Cortical reactivity to transcranial magnetic stimulation predicts risk of post-stroke delirium

Objective: Post-stroke delirium (PSD) is a frequent and with regard to outcome unfavorable complication in acute stroke. The neurobiological mechanisms predisposing to PSD remain poorly understood, and biomarkers predicting its risk have not been established. We tested the hypothesis that hypoexcitable or disconnected brain networks predispose to PSD by measuring brain reactivity to transcranial magnetic stimulation with electroencephalography (TMS-EEG). Methods: We conducted a cross-sectional study in 33 acute stroke patients within 48 hours of stroke onset. Brain reactivity to single-pulse TMS of dorsolateral prefrontal cortex, primary motor cortex and superior parietal lobule of the right hemisphere was quantiﬁed by response intensity, effective connectivity, perturbational complexity index (PCI ST ), and natural frequency of the TMS-EEG response. PSD development was clinically tracked every 8 hours before and for 7 days following TMS-EEG. Results: Fourteen patients developed PSD while 19 patients did not. The PSD group showed lower excitability, effective connectivity, PCI ST and natural frequency compared to the non-PSD group. The maximum PCI ST over all three TMS sites demonstrated largest classiﬁcation accuracy with a ROC-AUC of 0.943. This effect was independent of lesion size, affected hemisphere and stroke severity. Maximum PCI ST and maximum natural frequency correlated inversely with delirium duration. Conclusions: Brain reactivity to TMS-EEG can unravel brain network states of reduced excitability, effective connectivity, perturbational complexity and natural frequency that identify acute stroke patients at high risk for development of delirium. Signiﬁcance: Findings provide novel insight into the pathophysiology of pre-delirium brain states and may promote effective delirium prevention strategies in those patients at high risk.


Introduction
Delirium is an acute neuropsychiatric complication, which results in a fluctuating disturbance in attention, awareness and cognition (American Psychiatric Association, 2013;Inouye et al., 2014), increases mortality, prolongs hospitalization and adversely affects functional outcomes following critical illness ( et al., 1998;Khan et al., 2012;McCusker et al., 2002;Siddiq et al., 2006). Acute stroke is a known risk factor for the development of delirium (Gustafson et al., 1991;McManus et al., 2009). Although only limited published data on post-stroke delirium (PSD) exist, proportions of 10-27 % (Caeiro et al., 2004;Oldenbeuving et al., 2013) or higher (up to 48 %) (Dahl et al., 2010;Langhorne et al., 2000;Oldenbeuving et al., 2007) of PSD were indicated during the first critical weeks after stroke onset, making it a frequent and severe complication after stroke (Kiely et al., 2009;Shaw et al., 2019). Effective management, treatment and prevention of PSD may help to decrease incidence of mortality, disability (Rice et al., 2017;Song et al., 2018) and reduce clinical burden (Leslie et al., 2008;Maldonado, 2013). However, not much towards these aims has been achieved so far because detailed neural mechanisms underlying PSD are largely unclear, and markers to accurately predict PSD have not been developed.
Exploring the brain mechanisms underlying PSD will contribute to identifying patients who are at risk and aid clinicians in early initiation of delirium prevention and treatment. While the brain mechanisms of PSD are likely complex, several hypotheses from a molecular-biological perspective (Hshieh et al., 2008;van Munster et al., 2010) or an electrophysiological perspective (Wiegand et al., 2022) have been proposed to explain delirium development of mixed patient populations on the intensive care unit (ICU) (e.g., elderly patients post-surgery or with traumatic brain injury). In addition, neuroimaging studies highlighted altered functions of predominantly right-hemispheric cortical and subcortical networks for attention and arousal (Boukrina and Barrett, 2017) in delirium development. Still, the neurobiological basis and conditions involved in inducing a delirious brain state after stroke remain poorly understood.
Delirium prediction, which would facilitate early recognition of patients at high risk for delirium is, however, critical for clinical decision making and setting of priorities regarding the use of delirium preventive measures. In addition to usual precipitating factors for general ICU patients (van den Boogaard et al., 2014;Wassenaar et al., 2015), PSD is more likely to be dependent on stroke-related factors, such as stroke lesion location, lesion size, cerebral hypoperfusion and cerebral edema (Kostalova et al., 2012;Ojagbemi et al., 2017;Pasinska et al., 2018). Therefore, specialized neural assessment techniques (i.e., neurophysiology and neuroimaging) should be able to shed light on neural mechanisms underlying PSD and its prediction.
Transcranial magnetic stimulation (TMS)-EEG is a relatively novel technique, which provides a way for directly probing both local and widespread changes in brain neurophysiology, through the recording of TMS-evoked potentials and TMS-induced cortical oscillations. TMS-EEG can obtain important information regarding local excitability at the site of stimulation, and effective connectivity of both discrete and connected regions of the cortex, as well as providing insight into the oscillatory properties of brain networks (Tremblay et al., 2019). Compared to resting-state EEG, TMS-EEG enables direct and non-invasive exploration of cortical reactivity of the brain to external perturbation, with excellent temporal resolution. With regard to stroke patients, TMS-evoked brain responses were used to evaluate alterations in cortical reactivity (Sarasso et al., 2020), its reorganization (Pellicciari et al., 2018), and provided individual readout for prediction of motor recovery (Tscherpel et al., 2020).
Accordingly, we propose here that TMS-EEG-related neurophysiological markers of brain reactivity are capable of refining our understanding of the neurobiological basis of PSD. We hypothesize that patients with reduced excitability and connectivity are at high-risk for development of PSD (Shafi et al., 2017).

