Clinical brain network identification device

Abstract

Disclosed herein is a clinical brain network identification device that includes: a pre-processing module for conditioning sensed data from at least one patient; a storage medium for storing conditioned sensed data and corresponding patient metadata; a stochastic data selection module for selecting data from said storage medium that is conditioned sensed data and patient metadata relating to a set of patients; a brain network dynamics generator for receiving clinical scores and stimulus data and generating brain network dynamics being probabilities of a brain state being active at a point in time; a brain network pattern generator for generating brain network patterns, based on said selected data; a brain state inference module for determining inferred brain states based on sensed data and brain network patterns and brain network dynamics; and a therapeutic group prediction module for producing a set of therapeutic group assignments based on inferred brain states.

Description

CLINICAL BRAIN NETWORK IDENTIFICATION DEVICERelated Applications

[0001] This application is related to Australian Provisional Patent Application No.2022903932 titled “Closed-loop, non-invasive brain stimulation system and method relating thereto” and filed 21 December 2022, United Kingdom Patent Application No.2219341.1 filed 21 December 2022, International Patent Application No. PCT / AU2023 / 051354 filed 21 December 2023, and Australian Provisional Patent Application No.2023904187 titled “System and method for identifying brain network patterns” and filed 21 December 2023, the entire content of each of which is incorporated by reference as if fully set forth herein. Technical Field

[0002] The present disclosure relates to a system and method for identifying brain network derived biomarkers, including, for example, brain network patterns. In particular, the present disclosure relates to a system and method for identifying transient brain state attributes (brain network patterns and brain network dynamics) that differentiate therapeutic groups to facilitate making clinical diagnostic, prognostic, monitoring or screening predictions, or for the use of those transient brain network patterns as a therapeutic target in their own right, such as in a device of the type described in International Patent Application No. PCT / AU2023 / 051354. Background

[0003] The fields of psychiatry and neurology have struggled to identify brain network derived biomarkers that differentiate therapeutic groups for a number of medical illnesses. Many psychiatric and neurological illnesses are diagnosed and treated with reference to clinical and psychological assessments that focus on changes in subjective experience and behavioural expression arising from the illness.

[0004] Diagnostic assessments include, for example, questionnaires administered to diagnose and monitor mental health illnesses, such as depression and anxiety. Diagnostic assessments also include cognitive function tests that are performed to diagnose and monitor neurodegenerative disorders. Examples of clinical and psychological assessments used for diagnosing psychiatric and neurological illnesses include self-report rating scales (such as the K-10, BDI, DASS or PHQ-9), observer-rated scales (such as the MADRS or Hamilton rating scale), or clinical interview questionnaires.

[0005] The clinical scores generated in these assessments arise as a result of changes in underlying brain physiology, and as such reflect an indirect or proxy measure of disease progression. More direct measures, such as those reflected in neurophysiological datarecorded directly from the activity of large scale brain networks, could provide a valuable clinical decision support function and more directly inform individual diagnosis, prognosis, and therapy.

[0006] Identification of robust changes in the activity of large scale brain networks is challenging. A major reason for this is that differences in brain function are transient and arise only in specific contexts. For example, major depressive disorder has been linked to abnormal patterns of connectivity in the default mode network of a brain of a patient. These differences are only expressed during periods in which the default mode network of the brain is activated, which coincides with periods of internally oriented attention and mind wandering. Furthermore, a number of brain activity patterns may be present in health states and disease states, but those brain activity patterns may then endure for longer or shorter periods in a particular disease state.

[0007] Currently, few methods exist to guide the decision making of patients and clinicians with personalised information based on a patients’ own neurophysiology. This problem is most acute for patients who have several treatment options from which to choose. For example, in treatment-resistant depression, psychiatrists may offer patients the choice of undergoing electroconvulsive therapy or transcranial magnetic stimulation. These two treatments both present significant costs to the patient, as well as carrying different risk and efficacy profiles. In both cases, the likelihood of response is not much above 60%.

[0008] Currently, there is very little information available to patients and psychiatrists to support them in this choice, such as, for example, by assigning a patient to a therapeutic group based on recorded brain data from that patient that suggests the patient is more likely to respond to one or other course of treatment. Similarly, patients who have not responded to one form of medication (such as selective serotonin reuptake inhibitors) may or may not be well indicated to respond to another form of antidepressant medication, such as tricyclic antidepressants.

[0009] Thus, a need exists to provide a system and method for identifying brain network derived biomarkers. Summary

[0010] The present application provides a system and method for the identification of brain network derived biomarkers. In particular, the method and system identify brain state attributes that activate transiently in time in a manner that differentiates therapeutic groups. This enables identification of brain network derived biomarkers that can be embedded into clinical practice and inform individualised patient care.

[0011] Embodiments described herein provide a device and supervised learning method for identifying transient changes in neurophysiological data that differentiate therapeutic groups. The method models transient brain network patterns and corresponding brain network dynamics and anticipates long segments of data being statistically indistinguishable between therapeutic groups, interspersed with transient short-lived patterns emerging that distinguish the groups.

[0012] A first aspect of the present disclosure provides a clinical brain network identification device that includes: a pre-processing module for receiving sensed data from at least one patient and conditioning the sensed data; a first storage medium for storing the conditioned sensed data and corresponding patient metadata; a stochastic data selection module for selecting data from the first storage medium, the selected data being conditioned sensed data and patient metadata relating to a set of one or more patients; a brain network dynamics generator configured to receive clinical scores and stimulus data from the selected data and generate brain network dynamics being probabilities of a brain state being active at a point in time given the brain state activations preceding that point in time; a brain network pattern generator configured to receive the selected data and generate brain network patterns, based on the selected data; a brain state inference module configured to determine inferred brain states based on sensed data received from the stochastic data selection module and brain network patterns received from the brain network pattern generator and the brain network dynamics received from the brain network dynamics generator; and a therapeutic group prediction module for receiving inferred brain states from the brain state inference module and producing a set of therapeutic group assignments.

[0013] In some embodiments, the device applies a model that has been factorised into constituent components, wherein the factorisation entails: the brain network dynamics generator storing a probabilistic specification of brain network dynamics defining the probability of a brain network being active at a point in time given the brain state activations preceding that point in time; the brain network pattern generator storing a probabilistic specification of brain network patterns defining the spatio-spectral pattern of activity associated with recorded brain data from a given subject being attributed to a particular brain network; and the brain state inference module applying the brain network patterns and brain network dynamics to identify and store the brain network most likely to be active at any point in time.

[0014] In some embodiments, the sensed data includes neurophysiological data.

[0015] In some embodiments, the neurophysiological data is selected from the group consisting of: electroencephalogram (EEG) data), MEG data, OP-MEG data, and fNIRS data.

[0016] In some embodiments, the conditioning includes at least one of: buffering the sensed data, removing artefacts from the sensed data, and formatting the sensed data into a predefined format suitable for processing.

[0017] In some embodiments, the brain network dynamics generator and corresponding elements of the brain state inference module generate said brain network dynamics with a model selected from the group consisting of: a Time-Delay Embedded Hidden Markov Model (HMM); recurrent neural networks; temporal convolutional neural networks; and transformer architectures.

[0018] In some embodiments, the brain network pattern generator module and corresponding elements of the brain state inference module generate brain network patterns selected from the group consisting of: a spatio-spectral transform with multivariate gaussian observation model; a transformer architecture; a recurrent neural network architecture; variational autoencoder architecture; and a multilayer neural network architecture.

[0019] A second aspect of the present disclosure provides a system that includes: a clinical brain network device according to the above-described device, wherein the sensed data includes firs pre-existing sensed data relating to patient neurophysiological data derived from a patient population assigned to a set of therapeutic groups; and sensor equipment for acquiring newly sensed data from a new patient whose therapeutic group is unknown to present to the clinical brain network device; wherein the sensor equipment is selected from the group consisting of: an EEG headset, an MEG scanner, an OPM-MEG headset, and a fNIRS headset.

[0020] In some embodiments, the system further includes: an input device configured to receive input from the new patient; a display device; and a computer coupled to the input device and the display device, the computer including: a processor; and a memory for storing computer code instructions that when executed on the processor of the computer: display a game on the display device; record decisions made by the new patient at each point in time in response to the game; andtransmit game data to the clinical brain network device, wherein the game data include game events and corresponding patient data for each timepoint; wherein the clinical brain network device: processes newly sensed data sensed from the new patient by the sensor equipment and the game data to identify task-related transient expression of brain networks for that new patient.

[0021] In some embodiments, the clinical brain network device: determines, based on the identified task-related transient expression of brain networks for that new patient, the therapeutic group to which the new patient belongs for at least one of diagnostic, prognostic, monitoring or screening purposes.

[0022] In some embodiments, the game is any form of visual stimuli presented over time.

[0023] In some embodiments, the game is a game modelled on a computational psychiatry paradigm, in which a decision is required at different points in time, wherein the input device is configured to receive the decision from the patient.

[0024] In some embodiments, the input device is at least one of: a keyboard, a mouse, a stylus, a trackball, a steering wheel, or a drawing tablet.

[0025] A third aspect of the present disclosure provides a method for identifying brain network derived biomarkers, including the steps of: acquiring initial data relating to at least one of a neurological or psychiatric condition; training a model, using the initial data, to identify differences between a plurality of therapeutic groups, wherein the trained model generates trained brain state attributes; using the trained model brain state attributes to make a new prediction, wherein the new prediction is one of: (i) an application in which the new prediction relates to when in time a brain network derived biomarker is activated; and (ii) a prediction as to which of the therapeutic groups an individual belongs for at least one of diagnostic, prognostic, monitoring or screening purposes.

[0026] In some embodiments, the initial data includes neurophysiological data.

[0027] In some embodiments, the neurophysiological data is selected from the group consisting of: electroencephalogram (EEG) data), MEG data, OP-MEG data, and fNIRS data.

[0028] In some embodiments, the initial data includes patient metadata.

[0029] In some embodiments, the model is based on a Bayesian model that can be factorised into constituent components of the above-mentioned device.

[0030] According to another aspect, the present disclosure provides an apparatus for implementing any one of the aforementioned methods.

[0031] According to another aspect, the present disclosure provides a computer program product including a computer readable medium having recorded thereon a computer program that when executed on a processor of a computer implements any one of the methods described above.

