Postoperative delirium in hip fracture patients linked to epigenetic alterations in inflammatory and immune pathways: a genome-wide DNA methylation study

Abstract

Delirium is a common yet underdiagnosed condition in elderly hospitalized patients. The lack of effective diagnostic and therapeutic methods can be attributed to the limited understanding of its pathophysiology. Delirium has recently been reported to be linked to neuroinflammation and epigenetic changes. The aim of this study was to validate the pathways in a larger cohort with a uniform type of surgery, while rigorously adjusting for potential covariates. This study primarily investigated DNA methylation (DNAm) changes before and after surgery in postoperative delirium (POD) among older adults having undergone femoral fracture surgery. After propensity analysis, 65 subjects were divided into 2 subgroups; one consisted of subjects who had blood samples collected preoperatively and on postoperative day 0, and the other consisted of those who had samples collected preoperatively and on postoperative day 3. We performed differential methylation analysis and enrichment analysis on each subgroup. Enrichment analysis using CpGs that exhibited substantial DNAm changes between pre- and postoperative day 0 samples in POD cases showed inflammation- and immunity-related pathways such as “leukocyte mediated immunity” and “NF-kappa B signaling pathway.” Inflammation- and immunity-related pathways became less noticeable between pre- and postoperative day 3 samples. Inflammation- and immunity-related pathways showed in this study align with previous studies across diverse populations, reinforcing the role of inflammation- and immunity-related epigenetic mechanisms in delirium including POD. Notably, these DNAm changes were potentially transient, corresponding to the typical onset of delirium, suggesting their potential as biomarkers for early diagnosis.

Introduction

Delirium is a common and dangerous condition that frequently affects elderly hospitalized patients and can lead to poor outcomes. However, it is often underdiagnosed and undertreated [1–3]. The lack of effective diagnostic and therapeutic methods for delirium is partly attributed to the limited understanding of its pathophysiology. In the animal study of delirium using exogenous insults, an increase in pro-inflammatory cytokines (e.g., TNF-α, IL-1β, and IL-6) from activated microglia accompanied by cognitive deficits in behavioral tests has been repeatedly observed [4, 5]. Importantly, cognitive deficits in behavioral tests, the activation of microglia, and the increase in pro-inflammatory cytokines were more pronounced in aged animals than in young animals. This finding aligns with the clinical observation that aging is a major risk factor for delirium in humans, suggesting that heightened neuroinflammation may also apply to the pathophysiology of delirium in humans. However, the molecular mechanism by which the release of pro-inflammatory cytokines is enhanced in aged animals remains unclear. Aging is known to have a great impact on gene expression in the brain [6], and the expression of genes encoding pro-inflammatory cytokines may also change with aging. Such changes in gene expression are tightly controlled by epigenetic modifications, including DNA methylation (DNAm) [7]. DNAm plays an important role in the molecular mechanisms that control genome-wide gene expression [8] and changes dynamically with aging [9–11].

We have focused on DNAm and reported several studies that investigated the relationship between delirium and DNAm [12–16]. First, we reported that the TNF-α expression increased and DNAm levels at most of the CpGs on the TNF-α gene decreased with aging in blood from Grady Trauma Project cohort of 265 subjects with high prevalence of trauma exposure [12]. Furthermore, when using brain samples provided from a Neurosurgery cohort of 6 subjects, we found negative correlations between DNAm levels at CpGs on TNF-α gene and aging in glial cells as well [12]. This finding is consistent with the idea that enhanced inflammation with associated gene expression changes occurs in elderly humans. In our recent report focusing mainly on a Caucasian population, we compared pre- and postoperative blood DNAm levels of patients who developed delirium after neurosurgery [15]. After excluding the potential effect of common factors related to surgery and anesthesia between postoperative delirium (POD) cases and non-POD controls, we identified enrichment in pathways such as “immune response” and “T cell activation” [15]. Also, our study comparing pre- and postoperative blood DNAm levels of Japanese patients who developed delirium after gastrointestinal surgery similarly reported pathways for immune and inflammatory signals such as “T cell activation” [16]. These results indicate the possible involvement of immune and inflammatory functions in the pathophysiology of delirium, including POD, regardless of the etiology of delirium and ethnic background.

As mentioned above, we have been evaluating multiple cohorts to capture universal DNAm changes before and after surgery that occur with delirium. Indeed, as the patients in our latest study [16] underwent resection of abdominal organs, there were no direct surgical insults to their nervous system as in the case of neurosurgery, making the results of the study more applicable to delirium from common etiologies. However, the cohort included patients who underwent many different types of surgery, and the invasiveness of each surgery varied. In addition, the sample size was small, and various potential confounding factors such as “anesthesia type”, “operative time”, and “blood loss” were not considered as covariates. To overcome the aforementioned limitations, we conducted a prospective observational epigenome-wide association study (EWAS) focusing on patients who underwent hip fracture repair surgery to focus on the uniform type of surgery as a trigger of delirium with the largest sample size of EWAS study of POD to date with 65 subjects.

Materials and methods

Subjects

Patients admitted to the orthopedic ward for femoral fracture surgery at Kameda Medical Center, Japan, were approached for the study recruitment. A total of 148 subjects who provided written consent to participate in the study were enrolled between October 2020 and September 2023. Three subjects withdrew their consent, 4 subjects already had positive Confusion Assessment Method for the Intensive Care Unit (CAM-ICU) [17] results before surgery, and 7 subjects had samples or assessments on the scheduled day of sample collection or assessment that could not be obtained were excluded. Consequently, 134 subjects remained for final analysis. This study was approved by the Institutional Review Boards of Stanford University (study approval no.: 63791) and Kameda Medical Center (study approval no.: 19-171-230331).

