Digital twin system for predicting neuro-hemodynamic result and blood pressure prediction method using same

Abstract

The present invention relates to a digital twin system in which a computational processor predicts a nucleus tractus solitarii (NTS)-based neuro-hemodynamic result, the digital twin system comprising: an encoding unit for converting a hemodynamic value (BPT) acquired at time T into a state of a normalized NTS latent space; a calculation unit for calculating a next latent state from the latent space state converted by the encoding unit by using an artificial neural network; and a decoding unit for re-converting, into an updated hemodynamic value (BPT+1), the next latent state calculated by a stimulus-based dynamics calculation unit.

Description

Neuro-hemodynamic result prediction digital twin system and blood pressure prediction method using the same

[0001] The present invention relates to a digital twin system and method for predicting neuro-hemodynamic results, and more specifically, to a digital twin system based on the collective dynamics of a baroreflex-related NTS neuronal population involved in hemodynamic perturbation among visceral sensory stimuli, and as a result, to a digital twin system and method for predicting neuro-hemodynamic results capable of predicting hemodynamic responses based on visceral sensory stimuli.

[0002] Recent advancements in Brain-Machine Interface (BMI) technology are revolutionizing neural stimulation practices, providing promising solutions for restoring neurological dysfunction and promoting brain function.

[0003] In fact, advancements in BMI for neural stimulation have achieved significant success in various medical fields. In particular, applications for treating chronic diseases have been developed, such as deep brain stimulation to alleviate brain disorders like type 1 and type 2 epilepsy, and spinal cord stimulation for the rehabilitation of motor disorders resulting from spinal cord injury.

[0004] Despite these achievements, the application of BMI technology to visceral sensory nerve stimulation therapy to artificially control the functions of internal organs, such as the hemodynamic function of the cardiovascular system, has not yet been realized.

[0005] Although the treatment has shown potential to effectively regulate hemodynamic function, clinical interpretation is difficult due to the high interpersonal variability of the results following stimulation.

[0006] For example, baroreflex activation therapy faces clinical limitations such as excessive blood pressure (BP) reduction and inefficiency due to the open-loop stimulation method and heuristic selection of stimulation parameters (see Non-patent Literature 1, Non-patent Literature 2).

[0007] This has raised the need for personalized computational models, the concept of digital twins (patient-specific virtual replicas), as an aspect of BMI technology to individually predict stimulus-centered outcomes.

[0008] Accurate predictive modeling provided by digital twins is, in fact, an essential prerequisite for modern closed-loop feedback control systems. However, digital twins for complex stimulus-centered responses, such as the nonlinear evolution of blood pressure observed in therapy through baroreflex activation, are still difficult to find.

[0009] Designing digital twins to predictively model stimulus-based responses in internal organs, such as the cardiovascular system, requires a fundamental understanding of the anatomical and computational mechanisms by which stimulus inputs regulate the autonomic nervous system and internal organs. Classical and recent studies have established a neuroanatomical atlas of the autonomic nervous system.

[0010] The autonomic nervous system hemodynamically regulates visceral sensory afferents from the cardiovascular system to the brainstem. The brainstem is an essential intermediate processor between afferent and efferent pathways. In particular, the nucleus tractus solitarius (NTS) in the brainstem plays an important role in integrating visceral sensory information and coordinating the rostral ventrolateral medulla (RVLM) and the dorsal motor nucleus of the vagus (DMV).

[0011] Pre-autonomic nodes regulate the efferent pathways of the autonomic nervous system by regulating pre-autonomic neurons in spinal cord regions, such as the intermediolateral nucleus (IML), to provide feedback to peripheral viscera.

[0012] Despite this neuroanatomical clarity, the computational mechanisms related to the stimulus-centered dynamics of neural activity and hemodynamic function are not yet known.

[0013] [Prior Art Literature]

[0014] [Non-patent literature]

[0015] (Non-patent Document 1) Heusser, K. et al. Carotid baroreceptor stimulation, sympathetic activity, baroreflex function, and blood pressure in hypertensive patients. Hypertension 55, 619-626 (2010).

[0016] (Non-patent Document 2) Bisognano, JD et al. Baroreflex activation therapy lowers blood pressure in patients with resistant hypertension: results from the double-blind, randomized, placebo-controlled rheos pivotal trial. Journal of the American College of Cardiology 58, 765-773 (2011).

[0017] Therefore, the problem that the present invention aims to solve is to provide a digital twin modeling system based on a computational mechanism related to the stimulus-centered dynamics of neural activity and hemodynamic function.

[0018] To solve the above problem, the present invention is a digital twin system for predicting neuro-hemodynamic results based on the nucleus accumbens solitary (NTS) performed on a computationally capable processor, wherein the hemodynamic value (BP) obtained at time T T ) normalized NTS latent space( An encoding unit that converts the latent space state converted by the encoding unit into the next latent state ( A calculation unit that calculates ); and the next potential state calculated by the stimulation-based dynamics calculation unit, an updated hemodynamic value (BP). T+1 A digital twin system is provided that includes a decoding unit for converting )

[0019] In one embodiment of the present invention, the computation unit uses neural circuit modeling utilizing a recurrent neural network (RNN).

[0020] In one embodiment of the present invention, the recurrent neural network (RNN) successfully models the collective dynamics of the NTS neural circuit to replicate the measured single neural activity and neural trajectory.

[0021] In one embodiment of the present invention, the encoding unit converts a hemodynamic value (BP measurement) into a vector form representing the potential state of an NTS neuron population.

[0022] In one embodiment of the present invention, the decoding unit linearly combines each element of the latent space vector using a predefined decoding vector.

[0023] In one embodiment of the present invention, the decoding unit Each dimension value is multiplied by the weight of the decoding vector, and all of these are summed to calculate the Δ prediction value.

[0024] In one embodiment of the present invention, the hemodynamic value is blood pressure.