[0032] Other aspects of the present disclosure are also provided. Brief Description of the Drawings

[0033] One or more embodiments of the present disclosure will now be described by way of specific example(s) with reference to the accompanying drawings, in which:

[0034] Fig.1 is a diagram illustrating architecture of a Clinical Brain Network Identification Device in accordance with one or more embodiments of the present disclosure;

[0035] Fig.2A is a schematic block diagram representation of a pre-processing module in accordance with one or more embodiments of the present disclosure;

[0036] Fig.2B is a schematic block diagram representation of an alternative embodiment of a pre-processing module in accordance with one or more embodiments of the present disclosure;

[0037] Fig.3A is a schematic representation of a brain state inference module in accordance with one or more embodiments of the present disclosure;

[0038] Fig.3B is a schematic representation of an embodiment of a brain state inference module with a transformer architecture;

[0039] Fig.3C is a schematic representation of an embodiment of a brain state inference module with a Recurrent Neural Network (RNN);

[0040] Fig.4 is a schematic representation of an architecture of a brain dynamics generator on which one or more embodiments of the present disclosure may be practised;

[0041] Fig.5A is a schematic representation of an architecture of a brain network generator in accordance with one or more embodiments of the present disclosure;

[0042] Fig.5B is a schematic representation of an architecture of a brain network generator in accordance with one or more embodiments of the present disclosure;

[0043] Fig.6 is a schematic representation of an architecture of a brain network pattern generator in accordance with one or more embodiments of the present disclosure;

[0044] Fig.7 is a flow diagram illustrating a method in accordance with the present disclosure;

[0045] Fig.8 illustrates an example of patient stratification in accordance with one or more embodiments of the present disclosure;

[0046] Fig.9 is a schematic block diagram representation of a system that includes a general purpose computer on which one or more embodiments of the present disclosure may be practised;

[0047] Fig.10 shows images highlighting activation of a brain network. In particular, activated portions of a brain, corresponding to a brain network, are highlighted during an activated state;

[0048] Fig.11 shows spatio-spectral brain network patterns corresponding to the activation of a brain network from Fig.10;

[0049] Fig.12 shows brain state probability over time, based on the brain network patterns of Fig.11;

[0050] Fig.13 is a schematic representation of a mathematical model for one or more embodiments of the present disclosure;

[0051] Fig.14 is a schematic representation of a game displayed on a computer display to the patient during acquisition of EEG data; and

[0052] Fig.15 is a schematic representation of a system utilising a brain network identification device of the present disclosure.

[0053] Method steps or features in the accompanying drawings that have the same reference numerals are to be considered to have the same function(s) or operation(s), unless the contrary intention is expressed or implied. Detailed Description

[0054] The present disclosure provides a method and system for the identification of brain network derived biomarkers. Identifying transient brain state attributes that differentiate therapeutic groups enables the determination of brain network derived biomarkers, which may be used for a number of purposes. In particular, in circumstances in which brain network derived biomarkers are identified that separate therapeutic groups corresponding to healthy controls and patients with a specific illness, those brain network derived biomarkers may be used in diagnostic applications. That is, brain network activity of a patient is recorded and compared against brain network derived biomarkers associated with separate therapeutic groups to inform diagnosis for a particular illness for that patient.

[0055] In other applications, brain network derived biomarkers are used in prognostic applications in circumstances in which brain network derived biomarkers separate therapeuticgroups corresponding to responders and non-responders of a particular therapy. In such applications, a patient’s brain network activity is recorded and compared against brain network derived biomarkers to inform a determination as to whether the individual is a suitable candidate for a particular course of treatment.

[0056] In other applications, brain network derived biomarkers are used in monitoring applications in circumstances in which brain network derived biomarkers are predictive of the state of an illness. In such applications, a patient’s brain network activity is recorded at least once, and preferably at regular or scheduled intervals over a monitoring period, and compared against brain network derived biomarkers to inform an assessment of a current disease state of the patient and progression of the disease state over the monitoring period.

[0057] In other applications, brain network derived biomarkers are used in screening applications in circumstances in which brain network derived biomarkers are predictive of the likelihood of an asymptomatic individual developing a disease. In such applications, an asymptomatic individual’s brain network activity is recorded and compared against brain network derived biomarkers to inform early detection of said disease.

[0058] In other applications, brain network derived biomarkers that separate any form of disease state from a state of recovery are used directly in therapeutic applications. This includes, for example, methods that administer brain network targeted neuromodulation, where the method normalises the expression of this brain network pattern as an explicit form of therapy.

[0059] Brain network derived biomarkers should be understood to encompass a range of patterns in recorded brain data that can be probabilistically mapped against certain disease states. The term “brain network derived biomarkers” encompasses a set of attributes and not any single binary outcome measure, although the full set of attributes may be used to generate a single binary prediction when making therapeutic group predictions. This should be understood differently from more commonly used biomarkers, such as those in, for example, diagnostic pathology, which may apply a binary test for the presence or absence of an identified biomarker. Brain imaging derived biomarkers instead refer to the relative levels of expression of multiple attributes, which can be used to make a probabilistic assignment against certain therapeutic groups.

[0060] Therapeutic groups should be understood to encompass any subset of the full range of clinical classifications that can be assigned to patients or patient groups in the course of clinical practice. In particular, the term “therapeutic groups” encompasses diagnostic classifications such as patients with mild depression and healthy controls who have not been diagnosed ashaving mild depression; prognostic classifications such as responders and non-responders for a particular treatment; and screening classifications such as stratification of risk of developing some illness; and monitoring applications such as monitoring changes in biomarker expression over the course of an administered treatment.

[0061] The present disclosure relates to a method for: (i) selecting and processing relevant data; (ii) training a family of predictive models; and either one or both of (iii) applying those predictive models to some new subject for the purpose of making new predictions; or (iv) applying the output of this predictive model for use as a therapeutic outcome target.

[0062] Australian Provisional Patent Application No.2022903932 entitled “Closed-loop, non- invasive brain stimulation system and method relating thereto” and filed 21 December 2022, and United Kingdom Patent Application No.2219341.1 filed 21 December 2022, and International Patent Application No. PCT / AU2023 / 051354 filed 21 December 2023 describe the use of brain network patterns as therapeutic outcome targets, in which case new predictions about when the therapeutic outcome targets are detected can be used to activate a therapeutic intervention and the content of those applications is incorporated by reference as if fully set forth herein. Alternatively, the set of brain network patterns can be used to make predictions about some new individual, in particular the therapeutic group to which the individual most likely belongs. This has significant applications in psychiatry, where brain network patterns and the changes in the brain network patterns can be used to predict the best course of treatment for an individual (e.g., prognostic) or can be used to support diagnosis, screening or monitoring of more complex psychiatric indications. 1. Overall Device Architecture

[0063] Fig.1 is a diagram illustrating the architecture of a Clinical Brain Network Identification Device 100 that implements the method described herein. The device 100 takes as inputs sensed data 102 for a plurality of subjects (such as, for example, digital EEG data) and corresponding patient metadata 104 and outputs a set of Brain State Attributes 190 and a set of Therapeutic Group Assignments 195.

[0064] In the example of Fig.1, the sensed data 102 is digital EEG data, which is passed to a Preprocessing module 110, the output of which is stored in a database 120. The corresponding patient metadata 104 is stored directly in this database 120. The database 120 is queried by a stochastic data selection module 130, which then feeds a set of Clinical Scores and Stimulus Data to a Brain Network Dynamics Generator 140. The stochastic data selection module 130 also feeds a corresponding set of Non-Physiological Data to a Brain Network Pattern Generator 160 and the Pre-processed EEG Data to a Brain State Inference Module 150.

[0065] The Brain Network Dynamics Generator 140 presents an output to the Brain State Inference Module 150. The Brain Network Pattern Generator 160 outputs a set of brain network patterns and also presents that set of brain network patterns as an input to the Brain State Inference Module 150. The Brain State Inference Module 150 processes the inputs received from each of the Stochastic data selection module 130, the Brain Network Pattern Generator 160 and the Brain Dynamics Generator 140 to produce and output a set of inferred brain states to a Therapeutic Group Prediction Module 170. The Therapeutic Group Prediction Module 170 then outputs a set of Therapeutic Group Assignments 195.

[0066] The Clinical Brain Network Identification Device is configured to run two subroutines. These subroutines are called “Identify” and “Predict”.

[0067] The “Identify” subroutine outputs the set of brain state attributes 190, which include brain network patterns and brain network dynamics, that best differentiate therapeutic groups, based on the therapeutic group assignments given. The therapeutic group assignment associated with each patient is provided in an appropriate format in the input Patient Metadata 104. The architecture of the device 100 is designed to identify transient differences in brain patterns and builds on the neuroscientific observation that brain network activity is statistically indistinguishable between groups for long periods, with only transient periods of time where differences in Brain Network Patterns reliably emerge.

[0068] The “Predict” subroutine is performed by the Brain State Inference Module 150 and the Therapeutic Group Prediction Module 170 and outputs a set of Therapeutic Group Assignments 195 for an individual or a set of individuals. The set of Therapeutic Group Assignments 195 is predicted on the basis of the subject’s recorded neurophysiological data 102, such as EEG data or MEG data or OP-MEG data or fNIRS data, for example, and the established set of Brain State Attributes that differentiate therapeutic groups. The established set of Brain State Attributes may be stored in the database 120, be output by the Brain Network Pattern Generator 160 and the Brain Network Dynamics Generator 140, derived from an external source, or any combination thereof.

[0069] The “Predict” subroutine can only be run after the “Identify” subroutine. The Predict subroutine does not require the patient’s therapeutic group assignment to be included in the patient metadata, thus allowing diagnostic, prognostic, monitoring or screening predictions to be made for a new patient whose actual therapeutic group assignment is unknown. If a patient’s therapeutic group assignment is included in the patient metadata, this may instead be utilised for the purpose of validating the clinical predictive value of the Brain Network Derived Biomarkers 190 that have been identified.

[0070] A pre-processing module 110 of the Clinical Brain Network Identification Device 100 pre-processes the neurophysiological data 111. One or more embodiments of a suitable pre- processing module 110 are described in International Patent Application No. PCT / AU2023 / 051354, such as with reference to Fig.3 of PCT / AU2023 / 051354. The pre- processing performed by the pre-processing module 110 conditions the received data to be suitable for subsequent processing and may be utilised, for example, to buffer the received signals, remove artefacts, format the signals into a predefined format suitable for processing, filter extraneous data, or any combination thereof. The pre-processing performed on acquired data is identical over patient and control groups. Preferably, the pre-processing matches the pre-processing steps anticipated to be applied in any future intervention. For example, where the therapeutic targets are intended for use in a real-time EEG application, the preprocessing steps are limited to those steps that can be applied to real-time EEG and exclude steps that may be applied in offline settings.

[0071] The Pre-processed data is then stored in a database 120. The database 120 may be implemented using any computer-readable storage medium. Depending on the implementation, the database 120 may be integral with the Clinical Brain Network Identification Device 100, such as a hard disk drive. In alternative embodiments, the database 120 is an external storage device, such as an external hard disk drive, external server coupled via a communications network, or cloud-hosted storage.

[0072] The Database 120 is then accessed by a Stochastic Data Selection Module 130. In some embodiments designed to run on very large datasets, the Stochastic Data Selection Module 130 implements stochastic batch file reading, whereby each iteration within the “Identify” subroutine is run on a different randomly selected subset of the data in the database 120, wherein the selected data includes matching neurophysiological data and patient metadata. In some embodiments where the available data is smaller in size, the Stochastic Data Selection Module 130 selects all available data from the database 120 to be run on each iteration within the “Identify” subroutine.

[0073] A first output, ^^,^of the Stochastic Data Selection Module 130, representing the therapeutic group assignment for subject n, is passed to the Brain Dynamics Generator Module 140. The Brain Dynamics Generator Module 140 outputs parameters summarising the brain dynamics. In some embodiments, the output brain dynamics parameters are presented as amatrix of size ^ × ^, in which the (^, ^)th value is proportional to the expected log probability ofobserving a transition from state ^ to state ^ at time ^. In other embodiments, the output brain dynamics parameters are presented as a high-dimensional parameter set representing the internal parameters of a neural network, such as a recurrent neural network. The output braindynamics parameters are presented as an input to the Brain State Inference Module 150. One or more embodiments of a suitable Brain State Inference Module 150 are described in International Patent Application No. PCT / AU2023 / 051354, such as with reference to Fig.4 of PCT / AU2023 / 051354.