Data collection and clinical categorization

Doctors or nurses collected subject data through reviews of electronic medical records (EMRs) and patient interviews from preoperative to postoperative day 3. The Delirium Rating Scale-Revised-98 (DRS-R-98) [18] and CAM-ICU were used to assess the severity and presence of delirium, respectively. Delirium assessment was performed once at each time point from preoperative through postoperative day 3. When blood samples were collected, delirium assessment was performed at the same time point as blood collection, except for postoperative day 0, when blood samples were obtained immediately after surgery and delirium assessments were performed after return to the hospital room, typically within approximately 1 h after surgery. POD was defined by a DRS-R-98 severity score ≥ 19, a positive CAM-ICU result, or adjudication by anesthesiologists based on clinical descriptions in the EMRs that provided evidence of delirium [19]. Subjects who were diagnosed with POD at least once by postoperative day 3 were categorized as POD cases.

Thirty-two subjects out of the remaining 134 subjects were categorized as POD cases. Propensity analysis was performed to match non-POD controls to POD cases. The propensity scores for POD were calculated using logistic regression including 4 covariates; age, sex, and preoperative DRS-R-98 (severity and total) scores. Finally, 33 non-POD controls and 32 POD cases were selected (Fig. 1).

Fig. 1. Subject flowchart.

Abbreviations: CAM-ICU, the Confusion Assessment Method for the Intensive Care Unit; POD, postoperative delirium.

Demographic analysis

A t-test was performed to compare the mean values of age, DRS-R-98 severity score, DRS-R-98 total score, anesthesia duration, operative time, and blood loss. Fisher’s exact test was performed to compare the proportions of sex, dementia status, CAM-ICU result, and anesthesia type (general vs. spinal anesthesia). Postoperative DRS-R-98 (severity and total) scores and postoperative CAM-ICU results were based on the assessments conducted on the same day when the postoperative blood samples were collected. EZR [20], which is a graphical user interface for R (The R Foundation for Statistical Computing, Vienna, Austria) [21], was used for the data analysis.

Sample collection and processing

Blood samples were collected pre- and postoperatively. Preoperative blood sampling was performed in the hospital room in the afternoon of the day before surgery, or in the case of emergency surgery, several hours before surgery. Given that POD typically develops within the first 72 h after surgery, postoperative blood samples were deliberately collected at 2 separate time points: either immediately after surgery (i.e., on postoperative day 0) in the operating room or on postoperative day 3 in the hospital room.

Forty-eight subjects (non-POD controls = 24, POD cases = 24) out of 65 subjects provided their blood samples before surgery as well as on postoperative day 0. Seventeen subjects (non-POD controls = 9, POD cases = 8) out of 65 subjects provided their blood samples before surgery and on postoperative day 3. The difference in sample size between the 2 time points was due to the higher frequency of POD on postoperative day 0, for which we intentionally collected a larger number of samples. A total of 130 blood samples (65 pre-post surgery pairs) from 65 subjects was stored at −80 °C until DNA extraction.

DNA extraction and bisulfite conversion

DNA was isolated from blood samples with the QIAamp DNA Mini Kit (51306, QIAGEN, Germantown, MD, USA). DNA quantity was assessed with the Qubit^TM dsDNA Broad Range Assay Kit (Q32850, ThermoFisher Scientific, Waltham, MA, USA). DNA was stored at −80 °C until bisulfite conversion. 500 ng of DNA was bisulfite-converted with the EZ-96 DNA Methylation Kit (D5004, Zymo Research, Irvine, CA, USA). Bisulfite-converted DNA was stored at −80 °C.

Epigenome-wide analysis

DNAm of 130 bisulfite-converted blood DNA samples were analyzed with the Illumina Infinium MethylationEPIC V2 BeadChip array (G0209006, Illumina, San Diego, CA, USA). The raw methylation data was processed on R [21] with the R packages “ChAMP” [22, 23] and “minfi” [24, 25]. During the loading of data, probes were filtered out if they (i) had a detection p-value > 0.01, (ii) had < 3 beads in at least 5% of samples per probe, (iii) were non-CpG, SNP-related, or multi-hit probes, (iv) were located on chromosome X or Y. Samples were normalized by beta-mixture quantile dilation [26], and the batch effect was corrected by the ComBat [27] normalization approach before performing differential methylation analysis. Estimated cell proportions for CD8 T cells, CD4 T cells, natural killer cells, B cells, and monocytes were calculated by the DNAm Age Calculator available online [28, 29] using the method described in the literature [30]. Differential methylation analysis was conducted with the R package “Rnbeads” [31] using “limma” method [32] by specifying age, sex, anesthesia type, operative time, blood loss, and cell type proportions (CD8 T cells, CD4 T cells, natural killer cells, B cells, and monocytes) as covariates. Genome-wide significance was set at p-value < 5.00E-08.

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways were evaluated in enrichment analysis with the R package “missMethyl” [33] using the Gometh [34] by adjusting for the variable number of CpGs tested in each gene. For enrichment analysis, we compared (A) non-POD controls (n = 33) and POD cases (n = 32) blood DNAm levels of preoperative samples, (B) non-POD controls (n = 24) and POD cases (n = 24) blood DNAm levels of postoperative day 0 samples, (C) non-POD controls (n = 9) and POD cases (n = 8) blood DNAm levels of postoperative day 3 samples, (D) pre- (n = 24) and postoperative day 0 (n = 24) blood DNAm levels of POD cases, and (E) pre- (n = 8) and postoperative day 3 (n = 8) blood DNAm levels of POD cases (Figure S1). In the comparisons of (A), (B), and (C), differentially methylated CpGs were processed by (i) filtering with a p-value < 0.05, remaining only significant CpGs, (ii) extracting the top 1500 CpGs ranked by absolute β-value differences, which were then used for enrichment analysis. In the comparisons of (D) and (E), to obtain differentially methylated CpGs altered before and after surgery specifically based on the presence or absence of POD, we excluded the potential impact of common factors related to surgery and anesthesia. Specifically, we (i) filtered differentially methylated CpGs in both non-POD controls and POD cases with a p-value < 0.05, remaining only significant CpGs, (ii) excluded commonly significant CpGs between non-POD controls and POD cases from differentially methylated CpGs in POD cases, (iii) extracted the top 1500 CpGs in POD cases ranked by absolute β-value differences, which were then used for enrichment analysis. In other words, the cut-offs for absolute β-value differences were determined by obtaining the top 1500 CpGs.

