Method of obtaining personalized parameters for transcranial stimulation, transcranial stimulation system, method of applying transcranial stimulation

Abstract

The present disclosure relates to transcranial stimulation. In one arrangement, a method for obtaining personalized parameters for transcranial stimulation is provided. Baseline data is received regarding a test subject, the baseline data comprising information about the test subject taken prior to applying transcranial stimulation to the test subject. A Gaussian process model of performance of one or more training subjects is used to obtain personalized parameters for transcranial stimulation of the test subject based on the received baseline data. The Gaussian process jointly models the performance of the subject during and / or after transcranial stimulation as a function of i) parameters defining the transcranial stimulation; and ii) baseline data of the one or more training subjects.

Description

[0001] This invention relates to transcranial stimulation.

[0002] Transcranial stimulation is known in the art to be used to treat patients or to enhance cognition in healthy individuals. The optimal parameters for stimulation may vary between individuals, but a wide range of possible combinations needs to be selected. Therefore, a practical approach has been to use a generalized method in which the same set of parameters is used for multiple different subjects.

[0003] One objective of this invention is to improve transcranial stimulation.

[0004] According to one aspect of the invention, a computer-implemented method for obtaining personalized parameters of transcranial stimulation is provided, comprising: receiving baseline data about a test subject, the baseline data including information about the test subject acquired prior to the application of any transcranial stimulation to the test subject; and using a Gaussian process model of the performance of one or more training subjects to obtain personalized parameters of the transcranial stimulation of the test subject based on the received baseline data, wherein: the Gaussian process jointly models the subject performance during and / or after transcranial stimulation as a function of both: i) parameters defining the transcranial stimulation; and ii) the baseline data of the one or more training subjects.

[0005] Therefore, a probabilistic model is provided that inherently considers variations among individuals using baseline data. By considering baseline data in this way, the inventors have found that the model is better able to predict optimal parameters for transcranial stimulation, even when past data with identical or similar baseline data for no or very few subjects is available. The method is able to provide personalized parameters that vary among different subjects and, on average, outperforms generalized methods that ignore the fact that people are different from one another and that the individual brain is plastic and changes over time. This method allows for the effective customization of stimulation protocols to provide optimal results. The stimulation can be used for a variety of clinical and non-clinical purposes, including, for example, improving tremor in Parkinson's disease, improving sustained attention / focus, memory, mathematical performance, mood, training ability, and learning (e.g., language, mathematics, IT, factual information).

[0006] In one embodiment, baseline data represents performance unaffected by transcranial stimulation (e.g., cognitive and / or motor performance, subjective reports). The inventors have found that using this type of baseline data makes it particularly effective to obtain personalized parameters.

[0007] In one embodiment, obtaining personalized parameters further includes: obtaining test data representing the performance of the test subject (e.g., cognitive and / or motor performance, subjective reports) during and / or after the application of transcranial stimulation with multiple corresponding parameter combinations; refining the Gaussian process model using the obtained test data; and obtaining personalized parameters using the refined Gaussian process model. In one embodiment, the Gaussian process model is refined during an iterative Bayesian optimization process, which selects to sample the next step at each step, wherein the parameter combinations used for stimulation optimize the acquisition function. Using the acquisition function in this way allows the Gaussian process model to explore the most relevant parts of the available parameter space in an efficient manner, thereby facilitating rapid and reliable convergence to the optimal personalized parameters for the test subject.

[0008] Embodiments of the invention will now be described by way of example only with reference to the accompanying drawings, in which corresponding element symbols denote corresponding parts, and wherein:

[0009] Figure 1 This is a flowchart schematically depicting a method for obtaining personalized parameters for transcranial stimulation;

[0010] Figures 2 to 5 The Bayesian optimization process is illustrated schematically.

[0011] Figures 6 to 8 The predicted cognitive performance was plotted as a function of the frequency and current of the applied stimulus during stimulation of objects with three different baseline cognitive performances.

[0012] Figure 9 This is a graph showing the changes in optimized object performance as a function of baseline cognitive performance;

[0013] Figure 10 This is a graph showing the changes in personalized parameters (frequency (dashed line) and current (solid line)) as a function of baseline cognitive performance;

[0014] Figure 11 The optimization of the black-box function f(x) as a function of the number of iterations in the adjustment of the Gaussian process model is described;

[0015] Figure 12 This is a graph showing the results of a simulation analysis comparing the performance (upper curve) of an embodiment of the present disclosure with the performance based on random sampling of the parameter space;

[0016] Figure 13 This is a graph comparing the use of Bayesian optimization (lower curve) that models only the parameters used for the stimulus, with personalized Bayesian optimization (upper curve) that models both the parameters used for the stimulus and the baseline data of the object.

[0017] Figure 14 The transcranial stimulation system is schematically depicted;

[0018] Figure 15 It is a 3D visualization of a simulation function used to demonstrate the use of a method for obtaining personalized parameters for transcranial electrical stimulation used in the treatment of Attention Deficit Hyperactivity Disorder (ADHD).

[0019] Figure 16 It is a description Figure 15 A 2D visualization of the simulated function shown;

[0020] Figure 17 It is a description Figure 16 A 2D visualization of the simulation function, where the true optimal parameter for each neurophysiological value is plotted as a black dot (the true optimal parameter is unknown to scientists / clinicians).

[0021] Figure 18 yes Figure 17 The graph shows the suggested parameters from different algorithms superimposed (parameters suggested by personalized Bayesian optimization are shown by white squares, and parameters suggested by Bayesian optimization are shown by white dashed lines);

[0022] Figure 19 This is a 2D visualization depicting a simulation function with selected parameters plotted during iterations from personalized Bayesian optimization. Points randomly selected from the initial stage are shown in black, while points selected through personalized Bayesian optimization are shown in white; and

[0023] Figure 20 It is a graph comparing the performance of different methods.