[0025] The present invention also relates to a blood pressure prediction method for predicting a solitary nucleus accumulator (NTS)-based neuro-hemodynamic result performed on a computationally capable processor, wherein the hemodynamic value (BP) obtained at time T T ) normalized NTS latent space( An encoding step that converts the latent space state converted in the encoding step into the next latent state using an artificial neural network A calculation step for calculating ; the next potential state calculated in the calculation step is an updated hemodynamic value (BP). T+1 A blood pressure prediction method is provided, characterized by including a decoding step that converts ) back.

[0026] In one embodiment of the present invention, a neural circuit modeling utilizing a recurrent neural network (RNN) is used in the calculation step.

[0027] In one embodiment of the present invention, the recurrent neural network (RNN) successfully models the collective dynamics of the NTS neural circuit to replicate the measured single neural activity and neural trajectory.

[0028] In one embodiment of the present invention, the encoding step converts the hemodynamic value (BP measurement) into a vector form representing the potential state of the NTS neuron population.

[0029] In one embodiment of the present invention, the decoding step linearly combines each element of the latent space vector using a predefined decoding vector.

[0030] In one embodiment of the present invention, the decoding unit The Δ prediction value is calculated by multiplying each dimension value by the weight of the decoding vector and summing them all.

[0031] A blood pressure prediction method characterized in that, in one embodiment of the present invention, the hemodynamic value is blood pressure.

[0032] According to the present invention, a digital twin system based on the collective dynamics of a baroreflex-related NTS neuronal population involved in hemodynamic perturbation among visceral sensory stimuli is provided, and as a result, it is possible to predict hemodynamic responses based on visceral sensory stimuli.

[0033] Figure 1 is a diagram illustrating the structure of the neural population of the nucleus solitary (NTS) for predicting hemodynamic state.

[0034] Figure 2 illustrates the interactions and cross-correlation between neurons within the NTS and explains the collective dynamics of neural responses to stimulation.

[0035] Figure 3 is a diagram illustrating the modeling process and performance of predicting hemodynamic response (ΔBP) based on the activity of NTS nerve groups.

[0036] Figure 4 illustrates the process of analyzing consistent hemodynamic responses (ΔBP) in a common latency space by normalizing the neural trajectories of NTS neural populations observed in various individuals.

[0037] Figure 5 is a diagram illustrating the process of predicting stimulus-based hemodynamic responses using H-BIND (Hierarchical Brain-Inspired Neural Decoder).

[0038] FIG. 6 is a block diagram of a system for predicting blood flow information, such as blood pressure, from NTS neural information according to one embodiment of the present invention.

[0039] FIG. 7 is a step diagram of a blood pressure prediction method for predicting neuro-hemorrhagic results based on a nucleus isolate (NTS) performed on a computationally capable processor according to one embodiment of the present invention.

[0040] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings.

[0041] Before describing the present invention in detail, the terms and words used in this specification should not be interpreted as being unconditionally limited to their ordinary or dictionary meanings, and the inventor of the present invention may appropriately define and use the concepts of various terms to best describe their invention.

[0042] Furthermore, it should be understood that these terms or words should be interpreted in a meaning and concept consistent with the technical spirit of the present invention.

[0043] In other words, the terms used in this specification are used merely to describe preferred embodiments of the invention and are not intended to specifically limit the content of the invention.

[0044] It should be noted that these terms are defined in consideration of the various possibilities of the present invention.

[0045] Additionally, in this specification, singular expressions may include plural expressions unless the context clearly indicates a different meaning.

[0046] In addition, you should be aware that even if it is expressed in the plural, it may contain a singular meaning.

[0047] Throughout this specification, where it is stated that a component "includes" another component, unless specifically stated otherwise, this may mean that it does not exclude any other component but may include any other component.

[0048] Furthermore, in cases where it is stated that a component "exists inside or is installed in connection with" another component, this component may be installed in direct connection with or in contact with the other component.

[0049] In addition, they may be installed spaced apart at a certain distance, and in the case where they are installed spaced apart at a certain distance, there may be a third component or means for fixing or connecting the component to another component.

[0050] Meanwhile, it should be noted that the description of the third component or means mentioned above may be omitted.

[0051] On the other hand, if it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there is no third component or means.

[0052] Likewise, other expressions describing the relationship between each component, such as “between” and “right between”, or “adjacent to” and “directly adjacent to”, should be interpreted as having the same intent.

[0053] In addition, terms such as “one side,” “other side,” “one side,” “other side,” “first,” “second,” etc., in this specification are used to ensure that one component can be clearly distinguished from another component.

[0054] However, it should be noted that the meaning of the component is not used restrictively by such terminology.

[0055] In addition, positional terms such as "top," "bottom," "left," and "right" used in this specification should be understood as indicating the relative position of the corresponding component in the drawing.

[0056] Furthermore, unless an absolute location is specified regarding their positions, terms related to these locations should not be understood as referring to absolute locations.

[0057] Furthermore, in the specification of the present invention, terms such as “…part,” “…unit,” “module,” and “device,” if used, refer to a unit capable of handling one or more functions or operations.

[0058] You should be aware that this can be implemented in hardware, software, or a combination of hardware and software.

[0059] In the drawings attached to this specification, the size, location, connection relationships, etc., of each component constituting the present invention may be described in a partially exaggerated, reduced, or omitted manner for the convenience of explanation or to sufficiently clearly convey the concept of the present invention, and therefore, the proportions or scale may not be strictly accurate.

[0060] In addition, in describing the present invention below, detailed descriptions of components that are deemed to unnecessarily obscure the essence of the invention, such as known technologies including prior art, may be omitted.

[0061] In order to solve the aforementioned problem, the present invention systematically summarized the effects of NTS on neural population activity and hemodynamic response (BP), and then conducted experiments and analyses as follows to provide a digital twin system based thereon.

[0062] Examples

[0063] In this experiment, stimulus-based neural responses were analyzed by simultaneously measuring single-neuron activity and hemodynamic responses (blood pressure, BP) in the nucleus solitary tract (NTS) of rats. The experimental procedure is structured as follows.

