory failure, No. (%)	299 (1.6)	8356 (0.8)	0.073	299 (1.6)	246 (1.3)	0.024
Chronic kidney disease (CKD), No. (%)	585 (3.1)	17,462 (1.7)	0.094	583 (3.1)	532 (2.8)	0.016
Overweight or obesity, No. (%)	3518 (18.5)	125,142 (11.9)	0.187	3511 (18.5)	3290 (17.4)	0.030
Tobacco use, No. (%)	344 (1.8)	15,043 (1.4)	0.031	344 (1.8)	349 (1.8)	0.002
Alcohol-related disorders, No. (%)	509 (2.7)	17,470 (1.7)	0.071	509 (2.7)	500 (2.6)	0.003
Liver disease, No. (%)	878 (4.6)	40,846 (3.9)	0.037	878 (4.6)	736 (3.9)	0.037
Disorders of lipoprotein metabolism, No. (%)	1836 (9.7)	96,889 (9.2)	0.017	1834 (9.7)	1658 (8.7)	0.032
Cancer (any), No. (%)	115 (0.6)	5052 (0.5)	0.017	115 (0.6)	124 (0.7)	0.006
Thyroid disorders, No. (%)	1427 (7.5)	60,977 (5.8)	0.070	1427 (7.5)	1348 (7.1)	0.016
Noninfective enteritis and colitis, No. (%)	603 (3.2)	31,492 (3.0)	0.011	603 (3.2)	461 (2.4)	0.045
Streptococcal sepsis, No. (%)	35 (0.18)	729 (0.07)	0.032	35 (0.19)	31 (0.16)	0.005
Other sepsis, No. (%)	501 (2.64)	12,402 (1.18)	0.107	501 (2.64)	448 (2.36)	0.018
Viral hepatitis, No. (%)	142 (0.75)	7316 (0.69)	0.007	142 (0.75)	118 (0.62)	0.015
DIC (disseminated intravascular coagulation), No. (%)	16 (0.08)	371 (0.04)	0.020	16 (0.08)	18 (0.10)	0.004
Chronic pain, No. (%)	2846 (15.0)	137,089 (13.0)	0.058	2845 (15.0)	2788 (14.7)	0.008
Surgical Categories
Digestive tract procedures, No. (%)	3795 (20.0)	158,180 (15.0)	0.132	3794 (20.0)	3718 (19.6)	0.010
Musculoskeletal interventions, No. (%)	2846 (15.0)	126,544 (12.0)	0.088	2845 (15.0)	2750 (14.5)	0.014
Skin procedures, No. (%)	1898 (10.0)	147,635 (14.0)	0.124	1897 (10.0)	1935 (10.2)	0.007
Kidney, ureter, and bladder, No. (%)	1518 (8.0)	63,272 (6.0)	0.079	1517 (8.0)	1479 (7.8)	0.007
Cardiovascular operations, No. (%)	1328 (7.0)	42,181 (4.0)	0.133	1328 (7.0)	1290 (6.8)	0.008
Respiratory surgeries, No. (%)	1139 (6.0)	31,636 (3.0)	0.145	1138 (6.0)	1092 (5.8)	0.008
Neurosurgical interventions, No. (%)	759 (4.0)	21,091 (2.0)	0.116	759 (4.0)	721 (3.8)	0.010
Breast procedures, No. (%)	569 (3.0)	31,636 (3.0)	0.000	569 ((3.0)	550 (2.9)	0.006
Head and neck procedures, No. (%)	380 (2.0)	10,545 (1.0)	0.082	379 (2.0)	360 (1.9)	0.007
Ophthalmic procedures, No. (%)	190 (1.0)	52,727 (5.0)	0.237	190 (1.0)	208 (1.1)	0.010
Other surgical procedures, No. (%)	4747 (25.0)	369,088 (35.0)	0.217	4552 (24.0)	4865 (25.7)	0.039
Psychiatric and Neurological Disorders
Anxiety / stress-related disorders, No. (%)	1227 (6.5)	51,306 (4.9)	0.069	1225 (6.5)	1128 (6.0)	0.021
