Spinal cord electrical stimulation parameter determination system and method based on multi-modal data fusion

Abstract

The application discloses a spinal cord electrical stimulation parameter determination system and method based on multi-modal data fusion, and the system comprises a storage module, which stores a standard digital human model library; a pre-training processing module takes electrical stimulation parameters to be optimized as input, carries out large-scale simulation based on reinforcement learning in the standard digital human model library, establishes a state-action-response mapping space by simulating limb movement states, and obtains a general basic stimulation coding model; a personalized modeling module constructs a personalized digital twin model through transfer learning and fine-tuning in a personalized adaptation stage; and a strategy generation module migrates the general basic stimulation coding model to the personalized digital twin model, carries out iterative fine-tuning, and generates an initial stimulation strategy. The application avoids starting from zero to explore, and makes the personalized adaptation process on a real user faster and safer.

Description

Technical Field

[0001] This invention belongs to the field of medical robot technology, specifically involving the interdisciplinary technology of brain-computer interface (BCI), spinal cord stimulation (SCS) and artificial intelligence algorithms, and in particular, a system and method for determining spinal cord stimulation parameters based on multimodal data fusion. Background Technology

[0002] Spinal cord stimulation (SCS) is an important method for restoring motor function after spinal cord injury. Currently, existing spinal cord stimulation techniques are mainly divided into the following three categories: Open-loop spinal cord stimulation systems: These systems use fixed-parameter stimulation and cannot be adjusted according to the user's real-time physiological state. Their stimulation parameters are entirely set based on experience, resulting in limited functional improvement during stimulation and difficulty adapting to different activity states such as standing and walking.

[0003] A closed-loop stimulation system based on predefined rules detects motion states using sensors and switches stimulation modes according to preset rules. While it possesses a certain degree of state adaptability, it lacks the ability to continuously optimize the stimulation effect and cannot achieve personalized adjustments.

[0004] Closed-loop systems based on simple feedback adjust stimulus intensity by recording evoked compound action potentials (ECAPs) as feedback signals. These systems are primarily used to maintain stable activation levels and are less effective at optimizing complex motor function outputs.

[0005] Although cutting-edge technologies such as brain-computer interfaces have achieved initial closed-loop, the following significant bottlenecks still exist: the localization of stimulation targets depends on general anatomical atlases: the electrode implantation location and target selection are mainly based on the average anatomical structure of the population, failing to fully consider the specific differences of individuals in spinal cord geometry, injury sites and residual neural pathways.

[0006] Lack of personalized adaptive capabilities: The system cannot automatically adjust according to the user's training progress and dynamic changes in neural function, resulting in a decrease in stimulation effect over time. Summary of the Invention

[0007] The technical objective of this invention is to address the technical problems of current spinal cord stimulation protocols, such as reliance on general anatomical atlases for target localization and a lack of personalized adaptive capabilities, by providing a system and method for determining spinal cord stimulation parameters based on multimodal data fusion.

[0008] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution.

[0009] In a first aspect, embodiments of this application provide a spinal cord electrical stimulation parameter determination system based on multimodal data fusion, comprising: a storage module for storing a standard digital human model library and a knowledge base, wherein the standard digital human model library integrates a parameterized spinal cord geometric model, a biophysical model, and a neuromuscular skeletal model; and the knowledge base stores anatomical rules, neural activation criteria, and prior knowledge of dynamics for spinal cord electrical stimulation. The pre-training processing module is used to take the electrical stimulation parameters to be optimized as input, call the anatomical rules and neural activation criteria, and perform reinforcement learning-based simulation in the standard digital human model library. By simulating limb movement states, a state-action-response mapping space is established to obtain a general basic stimulus encoding model. The personalized modeling module is used to acquire the user's medical images and electrophysiological data, perform geometric registration of the spinal cord geometric model in the standard digital human model library using the anatomical rules, and assign conductivity parameters to the biophysical model to construct a personalized digital twin model. The strategy generation module is used to transfer the general basic stimulus encoding model to the personalized digital twin model, and to perform iterative fine-tuning by calling the biophysical model and the neuromuscular skeletal model through simulation to generate initial stimulus parameters for the target motor function. The initial stimulus parameters include initial stimulus target points and initial spatiotemporal stimulus codes.

[0010] Secondly, embodiments of this application provide a method for determining spinal cord electrical stimulation parameters based on multimodal data fusion, including: storing a standard digital human model library and a knowledge base, wherein the standard digital human model library integrates a parameterized spinal cord geometric model, a biophysical model, and a neuromuscular skeletal model; and the knowledge base stores anatomical rules, neural activation criteria, and prior knowledge of dynamics for spinal cord electrical stimulation. Using the electrical stimulation parameters to be optimized as input, the anatomical rules and neural activation criteria are invoked, and a reinforcement learning-based simulation is performed in the standard digital human model library. By simulating limb movement states, a state-action-response mapping space is established to obtain a general basic stimulation encoding model. The user's medical imaging and electrophysiological data are acquired, and the spinal cord geometric model in the standard digital human model library is geometrically registered using the anatomical rules. The conductivity parameter of the biophysical model is assigned to the model to construct a personalized digital twin model. The general basic stimulus encoding model is transferred to the personalized digital twin model. The biophysical model and neuromuscular skeletal model are iteratively fine-tuned through simulation to generate initial stimulus parameters for the target motor function. The initial stimulus parameters include initial stimulus target points and initial spatiotemporal stimulus encoding.

[0011] It should be understood that the summary section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description.