Finally, to validate the robustness of our findings and address the within-subject correlation structure more rigorously, we performed a sensitivity analysis using a linear mixed-effects modeling (LMM) approach with the R package “limma” [32]. We applied a repeated measures design using the duplicateCorrelation function to account for the correlation between pre- and postoperative samples from the same subject. In this model, postoperative day 0 and day 3 samples were combined into a single “post” category to maximize statistical power. The design matrix included a Time (pre vs. post) × Group (non-POD vs. POD) interaction term, adjusting for the same covariates used in the main analysis: age, sex, anesthesia type, operative time, blood loss, and estimated cell type proportions (CD8 T cells, CD4 T cells, natural killer cells, B cells, and monocytes). We evaluated the significance of the interaction coefficient to identify CpGs exhibiting POD-specific longitudinal changes.

Results

Subject demographics

There were no significant differences in demographics between non-POD controls and POD cases (Table S1).

The average age of the subjects who provided blood samples before surgery as well as on postoperative day 0 was 83.7 years (Standard Deviation, SD = 8.1) and 27 subjects (56.3%) were female (Table 1). Postoperative DRS-R-98 (severity and total) scores and postoperative CAM-ICU results were based on the assessments conducted on postoperative day 0. There were significant differences between the 2 groups (non-POD vs. POD) in postoperative DRS-R-98 (severity and total) scores (p < 0.001 for both), postoperative CAM-ICU positive (p < 0.001), and anesthesia types (p = 0.001). There were no significant differences in age, sex, dementia status, preoperative DRS-R-98 (severity and total) scores, anesthesia duration, operative time, and blood loss between the 2 groups.

Table 1.

Comparisons of mean values or proportions of subject characteristics between the non-POD and POD groups in 2 subgroups.

Characteristics	Postoperative day 0 subjects	Non-POD	POD	p-value	Postoperative day 3 subjects	Non-POD	POD	p-value
	N = 48	N = 24	N = 24		N = 17	N = 9	N = 8
Mean age (years)	83.7	83.0	84.4	0.562	83.8	82.7	85.0	0.572
SD	8.1	9.6	6.3		8.1	9.4	6.9
Female sex (n)	27	13	14	1	10	5	5	1
%	56.3	54.2	58.3		58.8	55.6	62.5
Asian race (n)	48	24	24		17	9	8
%	100.0	100.0	100.0		100.0	100.0	100.0
Dementia (n)	2	1	1	1	0	0	0
%	4.2	4.2	4.2		0.0	0.0	0.0
Preoperative DRS-R-98 severity score (points)	3.0	2.5	3.5	0.174	1.6	1.4	1.8	0.672
SD	2.6	2.0	3.1		1.4	1.1	1.8
Preoperative DRS-R-98 total score (points)	3.1	2.5	3.7	0.132	1.7	1.4	2.0	0.533
SD	2.8	2.0	3.3		1.8	1.1	2.3
Postoperative DRS-R-98 severity score (points)	8.1	2.6	13.6	< 0.001***	1.6	1.1	2.1	0.146
SD	9.1	2.0	10.1		1.4	0.6	1.9
Postoperative DRS-R-98 total score (points)	9.5	2.6	16.3	< 0.001***	1.6	1.1	2.1	0.146
SD	11.2	2.0	12.3		1.4	0.6	1.9
Postoperative CAM-ICU positive (n)	11	0	11	< 0.001***	0	0	0
%	22.9	0.0	45.8		0.0	0.0	0.0
General anesthesia (n)	26	7	19	0.001**	5	5	0	0.029*
%	54.2	29.2	79.2		29.4	55.6	0.0
Mean anesthesia duration (min)	155.0	144.0	166.1	0.097	122.7	118.1	127.9	0.613
SD	46.2	46.8	43.9		38.0	27.7	48.7
Mean operative time (min)	89.1	85.4	92.8	0.476	68.9	62.9	75.8	0.377
SD	35.3	37.5	33.4		28.9	14.5	39.6
Mean blood loss (ml)	143.3	123.3	163.3	0.331	100.0	60.0	145.0	0.113
SD	141.1	127.5	153.6		109.6	72.8	130.5

Postoperative DRS-R-98 (severity and total) scores and postoperative CAM-ICU results were based on the assessments conducted on the same day when the postoperative blood samples were collected.

CAM-ICU the Confusion Assessment Method for the Intensive Care Unit, DRS-R-98 the Delirium Rating Scale-Revised-98, POD, postoperative delirium, SD standard deviation.

*p < 0.05, **p < 0.01, ***p < 0.001.

The average age of the subjects who provided blood samples before surgery and on postoperative day 3 was 83.8 years (SD = 8.1) and 10 subjects (58.8%) were female. Postoperative DRS-R-98 (severity and total) scores and postoperative CAM-ICU results were based on the assessments conducted on postoperative day 3. There was a significant difference between the 2 groups (non-POD vs. POD) in anesthesia types (p = 0.029). There were no significant differences in other items.

Differential methylation analysis

Preoperative sample comparison between non-POD controls and POD cases

Table S2 presents the top 20 CpGs when comparing preoperative samples from non-POD controls and POD cases. No genome-wide significant signal was found.

Postoperative sample comparison between non-POD controls and POD cases

Table S3 presents the top 20 CpGs when comparing postoperative day 0 samples from non-POD controls and POD cases. No CpG signal was identified at the genome-wide significance level.

Table S4 presents the top 20 CpGs when comparing postoperative day 3 samples from non-POD controls and POD cases. No CpG signal was identified at the genome-wide significance level.