Bipolar disorder, No. (%)	133 (0.7)	5500 (0.5)	0.023	133 (0.7)	121 (0.6)	0.008
Schizophrenia, No. (%)	44 (0.2)	1520 (0.1)	0.020	44 (0.2)	41 (0.2)	0.003
Sleep disorders, No. (%)	1373 (7.2)	54,316 (5.1)	0.087	1370 (7.2)	1256 (6.6)	0.024
Rheumatoid arthritis / SLE, No. (%)	108 (0.6)	4353 (0.4)	0.033	108 (0.6)	110 (0.6)	0.003
Cardiovascular Procedures
Coronary therapeutic services and procedures, No. (%)	19 (0.10)	653 (0.06)	0.013	19 (0.10)	20 (0.11)	0.002
Coronary artery bypass (arterial grafts), No. (%)	10 (0.05)	24 (0.00)	0.030	10 (0.05)	10 (0.05)	0.000
Medication Use (Pre-/Perioperative)
Opioids, No. (%)	9660 (50.9)	245,839 (23.3)	0.596	9653 (50.9)	9118 (48.1)	0.056
Antiepileptics, No. (%)	3063 (16.1)	63,037 (6.0)	0.328	3056 (16.1)	3154 (16.6)	0.014
Antipsychotics, No. (%)	2137 (11.3)	29,639 (2.8)	0.335	2134 (11.3)	1903 (10.0)	0.040
Anxiolytics, No. (%)	2405 (12.7)	69,268 (6.6)	0.208	2404 (12.7)	2302 (12.1)	0.016
Hypnotics and sedatives, No. (%)	5945 (31.3)	84,817 (8.0)	0.613	5938 (31.3)	5284 (27.9)	0.076
Centrally acting sympathomimetics, No. (%)	295 (1.6)	11,966 (1.1)	0.037	295 (1.6)	338 (1.8)	0.018
Glucocorticoids, No. (%)	8128 (42.8)	215,047 (20.4)	0.497	8121 (42.8)	7609 (40.1)	0.055
Anthropometric Measures
Body mass index, mean ± SD (kg/m²)	29.82 ± 7.40	30.29 ± 7.61	0.062	29.82 ± 7.40	29.90 ± 7.57	0.011
BMI ≤ 25 kg/m², No. (%)	2932 (15.5)	205,914 (19.5)	0.107	2932 (15.5)	2548 (13.4)	0.058
25 ≤ BMI < 30 kg/m², No. (%)	3514 (18.5)	256,149 (24.3)	0.141	3514 (18.5)	2997 (15.8)	0.072
BMI ≥ 30 kg/m², No. (%)	4117 (21.7)	331,638 (31.5)	0.222	4117 (21.7)	3507 (18.5)	0.080
Albumin [Mass/volume] in Serum, Plasma or Blood, mean ± SD (g/dL)	4.07 ± 0.57	4.14 ± 0.53	0.129	4.07 ± 0.57	4.11 ± 0.55	0.067
eGFR, mean ± SD (mL/min/1.73 m²)	93.68 ± 28.63	92.09 ± 26.12	0.058	93.68 ± 28.64	93.82 ± 28.32	0.005
Laboratory Measures: Thyroid Function
Thyrotropin, mean ± SD (mIU/L)	2.55 ± 14.95	2.50 ± 18.31	0.003	2.55 ± 14.96	2.36 ± 13.57	0.013
Free thyroxine, mean ± SD (ng/dL)	1.13 ± 0.40	1.14 ± 0.37	0.017	1.13 ± 0.40	1.12 ± 0.36	0.022
Laboratory Measures: Hematology and Inflammation
Hemoglobin, mean ± SD (g/dL)	13.13 ± 2.14	13.35 ± 2.04	0.106	13.13 ± 2.14	13.05 ± 2.12	0.038
C-reactive protein, mean ± SD (mg/L)	30.79 ± 69.78	31.87 ± 61.71	0.016	30.82 ± 69.80	31.37 ± 63.77	0.008
Laboratory Measures: Metabolic
Hemoglobin A1c, mean ± SD (%)	5.67 ± 1.80	5.92 ± 1.61	0.149	5.67 ± 1.80	5.71 ± 1.69	0.026
Calcidiol, mean ± SD (ng/mL)	27.80 ± 15.52	29.39 ± 15.09	0.104	27.85 ± 15.52	26.43 ± 15.03	0.092

