Methods & systems for detection of dementia

Abstract

A computer-implemented method of assessing a human patient, the method comprising: receiving rs-fMRI data of the patient's brain; determining a measure of effective connectivity of the patient's default mode network based on the rs-fMRI data; and deriving a dementia prognosis for the patient based on the measure of effective connectivity.

Description

METHODS & SYSTEMS FOR DETECTION OF DEMENTIAField

[0001] The present invention relates to predicting the risk and time course for future onset dementia in individuals, in particular based on detecting changes in default mode network (DNM) seen in resting-state functional MRI (rs-fMRI) data and provides methods, systems and computer programs therefor.Background

[0002] There is currently intense interest in identifying strategies to reduce the growing population burden of dementia. Clinical syndromes of dementia are caused by multiple neuropathologies that typically co-occur within individuals. Alzheimer’s disease (AD) pathology is the most important contributor to dementia at population level, and is associated with distinct patterns of pathological protein deposition and altered neural function that precede the development of structural brain changes and clinical symptoms by a period of years.

[0003] Resting-state fMRI (rs-fMRI) is used as a tool for characterising connectomic biomarkers in AD as disclosed in Ref. 1. rs-fMRI measures endogenous fluctuations in blood oxygen level-dependent (BOLD) signal across the brain, which in turn reflect regional neural activation, whilst a participant lies in an MRI scanner at rest. By computing correlations between BOLD timeseries from different brain regions, a map of functional connectivity can be estimated as disclosed in Ref. 2. When rs-fMRI is applied to people with AD, or its precursor, mild cognitive impairment (MCI), there are significant changes in functional connectivity at group level, when contrasted with healthy controls as disclosed in Refs. 3-9. Similar changes have been identified in individuals who do not yet have MCI or AD but are considered high risk due to genetic polymorphisms, mutations for autosomal dominant AD, a family history of AD or a high burden of pathogenic amyloid and tau protein. Altered functional connectivity, measured with rs-fMRI, is therefore considered a potential preclinical biomarker of AD. However, it has not previously been shown to allow single-subject level identification of future dementia risk in a population-based cohort.

[0004] The brain regions most commonly implicated in altered functional connectivity in AD are those within the default-mode network (DMN), which is hypothesised to be selectively vulnerable to AD neuropathology as disclosed in Ref. 10. The DMN is typically described as having a core set of brain regions, which includes medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC) or precuneus, and bilateral inferior parietal cortices, as well as a set of supplementary brain regions, which includes the medial temporal lobes and temporal poles Ref. 11. The DMN was initially described as a network of regions that coactivate during a “task-negative” state in functional imaging studies Ref. 12. In other words, these brain regions appear more active when a participant is at rest. However, research shows that the DMN is implicated in several high-level cognitive processes, including social cognition and mental time-travel Refs. 11, 13, resulting in a contemporary view that the DMN furnishes an individual with their narrative sense of “self’ Ref. 14.

[0005] Whilst findings of altered functional connectivity in the DMN have led to claims that dementia is a syndrome of “dysconnectivity”, the exact connectivity changes observed are inconsistent across studies Refs. 1, 15 and are occasionally undetectable Ref. 16. This is perhaps due to methodological limitations associated with defining connectivity on the basis of timeseries correlations that overlook biophysical constraints and the established neurobiology of neural circuit function. An alternative approach in connectomics is to fit a neurobiologically-informed circuit model to the functional neuroimaging data, in order to characterise the excitatory and inhibitory connections between different brain regions, i.e. “effective connectivity” Ref. 17. Moving beyond correlations in brain activity, effective connectivity describes the causal influence of one brain region over another, by modelling the underlying neural signals that generated the observed data.

[0006] Effective connectivity mapping with DCM has been used to successfully discriminate between people with semantic dementia and healthy controls Ref. 18, and also to predict which people with Parkinson’s Disease are likely to experience hallucinations Ref. 19. A small number of studies have estimated effective connectivity differences between people with AD or MCI and healthy controls Refs. 20 - 25 and also detected differences in small samples of preclinical cohorts at high-risk for AD Ref. 26, 27.Summary

[0007] It is desirable to reliably detect early changes in neural function associated with AD, for example in order to provide a platform for the development of individualised dementia prevention strategies.

[0008] An embodiment of the present invention provides a computer-implemented method of assessing a human patient, the method comprising: receiving rs-fMRI data of the patient’s brain; determining a measure of effective connectivity of the patient’s default mode network based on the rs-fMRI data; and deriving a dementia prognosis for the patient based on the measure of effective connectivity.

[0009] Another embodiment of the present invention provides a computer-implemented method of training a model for use in deriving a dementia prognosis for a patient, the method comprising: receiving rs-fMRI data of the brains of a plurality of cases and controls; receiving dementia diagnosis data of the plurality of cases and controls; determining a measure of effective connectivity of the default mode network of each of the plurality of cases and controls based on the rs-fMRI data; and training the model based on the measures of effective connectivity and dementia diagnosis data for each of the plurality of cases and controls.

[0010] Accordingly, embodiments of the present invention detect effective connectivity changes in a constructed DMN which can be used to make early predictions about dementia incidence and prognosis of individuals in a population cohort.Brief Description of the Drawings

[0011] The invention will be described further below with reference to exemplary embodiments and the accompanying drawings, in which:Figure l is a schematic diagram of a process for determining effective connectivity in an embodiment;Figure 2 is a schematic diagram of an analysis pipeline of an embodiment;Figure 3 is a table showing effective connectivity in healthy controls;Figure 4 is a table showing Bayesian model average of the difference in effective connectivity between cases and controls;Figure 5 is a set of diagrams showing effective connectivity differences visualised in MNI space between cases and controls;Figure 6 is a graph of the receiver-operating characteristic (ROC) curve for a regularised logistic regression model trained on effective connectivity parameters to classify dementia cases (including preclinical cases) from controls;Figure 7 is a graph of the receiver-operating characteristic (ROC) curve for a regularised logistic regression model trained on grey matter volumes;Figure 8 is a table showing effective connectivity in dementia cases (including preclinical cases);Figure 9 is a table showing Bayesian model average of the difference in effective connectivity between cases associated with a longer time until diagnosis;Figure 10 is a set of diagrams showing effective connectivity differences visualised in MNI space between cases associated with a longer time until diagnosis;Figure 11 is a graph demonstrating the performance of a regularised linear regression model, trained on effective connectivity parameters, to predict time until diagnosis, in a cohort of 81 preclinical dementia cases;Figure 12 is a graph demonstrating the performance of a regularised linear regression model, trained on grey matter volumes, to predict time until diagnosis, in a cohort of 81 preclinical dementia cases;Figure 13 is a flowchart of a training method of an embodiment; andFigure 14 is a flowchart of a patient assessment method of an embodiment.

[0012] In the various drawings, like parts are indicated by like references.Exemplary Embodiments

[0013] An embodiment of the present invention seeks to assess a patient’s risk of developing dementia, including but not limited to dementia caused by Alzheimer’s disease, and to predict the time to development of symptoms on which a clinical diagnosis could be made.

[0014] Altered functional connectivity precedes structural brain changes and symptoms in dementia. Alzheimer’s disease is the largest contributor to dementia at population level, and disrupts functional connectivity in the brain’s default-mode network (DMN). A neurobiol ogical model of DMN effective connectivity is proposed to identify preclinical dementia at single-subject level. An embodiment was used to apply spectral dynamic causal modelling to resting-state fMRI data in a nested case-control group (n=l 133) from the UK Biobank, including individuals who developed dementia up to 9 years after imaging, and matched controls. Dysconnectivity predicted both future dementia incidence (AUC=0.81) and time to diagnosis (R=0.56), outperforming a model based on brain structure. Evaluation of associations between DMN dysconnectivity and major risk factors for dementia, revealed strong relationships with polygenic risk for Alzheimer’s disease and social isolation supporting the theoretical basis of the proposed methods. Therefore neurobiological models of effective connectivity may facilitate early detection of dementia at population level, supporting rational deployment of targeted dementia prevention strategies to individuals.