Assessment of the pre-postoperative changes in POD cases excluding overlapping CpGs observed in the pre-postoperative changes in non-POD controls

Table 2 presents the top 20 CpGs when comparing pre- and postoperative day 0 samples of POD cases, excluding CpGs that were commonly differentially methylated between pre- and postoperative day 0 comparison in non-POD controls and pre- and postoperative day 0 comparison in POD cases. No genome-wide significant signal was discovered.

Table 2.

Top 20 CpGs comparing pre- and postoperative day 0 samples of POD cases, excluding CpGs that were commonly differentially methylated between pre- and postoperative day 0 comparison of non-POD controls and pre- and postoperative day 0 comparison of POD cases.

Gene Name	Gene Group	CpG	Chromosome	mean pre	mean day 0	mean diff	diffmeth p-value
RGS4	TSS1500;TSS200	cg23309924	chr1	87.9%	91.5%		1.66E-05
		cg13084714	chr17	92.3%	93.9%		2.06E-05
IMPG1	Body	cg00713113	chr6	3.1%	2.2%		3.03E-05
N4BP2L1	TSS200	cg13424209	chr13	9.4%	8.3%		3.14E-05
SPAG17	Body	cg12851362	chr1	82.0%	85.0%		3.99E-05
SPATA13	TSS1500	cg05325643	chr13	90.7%	92.6%		4.60E-05
		cg17001881	chr16	90.6%	91.8%		6.73E-05
		cg13496119	chr14	82.9%	86.6%		8.33E-05
ERC1	Body	cg08817674	chr12	89.6%	91.4%		8.70E-05
SIRT1	Body	cg05443670	chr10	90.2%	92.2%		9.16E-05
		cg15561913	chr6	86.8%	88.8%		9.31E-05
MAST2	Body	cg16565164	chr1	75.6%	78.2%		1.02E-04
C5orf34	Body	cg12771068	chr5	92.8%	93.7%		1.07E-04
RAB33B	Body	cg02339519	chr4	89.6%	92.2%		1.15E-04
MAML2	Body	cg23581811	chr11	82.7%	85.0%		1.34E-04
PARD3B	Body	cg11530799	chr2	92.5%	93.7%		1.37E-04
		cg22680980	chr5	87.6%	89.4%		1.39E-04
		cg25973405	chr7	84.4%	86.9%		1.47E-04
		cg09669309	chr12	88.7%	90.7%		1.66E-04
		cg00791024	chr15	15.6%	13.7%		1.78E-04

Mean diff values highlighted in red indicate increased DNA methylation, whereas those in blue indicate a reduction.

Diff β-value differences in day 0 compared to pre, POD postoperative delirium.

Table 3 presents the top 20 CpGs when comparing pre- and postoperative day 3 samples of POD cases, excluding CpGs that were commonly differentially methylated between pre- and postoperative day 3 comparison in non-POD controls and pre- and postoperative day 3 comparison in POD cases. No genome-wide significant signal was discovered.

Table 3.

Top 20 CpGs comparing pre- and postoperative day 3 samples of POD cases, excluding CpGs that were commonly differentially methylated between pre- and postoperative day 3 comparison of non-POD controls and pre- and postoperative day 3 comparison of POD cases.

Gene Name	Gene Group	CpG	Chromosome	mean pre	mean day3	mean diff	diffmeth p-value
SEMA6A	TSS1500	cg11247562	chr5	4.9%	3.6%		5.06E-06
		cg05180767	chr13	90.3%	92.4%		5.55E-06
		cg18638536	chr19	94.6%	95.5%		7.84E-06
SND1;LRRC4	Body	cg14059988	chr7	96.0%	97.2%		8.86E-06
TIPRL	TSS1500	cg13700215	chr1	6.0%	5.0%		1.11E-05
LOC643339	Body	cg10433461	chr12	89.4%	90.9%		1.13E-05
IPMK	Body	cg18758937	chr10	85.2%	88.3%		1.46E-05
FGFR2	Body	cg04059337	chr10	79.5%	86.9%		1.58E-05
LDLR	TSS1500	cg10201616	chr19	92.9%	93.6%		1.96E-05
ZNF777	Body	cg07562382	chr7	94.3%	95.3%		2.11E-05
TMEM233	Body	cg11385769	chr12	13.1%	12.0%		2.16E-05
		cg09951607	chr10	88.1%	90.1%		2.20E-05
KLHL12	TSS1500	cg10422823	chr1	84.3%	86.4%		2.24E-05
		cg20874668	chr6	93.3%	95.6%		2.29E-05
		cg26210281	chr12	88.0%	90.4%		2.33E-05
PDS5B	5’UTR	cg09725202	chr13	93.8%	93.6%		2.41E-05
PAX8-AS1;PAX8	Body	cg01211054	chr2	89.3%	91.5%		2.57E-05
COL23A1	Body	cg10795898	chr5	8.6%	6.4%		2.64E-05
FBRSL1	Body	cg24621258	chr12	94.0%	95.1%		2.67E-05
		cg07084675	chr18	86.9%	89.0%		3.02E-05

Mean diff values highlighted in red indicate increased DNA methylation, whereas those in blue indicate a reduction.

Diff β-value differences in day 3 compared to pre, POD postoperative delirium.

Enrichment analysis

Preoperative sample comparison between non-POD controls and POD cases

Table S5 presents the top 20 pathways identified from the enrichment analysis comparing (A) blood DNAm levels of preoperative samples from non-POD controls (n = 33) and POD cases (n = 32). While not at the false discovery rate (FDR) significance level, GO analysis identified several terms related to cell adhesion such as “cell-cell adhesion via plasma-membrane adhesion molecules” and “homophilic cell adhesion via plasma membrane adhesion molecules” at nominal significance. KEGG analysis also identified several similar pathways such as “cell adhesion molecules” and “neuroactive ligand-receptor interaction” at nominal significance.