[0064] Surgery and anesthesia

[0065] The mice used in the experiment were adult male Sprague-Dawley mice, and anesthesia was administered by injecting a user-defined urethane solution into the abdominal cavity prior to the experiment. Subsequently, a polyethylene tube was inserted into the femoral artery to enable continuous recording of BP.

[0066] Measurement of single neuron activity

[0067] The skull was incised to expose the nucleus solitarius of the brainstem, and a 16-channel silicon probe was inserted into the nucleus solitarius to record the extracellular potential of a single neuron. Recorded neural activity was analyzed using Klusta software to automatically detect spikes and sort them into single neurons.

[0068] Blood pressure (BP) measurement

[0069] A catheter inserted into the femoral artery for BP measurement was connected to a pressure transducer and routed outside the Faraday cage to minimize electrical interference. The recorded BP data was saved in MATLAB format, and hemodynamic functions such as average BP and heart rate variability were calculated through post-processing by filtering unique frequency bands.

[0070] Stimulation protocol

[0071] To induce neural activity in the nucleus solitary tract, electrical stimulation was applied by targeting the solitary tract projected to the NTS. The stimulation was delivered in the form of a square wave set with a pulse width of 100 μs and a frequency of 20 Hz, and the stimulation intensity for each individual was adjusted to 150–250 μA. This stimulation induced a dynamic response including a rapid drop in blood pressure at the start of stimulation and a recovery phase.

[0072] Through this experimental method, the effects of stimulation on the neural population activity and hemodynamic response (BP) of NTS were systematically analyzed, and based on this, neuro-hemodynamic coupling was confirmed. In particular, the neural computation mechanism within NTS was identified from the perspective of the population dynamics of NTS neuronal populations related to baroreflexes involved in hemodynamic perturbations among visceral sensory stimuli, and a digital twin utilizing NTS neural information is provided.

[0073] Figure 1 is a diagram illustrating the structure of the neural population of the nucleus solitary (NTS) for predicting hemodynamic state.

[0074] Figure 1a illustrates the process of dimensionality reduction of NTS neural population data into a low-dimensional latent space to visualize neural trajectories appearing in various individuals (rats) and to normalize them. In the above process, neural trajectories are represented as a two-dimensional representation of the latent space, and a function is performed to normalize and position each trajectory in a common latent space to consistently analyze neural responses between individuals. This indicates that the hemodynamic response (Δ) corresponds linearly to a specific axis (y₂) of the latent space, providing the basis for the neuro-hemodynamic coupling mechanism.

[0075] Figure 1b illustrates a dynamic modeling process using a nucleus accumbens (NTS) neural circuit model to predict stimulus-centered hemodynamic responses.

[0076] Figure 1b includes the step of calculating neural population dynamics in an NTS neural circuit model according to visceral sensory input u(t) and predicting the hemodynamic state therefrom.

[0077] The digital twin model according to the present invention performs the function of predicting a hemodynamic response to a stimulus by encoding, dynamically calculating, and decoding a hemodynamic state in a low-dimensional latent space, and the computational device for this includes an encoding unit, a calculation unit, and a decoding unit for performing the above-described function.

[0078] As a result, it was demonstrated that the temporal evolution of NTS population states exists in a two-dimensional (2D) latent space where hemodynamic perturbations are linearly encoded (see Fig. 1a).

[0079] Referring to the results of FIG. 1, the present invention provides a digital twin framework for predictive modeling of visceral sensory stimulation-based hemodynamic responses using the collective dynamics of NTS that form the basis of neuro-hemodynamic coupling, said digital twin replicates the neural computation mechanism within NTS by implementing a biomimetic neural network model (see FIG. 1b).

[0080] In addition, through the combination of latent space and bidirectionally convertible hemodynamic perturbations, model parameters can be individually optimized using only hemodynamic records. Thus, the present invention can not only lay a fundamental foundation for BMI technology toward a closed-loop system capable of individually optimizing visceral sensory nerve stimulation to control the function of internal organs, but also advance the understanding of the neural basis of the autonomic nervous system for regulating organs.

[0081]

[0082] Stimulus-based population dynamics of NTS neuronal populations in latent space

[0083] Figure 2 illustrates the interactions and cross-correlation between neurons within the NTS and explains the collective dynamics of neural responses to stimulation.

[0084] Figures 2a to 2c explain how the stimulus response is formed through neural activity occurring in the NTS according to the stimulus input and cross-correlation between neurons, and Figure 2a records the change in Δ according to the stimulus and visualizes the neural activity of each neuron to show the association with the hemodynamic response.

[0085] Figure 2b shows the change in firing rate of 192 neurons over time to represent the collective response to the stimulus, and Figure 2c explains the statistical significance of the interaction between neurons by comparing the correlation coefficients and shared variances between the measured data and the dummy data.

[0086] Figures 2d, e to f show the results of analyzing the mechanism encoding stimulation-induced hemodynamic changes in NTS neural populations in a latent space. Figure 2d represents the stimulation-based neural trajectory of the NTS neural population in a two-dimensional latent space, visually distinguishing the 'activation', 'adaptation', and 'resting' states in response to stimulation, and explains the ring structure of the neural trajectory using continuous homology.

[0087] Figure 2e visualizes the pattern of neural group activity by comparing firing rates according to Δ and neural state stage θ, and Figure 2f shows the results of an analysis in terms of dispersion and information regarding the extent to which Δ and neural state stage effectively reflect hemodynamic changes.

[0088]

[0089] Referring to the drawings above, the present invention recorded the extracellular single-unit activity of NTS and simultaneously measured the femoral artery BP of a mouse (see Fig. 2a).

[0090] The electrical pulse train was delivered from a single pathway in the brainstem projected to the NTS as visceral sensory stimulus input.

[0091] Only NTS neurons responding to neural stimulation were selected for recording (n = 192, 72% of 266 neurons in 10 rats). Stimulus input resulted in the temporal evolution of heterogeneous mononeural responses (see Fig. 2b).