[0074] When set to the “Predict” mode, this output represents the predicted brain network dynamics as determined by the subject’s clinical group assignment ^^,^.

[0075] When set to the “Identify” mode, the internal parameters of the Brain Dynamics Generator Module 140 are optimised by the methods described below, such that the internal module parameter Φ identifies specific brain network dynamics that differentiate therapeutic groups.

[0076] A second output, ^^, of the Stochastic Data Selection Module 130, reflecting patient metadata that is expected to inform individual subject differences in how brain networks are expressed in neurophysiological patterns, is passed to the Brain Network Pattern Generator Module 160. The Brain Network Pattern Generator Module 160 implements a function f that predicts the subject specific brain network pattern, allowing calibration and warping of individual brain network patterns on the basis of individual differences expressed in ^^.

[0077] When set to the “Predict” mode, the Brain Network Pattern Generator Module 160 outputs the function ^(^^) that provides the best prediction of the set of ^ Brain Network Patterns for subject ^. In some embodiments intended for use in real-time brain state modulation, such as described in International Patent Application No. PCT / AU2023 / 051354, these Brain Network Patterns are a cell of numbers of dimension (nch + 1) x nch x K, representing the statistical mean and covariance associated with each brain state. In other embodiments, the Brain Network Patterns are a high-dimensional parameter set representing the internal parameters of a neural network, such as a transformer architecture.

[0078] When set to the “Identify” mode, the internal parameters of Brain Network Pattern Generator Module 160 are optimised by the methods described below, such that the output parameter ^(^^)specifically identifies brain network patterns that display different dynamics over the different groups.

[0079] A third output ^^,^of the Stochastic Data Selection Module 130, representing the pre- processed neurophysiological data vector for subject n at time t, is passed to the Brain State Inference Module 150. Consistent with implementations of brain state modulation, such as described in International Patent Application No. PCT / AU2023 / 051354, the Brain State Inference Module 150 outputs a variable ^^representing the probability of each brain network from the predefined set of brain networks being active at time ^.

[0080] The Stochastic Data Selection Module 130, Brain Network Dynamics Generator 140, Brain Network Pattern Generator 160, and Brain State Inference Module 150 are the core components of the “Identify” subroutine of the Clinical Brain Network Identification Device 100. In some embodiments, the “Identify” subroutine performs iterative variational updates that alternate between optimising the Brain Network Dynamics Generator 140 and optimising the Brain Network Pattern Generator 160. In these embodiments, optimisation is measured through the variational free energy.

[0081] In one iteration, the Brain Network Dynamics Generator 140 is optimised by minimisation of the variational free energy given the output Brain Network Pattern Generator 160. In a subsequent iteration, the Brain Network Pattern Generator 160 is optimised by minimisation of the variational free energy given the output of the Brain Network Dynamics Generator 140. This procedure is iterated until a stopping criterion is reached, signifying that the Brain State Attributes that differentiate therapeutic groups have been identified.

[0082] The Brain State Attributes that differentiate therapeutic groups encompass one or more Brain Network Derived Biomarkers, which can be isolated to explain the neurophysiological basis of differences between therapeutic groups, or utilised collectively in the context of the full set of Brain State Attributes.

[0083] This concludes the “Identify” subroutine.

[0084] The “Predict” subroutine can be called, which outputs a set of one or more Therapeutic Group Assignments 195 for an individual or a set of individuals. The core component of the “Predict” subroutine is the Therapeutic Group Prediction Module 170. The Therapeutic Group Prediction Module 170 receives as input the inferred brain states for a subject, as output by the Brain State Inference Module 150, and outputs the therapeutic group to which they are assigned, completing the “Predict” subroutine.

[0085] In one or more embodiments, an overall architecture of the Clinical Brain Network Identification device 100 of Fig.1 is motivated by the Bayesian model structure of Fig.13. The device 100 is designed to approximate solutions to a number of different models that follow this general overarching architecture, wherein the architecture factorises the model into constituent components representing different elements of the model, such as brain network patterns and brain network dynamics, in a manner that supports computationally feasible model inference on large datasets. It will be understood that other architectures of the Clinical Brain Network Identification device 100 may equally be practised within the spirit and scope of the present disclosure.

[0086] In some embodiments, the Clinical Brain Network Identification device is applied to EEG data recorded from patients for 5 minutes before and 5 minutes after undergoing their first session of Transcranial Magnetic Stimulation therapy. The device is run in the “Identify” subroutine, wherein the Stochastic Data Generation Module selects segments of EEG data ^^,^; corresponding therapeutic group assignment data ^^,^, which is a binary variable indicating whether that patient achieved a 50% reduction in clinical depression scores after undergoing 20 sessions of TMS therapy and is constant for all values of time ^; and corresponding metadata ^^containing that subject’s head-shape data obtained in a Magnetic Resonance Imaging scan.

[0087] The output of the “Identify” subroutine is a set of brain network dynamics Φ and corresponding brain network patterns ^(^^) that identify the specific Brain Network Patterns – adjusted for head shape information associated with a particular patient – that taken together as Brain State Attributes 190 indicate whether the patient is likely to achieve a 50% reduction in clinical depression scores. The Clinical Brain Network Identification device can then be applied in the “Predict” subroutine for any new patient to generate a prediction of whether or not that patient is likely to achieve a 50% reduction in clinical depression scores. That is, the “Predict” subroutine executing on the Clinical Brain Network Identification device generates a prediction of whether or not that patient is a good candidate to continue on Transcranial Magnetic Stimulation therapy. 2. Module Architecture

[0088] Fig.2 is a schematic block diagram representation of one embodiment of the Preprocessing module 110 of Fig.1. The pre-processing module 110 receives sensed neurophysiological data 102, such as EEG signal data sensed by EEG hardware. The neurophysiological data 102 is processed by a memory buffer 112, which serves as a temporary storage mechanism to ensure a steady and consistent flow of transmitted samples by accommodating variations in data arrival rates. In some implementations, this memory buffer module 112 also performs resampling with polyphase anti-aliasing filtering such that the samples are transmitted out of the memory buffer 112 at a lower sampling rate to that which they were received. In some embodiments, the sampling rate after down-sampling is 100 samples per second. The actual sampling rate will depend on the application and may be in the range of 50 samples per second to 16,000 samples per second, depending on needs and processing capabilities.

[0089] In some embodiments, the neurophysiological data is then filtered by a time domain filter 114. The time domain filter 114 applies linear time invariant filtering to the data recorded from each channel / electrode, in order to remove high frequency noise and low frequency line drift. Some embodiments utilise a 3rdorder Butterworth filter with a passband of 1-45Hz. Someembodiments utilise a notch filter with the notch frequency set to the frequency of main powerline noise in the country the device is used. Some embodiments additionally detrend the data to remove slow drifts. In some embodiments this is done by fitting a 5thorder polynomial and subtracting it from the signal.

[0090] Some embodiments pass the signal to an optional artefact identification module 116, which applies an algorithm to the filtered signal to identify artefacts, irrespective of whether that signal has been filtered. One embodiment computes the standard deviation of the signal across all channels at each point in time and classifies the signal as an artefact wherever this standard deviation exceeds the 99th percentile observed from configuration data (also referred to as calibration data). The artefact identification module 116 then removes identified artefacts. In some embodiments, artefacts are identified by a suitable decomposition method, such as independent component analysis (ICA), artefact subspace reconstruction (ASR), online recursive independent component analysis (ORICA) or any combination thereof. In some embodiments, the artefact identification module 116 removes part of the signal that is an artefact. In other embodiments, the artefact identification module 116 removes an entire signal or portion of signal, such as by sending an indication to downstream processing not to process the signal until the artefact has passed.

[0091] In some embodiments, a Co-registration Module 210 co-registers the data against subject-specific spatial coordinates that model the spatial position of each recording electrode relative to the subject’s cortex. In some embodiments, this uses data recorded from the patient’s scalp using fiducial markers. In some embodiments, this is co-registered to subject structural MRI data using a multiple local sphere forward model.

[0092] In some embodiments, a Common Space Mapping Module 220 then maps subject neurophysiological data into a common space that is more directly comparable across different subjects. In some embodiments, this entails beamforming the data using the spatial coordinates from the Co-registration Module in a Linearly Constrained Minimum Variance (LCMV) Beamformer projecting onto an 8mm MNI coordinate grid of cortical locations. In some embodiments, the beamformed data is then parcellated into anatomically defined regions of interest. In some embodiments, source leakage is then corrected for by orthogonalization, as outlined in Colclough et. al.2015: A Symmetric Multivariate Leakage Correction for MEG Connectomes.

[0093] The output of the artefact identification module 116 is optionally presented to a co-registration module 210. The output of the co-registration module 210 is presented as an input to a common space module 220. The output of the common space module 220 is presented as an input to the dimensionality reduction module 118.

[0094] In some embodiments, a dimensionality reduction module 118 processes the remaining signal to reduce the inherent dimensionality of the signal. In one embodiment, the dimensionality reduction module 118 applies principal component analysis to the configuration data (also referred to as calibration data) to identify the set of linear loadings that capture 90% of the data variance. The output of the dimensionality reduction module 118 is presented as output signal 119 to be stored in the database 120 of Fig.1.

[0095] Another embodiment of the Preprocessing Module 110 configured specifically for high sample rate real-time brain state modulation techniques is represented in Fig.3 of PCT / AU2023 / 051354, and omits the Co-registration 210 and Common Space Mapping 220 steps.

[0096] Fig.3A is a schematic block diagram representation of functional modules of one embodiment of the Brain State Inference Module 150 of Fig.1, when implemented using Time- Delay Embedded HMM. The Brain State Inference Module 150 implements a set of computations that are derived from an underlying mathematical model of how the recorded data relate to activation of brain networks. The assumptions of the model, which motivate each step in the brain state inference module, include: • When different brain networks activate, the respective brain networks result in scalp potentials that differ in both the spatial and spectral properties of the resulting timeseries. • It is assumed that we have already learned the different spatial and spectral patterns unique to the activation of each brain network. • In some embodiments, standard assumptions of Hidden Markov Models are applied, specifically: (1) that these brain networks are mutually exclusive with respect to time; and (2) the sequence of brain states forms a Markov chain, i.e. that a state zt is conditionally independent of zt-n ∀^ > 1, if zt-1 is known.

[0097] Based on these assumptions, the Brain State Inference Module 150 is implemented to answer a question at each time step. If the brain network that was activated one timestep prior (denoted by zt-1) is known, and the current spatial and spectral pattern of scalp potentials (denoted by Yt) are also known, then the question to be answered is what is the most likely brain network that is currently active (denoted by zt).

[0098] The embodiment of the Brain State Inference Module 150 depicted in Fig.3A answers this question in a Bayesian manner, which in the context of the broader Bayesian structure ofFig.15, represents a factorised component of this broader model corresponding to brain state inference: )

[0099] Eqn (1) is solved by the embodiment of the Brain State Inference Module 150 shown in Fig.3A using time delay embedding, dimensionality reduction, and state likelihood computation.

[0100] As shown in Fig.3A, the Brain State Inference Module 150 receives the vector signal Xtand performs time delay embedding to create a data vector that captures both the spatial patterns expressed over different channels as well as the spectral patterns, such as the frequency of a brainwave, which can only be observed by looking at the relationships between data over successive timepoints.