[0024] Transcranial alternating current stimulation (tACS) and similar techniques are thought to promote oscillatory activity through transcranial stimulation. Such techniques may modulate brain oscillations that contribute to cognitive processes. The effects of stimulation can be enhanced or disrupted by altering the parameters of the stimulation (e.g., current and frequency), for example, to target specific neural populations. However, it is impractical to explore the effects of each parameter combination on test subjects individually (e.g., on cognitive performance). Therefore, it has been difficult to optimally utilize transcranial stimulation techniques such as tACS. The embodiments of this disclosure address this problem.

[0025] Figure 1A framework for a method to obtain personalized parameters of transcranial stimulation is schematically depicted. The method can be implemented by a computer. Therefore, each step of one or more steps in each method can be performed by a computer. The computer can include various combinations of computer hardware, including, for example, a CPU, RAM, SSD, motherboard, network connection, firmware, software, and / or other elements known in the art that allow the computer hardware to perform desired computational operations. The desired computational operations can be defined by one or more computer programs. The one or more computer programs can be provided in the form of a medium or data carrier storing computer-readable instructions, optionally a non-transitory medium. When the computer reads the computer-readable instructions, the computer performs the desired method steps. The computer can consist of independent units, such as general-purpose desktop computers, laptops, tablets, mobile phones, or other smart devices. Alternatively, the computer can consist of a distributed computing system having multiple different computers connected to each other via a network, such as the Internet or an intranet. Therefore, the method can be performed partially or entirely by cloud computing facilities.

[0026] exist Figure 1 In step S1, the method includes receiving baseline data about the test subject. The baseline data may be received from a data storage device, directly from a device that measures baseline data, or based on subjective reports. Subject reports may be made by the subject (e.g., reporting how tired he / she feels) or by a third party (e.g., reporting how distracted the subject is). In one embodiment, the baseline data includes information about the test subject that was acquired before (and / or separately from and / or independently of) the application of transcranial stimulation to the test subject. In one embodiment, the baseline data represents the test subject's performance (e.g., cognitive and / or motor performance, subjective reports) unaffected by transcranial stimulation. Therefore, the baseline data may be obtained during a period when no transcranial stimulation was applied to the subject and / or when no transcranial stimulation was recently applied, such that the subject's performance was significantly affected by the stimulation. In some embodiments, the baseline data alternatively or additionally includes other personal information about the subject, such as neural data and / or demographic data (e.g., information about sleep patterns, lifestyle habits (e.g., smoking, drinking), age, gender, etc.).

[0027] Performance can be measured in various ways, thus providing a variety of possibilities for the format of baseline data (in embodiments where the baseline data includes information about performance). For example, accuracy and / or reaction time can be measured during testing. In some embodiments, the baseline data includes scalar values ​​that may optionally vary continuously as a function of cognitive performance (e.g., test score or IQ score). In some embodiments, the baseline data includes drift rate values ​​obtained by applying a diffusion decision model to the results of a test performed by the subject. The test may include any suitable activity that provides information about performance. In some embodiments, as illustrated below, the test includes mathematical tests, such as arithmetic tests. In some embodiments, arithmetic tests involving multiplication of a single-digit number by a multi-digit number are used. Modeling human behavior using a diffusion decision model helps to pinpoint the cognitive processes of interest rather than more peripheral processes, such as stimulus encoding or motor-related activities. This approach allows for the dissection of different components in the information processing chain by modeling the decision-making process and targeting components that reflect ability / task difficulty (e.g., drift rate, which combines response time and accuracy), rather than auxiliary components. Furthermore, the method allows for the simultaneous characterization of changes in accuracy and reaction time as a single metric.

[0028] The skills required to solve arithmetic problems vary considerably not only among individuals with learning disabilities but also in the general population. Similarly, a recent study highlighted the importance of individual differences in neural and behavioral aspects between individuals with high and low arithmetic skills. The left frontoparietal network has been suggested to play a significant role in arithmetic processing and can be readily targeted using transcranial stimulation techniques such as tACS. A recent tACS study demonstrated the importance of this network in the left hemisphere within the memory domain, which is also considered fundamental to the stepwise calculation of arithmetic problems. This approach can enhance the targeting of this network by providing personalized parameters for stimulation.

[0029] exist Figure 1 In step S2, the method includes using a Gaussian process model of the performance (e.g., cognitive and / or motor performance, subjective reports) of one or more training subjects to obtain personalized parameters for transcranial stimulation of the test subjects based on the baseline data received in step S1. The Gaussian process model will jointly model the performance (cognitive and / or motor performance, subjective reports) during and / or after transcranial stimulation as a function of both: 1) parameters defining the transcranial stimulation; and 2) baseline data of one or more training subjects.

[0030] In one embodiment, obtaining personalized parameters includes iteratively refining a Gaussian process model using measurements of the test subject's performance (e.g., cognitive and / or motor performance, subjective reports) during transcranial stimulation with multiple parameter combinations. In some embodiments, the parameters include one or more of the following: frequency, current, phase, duration, dose, and brain region. For example, different combinations of current in the 0.4-2 mA range and frequency in the 0.1-100 Hz range may be considered. In cases where the stimulation comprises a waveform more complex than a pure sine wave, additional parameters may be included to define relevant additional characteristics of the waveform (e.g., to define spectral aspects of the waveform, such as center frequency and bandwidth, band shape, pulse length, etc.). Figure 1 In the example program shown, the iterative refinement of the model includes the iterative execution of sub-steps S21, S22, and S23. In some embodiments, obtaining personalized parameters may include selection during individual treatment or between different individuals, between different overall stimulation modalities (e.g., between tACS, transcranial direct current stimulation (tDCS), transcranial random noise stimulation (tRNS), and / or other stimulation modalities).