Postoperative sample comparison between non-POD controls and POD cases

Table S6 presents the top 20 pathways identified from the enrichment analysis comparing (B) blood DNAm levels of postoperative day 0 samples from non-POD controls (n = 24) and POD cases (n = 24). While they were not at the FDR significance, KEGG analysis identified several pathways such as “neuroactive ligand-receptor interaction” and “cell adhesion molecules” at nominal significance.

Table S7 presents the top 20 pathways identified from the enrichment analysis comparing (C) blood DNAm levels of postoperative day 3 samples from non-POD controls (n = 9) and POD cases (n = 8). GO analysis again revealed similar differences in several terms such as “homophilic cell adhesion via plasma membrane adhesion molecules” and “cell-cell adhesion via plasma-membrane adhesion molecules” at the FDR significance level. Additionally, “cell-cell adhesion” was nominally significant in GO analysis.

Assessment of the pre-postoperative changes in POD cases excluding overlapping CpGs observed in the pre-postoperative changes in non-POD controls

Table 4 presents the top 20 pathways identified from the enrichment analysis comparing (D) pre- (n = 24) and postoperative day 0 (n = 24) blood DNAm levels of POD cases, excluding CpGs that were commonly differentially methylated between non-POD controls and POD cases. While the differences did not reach the FDR significance, the terms identified in GO analysis were almost all related to inflammation and immunity such as “leukocyte mediated immunity”, “immune effector process”, and “lymphocyte activation” at nominal significance. KEGG analysis also identified many pathways related to inflammation and immunity such as “NF-kappa B signaling pathway” and “cytokine-cytokine receptor interaction” at nominal significance.

Table 4.

Top 20 pathways identified from the enrichment analysis comparing pre- and postoperative day 0 samples of POD cases, excluding CpGs that were commonly differentially methylated between pre- and postoperative day 0 comparison of non-POD controls and pre- and postoperative day 0 comparison of POD cases.

GO (p < 0.05, absolute β differences ≥ 2.1%, 1500 CpGs)						KEGG (p < 0.05, absolute β differences ≥ 2.1%, 1500 CpGs)
TERM	Ont	N	DE	P.DE	FDR	Description	N	DE	P.DE	FDR
leukocyte mediated immunity	BP	389	19	4.99E-06	0.113	NF-kappa B signaling pathway	102	5	0.025	1
immune effector process	BP	627	24	2.56E-05	0.291	Complement and coagulation cascades	84	4	0.026	1
detection of biotic stimulus	BP	37	5	1.83E-04	0.919	Yersinia infection	137	6	0.027	1
adaptive immune response based on somatic recombination of immune receptors built from immunoglobulin superfamily domains	BP	303	14	2.33E-04	0.919	Homologous recombination	41	3	0.028	1
regulation of adaptive immune response	BP	205	11	2.86E-04	0.919	Viral protein interaction with cytokine and cytokine receptor	96	4	0.031	1
detection of external biotic stimulus	BP	25	4	3.10E-04	0.919	Synaptic vesicle cycle	78	4	0.034	1
regulation of lymphocyte mediated immunity	BP	185	10	3.31E-04	0.919	Pentose phosphate pathway	31	2	0.075	1
lymphocyte mediated immunity	BP	291	13	3.45E-04	0.919	Staphylococcus aureus infection	82	3	0.081	1
mononuclear cell differentiation	BP	478	19	3.64E-04	0.919	Cytokine-cytokine receptor interaction	292	7	0.086	1
leukocyte activation	BP	943	29	4.42E-04	0.943	Salmonella infection	248	7	0.094	1
cell activation	BP	1088	32	5.03E-04	0.943	Apoptosis - multiple species	32	2	0.099	1
adaptive immune response	BP	457	17	5.06E-04	0.943	Neutrophil extracellular trap formation	178	5	0.100	1
regulation of adaptive immune response based on somatic recombination of immune receptors built from immunoglobulin superfamily domains	BP	190	10	5.95E-04	0.943	ATP-dependent chromatin remodeling	113	4	0.103	1
regulation of biological process	BP	11866	209	6.60E-04	0.943	Hippo signaling pathway - multiple species	29	2	0.103	1
negative regulation of cell activation	BP	219	11	6.76E-04	0.943	Primary immunodeficiency	37	2	0.107	1
lymphocyte activation	BP	777	25	6.82E-04	0.943	Type I diabetes mellitus	40	2	0.118	1
lymphocyte differentiation	BP	426	17	7.05E-04	0.943	Rheumatoid arthritis	88	3	0.129	1
negative regulation of immune system process	BP	513	18	8.72E-04	1	Chemical carcinogenesis - DNA adducts	64	2	0.134	1
regulation of leukocyte mediated immunity	BP	246	11	9.15E-04	1	T cell receptor signaling pathway	122	4	0.146	1
negative regulation of leukocyte activation	BP	198	10	9.61E-04	1	Autophagy - animal	168	5	0.151	1

The cut-off for absolute β-value differences was determined by obtaining the top 1500 CpGs.

BP biological process, DE number of differentially expressed genes, which are the number of genes from the top 1500 CpGs, FDR false discovery rate, GO Gene Ontology, KEGG Kyoto Encyclopedia of Genes and Genomes, N number of genes included in each pathway, Ont ontology, POD postoperative delirium.

Table 5 presents the top 20 pathways identified from the enrichment analysis comparing (E) pre- (n = 8) and postoperative day 3 (n = 8) blood DNAm levels of POD cases, excluding CpGs that were commonly differentially methylated between non-POD controls and POD cases. GO analysis again identified a few terms related to inflammation and immunity such as “positive regulation of B cell activation” at nominal significance. KEGG analysis also identified a few pathways related to inflammation and immunity such as “cytokine-cytokine receptor interaction” at nominal significance.

Table 5.

Top 20 pathways identified from the enrichment analysis comparing pre- and postoperative day 3 samples of POD cases, excluding CpGs that were commonly differentially methylated between pre- and postoperative day 3 comparison of non-POD controls and pre- and postoperative day 3 comparison of POD cases.