[0101] Given a [P x 1] vector Xtat time t, where P is the number of neurophysiology data channels, the first step of the Time Delay Embedded HMM is to create a time embedding of thedata. This constructs a new vector %& ^ of dimension [PW x 1] where W is the length of theembedding.

[0102] The entire [P x 1] vector %'is passed through each time delay element, such that the output of 302 is a [P x 1] vector %'^(; the output of 304 is a [P x 1] vector %'^), etc.

[0103] The concatenation module 308 has W different vector inputs, being specifically %', %'^(, %'^), … %'^*+(. The concatenation module 308 then outputs a vector %,^of dimension[PW x 1], where %,^ = -%', ; %'^(; %'^); … %'^*+(0 (i.e., the row-wise concatenation of the inputvectors).

[0104] The new vector %,^is highly dimensional and is expected to contain much information that is superfluous or redundant. Accordingly, the dimensionality reduction module 310 projects the new vector %,^to a lower dimensional vector ^^of dimension [Q x 1]. In some embodiments,this is a linear dimensionality reduction operation, such that ^^ = 1 %,^, where B is a linearoperator of dimension [Q x PW] such that Yt reflects both the spatial and short term spectral profile of the neurophysiological signal. Referring to Fig.3A, the vector %,^is operated on by the B operator 310 to produce Yt. We refer to the group of operations performed to map the vector %'to the vector 2'as the spatio-spectral transform.

[0105] In some embodiments the B operator is learned through an appropriate method, such as principal component analysis. In other embodiments, the B operator is learned through an appropriate optimisation technique that optimises the spatio-spectral transform parameters for an objective function, such as therapeutic group discriminability. In some embodiments, this isa greedy optimisation algorithm that optimises the mean and variance of gaussian kernel functions in order to maximise the difference between brain network patterns obtained for the therapeutic groups. The processing of the spatio-spectral transformed data 2'then assumes a Hidden Markov Model with a multivariate Gaussian observation model: ^(^^|^^ = 3) = 4(56 ,^6) …Eqn (2)where Zt is the value of the latent state at timepoint t, μk is theis the covariance matrix for state k that is learned by an appropriate method, as described below.

[0106] Following the standard approach for real time HMMs, the process of inferring which brain state is active at time t is given by:

[0107] Some embodiments apply the variational Bayesian approximation to evaluate this, such that the Brain State Inference Module estimates the marginalised state probability at eachtimestep 8^^^ = 3^ ≈ ^^^^ = 3|^^ , ^^^^^ by evaluating:Where =>^"!#$^denotes the expectation with respect to the previous timestep’s approximate stateprobability. Referring to Fig. 3A, module 312 computes the value of log ^^^^|^^ = 3^,corresponding to the degree to which the current data matches the predefined brain networkpatterns, and module 314 computes the value = ^ log ^^^^^, corresponding tolikelihood of observing a brain state given the immediately preceding brain state (i.e., the brain dynamics).

[0108] In some embodiments, the Brain State Inference Module is run in an offline mode, such that the above computations can be run with access to data at all timepoints, rather than run on each timepoint given only the previous timepoint. These embodiments apply standard variational Bayesian methods for solving the approximate factorised posterior probability ateach timepoint 8^^^ = 3^ given all datapoints, such as the Baum Welch Forward-Backwardalgorithm.

[0109] It remains to specify the model of temporal dynamics ^(^^ = 3|^^^^^. The methodsoutlined in International Patent Application No. PCT / AU2023 / 051354 apply this model as a static estimate; the embodiments described herein instead implement a model of temporal dynamics that is non-stationary or informed by patient meta-data, in some embodimentsapplying the following model log ^^^^ = 3|^^^^, ^^,^^ = =>^?^ log ^^^^ = 3, Φ|^^^^, ^^,^^ for somelatent variable Φ modelled such as described in Table 1 and corresponding non-physiological data for patient n at time t given by ^^,^, or relax the Markov assumption and modellog ^^^^ = 3|^^:^^^, ^^,^^ = =>^?^ log ^^^^ = 3, Φ|^^:^^^, ^^,^^ such as described in Table 2.

[0110] The final output of the Brain State Inference Module 150 is a -^ × 10 vector AB^, the kthentry of which reflects the probability that state 3 is active at time ^, and in some embodiments is equal to the softmax function implemented by module 318, such that: CDEF G(H!|I!JK^L=M^I!#$^ DEF G^I!JK|I!#$,NO,!^8…Eqn (5)

[0111] In the example of Fig.3A, there are K known brain states. For each of the K brain states, the associated brain state attributes comprise an associated valuewhich isthe output of the Brain Network Pattern Generator 160, and a -^ × ^0 state transitionprobability matrix which is the output of the Brain Network Dynamics Generator 140. These values may be stored in the brain state inference module 150 or updated over time.

[0112] In some embodiments, module 318 instead computes the hardmax function, such that the final output AB^has a single entry which has probability 1 and all other entries with probability 0.

[0113] This value of AB^is the output of the brain state inference module 150 at each timepoint t, and corresponds to an inferred brain state probability that has been determined, from a set of K predefined brain states, for each point in time t.

[0114] The embodiments of the Brain State Inference Module 150 described above utilise a Time-Delay Embedded Hidden Markov Model to infer the brain state at each point in time. In some embodiments, such as those that are not intended for high sample rate real-time brain state modulation techniques, the brain state is inferred at each step in time using deep learning techniques.

[0115] Fig.3B represents an embodiment of the Brain State Inference Module 150 that is implemented by a Transformer architecture. In this embodiment, the functionality of embedding blocks 302, 304 and 306 are augmented by tokenizer blocks 362, 364 and 366. These tokenizer blocks 362, 364, 366 apply successive time delays as described previously with reference to the time delay elements 302, 304, 306 in the embodiment of Fig.3A, but also tokenise the data and apply a positional encoding 368, such that the data vector %,^, ofdimension [PW x 1], is a discretised and tokenised representation of the data suitable for subsequent presentation to a Transformer Encoder block 370.

[0116] The Transformer Encoder block 370 implements the dimensionality reductionfunctionality of module 310 using multi-head self-attention, such that ^^where Brepresents the internal parameters and weights of the transformer encoder block and ^^is the output layer of the Transformer Encoder Block which is of lower dimensionality than the data vector %,^and thereby represents a low dimensional latent embedding of the data at the timepoint ^.

[0117] In some embodiments, the transformer internal parameters B are trained by connecting an optional transformer decoder block 372, as in Fig.3B, to reconstruct the original signal ^^based on the inferred embedding, and applying backpropagation update methods based on the reconstruction error.

[0118] The output ^^of the Transformer Encoder Block 370 is presented to a Read-out feed-forward neural network block 374 such that ^^^^|^^ = 3^ = ^U^^^^, i.e., the kth output node of thefeed-forward neural network. Outputs of the Read-out forward neural network block 374 are presented to the brain network patterns modules 312a … k, wherein each respective brain network pattern module 312a…k determines a log likelihood estimation for respective brain network patterns 1…k. Outputs of the brain network patterns modules 312a…k are processed by brain network dynamics modules 314 in the same manner as described with reference to Fig.3A, which in some embodiments utilise the Hidden Markov Model architecture described above and in other embodiments utilise alternative architectures, such as recurrent neural networks described below.

[0119] In the embodiment of Fig.3B, the brain network parameters generated by the Brain Network Pattern Generator 160 correspond to the internal parameters B of the Encoder Blocks 370, Decoder Blocks 372 and read-out feed-forward neural network 374 that are used to compute the Brain Network Pattern Log Likelihoods. In one embodiment, this architecture has 6 encoder and 6 decoder blocks, each consisting of a multi-head self-attention mechanism, layer normalisation, residual connections, followed by a fully-connected layer of size 4 times the input size, and another normalisation layer.

[0120] Fig.3C represents an embodiment of the Brain State Inference Module 150 that is implemented using a recurrent neural network architecture. As for Fig.3A, received vector signal Xt, is fed to a set of time delay elements 302, 304, 306, the outputs of which are presented to a Recurrent Neural Network (RNN) 380. The RNN 380 implements the dimensionality reduction functionality of module 310 using recurrent computations, the outputlayer of which at time t is a signal ^^ where B represents the internal parameters andweights of the recurrent neural network block.

[0121] The output signal Ytcorresponds to an internal state of the RNN 380 captured by the dynamics of the system conditioned on incoming neurophysiological data, and is presented to a read out feed forward neural network 374. In some embodiments, the RNN block 380 comprises a single layer of 60 Long Short-Term Memory (LSTM) units. In this embodiment, the remaining blocks of the Brain State Inference Module are as described for the embodiment of Fig.3B.

[0122] It will be appreciated that alternative model architectures may equally be practised. In particular, consistent with the disclosure of International Patent Application No. PCT / AU2023 / 051354, the brain state inference module of Fig.3A may employ a probabilistic latent state model selected from the group consisting of: a Time Delay Embedded Hidden Markov Model (outlined above); state dynamics modelled using recurrent neural networks; non- Gaussian distributions over the raw data; and application of non-linear transformations to raw data.

[0123] Alternative Transformer architectures may include varying number and size of Encoder and Decoder blocks, depth and number of hidden units of the read-out neural network. Inside of each of encoder or decoder blocks alternative architectures within the spirit of the provided description may vary the number of self-attention heads, embedding dimensionality, sequence of normalisation layers, residual connections, and depth and size of the fully connected layer. Alternative Recurrent Neural Network architectures may vary in the size of each of recurrent layers, the number of such layers, and in type of computational units comprising the layers – this can include, for example, a Long Short-Term Memory (LSTM) unit, Gated Recurrent Unit (GRU), or an RNN unit without a gating mechanism.

[0124] Fig.4 is a schematic block diagram representing the functional composition of one embodiment of the Brain Dynamics Generator module 140 of Fig.1. The Brain Dynamics Generator module 140 generates brain dynamics patterns that are specified as a function of the therapeutic group to which a subject is assigned in a manner that enforces dynamics that are maximally dissimilar between therapeutic groups. This embodiment of the Brain Dynamics Generator module 140 includes two components, a Baseline dynamics modulator 410, and a Therapeutic bias modulator 420, which interact during the “Identify” subroutine to identify transient brain patterns that maximally distinguish therapeutic groups. The Baseline dynamics modulator 410 and Therapeutic bias modulator 420 provide summed inputs to an expected value generator 430, which provides the output of the Brain Dynamics Generator module 140.

[0125] With reference to the Bayesian model architecture specified in Fig.13, the output of the Brain Dynamics Generator module 140 represents a factorised component of this broader model corresponding to the predicted brain network dynamics, and is equal to:^^^^,^|^^,^:^^^ , Φ, ^^,^^ + log ^^ Φ^W…Eqn (6)where =>^?^denotes the expectation with respect to some probability distribution 8^Φ^, where Φ are the internal parameters of the Brain Dynamics Generator 140 and 8^Φ^ is the marginalised posterior distribution over the parameter Φ.

