Network connectivity and structural correlates of survival in progressive supranuclear palsy and corticobasal syndrome

Abstract There is a pressing need to understand the factors that predict prognosis in progressive supranuclear palsy (PSP) and corticobasal syndrome (CBS), with high heterogeneity over the poor average survival. We test the hypothesis that the magnitude and distribution of connectivity changes in PSP and CBS predict the rate of progression and survival time, using datasets from the Cambridge Centre for Parkinson‐plus and the UK National PSP Research Network (PROSPECT‐MR). Resting‐state functional MRI images were available from 146 participants with PSP, 82 participants with CBS, and 90 healthy controls. Large‐scale networks were identified through independent component analyses, with correlations taken between component time series. Independent component analysis was also used to select between‐network connectivity components to compare with baseline clinical severity, longitudinal rate of change in severity, and survival. Transdiagnostic survival predictors were identified using partial least squares regression for Cox models, with connectivity compared to patients' demographics, structural imaging, and clinical scores using five‐fold cross‐validation. In PSP and CBS, between‐network connectivity components were identified that differed from controls, were associated with disease severity, and were related to survival and rate of change in clinical severity. A transdiagnostic component predicted survival beyond demographic and motion metrics but with lower accuracy than an optimal model that included the clinical and structural imaging measures. Cortical atrophy enhanced the connectivity changes that were most predictive of survival. Between‐network connectivity is associated with variability in prognosis in PSP and CBS but does not improve predictive accuracy beyond clinical and structural imaging metrics.

In PSP and CBS, between-network connectivity components were identified that differed from controls, were associated with disease severity, and were related to survival and rate of change in clinical severity. A transdiagnostic component predicted survival beyond demographic and motion metrics but with lower accuracy than an optimal model that included the clinical and structural imaging measures. Cortical atrophy enhanced the connectivity changes that were most predictive of survival.
Between-network connectivity is associated with variability in prognosis in PSP and CBS but does not improve predictive accuracy beyond clinical and structural imaging metrics.

K E Y W O R D S
connectivity, corticobasal syndrome, fMRI, prediction, progressive supranuclear palsy, survival, tauopathies