Subjects
Thirty-three acute stroke patients (seventeen females; mean age ± SD: 78.4 ± 10.7 years, range: 52-93 years) (Table 1), who were admitted to the stroke unit of the department of Neurology & Stroke of the University Hospital Tübingen, were recruited into the study. Patients were enrolled based on the following inclusion criteria: (i) age >50 years in order to recruit a representative sample of stroke patients; (ii) acute ischemic or hemorrhagic stroke at supra-or infratentorial regions, verified by diffusion-weighted MRI (DWI); (iii) within the early acute stage, 48 hours after symptom onset; (iv) written informed consent obtained from the patients or their custodians prior to enrollment. Exclusion criteria were: (i) Richmond Agitation-Sedation Scale scores higher than 3 or lower than À4 at the time of the TMS-EEG measurement; (ii) PSD has already developed, or history of delirium, seizure or traumatic brain injury within the last three months; (iii) any contraindications to MRI or TMS (e.g., pacemakers, intracranial stents, epilepsy); (iv) central nervous system active drugs (e.g., benzodiazepines, neuroleptics) in the week prior to TMS-EEG measurement; (v) no detectable motor evoked potentials (MEP) in hand muscles with TMS of the primary motor cortex (M1) of either hemisphere, even at maximum stimulator output (MSO). The study was approved by the Ethics Committee of the University Hospital of Tübingen (protocol number 147/2020BO1).

PSD diagnosis
Delirium screening for the patients was conducted using the Intensive Care Delirium Screening Checklist (ICDSC), which was administered on admission and every 8 hours (morning, late and night shifts) for 7 days after the TMS-EEG measurement, by welltrained neurocritical care nurses. ICDSC scores of !4 for nonaphasic, or !5 for aphasic patients were considered indicative of PSD. Then, the diagnosis of PSD was confirmed in accord with the Diagnostic and Statistical Manual of the American Psychiatric Association (DSM-V) criteria by an independent neurologist, blinded for the ICDSC scores and TMS-EEG results. The delirium onset time, duration and end time were recorded (Table 1).