ASMD absolute standardized mean difference, BMI body mass index, eGFR estimated glomerular filtration rate, RA rheumatoid arthritis, SLE systemic lupus erythematosus.

After 1:1 propensity-score matching, all baseline covariates achieved adequate balance (ASMD < 0.10). Values are presented as No. (%) unless otherwise indicated.

Propensity score matching

To minimize confounding by indication, propensity scores estimating the probability of receiving sevoflurane versus propofol were calculated using multivariable logistic regression including all baseline covariates. Patients were matched 1:1 by nearest neighbor without replacement, using a caliper of 0.1 SD of the logit of the score. Covariate balance was evaluated with standardized mean differences (SMDs), with values < 0.1 indicating adequate balance. Matching was applied to the entire cohort, and all subsequent analyses were conducted using the balanced sample.

Statistical analysis

Incidence rates were expressed as new depression diagnoses per 10,000 person-years. Cox proportional-hazards models estimated hazard ratios (HRs) with 95% confidence intervals (CIs), and proportionality assumptions were verified using TriNetX diagnostics. All time-to-event analyses were conducted using a delayed-entry Cox model consistent with the lag-time landmark framework, with risk sets defined exclusively among individuals eligible at the 365-day landmark. Sensitivity analyses included a no-crossover model excluding patients later exposed to the alternate anesthetic, a repeated-exposure model assessing dose–response effects, and mortality-censored analyses restricted to survivors two to five years post-exposure. As TriNetX does not support competing risk regression, deaths were treated as censored events. All analyses were conducted within the TriNetX Live environment, which ensures deduplication and data integrity. Two-tailed P values < 0.05 were considered statistically significant.

Sensitivity and robustness analyses

Multiple sensitivity analyses were performed to assess the robustness of the observed association between sevoflurane and depression. These included no-crossover, repeated-exposure, and mortality-censored models, as well as outcome-control analyses using anxiety and postoperative nausea or vomiting as positive controls and appendicitis, irritable bowel syndrome, and the common cold as negative controls to evaluate residual confounding. E-values were calculated to quantify the minimum strength of an unmeasured confounder required to explain the observed hazard ratios.

Results

Baseline characteristics

Before matching, patients receiving sevoflurane anesthesia were generally older, had a higher proportion of Black individuals, and showed greater prevalence of hypertension, chronic respiratory or kidney disease, thyroid disorders, and anxiety or sleep disturbances. They were also more likely to use perioperative opioids, corticosteroids, sedative–hypnotics, or antipsychotics. In contrast, the propofol group had slightly higher mean BMI (≥30 kg/m²), higher albumin and HbA1c levels, and marginally lower eGFR, reflecting minor metabolic differences. After 1:1 propensity score matching (n = 18 968 per group), all demographic, clinical, pharmacologic, laboratory, and surgical category variables were well balanced (SMDs < 0.10), confirming exchangeability between cohorts and minimizing confounding by indication.

Primary outcome

During follow-up beginning one year after anesthesia, new-onset depression occurred in 2201 patients (11.6%) in the sevoflurane group and 1346 (7.1%) in the propofol group, corresponding to incidence rates of 323.0 and 221.5 per 10,000 person-years, respectively (Table 2). Sevoflurane exposure was associated with a higher risk of depression (adjusted hazard ratio [aHR], 1.51; 95% CI, 1.41–1.61; absolute risk difference, 4.51%). The Kaplan–Meier curves (Fig. 1) diverged soon after the one-year landmark and continued to separate throughout follow-up, indicating a sustained temporal association. The one-year latency design ensured that anesthetic exposure preceded outcome onset, reinforcing temporality. The calculated E-value for the point estimate (aHR 1.51) was 2.38, whereas the E-value for the lower confidence limit (1.41) was 2.17, indicating the minimum strength of association that an unmeasured confounder would require with both exposure and outcome to shift either the point estimate or the lower confidence bound to the null.

Table 2.

Association between type of anesthetic agent and postoperative depression, comparing sevoflurane with propofol (Reference).

Analysis Metric	Sevoflurane (n = 18,968)	Propofol (n = 18,968)	Between-Group Effect (95% CI)	p Value / E-value
Patients with depression, n (%)	2201 (11.6%)	1346 (7.1%)	—	—
Incidence rate (per 10,000 PY)	323.0	221.5	IRR 1.46	—
Depression-free survival probability†	88.4%	92.9%	—	—
Hazard ratio (Cox)	—	—	1.51 (1.41–1.61)	E-value 2.38 (LCL 2.17)
Absolute risk difference (%)	—	—	4.51 (3.92–5.09)	z = 15.08, p < 0.001

Values are adjusted for all covariates in Table 1. † Kaplan–Meier estimates at maximum available follow-up after the 365-day landmark.

Bold values indicate prespecified key outcome values and key between-group effect estimates highlighted for emphasis.

Fig. 1.

Cumulative incidence of de novo depression in the sevoflurane and propofol cohorts, beginning at the prespecified 365-day landmark index date (delayed-entry design).

Secondary outcome

All-cause mortality was slightly higher among patients receiving sevoflurane than propofol (0.91 vs. 0.62%; adjusted hazard ratio, 1.30 [95% CI, 1.03–1.64]) (Table 3). The survival curves (Supplementary Fig. 1) diverged gradually in favor of propofol. Although mortality was not on the causal pathway, the higher death rate among sevoflurane recipients suggests that patients most susceptible to depression may have been lost to follow-up, biasing results toward the null. Therefore, the persistence of excess depression risk despite higher mortality supports the robustness of the observed association rather than suggesting survival-related bias.

Table 3.

Association between type of anesthetic agent and all-cause mortality, comparing sevoflurane with propofol (Reference).