| INTRODUCTION
Progressive supranuclear palsy (PSP) and corticobasal syndrome (CBS) are characterised by short average survival, but with significant variability in individual outcome (Chiu et al., 2010;Coyle-Gilchrist et al., 2016). There is a pressing need to accurately predict rate of progression and survival time, to aid clinical management, assist stratification for clinical trials and to identify potential protective factors associated with better prognosis (Eimeren et al., 2019). Functional connectivity is a promising candidate to improve prognostication given the close association between functional organisation and changes in cognition with ageing and neurodegeneration (Chan et al., 2014;Tsvetanov et al., 2021).
PSP and CBS are distinct clinical disorders in their classical forms but nonetheless overlap in their genetic risk factors, (Höglinger et al., 2011;Kouri et al., 2015) pathologies, (Kovacs, 2015) clinical features, (Armstrong et al., 2013;Höglinger, 2018;Höglinger et al., 2017;Murley et al., 2020) prognostic indicators, (Lansdall et al., 2019;Murley et al., 2021) and in current and potential therapeutic agents (Bluett et al., 2021;Boxer et al., 2020;VandeVrede et al., 2020). We make a distinction between the clinical syndromes of PSP and CBS, and the clinicopathological four-repeat tauopathies of corticobasal degeneration (CBD) and PSP pathology. While the initial description of PSP, consisting of a supranuclear gaze palsy, axial predominant rigidity, and early falls, (Steele et al., 1964) is strongly predictive of PSP-pathology, (Litvan, 1997;Osaki et al., 2004) highly varied clinical features are associated with the same proteinopathy Kovacs et al., 2020). The pathological aetiology of corticobasal syndrome is heterogeneous; PSP-pathology is common as well as CBD (Koga et al., 2022). The presence of shared clinical features is recognised in the current diagnostic criteria for the clinical syndromes, (Armstrong et al., 2013;Höglinger et al., 2017) with operationalised definitions of PSP-CBS, (Höglinger et al., 2017) CBD-PSP, (Armstrong et al., 2013) and likely 4-R tauopathy (Höglinger et al., 2017). Many patients with a clinical diagnosis of either PSP or CBS will have features of both diagnostic criteria (Murley et al., 2020).
PSP and CBS have also been previously shown to share clinical determinants of survival, including apathy (Lansdall et al., 2019) and motor impairment (Murley et al., 2021). The degree of clinical convergence means that symptomatic therapeutic options are common between the two conditions (Bluett et al., 2021). Moreover, there is both growing interest in and completed examples of 'basket' designs for novel experimental agents, which recruit across the 4-R tauopathies VandeVrede et al., 2020). Therefore, considering PSP and CBS together can aid prognostication, to gain understanding of how and when pathophysiological processes converge to determine outcome, and to generate mechanistic biomarkers relevant to both conditions.
Temporally correlated brain networks are consistently observed in healthy adults, across the lifespan, and can be identified by functional magnetic resonance imaging at rest (Beckmann et al., 2005;Biswal et al., 1995;Damoiseaux et al., 2006;Yeo et al., 2011). Altered functional organisation, representing dysfunctional neurons and networks, maybe a more sensitive measure of underlying disease state than regional atrophy or cross-sectional performance on standardised clinical tasks. In neurodegenerative conditions network segregation is associated with maintained cognitive performance in the presence of pathology, (Ewers et al., 2021;Tsvetanov et al., 2021) with loss of network integrity and large-scale network change occurring at the point of symptom onset . It is therefore plausible that greater network disruption would imply poor longitudinal outcome.
Identifying connectivity markers of survival would enable in vivo mechanistic testing of the importance of different components of the neurodegenerative cascade for outcome.
A challenge when assessing the impact of connectivity on survival is that even in healthy controls, individual connections show poor reproducibility and vary on repeat scanning (Lynch et al., 2020;Noble et al., 2019). However, multivariate data-driven approaches to identify a small number of features, such as independent component analysis, significantly improve robustness of connectivity estimates (Elliott et al., 2018;Marek et al., 2022). This is important when considering the clinical syndromes of PSP and CBS where connectivity changes are diffuse, (Ballarini et al., 2020;Brown et al., 2017) in keeping with brain-wide synaptic loss observed in vivo  and at post-mortem Lipton et al., 2001). We, therefore, investigated the utility of functional connectivity to predict outcomes for individual diagnostic groups and transdiagnostically, adopting a whole-brain approach rather than focusing on individual connections.
Data reduction techniques to identify common patterns of connectivity change may not give the most sensitive survival predictors.
Machine learning approaches may be more successful in identifying predictors, but standard machine learning tools need to be modified when estimating time to death given the presence of censored data resulting from including individuals alive at the end of a follow-up period (Spooner et al., 2020). The simpler partial least squares (PLS) regression for Cox models (Bastien, 2008;Bastien et al., 2005Bastien et al., , 2015 Bertrand & Maumy-Bertrand, 2021) provides a promising approach that is adapted to explain maximal variance in survival, identifies patterns using all features, and is suitable for high-dimensional data.
We, therefore, used these methods to test whether connectivity changes are associated with poorer prognosis in PSP and CBS. We quantify connectivity through resting-state functional MRI and compare the predictive value of connectivity with clinical metrics and structural imaging. To assess generalisation, we used k-fold crossvalidation for data from two cohorts of PSP, CBS, and controls: from the Cambridge Centre for Parkinson-plus (CCPP) and the UK National PSP Research Network (PROSPECT-MR). We tested the following hypotheses: (i) between-network connectivity differs between participants with neurodegeneration and controls; (ii) more extensive changes in connectivity predict faster clinical deterioration and shorter survival; and (iii) changes in connectivity provide additive information to predict prognosis beyond clinical and structural imaging measures.