[0126] In some embodiments, motivated by Hidden Markov Model architectures amenable to high sample rate brain state modulation techniques such as described in International Patent Application No. PCT / AU2023 / 051354, the Brain Dynamics Generator Module applies a Markov assumption, i.e., letting log ^^^^,^|^^,1:^−1 , Φ, ^^,^^ = log ^^^^,^|^^,^−1 , Φ, ^^,^^. In theseembodiments, the Brain Dynamics Generator 140 outputs a K x K matrix, the {i,j}th entry of which is proportional to the log likelihood of transitioning from state i to state j at time t. The Brain Dynamics Generator 140 is envisaged to have at least three different modes, corresponding to different formats of the variable ^^,^. The Brain Dynamics Generator 140 accepts inputs for ^^,^that are either binary, categorical, or continuous. Each of these inputs has a corresponding mode of Baseline Dynamics Generator 410 and Therapeutic Bias Modulator 420, as indicated in Table 1 below. Each row corresponds to a different model variety reflecting the different formats of the variable ^^,^. Each row has a different form of Baseline Dynamics Generator 410^^,^^, i.e. how the current state ^^relates to the previous state ^^^^, the internal model parameters Φ, and the patient metadata ^^,^for subject ^ at timepoint ^.

[0127] To enforce that the Clinical Brain Network Identification Device 100 finds brain patterns that differentiate therapeutic groups, a Therapeutic bias modulator 420 is introduced. This applies a dissimilarity prior ^^Φ^which acts to penalise internal parameterisations that fail to differentiate the therapeutic groups, and rewards parameterisations that map different dynamic patterns between the therapeutic groups. This dissimilarity prior may take several forms, depending on the model input specification as detailed below. In all forms, the hyperparameter Z specifies the magnitude to which models are penalised that fail to differentiate therapeutic groups, and the hyperparameter A specifies noninformative values of the remaining parameters.

[0128] Corresponding forms for three embodiments of the Baseline Dynamics Generator 410 and the Therapeutic Bias Modulator 420 are outlined in table 1:Table 1: Modes of operation of the Brain Dynamics Generator

[00129] In the binary case, the input variable ^^,^ ∈ -0,10. Whilst the above utilises a subscriptfor time, in most cases it is expected this variable will be invariant over time but variable over subjects, where for example ^^,^ = 0 may denote that subject n has not been diagnosed with aparticular condition and ^^,^ = 1 may denote that subject n has been diagnosed with saidcondition.

[0130] In this embodiment, the internal model parameter Φ is a matrix of dimension (^ + 1) × ^. The first row of this matrix, the ^th column entry of which is denoted a^,b, denotesthe probability that a subject is assigned to a given therapeutic group, i.e. ^^,^ = 1, given thatstate i is active at time t. The remaining rows of the matrix, entries of which are denoted ab,_for ^∈ -1,2, … , ^0 and ^ ∈ -1,2,0, denote the probability of observing a transition from state i tostate j.

[0131] In this embodiment, the dissimilarity prior applied is a product of independent probability distributions. The method applies a product of independent Beta distributions over the first row of the matrix Φ, with shape and scale parameters both set to Z. The remaining rows of the matrix Φ are assigned independent Dirichlet distributions, with parameters given by the rows of a hyperparameter matrix A of size -^ × ^0. In some embodiments, A is set to amatrix of ones.

[0132] In this embodiment, during the ‘Identify’ subroutine, the probability distribution 8(Φ) is optimised in order to minimise the variational free energy. Since this embodiment usesconjugate priors, a further factorisation of 8(Φ)can be optimisedthrough variational inference, such that each marginal posterior distribution can be solved analytically. This allows optimisation of the overall Brain Dynamics Generator 140 to identify dynamics that best differentiate the therapeutic groups.

[0133] In the case where the input variable is categorical, with a total of G therapeutic groups,the variable ^^,^ ∈ -1, 2, … h0 denotes the therapeutic group to which subject n belongs.

[0134] In this embodiment, the brain network dynamics function can be read out from multiplelookup tables. In this embodiment, the internal parametersΦgis a [K x K] transition matrix,such that the i,jth entry of this matrix, written as ab,_,rdescribes the probability of observing a transition from state i at timepoint t-1 into state j at timepoint t for a subject in therapeutic group g. The core architectural assumption here is that the variables describing state transition dynamics are uniform over subjects within a therapeutic group, but different over therapeutic groups.

[0135] In this embodiment, the dissimilarity prior applied is formed as a penalised joint Dirichlet Distribution over the internal model parameters Φ:…Eqn (7) where the parameter λ is a penalty term that controls how strongly the method enforces the dissimilarity penalty. DKL(Φ||Φ) denotes the Kullblack-Liebler divergence between rows of the transition matrix, given by:…Eqn (8) and in which D(Φg|A) denotes the Dirichlet distribution, defined aswhere the matrix Beta function and the gamma function are defined as:…Eqns (10)(11)

[0136] In this embodiment, during the ‘Identify’ subroutine, the probability distribution 8(Φ^ is optimised in order to minimise the variational free energy. In some embodiments, this probability distribution is optimised by Markov Chain Monte Carlo sampling.

[00137] In the case where the input variable is a continuous vector, the input variable ^^,^ ∈ ℝk,i.e. the input variable is a vector of length -1 ^ h0 with entries taking real number values. In thisembodiment, the model predicts the state transition likelihood as a linear function of the input variable ^^,^. In some embodiments, the input vector ^^,^includes a constant term to model a constant offset.

[0138] In this embodiment, the input variable ^^,^can include clinical scores that relate to or reflect therapeutic groups. The input variable ^^,^can also include time varying information such as stimulus presentation information, which allows for common setups where data is recorded whilst the subject views some time-varying stimulus, such as a movie, or completes a task that is designed to elicit task-related transient brain networks that maximally differentiate the therapeutic groups being studied.

[00139] In this embodiment, the internal parameter Φ is an array of dimension -^ ^ ^ ^ h0,where G is the dimension of the input vector. The value ab,_,rreflects a linear coefficient, such that the log likelihood of observing a transition from state i to state j at time t is given by.

[0140] In this embodiment, the dissimilarity prior is implemented as an exponential distribution over the square of the coefficients:…Eqn (12)

[0141] In this embodiment, during the ‘Identify’ subroutine, the probability distribution 8^Φ^ is optimised in order to minimise the variational free energy. In one embodiment, this probability distribution is optimised by Markov Chain Monte Carlo sampling.

[0142] The above embodiments of the Brain Dynamics Generator 140 utilise a Hidden Markov Model, but it will be appreciated that alternative embodiments, in particular those that are not intended for high sample rate real-time brain state modulation techniques, use other means to generate brain dynamics patterns that are specified as a function of the therapeutic group to which a subject is assigned, such as recurrent neural networks, temporal convolutional neural networks or transformer architectures.

[0143] In some embodiments, the Brain Dynamics Generator implements the three modes of operation corresponding to binary, categorical or continuous values of the input variable ^^,^utilising a recurrent neural network. In this embodiment, the Baseline Dynamics Generator is implemented using a recurrent neural network configuration as outlined in Table 2, where RNN?,y,U(^^,^:^^^) denotes the value of the 3th output node at time ^ of a recurrent neural network with hidden state parameters Φ. In this embodiment the Therapeutic Bias Modulator outputs a corresponding function of a -h ^ ^0 matrix of observable parameters Φ^ (where byconvention h = 1 in the binary case). In some embodiments, the recurrent neural network is asingle layer comprising 60 Long Short-Term Memory units.Table 2: Embodiment of the Brain Dynamics Generator that are implemented using Recurrent Neural Networks

[0144] It will be appreciated that alternative model architectures may be employed within the spirit of the above description. Furthermore, whereas these model architectures are presented as distinct embodiments, they may be implemented in a combined or parallel setup, accomodating input vectors that mix multiple different formats.

[0145] The Brain Network Pattern Generator Module 160 implements a function f that predicts the subject specific brain network pattern, allowing calibration and warping of individual brain network patterns on the basis of individual differences expressed in ^^. Fig.5A is a schematic diagram of one embodiment of the Brain Network Pattern Generator Module 160 of Fig.1. Fig 5B is a schematic diagram of an alternative embodiment of the Brain Network Pattern Generator Module 160 of Fig.1.

[0146] In the context of the Bayesian model architectures of Fig.13, the Brain Network Pattern Generator Module 160 represents a factorised component of this broader model, corresponding to the brain network patterns, modelled as the expected value of the factorised posterior distribution over the function f, equal to the following: =>^^^Vlog ^^^^,^|^^,^ , ^, ^^) + log ^(^)W …Eqn (13)denotes the expectation with respect to some probability distribution 8(^), which approximates the posterior distribution over brain network pattern function. This function takes several forms depending on the model specification as detailed below.

[00147] In some embodiments, the input data per subject, ^^, is a vector of size -1 × ^0,reflecting patient data that is expected to inform individual subject differences in how brain networks are expressed in neurophysiological patterns. For example, this input data may reflect information about head shapes, EEG electrode positions, EEG electrode montages, and other such information that can be derived from EEG equipment or other sensor equipment used to sense patient data. In another embodiment, this patient data may reflect skull thickness, cortical folding, and other anatomical and functional information that may be obtained from another imaging modality, such as magnetic resonance imaging, functional magnetic resonance imaging, diffusion tensor imaging, positron emission tomography, or any other such imaging modality. In some embodiments, the input data vector is inferred from a subject’s data using a latent space embedding such as that applied by Shneider et al 2023 (https: / / www.nature.com / articles / s41586-023-06031-6).

[0148] In one embodiment, this function is constant over all subjects, such that no individual differences are modelled: ^6(^^) = ^6 …Eqn (14)

[0149] In another embodiment, where ^^corresponds to information of EEG spatial locations, either given as montage labels conforming to some international standard such as the 10-20 international standard, or as spatial references within a defined spatial coordinate system, thefunction f may correspond to a spatial interpolation function that maps from individual differences in recording locations the recorded data into a common space. This is represented in the right hand side of Figure 5, with the input ^^mapping onto a deterministic warpfield calculation, which in some embodiments applies a cubic spline interpolation function over the 3D spatial coordinates. This produces a set of linear warpfield coefficients ^^, such that the output is instead modelled as:

[0150] In another embodiment, this function is a multilayer neural network, which has an input vector given by ^^which is defined as a one-hot subject encoding vector of length N. This input is passed to a low dimensional embedding layer, replicating a variational autoencoder architecture, such that this layer represents the mean and standard deviation of a latent embedding space that represents inferred subject differences in brain network patterns.

[0151] This low dimensional embedding space is then scaled by a randomly sampled standard normal variable and passed through a multilayer feed forward neural network. The ^^^+^^ output layer of this network, denoted by ^^, is of dimension ^ ; it is then added to a static layer ^^of the same dimension applied to regularise the variance over subjects. This architecture is replicated K times for the K different states, with the outputs representing K unique sets of lower triangular Cholesky decomposition parameters, applying for each brain network and each individual.

[0152] In the embodiment of Figure 5, the above two functions are combined, such that the output of this embodiment can be written as:Where ^^,6is the output of the neural network for the nth subject and kth state, and ^^is the output of the warpfield mapping function for the nth subject.

[0153] In some embodiments, the method then assigns a prior distribution over the brain network pattern parameters μ1:K and f1:K. In some embodiments the mean parameter is zero in all states; this assumption is written as:…Eqn (17)

[0154] An alternative implementation assigns a Gaussian distribution over this parameter, with hyperparameters m and S representing the mean and covariance respectively:…Eqn (18)

[0155] In one embodiment, in which the output function f represents a Cholesky decomposition of the data covariance such that the product ^3 is a positive semi-definite matrix, the method assumes a Wishart distribution for the prior over the function ^ ^ 3^3 , with shape and scale hyperparameters W and p as follows: ^^^^:d|^, ^) = ∏d67^ ^(^6^^6 |^, ^) …Eqn (19)where p is set equal to the number of dimensions and W equal to the identity matrix of dimension p.