[0092] In one embodiment of the present invention, the NTS neuron population consisted of 51% (n = 98), 29% (n = 55), and 20% (n = 39) complex responses out of a total of 192 neurons, and the effect of interneuronal dialogue on these heterogeneous stimulus-based responses was investigated by analyzing the cross-correlation and shared variance between the observed neurons. The pairwise cross-correlation between NTS neurons was higher than that of the dummy data, which consisted of a randomly shuffled time course of the measured neuron activity. The absolute values ​​of the cross-correlation (log-scaled) were as follows: 0.0533 (Interquartile Range [IQR]: 0.046–0.060) and 0.59 (IQR: 0.48–0.64) for the dummy and measured data, respectively; p < 1.83×10⁻⁶ -4.

[0093] The shared variance of each neuron's response was highly explained together with other neurons: 27% (IQR: 22–35%) and 83% (IQR: 76–87%) for dummy and measured data, respectively; p < 1.83×10⁻⁶ -4 (Fig. 2c).

[0094] These connections between neurons suggest that heterogeneous responses originate from interconnections formed within NTS neuron populations, which supports the idea that NTS entail self-modulating circular neural circuits through interconnections.

[0095] The results of the present invention verify that local neural circuits within a population of NTS neurons derive population dynamics that simplify the complex neural activities of numerous single neurons into the dynamics of some effective variables called latent state variables. This approach is based on an understanding of the dynamic characteristics of neural circuits, whereby neural circuits contain large-scale information stored in numerous single neurons, but each neural activity of a single neuron is a dimensional state space considered as 1D dynamics.

[0096] However, the interneuronal connections resulting from the network structure of such circuits can restrict collective dynamics to a hidden dynamic state space of much lower dimension than the original number of neurons.

[0097] In this specification, restricted dynamics are referred to as collective dynamics, and the hidden state space containing collective dynamics is referred to as latent space.

[0098] To examine the population dynamics of NTS neuron populations, the dimensionality of stimulus-based neural activity of NTS neurons was reduced using IOSMAP, a nonlinear and uncontrolled dimensionality reduction technique, which was effective in providing a concise interpretation and visualization of population dynamics (see Chaudhuri, R., Gerek, B., Pandey, B., Peyrache, A. & Fiete, I. The intrinsic attractor manifold and population dynamics of a canonical cognitive circuit across waking and sleep. Nature neuroscience 22, 1512-1520 (2019)).

[0099] After reducing the dimensionality, the phase of the neural trajectory driven by the group dynamics and the effective dimension of the latent space were distinguished, and the characteristics of the group dynamics and the latent space were analyzed geometrically (topologically).

[0100] The topology of each neural trajectory was determined by the homology group (H0, H1, H2) to which they belonged. The homology group represents a unique topology of geometric structure (H0: point or solid sphere, H1: ring, H2: Halloween sphere or toroid; Fig. S2a), which was investigated using the method of continuous homology 39.

[0101] Continuous homology generated Betti barcodes visualizing the presence of each homology group (see "Continuous Homology" in the Methods section). The invention confirmed that stimulus-based neural trajectories generated from NTS neural populations followed a ring topology, notably showing a rapid correlation with long barcode lifetimes (meaning 'bar length') only in H1, as shown in Fig. 2d, whereas there were no long bars in H2 (statistical analysis comparing z-score lifetimes of the longest bars: 12.38 for H1 [IQR: 10.33-14.01], 2.252-3.415 for H2], p < 1.83×10⁻¹⁰ -4 The effective dimension of the latent space was specified based on how many principal components (PCs) of the latent space are needed to sufficiently capture the variance explained by the trajectory.

[0102] According to the results of the present invention, the first and second components were found to explain more than 86% of the dynamic variance (mean ± standard deviation (STD)) of the explained variance (65.2 ± 8.05% for the first PC and 21.4 ± 6.21% for the second PC).

[0103] Therefore, it was found that the effective latent dimension belongs to the 2D dimension. In addition, the correlation dimension was measured by the gradient value of the change in the average number of neighbor points around a neural data point with respect to the change in the threshold distance defining the neighbor. The order of increase of neighbors relative to the threshold distance was close to 1 (average ± STD of the gradient of the number of neighbors relative to distance: 0.84 ± 0.031, Fig. S2d). This corresponds to the characteristics of a ring-shaped topology.

[0104] The collective dynamics of the latent space effectively captured both the information and variance inherent in the single neuron activity of NTS neurons, as visually illustrated in Figure 2e, thereby effectively summarizing the original high-dimensional dynamics of NTS neuron collective dynamics.

[0105] Neural state stages, which are internal variables that quantify the state along a loop-shaped trajectory, distinguished each tuning curve specific to NTS neurons, indicating that all characteristic behaviors of individual neurons were successfully captured by collective dynamics.

[0106] In particular, when comparing the tuning curve aligned with changes in BP (ΔBP) and the tuning curve aligned with the neural state stage (θ), the latter captured the characteristic behavior of the neuron more accurately. These results were further supported by statistical analysis, which showed that in single-neuron activity, ΔBP (fraction of variance: 0.54 [IQR: 0.45-0.65] and 0.78 [IQR: 0.69-0.84], mutual information: 0.82 [IQR: 0.70-0.87] and 1.0 [IQR: 0.91-1.1]) was distinguished from the BP and neural state stages, respectively (see Fig. 2f).

[0107] Therefore, these results show that the dynamic and informational characteristics of NTS single neurons are more accurately represented through population dynamics rather than external variables in the physical domain, such as hemodynamic functions, suggesting that population dynamics can characterize the activity patterns of NTS neuron populations while minimizing the loss of features of single neuron activity.

[0108]

[0109] Neuro-hemodynamic coupling through collective dynamics in latent space

[0110] The present invention discovered a surprising and unique finding that the temporal dynamics of NTS neural activity tend to be similar to stimulus-induced BP among visceral sensory nerve stimuli.