| Participants
We recruited 146 participants with MDS-PSP criteria probable or possible PSP, (Höglinger et al., 2017) 82 participants with the clinical phenotype of corticobasal syndrome, (Armstrong et al., 2013) and 90 age-matched healthy controls from the CCPP and the PSP-Corticobasal Syndrome-Multiple System Atrophy-UK (PROSPECT-MR) study (Jabbari et al., 2020). Clinical assessments for the two cohorts included the PSP rating scale (PSPRS), (Golbe & Ohman-Strickland, 2007) the Cambridge Behavioural Inventory-Revised (CBIR) (Wear et al., 2008) and the Addenbrooke's Cognitive Examination-Revised (ACER) (Mioshi et al., 2006). A total of 49 participants with PSP, 11 participants with CBS, and 9 healthy controls were excluded following assessment for motion (see below). We recorded survival and longitudinal neurocognitive assessments for participants up to 12 years from baseline imaging. We recorded date of death from participants' NHS Summary Care Record. Demographic details and summary scores for included participants are described in Table 1.
Twenty-seven participants included with a clinical diagnosis of PSP proceeded to autopsy, with a predominant neuropathological diagnosis of PSP in 26, and 1 predominant argyrophilic grain disease.
Sixteen of the participants included with corticobasal syndrome donated their brains. As expected in CBS, the underlying neuropathology was heterogenous with a final pathological diagnosis of CBD T A B L E 1 Demographic details for participants at baseline scan. (n = 6), Alzheimer's disease (n = 5), mixed CBD/PSP (n = 1), PSP (n = 2), Pick's disease (n = 1) and multiple system atrophy (n = 1).

| MRI acquisition and preprocessing
Participants at CCPP underwent fMRI imaging at 3T (TR 2-2.5 s, TE 30 ms, 3 Â 3 Â 3.5 mm/3 Â 3 Â 3.75 mm voxels, 140-305 volumes). We adapted the FSL preprocessing pipeline  with the addition of wavelet despiking (Patel et al., 2014) given higher in-scanner movement in participants with neurodegenerative diseases. For initial fMRI preprocessing the T1 structural images were cropped to remove non-brain tissue followed by brain extraction using FSL's Brain Extraction Tool. We then used FSL's FEAT with the following steps: motion correction using MCFLIRT; spatial smoothing using a Gaussian kernel of 5 mm FWHW; grand-mean intensity normalisation of the 4D dataset by a single multiplicative factor; and 100 Hz high-pass temporal filtering. Structured artefacts were removed using independent component analysis denoising using FSL's MELODIC together with FIX, following hand-training. Registration to high-resolution structural and standard space images was carried out using FLIRT. Registration from high-resolution structural to MNI space was then further refined using FNIRT nonlinear registration. Normalized images were inspected to ensure adequate registration. We did not use global signal regression, given the potential to remove neural signals and introduce anti-correlations. (Murphy & Fox, 2017) Wavelet despiking was used for further removal of motion artefact.
Since in-scanner participant motion in resting state fMRI has the potential to bias connectivity estimates, (Power et Table S1.

| Structural parcellation
We derived subcortical volumes and cortical thickness for parcels of the Brainnetome Atlas (Fan et al., 2016) using Freesurfer 7.1.0. (Dale et al., 1999). Subcortical volumes were adjusted for total intracranial volume by deriving residuals from linear regression between parcel volume and total intracranial volume (Voevodskaya et al., 2014). Volumes and thicknesses were averaged over the 48 larger regions and gyri to reduce number of features for model fitting. We additionally calculated volumes for four brainstem structures (medulla, pons, midbrain, and superior cerebellar peduncle) (Iglesias et al., 2015).