[0156] Upon conclusion of the “Identify” subroutine, the Brain State Attributes 190, which comprise brain network patterns and brain network dynamics, encompass one or more Brain Network Derived Biomarkers which explain transient differences in brain state activation between therapeutic groups. Depending on the desired application, these Brain Network Derived Biomarkers can be utilised collectively in the context of the full set of Brain State Attributes (as described below in equations 20 - 21) or isolated for use as a therapeutic outcome target (as described below in equations 22 - 23).

[0157] Fig.6 is a schematic block diagram representation of one embodiment of the Therapeutic Group Prediction Module 170 of Fig.1. For some new individual patient with data ^^^, ^^^for whom the therapeutic group is unknown, the sequence of brain states AB^is computed by the Brain State Inference Module and provided as input to the Therapeutic Group Prediction Module of Fig.6.

[0158] In some embodiments, the patient’s unknown therapeutic group refers to the variable ^^,^, in which case the Therapeutic Group Prediction Module outputs the variable ^^,^for the new individual patient. In other embodiments, the patient’s unkown therapeutic group may be only one dimension of ^^,^, with other dimensions containing known information such as the data output from a computer game such as that of Figure 14, in which case the known data dimensions of ^^,^are provided as input data and the Therapeutic Group Prediction Module outputs only the unknown data dimensions of ^^,^.

[0159] In some embodiments, these brain states are accumulated into a vector of dimension[^^ × ^] which is provided as input to a Feature Extractor. The feature extractor extracts a setof features that describe the dynamics of the inferred state timecourse for the new subject, where these features may include but are not limited to the frequency of state transitions between each pair of states; the fractional occupancy per state; the mean and / or median and / or percentile values of the interval time per state; the mean and / or median and / or percentile values of the duration time per state; and the cyclical dynamic statistics, as outlined in Van Es, Higgins et al 2023 (https: / / doi.org / 10.1101 / 2023.07.25.550338 ).

[0160] In some embodiments, the features are then provided as input to a Classifier to make a therapeutic group prediction. The therapeutic group prediction is made by the classifier, which compares these statistics with a reference dataset, the reference dataset being the same set of Features from the inferred brain state dynamics for the subject data utilised in the “Identify” subroutine, and those subjects’ corresponding therapeutic group assignments.

[0161] In some embodiments in which the variable ^^,^is a binary or categorical outcome measure invariant over time, this is applied by a machine learning classifier, such as logistic regression classifier, support vector machine with Gaussian kernel radial basis function classifier, K-nearest neighbour classifier, a random forest classifier or a linear discriminant analysis classifier.

[0162] In some embodiments in which the variable ^^,^is a continuous outcome measure invariant over time, this is applied by a machine learning prediction model such as a linear regression model, a multilayer neural network model, a support vector machine, or gaussian process model.

[0163] The Therapeutic Group Prediction Module is designed to be modular in its application, potentially taking inputs from a set of parallel Clinical Brain Network Identification Devices. This modularity overcomes potential shortcomings of using a single Clinical Brain Network Identification Device, including the risk of local minima convergence leading to sub optimal brain network identification. In this embodiment, the “Predict” subroutine is run in parallel on eachClinical Brain Network Identification Device, with a single Therapeutic Group Prediction Module applying a suitable ensemble classification algorithm such as a random forest classifier, to assign the therapeutic group on the basis of the combined outputs of multiple Clinical Brain Network Identification Devices.

[0164] In some embodiments, the method infers the probability of each therapeutic group by implementing a Bayesian model inversion. Specifically, the method seeks to determine: ^

[0165] It can be shown that the probability of assigning this subject, being the new individual patient, to each of the G groups is:…Eqn (21)

[0166] The goal of the Therapeutic Group Prediction Module is to make an effective therapeutic group prediction for the new subject. Whilst the embodiments described herein achieve this through a Bayesian model inversion, any machine learning method designed for prediction can be utilised.

[0167] This concludes the “Predict” sub routine, with the output Therapeutic Group Assignment identified.

[0168] In some embodiments, EEG data is recorded from subjects whilst they complete a game on a computer apparatus, wherein said game has been specifically designed to include tasks shown to be processed differently by patients assigned to different therapeutic groups, and in which the task may be expected to elicit task-related transient differences in brain network expression.

[0169] In one application, patients assigned to a therapeutic group related to major depressive disorder have their EEG data recorded while they complete with a computer game modelled on computational psychiatry paradigm, motivated by the fact that patients with major depression make less optimistic decisions in this paradigm than those without major depression. The input variable ^^,^encodes the decisions made by a patient at each point in time, either as raw triggerinformation or following some processing step to convert the triggers into interpretable behavioural parameters such as optimism level or learning rate. An example of such a computational psychiatry paradigm is an exploit / explore paradigm, which is believed to elicit task-related changes in transient expression of brain networks.

[0170] Fig.15 is a schematic representation of a system 1500 utilising a brain network identification device 1550 of the present disclosure. A patient 1505 wears an electroencephalogram (EEG) headset 1510 that includes a set of electrodes for acquiring and recording task-related EEG data from the patient 1505.

[0171] In the system 1500, a visual stimulus is presented to the patient on a display 1525, which is positioned so that the patient can readily view the display 1525. The display 1525 may be, for example, a computer monitor, computing tablet, smartphone, laptop screen, or the like. A computer 1520 programmed to display stimuli on the display 1525 is coupled to the display 1525. It will be appreciated that the computer 1520 and display 1525 may be integral, such as in a laptop or tablet computing device, or separate components, such as a personal computer coupled to a monitor. The computer 1520 may be co-located with the display 1525 or remotely located, such as may occur if the computer 1520 is implemented as a remote server coupled to the display 1525 via a communications network, such as the Internet.

[0172] The system 1500 includes an input device 1530 for receiving input from the patient 1505. Depending on the implementation and application, the input device 1530 may include a keyboard, computer mouse, stylus, joystick, trackball, steering wheel, drawing tablet, or the like, or any combination thereof. In some embodiments, visual stimuli presented to the patient 1505 via the display 1525 are embodied in the form of a game, with software for implementing the game executing on a processor of the computer 1520. During gameplay, the patient 1505 uses the input device 1505 to make decisions.

[0173] Fig.14 is a flow chart representation of one embodiment of an exploit / explore paradigm 1400 commonly used in computational psychiatry experiments (and more fully explained in Cremer et. al.2023; Neuropsychopharmacology), in which a subject must repeatedly decide whether to ‘exploit’ from a known resource that is being depleted (in this case where the resource is represented by a tree and the decision to exploit represented by the action of harvesting), or ‘explore’ by going in search of some new resource which is unknown (in this case where the decision to explore is represented by the action of switch). The exploit / explore paradigm show in Fig.14 is an example of a stimulus / game displayed on a computer display 1525 to the patient 1505. The game 1400 includes a number of states over time, with time points displayed from left to right. The patient 1505 uses the input device 1530 to make exploit or explore decisions at various points in the game 1400. Computationalpsychiatry tasks such as this exploit / explore paradigm are selected on the basis that Brain Network Related Biomarkers may emerge transiently in response to controlled events in a task- related EEG design.

[0174] The computer 1520 transmits game data Un,t1535 to the clinical brain network identification device 1550. The game data 1535 includes game events and participant decisions made by the patient 1505 using the input device 1530 at each timepoint. Similarly, the EEG headset 1510 transmits EEG data Xn,t 1515 to the clinical brain network identification device 1550 for each point in time.

[0175] The clinical brain network identification device 1550 processes the received game data 1535 and EEG data 1515 to identify task-related transient brain network patterns for that patient 1505. The identified task-related transient brain network patterns may be utilised to predict when in time a therapeutic target is activated to guide an automated and immediate action (such as brain state modulation, or some other intervention, such as a stimulus delivered to a computer display (including during a computer game)), and / or predict to which therapeutic group that patient belongs for at least one of diagnostic, prognostic, monitoring or screening purposes.

[0176] The input variable ^^,^can also include time-varying or time invariant information relating to clinical confounds. This reflects cases in which it is desirable to to control for certain variables that may affect the state transition probability, such as if the scan is taken before or after administration of some intervention, so as to obtain a more accurate estimate of the effect that can be isolated to the therapeutic grouping being studied. 3. Overall Method Description

[0177] Fig.7 is a flow diagram illustrating a method 700 in accordance with the present disclosure. A first step 710 of the method 700 acquires and processes suitable data relating to a neurological or psychiatric condition. In the example of Fig.7, the first step 710 includes a therapeutic group selection and stratification step 712, a protocol matching 714, a data acquisition step 716, and a data storage step 718.

[0178] Control passes from step 710 to step 720, which utilises the data from step 710 as input to a Clinical Brain Network Identifier Device set to the ‘Identify’ subroutine to identify candidate Brain State Attributes 190 that differentiate therapeutic groups. In step 730, the brain State Attributes are used as input to a Clinical Brain Network Identifier Device in ‘Predict’ mode to predict which therapeutic group an individual belongs to for diagnostic, prognostic, monitoring or screening purposes. In step 740, the method uses the Brain Network Patterns output from step 720 as therapeutic outcome targets for use in an automated device such as aNeuromodulatory Outcome Optimisation Device outlined in International Patent Application No. PCT / AU2023 / 051354.

[0179] The first step 710 of method 700 entails a set of preliminary steps to implement a parallel pipeline for all therapeutic groups that is identical in all steps; that removes any systematic variation over groups and ensures that variations in the data processed are reflective of the therapeutic condition and not of differences in preliminary processing steps.

[0180] The following preliminary steps within the first step 710 are present in one or more embodiments: 712. Therapeutic group selection and stratification 714. Protocol Matching 716. Data acquisition 718. Data storage 712. Therapeutic group selection and stratification

[0181] An initial step relates to identifying patients who fit a profile associated with therapeutic targeting. In one embodiment, these patients fit into a minimum of two different groups, representing at least one clinical condition and at least one control group. In alternative embodiments, each patient has a clinical score against some clinical outcome measure, such as a depression index score, such that the patients represent a reasonably well balanced sampling from across the spectrum of anticipated clinical scores.

[0182] In the context of identifying therapeutic targets, the approach is to derive the therapeutic targets from the data such that the targets reflect transient patterns that emerge in clinical condition groups that differentiate those targets from the control group. This is intended for use in interventions that seek to ’normalise’ these target patterns in individual subjects experiencing the clinical condition.

[0183] The selection of patients for each group is crucial and may differ depending on the desired intervention. For example, an intervention targeting mild to moderate depression sufferers may stratify patients by their clinical scores on a clinically standardised depression questionnaire that assigns each subject a depression score. Alternatively, an intervention may collate a single group based on a clinical diagnosis of mild to moderate depression. In some embodiments, there is a need for a control group of subjects reflecting a healthy condition or desired end-state for the clinical group. A control group may comprise, for example: people with low clinical scores on the same depression questionnaire; people who have never been diagnosed with mild to moderate depression; or people who have recently recovered after a diagnosis of mild to moderate depression. An alternative therapeutic grouping may assignpatients longitudinally, based on how far the respective patients have progressed through treatment.