[0015] Correlations in BOLD activity among two brain regions in a network can be explained by an enormous number of possible underlying neural circuitries. With dynamic causal modelling (DCM), multiple putative circuit models of effective connectivity can be compared with each other using model comparison procedures, and the best explanation for the observed data is identified. Thus, effective connectivity provides a more nuanced description of neural connectivity and is likely to detect features that would otherwise be missed when mapping functional connectivity derived from observed BOLD signal. These connectomic subtleties, effective connectivity parameters, are likely to afford discriminative and predictive clinical value, for individualised precision medicine, over and above the correlations measured in functional connectivity Ref. 28.

[0016] In neurodegeneration and ageing, neural connections and neurovascular coupling are both impacted Ref. 29 and here DCM becomes particularly useful. Because functional connectivity is a multiplexed signal of neural, haemodynamics and noise components, any observed changes in functional connectivity don’t differentiate if these are caused by changes in neural circuitry, haemodynamics or both. DCM, on the other hand, models the neural, haemodynamics and noise components of BOLD separately.

[0017] Embodiments use effective connectivity changes in the DMN to make early predictions about dementia incidence and prognosis of individuals in a population cohort. This has been validated by a nested case-control study using the UK Biobank cohort, among which a sample have developed incident dementia in the years since neuroimaging data acquisition. To ensure that the analysis was ecological and reflected the range of dementia pathologies within the population, all-cause dementia outcomes were used rather than restricting the analysis to those with Alzheimer’s disease. rs-fMRI data from individuals who developed dementia and a large sample of matched controls was analysed. Spectral DCM Ref. 28, a technique that fits generative neural and haemodynamic models to the cross-spectra of BOLD timeseries from rs-fMRI data was applied in order to estimate effective connectivity. It was predicted that there would be detectable differences in DMN effective connectivity years before people were diagnosed with dementia, and that these differences would be large enough to make meaningful out-of-sample predictions about future dementia incidence. It was predicted that these early patterns of dysconnectivity would be associated with exposure to known risk factors, particularly polygenic risk for AD as the key driver of AD pathological change, and social isolation due to the role of the DMN in social cognition Ref. 11.

[0018] UK Biobank sample selection

[0019] We identified all participants recruited into the UK Biobank (UKB), who have ever had a dementia diagnosis on their GP record, as of the latest UKB data update in May 2023, and who also had resting-state functional MRI (rs-fMRI) data available on the UKB database. For each of these dementia cases, we identified 10 control participants who were matched with the dementia case on age, sex, handedness, ethnicity, and geographical location of MRI scanning centre. This yielded an initial sample of 148 dementia cases and 1480 controls. After excluding participants who failed the preprocessing stage (e.g. excessive head motion) and replacing failed controls with new matched controls we were left with a final usable sample of 103 cases and 1030 matched controls. Of these 103 cases, 81 did not have a dementia diagnosis at the time of MRI data acquisition, whilst 22 already had prevalent dementia.

[0020] MRI data acquisition

[0021] MRI data were acquired between 2006 and 2010 as part of the UKB prospective cohort study, across multiple sites in the United Kingdom (Manchester, Newcastle and Reading). The scanner was a Siemens Skyra 3T with a Siemens 32-channel RF receive headcoil. Each participant underwent a 35-minute scanning session, during which the following data were acquired: T1 -weighted structural image, rs-fMRI timeseries, T2-weighted FLAIR structural image, diffusion MRI structural image, susceptibility-weighted image and taskbased fMRI timeseries data. For our analyses, we only used the T1 image and the rs-fMRI data.

[0022] The T1 -weighted image was acquired in a five-minute 3D MPRAGE sequence with a resolution of 1 mm isotropic. The rs-fMRI data were acquired using a six-minute GE-EPI sequence with x8 multislice acceleration. Resolution 2.4 mm isotropic, TR 0.735 s, TE 39 ms, flip angle 52°.

[0023] MRI data preprocessing

[0024] Preprocessing was performed on the raw UKB imaging data in SPM12 using batch scripts in MATLAB R2023a. Firstly the T1 -weighted structural image was segmented into tissue subtypes, skull-stripped and then warped into MNI (Montreal Neurological Institute) space. The rs-fMRI data were spatially realigned to the single-band reference scan that was acquired in addition to the multi -band EPI sequence. Volumes were then co-registered to the skull-stripped T1 image, normalised to MNI space and spatially smoothed using a 6 mm isotropic Gaussian kernel.

[0025] In-scanner head motion was estimated for each subject by computing frame-wise displacement for each subject using the three translational and three rotational motion parameters (assuming rotation around the surface of a sphere with radius 50 mm). Subjects were excluded from further analysis if their maximum framewise displacement exceeded the voxel resolution of 2.4 mm. This threshold was chosen because it was the voxel resolution of the dataset.

[0026] Timeseries extraction

[0027] A default-mode network (DMN) was constructed by pre-defining 10 regions-of- interest (ROI) based on pre-existing literature. This number of ROIs was chosen to compromise between anatomical detail and feasible computation time when fitting dynamic causal models. Fewer ROIs, e.g. 8 or 9, may be used. More ROIs - e.g. up to 12, 15, or 20 - may also be used. The 10-node network comprised a core DMN of anterior medial prefrontal cortex (amPFC), precuneus (PRC) and left / right intraparietal cortex (UPC / rlPC). These four ROIs were centred around the following co-ordinates respectively: (x = 2, y = 56, z = -4), (x= 2, y = -58, z = 30), (x = -44, y = -60, z = 24), (x = 54, y = -62, z = 28). We included the following additional ROIs in our DMN network, using co-ordinates from a previous study on DMN connectivity in amnestic cognitive impairment by Dunn et al. Ref. 4: ventromedial prefrontal cortex (vmPFC), dorsomedial prefrontal cortex (dmPFC), left / right lateral temporal cortex (ILTC / rLTC), and left / right parahippocampal formation (IPHF / rPHF), centred on the following co-ordinates respectively: (x = 0, y = 26, z = 18), (x = 0, y = 52, z = 26), (x = -60, y = -24, z = 18), (x = 60 y = -24, z = 18), (x = -28, y = -40, z = -12), (x = 28 y = -40, z = - 12). Other regions that could be selected include hippocampal formations (left x=-25, y=-23, z=-14; right x=26, y=-21, z=-14), temporal pole (e.g. left x=-40, y=l 1, z=-30; right x=41, y=13, z=-30), and temporoparietal junction (e.g. left: x = -52, y = -54, z = 23; right: x = 51, y = -54, z = 26). It will be appreciated that if a different brain space is used, the numerical coordinates of the various regions of interest will be different than set out above.

[0028] Signal from each ROI was estimated by estimating a general linear model (GLM) containing a discrete cosine basis set with frequency range 0.0078 - 0.1 Hz as well as the following nuisance regressors: six head motion regressors, a regressor for cerebrospinal fluid (CSF) signal (principal eigenvariate of 5 mm radius sphere centred in the 3rd ventricle at (x = 0, y = -40, z = -5)), a regressor for white matter (WM) signal (principal eigenvariate of 6 mm radius sphere centred in the brainstem at (x = 0, y = -24, z = -33)). Global signal regression was not performed as there is evidence it does not significantly impact results in small network analyses. An F-contrast was specified across all components of the discrete cosine basis set, yielding a BOLD timeseries of low-amplitude fluctuations in each voxel within a 10 mm radius sphere centred on each of the 10 ROI co-ordinates listed above.

[0029] For each ROI, a new 8 mm sphere was then centred on the peak intensity voxel. A summary signal for the ROI was computed as the principal eigenvariate of all suprathreshold voxels (uncorrected alpha 0.05) that lay in the conjunction space of the first 10 mm sphere and the second 8 mm sphere. These were voxels with evidence for low frequency BOLD fluctuations. Note that the principal eigenvariate across voxels is used, rather than the mean, so that negative and positive signals do not negate each other and that the extreme values don’t bias the mean estimate. If any of the 10 ROIs yielded no suprathreshold voxels then the subject was excluded from further analysis.