TMS-EEG data recording and navigation
TMS-EEG data was acquired by a TMS-compatible EEG system (BrainAmp 64 actiCHamp Plus, BrainProducts GmbH, Munich, Germany). The EEG cap was equipped with TMS-compatible C-ring slit Ag/AgCl pin electrodes arranged in the International 10-20 montage. The EEG amplifier was set with a hardware filter at DC to 1 kHz and a sampling rate of 5 kHz. The skin/electrode impedances of all electrodes were maintained below 5 kX throughout the data recordings. TMS pulses were delivered through a MagVenture (MagPro Compact, MagVenture A/S, Denmark) magnetic stimulator with a monophasic current waveform. A stereoscopic neuronavigation system (Localite GmbH, St. Augustin, Germany) was used, based on individual anatomical MRI to enable precise positioning of the TMS coil relative to the individual brain anatomy. To define the standardized localization of the TMS targets, the individual brains were projected to a template according to the Montreal Neurological Institute (MNI) coordinate system based on the anatomical positions of the anterior commissure, posterior commissure and one point on the falx cerebri.
Prior to the TMS-EEG recordings, the individual resting motor threshold (RMT) in the right M1 was determined by MEP recordings (Tscherpel et al., 2020) (see details in Supplementary Materials). RMT of 3 patients (Table 1) was determined in the left Right-hemispheric cortical regions, including dorsolateral prefrontal cortex (DLPFC), primary motor cortex (M1) and superior parietal lobule (SPL) were shown to have a close association with delirium (Boukrina and Barrett, 2017). Accordingly, three righthemispheric TMS targets were defined and set at the individually MNI-fitted images according to the MNI coordinates (DLPFC: x = 38, y = 19, z = 51; M1: x = 51, y = À8, z = 44; SPL: x = 19, y = À54, z = 64; marked as red dots in Fig. 1 A-C) to make stimulation sites comparable between patients. Importantly, the individual target sites were evaluated prior to each TMS measurement by a neurologist according to the individual DWI information on the ischemic stroke lesion. The individual target sites close (<2 cm) to lesions (M1 in one non-PSD and one PSD patient, DLPFC in one non-PSD patient) were skipped in order to avoid impact on our findings from perilesional cortical off-periods (Sarasso et al., 2020) or inexcitable cortex (Gosseries et al., 2015).
During the TMS-EEG recordings, single-pulse TMS at an intensity of 90 % RMT was delivered at the three targets in separate blocks of 200 trials, with jittered inter-trial intervals of on average 2 s. The order of the targets was pseudo-randomized and balanced across patients. TMS coil was set and optimized at each target under online neuronavigation to induce an electric field perpendicular to the main axis of the targeted gyrus. To investigate whether the stimulation intensity was comparable within and between patients, distribution and intensity of the intracranial electric field induced by TMS were calculated at each target site of each patient using the SimNIBS toolbox (Saturnino et al., 2019). The mean and navigated TMS target sites (red dots). Spatial distribution plots represent average cortical activity of all patients (absolute z-transformation relative to the baseline [À300 ms À50 ms]) elicited over the first 100 ms after the TMS pulse. Butterfly plots: group average cortical reactivity (y-axes: linearly constrained minimum variance (LCMV) values) evoked by TMS (time, 0 ms) of DLPFC (A), M1 (B) and SPL (C) of non-post-stroke delirium (non-PSD) group (upper plots) and the PSD group (lower plots). The black traces indicate TMS evoked response in the target regions underneath the TMS coil (defined as dark blue in the left column). The response intensity was defined as summarized absolute values of significant (bootstrap nonparametric statistics with p < 0.01) cortical responses between 20 ms to 300 ms. Each gray trace indicates a TMS evoked responses from one brain region. Spatial activation maps show the difference of cortical response intensity between non-PSD and PSD groups evoked by TMS of DLPFC (D), M1 (E) and SPL (F). Yellow colors indicate higher values in the non-PSD group. Black colors in the underneath cortex plots show regions with significant difference (Mann-Whitney U-test, FDR correction with p < 0.05) of response intensity between non-PSD vs PSD groups.
Clinical Neurophysiology xxx (xxxx) xxx ( 1SD) strengths of the electric fields were calculated at the defined targets and compared between non-PSD vs PSD groups. Importantly, there were no significant group differences (non-PSD vs PSD): 99.8 17.2 V/m vs 100.1 12.6 V/m at DLPFC, 91.4 13.7 V/ m vs 96.6 15.4 V/m at M1, 81.6 12.7 V/m vs 85.9 14.5 V/m at SPL, and all individual electrical field strengths were sufficient (>40 V/m) to evoke distinct cortical reactivity (Rosanova et al., 2009). The methodological details are given in the Supplementary Materials ( Figure S5 and related content). Although noise masking is always suggested, long-time exposure with high-decibel (up to 90 dB) white noise turned out to be intolerable for most of the acute stroke patients, and might even constitute a risk factor for inducing delirium (Kalish et al., 2014). Therefore, we chose earplugs but not noise masking during TMS-EEG recording. Before performing further analysis on TMS-EEG, we investigated the cortical response elicited by the TMS clicks, which is represented by a wave component in the central region between 100-200 ms (Rocchi et al., 2021). We extracted the cortical auditory evoked potential (AEP, N1: 80-120 ms and P2: 160-200 ms) by averaging the TMS evoked potentials recorded over FCz and the eight surrounding electrodes (F1, Fz, F2, FC1, FC2, C1, Cz and C2) (Vallesi et al., 2021). The N1 (80-120 ms following the TMS pulses) and P2 (160-200 ms) components of the AEP and their topographies with TMS of DLPFC, M1 and SPL are shown in Supplementary Material ( Figure S6). N1 and P2 were reconstructed in source space. Cortical areas of interests for cortical AEP analysis, including Heschl's gyrus, Brodmann area 22 and planum temporale, were identified in MNI coordinates, in accord with previous AEP localizing studies (Gascoyne et al., 2016;Godey et al., 2001) and mapped onto the Brainnetome Atlas (Supplementary Materials, Figure S7A). Then, the absolute values of cortical AEP amplitudes in these cortical areas of interests were reconstructed for each patient using a linearly constrained minimum variance (LCMV) beamforming method (Sekihara and Nagarajan, 2008). The AEP-related cortical responses were extracted, by averaging values in the areas of interests, within the period 80-120 ms for N1, and 160-200 ms for P2.

Data analysis
Preprocessing of TMS-EEG (Rogasch et al., 2014) were performed using customized analysis scripts on MATLAB (Version 2017b, MathWorks Inc., Natick, USA) and EEGLAB 14.1.2b. A LCMV beamforming method (Sekihara and Nagarajan, 2008) was used to perform source reconstruction based on the FieldTrip toolbox and FreeSurfer (see methodology details of data preprocessing and source reconstruction in Supplementary Materials).