[0111] This direct correlation between NTS activity and hemodynamic output suggested the possibility that neural computations embedded in NTS could have a significant impact on the temporal development of stimulus-centered hemodynamics during neural stimulation.

[0112] This hypothesis introduced and utilized a new phenomenon in which the baroreflex plays a pivotal role in quantitatively guiding the regulation of the circulatory system based on the integration of barosensory input to explain the neural computational mechanism underlying the coupling between NTS neural activity and visceral sensory stimulus-centered hemodynamic results in the following series of experiments.

[0113] Figure 3 is a diagram illustrating the modeling process and performance of predicting hemodynamic response (ΔBP) based on the activity of NTS nerve groups.

[0114] Figures 3a and 3b illustrate the single-neuron prediction performance for predicting ΔBP based on the firing rate of individual neurons. Figure 3a displays the change in firing rate of each neuron according to stimulation and compares the predicted ΔBP with actual measured values, while Figure 3b shows the prediction accuracy of each neuron as a histogram. This suggests that the ΔBP prediction performance is low when relying solely on individual neurons. Figures 3c and 3g explain a method for more accurately predicting ΔBP through a linear decoding process using the latent space of an NTS neuron population, which is as follows.

[0115] Figure 3c illustrates a linear decoding protocol that encodes ΔBP into a latent space and reconstructs ΔBP using a decoding vector, showing the coupling between neural trajectories and ΔBP within the latent space. Figure 3d visualizes high prediction accuracy by comparing the decoded prediction results with actual ΔBP measurements, and Figures 3e and 3f show that linear decoding has superior performance by statistically comparing the error distribution and prediction accuracy between the decoding results and single-neuron-based prediction results. Figure g further quantifies the decoding prediction performance based on the Mean Squared Error (MSE) to visualize the improved performance compared to single-neuron-based prediction.

[0116] Referring to FIG. 3, the digital twin model according to the present invention explains that a linear decoding mechanism that predicts ΔBP by utilizing the latent space of NTS neural populations provides high accuracy in predicting stimulus-centered hemodynamic responses.

[0117] In other words, the above results are consistent with previous results (Fig. 2) regarding various activity patterns of NTS single neurons due to local neural circuits, which means that only a small subset of these neurons can derive a high correlation with BP in trivial parts.

[0118] Furthermore, this heterogeneity among NTS neurons reflected the characteristics of NTS neuron populations in which many neurons regulate baroreflexes that can be regulated by internal processes hidden within the NTS local circuit, rather than hemodynamic functions in the external physical domain. These hidden processes suggest that while individual neurons do not have a direct correlation with external hemodynamics, their collective dynamics can provide a more comprehensive understanding of the relationship between NTS and hemodynamics. Accordingly, the present invention analyzed the correlation between the latent space inhabited by the collective dynamics of NTS neuron populations and hemodynamic perturbations caused by stimulation.

[0119] One embodiment of the present invention assumed that hemodynamic functions can be encoded by the latent space of NTS neuronal populations, as in cognitive and motor functions, with reference to Vyas, S., Golub, MD, Sussillo, D. & Shenoy, KV Computation through neural population dynamics. Annual review of neuroscience 43, 249-275 (2020).

[0120] As a result, the present invention discovered that the 2D latent space is linearly coupled to BP (Fig. 3c).

[0121] Interestingly, each point along the ring-shaped neural trajectory (neural state stage) did not exhibit a unique BP value, but the BP was displayed in a linear subspace of the latent space called the “decoding space” by the present invention.

[0122] This linear decoding space suggests that the NTS latent space can be converted to BP using a linear transformation protocol. Predictions of BP based on the linear protocol showed high prediction accuracy of -0.799 and -1.96 for single-neuron and neural decoding, respectively (log scale of mean squared error: -0.799 and -1.96 for single-neuron and neural decoding, respectively; Fig. 3e).

[0123] Therefore, the results according to the present invention indicate that stimulus-induced BP is encoded within the collective dynamics of the NTS latent space.

[0124] The statistical analysis of the present invention provided reliable evidence for these findings by showing p <1.83×10⁻⁴; Fig. 3f> and smaller prediction errors (log scale of mean squared error: -0.75 [IQR: -0.84 to -0.62]) and -1.22 [IQR: -1.78 to -1.01] for single neurons and neural decoding, respectively, and by presenting much higher prediction accuracy of the decoding latent space (named neural decoding) for single neurons and neural decoding.

[0125]

[0126] Inter-individual consistency of group dynamics in NTS latent space

[0127] As previously shown in Figures 2 and 3, the linear protocol for the ring-shaped topology and neuro-hemorrhagic coupling was common to all rats, but each rat subject showed distinct neural trajectories reflecting inter-individual variability in both population dynamics and coupling protocols, which is explained in Figure 4.

[0128] Figure 4 illustrates the process of analyzing consistent hemodynamic responses (ΔBP) in a common latency space by normalizing the neural trajectories of NTS neural populations observed in various individuals.

[0129] In Figures 4a to 4c, neural trajectories of individual mice were visualized in latent space, and phase alignment and consistency before and after normalization were analyzed. Figure 4a shows neural trajectories from multiple individuals and indicates changes in neural trajectories by classifying them into pre-stimulation, stim., and post-stimulation.

[0130] Figure 4b visualizes the coupling with Δ and highlights that the normalized latent space provides a consistent interpretation of stimulus-centered hemodynamic responses.

[0131] Figure 4c compares the phase lock value (PLV) and standard deviation (STD) before and after normalization, showing that neural trajectories have higher consistency in the normalized latent space.

[0132] Figures 4d through 4f compare the standard deviation of neural trajectories and prediction accuracy by normalizing the latent space to predict Δ. Figure 4d visually demonstrates that Δ can be represented more precisely in the normalized latent space and shows consistency by statistically comparing the standard deviations of Δ before and after normalization. Figure 4e indicates that Δ corresponds linearly to a specific axis (y₂) in the normalized latent space, and Figure 4f quantitatively evaluates the prediction accuracy in the normalized latent space, suggesting that the normalized model provides more consistent prediction performance. The figures illustrate a neural computational mechanism capable of consistently predicting and analyzing stimulus-centered hemodynamic responses across individuals through the normalized latent space.