[0030] Estimating effective connectivity

[0031] Effective connectivity was estimated using spectral dynamic causal modelling (DCM) using the DCM12 toolbox in SPM12 published by Functional Imaging Laboratory, UCL Queen Square Institute of Neurology, London, UK. Spectral DCM fits a biophysical state-space model to the observed cross-spectra of BOLD signals, to estimate underlying neuronal states and the rate of change in neural activity in each region (in Hz) as a function of activity in other regions (i.e. effective connectivity). For each subject we fitted a fully- connected DCM with a connectivity parameter for every possible pair of the 10 ROIs, including auto-inhibitory self-connections. This model thus comprised 100 connectivity parameters. The DCM software uses the Variational Laplace algorithm to invert the model and estimate these connectivity parameters by minimising negative free energy. We used the software’s default priors. Each subject’s DCM fit was screened for convergence by ensuring it met the following criteria: Explained variance of BOLD signal greater than 10%, at least one connection (excluding self-connections) with an absolute connection strength of greater than 1 / 8 Hz, and at least one effectively estimated parameter (based on the Kullback-Leibler divergence of posterior from prior). All subjects met these criteria for model convergence.

[0032] To estimate an average connectivity matrix across subjects, and to estimate the difference in connectivity between cases and controls, we fit a parametric empirical Bayes (PEB) model to the full set of subject-specific DCMs. The PEB technique enables us to estimate group-level connectivity strengths by fitting a hierarchical model to the estimated connectivity parameters of each individual and the precisions of those parameters. We specified a between-subjects design matrix that contained five columns. A column of ones, to model the average connectivity strengths across all subjects, a column of ones and zeros to model the differences in connectivity between cases and controls, as well as three columns to model covariates of no-interest (age, sex and mean framewise displacement to model any effects attributable to head motion). The latter three columns were mean-centred. Rather than estimating a full covariance matrix across connectivity parameters, a single precision component was shared across connectivity parameters, to permit model estimation within a reasonable amount of time. The resulting PEB model comprised 500 connectivity parameters, a 10x10 connectivity matrix for each of the five columns of the between-subjects design matrix.

[0033] Finally, we used exploratory Bayesian model reduction and Bayesian model comparison to find the best (and simplest) model to explain the data. In this procedure an automatic greedy search over reduced models iteratively discards parameters that don’t contribute to model evidence. A Bayesian model average of parameters is then calculated over the 256 models from the final iteration of the greedy search (default settings of DCM software).

[0034] The details of the biophysical model used in DCM, model inversion at subject- and group-level, and Bayesian model reduction have already been extensively documented and will not be reproduced here.

[0035] Case-control classifier

[0036] Of the group-level parameters that model differences in effective connectivity between cases and controls, we selected all parameters with a posterior probability of at least 99% of being non-zero. This way we identified a set of statistically plausible connections to use as data features for our classifier.

[0037] We trained an elastic-net regularised logistic regression model on these features to classify cases from controls using the glmnet toolbox for MATLAB. To accommodate for the 10: 1 imbalance in class size, observation weights were applied so that cases were weighted 10 times more than controls. A leave-one-out nested cross-validation (CV) scheme was applied for tuning two hyperparameters: elastic mixing parameter a and regularisation penalty X. For each outer fold of CV one subject was selected as a test subject. Among all remaining training subjects, each was iteratively selected as a validation subject for multiple inner folds of CV. Each inner fold was repeated for a different value of a (0, 0.2, 0.4, 0.6, 0.8, 1). Glmnet automatically uses a range of 100 values every time a model is estimated. For each possible combination of hyperparameters, a ROC curve was computed from the predictions of each inner fold in the training set. The ROC with the highest area under the curve (AUC) was identified. The logistic regression model was then trained on the full training set, using only the hyperparameters that yielded the winning ROC within the training set. This fully trained model was finally applied to the left-out test subject to generate a probability P(case) and a final ROC curve.

[0038] Prognosticator

[0039] To test whether effective connectivity features could also be used to predict when these individuals got their dementia diagnosis, we trained a prognosticator model. A group- level effective connectivity matrix was computed, using the PEB framework with Bayesian model reduction, as described above, but this time only the dementia cases were included in the analysis. The second column in the between-subjects design matrix was not a column of ones and zeros to represent cases and controls, but rather a continuous variable that was computed as date of MRI acquisition subtracted from date of dementia diagnosis (i.e. how long, in years, until dementia diagnosis). One subject, with prevalent dementia at the time of data collection, was excluded from this analysis as there was no reliable date of prior dementia diagnosis. Of the group-level parameters that model differences in effective connectivity as a function of time until diagnosis, we selected all parameters with a posterior probability of at least 99% of being non-zero.

[0040] We then trained an elastic-net regularised linear regression model using exactly the same procedure as described above for the classifier. The only difference here was that hyperparameters were tuned by minimising the squared error between predictions and true values. Performance was evaluated as the Spearman correlation coefficient between final model predictions are true values.

[0041] Volumetric data comparison

[0042] To see how useful effective connectivity parameters were at making predictions about dementia compared to other MRI-based features, we repeated the above analyses, but this time using volumetric data features from structural MRI instead. We used pre-existing grey matter volume data from UKB’s imaging-derived phenotype (TDP) database. This data comprises 139 regional grey matter volumes segmented using FAST (FMRIB'S Automated Segmentation Tool) and 14 subcortical volumes segmented using FIRST (FMRIB's Integrated Registration and Segmentation Tool). We summarised all of this volumetric data into 32 features as follows: For each left and right cortical lobe (frontal, parietal, temporal and occipital) we computed the sum of grey matter volumes for regions within that lobe, yielding 8 features. An additional 22 features comprised grey matter volumes for left and right amygdala, subcallosal cortex, cerebellum, caudate, pallidum, putamen, ventral striatum, hippocampus, parahippocampal gyrus, thalamus and insula. Finally, an additional 2 midline structures (vermis and brainstem) were included, yielding a total of 32 features. Each featurewas normalised by total intracranial volume. We then trained regularised logistic regression and linear regression models on this volumetric data using exactly the same procedures that we used for the effective connectivity data features, as described above.

[0043] The results are shown in Figures 6 and 7. Figure 6 show receiver-operating characteristic (ROC) curve for a regularised logistic regression model trained on effective connectivity parameters to classify dementia cases (including preclinical cases) from controls. Figure 7 is the same as Figure 6 but for a model trained on grey matter volumes. In both panels the grey lines show 95% confidence intervals for the true and false positive rates, generated with bootstrapping. It can clearly be seen that the model trained on EC has better results (AUC=0.81 vs AUC = 0.65).

[0044] Modifiable risk factors

[0045] We investigated which modifiable risk factors were associated with dementia-related changes in DMN effective connectivity using multiple multivariable linear regression models. We constructed a variable for each of the 12 modifiable risk factors identified in the 2020 Lancet commission on dementia. History of hypertension, diabetes, smoking, depression, physical inactivity, traumatic brain injury and hearing loss, absence of secondary education, and residence in a highly polluted neighbourhood (top decile) were coded as binary variables. BMI, weekly alcohol consumption and social isolation were coded as continuous numerical variables. The social isolation variable was constructed with data from 3 questions, which participants answered as part of the touchscreen session at baseline data collection. These 3 questions assessed: (1) Weekly attendance at social leisure activities (binary), (2) Estimated number of visits from friends or family within a year (continuous numerical), and (3) Estimated number of times the participant felt able to confide in someone close to them within a year (continuous numerical). We ran a principal components analysis on these 3 variables and took individual scores for the first principal component, which loaded negatively on all 3 variables (i.e. a higher score on this principal component indicated greater social isolation). Traumatic brain injury was excluded from the subsequent regression analyses as there were only 9 positive cases across the entire sample. This left 11 modifiable risk factors for analysis. For all variables, missing data was imputed with the median across all participants.

[0046] For each of the 11 modifiable risk factors, as well as AD polygenic risk score (PRS), a weighted linear regression model was estimated using fitglm in MATLAB, where the riskfactor of interest was the predictor variable, and “effective connectivity index (ECI)” was the response variable. ECI is simply the probability of dementia outputted from the previous classification analysis. A higher value here indicates that the participant’s overall effective connectivity pattern conforms more to a “dementia-like” phenotype than a “control-like” phenotype. Age, sex and Townsend social deprivation score were included as covariates of no-interest in each of the 12 linear regression models. Dementia subjects were upweighted and control subjects were downweighed in the linear regression models, such that cases and controls made equal contributions to the regression models. A p value was estimated for each of the 11 modifiable risk factors and for PRS, which was corrected for multiple comparisons using the Holm-Bonferroni method, to maintain a family-wise error rate (FWER) of 0.05.