Analysis Metric	Sevoflurane	Propofol	Between-Group Effect (95% CI)	p Value / E-value
Patients with mortality, n (%)	173 (0.91%)	117 (0.62%)	—	—
Incidence rate (per 10k person-y)	25.39	19.25	IRR 1.32	—
Survival probability	96.0%	97.1%	—	—
Log-rank χ² (df = 1)	—	—	4.73	p = 0.0296
Hazard ratio (Cox)	—	—	1.30 (1.03–1.64)	E-value 1.92 (LCL 1.21)
Absolute risk difference (%)	—	—	0.30 (0.12–0.47)	p = 0.001
Risk ratio	—	—	1.48 (1.17–1.87)	E-value 2.32 (LCL 1.62)
Odds ratio	—	—	1.48 (1.17–1.88)	E-value 2.33 (LCL 1.62)

Values are adjusted for all covariates in Table 1.

Severity-stratified analyses

The association between sevoflurane exposure and subsequent depression persisted across severity strata. For mild depression (diagnostic codes only), the adjusted hazard ratio was 1.77 (95% CI, 1.26–2.49), and for moderate-to-severe depression confirmed by concurrent antidepressant use, it was 1.51 (95% CI, 1.41–1.62) (Table 4). The cumulative-incidence curves (Supplementary Figs. 2 and 3) showed persistent separation favoring propofol, most evident in pharmacologically confirmed cases. Restricting outcomes to confirmed diagnoses enhanced construct validity and reduced misclassification bias. The consistent and stronger associations under stricter definitions, along with the widening Kaplan–Meier divergence, support a genuine pharmacologic or neurobiologic effect rather than diagnostic artifact or residual confounding.

Table 4.

Stratified analysis of depression severity following anesthesia, comparing sevoflurane with propofol (Reference).

Analysis Metric	Sevoflurane	Propofol	Between-Group Effect (95% CI)	p Value / E-Value
(A) Mild Depression
Patients with outcome, n (%)	102 (0.54%)	50 (0.26%)	—	—
Incidence rate (per 10,000 person-years)	14.97	8.23	IRR 1.82	—
Log-rank χ² (df = 1)	—	—	11.30	p = 0.0008
Hazard ratio (Cox)	—	—	1.77 (1.26–2.49)	E-value 2.94 (LCL 1.83)
Absolute risk difference (%)	—	—	0.27 (0.15–0.40)	p < 0.001
Risk ratio / Odds ratio	—	—	2.05 (1.46–2.87)	E-value 3.52 (LCL 2.28)
(B) Moderate-to-Severe Depression
Patients with outcome, n (%)	1981 (10.44%)	1202 (6.34%)	—	—
Incidence rate (per 10,000 person-years)	290.7	197.8	IRR 1.47	—
Log-rank χ² (df = 1)	—	—	128.72	p < 0.001
Hazard ratio (Cox)	—	—	1.51 (1.41–1.62)	E-value 2.39 (LCL 2.17)
Absolute risk difference (%)	—	—	4.11 (3.55–4.66)	p < 0.001
Risk ratio / Odds ratio	—	—	1.72 (1.60–1.86)	E-value 2.83 (LCL 2.58)

Values are adjusted for all covariates in Table 1.

Bold values indicate prespecified key outcome values and key between-group effect estimates highlighted for emphasis.

Subgroup analyses

Across demographic, clinical, and surgical subgroups, the association between sevoflurane exposure and long-term depression remained directionally consistent, with nearly all adjusted hazard ratios exceeding 1.0 (Fig. 2). Higher risks were observed among patients undergoing cardiovascular, musculoskeletal, and head-and-neck surgeries, suggesting that the effect reflects a drug-related rather than procedural mechanism. The persistence of elevated risk in patients without prior psychiatric disorders or psychotropic use further argues against confounding by indication. This cross-context consistency—spanning sex, age, and comorbidity strata—supports the robustness of the association across diverse clinical contexts.

Fig. 2. Subgroup Analysis of the Association Between Anesthetic Type and Depression.

(A, top panel) Clinical and surgical subgroup analyses according to demographic characteristics, comorbidities, medication use, and surgical categories. (B, bottom panel) Biochemical subgroup analyses according to baseline laboratory strata, including thyrotropin, estimated glomerular filtration rate, C-reactive protein, hemoglobin A1c, and calcidiol. Adjusted hazard ratios (aHRs) with 95% confidence intervals compare sevoflurane with propofol as the reference group. The vertical line indicates no association (aHR = 1.0), and horizontal bars indicate 95% confidence intervals.

