Dysfunctional resting state network connectivity predicts postoperative delirium after major surgery

Abstract

Background:

Postoperative delirium is associated with increased morbidity, mortality, future cognitive decline and/or dementia. Understanding the neural mechanisms that differentiates an individual’s brain vulnerabilities is critical for future therapeutic development and prevention of postoperative delirium. In this research study, we investigated the hypothesis that impaired resting-state functional connectivity (FC) would explain predisposition to delirium.

Methods:

Preoperative blood oxygen level dependent functional MRI data were collected from 120 participants (>65yrs, 52 Female). Denoised blood oxygen level dependent signal time-series for 400 cortical regions were used to calculate resting-state FC within and between canonical resting state networks. We used a support vector machine to determine whether resting state FC across higher-order cortical networks were predictive of postoperative delirium.

Results:

Group comparisons revealed significantly decreased within-network connectivity in salience-ventral attention, cognitive control and default mode networks for participants with postoperative delirium (n=31) (non-parametric permutation test, 1000 iterations p<0.05). We found overall weaker connectivity within the default mode network and specific differences across the sub-networks of the default mode which overlap with higher-order cognitive processing. Supervised machine learning identified that the visual and salience-ventral attentional networks predicted postoperative delirium incidences with an accuracy of 68%.

Conclusions:

These results demonstrate that resting state FC is a neural correlate of vulnerability to postoperative delirium. In particular, revealing weaker connectivity across higher-order cognitive and salience attention networks are sites of vulnerability.

Introduction

Postoperative delirium is an acute state of cognitive impairment, attention, and/or arousal which has significant impact on overall long-term cognitive outcomes and associated with an increased risk of dementia.¹ Critically it is associated with increased morbidity, mortality and costs.² Amongst the prevailing theories, one posits that delirium is a ‘cognitive disintegration’ due to impaired cortical connectivity.³ The theory predicts that (1) pre-delirium connectivity will predict vulnerability to delirium and (2) that delirium results when the network disintegrates below a threshold of connectivity. While the latter is supported by recent EEG studies, showing a reduction in feedback cortical connectivity⁴ and resting state functional connectivity (rsFC) in delirium^5-7, data linking connectivity to vulnerability are less forthcoming.

Several studies have revealed that global and regional cortical atrophy, white-matter hyperintensities and structural based differences may predict postoperative delirium.^{5, 8-10} A systematic review of brain functional connectivity (FC) changes has revealed that delirium participants consistently have disrupted FC, lower EEG connectivity strength¹¹ and risk factors associated with delirium have lower structural connectivity.¹²

An fMRI study suggested that during, and after, delirium there were FC differences across the default mode network (DMN), however they could not exclude these differences were present pre-morbidly.¹³ This is especially important as the rsFC changes outlasted the delirium symptoms. Altered FC in the DMN is associated with disorders of cognition¹⁴ and Alzheimer's disease^15-17 adding plausibility to the findings of Choi et al.¹³ The DMN has been implicated in higher-order cognitive processing relating to self-relevant information¹⁸, social cognition¹⁹ and co-activation in cognitive transitions.²⁰ The DMN consists of the posterior cingulate cortex (PCC), precuneus, bilateral inferior parietal cortices (IPC), medial prefrontal cortex (mPFC) and sub-set of medial temporal regions²¹, which has been functionally differentiable into sub-networks.²¹ We sought to establish if preoperative rsFC may predict vulnerability to postoperative delirium. We hypothesized that the DMN would have altered preoperative FC (weaker) in the postoperative delirium population compared the non-delirious population. As secondary analyses, we explored the role of other networks in vulnerability to delirium. Notably, given inattention is a core feature of delirium diagnosis, we were interested in determining if attention networks showed impaired connectivity preoperatively.