| Between-network connectivity
To identify between-network connectivity patterns we employed the pipeline used by Elliot and colleagues (Elliott et al., 2018) (Figure 1).
We adopted this approach as it captures multivariate large-scale connectome patterns with improved test-retest reliability, important in these heterogenous conditions where widespread connectivity change and synaptic loss  suggest that isolated connections are unlikely to be reliably related to survival. Additionally, connections between large-scale networks using group-independent component analysis are more robust to spatial variability or differences between participants in alignment (Allen et al., 2012). We per- FSLNets (Smith, Vidaurre, et al., 2013). We adjusted for scanner and site differences through an empirical Bayes framework using ComBat (Johnson et al., 2007;Yu et al., 2018). We compared the adjusted between-component connectivity between patient groups and healthy controls in a linear model with age and sex as covariates of no interest, using the Benjamini-Hochberg method (Benjamini & Hochberg, 1995) to control the false discovery rate.
We then performed a further independent component analysis (Hyvarinen, 1999) to identify a small number of components capturing between-network connectivity patterns. We set a maximum model order of four since even in a large dataset only four components could be robustly inferred, (Elliott et al., 2018)  beyond baseline severity, we performed stepwise regression using the Akaike information criteria. In the initial model, estimated slope was the dependent variable, with the two connectivity components, baseline clinical score, and total grey matter volume as independent variables. Age, sex, motion, and total intracranial volume were covariates of no interest, and not stepped out of the model.
We used a Cox proportional hazards regression analysis to assess the relationship between component score and time from scan until death with age and sex as covariates, an approach that enabled us to include participants alive at the end of the assessment period. Given the importance of in-scanner motion as a potential confounder in quantifying connectivity, we additionally report the relationship between mean framewise displacement and time from scan until death.

| Partial least squares for cox models
We proceeded to compare different potential predictors of survival in PSP and CBS. An independent component analysis finds statistically independent connectivity changes, but these may not be the best survival predictors. We, therefore, used PLS for Cox models (Bastien et al., 2005(Bastien et al., , 2015Bertrand & Maumy-Bertrand, 2021) to maximize covariance of the predictor to censored survival data. This finds broad connectomic patterns most predictive of survival and likely to improve reliability beyond focusing on individual connections.
We used a transdiagnostic approach with PLS regression for Cox models performed with all baseline patient scans as a single group.
We derived models with different predictors to determine indicators To determine the best survival predictors, we used 20 repeats of five-fold cross-validation comparing: a baseline model ( survival data (Heagerty et al., 2000).
While partial least squares regression as a data reduction technique is designed to handle high-dimensional data with multicollinearity, where collinearity is very high there is a risk of model misspecification and overfitting, (Bastien et al., 2015) particularly with a large number of predictors (Chun & Keleş, 2010). Since structural measures were found to have be strongly collinear, we compared the predictive accuracy and coefficients from the PLS regression with regularized regression for Cox models. We used an elastic net penalty with the glmnet function from the glmnet package (Friedman et al., 2010;Simon et al., 2011) in R. We performed 100 repeats of five-fold cross-validation, comparing mean concordance at the regularization term with highest cross-validation performance with mean concordance from the PLS regression model.
On a post-hoc basis, we repeated model comparison with PSP and CBS individually. We compared the same models as in our transdiagnostic assessment, with the addition of a combination of the baseline model with clinical scores and a composite of thalamic, pontine, and midbrain volume, given the risk of overfitting with higher feature number to participant ratio in these subgroups.

| Baseline atrophy and longitudinal connectivity change
We tested whether baseline focal atrophy influenced the longitudinal change in connectomic predictors of survival for the subset of patients with repeat imaging. We first derived PLS connectivity component scores for scans after the baseline visit. We tested the relationship between connectivity and time from baseline imaging session in a linear mixed-effect model with PLS connectivity component score as the dependent variable, time from baseline scan as a fixed effect, and a random intercept for each participant. We then refitted the model including an interaction term with time from baseline imaging and focal atrophy (mean cortical thickness or subcortical volume).
We proceeded to perform mediation analysis using the mediation (Tingley et al., 2014) package in R using bootstrapping with 100,000 draws, with the PLS connectivity component as a mediator, mean cortical thickness or subcortical volume as predictors and age, sex and the remaining atrophy marker as a covariate of no interest.

| Data sharing
The PROSPECT-MR dataset reported here may be available subject to a PROSPECT data-sharing agreement, after review by the PROS-PECT data access committee. For details on how to apply, please contact the senior authors. Data from the CCPP is available on reasonable request to the senior authors but may be subject to restrictions that protect confidentiality, and a data transfer agreement may be required according to the nature of the request.
PSP between the thalamus and a dorsal attention network component and between sensorimotor and visual regions ( Figure S1). There were no significant differences between PSP and CBS after correction for multiple comparisons.