Biochemical subgroup analyses (thyrotropin, eGFR, C-reactive protein, HbA1c, and calcidiol) showed that the increased risk persisted even among patients with normal metabolic or inflammatory profiles, indicating that sevoflurane’s effect is independent of systemic dysfunction and likely mediated through central neuroinflammatory pathways. Together, these findings reinforce the internal validity and biological plausibility of the observed association between sevoflurane exposure and subsequent depression.

Sensitivity analyses

Across all sensitivity analyses, the association between sevoflurane exposure and depression remained stable and directionally consistent (Fig. 3). In the no-crossover analysis excluding patients who later received both anesthetics, the adjusted hazard ratio (aHR 1.61; 95% CI, 1.07–2.40) was nearly identical to the primary estimate, confirming exposure specificity. The repeated-exposure model revealed a graded increase in risk among patients receiving sevoflurane once versus twice or more within five years (aHR 1.58 [1.46–1.70] and 1.75 [1.46–2.09]), satisfying Hill’s criterion of biologic gradient and supporting a dose-response mechanism. In mortality-censored analyses restricted to survivors at 2–5 years, the association persisted (aHR range 1.48–1.55), indicating that competing risk bias did not account for the findings.

Fig. 3. Association of Sevoflurane Versus Propofol With Depression in No-Crossover Cohorts, by Repeated Exposure, and in Mortality-Censored Survivors.

Footnote: Adjusted hazard ratios (aHRs) with 95% confidence intervals for depression are shown under three sensitivity analyses. In the no-crossover analysis, patients receiving both anesthetics during follow-up were excluded to isolate single-agent effects. The repeated-exposure analysis assessed dose–response among sevoflurane users with single or ≥2 exposures within 5 years. The mortality-censored analyses included only patients surviving through 2, 3, 4, or 5 years to address competing mortality. Horizontal bars indicate 95% confidence intervals.

Outcome-control and specificity checks

Outcome-control analyses (Fig. 4) showed positive associations for anxiety and postoperative nausea or vomiting—plausible anesthetic-related outcomes—but null results for unrelated events such as appendicitis, irritable bowel syndrome, or the common cold. This pattern of biologically consistent positives and null negatives reinforces the specificity and validity of the observed association.

Fig. 4. Sensitivity Analyses Comparing Sevoflurane With Propofol (Reference) Across Primary, Secondary, and Specificity Outcomes and Positive and Negative Outcome Controls.

Footnote: Adjusted hazard ratios (aHRs) with 95% confidence intervals are shown for sensitivity analyses evaluating the robustness of the association between sevoflurane and depression compared with propofol (reference). The analyses included the primary outcome (depression), secondary outcome (all-cause mortality), and specificity outcomes (mild and moderate-to-severe depression), as well as positive outcome controls (anxiety and postoperative nausea and vomiting) and negative outcome controls (common cold, burn, appendicitis, and irritable bowel syndrome). Horizontal bars represent 95% confidence intervals.

Discussion

In this large, multinational pharmacoepidemiologic study using the Global Federated Network, we provide the first population-based evidence directly comparing sevoflurane and propofol in relation to the long-term risk of new-onset depression following general anesthesia. Previous investigations into anesthetic neurotoxicity have largely focused on short-term outcomes such as postoperative delirium and cognitive dysfunction, often with small sample sizes and brief follow-up durations [11, 12, 15, 16, 34, 35]. Extending beyond these acute endpoints [11, 12, 15, 16, 34, 35], our study demonstrates that sevoflurane exposure was associated with a significantly higher incidence of depression than propofol (adjusted hazard ratio [aHR], 1.51; 95% CI, 1.41–1.61). This finding addresses a critical gap and a “clinical blind spot” in perioperative care, where delayed neuropsychiatric consequences have historically been overlooked in favor of short-term recovery metrics. By redefining time zero at the one-year landmark, our delayed-entry design establishes clear temporal ordering while mitigating immortal time bias and reverse causation. When restricted to clinically validated moderate-to-severe depression (diagnostic code plus antidepressant use), the association remained robust (aHR, 1.51; 95% CI, 1.41–1.62), reinforcing diagnostic validity and internal coherence [36]. A biologic gradient was evident, as repeated sevoflurane exposure conferred progressively greater risk. These associations were consistent across demographic and surgical subgroups and specific to psychiatric outcomes biologically related to mood regulation, rather than nonspecific clinical events