[0133] In addition, despite inter-individual variability, the rat-specific NTS latent space could be normalized across all rats.

[0134] Each neural trajectory in the latent space (z1, z2) for each rat was normalized by aligning it to a unit circle centered at the origin, and the new space containing the unit circuit was named the normalized latent space (y1, y2). In the normalized latent space, changes in the neural state stage (θ) were found to increase commonality in the population dynamics of the entire rat population.

[0135] Normalization improved the commonality of dynamic characteristics of the entire rat by increasing the phase locking value (PLV) of the resting point distribution before stimulation and the recovery point distribution after stimulation (resting: from 0.26 to 0.83 before stimulation, recovery: from 0.68 to 0.86 after stimulation).

[0136] In this invention, a common decoding space among rats was selected as the vertical axis to decode a normalized latent space for simple calculation (Fig. 4e). The normalized neural trajectory significantly improved the consistency among rats in neuro-hemodynamic coupling by reducing the variability among rats in the latent space representation of ΔBP for each θ (Fig. 4e).

[0137] For the denormalized trajectory and the normalized trajectory, the prediction accuracy was 0.52 [IQR: 0.40-0.57] and 0.37 [IQR: 0.24-0.45], respectively, with p = 0.00258 (Fig. 4d). Then, the effect of normalization on the accuracy of BP prediction was evaluated by comparing the decoding accuracy of the normalized latent space with the decoding accuracy of the rat-specific denormalized latent space.

[0138] In order to decode a normalized latent space for simple calculation, the present invention selected the common decoding space between rats as the vertical axis (Fig. 4e).

[0139] There was no significant difference in prediction accuracy between the unnormalized latent space and the normalized latent space (p = 4.232). This indicates that individual-to-individual variable hemodynamics among visceral sensory nerve stimuli can be adequately predicted using only a general neuro-hemodynamic coupling protocol that relies on a predetermined normalized latent space.

[0140] As indicated by varying prediction accuracy, the variability of stimulus-based neural trajectories observed across individual rats raised essential ambiguities regarding the properties distinguishing individual (rat-specific) latent spaces from normalized latent spaces. To resolve these ambiguities, we examined in detail the underlying factors of variability in neural trajectory shapes and associated prediction accuracy. One simple explanation is that the variability may arise from inter-individual differences in the network structure of NTS neural circuits. However, network structures, such as the rank of the connection matrix, are known to affect only the dimensionality of the latent space. The analysis of the NTS latent space by the present invention refuted this explanation by demonstrating a consistent 2D latent space across rats, as shown in Fig. 2d.

[0141] Another plausible hypothesis is that similar NTS neural circuits corresponding to similar latent spaces are shared among rats, but differences in the recorded subgroups for each rat, known as "recording instability," may cause variability. To test this hypothesis, the present invention first demonstrated that when subgroups were extracted from a given neuron population (n = 100 random subgroups generated by randomly sampling 14 neurons from 28 neurons), these subgroups exhibited different neural trajectories in the latent space.

[0142] These differences included shape distortions from the neural trajectories of the original population and varying prediction accuracy. Then, the present invention demonstrated that these heterogeneous neural trajectories could be normalized to a unit circle without altering the prediction accuracy for each subgroup. Considering the similarity of the observed rat and subgroup trajectories, the present invention determined that the heterogeneous neural trajectories did not originate from individual features of the NTS neural circuit, but rather from the records of different subgroups of each rat, despite the fact that the rats shared similar NTS neural circuits.

[0143] Furthermore, the present invention calculated the neural trajectory of the entire population including all recorded neurons of different mice, and the entire trajectory still exhibits a 2D loop shape. This preservation of the loop topology in the entire population and in each mouse-specific population is consistent with the above relationship between the original population and the subgroups. Additionally, it was verified that the neural trajectory of a given mouse can be (appropriately) replicated by subsampling the entire population that does not include the neurons of that mouse, indicating that similar single-neuron responses were shared across the entire mouse population.

[0144] Therefore, these results demonstrate that NTS neural circuits and latent spaces are consistent among individual rats, and that the measured differences in the accuracy of neural trajectories and neuro-hemodynamic coupling of population dynamics depend significantly on actual recording characteristics. In other words, the present invention constituted the consistent characteristics of NTS population dynamics among rats during visceral sensory nerve stimulation as follows:

[0145] 1) Stimulus-driven hemodynamics follows a ring-shaped neural trajectory in 2D latent space, and

[0146] 2) Stimulus-driven hemodynamic results can be represented in a normalized latent space of the rat liver, and

[0147] 3) The neural computational mechanism embedded in the NTS neural circuit is similar to that of a rat liver.

[0148]

[0149] Predictive modeling of stimulus-based hemodynamics based on neural computational mechanisms

[0150] Through the above results, we were able to identify the dynamic system underlying hemodynamic perturbations during visceral sensory nerve stimulation and develop a digital twin that replicates neuro-hemodynamic responses. In particular, the digital twin was designed as a predictive modeling framework that predicts stimulation-based neuro-hemodynamic results based on the NTS neural circuit model.

[0151] The circuit model includes a system having a neuro-hemodynamic combined protocol capable of linearly transforming from collective dynamics to hemodynamics by reflecting collective dynamics in the NTS latent space. The digital twin framework according to the present invention was named the sensor input-based neurodynamic model (H-BIND, FIG. 1b and 5a).

[0152] Figure 5 is a diagram illustrating the process of predicting stimulus-based hemodynamic responses using H-BIND (Hierarchical Brain-Inspired Neural Decoder).