[0184] In each case, the different control group will produce slightly different therapeutic targets. Alternative implementations include stratifying groups by whether or not members of a group have responded to treatment, with the brain state attributes being used to predict a new subject’s likelihood of responding to the particular treatment (prognostic application), and the patterns being used thereafter therapeutically in those who have responded to treatment as a form of maintenance therapy (i.e., to maintain the effects of the effective intervention and prevent relapse). 714. Protocol Matching

[0185] The method requires neurophysiological data recorded from groups of patients belonging to the therapeutic groups identified above. This neurophysiological data may be acquired directly through patient recruitment and recording, or may be acquired indirectly through the use of large public datasets of suitably anonymised neurophysiological data.

[0186] Fig.8 illustrates an example of patient stratification. An initial cohort 810 of patients assigned to particular therapeutic groups includes a diverse range of patients. In the example of Fig.8, the initial cohort 810 is a random sample of the population with people belonging to either Therapeutic Group A (grey) or Therapeutic Group B (black). The initial cohort 810 is stratified by key indicators to ensure a propensity matched cohort dataset 820 that is age and gender matched. Patients from the initial cohort 810 that do not satisfy the criteria associated with the propensity matched cohort dataset 820 form an excluded cohort 830.

[0187] In order to acquire relevant data, the recording protocol must match the future intervention protocol. That is, where the transient therapeutic targets are intended for use in a clinical setting in which a patient is at rest while electroencephalograph (EEG) data is acquired, then training data should be acquired in a clinical setting where a subject is at rest while their EEG data is acquired. Similarly, if the intervention is desired to be applied whilst a patient completes a cognitive task in order to target task-related transient expression of brain networks, then the training data should be acquired while participants complete the same cognitive task. 716. Data acquisition

[0188] For the patient groups identified in the therapeutic group selection and stratification described in 3.1 above, neurophysiological data is acquired. Different methodologies may be utilised to acquire the neurophysiological data. Such methodologies include non-invasive neurophysiological means, such as EEG, magnetoencephalography (MEG), functional near- infrared (IR) spectroscopy (fNIRS), and optically-pumped magnetometer based MEG(OP-MEG). Alternative methodologies include invasive neurophysiology mechanisms, such as intracranial EEG or subcutaneous EEG. Data can either be acquired directly by administration of a suitable recording device in a suitably controlled setting, or alternatively by indirect acquisition, such as through the use of large public datasets. 718. Data storage

[0189] Data is transferred and stored in a secure environment, such as a local database or cloud-based database. It will be appreciated that the data may be acquired from one or more sources and any combination of sources is possible. In some embodiments, the data is acquired by sensing data from patients and then transferred for storage. In other embodiments, the data is retrieved from external stored datasets and then transferred for local storage. Stored data may include, for example, raw neurophysiological data, derivative data, clinical scores informative of therapeutic groups, and trained model parameters.

[0190] Returning to Fig.7, method 700 includes a second step 720 that follows step 710. This involves running the Clinical Brain Network Identification Device using the “Identify” subroutine, using the EEG data available from step 710 and the corresponding Patient Metadata, which must include therapeutic group assignment information. In some embodiments, this therapeutic group assignment information may take the form of a binary classification, such as Depression vs Healthy Controls; in another embodiment, the therapeutic group assignment information may take the form of a discrete classification, such as {healthy control, mild, moderate, severe}; in another embodiment, the therapeutic group assignment information may take the form of a clinical score or set of clinical scores on a scale, such as the scores obtained from a DASS and / or K10 questionnaire.

[0191] Method 700 includes a third step 730. Step 730 involves running the Clinical Brain Network Identification Device using the “Predict” subroutine, as described herein.

[0192] Method 700 includes a fourth step 740. Step 740 involves using the Brain Network Derived Biomarkers identified during the “Identify” subroutine as therapeutic outcome targets.

[0193] The Brain Network Derived Biomarkers are then identified from the Brain State Attributes. In some embodiments, this is done by first taking the posterior distributions over the transition probabilitydefined for each therapeutic group, and computing the difference in dynamic patterns attributable to each of the K states. Specifically, assuming group 1 represents the desired target for group g > 1, the method defines the difference matrix D as follows:…Eqn (22)

[0194] Some embodiments then define the Brain Network Derived Biomarker as the value of k that displays maximally different temporal dynamics between therapeutic groups as follows:…Eqn (23)

[0195] Other embodiments define multiple Brain Network Derived Biomarkers by applying athreshold to the measure er^^, ^) or taking the highest n values for some integer n.

[0196] The therapeutic target pattern for future use in methods that seek to elicit or suppress transient patterns of activity is given by the associated parameters of the above Brain Network Derived Biomarker, which in some embodiments is

[0197] While this brain state and its associated set of brain network patterns are identified as the therapeutic outcome target, practical implementations typically involve the utilization of the complete set of Brain State Attributes. This approach is necessary because these therapeutic outcome targets are more easily identifiable relative to the remaining identified brain network patterns. Consequently, the output of this method encompasses all Brain State Attributes, with the therapeutic targets clearly delineated to guide interventions. In some embodiments, the therapeutic outcome target is itself the Brain Network Derived Biomarker; in other embodiments, the Brain Network Derived Biomarker refers collectively to the relative levels of expression over all identified Brain State Attributes 190.

[0198] In some embodiments intended for use in realtime brain state modulation, such as described in International Patent Application No. PCT / AU2023 / 051354, the therapeutic target is then programmed into the reward function of the Reinforcement Learning module that adaptively learns the best stimulation protocol to elicit or inhibit said therapeutic target. In some embodiments this is achieved by setting the Reinforcement Learning module’s reward function to the following: ^^ = ∑^ b7^ ^b AB^^+b^ …Eqn (24)Where ^ of dimension [K x 1] is set to a one-hot vector encoding of the therapeutic target is the output of the brain state inference module at timepoint ^; ^ is a discount factor taking values 0 < ^ < 1, which in some embodiments is set to 0.98. 4. Detailed Overview of the “Identify” subroutine

[0199] The “Identify” subroutine of the present disclosure applies a customised variational inference framework to identify transient brain states that exhibit different temporal dynamics over the therapeutic groups. 4.1 Relation to Bayesian model

[0200] The architecture of the Clinical Brain Network Identification Device is designed such that different elements can be factorised and each reflect different components of a Bayesian model illustrated in Fig.13. Specifically, the identify subroutine applies variational Bayesian inference to approximate the following posterior probability equation:

[0201] This is approximated through variational Bayesian inference. We let the above equation be approximated as a product of factorised functions as follows:

[0202] In the Clinical Brain Network Identification Device, the output of the Brain State Inference Module at each timepoint is equal to 8^^^,^); the output of the Brain Pattern Generator is equal to the expected value of 8(^|^^); and the output of the Brain Dynamics Generator is equal to the expected value of log 8^Φ|U^,^^. The overall method proceeds by iterativelyoptimising each of these functions in turn in a variational manner. 4.2. Model initialisation

[0203] All model parameters are initialised to predefined values. In one embodiment of the method, these values are derived from established unsupervised methods for equivalent models. In another embodiment, the values may be initialised to equivalent values obtained forexample in published literature; in another embodiment, the values may be initialised to random values derived from a random number generator in a suitable configuration. 4.3. Training the brain state inference module

[0204] In some embodiments, the brain state inference module 150 is trained by minimisation of the variational free energy. This training employs the variational Bayesian forward-backward algorithm, applying the Baum-Welch algorithm, using as inputs the data chain ^^:^,^:^O, and alsothe outputs of the brain dynamics generator (which models the expected value ofand the brain network pattern generator (which models the expected value of 8(^|^^)). 4.4. Training the brain network pattern generator module

[0205] In some embodiments, the brain network pattern generator module 160 is then trained by minimisation of the variational free energy. In one or more embodiments, this is achieved through the Adam Optimisation Algorithm for stochastic objective function minimisation, given the input data ^^for each subject. 4.5. Training the brain dynamics generator module

[0206] In some embodiments, the brain dynamics generator module 140 is then trained by minimisation of the variational free energy. In one or more embodiments, this uses Markov Chain Monte Carlo sampling to estimate the expected value of the marginal posterior function. 4.6. Iterate until convergence

[0207] Steps 4.3 to 4.5 are then repeated over the factorised components of the braoder Bayesian model of Fig.13 until a stopping criteria has been reached. In some embodiments, the stopping criteria is defined as the point in time where successive iterations no longer lead to changes in the output of the brain state inference module 150. In other embodiments, steps 4.3 to 4.5 are repeated for a set number of cycles, after which a new batch of data is randomly selected by the Stochastic Data Selection Module 130. Steps 4.3 to 4.5 are then repeated for another set number of cycles, with the model parameters updated by an appropriate stochastic update rule. After this, a new batch of data is randomly selected, and this process of stochastic data selection continues until a stopping criteria has been satisfied. 4.7 Training of Spatio Spectral Transform

[0208] In some embodiments, the spatio spectral transform operator B is also optimised by a greedy optimisation algorithm. In some embodiments, this optimisation routine is applied in between each iteration of steps 4.3 and 4.4. In some embodiments, this optimisation routine is instead applied after each stochastic batch has been processed in step 4.5.5. Computing Device and infrastructure

[0209] The brain pattern identification system of the present disclosure may be practised using a computing device, such as a general purpose computer or computer server. Fig.9 is a schematic block diagram of a system 900 that includes a general purpose computer 910. The general purpose computer 910 includes a plurality of components, including: a processor 912, a memory 914, a storage medium 916, input / output (I / O) interfaces 920, and input / output (I / O) ports 922. Components of the general purpose computer 910 generally communicate using one or more buses 948.

[0210] The memory 914 may be implemented using Random Access Memory (RAM), Read Only Memory (ROM), or a combination thereof. The storage medium 916 may be implemented as one or more of a hard disk drive, a solid state “flash” drive, an optical disk drive, or other storage means. The storage medium 916 may be utilised to store one or more computer programs, including an operating system, software applications, and data. In one mode of operation, instructions from one or more computer programs stored in the storage medium 916 are loaded into the memory 914 via the bus 948. Instructions loaded into the memory 914 are then made available via the bus 948 or other means for execution by the processor 912 to implement a mode of operation in accordance with the executed instructions.

[0211] One or more peripheral devices may be coupled to the general purpose computer 910 via the I / O ports 922. In the example of Fig.9, the general purpose computer 910 is coupled to each of a speaker 924, a camera 926, a display device 930, an input device 932, a printer 934, and an external storage medium 936. The speaker 924 may be implemented using one or more speakers, such as in a stereo or surround sound system. In the example in which the general purpose computer 910 is utilised to implement a brain pattern identification system, one or more peripheral devices may relate to data acquisition devices, such as EEG or MEG or OP-MEG or fNIRS devices, or the input device 1530 of Fig.15, connected to the I / O ports 922.

[0212] The camera 926 may be a webcam, or other still or video digital camera, and may download and upload information to and from the general purpose computer 910 via the I / O ports 922, dependent upon the particular implementation. For example, images recorded by the camera 926 may be uploaded to the storage medium 916 of the general purpose computer 910. Similarly, images stored on the storage medium 916 may be downloaded to a memory or storage medium of the camera 926. The camera 926 may include a lens system, a sensor unit, and a recording medium.