[0047] A mediation analysis was performed, with social isolation as a predictor, ECI as a mediator and dementia incidence as a response variable. Each regression model estimated in the mediation analysis included age, sex and Townsend social deprivation score as covariates of no-interest, and used weighted observations such that cases and controls contributed equally to the model. A p value was estimated for the significance of the indirect path coefficient by generating a permutation-based null distribution. For each permutation, the dementia incidence variable was randomly shuffled and an indirect path coefficient was estimated. This was repeated 1,000 to generate a null distribution of indirect path coefficients with which to evaluate the true indirect path coefficient magnitude.

[0048] Results

[0049] After exclusions for image quality and excessive in-scanner head motion (see Methods), our final usable sample included 103 dementia cases (22 with prevalent dementia and 81 who later developed incident dementia) and 1030 matched controls. The 81 preclinical cases had a median time to diagnosis of 3.7 years (range: 0.4 - 8.5). The total sample had a mean age of 70.4 at the time of MRI data acquisition. Cases and controls were matched on age, sex, ethnicity, handedness and geographical location of testing centre.

[0050] The analysis pipeline is illustrated in Figure 1. For each subject, BOLD timeseries were extracted from 10 pre-defined regions-of-interest (ROI), which together defined our default mode network (DMN). The network included 4 midline ROIs (precuneus (PRC), anterior medial prefrontal cortex (amPFC), dorsomedial prefrontal cortex (dmPFC) and ventromedial prefrontal cortex (vmPFC)), 1 ROI in each medial temporal lobe, in the left andright parahippocampal formations (IPHF / rPHF), and 4 lateral ROIs (right intraparietal cortex (rIPC), left intraparietal cortex (UPC), right lateral temporal cortex (rLTC) and left lateral temporal cortex (1LTC). Further details on these ROIs can be found in the Methods.

[0051] A fully connected dynamic causal model (DCM) was fitted to the cross-spectra of these timeseries data (spectral DCM) in order to estimate the effective connectivity between each and every pair of ROIs in the 10-node network shown in Figure 1. There are 4 mid-line ROIs shown in the sagittal section, 2 medial temporal ROIs shown in the coronal section and 4 lateral temporal and parietal ROIs shown in the axial section. For each participant, voxels were only selected within the sphere if supra-threshold activation was detected. Lighter shades of blue indicate voxels that were selected more frequently across participants. BOLD timeseries were extracted from each of the 10 ROIs.

[0052] A spectral dynamic causal model (DCM) was fitted to these BOLD timeseries data. The DCM optimises effective connectivity parameters, to find the best explanation for the observed BOLD timeseries, in terms of excitatory (purple) and inhibitory (grey) neural connections, and altered blood flow that would be expected to result from this neural activity. Each participant’s EC pattern is estimated separately and is represented as a 10 by 10 EC matrix, where each cell in the matrix shows the magnitude and valence (excitatory or inhibitory) of a connection between a pair of ROIs. Bayesian model reduction is applied to the EC matrix to eliminate unnecessary parameters and find the most parsimonious model to explain the observed data, at the group-level. The resulting sparse EC patterns are used to train regularised logistic regression models to predict dementia incidence using leave-one-out cross-validation as depicted in Figure 2. A single participant is left out of the analysis as a test subject, highlighted in green. All remaining participants constitute a training set. The hyperparameters of the regression model are optimised on this training with new nested test and train sets within the outer training set. Once optimal hyperparameters are selected, the optimised model is trained on the full outer training set and tested on the original left-out participant. This procedure is iterated such that every participant is used as a test subject once.

[0053] Effective connectivity can predict who is more likely to get dementia

[0054] Bayesian model reduction and averaging were applied (see Methods) to estimate the simplest effective connectivity map to explain group-level differences between dementia cases and controls (Figures 3 to 5), whilst controlling for age, sex and in-scanner head motion.There was very strong evidence (posterior probability > 0.99) for 15 connectivity parameters that differed between cases and controls. The three largest connectivity changes seen in the dementia cases were: increased inhibition from vmPFC to 1PHF, increased inhibition from 1IPC to 1PHF and attenuated inhibition from rPHF to dmPFC.

[0055] These 15 connectivity parameters were used to train an elastic-net logistic regression model to predict dementia incidence in leave-one-out cross validation (Figure 2). Using a receiver-operating characteristic (ROC) analysis we found the model to have excellent discriminative performance (Figure 6) with an area under the curve (AUC) of 0.81 (95% CI: 0.76 - 0.86). The point biserial correlation between true class labels (case or control) and predicted class labels was 0.36 (p = 9.9 x 10-36).

[0056] To ensure that the model had true predictive validity in a preclinical cohort, we repeated the above analysis, excluding the 22 cases with prevalent dementia and their 220 matched controls. This analysis yielded comparable results, with an AUC of 0.8 (95% CI: 0.73 - 0.84) and a point biserial correlation between true and predicted class labels of 0.33 (p = 3.7 x 10-24).

[0057] Figure 3 solves the Bayesian model average of effective connectivity in healthy controls. Each cell shows the effective connectivity, in Hz, between a pair of regions. Grey indicates an inhibitory connection. Purple indicates an excitatory connection. Cells along the diagonal represent auto-inhibitory connections, as unitless scaling parameters. Only parameters with a posterior probability of being non-zero of at least 99% (amounting to a very strong evidence) are shown. Figure 4 shows the Bayesian model average of the difference in effective connectivity between cases and controls. Grey indicates a change towards increased inhibition (reduced excitation). Purple indicates a change towards increased excitation (reduced inhibition). Only parameters with a posterior probability of being non-zero of at least 99% are shown. Figure 5 shows effective connectivity differences between cases and controls visualised in MNI space. Each tube represents a connection change. Solid tubes represent connections that are strengthened in cases compared to controls. Dashed tubes represent connections that are attenuated in cases compared to controls. The thickness of the tube represents the magnitude of the connection change. The colour of the tube represents the brain region from where the connection originates. The top row and bottom row show the same data but from two different angles. The four columns display the following respectively:Attenuated excitatory connections, strengthened excitatory connections, attenuated inhibitory connections, strengthened inhibitory connections.

[0058] Effective connectivity can predict time to dementia diagnosis in preclinical cases

[0059] In order to assess the potential role of DMN effective connectivity in prognostication, we ran an analysis that only used the case cohort. We used Bayesian model reduction and averaging (see Methods) to estimate the simplest effective connectivity map to explain inter-individual variation associated with the time until future dementia diagnosis, whilst controlling for age, sex and in-scanner head motion. There was a very strong evidence (posterior probability > 0.99) for 37 connectivity parameters that were associated with time until diagnosis (Figures 8 to 10), including the 3 connections, described above, that showed the largest difference between cases and controls (Figures 3 to 5).

[0060] These 37 connectivity parameters were used to train an elastic-net regularised linear regression model to predict time until diagnosis in leave-one-out cross validation. There was a positive correlation between actual time until diagnosis and predicted time until diagnosis (Spearman’s rho = 0.56, p = 1.6 x 10-9). The analysis was also repeated, using only the 81 preclinical cases (Figures 11 and 12), which yielded comparable results (Spearman’s rho = 0.55, p = 2 x 10-7).

[0061] Effective connectivity changes associated with time to dementia diagnosis are shown in Figures 8 to 10.

[0062] Figure 8 shows Bayesian model average of effective connectivity in cases. Each cell shows the effective connectivity, in Hz, between a pair of regions. Grey indicates an inhibitory connection. Purple indicates an excitatory connection. Cells along the diagonal represent auto-inhibitory connections, as unitless scaling parameters. Only parameters with a posterior probability of being non-zero of at least 99% are shown. Figure 9 shows Bayesian model average of the changes in effective connectivity, amongst cases, associated with a longer time until diagnosis. Grey indicates a change towards increased inhibition (reduced excitation). Purple indicates a change towards increased excitation (reduced inhibition). Only parameters with a posterior probability of being non-zero of at least 99% are shown. Figure 10 shows effective connectivity changes. Visualisation follows the same format as Figure 5, but with connectivity changes associated with time until diagnosis, rather than differences between cases and controls.

[0063] Figure 11 illustrates the performance of a regularised linear regression model, trained on effective connectivity parameters, to predict time until diagnosis, in a cohort of 81 preclinical dementia cases. Figure 12 is the same, but for a model trained on grey matter volumes. It can clearly be seen that the model trained on effective connectivity has better predictive power.