[0031] In the example shown, in the initial instance of substep S21, initial parameters for stimulating the test subject are obtained using only a pre-trained Gaussian process model and baseline data. In this embodiment, therefore, initial operating parameters are obtained during and / or after the application of transcranial stimulation to the test subject, before any measurement of performance (e.g., cognitive and / or motor performance, subjective reports) has been performed. The pre-trained Gaussian process model can be trained on a range of previously collected data. This data can come from random assignments of stimulation parameters. Technically, the Gaussian process (GP) does not explore different combinations; it simply models existing data and provides estimates of the responses to various stimulation parameters, as well as estimates of uncertainty. Exploration is accomplished by an acquisition function, which is the next stage of the Bayesian optimization process.

[0032] In the subsequent sub-step S22, the performance of the test subject during and / or after transcranial stimulation (e.g., cognitive and / or motor performance, subjective reports) is measured. Transcranial stimulation is applied using the parameters provided in sub-step S21. The measured performance forms test data. During iterations, sub-step S22 is repeated to obtain test data representing measurements of performance during and / or after transcranial stimulation with multiple corresponding combinations of operating parameters (e.g., operating parameters having different combinations of values ​​during each iteration).

[0033] In the subsequent sub-step S23, the test data obtained in sub-step S22 is used to adjust the Gaussian process model used in the previous instance of sub-step S21. Sub-steps S21, S22, and S23 are then repeated iteratively until the measured performance (e.g., cognitive performance and / or motor performance, subjective reports) obtained in sub-step S22 converges to an optimal value (e.g., when the improvement rate falls below a predetermined threshold) and / or a predetermined number of iterations have been performed. Therefore, the Gaussian process model is refined using the obtained test data (in the illustrated example, through multiple iterations of sub-steps S21, S22, and S23). This optimization process can be performed across multiple different users (each user using the process at least once) and / or by each of one or more individual users multiple times.

[0034] The refined Gaussian process model at convergence provides a prediction of how expected subject performance (e.g., cognitive and / or motor performance, subjective reportage) varies as a function of the parameters used to stimulate the test subject. Therefore, the parameter combination that yields the maximum performance predicted by the refined Gaussian process model can serve as the personalized parameter output for the test subject. Thus, the refined Gaussian process model is used to obtain personalized parameters.

[0035] Gaussian process models can be refined during iterative Bayesian optimization (BO), such as... Figures 2 to 5 The diagram illustrates this. Boolean optimization (BO) is an active machine learning technique that aims to find the global optimum of a black-box function f(x) through a series of evaluations. Figures 2 to 5 In the diagram, the vertical axis represents f(x) and the horizontal axis represents x (for ease of illustration, it is shown in a schematic one-dimensional form, but it should be understood that in embodiments of this disclosure, x will be multidimensional, including, for example, one dimension for each type of parameter used in the simulation (e.g., current and frequency), and one dimension for the baseline data).

[0036] like Figure 2 The described BO process may initially include fitting a Gaussian process model to the available data points (in... Figures 2 to 5 Representative points are marked as "2"). The fitted Gaussian process model has associated estimated mean (marked as "4") and uncertainty (marked as "6"). The example objective function is marked as "8".

[0037] The Gaussian process model can be refined by providing additional data points. In the methods of this disclosure, additional data points may include measurements of the test subject's performance (e.g., cognitive and / or motor performance, subjective reports) during and / or after the application of transcranial stimulation with different parameter combinations, for example, by... Figure 1The test data generated during each iteration of sub-step S22, or provided from test data provided by other test objects.

[0038] In each step, the iterative BO process can be configured to select the next sample, where x is optimized (e.g., maximized) to obtain function 10, such as... Figure 3 As depicted. In embodiments of this disclosure, the BO process can therefore select to sample the next stimulus, wherein the combination of parameters used for the stimulus is optimized (e.g., maximized) for the acquisition function 10. The acquisition function 10 is derived based on a Gaussian process model but is configured to be computationally cheaper to optimize than the objective function of the Gaussian process model. The acquisition function can be selected to facilitate guiding sampling in the direction most likely to improve the current best assessment (e.g., highest cognitive performance). There are no particular limitations on the nature of the acquisition function. The acquisition function can, for example, include known acquisition functions such as the Gaussian process upper confidence bound (GP-UCB) and / or expected improvement (EI). More illustrative details regarding acquisition functions are given below.

[0039] like Figure 4 What is described, in Figure 3 The value of x at the optimal point (the maximum value in the example shown) of the acquisition function 10 (e.g., corresponding to a specific combination of parameters used for the stimulus) is selected as the next data point to be sampled in the BO process (labeled "12"). The Gaussian process model is adjusted (e.g., by adjusting the hyperparameters) to account for the new data point 12, resulting in a new estimated mean 4 and uncertainty 6. The BO process continues by gradually adding additional data points until the potential improvement is considered negligible (e.g., convergence). Figure 5 The convergence state is schematically depicted, in which the Gaussian process models the objective function 4 more accurately with a relatively narrow uncertainty band. In embodiments of this disclosure, parameters for stimulation corresponding to the maximum value of the estimated average performance in the converged Gaussian process model can be output as personalized parameters for the test subject.

[0040] The process described above for iteratively refining the Gaussian process model can also be used to provide a Gaussian process model that is initially input to... Figure 1Step S21. Training the Gaussian process model can, for example, use training data representing the performance (e.g., cognitive and / or motor performance, subjective reports) of each of a plurality of training subjects during and / or after the application of transcranial stimulation having multiple corresponding parameter combinations for each training subject. As described above, the Gaussian process model can be trained in an iterative Bayesian optimization process, which selects to sample the next step at each step, wherein the parameter combination for stimulation is optimized (e.g., maximized) to obtain the function.