GO (p < 0.05, absolute β differences ≥ 5.1%, 1500 CpGs)						KEGG (p < 0.05, absolute β differences ≥ 5.1%, 1500 CpGs)
TERM	Ont	N	DE	P.DE	FDR	Description	N	DE	P.DE	FDR
sequestering of TGFbeta in extracellular matrix	BP	4	3	7.25E-05	1	Intestinal immune network for IgA production	43	4	0.004	1
maintenance of protein location in extracellular region	BP	8	3	4.61E-04	1	Glycolysis / Gluconeogenesis	67	5	0.006	1
sequestering of extracellular ligand from receptor	BP	11	3	9.92E-04	1	Carbon metabolism	115	6	0.014	1
positive regulation of B cell activation	BP	84	7	0.001	1	African trypanosomiasis	36	3	0.018	1
protein sequestering activity	MF	23	4	0.001	1	Cytokine-cytokine receptor interaction	292	9	0.026	1
molecular sequestering activity	MF	40	5	0.001	1	Primary immunodeficiency	37	3	0.027	1
endoplasmic reticulum membrane organization	BP	13	3	0.001	1	Glycosaminoglycan biosynthesis - keratan sulfate	14	2	0.030	1
extracellular regulation of signal transduction	BP	13	3	0.002	1	Amoebiasis	102	5	0.043	1
extracellular negative regulation of signal transduction	BP	13	3	0.002	1	Biosynthesis of various nucleotide sugars	19	2	0.046	1
regulation of B cell activation	BP	129	8	0.003	1	Malaria	49	3	0.048	1
negative regulation of macrophage activation	BP	18	3	0.003	1	Other types of O-glycan biosynthesis	47	3	0.054	1
cellular response to fructose stimulus	BP	4	2	0.003	1	Pentose and glucuronate interconversions	36	2	0.068	1
endoplasmic reticulum tubular network membrane organization	BP	4	2	0.003	1	HIF-1 signaling pathway	110	5	0.072	1
regulation of plasminogen activation	BP	18	3	0.005	1	N-Glycan biosynthesis	55	3	0.073	1
positive regulation of adaptive immune response based on somatic recombination of immune receptors built from immunoglobulin superfamily domains	BP	122	7	0.005	1	Leukocyte transendothelial migration	115	5	0.073	1
negative regulation of transmembrane receptor protein serine/threonine kinase signaling pathway	BP	178	9	0.005	1	Virion - Human immunodeficiency virus	5	1	0.073	1
endoplasmic reticulum tubular network membrane	CC	5	2	0.005	1	Cell adhesion molecules	154	6	0.076	1
UDP-glucose metabolic process	BP	6	2	0.005	1	Hepatitis C	158	6	0.084	1
regulation of immune effector process	BP	384	14	0.006	1	Glycosphingolipid biosynthesis - lacto and neolacto series	28	2	0.096	1
estrone sulfotransferase activity	MF	1	1	0.006	1	Cell cycle	158	6	0.100	1

The cut-off for absolute β-value differences was determined by obtaining the top 1500 CpGs.

BP biological process, CC cellular component, DE number of differentially expressed genes, which are the number of genes from the top 1500 CpGs, FDR false discovery rate, GO Gene Ontology, KEGG Kyoto Encyclopedia of Genes and Genomes, N number of genes included in each pathway, Ont ontology, POD postoperative delirium.

Sensitivity analysis using the interaction model

We also conducted a sensitivity analysis using the Time × Group interaction model to verify the robustness of our findings. This analysis identified one CpG site reaching genome-wide significance (cg01534316; p = 2.41E-08, FDR = 0.021). Notably, although this site was not annotated in the standard array manifest, manual genomic inspection (hg38) localized it to the TMIGD3 gene, situated at the TMIGD3/ADORA3 locus. Two additional sites showed suggestive associations (both FDR = 0.075): cg05325643 (located in the upstream regulatory region of SPATA13) and cg20382094 (located within the COL13A1 gene). These sites exhibited distinct longitudinal trajectories between the non-POD and POD groups (Figure S2). Furthermore, although pathway analysis based on this model did not yield FDR significant terms due to limited power (Table S8), the top pathways extracted by nominal significance (p < 0.05) largely overlapped with those identified in our main analysis. Specifically, inflammation- and immunity-related pathways such as “leukocyte activation” and “lymphocyte activation” (GO), as well as “primary immunodeficiency” (KEGG), were consistently observed.

Discussion

This study is, to the best of our knowledge, EWAS with the largest sample size to investigate the role of DNAm in POD. We focused exclusively on patients who underwent repair surgery after femoral fractures, allowing us to narrow down the trigger of delirium to limit the potential confounders. We also selected non-POD controls matched to POD cases based on age, sex, and preoperative DRS-R-98 (severity and total) scores. We further accounted for various potential confounding factors that may influence POD vulnerability (e.g., anesthesia type: general vs. spinal anesthesia) as covariates. The absence of differences in POD assessment items between non-POD controls and POD cases on postoperative day 3 (i.e., postoperative day 3 DRS-R-98 (severity and total) scores and postoperative day 3 CAM-ICU results) suggests that the severity of POD had subsided by that time.