[0064] Association between dementia risk factors and effective connectivity

[0065] Finally, we conducted an exploratory analysis to investigate whether the effective connectivity changes might represent the effects of major risk factors for dementia. We first defined an effective connectivity (EC) index for each participant, which was simply the probability of dementia, outputted by the case-control classifier trained on effective connectivity parameters. This value summarises the extent to which an individual’s DMN effective connectivity pattern conforms to a dementia-like phenotype rather than a control-like phenotype, where a value of 1 indicates a dementia-like pattern and a value of 0 indicates a control-like pattern.

[0066] For each individual, we then extracted data from UKB describing the modifiable risk factors identified in the 2020 Lancet commission on dementia (see Methods) as well as each subject’s AD polygenic risk score (PRS). For each risk factor, we ran a separate weighted linear regression model, across the entire cohort of cases and controls, to measure the association between EC index and that specific risk factor, controlling for age, sex and social deprivation score. After correcting for multiple comparisons, AD PRS was strongly associated with EC index (beta: 0.053, p = 3.7 x 10-12), and this association was substantially stronger than any association between EC index and a modifiable risk factor. This suggests that the EC changes we observed are likely to represent a pathological change due to Alzheimer’s disease, rather than a more general reflection of brain health.

[0067] We constructed a mediation model to see whether EC index mediated any of the relationship between PRS and dementia incidence. By including EC index as a mediator, the direct path coefficient from PRS to dementia incidence was reduced from 0.5 (p = 0.0007) to 0.45 (p = 0.017). There was a significant indirect mediated path (beta: 0.07, p < 0.001), which explained away 10% of the association between PRS and dementia incidence. These results indicate that DMN effective connectivity partially mediates the role of genetic risk in dementia pathogenesis.

[0068] For the modifiable risk factors, social isolation was the only variable that showed a significant association with EC index (beta: 0.025, p = 0.003). This association demonstrated that individuals with more self-reported self-isolation were more likely to have a “dementialike” pattern of DMN effective connectivity.

[0069] We constructed a mediation model to test whether EC index might mediate the known association between social isolation and dementia incidence. After accommodating for a hypothesised mediating effect of EC index, we detected a significant indirect path from social isolation to dementia, mediated by EC index (p < 0.001). Furthermore, an association between social isolation and dementia incidence (p = 0.037) was rendered non-significant (p = 0.07), after accounting for this mediator. To further test the biological plausibility of this model, we repeated this mediation analysis, excluding the 22 prevalent cases and their 220 matched controls. This yielded comparable results, with a significant indirect mediation path (p < 0.001), and a direct path from social isolation to dementia (p = 0.043) that was rendered non-significant by including the mediator (p = 0.09). These results show that preclinical effective connectivity changes, in the DMN, mediate an association between premorbid social isolation and subsequent dementia incidence. Taken together, preclinical DMN dysconnectivity appears to be a consequence of both genetic and environmental risk factors.

[0070] Our findings show that a neurobiol ogically-informed model of DMN effective connectivity can enable accurate predictions about whether and when an individual will develop dementia. The performance of our effective connectivity-based classifier exceeded that of classifiers based on volumetric data from structural MRI, both in our analysis, and also when comparing it to prior work using structural MRI data as a unimodal predictor of future conversion to dementia.

[0071] From a clinical perspective, this suggests that rs-fMRI could become a tool for identifying a neural network signature of dementia risk among asymptomatic individuals. This type of non-invasive early detection of dementia is an increasingly valuable goal, particularly with the arrival of disease-modifying drugs. Recent clinical trials have shown promise for amyloid beta-targeting monoclonal antibodies, which are modifying the disease trajectory in AD for the first time, supposedly with greater therapeutic potential when started earlier in the disease process. Early detection of dementia risk is also important in the context of targeted risk reduction strategies irrespective of underlying pathology. Whereas pathology-specific biomarkers can guide disease-modifying molecular therapies, non-specific biomarkers for all-cause dementia, like that developed in the current study, will be useful for identifying who is most likely to benefit from lifestyle changes and public health interventions, and when these interventions are likely to have the biggest impact.

[0072] Recent research on early detection of dementia has tended to prioritise biomarkers that directly reflect pathogenic protein deposition in AD, such as cerebrospinal fluid analysis for amyloid beta and tau proteins. However, these markers have limited predictive ability among healthy population cohorts because a majority of individuals remain asymptomatic during follow up (e.g. >90% of those with amyloid beta positivity remain asymptomatic over 5 years). It is likely that rational use of anti-amyloid therapies among asymptomatic individuals would be enhanced by the addition of a proximity marker based on early neural dysfunction, and our results suggest that effective connectivity could be an ideal candidate for this, especially because they demonstrate that effective connectivity can be used to make predictions, not only about who will develop dementia, but also the time until future diagnosis. These predictions were more accurate than previous prognostic models trained on structural MRI data and functional connectivity features.

[0073] Our use of a population cohort of all-cause dementia, rather than a well-phenotyped AD-specific cohort is both a strength and a limitation of this work. Dementia is typically due to mixed pathologies, and syndromic diagnoses in life are frequently found to be incorrect at post mortem. From a pragmatic population health standpoint, the ability to accurately predict all-cause dementia is therefore desirable, and makes it likely that the results of this study would be generalisable to real-world settings. However, we are limited in our ability to make pathology-specific inferences from these results beyond the strong association between DMN effective connectivity and AD polygenic risk score, which suggest that these effective connectivity changes do at least partially represent pathological changes specific to AD. Stronger evidence for a specific relationship with AD pathology could be obtained through future work incorporating biomarkers of AD proteinopathies. Indeed, in previous work, classifiers have made improved predictions on preclinical cohorts when multimodal data were used, for instance by combining structural MRI, genetic data, cerebrospinal fluid (CSF) assays and cognitive assessments. We anticipate that, when combined with other datamodalities such as amyloid beta and tau markers, effective connectivity would be likely to yield improved predictive performance.

[0074] In an exploratory analysis of modifiable risk factors, we found that social isolation had a unique and strong association with the effective connectivity changes in the dementia cohort. This finding has important implications for our understanding of why DMN dysconnectivity is so frequently observed in clinical and preclinical dementia. There is a significant overlap between the DMN and what is typically described as a “social cognition” network Ref. 11. The mPFC, temporal poles, precuneus and temporo-parietal junction (TP J) consistently activate during cognitive tasks where participants are required to think about another person’s intentions or beliefs (i.e. Theory of Mind). There is emerging evidence that this network of brain regions is highly sensitive to one’s social environment. Social isolation is a well-established risk factor for dementia. Psychosocial interventions, such as cognitive stimulation therapy (CST) can improve symptom burden and may also reverse some of the changes in DMN functional connectivity that are seen in AD. These interventions are thought to weaken the link between underlying dementia pathology and cognitive decline, by promoting compensatory brain changes and expanding “cognitive reserve”.

[0075] From a neurobiological perspective, DMN dysconnectivity is thought to be a consequence of tau accumulation in the posterior cingulate cortex (PCC), which then spreads throughout the network in an activity-dependent manner. In the current study, we found that effective connectivity changes in the DMN mediated an association between social isolation and dementia incidence. This finding is consistent with a theory that social isolation triggers the DMN dysconnectivity observed in dementia.

[0076] One further limitation of our study is that we are unable to determine which DMN effective connectivity changes are pathological and which are compensatory. We identified multiple changes to both inhibitory and excitatory connections. Some of these connections were strengthened whilst others were attenuated. Interestingly, there was an overlap between the connections that discriminated between cases and controls and the connections that changed as a function of time until diagnosis.

[0077] A method of training a classifier according to an embodiment of the invention is depicted in Figure 13. Firstly, resting state functional MRI data as training data is obtained SI for a plurality of subjects. The rs-fMRI data may be obtained specially for the purpose oftraining the classifier or may be selected from pre-existing data. The training data should include data obtained from patients having dementia at the time of scanning and / or diagnosed with dementia after scanning (cases) as well as patients who do not have dementia (controls). Desirably, the controls are matched with the cases in relevant demographic factors, such as age, sex, and other characteristics known to be risk factors for dementia. Any suitable, conventional scanning protocol may be employed.