At the mechanistic level, our findings align with extensive preclinical evidence showing that sevoflurane induces neuroinflammation, oxidative stress, microglial activation, and neuronal apoptosis, all of which overlap with the biological pathways of depression [20–23, 37, 38]. Animal models have demonstrated hippocampal injury, mitochondrial dysfunction, and synaptic loss after sevoflurane exposure—hallmarks of depressive phenotypes [21–23, 38–41]. Although subanesthetic doses have transient antidepressant-like effects in acute stress models, these are short-lived and context-dependent [29, 30]. In contrast, propofol exhibits anti-inflammatory and antioxidant properties that may confer neuroprotection [24, 25, 27, 28]. Randomized trials and meta-analyses have suggested that propofol reduces postoperative delirium or delayed neurocognitive recovery [10, 32, 42], and small experimental studies in treatment-resistant depression have shown short-term mood improvement [43, 44]. Our findings extend these observations [10, 32, 42–44], suggesting that propofol’s neuroprotective effects may translate into clinically meaningful long-term mental health advantages. Collectively, by integrating PSM, a one-year latency washout, and multinational longitudinal data, this study satisfies multiple Hill’s criteria [45, 46]—including temporality, consistency, biologic gradient, and specificity—providing strong epidemiologic evidence supporting a potential causal relationship, while appropriately acknowledging the inherent limitations of observational inference.

The observed association between sevoflurane exposure and an elevated long-term risk of depression is strongly supported by biologic plausibility [20–23, 38]. Sevoflurane and propofol exhibit distinct pharmacodynamic profiles in the central nervous system that could differentially influence neuroinflammatory and neurotoxic cascades [20, 25, 28, 47–49]. Rather than representing isolated mechanisms, these findings can be conceptualized as a sequential biological cascade linking acute anesthetic exposure to sustained affective vulnerability. Experimental studies demonstrate that sevoflurane—particularly when repeated or prolonged—triggers systemic inflammatory activation [50, 51], which propagates centrally and results in sustained microglial activation [10, 52]. This neuroimmune response is accompanied by upregulation of proinflammatory cytokines, including IL-6, TNF-α, and IL-1β [10, 33]. Persistent cytokine signaling has been shown to impair mitochondrial function and increase oxidative stress through activation of JNK/p38 and inhibition of ERK pathways [11, 19], resulting in disrupted neuronal bioenergetics and redox imbalance. These upstream disturbances converge on synaptic circuitry: sevoflurane disrupts NMDA receptor regulation and glutamate transporter balance [29, 30, 53], thereby enhancing excitotoxic signaling, reducing synaptic plasticity, and promoting hippocampal neuronal apoptosis—alterations closely aligned with the neurobiology of major depressive disorder. Within this cascade framework, repeated exposure may incrementally amplify neuroinflammatory burden, oxidative injury, and synaptic destabilization, offering a mechanistic explanation for the cumulative risk pattern observed in our epidemiologic analyses. In contrast, propofol appears to modulate several nodes within this cascade. Beyond its GABAergic properties, propofol inhibits NF-κB activation and attenuates cytokine expression [54, 55], thereby dampening the initial inflammatory trigger. Its intrinsic antioxidant properties scavenge reactive oxygen species and preserve mitochondrial integrity [11, 19], limiting downstream synaptic injury. By interrupting both inflammatory and oxidative amplification loops, propofol may mitigate progression from acute neuroimmune activation to sustained mood dysregulation, providing a biologically coherent counterpoint to volatile anesthetic exposure. From an epidemiologic perspective, our findings reinforce this mechanistic directionality. The dose–response gradient observed in the repeated-exposure analyses—where patients undergoing multiple sevoflurane anesthetics exhibited progressively greater risk of depression (aHR 1.58 for single vs. 1.75 for multiple exposures)—provides empirical support for a biologic gradient, one of Hill’s causal criteria [45, 46]. This monotonic increase is compatible with the concept of cumulative neurotoxicity, whereby repeated volatile anesthetic exposure may incrementally amplify neuroinflammatory cascades, oxidative stress burden, and synaptic vulnerability over time. Experimental models demonstrating dose-dependent neuronal injury following repeated exposure parallel the graded epidemiologic risk observed in our cohort, strengthening the coherence between biological mechanism and population-level effect [11, 20, 23, 26, 56]. Taken together, the alignment between epidemiologic evidence and biologic mechanisms offers a coherent framework linking sevoflurane-induced neuroinflammation and synaptic dysregulation to the increased risk of depression, while highlighting the potential neuroprotective profile of propofol as a plausible counterpoint.