TMS evoked potentials (TEPs)
To estimate the neuronal responses elicited by TMS, we projected the source-reconstructed responses into the human Brainnetome Atlas (Fan et al., 2016) containing 246 regions of interest (ROIs) across both hemispheres. The activity for each of these ROIs was estimated by taking the first component from a principal component analysis performed on the time course of dipoles included in the ROIs. Based on the Atlas, we defined the DLPFC, M1 and SPL, according to their spatial positions (marked as dark blue areas in Fig. 1A-C). In order to validate the accuracy of the navigation procedure and support the choice of the ROIs, we estimated the intensity distribution of the neuronal responses elicited by TMS. The neuronal response of each dipole after source reconstruction was firstly normalized by z-transformation relative to the baseline (À300 ms to À20 ms) and then averaged over the first 100 ms (21 ms to 100 ms, first 20 ms discarded in order to avoid possible non-neuronal activities) after TMS.

Effective connectivity
TMS evoked effective connectivity was measured by detecting directional information flow elicited (time-locking) by the TMS pulse. Symbolic transfer entropy was calculated in pairs of TEPs (21 ms to 400 ms) of ROIs with recommended parameter settings (number of symbols: 4, step interval: 16 ms, forward step: 50 ms) (Ye et al., 2020). A bootstrap procedure was conducted to exclude spurious connectivity in each individual. It generated 1000 surrogating information flow matrices by calculating symbolic transfer entropy on shuffled TEPs. The connectivity strength was set to zero if they did not exceed 95 % of surrogating strength.

Perturbational complexity index (PCI ST ) and natural frequencies
We measured the cortical reactivity evoked (i.e., phase-locked, TEPs) and induced (i.e., non-phase-locked, time-frequency representations, TFRs) by TMS. The PCI ST measures the ability of the whole cortex to engage in complex patterns of causal interactions by quantifying the non-redundant state transitions across all principal components of the evoked perturbation signals (Comolatti et al., 2019). The PCI ST values were calculated on averaged source-reconstructed TEPs of each patient, and group averaged source-reconstructed TEPs of PSD and non-PSD groups, separately for TMS of each target site (DLPFC, M1 and SPL). The principal components were selected so as to account for at least 99 % of variance of the response amplitude, and components with low signal-tonoise ratio (SNR 1.1) were removed. The average number of state transitions in the matrices of the response (20 ms to 300 ms) was compared with that of the baseline (À300 ms to À20 ms). A parameter k (set to 1.2 in this study) was used to control the relative weight of state transitions between baseline and response. More details of the computing pipeline and additional results of TEPs and PCI ST at the sensor level are given in the Supplementary Materials (Figures S2-S3 and related content).
To explore the oscillatory information induced by TMS, we performed event-related spectral perturbation (ERSP) analysis. The induced neuronal responses were isolated by subtracting the individual time-domain average from each trial (Cohen and Donner, 2013). Time-frequency representations (TFRs) of TMS-related oscillatory power were calculated, separately for each ROI at the single trial level, by means of a Hanning taper windowed FFT with frequency dependent window length (width: 3.5 cycles per time window, time steps: 10 ms, frequency steps: 0.25 Hz from 4 to 45 Hz). Data from À1000 ms to 1000 ms around the TMS pulse was selected to ensure a sufficient time and frequency resolution of the ERSPs. We performed single-trial normalization by ztransforming the TFRs and baseline correction (subtracting the average of the À300 ms to À20 ms period) of each trial for each frequency (Premoli et al., 2017). Then, the ERSPs were extracted by cropping the TFRs during the time of interest (À100 ms to 400 ms) for further statistical analysis.
Natural frequency of each ROI was assessed by estimating the main frequency of the local TMS-induced oscillations. We calculated the power spectrum profiles by averaging the oscillatory power between 21 ms to 400 ms of the ERSPs at each target ROI. Then the natural frequency was defined corresponding to the maximum peak of the power spectrum profile (Rosanova et al., 2009;Tscherpel et al., 2020).

Lesion maps
Individual lesion maps were manually created and crossvalidated by three well-trained physicians on DWI images through the software MRIcron (https://www.nitrc.org/projects/mricron). Then, the individual lesion maps and DWI images were coregistered to the individual T1-weighted images by Statistical Parametric Mapping (SPM12, https://www.fil.ion.ucl.ac.uk/ spm/software/spm12/). Lesion size was measured by the number of damaged voxels. To investigate possible associations between specific damaged voxels and development of PSD, we normalized the individual T1-weighted images and lesion maps to the T1weighted MNI-template implemented in SPM12.