[0078] A preprocessing step S2 is carried out to normalise the scans, e.g. by remapping to a standard space, and to exclude any unsuitable scans, e.g. where there has been excessive head movement during the scan.

[0079] A timeseries is extracted S3 for each of a plurality of regions of interest from each scan of the training data. There may be at least 8, desirably at least 10 regions of interest. In particular the set of regions of interest desirably comprises a plurality of core default mode network regions and at least one medial temporal lobe region. In a specific embodiment, the set of regions of interest comprises four midline RO Is (e.g. precuneus (PRC), anterior medial prefrontal cortex (amPFC), dorsomedial prefrontal cortex (dmPFC) and ventromedial prefrontal cortex (vmPFC)), one ROI in each medial temporal lobe, one ROI in each of the left and right parahippocampal formations (IPHF / rPHF), and for lateral ROIs (e.g. right intraparietal cortex (rIPC), left intraparietal cortex (UPC), right lateral temporal cortex (rLTC) and left lateral temporal cortex (1LTC)). The timeseries may comprise a series of samples of a BOLD signal.

[0080] Effective connectivity is estimated S4 for each case and control in the training data based on the time series. Dynamic causal monitoring may be used to obtain a measure of effective connectivity as described above. The measure of effective connectivity may comprise a matrix of values indicating connectivity between each of the regions. For example, each cell in the matrix shows the magnitude and valence (excitatory or inhibitory) of a connection between a pair of ROIs.

[0081] Model reduction S5 is performed on the complete set of effective connectivity data to extract parameters that contribute to explaining the incidence of dementia in the cases versus the controls. This may result in a reduced matrix and reduces the computational load required to train the classifier whilst improving its discriminatory power. A parametrical empirical Bayes (PEB) method may be employed.

[0082] The reduced model is then used to train S6 a classifier, such as a regularised logistic regression model or an artificial neural network. Any suitable classifier may be used.

[0083] A method of assessing a patient is illustrated in Figure 14. An rs-fMRI scan of the patient is obtained SI 1. Desirably this is obtained using the same or a comparable protocol as was used for training data. The patient rs-fMRI data is pre-processed S12, again in the same or a comparable manner as applied to the training data. Timeseries for the regions of interest are obtained S13 and a measure of effective connectivity is calculated S14 using dynamic causal monitoring, including a model reduction step is performed to select the effective connectivity pattern with strongest evidence for that subject.

[0084] The measure of effective connectivity is applied S14 to the trained classifier so as to obtain a prognosis for the patient S16. The patient prognosis may comprise: a diagnosis of dementia; a risk of dementia; and / or a prediction of a time to develop symptoms.Conclusion

[0085] In summary, we found that effective connectivity in a constructed DMN can be used as a non-invasive population based preclinical biomarker for predicting future dementia incidence at the level of individual patients. This biomarker, using rs-fMRI data, is superior to using structural MRI data. The connectivity changes in the DMN are strongly associated with AD polygenic risk and social isolation, a risk factor that might accelerate the effects of pathological protein in the DMN.

[0086] Exemplary embodiments of the invention are described below:

[0087] 1. A computer-implemented method of assessing a human patient, the method comprising: receiving rs-fMRI data of the patient’s brain; determining a measure of effective connectivity of the patient’s default mode network based on the rs-fMRI data; and deriving a dementia prognosis for the patient based on the measure of effective connectivity.

[0088] 2. A computer-implemented method according to embodiment 1 wherein the measure of effective connectivity is determined based on a set of regions of interest comprising at least 8, desirably at least 10 brain regions.

[0089] 3. A computer-implemented method according to embodiment 1 or 2 wherein the measure of effective connectivity is determined based on a set of regions of interest consisting of no more than 20, desirably no more than 15, more desirably no more than 10 brain regions.

[0090] 4. A computer-implemented method according to embodiment 2 or 3 wherein the set of regions of interest comprises a plurality of core default mode network regions and at least one medial temporal lobe region.

[0091] 5. A computer-implemented method according to embodiment 2 or 3 wherein the set of regions of interest is selected from the group consisting of anterior medial prefrontal cortex (amPFC); precuneus (PRC); left / right intraparietal cortex (UPC / rlPC); ventromedial prefrontal cortex (vmPFC); dorsomedial prefrontal cortex (dmPFC); left / right lateral temporal cortex (ILTC / rLTC); left / right parahippocampal formation (IPHF / rPHF); hippocampal formations; temporal pole; and temporoparietal junction.

[0092] 6. A computer-implemented method according to embodiment 2, 3 or 4 wherein the set of regions of interest comprises four midline ROIs (e.g. precuneus (PRC), anterior medial prefrontal cortex (amPFC), dorsomedial prefrontal cortex (dmPFC) and ventromedial prefrontal cortex (vmPFC)), one ROI in each medial temporal lobe, one ROI in each of the left and right parahippocampal formations (IPHF / rPHF), and for lateral ROIs (e.g. right intraparietal cortex (rIPC), left intraparietal cortex (1IPC), right lateral temporal cortex (rLTC) and left lateral temporal cortex (1LTC)).

[0093] 7. A computer-implemented method according to any one of the preceding embodiments wherein the measure of effective connectivity is determined using dynamic causal modelling.

[0094] 8. A computer-implemented method according to embodiment 7 wherein the dynamic causal modelling is applied to a blood oxygen level-dependent (BOLD) signal for a plurality of ROIs.

[0095] 9. A computer-implemented method according to any one of the preceding embodiments wherein the measure of effective connectivity comprises a matrix, where each cell in the matrix shows the magnitude and valence (excitatory or inhibitory) of a connection between a pair of ROIs.

[0096] 10. A computer-implemented method according to embodiment 9 wherein determining a measure of effective connectivity comprises obtaining a reduced matrix as the measure of effective connectivity, the reduced matrix representing parameters obtained by applying Bayesian model reduction to training data.

[0097] 11. A computer-implemented method according to any one of the preceding embodiments wherein deriving a dementia prognosis comprises inputting the measure of effective connectivity to a trained classifier.

[0098] 12. A computer-implemented method according to embodiment 11 wherein the trained classifier is a regularised logistic regression model.

[0099] 13. A computer-implemented method according to any one of the preceding embodiments wherein the dementia prognosis comprises a dementia risk value and a time to symptoms value.

[0100] 14. A computer-implemented method of training a model for use in deriving a dementia prognosis for a patient, the method comprising: receiving rs-fMRI data of the brains of a plurality of cases and controls; receiving dementia diagnosis data of the plurality of cases and controls; determining a measure of effective connectivity of the default mode network of each of the plurality of cases and controls based on the rs-fMRI data; and training the model based on the measures of effective connectivity and dementia diagnosis data for each of the plurality of cases and controls.

[0101] 15. A computer-implemented method according to embodiment 14 wherein the model is a classifier, such as a regularised logistic regression model or an artificial neural network.

[0102] 16. A computer-implemented method according to embodiment 15 wherein the model is trained using leave-one-out cross-validation.

[0103] 17. A computer-implemented method according to embodiment 14, 15 or 16 wherein the measure of effective connectivity is determined based on a set of regions of interest comprising at least 8, desirably at least 10 brain regions.

[0104] 18. A computer-implemented method according to any one of embodiments 14 to 17 wherein the measure of effective connectivity is determined based on a set of regions ofinterest consisting of no more than 20, desirably no more than 15, more desirably no more than 10 brain regions.

[0105] 19. A computer-implemented method according to embodiment 17 or 18 wherein the set of regions of interest comprises a plurality of core default mode network regions and at least one medial temporal lobe region.

[0106] 20. A computer-implemented method according to embodiment 17 or 18 wherein the set of regions of interest is selected from the group consisting of anterior medial prefrontal cortex (amPFC); precuneus (PRC); left / right intraparietal cortex (UPC / rlPC); ventromedial prefrontal cortex (vmPFC); dorsomedial prefrontal cortex (dmPFC); left / right lateral temporal cortex (ILTC / rLTC); left / right parahippocampal formation (IPHF / rPHF); hippocampal formations; temporal pole; and temporoparietal junction.