[0041] like Figure 14 As schematically depicted, a data processing device 22 may be provided for performing any of the computer-implemented methods described above. Figure 14 The data processing device 22, as depicted, can be provided as a standalone device or integrated into the transcranial stimulation system 20. The transcranial stimulation system 20 may include a stimulation unit 24 for providing stimulation to a subject. Transcranial stimulation may include any one or more of the following: transcranial alternating current stimulation; transcranial random noise stimulation; transcranial direct current stimulation; transcranial magnetic stimulation; transcranial focused ultrasound; and transcranial optical stimulation. The stimulation unit 24 includes suitable hardware for applying the relevant stimulation (e.g., one or more electrodes 26 for transcranial electrical stimulation in the illustrated example). The data processing device 20 controls the operating parameters of the stimulation unit 24 to define parameters for stimulation. Parameters for stimulation may include one or more of the following: frequency, current, phase, duration, dose, and brain region. The data processing device 20 may perform any of the computer-implemented methods described above, which provide personalized parameters for the subject and control the operating parameters of the stimulation unit 24 based on the obtained personalized parameters of the subject.

[0042] Detailed example

[0043] More details on Bayesian optimization

[0044] The function f can be used to represent an unknown black-box function. We do not have a closed-form expression for this black-box function, but we can have an infinite number of queries. Furthermore, evaluating such a black-box function is very expensive and time-consuming. Formally, let... (representing a set of real numbers from -∞ to +∞) is defined in the subset A well-performing function on , where d is the dimension. In the case of f representing cognitive performance, solving the following global optimization problem can provide improved optimization parameters x for applying transcranial stimulation. *

[0045]

[0046] By conducting a series of evaluations x1, x2, ..., x T The BO algorithm can be used, for example, to find the global optimum of the arithmetic performance of a black-box function f(x), as indicated by the drift rate. However, this method does not optimally account for variations in baseline capabilities (e.g., baseline cognitive performance) between different objects.

[0047] Consider different baseline data (e.g., cognitive performance).

[0048] Each subject has individual baseline data, representing, for example, baseline cognitive performance (e.g., arithmetic ability), which can be represented by a baseline data value c. The baseline data value c is provided individually for each subject. The inventors have discovered that the optimal parameters for transcranial stimulation differ significantly between subjects with different baselines. That is, the optimal parameter x... * The value of c is not unique for different baseline data. The inventors have recognized that the method for obtaining the parameters used for stimulation can be improved by adapting the above-described global optimization problem to inherently consider the baseline data.

[0049] The consideration of baseline data can be accomplished by fundamentally altering the global optimization problem, so that it can be formally defined as:

[0050]

[0051] Where c is the baseline data value. Optimal parameter x * It's not defined globally, but specifically for the variable c.

[0052] Gaussian process for jointly modeling parameters used for stimulation and baseline data.

[0053] Traditional Boolean operation (BO, disregarding baseline data value c) infers f by constructing a Gaussian process (GP) via evaluating f ~ GP(m,k), where m is the prior mean function and k is the covariance function. This flexible distribution allows us to correlate normally distributed random variables at every point in the continuous input space. Therefore, we obtain a predicted distribution of the new observations, which also follows a Gaussian distribution, and whose mean (μ) and variance (σ) are... 2 The following formula is given:

[0054] μ(x′)=k(x′;X)K(X;X) -1 y

[0055] σ 2 (x′)=k(x′;x′)-k(x′;X)K(X;X) -1 k(x′;X) T

[0056] Where K(U;V) is the covariance matrix, and its elements (i;j) are calculated as k i,j =k(x i ;x j ), where x i ∈U and x j ∈V. Behavioral observations are often associated with noise that can be accommodated in a Gaussian process model. That is, each f(x) has its own additional covariance because the noise is assumed to be independent, including an amplitude equal to the noise variance:

[0057] y i =f(x) i )+∈ i in in It is the noise variance.

[0058] When noise is taken into account, the output follows a Gaussian process, which is... Where i = j is Kronecker's Δ, then δ i,j =1. The covariance function of a noisy process becomes the sum of the signal covariance and the noise covariance. Specifically, the covariance function between two observations can be calculated as:

[0059]

[0060] To account for the baseline data value c, one possible approach would be to establish a Gaussian process and optimize for each baseline data value c separately. However, this approach suffers from a critical data efficiency problem. Specifically, the number of data samples is insufficient to estimate each baseline data value c individually. In embodiments of this disclosure, this deficiency is overcome by extending the Gaussian process alternative to jointly model our objective function f and the additional baseline data value dimension c. In this case, the GP covariance becomes;

[0061] k({x i ,c i};{x j ,c j})=k(x i ,x j )×k(c i ,c j )

[0062] Where k(x) i ,x j As defined above, and These covariance functions correspond to the parameters used for the stimulus and baseline data values, respectively. In k(c i ,c j The length scale parameter used in ) Unlike in k(x)i ,x j Used in ) For example, if the baseline data length scale If it is very large, it means that the performance function has not changed relative to the baseline data (e.g., baseline performance). On the other hand, if If the value is very small, it means the performance function is changing rapidly with the baseline data. We then use maximum marginal likelihood to estimate these length-scale parameters directly from the data. Based on our modification of GP to consider the baseline data value c, we can estimate the predicted mean and predicted variance as follows:

[0063] μ(x,c)=k({x,c};Z)K(Z;Z) -1 y

[0064] σ 2 (x,c)=k({x,c};{x,c})-k({x,c};Z)K(Z;Z) -1 k({x,c};Z) T

[0065] Here we denote Z = [X, C], and the context covariance matrix k is defined above.

[0066] Get function

[0067] To select the next point to evaluate, the acquisition function α(x) will be chosen to construct a utility function based on the Gaussian process model described above. Instead of maximizing the expensive original function f, we maximize the cheaper acquisition function to select the next optimal point:

[0068]

[0069] In this auxiliary maximization problem, the form of the obtained function is known and can be easily optimized using standard numerical techniques. One option for the obtained function is the Gaussian process confidence upper bound (GP-UCB):

[0070] α(x,c)=μ(x,c)+κ×σ(x,c)

[0071] Where μ(x,c) and σ(x,c) are the mean and variance of the GP predictions as defined above. In the above equation, κ is a parameter controlling the trade-off between exploration and exploitation. The value of κ can be theoretically chosen according to the method proposed by Srinivas, Niranjan, Andreas Krause, Sham Kakade, and Matthias Seeger in the Proceedings of the 27th International Conference on Machine Learning, pp. 1015-1022.