[0153] Figure 5a illustrates the overall operating mechanism of H-BIND. The stimulus input u(t) induces a hemodynamic change, which is encoded into a latent space to predict the next state through an NTS neural circuit model. This neural circuit model uses a circular feedback structure to calculate dynamic state changes within the latent space and finally derives the predicted hemodynamic state +1 through a decoding process. These steps are described by dividing them into three main processes: encoding, dynamic calculation, and decoding.

[0154] Figure 5b shows the Δ prediction accuracy of individuals (rats) by comparing H-BIND and neural decoding. For the two individuals (Rat7 and Rat8), it can be seen that H-BIND shows a high degree of agreement with the actual measurements and has higher prediction accuracy than the neural decoding method (see b in figure).

[0155] In addition, c in Fig. 5b illustrates the process by which H-BIND predicts changes in the stimulus-centered neural state in a normalized latent space. By visualizing the trajectory and phase changes of each state (resting and adaptive states), the process by which H-BIND accurately tracks changes in the neural state is explained, and the phase error and accuracy according to the changes in the neural state are compared. In Figs. 5d to 5e, the predictive performance of H-BIND is quantitatively evaluated by comparing it with neural decoding.

[0156] Figure 5c shows that H-BIND exhibits higher accuracy and lower mean squared error (MSE) by comparing Δ prediction accuracy and prediction error, and Figure 5d visually demonstrates the superior performance of H-BIND by evaluating phase error and radius error.

[0157] The above figure illustrates that H-BIND performs the function of effectively predicting stimulus-based hemodynamic responses and precisely tracking changes in neural state.

[0158] To explain this more specifically, hemodynamic perturbation prediction by H-BIND, a digital twin system according to the present invention, is based on the current hemodynamic value (BP). T ) normalized NTS latent space( It starts by transitioning to the state of ) (the "encoding" process by the encoding unit). Then, H-BIND uses an artificial neural network to reflect an NTS neural circuit model that incorporates collective dynamics (the "stimulus-based dynamics computation" process by the stimulus-based dynamics computation unit) to the next latent state ( It calculates ). Finally, H-BIND inversely transforms the updated latent state (the "decoding" process by the decoding unit) to obtain the updated hemodynamic value (BP). T+1 ) is obtained. For this H-BIND design, the system identification process was subdivided to configure an NTS neural circuit model-based computation unit and a neuro-hemorrhagic encoding unit / decoding unit.

[0159] In one embodiment of the present invention, the NTS neural circuit of H-BIND utilized a recurrent neural network (RNN) structure. The present invention first assumed that population dynamics utilizing RNNs could be characterized as a linear combination of individual unit activities. This assumption regarding linearization was supported by empirical findings demonstrating that linear regression of NTS single neural activity effectively captures stimulus-based changes in each major component of the NTS latent space. An RNN formulated with a linear relationship between single neural activity and the latent space successfully reproduced stimulus-based neural trajectories in the NTS latent space, and the single neural activity patterns by the RNN unit could reflect the observed single neural activity patterns of the NTS.

[0160] The actual trained RNN successfully modeled the collective dynamics of the actual NTS neural circuit, replicating both the measured single neural activity and the neural trajectory. For the neuro-hemodynamic encoder / decoder, the bidirectional properties of the linear neuro-hemodynamic coupling protocol suggest that stimulus-based hemodynamic perturbations can be converted into stimulus-based neural trajectories in a predetermined normalized latent space.

[0161] By estimating neural trajectories in latent space from hemodynamic perturbations through bidirectionality, it was possible to train an RNN without neural recordings (Fig. 5c). Thus, H-BIND can be individually optimized using only hemodynamic recordings, and offers the advantage of not requiring severe invasive NTS recordings and minimizing the aforementioned recording quality issues, neural trajectory distortions, and low prediction accuracy.

[0162] The present invention evaluated the performance of predicting stimulus-based hemodynamics by comparing the prediction accuracy of H-BIND with that of decoding measured NTS activity (Fig. 5b). Individually optimized H-BINDs accurately predicted hemodynamic perturbations in two different rats, in contrast to the relatively low accuracy observed in the presented neurodecoding. The reduced performance of neurodecoding may be due to poor recording characteristics (see "Individual consistency of population dynamics in latent space" in the results section).

[0163] In addition, H-BIND predicted neural trajectories reconstructed from hemodynamic records using the dual-directional nature of the neuro-hemodynamic coupling more accurately than neurodecoding (Fig. 5c). These results indicate that H-BIND is superior not only in performance but also in the stability of model-based predictions compared to predictions based on the direct neurodecoding approach.

[0164] Next, we propose a closed-loop system design for precisely controlling hemodynamic outcomes during visceral sensory nerve stimulation using a trained H-BIND, and provide in silico simulations for proof of concept. For the control strategy, we utilized Nonlinear Model Predictive Control (NMPC), a method that optimizes stimulation by predicting stimulus-based effects at several future steps based on an accurate nonlinear prediction model. The controller design for NMPC is configured to receive the state of the 2D latent space of the H-BIND RNN as input and output and provide feedback on the stimulation. NMPC requires an objective function (cost function) to quantify control performance in order to optimize stimulation feedback. In this invention, the 2D Euclidean distance between the target and the current state of the latent space was selected as the cost matrix. This approach, which uses the latent space to monitor stimulus-centered responses, significantly reduces the computational burden and bypasses the need to consider the entire unit of the H-BIND RNN. The closed-loop system leverages the ability of the NTS neural computation mechanism to design a computationally efficient control system by simplifying a complex stimulus-centered system into manageable dynamics within a 2D latent space.

[0165] To accurately identify the appropriate target within the potential space for the controller, the topological portrait of the potential space was analyzed, and a stable node (not a resting point) during equilibrium was selected as the target.

[0166] The developed control system successfully induced a target state within the latent space, enabling precise control of the stimulus-centered response of BP by utilizing the bidirectionality in neuro-hemodynamic coupling. In contrast, uncontrolled open-loop stimulation could not induce a target state or δ response. Statistical analysis confirms the control performance of the developed system.