[0107] 21. A computer-implemented method according to embodiment 17 or 18 wherein the set of regions of interest comprises four midline ROIs (e.g. precuneus (PRC), anterior medial prefrontal cortex (amPFC), dorsomedial prefrontal cortex (dmPFC) and ventromedial prefrontal cortex (vmPFC)), one ROI in each medial temporal lobe, one ROI in each of the left and right parahippocampal formations (IPHF / rPHF), and for lateral ROIs (e.g. right intraparietal cortex (rIPC), left intraparietal cortex (1IPC), right lateral temporal cortex (rLTC) and left lateral temporal cortex (1LTC)).

[0108] 22. A computer-implemented method according to any one of embodiments 14 to 21 wherein the measure of effective connectivity is determined using dynamic causal modelling.

[0109] 23. A computer-implemented method according to embodiment 22 wherein the dynamic causal modelling is applied to a blood oxygen level-dependent (BOLD) signal for a plurality of ROIs.

[0110] 24. A computer-implemented method according to any one of embodiments 14 to 23 wherein the measure of effective connectivity comprises a matrix, where each cell in the matrix shows the magnitude and valence (excitatory or inhibitory) of a connection between a pair of ROIs.[oni] 25. A computer-implemented method according to embodiment 24 wherein determining a measure of effective connectivity comprises obtaining a reduced matrix as themeasure of effective connectivity, the reduced matrix representing parameters obtained by applying Bayesian model reduction to training data.

[0112] 26. A computer program comprising instructions that, when executed by a computer system, cause the computer system to perform the method of any preceding embodiment.

[0113] 27. A computer program according to embodiment 26 that is an update or addon to operating software of an MRI scanner.

[0114] 28. A non-transitory computer-readable medium having stored thereon a computer program according to embodiment 26.

[0115] 29. An MRI scanner comprising one or more processors and a memory, the memory storing instructions that, when executed by the processor(s), cause the MRI scanner to perform the method of any of embodiments 1 to 12.

[0116] The methods of the present invention may be performed by computer systems comprising one or more computers. A computer used to implement the invention may comprise one or more processors, including general purpose CPUs, graphical processing units (GPUs), tensor processing units (TPU) or other specialised processors. A computer used to implement the invention may be physical or virtual. A computer used to implement the invention may be a server, a client or a workstation. Multiple computers used to implement the invention may be distributed and interconnected via a network such as a local area network (LAN) or wide area network (WAN). Individual steps of the method may be carried out by a computer system but not necessarily the same computer system. Results of a method of the invention may be displayed to a user or stored in any suitable storage medium. The present invention may be embodied in a non-transitory computer-readable storage medium that stores instructions to carry out a method of the invention. The present invention may be embodied in a computer system comprising one or more processors and memory or storage storing instructions to carry out a method of the invention. The present invention may be incorporated into an MRI device or into software updates or add-ons for such a device.

[0117] Having described the invention it will be appreciated that variations may be made on the above described embodiments, which are not intended to be limiting. The invention is defined in the appended claims and their equivalents.References1 Ibrahim, B. et al. Diagnostic power of resting-state fMRI for detection of network connectivity in Alzheimer's disease and mild cognitive impairment: A systematic review. Hum Brain Mapp 42, 2941-2968, doi: 10.1002 / hbm.25369 (2021)2 van den Heuvel, M. P. & Hulshoff Pol, H. E. Exploring the brain network: a review on resting-state fMRI functional connectivity. Eur Neuropsychopharmacol 20, 519-534, doi: 10.1016 / j.euroneuro.2010.03.008 (2010).3 Berron, D., van Westen, D., Ossenkoppele, R., Strandberg, O. & Hansson, O. Medial temporal lobe connectivity and its associations with cognition in early Alzheimer's disease. Brain 143, 1233-1248, doi: 10.1093 / brain / awaa068 (2020).4 Dunn, C. J. et al. Deficits in episodic memory retrieval reveal impaired default mode network connectivity in amnestic mild cognitive impairment. Neuroimage Clin 4, 473-480, doi: 10.1016 / j.nicl.2014.02.010 (2014).5 Khazaee, A., Ebrahimzadeh, A., Babajani-Feremi, A. & Alzheimer's Disease Neuroimaging, I. Classification of patients with MCI and AD from healthy controls using directed graph measures of resting-state fMRI. Behav Brain Res 322, 339-350, doi: 10.1016 / j.bbr.2016.06.043 (2017).6 Liu, Y. et al. Impaired long distance functional connectivity and weighted network architecture in Alzheimer's disease. Cereb Cortex 24, 1422-1435, doi: 10.1093 / cercor / bhs410 (2014).7 Putcha, D. et al. Tau and the fractionated default mode network in atypical Alzheimer's disease. Brain Commun 4, fcac055, doi: 10.1093 / braincomms / fcac055 (2022).8 Rombouts, S. A., Barkhof, F., Goekoop, R., Stam, C. J. & Scheltens, P. Altered resting state networks in mild cognitive impairment and mild Alzheimer's disease: an fMRI study. Hum Brain Mapp 26, 231-239, doi: 10.1002 / hbm.20160 (2005).9 Badhwar, A. et al. Resting-state network dysfunction in Alzheimer's disease: A systematic review and meta-analysis. Alzheimers Dement (Amst) 8, 73-85, doi: 10.1016 / j.dadm.2017.03.007 (2017).10 Warren, J. D. et al. Molecular nexopathies: a new paradigm of neurodegenerative disease. Trends Neurosci 36, 561-569, doi: 10.1016 / j . tins.2013.06.007 (2013).11 Mars, R. B. et al. On the relationship between the "default mode network" and the "social brain". Front Hum Neurosci 6, 189, doi: 10.3389 / fnhum.2012.00189 (2012).12 Shulman, G. L. et al. Common Blood Flow Changes across Visual Tasks: II. Decreases in Cerebral Cortex. J Cogn Neurosci 9, 648-663, doi: 10.1162 / jocn.1997.9.5.648 (1997).13 Yeshurun, Y., Nguyen, M. & Hasson, U. The default mode network: where the idiosyncratic self meets the shared social world. Nat Rev Neurosci 22, 181-192, doi : 10.1038 / s41583 -020-00420-w (2021 ).14 Menon, V. 20 years of the default mode network: A review and synthesis. Neuron, doi: 10.1016 / j. neuron.2023.04.023 (2023).15 Eyler, L. T. et al. Resting State Abnormalities of the Default Mode Network in Mild Cognitive Impairment: A Systematic Review and Meta- Analysis. J Alzheimers Dis 70, 107- 120, doi: 10.3233 / JAD-180847 (2019).16 Ibnidris, A. et al. Investigating the Association Between Polygenic Risk Scores for Alzheimer's Disease With Cognitive Performance and Intrinsic Functional Connectivity in Healthy Adults. Front Aging Neurosci 14, 837284, doi: 10.3389 / fnagi.2022.837284 (2022).17 Stephan, K. E. & Friston, K. J. Analyzing effective connectivity with functional magnetic resonance imaging. Wiley Interdiscip Rev Cogn Sci 1, 446-459, doi: 10.1002 / wcs.58 (2010).18 Benhamou, E. et al. The neurophysiological architecture of semantic dementia: spectral dynamic causal modelling of a neurodegenerative proteinopathy. Sci Rep 10, 16321, doi : 10.1038 / s41598-020-72847- 1 (2020).19 Thomas, G. E. C. et al. Changes in both top-down and bottom -up effective connectivity drive visual hallucinations in Parkinson's disease. Brain Commun 5, fcac329, doi: 10.1093 / braincomms / fcac329 (2023).20 Wu, X. et al. Altered default mode network connectivity in Alzheimer's disease— a resting functional MRI and Bayesian network study. Hum Brain Mapp 32, 1868-1881, doi: 10.1002 / hbm.21153 (2011).21 Li, R. et al. Bayesian network analysis reveals alterations to default mode network connectivity in individuals at risk for Alzheimer's disease. PLoS One 8, e82104, doi: 10.1371 / joumal. pone.0082104 (2013).22 Nie, Y. et al. Spectral dynamic causal modelling of resting-state fMRI: an exploratory study relating effective brain connectivity in the default mode network to genetics. Stat Appl Genet Mol Biol 19, doi: 10.1515 / sagmb-2019-0058 (2020).23 Scherr, M. et al. Effective connectivity in the default mode network is distinctively disrupted in Alzheimer's disease-A simultaneous resting-state FDG-PET / fMRI study. Hum Brain Mapp 42, 4134-4143, doi: 10.1002 / hbm.24517 (2021).24 Huang, J., Jung, J. Y. & Nam, C. S. Estimating effective connectivity in Alzheimer's disease progression: A dynamic causal modeling study. Front Hum Neurosci 16, 1060936, doi: 10.3389 / fnhum.2022.1060936 (2022).25 Mohammadian, F. et al. Effective Connectivity Evaluation of Resting- State Brain Networks in Alzheimer's Disease, Amnestic Mild Cognitive Impairment, and Normal Aging: An Exploratory Study. Brain Sci 13, doi: 10.3390 / brainscil3020265 (2023).26 Luo, X. et al. Altered effective connectivity anchored in the posterior cingulate cortex and the medial prefrontal cortex in cognitively intact elderly APOE epsilon4 carriers: a preliminary study. Brain Imaging Behav 13, 270-282, doi: 10.1007 / sl l682-018-9857-5 (2019).27 Penny, W., Iglesias-Fuster, J., Quiroz, Y. T., Lopera, F. J. & Bobes, M. A. Dynamic Causal Modeling of Preclinical Autosomal-Dominant Alzheimer's Disease. J Alzheimers Dis 65, 697-711, doi: 10.3233 / JAD-170405 (2018).28 Friston, K. J., Kahan, J., Biswal, B. & Razi, A. A DCM for resting state fMRI. Neuroimage 94, 396-407, doi : 10.1016 / j . neuroimage.2013.12.009 (2014).29 Tsvetanov, K. A. et al. Activity and Connectivity Differences Underlying Inhibitory Control Across the Adult Life Span. J Neurosci 38, 7887-7900, doi:10.1523 / JNEUROSCI.2919-17.2018 (2018).