[0213] The display device 930 may be a computer monitor, such as a cathode ray tube screen, plasma screen, or liquid crystal display (LCD) screen. The display 930 may receive information from the computer 910 in a conventional manner, wherein the information ispresented on the display device 930 for viewing by a user. The display device 930 may optionally be implemented using a touch screen to enable a user to provide input to the general purpose computer 910. The touch screen may be, for example, a capacitive touch screen, a resistive touchscreen, a surface acoustic wave touchscreen, or the like.

[0214] The input device 932 may be a keyboard, a mouse, a stylus, joystick, trackball, steering wheel, drawing tablet, or the like, or any combination thereof, for receiving input from a user. The external storage medium 936 may include an external hard disk drive (HDD), an optical drive, a floppy disk drive, a flash drive, solid state drive (SSD), or any combination thereof and may be implemented as a single instance or multiple instances of any one or more of those devices. For example, the external storage medium 936 may be implemented as an array of hard disk drives.

[0215] The I / O interfaces 920 facilitate the exchange of information between the general purpose computing device 910 and other computing devices. The I / O interfaces may be implemented using an internal or external modem, an Ethernet connection, or the like, to enable coupling to a transmission medium. In the example of Fig.9, the I / O interfaces 922 are coupled to a communications network 938 and directly to a computing device 942. The computing device 942 is shown as a personal computer, but may be equally be practised using a smartphone, laptop, or a tablet device. Direct communication between the general purpose computer 910 and the computing device 942 may be implemented using a wireless or wired transmission link.

[0216] The communications network 938 may be implemented using one or more wired or wireless transmission links and may include, for example, a dedicated communications link, a local area network (LAN), a wide area network (WAN), the Internet, a telecommunications network, or any combination thereof. A telecommunications network may include, but is not limited to, a telephony network, such as a Public Switch Telephony Network (PSTN), a mobile telephone cellular network, a short message service (SMS) network, or any combination thereof. The general purpose computer 910 is able to communicate via the communications network 938 to other computing devices connected to the communications network 938, such as the mobile telephone handset 944, the touchscreen smartphone 946, the personal computer 940, and the computing device 942.

[0217] One or more instances of the general purpose computer 910 may be utilised, when suitably programmed so as to provide an improved computing device, to implement a server acting as a pre-processing module, a data separation module, a brain network pattern generator, a brain network dynamics generator, or the like, in order to implement a brain pattern identification system in accordance with the present disclosure. In such an embodiment, thememory 914 and storage 916 are utilised to store data relating to inference data, parameters, sensed data, patient metadata, and the like. Software for implementing the brain pattern identification system is stored in one or both of the memory 914 and storage 916 for execution on the processor 912. The software includes computer program code for implementing method steps in accordance with the method of identifying brain patterns and associated applications described herein. One or more instances of the general purpose computer 910 may be utilised, when suitably programmed so as to provide an improved computing device, to implement the stimulus computer 1520 or Clinical Brain Network Identification Device 1550 of Fig.15.

[0218] Fig.10 shows images highlighting activation of a brain network. In particular, activated portions of a brain, corresponding to a brain network, are highlighted during an activated state.

[0219] Fig.11 shows spatio-spectral brain network patterns corresponding to the activation of a brain network from Fig.10.

[0220] Fig.12 shows brain state probability over time, based on the brain network patterns of Fig.11. Industrial Applicability

[0221] The arrangements described are applicable to the computing, health, and medical industries.

[0222] Although the invention has been described with reference to specific examples, it will be appreciated by those skilled in the art that the invention may be embodied in many other forms. The foregoing describes only some embodiments of the present invention, and modifications and / or changes can be made thereto without departing from the scope and spirit of the invention, the embodiments being illustrative and not restrictive.

[0223] In the context of this specification, the word “comprising” and its associated grammatical constructions mean “including principally but not necessarily solely” or “having” or “including”, and not “consisting only of”. Variations of the word "comprising", such as “comprise” and “comprises” have correspondingly varied meanings.

[0224] As used throughout this specification, unless otherwise specified, the use of ordinal adjectives "first", "second", "third", “fourth”, etc., to describe common or related objects, indicates that reference is being made to different instances of those common or related objects, and is not intended to imply that the objects so described must be provided or positioned in a given order or sequence, either temporally, spatially, in ranking, or in any other manner.

[0225] Reference throughout this specification to “one embodiment,” “an embodiment,” “some embodiments,” or “embodiments” means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrases “in one embodiment” or “in an embodiment” in various places throughout this specification are not necessarily all referring to the same embodiment, but may. Furthermore, the particular features, structures or characteristics may be combined in any suitable manner, as would be apparent to one of ordinary skill in the art from this disclosure, in one or more embodiments.

[0226] While some embodiments described herein include some but not other features included in other embodiments, combinations of features of different embodiments are meant to be within the scope of the invention, and form different embodiments, as would be understood by those skilled in the art. For example, in the following claims, any of the claimed embodiments can be used in any combination.

[0227] Furthermore, some of the embodiments are described herein as a method or combination of elements of a method that can be implemented by a processor of a computer system or by other means of carrying out the function. Thus, a processor with the necessary instructions for carrying out such a method or element of a method forms a means for carrying out the method or element of a method. Furthermore, an element described herein of an apparatus embodiment is an example of a means for carrying out the function performed by the element for the purpose of carrying out the invention.

[0228] In the description provided herein, numerous specific details are set forth. However, it is understood that embodiments of the invention may be practised without these specific details. In other instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

[0229] Note that when a method is described that includes several elements, e.g., several steps, no ordering of such elements, e.g., of such steps is implied, unless specifically stated.

[0230] The term “coupled” should not be interpreted as being limitative to direct connections only. The terms “coupled” and “connected,” along with their derivatives, may be used. It should be understood that these terms are not intended as synonyms for each other, but may be. Thus, the scope of the expression “a device A coupled to a device B” should not be limited to devices or systems wherein an input or output of device A is directly connected to an output or input of device B. It means that there exists a path between device A and device B which may be a path including other devices or means in between. Furthermore, “coupled to” does not imply direction. Hence, the expression “a device A is coupled to a device B” may besynonymous with the expression “a device B is coupled to a device A”. “Coupled” may mean that two or more elements are either in direct physical or electrical contact, or that two or more elements are not in direct contact with each other but yet still co-operate or interact with each other.

Claims

We claim:

1. A clinical brain network identification device comprising: a pre-processing module for receiving sensed data from at least one patient and conditioning said sensed data; a first storage medium for storing said conditioned sensed data and corresponding patient metadata; a stochastic data selection module for selecting data from said first storage medium, said selected data being conditioned sensed data and patient metadata relating to a set of one or more patients; a brain network dynamics generator configured to receive clinical scores and stimulus data from said selected data and generate brain network dynamics being probabilities of a brain state being active at a point in time given the brain state activations preceding that point in time; a brain network pattern generator configured to receive said selected data and generate brain network patterns, based on said selected data; a brain state inference module configured to determine inferred brain states based on sensed data received from said stochastic data selection module and brain network patterns received from said brain network pattern generator and said brain network dynamics received from said brain network dynamics generator; and a therapeutic group prediction module for receiving inferred brain states from said brain state inference module and producing a set of therapeutic group assignments.

2. The device according to claim 1, wherein the device applies a model that has been factorised into constituent components, wherein said factorisation entails: the brain network dynamics generator storing a probabilistic specification of brain network dynamics defining the probability of a brain network being active at a point in time given the brain state activations preceding that point in time; the brain network pattern generator storing a probabilistic specification of brain network patterns defining the spatio-spectral pattern of activity associated with recorded brain data from a given subject being attributed to a particular brain network; and the brain state inference module applying the brain network patterns and brain network dynamics to identify and store the brain network most likely to be active at any point in time.

3. The device according to either one of claim 1 or claim 2, wherein said sensed data includes neurophysiological data.

4. The device according to claim 3, wherein said neurophysiological data is selected from the group consisting of: electroencephalogram (EEG) data), MEG data, OP-MEG data, and fNIRS data.

5. The device according to any one of claims 1 to 4, wherein said conditioning includes at least one of: buffering said sensed data, removing artefacts from said sensed data, and formatting said sensed data into a predefined format suitable for processing.

6. The device according to any one of claims 1 to 5, wherein said brain network dynamics generator and corresponding elements of the brain state inference module generate said brain network dynamics with a model selected from the group consisting of: a Time-Delay Embedded Hidden Markov Model (HMM); recurrent neural networks; temporal convolutional neural networks; and transformer architectures.

7. The device according to any one of claims 1 to 6, wherein said brain network pattern generator module and corresponding elements of the brain state inference module generate brain network patterns selected from the group consisting of: a spatio-spectral transform with multivariate gaussian observation model; a transformer architecture; a recurrent neural network architecture; variational autoencoder architecture; and a multilayer neural network architecture.

8. A system comprising: a clinical brain network device according to any one of claims 1 to 7, wherein said sensed data includes magnetoencephalography pre-existing sensed data relating to patient neurophysiological data derived from a patient population assigned to a set of therapeutic groups; and sensor equipment for acquiring newly sensed data from a new patient whose therapeutic group is unknown to present to said clinical brain network device; wherein said sensor equipment is selected from the group consisting of: an EEG headset, an MEG scanner, an OPM-MEG headset, and a fNIRS headset.

9. The system according to claim 8, further comprising: an input device configured to receive input from the new patient; a display device; anda computer coupled to the input device and the display device, the computer including: a processor; and a memory for storing computer code instructions that when executed on the processor of the computer: display a game on said display device; record decisions made by the new patient at each point in time in response to said game; and transmit game data to said clinical brain network device, wherein said game data include game events and corresponding patient data for each timepoint; wherein said clinical brain network device: processes newly sensed data sensed from said new patient by said sensor equipment and said game data to identify task-related transient expression of brain networks for that new patient.

10. The system according to claim 9, wherein said clinical brain network device: determines, based on said identified task-related transient expression of brain networks for that new patient, the therapeutic group to which the new patient belongs for at least one of diagnostic, prognostic, monitoring or screening purposes.

11. The system according to either one of claims 9 or 10, wherein said game is any form of visual stimuli presented over time.

12. The system according to any one of claims 9 to 11, wherein said game is a game modelled on a computational psychiatry paradigm, in which a decision is required at different points in time, wherein the input device is configured to receive said decision from the patient.

13. The system according to any one of claims 9 to 12, wherein the input device is at least one of: a keyboard, a mouse, a stylus, a trackball, a steering wheel, or a drawing tablet.

14. A method for identifying brain network derived biomarkers, comprising the steps of: acquiring initial data relating to at least one of a neurological or psychiatric condition; training a model, using said initial data, to identify differences between a plurality of therapeutic groups, wherein said trained model generates trained brain state attributes;using the trained model brain state attributes to make a new prediction, wherein said new prediction is one of: (iii) an application in which the new prediction relates to when in time a brain network derived biomarker is activated; and (iv) a prediction as to which of said therapeutic groups an individual belongs for at least one of diagnostic, prognostic, monitoring or screening purposes.

15. The method according to claim 14, wherein said initial data includes neurophysiological data.

16. The method according to claim 15, wherein said neurophysiological data is selected from the group consisting of: electroencephalogram (EEG) data), MEG data, OP-MEG data, and fNIRS data.

17. The method according to any one of claims 14 to 16, wherein said initial data includes patient metadata.

18. The method according to any one of claims 15 to 17, wherein said model is based on a Bayesian model that can be factorised into constituent components of the device of any one of claims 1 to 7.

Images