[0072] In the examples executed in this study, the following acquisition function α(x) was used:

[0073]

[0074] Here, Vstim is the drift rate over 50 trials of the arithmetic multiplication block, normalized by Vbase, which is the drift rate calculated over the first 25 trials starting from the baseline task. To determine the improvement between the baseline and the stimulus block, a second Vbase is calculated for the remaining 25 trials starting from the baseline task. Pearson correlation is calculated to determine whether there is no difference between the two drift rates used as an index of improvement (r = .57, p < .001). The acquisition function is designed to balance the exploration of the search space with the development of currently promising regions. A burn-in phase using 60 random tACS frequency-current combinations, assigned to the first 20 participants in the BO design, is used to determine the amount of change induced by the stimulus. We decided to use a large burn-in in our paradigm to reliably design a Bayesian optimization algorithm based on large amounts of data.

[0075] To select the hyperparameters (estimated parameters without using the observed data) for the next stimulus block, the expected improvement (EI) to the current best result or GP-UCB can be optimized. In short, UCB is more likely to select an evaluation with both high mean and high variance. Both EI and UCB have been shown to be effective in determining the number of function evaluations required to find the global optimum of many multimodal black-box functions. During this study, EI was applied to find the optimal value for arithmetic performance. Finally, we decided to remove a drift rate value of 3.6 during the experimental procedure because it might have an upper bound effect on BO. Furthermore, one data point was not included in the BO procedure due to technical issues. In total, we obtained 148 iterations.

[0076] result

[0077] Group-level baseline personalized Bayesian optimization

[0078] Figures 6 to 8 The graph depicts predicted cognitive performance as a function of the frequency (vertical axis) and current (horizontal axis) of the applied stimulus during three groups of corresponding subjects with different baseline cognitive performances. The predicted cognitive performance can be obtained using a Gaussian process model that jointly models performance as a function of the parameters used for the stimulus and the baseline data value c, as referenced above. Figures 1 to 5 As stated above. Figure 6 A model was depicted for relatively poor baseline performance (1 SD below the average drift rate). Figure 7 A model was drawn for the average baseline performance (mean = 0.055). Figure 8 Models for relatively high baseline performance (1 SD above the average drift rate) are depicted. The vertical axis represents the frequency range of the applied stimulus (0–50 Hz), and the horizontal axis represents the stimulus current (0–1.6 mA). Arithmetic performance is indicated by shading based on the calculated drift rate (tACS block / baseline block). Low drift rates are shown in darker shading, and high drift rates in brighter shading. The optimal inference point for arithmetic performance, which can be used as a basis for providing the optimal personalized parameters for the stimulus, is indicated by white squares in each figure.

[0079] Figures 6 to 8 This shows a significant shift in baseline performance from higher to lower frequencies and from lower (poor) to higher (strong) currents: for Figure 6 For the average value of -1 SD example, the optimal operating parameters are 38.67 Hz and 0.97 mA peak-to-peak; for Figure 7For the average example, the optimal operating parameters are peak-to-peak at 16.67 Hz and 0.88 mA, and for the average +1 SD example, the optimal operating parameters are peak-to-peak at 18 Hz and 0.60 mA. In this example, the baseline data values ​​are continuous parameters, so the optimal inferred parameter combination is expected to vary along this continuum. It should also be noted that... Figures 6 to 8 The results were derived from tests that included a larger number of subjects with below-average baseline data values ​​(indicating below-average cognitive performance) rather than those with above-average baseline data values. Therefore, the estimation of the optimal frequency-current parameter combination... Figure 6 The result is better than Figure 8 The results have a higher confidence level.

[0080] Verification of arithmetic performance through Bayesian optimization

[0081] Figure 9 This is a graph showing the change in optimized object performance (intermediate curve) as a function of baseline cognitive performance. The 95% confidence intervals are indicated by shading. Figure 9 The data shows fluctuations between objects with low and high baseline data values, but all objects showed significant improvement, with the stimulus not being significantly more effective for any particular group of objects (range of baseline data values).

[0082] Figure 10 The variation of the personalized tACS parameter (vertical axis) as a function of baseline cognitive performance, as proposed by the BO algorithm, is plotted. The solid line represents the variation of stimulus current with baseline cognitive performance, and the dashed line represents the variation of stimulus frequency with baseline cognitive performance. This curve confirms that as baseline ability increases, objects with low baseline ability shift from higher frequencies and currents to lower frequencies and currents, in line with... Figures 6 to 8 The group-level BO model shown is consistent.

[0083] Figure 11 The optimization of the black-box function f(x) using EI as the acquisition function is depicted as the number of iterations increases. The stable platform shows that the black-box function f(x) is reliably optimized after each iteration.

[0084] Personalized Bayesian optimization

[0085] To the inventors' knowledge, this research is the first attempt to optimize complex human behavior for transcranial stimulation parameters using a personalized BO approach. The BO is configured to jointly model the parameters used for stimulation and baseline data representing personal information about the subject, such as baseline cognitive performance, independent of any transcranial stimulation. In a specific example application studied, arithmetic performance was improved by applying different frequency-current tACS combinations to the left frontoparietal network.

[0086] Figure 12 The results of the simulation analysis (Hartmann 3D simulation—further details provided below) are shown to compare the performance of our method for finding the optimal parameters for transcranial stimulation (upper curve, based on the EI acquisition function) with the performance obtained by random sampling of the parameter space (lower curve). Within the depicted 60-iteration window, it can be seen that our method improves the optimal discovery value faster than random sampling, and reaches a significantly higher level earlier than the random sampling method within the iteration window.