Claims

CLAIMS1. A computer-implemented method of assessing a human patient, the method comprising: receiving rs-fMRI data of the patient’s brain; determining a measure of effective connectivity of the patient’s default mode network based on the rs-fMRI data; and deriving a dementia prognosis for the patient based on the measure of effective connectivity.

2. A computer-implemented method according to claim 1 wherein the measure of effective connectivity is determined based on a set of regions of interest comprising at least 8, desirably at least 10 brain regions.

3. A computer-implemented method according to claim 1 or 2 wherein the measure of effective connectivity is determined based on a set of regions of interest consisting of no more than 20, desirably no more than 15, more desirably no more than 10 brain regions.

4. A computer-implemented method according to claim 2 or 3 wherein the set of regions of interest comprises a plurality of core default mode network regions and at least one medial temporal lobe region.

5. A computer-implemented method according to claim 2 or 3 wherein the set of regions of interest is selected from the group consisting of anterior medial prefrontal cortex (amPFC); precuneus (PRC); left / right intraparietal cortex (UPC / rlPC); ventromedial prefrontal cortex (vmPFC); dorsomedial prefrontal cortex (dmPFC); left / right lateral temporal cortex (ILTC / rLTC); left / right parahippocampal formation (IPHF / rPHF); hippocampal formations; temporal pole; and temporoparietal junction.

6. A computer-implemented method according to claim 2, 3 or 4 wherein the set of regions of interest comprises four midline ROIs (e.g. precuneus (PRC), anterior medial prefrontal cortex (amPFC), dorsomedial prefrontal cortex (dmPFC) and ventromedialprefrontal cortex (vmPFC)), one ROI in each medial temporal lobe, one ROI in each of the left and right parahippocampal formations (IPHF / rPHF), and for lateral ROIs (e.g. right intraparietal cortex (rIPC), left intraparietal cortex (UPC), right lateral temporal cortex (rLTC) and left lateral temporal cortex (1LTC)).

7. A computer-implemented method according to any one of the preceding claims wherein the measure of effective connectivity is determined using dynamic causal modelling.

8. A computer-implemented method according to claim 7 wherein the dynamic causal modelling is applied to a blood oxygen level-dependent (BOLD) signal for a plurality of ROIs.

9. A computer-implemented method according to any one of the preceding claims wherein the measure of effective connectivity comprises a matrix, where each cell in the matrix shows the magnitude and valence (excitatory or inhibitory) of a connection between a pair of ROIs.

10. A computer-implemented method according to claim 9 wherein determining a measure of effective connectivity comprises obtaining a reduced matrix as the measure of effective connectivity, the reduced matrix representing parameters obtained by applying Bayesian model reduction to training data.

11. A computer-implemented method according to any one of the preceding claims wherein deriving a dementia prognosis comprises inputting the measure of effective connectivity to a trained classifier.

12. A computer-implemented method according to claim 11 wherein the trained classifier is a regularised logistic regression model.

13. A computer-implemented method according to any one of the preceding claims wherein the dementia prognosis comprises a dementia risk value and a time to symptoms value.

14. A computer-implemented method of training a model for use in deriving a dementia prognosis for a patient, the method comprising: receiving rs-fMRI data of the brains of a plurality of cases and controls; receiving dementia diagnosis data of the plurality of cases and controls; determining a measure of effective connectivity of the default mode network of each of the plurality of cases and controls based on the rs-fMRI data; and training the model based on the measures of effective connectivity and dementia diagnosis data for each of the plurality of cases and controls.

15. A computer-implemented method according to claim 14 wherein the model is a classifier, such as a regularised logistic regression model or an artificial neural network.

16. A computer-implemented method according to claim 15 wherein the model is trained using leave-one-out cross-validation.

17. A computer-implemented method according to claim 14, 15 or 16 wherein the measure of effective connectivity is determined based on a set of regions of interest comprising at least 8, desirably at least 10 brain regions.

18. A computer-implemented method according to any one of claims 14 to 17 wherein the measure of effective connectivity is determined based on a set of regions of interest consisting of no more than 20, desirably no more than 15, more desirably no more than 10 brain regions.

19. A computer-implemented method according to claim 17 or 18 wherein the set of regions of interest comprises a plurality of core default mode network regions and at least one medial temporal lobe region.

20. A computer-implemented method according to claim 17 or 18 wherein the set of regions of interest is selected from the group consisting of anterior medial prefrontal cortex (amPFC); precuneus (PRC); left / right intraparietal cortex (UPC / rlPC); ventromedial prefrontal cortex (vmPFC); dorsomedial prefrontal cortex (dmPFC); left / right lateral temporal cortex (ILTC / rLTC); left / right parahippocampal formation (IPHF / rPHF); hippocampal formations; temporal pole; and temporoparietal junction.

21. A computer-implemented method according to claim 17 or 18 wherein the set of regions of interest comprises four midline ROIs (e.g. precuneus (PRC), anterior medial prefrontal cortex (amPFC), dorsomedial prefrontal cortex (dmPFC) and ventromedial prefrontal cortex (vmPFC)), one ROI in each medial temporal lobe, one ROI in each of the left and right parahippocampal formations (IPHF / rPHF), and for lateral ROIs (e.g. right intraparietal cortex (rIPC), left intraparietal cortex (1IPC), right lateral temporal cortex (rLTC) and left lateral temporal cortex (1LTC)).

22. A computer-implemented method according to any one of claims 14 to 21 wherein the measure of effective connectivity is determined using dynamic causal modelling.

23. A computer-implemented method according to claim 22 wherein the dynamic causal modelling is applied to a blood oxygen level-dependent (BOLD) signal for a plurality of ROIs.

24. A computer-implemented method according to any one of claims 14 to 23 wherein the measure of effective connectivity comprises a matrix, where each cell in the matrix shows the magnitude and valence (excitatory or inhibitory) of a connection between a pair of ROIs.

25. A computer-implemented method according to claim 24 wherein determining a measure of effective connectivity comprises obtaining a reduced matrix as the measure of effective connectivity, the reduced matrix representing parameters obtained by applying Bayesian model reduction to training data.

26. A computer program comprising instructions that, when executed by a computer system, cause the computer system to perform the method of any preceding claim.

27. A computer program according to claim 26 that is an update or add-on to operating software of an MRI scanner.

28. A non-transitory computer-readable medium having stored thereon a computer program according to claim 26.

29. An MRI scanner comprising one or more processors and a memory, the memory storing instructions that, when executed by the processor(s), cause the MRI scanner to perform the method of any of claims 1 to 12.

Images