We conducted enrichment analysis using CpGs unique to POD cases comparing pre- and postoperative samples after extracting CpGs commonly seen both in POD cases and non-POD controls to remove the potential influence of surgery and anesthesia. Comparison between preoperative samples and those from postoperative day 0 of POD cases showed several GO terms related to inflammation and immunity such as “leukocyte activation”, “adaptive immune response”, and “lymphocyte activation” in the top 20. It is worth noting that these GO terms are highly consistent with those we discovered in pre- and postoperative comparisons of POD cases from our previous independent cohorts, ranging from neurosurgery patients [15] to even Japanese gastrointestinal surgery patients [16]. Furthermore, GO pathways identified in this study, including “leukocyte mediated immunity”, “immune effector process”, “leukocyte activation”, and “cell activation” are consistent with those comparing non-delirium and delirium in our previous cohort of a diverse patient population, primarily from general medicine inpatients [13]. This consistency further validates our analytic approach that exclude DNAm changes induced by surgery and anesthesia from those associated with, and potentially responsible for, the development of delirium. In KEGG analysis, inflammation- and immunity-related descriptions such as “NF-kappa B signaling pathway” and “T cell receptor signaling pathway” were included in the top 20. These KEGG pathways are also highly consistent with our previous report from the different independent cohort [15, 16]. Our sensitivity analysis using the standard interaction model (Time × Group) successfully captured a coherent epigenetic signature related to inflammation and immunity. This suggests that the epigenetic alterations identified in this study are robust across different statistical methodologies, further corroborating the biological validity of our findings. Notably, manual genomic inspection revealed that the top significant CpG site (cg01534316) is situated at the gene body of the TMIGD3/ADORA3 locus. Modulation of TMIGD3 via gene body methylation could influence inflammatory responses through suppression of NF-κB [35]. ADORA3 is a key regulator of anti-inflammatory signaling through deregulation of the NF-κB signaling pathway, leading to inhibition of pro-inflammatory cytokines (e.g., TNF-α) [36]. ADORA3 is also distributed on the neutrophils cell membrane and contribute to their chemotaxis and migration [37]. In this study, hypermethylation of the TMIGD3/ADORA3 gene body was observed in the POD group. While methylation in promoter regions typically suppresses gene expression, methylation in the gene body is often positively correlated with active transcription and is thought to maintain genomic stability by preventing spurious transcription initiation, although it remains unclear. [38, 39]. Based on this framework, the observed alteration may not represent gene silencing but rather reflect an active regulation or stabilization of immune-related pathways in response to surgical stress. Other recent studies have also demonstrated that DNAm changes modulate inflammatory and immune responses, neurotransmission, and synaptic dysfunction, thereby potentially predisposing individuals to perioperative neurocognitive disorders (PND) [40–42]. A fundamental study using a laparotomy PND model in aged mice reported a significant reduction in DNMT3a, resulting in decreased methylation in the LRG1 promoter region [41]. This change led to activation of TGF-β signaling by the increase in LRG1 level, ultimately impacts the synaptic function [41]. These findings further support the biological plausibility of the epigenetic pathways identified in the present study.

Intriguingly, when comparing preoperative samples and those taken on postoperative day 3, there were much fewer pathways related to inflammation and immunity. Such rapid changes in DNAm levels at CpGs associated with inflammation- and immunity-related pathways observed on postoperative day 0, followed by the swift disappearance of the differences in these pathways by postoperative day 3, align with the typical onset of delirium within 72 h after surgery [43]. The time course of DNAm changes discovered here in the study with the largest sample size has also been confirmed in our previous study with smaller sample size [16]. In fact, this time course temporally coincides with the clinical time course of delirium as well. This result suggests that DNAm has the potential to serve as a biomarker for delirium to detect its onset and monitor treatment response, provided that technological advancements make DNAm measurement more accessible and rapid. Of note, one study similarly reported that major surgery induced acute changes in DNAm associated with immune response pathway, although they were relatively stable in the postoperative period and generally persisting until discharge from the hospital [44]. Because they did not have information about DNAm differences between patients who developed POD and those who did not, these results cannot be directly compared to our data presented here. Thus, further investigation is required to determine the typical time course of DNAm changes after surgery.

We also performed enrichment analysis using top CpGs, where DNAm levels differed between POD cases and non-POD controls. In the comparisons between non-POD controls and POD cases of pre- and postoperative day 3 samples, both “homophilic cell adhesion via plasma membrane adhesion molecules” and “cell-cell adhesion via plasma-membrane adhesion molecules” were consistently presented in the top 20 GO terms. These terms are consistent with GO terms we previously reported in comparisons between non-POD controls and POD cases of pre- and postoperative samples from the gastrointestinal surgery cohort [16]. In KEGG analysis, while not significant at the FDR level, “cell adhesion molecules” were included in the top 20 KEGG descriptions for pre- and postoperative day 0 samples. This pathway is also consistent with our previous report about the comparisons between non-POD controls and POD cases of both pre- and postoperative samples from the neurosurgery cohort [15]. Notably, our sensitivity analysis identified COL13A1, a gene mediating cell-matrix adhesion, as a suggestive hit, further supporting the involvement of this pathway. These replicated findings suggest that cell adhesion may be a process that distinguishes future non-POD and POD even before surgical invasion and the difference in the process may also continue even after invasion. Cell adhesion molecules are known to regulate the integrity of the blood-brain barrier (BBB) and interactions between neurons [45]. They are also closely associated with inflammatory responses. Pro-inflammatory cytokines such as TNF-α and IL-6 modulate the expression of cell adhesion molecules, thereby increasing BBB permeability and promoting the progression of neuroinflammation [46]. Recent experimental evidence further supports this mechanism: a study demonstrated that BBB permeability induced by IL-1β was suppressed by omega-3 fatty acids, both in a human-derived BBB-on-chip model and in a mouse model of postoperative neuroinflammation [47]. Given that aging is associated with reduced BBB integrity [48], and increased BBB permeability has been independently associated with higher rates of delirium [49], the observed DNAm changes in cell adhesion molecules may serve as a novel evidence of molecular mechanism how BBB vulnerability contributes to POD pathophysiology. Further, “Neuroactive ligand-receptor interaction” was in the top 20 KEGG descriptions for pre- and postoperative day 0 samples in this study. This pathway was reported in our previous report from the gastrointestinal surgery cohort as well [16], Thus, this pathway may be involved in the pathogenesis of delirium after surgical invasion.