[0167] FIG. 6 is a block diagram of a system for predicting blood flow information, such as blood pressure, from NTS neural information according to one embodiment of the present invention.

[0168] Referring to FIG. 6, a system according to one embodiment of the present invention includes an encoding unit (100), a stimulus calculation unit (200); and a decoding unit (300).

[0169] In one embodiment of the present invention, the encoding unit obtains hemodynamic data (BP) acquired at time T. T It performs the function of converting ) into a low-dimensional latent space state of NTS. That is, the encoding unit (100) converts hemodynamic data (BP measurements) into a vector form representing the latent state of an NTS neuron population. In this conversion process, the complex high-dimensional signal of BPT is compressed into a low-dimensional latent space to simplify the overall dynamic state of the neuron population and extract key features. Through this, the latent space state required for neuron computation It plays the role of generating and passing it to the next calculation unit.

[0170] The above calculation unit (200) uses an artificial neural network to determine the next potential state from the current potential space state ( Calculate ). At this time, a Recurrent Neural Network (RNN) is used to calculate the current latent space state and the latent state of the next time step based on stimulus input u(t) ( The above computational unit models the collective dynamics of the nucleus accumbens (NTS) neural circuit and is designed to predict the response of a neural population to a stimulus, performs the role of calculating the stimulus-centered hemodynamic response in real time, and transmits the predicted next state to the decoding unit.

[0171] The above decoding unit (300) calculates the next potential state ( ) again hemodynamic value (BP T+1)It performs the function of converting into and providing prediction results. At this time, the decoding unit uses the next latent state generated by the stimulus-based dynamics computation unit ( It reconstructs ) into the original hemodynamic value (BP). In this process, the decoding unit utilizes a predefined decoding vector to linearly combine each element of the latent space vector. Specifically, Each dimension value is multiplied by the weight of the decoding vector and summed to calculate the predicted ΔBP value. Through this, hemodynamic changes induced by stimulation are predicted, and it can be utilized for simulation of BP state changes and feedback control.

[0172] The present invention also provides a method for predicting hemodynamic values, such as blood pressure, through the digital twin described above.

[0173] FIG. 7 is a step diagram of a blood pressure prediction method for predicting neuro-hemorrhagic results based on a nucleus isolate (NTS) performed on a computationally capable processor according to one embodiment of the present invention.

[0174] Referring to FIG. 7, a prediction method according to one embodiment of the present invention first obtains a hemodynamic value (BP) at time T. T ) normalized NTS latent space( An encoding step that converts the latent space state converted in the encoding step into the next latent state ( A calculation step for calculating ); the next potential state calculated in the calculation step, an updated hemodynamic value (BP). T+1 It includes a decoding step that converts ) back.

[0175] As described above, the present invention provides a digital twin system based on the population dynamics of NTS neuronal groups related to baroreflex that are involved in hemodynamic perturbations among visceral sensory stimuli, and as a result, it is possible to predict hemodynamic responses based on visceral sensory stimuli.

Claims

1. A digital twin system for predicting neuro-hemorrhagic results based on the nucleus isolate (NTS) performed on a computationally capable processor, Hemodynamic values ​​(BP) obtained at time T T ) normalized NTS latent space( Encoding unit that converts to the state of ); The latent space state converted by the above encoding unit is used to create the next latent state ( A calculation unit that calculates ); and The next potential state calculated by the above stimulus-based dynamics calculation unit is the updated hemodynamic value (BP). T+1 A digital twin system characterized by including a decoding unit that converts ) 2. In Paragraph 1, A digital twin system characterized by the above-mentioned computational unit using neural circuit modeling utilizing a recurrent neural network (RNN).

3. In Paragraph 2, A digital twin system characterized by the above-mentioned recurrent neural network (RNN) successfully modeling the collective dynamics of the above-mentioned NTS neural circuit to replicate measured single neural activities and neural trajectories.

4. In Paragraph 1, A digital twin system characterized by the above encoding unit converting hemodynamic values ​​(BP measurements) into a vector form representing the potential state of an NTS neuron population.

5. In Paragraph 1, A digital twin system characterized by the above-described decoding unit linearly combining each element of a latent space vector using a predefined decoding vector.

6. In Paragraph 5, The above decoding unit A digital twin system characterized by multiplying each dimension value by the weight of a decoding vector and summing them all to calculate a Δ-predicted value.

7. In Paragraph 1, A digital twin system characterized in that the above hemodynamic value is blood pressure.

8. A blood pressure prediction method for predicting neuro-hemorrhagic results based on a nucleus isolate (NTS) performed on a computationally capable processor, Hemodynamic values ​​(BP) obtained at time T T ) normalized NTS latent space( Encoding step that converts to the state of ); The latent space state transformed in the above encoding step is used with an artificial neural network to create the next latent state Calculation step for calculating; The next potential state calculated in the above calculation step is the updated hemodynamic value (BP). T+1 A blood pressure prediction method characterized by including a decoding step that converts ) 9. In Paragraph 8, A blood pressure prediction method characterized by using neural circuit modeling utilizing a recurrent neural network (RNN) in the above calculation step.

10. In Paragraph 9, A blood pressure prediction method characterized by the above-mentioned recurrent neural network (RNN) successfully modeling the collective dynamics of the above-mentioned NTS neural circuit to replicate the measured single neural activity and neural trajectory.

11. In Paragraph 8, A blood pressure prediction method characterized by the above encoding step converting hemodynamic values ​​(BP measurements) into a vector form representing the potential state of an NTS neuron population.

12. In Paragraph 8, A blood pressure prediction method characterized by the above-mentioned decoding step of linearly combining each element of a latent space vector using a predefined decoding vector.

13. In Paragraph 12, The above decoding unit A blood pressure prediction method characterized by multiplying each dimension value by the weight of a decoding vector and summing them all to calculate a Δ-predicted value.

14. In Paragraph 8, A blood pressure prediction method characterized in that the above hemodynamic value is blood pressure.

Images