[0087] Figure 13 The use of boolean modeling (BO) that models only the parameters used for the stimulus (lower curve) and personalized BO that models both the parameters used for the stimulus and baseline cognitive performance (upper curve) were compared. It can be seen that personalized BO (upper curve) can efficiently find higher discovery values ​​along 100 iterations, while conventional BO (lower curve) that models only the parameters used for the stimulus cannot. In other words, a higher level of cognitive performance (e.g., drift rate value) is obtained when using the method of this disclosure, which is based on jointly modeling both the parameters used for the stimulus and baseline data, compared to random sampling of parameters. This finding applies regardless of whether the EI or GP-UCB acquisition function is used.

[0088] EEG and baseline capabilities

[0089] Previous literature has shown a positive correlation between frontoparietal θ (4–7.5 Hz) connectivity and arithmetic baseline ability, due to the importance of this frequency band in higher cognitive processing, including arithmetic. However, no correlation was found when running regression models for connectivity in the 4–8 Hz range (all p > .3). The same applies to β (13–40 Hz) connectivity (all p > .1) and frontoparietal θ power (p = .88). However, our findings from the BO model highlight a strong performance effect in the β frequency range (13–40 Hz) for subjects with average and high baseline abilities. Therefore, we examined the relationship between baseline ability and baseline β strength in an exploratory manner. We found a positive correlation between the two (a non-parametric (spearman) correlation: r s =.29, p =.03). In other words, objects with higher arithmetic baseline ability have higher β strength compared to objects with lower arithmetic baseline skills, which confirms the above reference. Figures 6 to 8 The results of the group-level BO model are discussed.

[0090] Further details about the example method

[0091] Target and ethical license

[0092] Fifty participants provided written consent prior to the start of the study. Each individual received a financial compensation of £50. In addition to this compensation, participants had the opportunity to win an additional £50 based on their performance. Behavioral data from all 50 participants aged 18–30 years were used for Bayesian optimization (mean age = 22.52 ± standard deviation (SD = 4.09), 31 women, all right-handed, with educational attainment levels: 1 GCSE, 14 A grades, 17 undergraduates and 18 postgraduates; in the UK education system, GCSE refers to secondary education and A grades refer to advanced levels leading to university). All participants reported no contraindications to electrical stimulation (see supplemental data in the full screening list) and no history of calculation difficulties, dyslexia, or attention deficits. The proposed study received ethical approval from the University of Oxford Medical Sciences Interdivisional Research Ethics Committee (Agreement No.: MSD-IDREC-C2-2014-033).

[0093] An overview of experimental paradigms and stimuli

[0094] During the experiment, participants completed four blocks of 50 multiplication problems, each block consisting of a baseline block (e.g., for obtaining baseline metrics) and three stimulus blocks (e.g., for obtaining test data to refine the Gaussian process model). After recording an initial 4-minute resting-state EEG (rs-EEG), the task was explained to the participants, and they completed 10 practice trials, followed by the baseline block during which no tACS was applied. Participants then went through three multiplication blocks, in which they received tACS. Before each tACS block, the BO algorithm was run to determine the stimulus parameters (current intensity and frequency) to be delivered during the upcoming experimental block based on the individual participant's performance in the baseline block (i.e., based on the participant's baseline data values). The stimulus parameters were automatically selected by the algorithm and applied while maintaining blinding between the participant and the experimenter. Rs-EEG was recorded again after each stimulus block.

[0095] Behavioral stimuli

[0096] Arithmetic performance was tested using an arithmetic computation paradigm consisting of problems involving multiplying a one-digit number by a two-digit number, resulting in a three-digit number. The computation paradigm was used instead of the extraction paradigm because the computation is already associated with increased activation in the frontoparietal network. No multiplication operation included operands with the digits 0, 1, and 2 to account for variations in difficulty. Furthermore, the number of operands for the two-digit number was no less than 15, and no duplicate digits were present. The multiplication problems were visually presented to the object on a screen, with correct and incorrect answers placed to the left and right sides below the problem. The positions of the correct and incorrect answers were randomly assigned to the left and right sides of the screen, and they were always 10 numbers apart. Finally, each arithmetic multiplication problem was composed of a new problem.

[0097] Measurement of baseline capability

[0098] An arithmetic baseline task containing 50 different arithmetic multiplication operations was proposed to measure an individual's arithmetic ability in terms of response time and accuracy. The baseline drift rate for each subject was then calculated based on a dual-selection EZ-diffusion model. This model was chosen to reliably combine response time and accuracy in a single outcome, which could be optimized using a Bayesian optimization procedure. The 50 trials completed in the baseline block were randomly split into two, and two independent drift rates were calculated. One was used as a measure of the participant's baseline ability, while the other was used to normalize the drift rate calculated during the optimization phase (e.g., during the Bayesian optimization experimental procedure). This was done to eliminate the correlation between the participant's baseline ability score and the normalized score in each stimulus block. To reduce fatigue, subjects rested for 30 seconds after every 10 trials. After completing the baseline task, subjects had a short rest (approximately 3 minutes) before continuing with the Bayesian optimization experimental procedure.

[0099] Bayesian optimization experimental procedure

[0100] Before each stimulus block, a Bayesian optimization procedure was run to determine the stimulus parameters to be used. A total of 150 different multiplication problems were applied during the experimental procedure (three blocks of 50 trials each). The combination of tACS parameters (frequency and current) depended on the Bayesian optimization and varied for each subject and between blocks. Furthermore, performance from the baseline task was input as a contextual variable (e.g., as a baseline metric c). Therefore, behavioral performance optimization relied on the frequency and current of the tACS and baseline cognitive ability, as indicated, for example, by the drift rate as a baseline metric. Each subject received three different frequency-current tACS combinations while simultaneously performing mental arithmetic problems. The subject then had to indicate which answer was correct as quickly as possible. The correct answer was indicated by pressing a left or right button on a response box located in front of the participant corresponding to the answer's position on the screen. Participants were explained that they must answer as accurately and quickly as possible. Immediately after each block, the performance drift rate was calculated, and another rs-EEG was measured for 4 minutes.