Data availability
The data that support the findings of this study and all custom written MATLAB codes are available from the corresponding author upon reasonable request.

Statistics
Bootstrap sampling statistics was used to determine significant cortical responses (Casali et al., 2013). Brain excitability was determined as the summed absolute values of significant cortical responses (TEPs) between 20 ms to 300 ms. Excitability of brain regions (defined by the Brainnetome Atlas) was compared between PSD vs non-PSD groups by non-parametric Mann-Whitney U-tests and FDR correction with p < 0.05. We conducted non-parametric permutation tests (p < 0.05) to indicate significantly different information flow matrix between PSD and non-PSD groups. Information sending/receiving of ROIs were summarized from rows/columns of information flow matrix in each individual. ROIs with significantly different information sending/receiving between PSD and non-PSD groups were detected by Mann-Whitney U-test and FDR correction with p < 0.05.
Two-way repeated measures analyses of variance (rmANOVAs) were conducted for the cortical AEP components N1 and P2 in source space, with the repeated effect of TMS target (3 levels: DLPFC, M1 and SPL) and the effect of group (2 levels: non-PSD and PSD), after verification that the data were normally distributed. Post-hoc independent-sample two-tailed t-tests were performed in case of a significant effect of group, or interaction of group with TMS target, to compare non-PSD vs PSD. P values were Bonferroni-corrected for multiple comparisons.
To compare cortical reactivity (PCI ST and natural frequency) between PSD and non-PSD groups, we conducted a two-way rmA-NOVA using the software SPSS (version 25), after verifying normal distribution of all data. Repeated effect of TMS target (3 levels: DLPFC, M1 and SPL) and effect of group (2 levels: non-PSD and PSD) were investigated. Post-hoc independent-sample two-tailed t-tests were performed in case of a significant effect of group to compare the PSD vs non-PSD group. P values after multiple comparisons were adjusted using Bonferroni correction. Nonparametric Mann-Whitney U-test was performed to compare lesion size between PSD and non-PSD group, since lesion size did not follow a normal distribution. Receiver operating characteristic curve (ROC) and area under ROC (AUC) were used to assess classification ability of TMS-EEG characteristics and lesion size in distinguishing PSD vs non-PSD.
Relationships of maximum values of PCI ST and natural frequency with delirium duration were tested by Pearson correlation analyses and simple linear regressions with scatter-plots. In order to investigate the associations between specific stroke-affected voxels and delirium and TMS-EEG characteristics, we conducted a voxel lesion symptom mapping (VLSM) analysis. Based on normalized T1-weighted lesion maps, voxels that were damaged in at least two patients in either group were included for statistical analysis (Tscherpel et al., 2020). VLSM was calculated with the NiiStat software (https://github.com/neurolabusc/NiiStat) running on the MATLAB environment and displayed on a T1-weighted MNI-template head by MRIcroGL. Non-parametric permutation tests (2000 permutations) were used to correct for multiple comparisons.
We performed a stepwise logistic regression analysis and a Wald statistic to investigate the strongest predictor of risk of delirium among all the neurophysiologic, imaging and clinical markers, including PCI ST and natural frequency obtained at each of the three TMS targets, maximum PCI ST and maximum natural frequency across the three TMS target sites, and lesion size and NIHSS score.

Results
Fourteen patients (8 males, 6 females; mean 1SD age, 81.1 9. 0 years) developed delirium after the TMS-EEG measurement (range of onset after the TMS-EEG measurement, 8-156 hrs according to longitudinal ICDSC assessment), and were defined as the PSD group (DSM-V). Nineteen patients (8 males, 11 females; mean 1SD age, 77.1 11.6 years), who did not develop delirium over the 7 days following the TMS-EEG measurement, were defined as the non-PSD group (Table 1).

TMS-evoked potentials (TEPs)
The TMS-evoked EEG responses (absolute z-values relative to the baseline, averaged over all patients, in a window 21 -100 ms after the TMS pulse) occurred mainly at the site of the TMS targets (red dots indicate neuronavigated targets in DLPFC, M1 and SPL) and ROIs (dark blue areas) defined by the Brainnetome Atlas (Fig. 1A-C). This verifies that the intended cortical TMS targets were successfully activated.
Temporal and spatial characteristics of TEPs of DLPFC (Fig. 1A), M1 (Fig. 1B) and SPL (Fig. 1C) were different in the PSD vs non-PSD group averages. The spatial activation maps of TMS-evoked response intensity (absolute sum of significant responses 20 -300 ms post-stimulus, bootstrap non-parametric statistics with 1000 times shuffling) showed widely decreased cortical responses in PSD compared to non-PSD group. The significantly decreased responses (non-parametric Mann-Whitney U-test, FDR corrected with p < 0.05) involved, in addition to local activation, several cortical sources distant from the stimulated sites ( Fig. 1D-F).