The robustness of our findings is supported by the severity-stratified and subgroup analyses (Table 4, Fig. 2). The association between sevoflurane exposure and long-term depression persisted across outcome definitions and was strongest among patients with moderate-to-severe depression defined by diagnostic codes plus antidepressant prescriptions (aHR = 1.51; 95% CI, 1.41–1.62). This graded pattern—where stricter diagnostic criteria yield clearer associations—enhances construct validity and reduces misclassification bias, suggesting that the observed risk elevation reflects a genuine pharmacologic or neurobiologic effect rather than diagnostic artifact [36]. Subgroup analyses across surgical categories demonstrated consistent risk elevations, with nearly all stratum-specific aHRs > 1.0, reinforcing internal validity and arguing against confounding by surgical indication. Because surgical categories were incorporated directly into the propensity score model and achieved adequate balance after matching, these subgroup analyses illustrate consistency across pre-balanced procedural contexts rather than correction of residual imbalance. The persistence of risk among patients without preexisting psychiatric conditions or psychotropic use further indicates that the association is not driven by selection bias or baseline mental health differences. Biochemical subgroup analyses based on renal (eGFR), metabolic (HbA1c), inflammatory (CRP), and nutritional (calcidiol) parameters also confirmed the independence of this association from systemic inflammation or metabolic dysfunction. The consistent risk observed in physiologically normal individuals supports the hypothesis that sevoflurane’s long-term effects may act through central neuroinflammatory pathways rather than peripheral metabolic alterations. Collectively, these convergent results across disease severity, surgical context, and biochemical strata strengthen the internal validity of the association and its biological plausibility.

This study’s methodological rigor lies in its advanced pharmacoepidemiologic framework, designed to minimize bias and strengthen causal inference between anesthetic exposure and long-term depression risk. Leveraging the Global Federated Network, which integrates longitudinal electronic health records from more than 200 healthcare organizations across three continents and represents over 250 million patients, we achieved exceptional external validity and generalizability across diverse populations and surgical contexts. To ensure temporal validity, a one-year latency washout excluded patients with depression diagnoses or antidepressant use within one year before or after anesthesia, eliminating reverse causation and transient postoperative mood effects while establishing a clear temporal sequence. PSM yielded near-perfect baseline balance (all standardized mean differences <0.1), ensuring exchangeability and mitigating confounding by indication. A suite of robustness analyses provided multi-level causal triangulation: The association persisted in no-crossover models, confirming specificity; the repeated-exposure analysis revealed a dose–response gradient consistent with biologic plausibility; mortality-censored models excluded survival bias; and outcome controls demonstrated specificity to psychiatric endpoints but not unrelated events. Finally, the E-value (2.38) indicated that an unmeasured confounder would require a ≥ 2.4-fold association with both exposure and outcome to fully explain the observed effect—an implausible magnitude. Collectively, the combination of temporal separation, balanced comparability, multidimensional sensitivity testing, and quantitative bias evaluation represents a rigorously structured observational framework within observational pharmacoepidemiology, substantially reinforcing the internal validity of the observed association and strengthening its epidemiologic credibility.

The present findings indicate that anesthetic selection may exert durable neuropsychiatric consequences that extend well beyond the perioperative window. Depression—one of the leading causes of global disability—has historically not been incorporated into anesthetic risk stratification frameworks [57]. In this large multinational cohort, exposure to sevoflurane was associated with an approximately 50% higher long-term risk of incident depression compared with propofol (aHR ≈ 1.5). From a clinical perspective, anesthetic selection may represent a potentially modifiable factor influencing long-term brain health, particularly in populations at increased risk for neuroinflammatory sequelae—such as older adults and patients undergoing repeated procedures. The observed biologic gradient—a progressive increase in risk with cumulative sevoflurane exposure—suggests a dose-dependent vulnerability consistent with its known neuroinflammatory and oxidative properties. These findings support careful and individualized risk–benefit assessment. This is particularly critical for individuals requiring serial operations, where the cumulative neuroinflammatory burden of repeated sevoflurane exposure warrants careful evaluation relative to the potential neuroprotective properties of propofol-based total intravenous anesthesia (TIVA) [10, 11, 58, 59]. Conversely, propofol-based TIVA may offer relative neuropsychological protection, aligning with its anti-inflammatory and antioxidant mechanisms. Prior trials showing lower rates of postoperative delirium and delayed neurocognitive recovery with propofol now find parallel support in long-term psychiatric outcomes [10, 42–44, 58, 60], suggesting a possible protective effect against late-onset depression [43, 44, 60]. From a broader perioperative medicine standpoint, these findings may inform consideration of long-term neuropsychiatric outcomes alongside traditional short-term recovery endpoints. Given the observational design, these findings should be interpreted as hypothesis-generating; nevertheless, the magnitude, consistency, and graded association observed support consideration of long-term neuropsychiatric outcomes within anesthetic decision-making frameworks. Within a precision perioperative medicine paradigm, anesthetic selection—particularly in high-risk or multi-exposure patients—should be considered part of a broader strategy aimed at preserving long-term cognitive and emotional health. Grounded in multinational real-world data, this study contributes to the evolving discussion of how anesthetic strategies may intersect with long-term mental health outcomes. In addition to its potential neuropsychiatric implications, anesthetic selection increasingly intersects with considerations of environmental sustainability. Volatile anesthetics, including sevoflurane, are recognized greenhouse gases with measurable global warming potential and contribute to healthcare-related carbon emissions. In contrast, propofol-based TIVA avoids direct atmospheric release of inhalational agents and aligns more closely with emerging “Green Anesthesia” initiatives. While environmental impact was not evaluated in the present study, integrating clinical safety signals with ecological responsibility may further inform anesthetic decision-making in contemporary perioperative practice.