Materials and methods

Clinical Data

All data was collected as part of the perioperative cohort study, Interventions for Postoperative Delirium: Biomarker-3 (IPOD-B3), approved by University of Wisconsin-Madison Institutional Review Board (2015-0374, NCT01980511 and NCT03124303). All participants provided informed consent, and work was conducted in accordance with the Code of Ethics of World Medical Association (Declaration of Helsinki) for experiments involving humans. 5519 individuals were screened, and 296 patients were enrolled for participation in the perioperative cohort study (see supplementary methods for details of inclusion). All participants were aged 65yrs or older and underwent major elective non-intracranial surgery. Exclusion consisted of any documented dementia or residing in a nursing facility. Preoperatively participants underwent cognitive assessment, blood sample collection, and preoperative MRI scanning (optional - see supplementary methods). Cognitive assessment included a battery of tests comprising the Montreal cognitive assessment (MoCA), trails making tests A and B, and the digital substitution test (DSST). Other data, including preoperative EEG data, were collected but are not further described herein^4,5. After surgery, participants underwent delirium assessments twice daily using the 3D-Confusion Assessment Method (3D-CAM or CAM-ICU (dependent on whether patients were intubated or not) to define delirium. Additionally, in patients who were able to verbalise, delirium severity was assessed by delirium rating scale 98 (DRS) and 3D-CAM-severity score. All assessors were trained in 3D-CAM²², CAM-ICU²³ and delirium rating scale delirium assessments. Assessments were discussed in weekly meetings to review diagnoses and harmonise delirium scoring. Sedation (i.e., coma) was assessed by Richmond Agitation/Sedation Scoring (RASS).²⁴ All patients who scored positive for coma were marked as delirious, and later scored as delirious. Daily blood sample collection was also conducted (see supplementary methods).

Nomenclature

These analyses compare brain network connectivity before surgery (preoperative) between participants who later developed postoperative delirium (termed delirious) and who did not develop postoperative delirium (termed non-delirious). To be clear, the scanning was conducted prior to surgery, when no patient was actually delirious, and the data are used to predict whether a patient later became delirious or not.

Functional MRI acquisition and preprocessing

Participants were recruited into an optional substudy of the cohort and scanned across five different sites (see supplementary methods). An MRI scanner was used to acquire T1-weighted axial structural images at beginning of each scan using FSPGR BRAVO sequence; repetition time = 8.132ms; echo-time = 450ms, 256 x 256 matrix, 156 slices, flip-angle = 60°, and slice thickness = 1mm. A T₂-weighted echo-planar functional images, acquired in sequential order with: repetition time = 2600ms; echo-time = 22ms; flip-angle = 60°; 40 axial slices covering the entire brain; interslice gap = 3.5 mm; field of view = 224mm; and the raw voxel size = 1 x 1 x 1mm. A high-resolution 3D T₁-weighted anatomical image with voxel size = 3.5 x 3.5 x 3.5mm was obtained for co-registration with functional scans. Pre-processing of images was performed using fMRIPrep (version stable 20.0.2²⁵), run through NiPype (version 1.8.1²⁶) (see supplementary methods). All BOLD signal was denoised by regressing out the following noise parameters 12 measures of head motion and their corresponding derivatives, white-matter and CSF measures and frame-wise displacement and a high and low-pass filter (0.01 Hz & 0.1 Hz) was applied.

Functional MRI Analysis

Region of Interest Selection and Functional connectivity

The denoised functional data underwent BOLD time-series extraction for the 17-network 400-regions of interest (ROI) defined by the Schaefer cortical parcellation.²⁷ We categorised the regions into Yeo’s 7-network grouping of the 400 Schaefer atlas parcellation (with an additional network for tempo-parietal regions from the 17-network label). To determine the FC for each participant, the Pearson’s correlation was calculated for each average BOLD time-series for each unique ROI-to-ROI connection.

We analysed whether there were average group differences in within DMN FC for postoperative delirium compared to non-delirium. Additionally, we categorised the DMN into sub-networks defined by; into the core, dorsomedial PFC and medial temporal lobes²¹.

Cortical Network Connectivity Analysis

For each participant, we calculated the average FC within and across the above-mentioned predefined cortical networks. Group-wise differences in within- and across-network connectivity were assessed by comparing average connectivity values between the two groups. Importantly, we focused only on the functional connections across the entire cortex that showed significant differences between groups in the permutation test. Additionally, we calculated (for each subject) the average significant within-network connectivity and generated boxplots to depict the average differences in within-network connectivity for the predisposed and non-predisposed group.