Effective connectivity
TMS evoked effective connectivity was significantly decreased in PSD compared to non-PSD (non-parametric permutation tests, p < 0.05), particularly when TMS targeted DLPFC and M1 (Fig. 2-A-C). This decrease in effective connectivity in PSD was mainly caused by a significant decrease of receiving information in distributed brain regions (Mann-Whitney U-test, FDR correction with p < 0.05) rather than a decrease in sending information ( Fig. 2A-C).

Perturbational complexity index
The PCI ST values (including data of all three TMS targets) of the PSD group (mean 1SD: 52.3 12.3) were significantly lower (p 0.001, t = À7.3, independent two-sample t-test) than the values of the non-PSD group (mean 1SD: 73.7 15.5) (Fig. 3A). Two-way rmANOVA revealed a significant effect of group, with lower PCI ST values in the PSD group compared to the non-PSD group [main effect of group: F (1,90) = 54.4, p < 0.001], without significant interaction with TMS target [F (2,90) = 2.4, p > 0.05]. Post hoc t-tests disclosed a between group difference with significantly lower PCI ST values in the PSD group compared to the non-PSD group for each of the three TMS targets (mean 1SD of PSD vs non-PSD at DLPFC: 52.3 12.3 vs 78.9 17.3; at M1: 52.0 13.2 vs 72.0 13.6; at SPL: 50.3 13.0 vs 70.4 14.9) (all p < 0.001) (Fig. 3B). The PCI ST values with TMS of either DLPFC (AUC = 0.910), M1 (AUC = 0.790) or SPL (AUC = 0.850) could effectively distinguish between PSD and non-PSD (Fig. 3C). When taking the maximum PCI ST across the three TMS targets in each patient, the classification reached an AUC of 0.943 (Fig. 3C), higher than the AUCs of resting-state power spectral density analyses ( Figure S1 and related content in Supplementary Materials). In addition, these PCI ST data analyzed at source level were validated by a PCI ST analysis at sensor level that provided virtually identical findings ( Figures S2-S3 and related content in Supplementary Materials).
Lesion size (AUC = 0.737) or NIHSS (AUC = 0.852) alone did not achieve better classification between PSD and non-PSD than TMS-EEG characteristics. Furthermore, the Bayesian analysis of covariance (Supplementary Materials, Figure S4) did not provide evi- Cortex plots show regions sending (right of the matrix) and receiving information (above the matrix), which was significantly weaker (Mann-Whitney U-test, FDR correction with p < 0.05) in PSD than non-PSD group. dence for the inclusion of the stroke-affected hemisphere as a significant factor in predicting post-stroke delirium in conjunction with PCI ST or natural frequency.
VLSM analysis revealed that the PCI ST at SPL was associated with small lesion clusters in the subcortical white matter of the corona radiata (Fig. 6A). Natural frequency at DLPFC was associated with small lesion clusters in frontal cortex (Fig. 6B).
A stepwise logistic regression analysis indicated that the strongest predictor (Chi2Stat = 26.08, p < 0.001) of delirium risk was maximum PCI ST . Its predictive effect was confirmed by Wald statistic (Wald statistic = 6.16, p = 0.013) (Fig. 6C). The Wald statistic indicated that only NIHSS score (Wald statistic = 4.13, p = 0.042) potentially contributed to the predictive model of maximum PCI ST .