| Structural metrics
Differences between groups in cortical thicknesses and subcortical volumes are presented in Figure S2. In PSP the greatest atrophy, when compared to controls, was found in subcortical regions and the frontal lobe. The largest effect sizes in CBS (vs. control participants) were in the frontal and parietal lobes, thalamus, and basal ganglia.
Comparing PSP and CBS, atrophy was greater across cortical regions in CBS, with larger reductions in thalamic and brainstem volumes in PSP.

| Connectivity relates to clinical severity
We took the between-network connections to an independent component analysis to capture broad patterns of connectivity to compare with clinical severity and progression. We found that using four CBS versus Control Tukey-adjusted p = .014). The same components differed by group with mean framewise displacement included in the model (Component 1 F = 12.6, p = 3 Â 10 À5 ; Component 2 F = 4.9, p = .044). In Component 1 the disease state was associated with predominantly decreased connectivity but with increased connectivity between task-positive, motor, and subcortical regions (Figure 3b).
Lower scores in Component 2, as observed in CBS, were associated with relatively increased connectivity between the default mode, dorsal attention, and motor networks and decreased connectivity within these networks.
We considered the components that differed from controls in either disease group in subsequent analysis. Component 1 scores were associated with the PSPRS (Figure 3c Std Beta = À.31, p = .0007) and the CBIR (Std Beta = À.29, p = 0.002), with similar but weaker associations found with Component 2 (PSPRS Std

| Connectivity and disease progression
We tested whether baseline component scores were associated with a subsequent decline in neuropsychological assessments. Linear mixed-effect models indicated an effect of time for all measures (Table S2). We found that baseline Component 1 score was associated We used stepwise regression to investigate if connectivity components were included in the best model of progression when incorporating baseline severity and total grey matter volume. For the PSPRS, Component 1 was included in the final model (Table S3), with Component 2 in the final model for ACER (Table S4), and both components for the CBIR (Table S5). Between-network connectivity differences were associated with a more rapid decline in severity beyond baseline clinical scores and global atrophy.

| Connectivity and survival
We found that a lower Component 1 score was a significant predictor of survival using Cox proportional hazards regression

| Comparing transdiagnostic models to predict survival
We proceeded to investigate the optimal predictors of survival in patients with PSP and CBS. Since the most important connectivity changes for determining outcome may differ from the patterns of changes most common in disease, we used PLS for Cox models to maximise covariance between predictor and survival. We identified a connectivity component with covariance maximised to predict survival (Figure 5a), with worse survival related to relatively increased connectivity between task-positive regions, from the thalamus to sensorimotor regions and from the default mode network to visual regions, representing a loss of segregation between these large-scale networks, with decreased connectivity elsewhere. We found component scores differed between patient groups and controls (PSP versus Controls t = 3.8 Tukey-adjusted p = .0005; CBS versus Control t = 3.6, p = .001), with no difference between PSP and CBS (Figure 5c t = 0.1, p = .99). We also identified two structural components predictive of survival ( Figure S4). The highest absolute weights for the first component were for the thalamus, pons, and midbrain, with significant contributions from limbic and frontotemporal cortical regions. Lower scores in the second component, associated with worse survival, were for participants with thalamic and brainstem atrophy but relatively preserved cortical thickness. The highest component weight in a clinical component was for the PSPRS (Table S6).
We compared transdiagnostic predictive models of survival using repeat five-fold cross-validation to a baseline model consisting of age, sex, and mean framewise displacement from the fMRI scanning session since the latter is predictive of survival ( Figure S2). We found that combining connectivity with the baseline model showed moderate improvement in predictive power, but that this was outperformed by F I G U R E 5 Identifying a transdiagnostic component predictive of outcome. We used partial least squares regression for Cox models to find a component (a) that maximised the covariance between connectivity and censored time to death. Connections represent PLSR weights, so that for higher scoring subjects red indicates stronger connections and blue weaker. This component did not differ between participants with PSP and those with CBS (c). Using five-fold cross-validation with outcome assessed using concordance analysis and integrated area under the curve, we found that connectivity provided additional information above patient's demographic information and inpatient motion, but with a combination of structural, clinical and baseline metrics providing the best predictive accuracy (d-e To consider the potential impact of multiple collinear structural features we tested a further model with baseline and clinical measures, and the sum of volumes from the thalamus, pons, and midbrain. This post-hoc model showed a modest improvement in performance over other models (mean concordance 0.7; mean iAUC 0.69).
We further tested survival predictors in each diagnostic group individually. In PSP, for a component derived using PLS regression with all predictors, highest weights were for the PSPRS, pons, midbrain, and thalamic volumes and bilateral superior temporal gyri thicknesses (Table S7). In CBS largest component weights were for the PSPRS, right thalamus, pons, and midbrain, with hippocampal atrophy also predictive of poor survival (Table S8). In addition, in CBS connectivity between posterior networks (posterior default mode network, dorsal attention, and visual) were also weighted highly. In both PSP and CBS best model performance was with baseline and clinical predictors, together with the composite thalamic, pons, and midbrain volume (PSP mean concordance 0.68, mean iAUC 0.68; CBS mean concordance 0.72 mean iAUC 0.69).