Predicting Delirium Using Support Vector Machine

To further explore which specific functional networks are predictive of postoperative delirium, we employed a support vector machine (SVM) with a linear kernel (using scikitlearn, Python Version 3.12) for ease of interpretability. The SVM was trained to estimate which features of the cortical functional networks were most predictive of delirium. We trained the model on a held-out subset of the participant data and tested on the 40% remaining held-out set of participant data (n=48). After selecting the appropriate hyperparameter (see supplementary methods), we ran the model to find the overall accuracy of the linear SVM in predicting the postoperative delirium outcome from the average within-FC across cortical networks. The model provided a ranking of functional cortical networks connections based on their predictive value, with a strength in the coefficient (β coefficient value), being an indicator of the predictive value of the average within network connectivity. We used an arbitrary selection of the peak ß coefficient values to assess the following network-specific analysis.

Network-Specific Analysis

After identifying significant group differences in FC between visual, salience-ventral attention (SVN), and DMN, we performed additional analyses to examine the specific connections between these cortical networks and the rest of the brain. This allowed us to investigate the global FC patterns associated with these networks, providing a deeper understanding of their role in postoperative delirium.

Statistical Analysis

To determine significant differences in FC between the two groups, we performed a permutation-based independent samples t-test. Group labels were randomly shuffled, and the t-test was conducted for each permutation. The process was repeated for a large number of iterations (e.g., 1,000 permutations) to generate a null distribution, allowing us to identify significant group-wise differences in connectivity at each unique ROI-to-ROI edge. This approach controlled for multiple comparisons across all pairwise connections between cortical regions.

Results

Of a total of 142 scanned participants, 22 participants were removed from further analysis due to withdrawal, incomplete assessments and substantial denoising errors in fMRI (see Fig. 1 and supplementary methods). Among scanned participants, most patient characteristics were not different between delirious and non-delirious participants (see Supplementary Table 1), except for NSQIP-SC, NSQIP-D (indicating more severe surgery), surgical timing and Overall Peak DRS (highest received DRS score for multi-day assessment). Preoperative MRI scanning occurred in 142 participants in the entire cohort, the non-scanned participants (n=153) had significantly higher ASA (P=0.007), NSQIP-SC (P<0.001) and NSQIP-D (P=2.93x10⁻⁵) scores compared to non-scanned participants (see Table 1 and Supplementary Table 2).

Figure 1: Strobe plot of relevant participants in study.

Strobe figure depicts the relevant participant data from recruitment to participants that have been analysed in this study.

Table 1:

Patient Clinical data for entire recruitment for study divided into postoperative delirium or non-delirious for both MRI scan and not scanned participants.