This study has limitations inherent to its observational design, though multiple safeguards were applied to mitigate bias. Residual confounding cannot be fully excluded. Although our propensity-score matching model accounted for chronic pain diagnoses, opioid exposure, and other pain-related pharmacotherapy, these variables represent proxy measures derived from clinical coding and prescriptions. Perioperative pain intensity and its subjective psychological burden—factors known to influence long-term depression risk—are not directly measurable within structured EHR data. Although opioid exposure and chronic pain diagnoses serve as objective surrogates, unmeasured variation in pain severity, persistence, and affective distress may have contributed to residual confounding. Given the well-recognized bidirectional relationship between pain and depression, such residual confounding cannot be entirely excluded. Nevertheless, rigorous PSM balanced over 100 demographic, clinical, pharmacologic, and surgical covariates. The E-value was 2.38 for the point estimate and 2.17 for the lower confidence limit, indicating the minimum strength of association an unmeasured confounder would require with both exposure and outcome to shift the observed association to the null. Depression ascertainment relied on diagnostic codes and antidepressant prescriptions rather than structured psychiatric interviews, which may underdetect mild or transient symptoms. However, severity-stratified analyses restricted to moderate-to-severe depression consistently reproduced the main effect, strengthening construct validity. Although competing-risk modeling was not feasible within the TriNetX platform, mortality-censored analyses produced nearly identical results, arguing against substantial survival bias. The use of a federated, multinational EHR network also introduces coding heterogeneity; however, this enhances external validity by demonstrating consistency across healthcare systems. Finally, while causal inference from real-world data remains inherently limited, a randomized trial powered to evaluate multi-year neuropsychiatric outcomes of different anesthetic agents would pose substantial logistical and ethical challenges. Accordingly, rigorously conducted real-world evidence remains essential for investigating potential long-term anesthetic effects and informing future mechanistic and prospective studies.

Conclusions

In this large, multinational, propensity score–matched cohort study, sevoflurane exposure was associated with a significantly higher long-term risk of new-onset depression compared with propofol (aHR 1.51; 95% CI, 1.41–1.61). This association persisted across definitions, subgroups, and sensitivity analyses, and the dose–response pattern with repeated exposure supports a biologically plausible relationship. These findings indicate that anesthetic choice may be associated with lasting neuropsychiatric differences beyond short-term postoperative recovery. Given the global burden of depression, this signal warrants clinical awareness and further mechanistic validation. While randomized trials for such comparisons are ethically and logistically challenging, rigorously conducted real-world evidence currently represents an important and ethically feasible approach to inform anesthetic decision-making and future research into long-term mental health outcomes.

Author contributions

MS and YW were responsible for study conception, design, data collection, and drafting of the initial manuscript. ZL, J. Wang, LJ, RZ, and TM contributed to data curation, statistical analysis, and interpretation of findings. MM, LZ, and SF assisted in patient recruitment, clinical data acquisition, and validation of results. WMC provided methodological guidance, critical revision of the manuscript, and supervision of analytic approaches. SYW served as senior author, supervised the study design and execution, provided major intellectual input, and critically revised the manuscript for important intellectual content. JZ served as senior author, contributed to overall study conception and design, provided oversight of clinical components, and critically revised the manuscript. All authors have read and approved the final version of the manuscript.

Research funding

This work was supported by the National Key Research and Development Program of China (Funding Number: 2023YFC25069003), which funded Zhang’s research, and by the Lo-Hsu Medical Foundation, Lotung Poh-Ai Hospital, which supported the work of Szu-Yuan Wu (Funding Numbers: 11403 and 11404).

Data availability

The data supporting the findings of this study were obtained from the TriNetX Global Federated Network. These data are de-identified and aggregated in accordance with HIPAA and GDPR regulations to ensure patient privacy. Restrictions apply to the availability of these data, which were used under a license agreement for the current study and are not publicly available. However, data access can be requested directly through the TriNetX platform (https://trinetx.com/) by member healthcare organizations or approved researchers.

Competing interests

The authors declare no competing interests.

Ethics approval and consent

This study utilized the TriNetX platform, which exclusively provides de-identified and aggregated electronic health record data. In accordance with the Health Insurance Portability and Accountability Act (HIPAA), such data do not allow for individual identification and thus do not meet the federal definition of human subjects research. Consequently, research conducted using this platform is exempt from Institutional Review Board (IRB) oversight and does not require informed consent.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-026-04075-6.