| Focal atrophy and its relationship to connectivity
Since connectivity was only a moderate survival predictor, we investigated whether connectivity change may be driven by focal pathology.
We considered the relationship between connectivity and cortical and subcortical atrophy, given that subcortical parcels had high loadings in the best survival model.  In summary, we have found that cortical rather than subcortical atrophy modulates the connectivity changes that are more strongly predictive of survival. However, the effects of subcortical atrophy on survival (primarily thalamic, pontine, and midbrain) are predominantly not mediated by changes in between-network connectivity.

| DISCUSSION
In this study of two independent cohorts, we have found that functional connectivity and focal atrophy predict disease trajectory for people with PSP and CBS, including their rate of progression and survival. There are connectivity changes associated with shorter time to death that are shared between the diseases, but these provide less robust predictions than simple clinical and structural imaging metrics.
In the most accurate model for survival prediction, the greatest weights were for the PSPRS and thalamic, midbrain, and pontine volume. Cortical rather than subcortical volume at baseline was associated with subsequent progressive change in the functional connectivity that was predictive of survival. In contrast, the prognostic value of subcortical atrophy on survival is largely independent of the changes in network connectivity.
We found patterns of connectivity and structural change associated with poor survival that were shared between PSP and CBS. This is in keeping with the clinical, molecular, and pathological overlap between the diseases, (Höglinger, 2018;Murley et al., 2020)  The relationship between network connectivity and clinical severity is in keeping with findings that connectivity changes are closely associated with cognitive status in ageing (Chan et al., 2014) and in presymptomatic carriers of dementia-causing mutations Tsvetanov et al., 2021). Our whole brain approach shows that connectivity changes that predict survival similarly represent a disruption to functional organisation rather than simply connectivity loss. Between-network connectivity was predominantly decreased in participants with CBS and PSP, with increased connectivity also occurring across network hierarchies (Gotts et al., 2020;Margulies et al., 2016). Greater connectivity increased scores of a component with covariance maximised to predict survival, notably between taskpositive multimodal networks, from the thalamus to sensorimotor regions and from the default mode network to visual regions. The finding that relative regional increases in connectivity contribute to poor survival supports studies demonstrating an association between increased connectivity of higher cognitive networks in health and poor cognitive function, (Chan et al., 2014;Geerligs et al., 2017) and suggests that these connectivity differences indicate network inefficiency rather than compensatory changes. Cell death and the widespread cortical synaptic loss in PSP and CBS ) may cause loss of segregation between distinct networks, such as the dorsal and ventral attention networks, with network segregation important in maintaining performance on cognitive tasks despite pathological change (Ewers et al., 2021;Tsvetanov et al., 2021). Functional brain organisation at rest relates to task-based network changes (Cole et al., 2014(Cole et al., , 2016. Altered connectivity between multimodal networks at rest in PSP and CBS may indicate task-based network dysfunction, with behavioural and cognitive consequences relevant for disease progression (Lansdall et al., 2019;Murley et al., 2021).
Cortical atrophy and cortical network connectivity are interconnected, demonstrated by the finding that 'epicenter' regions of maximal atrophy can be used as seeds to select functional networks associated with neurodegenerative disease (Seeley et al., 2009;Zhou et al., 2012). Our findings support this observation, suggesting that connectivity change potentially mediates the survival effects of cortical atrophy. However, the largest effects on connectivity for structural measures were for the thalamus, pons, and midbrain. The importance of thalamic atrophy may be surprising given that in PSP cortical pathology defines the later stage of PSP tauopathy (Kovacs et al., 2020) while in CBS cortical