	Non-scanned		All Scanned
	Delirious (n=40)	Non-delirious (n=108)	Delirious (n=33)	Non-delirious (n=102)
Age, YR	71 (67.75 – 74.25)	71 (68 - 75)	71 (69 - 75)	71 (68 – 75)
Sex	15 Female	35 Female	16 Female	41 Female
Education
<12yr	0	3	2	2
12 yr	10	29	8	18
>12yr	28	76	23	82
TMT-A	35 (29.5 – 45.5)	32 (27 – 40.75)	34.50 (30.75 – 41.50)	34 (28 - 43)
TMT-B	81.5 (± 57.75 – 97.75)	66 (52.50 – 89.50)	81 (66 - 102)	66 (53.25 - 94)
MoCA	23 (22 - 24)	23 (22 - 25)	24 (22 - 26)	23 (21 - 26)
DSST	40 (34 - 52) *	48.5 (40.75 - 55) *	46 (37 - 52)	47 (37.25 – 54.75)
ASA	3 (3 – 3.25)	3 (3 -3)	3 (2 - 3)	3 (2 - 3)
IL-8	15.69 (5.43 – 53.19) * (n= 22)	2.86 (−0.105 – 10.93) * (n=28)	7.98 (1.30 – 21.15) * (n = 24)	2.40 (−0.2 – 6.24) * (n= 65)
IL-10	28.23 (6.27- 74.60) * (n= 22)	3.05 (0.58 – 13.64) * (n=28)	10.74 (6.02 – 40.80) * (n= 24)	1.56 (−0.31 – 7.37) * (n= 65)
Surgical Operative Duration (mins)	469 (382.25 – 626.75) **	300 (226.50 - 394) **	331 (221 - 529) *	218 (173 - 341) *
Surgical Categories	V = 13; T =1; G= 5; O = 4; U = 2; C = 15	V= 24; T = 11; G = 20; O = 6; S = 4; U = 23; C=23	V = 11; T =0; G=3; O =8; U = 2; C = 8	V = 23; T = 4; G = 14; O = 39; S = 2; U = 4; C = 18
NSQIP-SC	24.2 (13.10 – 33.40) *	16.3 (11.25 – 25.15) *	22.25 (11.55 – 31.38) *	11.2 (7.30 – 15.80) *
NSQIP-D	3.6 (1.80 – 6.10) *	1.35 (0.68 – 3.43) *	2.7 (0.675 – 5.43) *	0.7 (0.20 – 2.60) *
Overall Peak DRS	16.5 (15 – 21.50) *	10 (9 - 12) *	19 (13 - 22) *	10 (8 - 12) *

Clinically relevant data listed. Standard Deviation (SD). Years (YR). Montreal Cognitive Assessment (MoCA). Trail Making Test A/B (TMT-A and TMT-B). Digit Symbol Substitution Test (DSST). American Society of Anaesthesiologists Physical Status Classification (ASA). Interleukin-8 (IL-8) and Interleukin-10 (IL-10), calculated as postoperative day 1 minus baseline recording. National Quality Improvement Program for surgical risk of serious complications (NSQIP-SC). American College of Surgeons’ National Quality Improvement Program for surgical risk of death (NSQIP-D). Note inflammatory biomarkers were calculated as the difference in postoperative and baseline within a subset of participants. Surgical Categories as follows; V= Vascular, T= Thoracic, G = General, O = Orthopedic, S = Spinal, U= Urological/Gynaecological, C = Cardiac.

^*p<0.05 significance determined by Mann Whitney U-Test (as one set of data was not normally distributed), for between not scanned postoperative delirious and non-delirious in DSST scores p=0.01784, IL-8 difference p=0.0008, IL-10 difference p=0.0006, overall peak DRS p= 2.2e⁻¹⁶, NSQIP-D p=0.001, NSQIP-SC p=0.013. For between scanned postoperative delirious and non-delirious IL-8 difference p=0.022, IL-10 difference p=6.29e⁻⁵, surgical operative duration p=0.0002, overall peak DRS p= 2.76e⁻³⁶, NSQIP-D p=0.003, NSQIP-SC p=0.0002.

^**p<0.05 significance determined by Two Sample T-Test (data normally distributed), for between not scanned postoperative delirious and non-delirious surgical operative duration p=2.42e⁻⁷.

Default Mode Network Connectivity Differences

The analyses compare brain network connectivity before surgery (preoperative) between participants who developed postoperative delirium (delirious) and who did not develop postoperative delirium (non-delirious). We found an overall weaker within-DMN functional connections (Fig. 2B-C) in delirious participants. Specifically, between sub-regions of the DMN from the dorsal prefrontal cortex (PFC) to temporal, and dorsal PFC to medial PFC (Fig. 2C). There are differences within sub-networks DMN connectivity (Fig. 2D); with weaker functional connections within the dorsomedial PFC for delirious participants compared to non-delirious (Fig. 2D). In addition, there are increased functional connections across the medial temporal sub-networks of the DMN (Fig. 2E); specifically, between the retrosplenial and inferior parietal lobes of the medial temporal sub-network (Fig. 2C).

Figure 2: Default Mode Network Connectivity differences between delirious and non-delirious participants.