Discussion
EEG and functional MRI research have verified that delirium is a disconnection syndrome, i.e., a consequence of a breakdown of connectivity in brain networks (Sanders, 2011;van Dellen et al., 2014). Thus, measuring connectivity of brain networks will be a more direct approach to probe the underlying mechanism of delirium, and possibly provide the opportunity to predict delirium prior to its onset (Shafi et al., 2017). However, resting-state EEG and fMRI passively record brain activity and, therefore, are limited in their capacity to make inferences about brain function. In contrast, TMS-EEG provides a powerful means to directly measure the cerebral response to a defined perturbation, which allows testing of effective connectivity. Although preliminary, our study indicates, to the best of our knowledge for the first time, that TMS-EEG can be used to predict the risk of PSD. We present evidence that abnormalities of cortical reactivity to TMS, quantified by evoked responses, evoked connectivity and induced oscillations, are associated with the risk of delirium development during the following days. Specifically, those acute stroke patients, who presented with low maximum PCI ST , had a high risk of PSD.
TEPs reflect spatial and temporal summation of excitatory and inhibitory post-synaptic potentials, time-locked to the TMS pulse, and originating from the activity of a large population of cortical pyramidal neurons and interneurons (Hill et al., 2016). Abnormal TEP morphologies are linked with altered brain states caused by, e.g., severe psychiatric or neurological disorders (Tremblay et al., 2019). A high-amplitude and low-complexity early component was demonstrated as the TEP characteristic in stroke patients (Sarasso et al., 2020), and was associated with severity of initial neurological deficit and functional outcome at 90-day follow-up (Tscherpel et al., 2020). The present study revealed that PSD patients exhibited TEPs with reduced amplitude and number of deflections compared to non-PSD patients (Fig. 1). Considering that the amplitude of TEPs reflects information on the excitability of the local underlying cortical networks, and is sensitive to state changes (Massimini et al., 2005), the decreased TEP amplitude (response intensity, Fig. 2D-F) observed in the PSD group might represent a  (M1), and (C) superior parietal lobule (SPL), of one representative patient with non-post-stroke delirium (non-PSD) (left column) and another patient with PSD (middle column). The gray area plotted at the right of each time-frequency plot depicts the power spectrum profile elicited during the first 400 ms after transcranial magnetic stimulation (TMS). The horizontal dashed lines highlight maximum power corresponding to the natural frequency (indicated in Hz also in the time-frequency plots). Right panel shows individual natural frequencies (means 1SD) at DLPFC, M1 and SPL. * indicate statistical significance (twoway rmANOVA with post-hoc independent two-sample t-tests, Bonferroni-corrected for multiple comparisons) between non-PSD (red circles) and PSD (gray circles) groups.  direct measure of neuronal dysfunction prior to delirium onset. Furthermore, the decrease of response intensity in PSD patients was distributed widely throughout extensive bihemispheric regions distant from the stimulated sites (cf. Fig. 2D-F), suggesting a suppressed propagation of the neuronal responses to other areas of the brain beyond local hypoexcitability at the sites of stimulation. Our study provides evidence that such a breakdown of effective connectivity might link with an abnormal brain state facilitating delirium development.
Disturbances in the organization of brain networks result in cognitive deficits and altered levels of attention and awareness (Chennu et al., 2017), which are typical core symptoms of delirium. Delirium has been hypothesized to be a disconnection syndrome (Sanders, 2011;van Dellen et al., 2014). In our study, to quantify the ''network property" of TEPs, we calculated the symbolic transfer entropy and PCI ST . Methodologically, symbolic transfer entropy measures information-theoretic causal relationship of TEPs and PCI ST measures the spatiotemporal dynamics of TEPs and reflects the joint presence of integration and differentiation in thalamocortical brain networks. The significantly decreased information flow and lower PCI ST values in PSD patients compared to non-PSD patients indicates reduced integration of TMS-evoked responses across cortical areas or a lack of differentiation of cortical responses (stereotypical activity). Therefore, both the suppressed propagation of TMS evoked responses, blocked information flow and the low PCI ST reflect disturbed effective network connectivity of PSD patients, which is consistent with the functional and structural network findings predisposing to delirium (Sanders, 2011;van Montfort et al., 2019).
From a structural prospective, white matter is considered as main propagation pathway of TMS-evoked signals. White matter disintegrity, which has been considered closely associated with delirium development (Hatano et al., 2013;Morandi et al., 2012a), should have significant influence on network integration (i.e., PCI ST ) evoked by TMS. This is consistent with our observed association of PCI ST with lesion voxels in deep white matter (Fig. 6A). Besides white matter lesion load, neurotransmitter and neuroendocrine dysregulation, inflammation, aging, oxidative stress, diurnal dysregulation, all of which have been considered as predisposing delirium risk factors, can affect the integrity of brain networks in stroke patients (Maldonado, 2013). Therefore, we speculate that a breakdown of effective brain network connectivity creates a vulnerable brain condition that lowers the threshold for transition from a normal state to a cognitive dysfunctional or unawareness state. This relation may be of particular relevance for advancing our understanding of the pathophysiology of delirium development in the first days after a stroke event.
Natural frequency reflects the predominant frequency of synchronization of neuronal firing in a brief period following the TMS pulse (Herring et al., 2015) and is presumably mediated through cortico-subcortical networks (Rosanova et al., 2009). Our study reports decreases of natural frequencies at each of the three TMS targets (DLPFC, M1 and SPL) of the stroke patients (Fig. 4), when compared to the natural frequency of healthy subjects reported in previous studies (Rosanova et al., 2009). The reduced TMS-induced oscillation frequencies are consistent with findings in sub-acute stroke patients reported in (Tscherpel et al., 2020). More importantly, we demonstrated that patients in the PSD group exhibited significantly lower natural frequencies compared to those in the non-PSD group. Together, slowing oscillations in resting-state EEG and reduction of natural frequency in TMS-EEG might index a predisposing brain state of delirium.
In contrast to the network integration index PCI ST , TMS-induced natural frequencies are a local-region specificity index. Each region tends to resonate at approximately its own characteristic fre-quency to TMS (Rosanova et al., 2009;Vallesi et al., 2021). Damage of gray matter and its cortico-subcortical connectivity would reduce the natural frequency (Fig. 6B) (Sarasso et al., 2020). Our natural frequency findings highlighted the DLPFC and SPL but not M1 in predicting risk of delirium development. Lesion studies and functional brain imaging offered clues that the prefrontal and parietal cortex, and the surrounding white matter, are correlated with delirium (Committeri et al., 2007;van Montfort et al., 2019). Prefrontal cortex has a unique role as the executive area of the brain for higher associative and integrative activities. Patients with delirium show a positive correlation between activity in the DLPFC and the posterior cingulate cortex compared to healthy controls who demonstrated inverse correlation (Choi et al., 2012). Prefrontal cortex has a 'supramodal executive status' for information processing and has wide-ranging effects on behavior and cognition, since it exhibits rich interconnectedness with cortical association areas, limbic cortex, and ascending brainstem neurotransmitter pathways. It is, therefore, not surprising that the dysfunctional prefrontal cortico-subcortical pathway, presented by reduction of natural frequency, was included as a predisposing risk of delirium. In addition to frontal regions, the parietal cortex also plays an important role in forming cortical top-down attention networks (Corbetta and Shulman, 2011). Strokeinduced impairments in the functioning of cortical attention networks have been associated with behavioral signs of delirium (He et al., 2007;Karnath et al., 2001). Furthermore, the dorsal and ventral parietal cortices project to and modulate the Ascending Reticular Activating System (ARAS), a system which initiates and maintains wakefulness and arousal. The ARAS was proposed a specific brain network associated with delirium (Boukrina and Barrett, 2017). The delirious patients had an acute reversible disruption of connectivity in ARAS and returned to normal after resolution of delirium (Choi et al., 2012). Therefore, functioning of the parietal cortex, participating in the top-down attention network and bottom-up (afferent) projections of ARAS, should be closely related to the neuronal mechanisms of delirium. Consistently, strokes patients with lesions at posterior parietal cortex present with severe delirium as the main clinical manifestation (Boukrina et al., 2021;Naidech et al., 2016).
Stroke lesion size (Fig. 5A), affected hemisphere, and stroke severity (Fig. 5D) have been considered risk factors for delirium (Kostalova et al., 2012;Ojagbemi et al., 2017;Pasinska et al., 2018). Although the PCI ST values showed inverse correlations with lesion size (Fig. 5B-C) and NIHSS (Fig. 5E-F), the covariate analyses and the stepwise logistic regression analysis including Wald statistic verified that PCI ST was independently associated with PSD, irrespective of lesion size, affected hemisphere or stroke severity.
Importantly, measures of cortical reactivity were significantly related to delirium duration. Lower complexity and slower TMSinduced oscillations (maximum values across the three TMS targets) were associated with longer time staying in the delirium state ( Fig. 6D-E). Considering the impact of whiter matter integrity on complexity and natural frequency of TMS-evoked neural activity, these findings are in line with previous results that longer delirium duration correlated with decreased white matter integrity (Morandi et al., 2012b).
A further aim of the present study was to establish a novel approach to predict delirium risk in acute stroke patients. Although the number of studies is still limited, TMS-EEG has been shown a potentially useful technique to identify neurophysiological changes after stroke Gray et al., 2017;Pellicciari et al., 2018;Tscherpel et al., 2020). However, no TMS-EEG study so far has been applied to delirium research, probably because the technique is complex and difficult to handle at bedside.