[0101] Transcranial alternating current stimulation (tACS)

[0102] Alternating current stimulation was always applied to the left frontoparietal network. The tACS procedure consisted of two stimulation electrodes (3.14 mm in diameter) using NGPistim Ag / AgCl electrodes (F3 and P3) and a return electrode (Cz) using Starstim 32 (Neuroelectrics, Barcelona). Electrode impedance was maintained at <10 kΩ. For the burn-in phase of the study, stimulation intensity ranged from 0.1 mA to 1.6 mA in 0.1 mA steps, peak-to-peak. For optimization, 0 mA was also added to control for possible spurious effects. We selected this maximum stimulation intensity based on a small preliminary study of 3 subjects to determine the maximum comfort intensity.

[0103] Stimulation was administered in a double-blind manner during three experimental blocks, each lasting a maximum of 10 minutes. Stimulation began 45 seconds before the start of each block and changed after each block. If a subject received a 0 mA stimulus intensity (sham stimulus) during a block, a 30-second rise and fall was initiated to provide initial skin sensation during the stimulation to ensure blinding. When a subject completed a block within 10 minutes, the stimulus was lowered for 30 seconds, and the subject continued rs-EEG for 4 minutes. It should be noted that subjects who completed the task faster than 10 minutes did not receive the full length of stimulation (24 out of 150 stimulation blocks lasted less than 10 minutes but longer than 7.84 minutes, and 126 stimulation blocks lasted longer than 10 minutes). This is not a problem, as the current study only investigated the online effects of tACS. After completing a block associated with one tACS combination, participants completed a questionnaire asking them several questions designed to estimate the level of sensation experienced during the stimulation (see supplemental information for the full questionnaire). We use this data to assess the relationship between the intensity level of each sensation and the tACS amplitude.

[0104] Simulation Analysis

[0105] Several simulations were run to validate the cBO procedure during arithmetic performance and tACS. This analysis aims to show that BO can outperform random sampling of different frequency-current tACS combinations. Furthermore, we wanted to compare the effects of different acquisition functions (EI and UCB) and different dimensions in the range of 3 to 6. Note that the current study includes three dimensions: frequency, current, and personalized baseline capability (baseline metric). Therefore, we ran 60 iterations using the Hartmann function, which includes four local minima across three dimensions, similar to our experimental procedure for BO, where we have three dimensions. The same applies to the Hartmann function with 6 dimensions, the Ackley function with 5 dimensions, and the G function with 4 dimensions. These simulations were noise-free, as BO is primarily designed for noise-free environments. In contrast, human studies are prone to noisy evaluations. Therefore, we decided to compare the effects of different noise variation values. The same Hartmann 3D function was run to introduce noise into the simulations. Performance in these simulations was compared based on distance from the known optimizer location and Euclidean distance as a metric. Finally, we used the Hartmann function to compare a BO that models only the parameters used for the stimulus with a BO that models both the parameters used for the stimulus and the baseline data.

[0106] EEG analysis of spectral power and frontoparietal theta connectivity

[0107] The remaining rs-EEG data were divided into 2-second segments with a 1-second overlap, and windowed using a Hann window. Subsequently, the data were transformed to the frequency domain using a Fast Fourier Transform (FFT). The θ (4-7.5Hz) and β (14-30Hz) bands were determined based on their relative power (μV). 2 The power was calculated and normalized by dividing the absolute frequency power of each band by the average absolute power in the range of 1.5–30 Hz. Furthermore, we decided to normalize the power by dividing the absolute frequency power of the applied tACS frequency value by the average absolute power in the range of 4–50 Hz. The weighted phase lag index (wPLI) in the ranges of θ and β was calculated to determine the phase lag synchronization between the left frontal and parietal regions at baseline and after each tACS block. This calculation was performed for complementary channels F3 and P3. θwPLI was calculated for 4–8 Hz in 1 Hz steps, and βwPLI was calculated from 14–30 Hz in 4 Hz steps. Furthermore, we normalized wPLI by calculating the wPLI at a specific tACS frequency and dividing it by the baseline wPLI at the same frequency.

[0108] First, outliers were removed using Cook distance before running the statistical models. To focus on the relationship between arithmetic baseline capability and spectral power, separate regression models were run using θ and β power as dependencies. Similarly, we tested the relationship between frontoparietal θ and β connectivity scores by running several regression models for θwPLI with steps of 1 Hz and then 4 Hz.

[0109] Experimental work on ADHD

[0110] The following describes the work demonstrating the effectiveness of the disclosed method for obtaining personalized parameters for transcranial electrical stimulation (hereinafter “tES”) in the treatment of attention deficit hyperactivity disorder (ADHD).

[0111] A simulation function was designed to demonstrate the utility of considering ADHD heterogeneity in order to tailor the optimal tES parameters using personalized Bayesian optimization. The simulation function was set with two input dimensions: (1) tES current intensity (mA) in the range of 0.1 to 2.0, and (2) neurophysiology (TBR) in the range of 0.6 to 3.0. Neurophysiology was considered a personalized variable, specific to each participant. The output of the simulation function was the clinical outcome, which, for illustrative purposes, ranged from -1 (adverse outcome) to 1 (desired outcome).

[0112] Figure 15 (for 3D) and Figure 16(The simulation function is shown in 2D, with peak points indicated by white circles.) Based on previous findings, it is expected that those with higher TBR will require higher currents than those with lower TBRs. Given personalized variables of neurophysiology, the inventors selected current intensities, tested the function, and observed clinical outcomes. It should be noted that it is neither possible to control nor optimize these personalized variables.