A) The anatomical representation of the different regional divisions of the DMN. The core sub-network includes posterior cingulate cortex and precuneus; the dorsomedial prefrontal cortex sub-network includes the dorso-medial PFC, temporoparietal junction, lateral temporal cortex and temporal pole; and the medial temporal lobes sub-network includes the ventral medial PFC, posterior inferior parietal lobule, retrosplenial cortex and parahippocampal regions. B) Average within FC for the delirium (orange) and non-delirious participants (blue). C) Chord plot showing the significant difference in functional connections, delirium minus non-delirious, negative relationship indicated by darkened lines of the region. D) The anatomical representation of the sub-networks of the DMN. E) Significantly different FC clustered into the sub-networks of the DMN, delirium minus non-delirious participants. Significance determined by permutation test with 1000 iterations (P<0.05).

Whole-Brain Functional Network Connectivity Differences

We expanded our analysis to determine whether there are significant differences in the ROI-to-ROI FC across all 400 regions in the Schaefer parcellation between the delirious and non-delirious participants. We found weaker FC across several cortical networks (Fig. 3A, indicated by pink colours), including the somato-motor, salience-ventral attention, cognitive control and DMN for participants who developed delirium compared to those that did not. There was a stronger FC within the visual network for participants who developed delirium (Fig. 3A). Despite no significant differences in the attentional cognitive tasks between the participants who developed delirium and those that did not (see Supplementary Fig. 1A); we did observe a reduced FC with the salience-ventral attentional networks related to poorer performance on both the attentional cognitive tasks (see Supplementary Fig. 1B); further highlighting the salience-ventral attentional networks importance in cognitive processing.

Figure 3: Whole-Brain FC Edges Significantly related to delirium outcome & Within Network Connectivity.

A) Connectivity matrix of the significant average difference in FC for each region (delirium – non-delirious), (permutation, 1000 iterations, P<0.05). B) The weighted coefficients from a linear supervised vector machine (c=0.1), of the average within/across functional connections of cortical networks features that are predictive of delirium outcome. For the following networks; Visual (C), Somato-motor (D), Salience Ventral Attention (E), Dorsal Attention (F), Control (G), Limbic (H), Temporal-parietal regions (I). i) Anatomical brain plots of specific regions within the network. ii) Boxplot of the average within FC of the network for each participant split into groups; non-delirious indicated by blue and delirium indicated by orange. iii) Chord plots of the difference of average functional connections between sub-regions of the network for predisposed minus non-predisposed (significance determined through permutation testing P<0.05); note – darkened colour lines indicate negatively weighted connections, directionality of these connections plotted is not inferred.

To further determine which specific cortical networks are more predictive of delirium, we used a linear SVM to determine which cortical networks were features in determining postoperative delirium (Fig. 3B). The overall accuracy of the linear SVM was 68% in predicting postoperative delirium outcome across cortical networks connectivity. We found that the visual network had a positive β coefficient of 0.47, indicating that an average stronger connectivity within the visual network is predictive of delirium. The salience-ventral attentional network was negatively associated (β = −0.26) indicating that a weaker connectivity within the salience-ventral attentional network is more predictive of delirium (Fig. 3B). Additionally, we ran the linear SVM with an additional clinical predictor, the inflammatory marker IL-8 & IL-10 (Day 1 postoperative IL-8 & IL-10). The model’s overall accuracy was 56% and 70% respectively; and the overall strength in the β coefficients for across the cortical networks were likely reduced due to the IL-8/IL-10 inflammatory marker having a stronger predictive powers (especially in the case of IL-10 which dominated with a weighted β coefficients of 1.022); however the SVM still predicted that both the visual and the salience-ventral attentional network were more predictive with these inflammatory biomarker inclusion (see Supplementary Fig. 2).