Limitations
We provide first evidence that measures of cortical reactivity calculated from TMS-EEG data could be used as potential biomarkers for predicting delirium risk, but our work is not without limitations. First, although we included a representative sample of the stroke patients, the sample size is small. Therefore, the findings will have to be validated in a larger sample. Second, we used earplugs rather than noise masking as pilot testing proved auditory masking to be stressful in acute stroke patients. Further, we wanted to avoid exposure of patients to an additional risk factor for delirium. However, as we have indicated earlier, we conducted more extensive analyses compared to previous studies (Tscherpel et al., 2020;Vallesi et al., 2021) to render a major confound of our findings by auditory evoked potentials unlikely. Finally, patients without detectable MEPs were excluded from the study. Such patients could be recruited into future studies with advanced navigation systems that allow estimation of the induced electric field for determining individual TMS intensity (Sarasso et al., 2020).

Conclusions
EEG responses to TMS can unravel brain network states of reduced excitability, effective connectivity, complexity and natural frequency that identify acute stroke patients at high risk for development of delirium. Findings provide novel insight into the pathophysiology of pre-delirium brain states, and may promote targeted effective delirium prevention strategies in those patients at high risk. Moreover, TMS-EEG is a relatively demanding technology that cannot be easily broadly applied. Given the increasing clinical interest in TMS-EEG, we expect that this will drive technical development and simplification to make this important technology more widely available soon.

Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.