[0113] The following three methods were run and compared over 30 iterations: 1) random search, 2) non-personalized Bayesian optimization (hereinafter referred to as Bayesian optimization), and 3) personalized Bayesian optimization (which can be implemented using any of the methods described above for obtaining personalized parameters). These over 30 iterations included 6 randomly selected points at the beginning. The program is implemented in Python. Each iteration takes two seconds to suggest new parameters.

[0114] The true optimal parameters for each neurophysiological value (which are unknown to scientists / clinicians) are in Figure 17 The dots are shown in black. It can be seen that different neurophysiological values ​​will require different current intensities to allow for optimal clinical outcomes.

[0115] The optimal parameters estimated at the final iteration through Bayesian optimization and personalized Bayesian optimization are... Figure 18 Visualization is provided. Parameters suggested by personalized Bayesian optimization are depicted with white squares. The true optimal parameters are depicted with black dots. The result of Bayesian optimization is depicted with a white dashed line. It can be seen that the parameters suggested by personalized Bayesian optimization (white squares) are very close to the true optimal parameters (black dots). In contrast, Bayesian optimization without considering personalized neurophysiological values ​​suggests a flat line (white dashed line).

[0116] Figure 19 The parameters selected through Personalized Bayesian optimization during iteration are depicted. Points randomly selected from the initial stage are represented by black squares, while points selected by Personalized Bayesian optimization are represented by white squares. These points are adapted based on neurophysiological values, as indicated by the white dashed lines. It becomes apparent that Personalized Bayesian optimization can identify high-performance regions because most of its selected parameters remain within high-value regions (dark gray). A desirable property of Personalized Bayesian optimization is that while it primarily selects points in high-value regions, it retains some probability of exploration to obtain information about the fundamental functions.

[0117] Figure 20 The performance of different methods was compared. Figure 20The results show that personalized Bayesian optimization outperforms both Bayesian optimization and random search in delivering optimal clinical outcomes. In the early stages (at iteration 10), the performance of personalized Bayesian optimization is only slightly slower than that achieved by Bayesian optimization and random search (after 20 and 30 iterations, respectively). In other words, in this experiment, personalized Bayesian optimization is twice as fast as Bayesian optimization and three times as fast as random search, thus requiring significantly fewer resources.

Claims

1. A computer-implemented method for obtaining personalized parameters of transcranial stimulation, comprising: Receive baseline data about the test subject, the baseline data including information about the test subject acquired before transcranial stimulation is applied to the test subject; as well as A Gaussian process model of the performance of one or more training subjects is used to obtain personalized parameters for transcranial stimulation of the test subjects based on received baseline data, wherein: The Gaussian process will jointly model the object's performance during and / or after transcranial stimulation as a function of both of the following: i) Define the parameters of the transcranial stimulation; and ii) Baseline data of the one or more training objects.

2. The method of claim 1, wherein the baseline data represents performance unaffected by transcranial stimulation.

3. The method according to claim 1 or 2, wherein obtaining the personalized parameters further comprises: Obtain test data representing the performance of the test subject during and / or after the application of transcranial stimulation with multiple corresponding parameter combinations; The Gaussian process model was refined using the obtained test data; and The personalized parameters are obtained using a refined Gaussian process model.

4. The method of claim 3, wherein the Gaussian process model is refined in an iterative Bayesian optimization process, wherein the iterative Bayesian optimization process selects to sample the next step in each step, wherein the combination of parameters used for stimulation optimizes the acquisition function.

5. The method of claim 4, wherein the acquisition function is configured to be computationally cheaper to optimize than the objective function of the Gaussian process model.

6. The method of claim 5, wherein the acquisition function comprises one or more of the following: a Gaussian process confidence upper limit; and a desired improvement.

7. The method of claim 1, further comprising training a Gaussian process model to provide the Gaussian process model for obtaining the personalized parameters.

8. The method of claim 7, wherein training the Gaussian process model includes using training data representing the performance of each of a plurality of training subjects during and / or after the application of transcranial stimulation having a plurality of corresponding parameter combinations for each training subject.

9. The method of claim 8, wherein the Gaussian process model is trained in an iterative Bayesian optimization process, wherein the iterative Bayesian optimization process selects to sample the next step at each step, wherein the combination of parameters used for stimulation optimizes the acquisition function.

10. The method of claim 9, wherein the acquisition function is configured to be computationally cheaper to optimize than the objective function of the Gaussian process model.

11. The method of claim 10, wherein the acquisition function comprises one or more of the following: a Gaussian process confidence upper limit; and a desired improvement.

12. The method of claim 1, wherein the transcranial stimulation comprises any one or more of the following: transcranial alternating current stimulation; transcranial random noise stimulation; transcranial direct current stimulation; transcranial magnetic stimulation; transcranial focused ultrasound; and transcranial optical stimulation.

13. The method of claim 1, wherein the parameters for transcranial stimulation include one or more of the following: one or more frequencies, currents, phases, durations, doses, and brain regions.

14. The method of claim 1, wherein the performance is represented by one or more of the following: accuracy; reaction time; test score; and subjective report.

15. The method of claim 1, wherein the performance is represented by a drift rate value, which is obtained by applying a diffusion decision model to the results of tests performed on the respective objects.

16. A data processing apparatus comprising a processor configured to perform the method according to any of the preceding claims.

17. A method for applying transcranial stimulation, comprising: Perform the method according to any one of claims 1-15 to obtain personalized parameters of transcranial stimulation for the test subject; as well as Transcranial stimulation was applied to the test subject using the personalized parameters.

18. A computer program product comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1-15.

19. A computer-readable data carrier having a computer program according to claim 18 stored thereon.

20. A transcranial stimulation system, comprising: A stimulation unit configured to provide transcranial stimulation to a subject; as well as A data processing device configured to control the operating parameters of the stimulation unit to define the parameters of the transcranial stimulation, wherein: The data processing device is configured to perform the method according to any one of claims 1-15 to obtain personalized parameters of the object, and to control the operating parameters of the stimulation unit based on the obtained personalized parameters.

Images