To reduce the dimensionality of the data for interpretability, we calculated the average FC in each cortical network (FC to regions within the network) that had passed significance from above-mentioned whole-brain FC. Then, we determined group differences between the participants who developed delirium compared to those that did not by group t-tests. There is a slight increase in average within-network connectivity across the visual network for participants who developed delirium (Fig. 3C ii), from frontal eye fields to posterior central, parietal to temporal occipital regions (see light grey colours in chord plot Fig. 3C iii) (P>0.05, group comparison). The somato-motor has on average a weaker within-network connectivity in participants who developed delirium (Fig. 3D ii, iii). The average within FC of the salience-ventral attentional network is significantly weaker for participants who developed delirium compared (Fig. 3E ii, iii). The dorsal attention network had no-relative difference, but for significantly plotted differences (for within network connections) there is a relative stronger connectivity for participants who developed delirium compared to those that did not (Fig. 3F, iii indicated by light grey colours in chord plot), despite an overall average weaker within-network connectivity for the delirium group (Fig. 3F ii). The cognitive control network on average had a weaker within-network connectivity for participants who developed delirium (Fig. 3G ii), across multiple regions (Fig. 3G iii). The limbic network and tempero-parietal regions did not have any noticeable differences (Fig. 3H and G).

Select Cortical Networks to the Rest of the Brain Connectivity

To further establish the role of dysFC in postoperative delirium we extended the FC analysis to focus on three specific cortical networks that were both shown to be predictive of delirium in the linear support vector machine (selected based on their peak β coefficient value) and had plausible neurobiological mechanism of preoperative vulnerability. The significantly different connections from the DMN to the rest of the brain plotted across the cortical surface indicates an overall weaker connectivity for the participants who developed delirium compared to those that did not (Fig. 4A i). The DMN had weaker connectivity across all other cortical networks (Fig. 4A ii), specifically the dorsal and medial pFC to the visual network, somato-motor and across both the salience-ventral and dorsal attentional networks (Fig. 4A iii, iv). Similarly, the average significantly different connections from the salience-ventral attention network (SVN) to the rest of the cortex indicates an overall weaker connectivity for participants who developed delirium compared to those that did not (Fig. 4B i). The SVN had specific decreased connectivity from the insula and parietal-medial regions of the SVN to the dorsal attentional network (Fig. 4B iii, iv); and the lateral pFC region of the SVN to the visual network (Fig. 4Biii, iv). Lastly, the entire visual network connectivity to the rest of the brain had weaker connectivity, specifically to SVN and the DMN (Fig. 4Ci ii) in the participants who developed delirium compared to those that did not, but had stronger connectivity with the dorsal attentional network (Fig. 4C iv).

Figure 4: Cortical Network Connectivity differences across the whole brain.

A)The significantly different functional connections between the Default Mode network to the other cortical networks. B) Significantly different functional connections between Salience Ventral Attentional networks to other networks. C) Significantly different functional connections between Visual Network to other networks. i) The average differences in functional connections between delirium minus non-delirious for a specific network connections to the whole brain. ii) The average differences in FC matrix from specific network to whole brain. iii) Chord plot of the significantly different connections for specific network to whole brain; note – darkened colours indicate negatively weighted connections. iv) Chord plot of the significantly different connections for the specific network to the whole brain, darken grey lines indicate negative connnections (significance determined through permutation testing P<0.05).

Discussion

In this study, we found that the DMN had significantly reduced within- and across network FC in the participants who developed delirium compared to those that did not. Upon further investigation of cortical FC, we established several other cortical networks that had significantly differentiated the groups, particularly the salience-ventral attentional network.

The finding of reduced within cortical network connectivity is consistent across other studies that have evidenced reduced connectivity within networks associated with cognitive impairment.²⁸ Our findings are compatible with the theory of ‘cognitive disintegration’ which conceptualises that delirium patients have a pre-existing neural vulnerability due to instability in their cortical network architecture.³ In the context of postoperative delirium onset, it suggests that these individuals develop delirium because of a prior neuronal vulnerability, this fits with previous work that revealed decreased structural connectivity in patients at risk of postoperative delirium^{5, 8, 12}, and implies a prior underpinning of cortical ‘disintegration’ across the brain’s connectome could lead to postoperative delirium. However, our data do call for refinement of the cognitive disintegration theory, as our data suggest decreased rsFC within some networks, while other networks show increases in connectivity.