rather than thalamic atrophy is a major imaging correlate (Boxer et al., 2006;Whitwell et al., 2010). The thalamus, pons, and midbrain contain fibres and nuclei important in diverse neuronal systems, (Roy et al., 2022) including in core motor functions that have been linked to survival in PSP and CBS (Glasmacher et al., 2017;Murley et al., 2021). While thalamocortical connections have been shown to be disrupted in primary tauopathies (Whitwell et al., 2011) our data suggest that the majority of the effect of subcortical atrophy on survival is not mediated by disruption to between-network connectivity. Instead, the contribution of subcortical atrophy to survival is relatively independent of cortical atrophy or connectivity.
Our work highlights some of the barriers that limit betweennetwork connectivity from resting-state functional MRI as a dementia biomarker. Network connectivity satisfies criteria for a biomarker of progression, anticipating clinical deterioration with a mechanistic rationale for a causal relationship (Eimeren et al., 2019). Yet even when adopting a methodology designed to increase reliability, the failure of connections to appear repeatedly in imaging means that results are insufficiently robust to provide accurate single-subject survival predictions or to operate as an intermediate endpoint for clinical trials (D'Esposito, 2019). We selected a small number of independent components to assess between-network connectivity, but this approach may fail to identify important functional connectivity or activation patterns relevant for survival. There is a range of alternative approaches to analysing functional data, including graph metric, dynamic connectivity, voxel-wise, and gradient-based analyses which may also capture characteristic differences predictive of survival. Further work is needed to determine whether these methods are more robust and with better test-retest reliability in neurodegenerative conditions with diffuse connectivity change and synaptic loss. One important consideration is the relevance of brainstem and thalamic structures in survival in PSP and CBS. Estimates of functional connectivity in these regions are affected by high physiological noise and other analytic approaches may be considered (Beissner et al., 2014).
There are other limitations to our study. We found that inscanner motion itself predicts survival in PSP and CBS. Despite adopting a principled preprocessing pipeline and not including motion confounds as a regressor in higher-order regressions, (Geerligs et al., 2017) there is an inevitable compromise between over-zealous preprocessing removing connectivity indicative of poor survival, and the failure to remove spurious connectivity deriving from motion (Power et al., 2012). To reduce the risk of motion biasing our assessments of connectivity we excluded significant numbers of participants, so it is possible that our conclusions do not apply to the excluded members of the cohort. We have used cross-validation to assess the accuracy of our survival predictions across sites but have not tested results in a third, out-of-sample, cohort that varies by scanner and protocol (Yu et al., 2018). Although we present data from a sizeable cohort of participants, increasing study power would allow for model fine-tuning and to compare machine learning approaches.
We found only uncorrected differences between PSP and CBS and differential effects of connectivity on cognitive performance. We adopted an approach to analysis designed to detect diffuse changes in connectivity that might be associated with poor survival. Alternative methodological choices, such as completing analysis only with patient groups, may better capture between-group differences and be useful to test if these differences are important in predicting survival. Recent developments (Horie et al., 2022) in fluid biomarkers may help improve in vivo prediction of pathological aetiology in tauopathies, which has the potential to assist prognostication.
In conclusion, between-network functional brain connectivity predicts clinical deterioration and survival in PSP and CBS, with prediction in terms of cross-validation and in terms of future changes after baseline scanning. However, functional connectivity provides less accurate predictions of survival than simpler measures of focal subcortical atrophy and baseline clinical severity.

FUNDING INFORMATION
The study was co-funded by the National Institute for Health