This study had some limitations. First, the definition of delirium is not based on Diagnostic and Statistical Manual of Mental Disorders-5 criteria, which is the gold standard psychiatric assessment [50], and delirium assessment was not performed twice daily; therefore, fluctuating or brief episodes may have been missed, potentially leading to underdiagnosis. Furthermore, a double-review or blinded adjudication process was not implemented. Second, dementia is known to be an important risk factor for delirium. To avoid reducing the sample size, we included these patients in our analysis after confirming that there was no significant difference in the proportion of dementia patients between the non-POD controls and POD cases (Table 1). Third, although we adjusted for major perioperative factors (i.e., anesthesia type, operative time, and blood loss) and demographic variables, there are still additional confounders potentially influencing the vulnerability for POD that were not included in this study due to the limited sample size. These include long-term medication use, specific comorbidities, alcohol consumption, and nutritional status. In addition, the timing of preoperative blood sampling was not fully standardized, and the exact interval from blood collection to surgical incision was not uniformly available; therefore, we were unable to include “time from blood collection to surgical incision” as a covariate. Although the urgent nature of femoral fracture surgery constrains this interval within a relatively narrow clinical window (patients typically underwent surgery within 48 h of injury), we cannot exclude residual confounding related to perioperative timing (e.g., circadian or stress-related effects on DNAm). However, our previous data from a more diverse group of inpatients with and without delirium from various etiologies [13] and our previous cohorts of surgical patients [15, 16] showed consistent results, suggesting that our present findings have a certain level of generalizability and replicability to other populations with delirium. Fourth, the final sample size remains relatively small for an epigenome-wide analysis. Although we employed LMM in our sensitivity analysis to maximize statistical power, the limited sample size generally constrains the statistical power to detect associations with smaller effect sizes (Type II error) and warrants caution regarding the potential for false positives. Moreover, while we identified specific CpGs with genome-wide significance, we did not perform predictive modeling due to the limited sample size. Thus, future studies with larger cohorts are needed to validate the predictive performance of these DNAm signatures for clinical use. Fifth, because our cohort predominantly comprised very old adults, the generalizability of these findings to younger populations remains uncertain. Future studies including younger age groups are warranted. Furthermore, all patients who underwent general anesthesia were managed using a uniform anesthetic regimen, consisting of desflurane in combination with remifentanil and fentanyl, so we were unable to evaluate differential effects of individual anesthetic agents. The generalizability of our findings to other anesthetic regimens remains uncertain. Sixth, interpreting the functional impact of DNAm requires caution regarding directionality. While we identified the pathways enriched with differentially methylated positions, these pathways included both hyper- and hypomethylated CpGs. Given that the relationship between methylation and gene expression varies by genomic context (e.g., suppression in promoters vs. activation in gene bodies), we interpret our findings as evidence of epigenetic involvement targeting specific pathways rather than uniform activation or inhibition of the pathways. Future studies incorporating multi-omics data are essential to precisely delineate the net functional outcome of these methylation changes on pathway activity. Finally, although this report suggests an association between epigenetic signals and POD, it does not necessarily indicate a causal relationship. In addition, the pathways identified in this study, such as “leukocyte mediated immunity” and “NF-kappa B signaling pathway,” are biologically plausible but not specific to delirium. Therefore, even though we adopted an approach that sought to exclude the potential effect of common factors related to surgery and anesthesia, these may reflect broader inflammatory and stress-related biological processes.

In summary, we have shown that inflammation- and immunity-related epigenetic signals are present in older adult patients with POD after femoral fracture surgery. These data are consistent with our previous investigations using independent cohorts including non-surgical and surgical patients from different ethnic groups, therefore providing further support for the involvement of inflammation- and immunity-related epigenetic mechanisms in the pathogenesis of delirium including POD. We also suggest that these signals may serve as a potential biomarker for POD, although further validation studies are warranted to establish their clinical utility.

Author contributions

T.S. organized and analyzed data and wrote the original and final manuscript. S.N. processed samples, analyzed data, critically reviewed the manuscript, and wrote the final manuscript. Y.N. organized data and critically reviewed the manuscript. K.Y., A.S., T.N., T.I., and B.A. critically reviewed the manuscript. T.A.S., N.J.P., N.G., and H.D.N. edited the manuscript. R.T. collected and organized data and critically reviewed the manuscript. T.K. and T.I. critically reviewed the manuscript. D.S. and K.U. collected samples and data, organized data, and critically reviewed the manuscript. T.Y. and G.S. conceived the study’s design, organized study structure including data collection, critically reviewed the manuscript, and edited the final manuscript.

Funding

This work was supported by research grants from the National Institute of Mental Health, United States (R01 MH119165 and R01 AG084710). G.S. also received research grant support from Sumitomo Pharma Co., Ltd. The supporters had no role in the design, analysis, interpretation, or publication of this study.

Data availability

The methylation array data generated in this study have been deposited in the NCBI Gene Expression Omnibus (GEO) database under accession number GSE330869. The analysis code used in this study is available from the corresponding author upon reasonable request.

Competing interests

G.S. is a co-founder of Delight Health Inc and has pending patents: “Non-invasive device for predicting and screening delirium,” PCT application no. PCT/US2016/064937 and US provisional patent no. 62/263,325; “Prediction of patient outcomes with a novel electroencephalography device,” US provisional patent no. 62/829,411; and “Epigenetic Biomarker of Delirium Risk,” PCT application no. PCT/US19/51276 and US provisional patent no. 62/731,599. T.I. is employed by Sumitomo Pharma Co., Ltd, and currently affiliated with RACTHERA Co., Ltd. The other authors declare no actual or potential conflicts of interest associated with this study.

Ethics approval and consent to participate

All methods were performed in accordance with the relevant guidelines and regulations. This study was approved by the Institutional Review Boards of Stanford University (study approval no.: 63791) and Kameda Medical Center (study approval no.: 19-171-230331). Written informed consent was obtained from all participants.

Supplementary information

The online version contains supplementary material available at 10.1038/s41398-026-04067-6.