Previous literature has established that the DMN is an important cortical network that is involved in several high-ordered processes including self-ideation, imagination, self-generated cognition and mind-wandering.¹⁹ Interestingly, our data suggest the dorsomedial PFC region of the DMN appears to be the critical sub-network of the DMN that differentiates participants who developed delirium compared to those that did not – this is of interest given that the dorsomedial PFC region of the DMN is involved in metacognition, self-referential, and social cognitive processes.²¹

In addition to differences in the DMN, we observed differences in the salience ventral attentional networks (SVN) which is consistent with previous evidence of connectivity differences in the SVN and DMN during postoperative delirium.²⁹ This is important as the SVN mediates different modes of attention, specifically in allocating adequate attentional resources to high-order processing of context dependent sensory information.³⁰ Given the central role of attention in delirium diagnosis, this finding – and the impaired connectivity of the SVN with other networks, could be key to understanding how reorientation of attention and cognition is impaired in delirium. We further investigated whether changes in rsFC within the attentional networks were related to performance on attentional cognitive tasks. Although we did not specifically test attention in a task-based paradigm in the scanner, we did reveal that significantly reduced FC within the salience ventral attentional network related to poorer performance on both attentional cognitive tasks (see Supplementary Fig. 1B).

The visual network had significantly increased within network connectivity in the participants who developed delirium compared to those that did not. Notably, we previously observed that participants who were vulnerable to delirium showed exaggerated feedforward responses from primary auditory cortex in response to sensory stimulation. ⁴ In concert with these findings, hyper-connected primary sensory regions may indicate vulnerability to delirium.

A limitation of this study was the selection bias in the participants that were scanned versus non-scanned, which indicates that non-scanned participants are likely undergoing more severe surgery, and/or have more severe underlying comorbidities. While this primarily relates to participant choice, some of the exclusions from the imaging sub-study related to contraindications to scanning, however this still limits the generalisability of our work. Adjustment for inflammation, as a key precipitant of delirium, mitigates the risk of confounding from varying surgeries which is especially important given the imbalance of surgical severity (and therefore inflammation between the groups). Additionally, there are limitations relating to other unknown confounds that may affect the results. Furthermore, a range of scanners were used in this study, and this can introduce further confounds in the results. However, in terms of other cognitive metrics that have previously been shown to predict delirium³¹ there are no major differences both between the participants who developed delirium compared to those that did not (see Supplementary Table 1) and the scanned/non-scanned participants (see Table 1). Hence, this maybe considered a strength of this study as we could elucidate whether there are inherent rsFC vulnerabilities prior to undergoing surgery with matched cognition. In addition, our findings are contradictory to previous work that has shown that predisposing risk factors (e.g. age, cognitive impairment, depression) for delirium were not associated with delirium-related fMRI network changes.³² However, it is important to note that that study used a minimum spanning tree metric which resulted in only positive correlations being quantified in functional networks. Given these differences in methodological approaches, it is difficult to compare directly between our findings and theirs. Future studies should attempt to incorporate multi-modal imaging, such as simultaneous PET-fMRI and structural imaging to determine the potential underlying neuropathological pathways that could drive the differences in preoperative FC between the participants who developed delirium compared to those that did not

Overall, we provide supportive evidence for dysfunctional cortical network connectivity in individuals prior to delirium onset. This is consistent with previous literature and highlights the importance of considering the specific neuronal pathways that contribute to postoperative delirium when developing mitigation strategies. Future studies should investigate plausible neurobiological mechanisms of this vulnerability, specifically with the use of modern PET-scanning for amyloid/tau-pathology.

Abbreviations:

rsFC

resting-state FC

fMRI

functional MRI

DMN

Default mode network

BOLD

Blood oxygen level dependent

Data availability

The data that support the findings of this study are available upon reasonable request from the corresponding author. The data are not publicly available due to clinically relevant data under ethics restrictions. All code for the analysis is available: https://github.com/NatashaLTaylor/Preoperative_fMRI_FunctionalNetworks