[0012] Compared with the prior art, the beneficial technical effects achieved by the embodiments of this application are as follows: through the pre-learning of the standard digital human model, the system obtains rich prior knowledge about neural regulation, avoids exploring from scratch, and makes the personalized adaptation process on real users faster and safer, significantly shortening the training cycle.

[0013] Personalized digital twin models built based on user-specific data enable precise customization of stimulation targets and parameters, overcoming the problem of poor stimulation effects caused by anatomical differences.

[0014] Non-contact, objective assessment: Motion assessment using video skeletal animation is a non-contact, objective, and information-rich assessment method that replaces subjective clinical scales and bulky physical sensors, improving the continuity and precision of the assessment.

[0015] High safety: A multi-layered safety monitoring mechanism ensures the safety of the training process, laying the foundation for large-scale clinical application. Attached Figure Description

[0016] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of this application in any way. Furthermore, the shapes and scales of the components in the drawings are merely illustrative to aid in understanding this application and do not specifically limit the shapes and scales of the components. Those skilled in the art, guided by the teachings of this application, can select various possible shapes and scales to implement this application according to specific circumstances. In the drawings: Figure 1 A schematic diagram of the system framework for determining spinal cord electrical stimulation parameters based on multimodal data fusion provided in this embodiment; Figure 2 A schematic flowchart of a method for determining spinal cord electrical stimulation parameters based on multimodal data fusion, provided for an embodiment; Figure 3 This is a schematic diagram of the pre-training process in the embodiment; Figure 4 This is a schematic diagram illustrating the personalized digital twin model construction process in the embodiment; Figure 5 This is a schematic diagram of the real-time optimization stage in the embodiment. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0018] It should be fully understood that the EEG signals, medical images, and electrophysiological data of users involved in this application are all information and data authorized by the users or fully authorized by all parties. The use of user information should comply with the privacy policies and practices of the industry that are generally considered to meet or exceed the requirements for protecting user privacy. The collection, use, and processing of related data should comply with relevant laws, regulations, and standards, and provide corresponding operation access points for users to choose to authorize or refuse.

[0019] Current technologies cannot address the individual differences in stimulation effects caused by variations in spinal cord anatomy, injury site specificity, and neural pathway remodeling. Universal stimulation targets are imprecise and cannot meet personalized needs.

[0020] This application aims to overcome the above shortcomings and provide a system and method for determining spinal cord electrical stimulation parameters based on multimodal data fusion. By pre-learning through a standard digital human model, a personalized spinal cord digital twin model can be quickly established for the user. Furthermore, by utilizing a computer vision-based reinforcement learning algorithm, stimulation parameters can be adjusted in real time to minimize the difference between the user's actual movement and the target movement, thereby achieving a truly personalized and adaptive spinal cord stimulation coding control method.

[0021] Example 1: A system for determining spinal cord electrical stimulation parameters based on multimodal data fusion, such as... Figure 1 As shown, it includes a storage module, a pre-training processing module, a personalized modeling module, and a policy generation module.

[0022] The storage module is used to store the standard digital human model library and the knowledge base. The standard digital human model library integrates a parametric spinal cord geometric model, a biophysical model, and a neuromuscular skeletal model. The knowledge base stores anatomical rules, neural activation criteria, and prior knowledge of dynamics for spinal cord electrical stimulation.

[0023] The pre-training processing module takes the electrical stimulation parameters to be optimized as input, calls anatomical rules and neural activation criteria, and performs reinforcement learning-based simulations in a standard digital human model library. By simulating limb movement states, it establishes a state-action-response mapping space and obtains a general basic stimulus encoding model.

[0024] The personalized modeling module is used to acquire the user's medical images and electrophysiological data, perform geometric registration of the spinal cord geometric model in the standard digital human model library using anatomical rules, and assign conductivity parameters to the biophysical model to construct a personalized digital twin model.

[0025] The strategy generation module is used to migrate the general basic stimulus coding model to the personalized digital twin model. It uses simulation to call the biophysical model and neuromuscular skeletal model for iterative fine-tuning to generate initial stimulus parameters for the target motor function. The initial stimulus parameters include the initial stimulus target and the initial spatiotemporal stimulus code.

[0026] The current system lacks a quantitative understanding of the relationship between stimulus effects and motor function output, and cannot automatically adjust parameters to optimize motor performance. The stimulus parameters are mismatched with motor function, requiring users to manually adjust parameters multiple times, which is inefficient. As training progresses, the user's neural pathways change, but the existing system cannot automatically adapt to these changes, leading to a decline in effectiveness later on, thus lacking long-term adaptive capabilities.

[0027] In some embodiments, the system further includes a real-time optimization module, which includes a state awareness module, a neural acquisition module, a data fusion module, an evaluation and control module, and a parameter fine-tuning module.

[0028] The state awareness module is used to acquire real-time motion images of users during the real-time optimization phase, and extract kinematic state features that represent the real-time position and posture of limbs based on the skeletal animation data generated from the real-time motion images.

[0029] The neural acquisition module is used to collect users' neural state data in real time.

[0030] The data fusion module is used to fuse kinematic state features with neural state data to construct a multimodal state vector that represents the user's current physiological and motor performance.

[0031] The evaluation and control module is used to evaluate the current motion quality in real time by taking the initial stimulus parameters as the iterative benchmark, the multimodal state vector as the input variable, the fine-tuned general basic stimulus encoding model for reinforcement learning closed-loop control, and the dynamic prior knowledge.

[0032] The parameter fine-tuning module is used to dynamically adjust the initial spatiotemporal stimulus code online based on the motion quality assessment results, so as to obtain optimized personalized stimulus parameters.

[0033] In some embodiments, the parameter fine-tuning module obtains optimized personalized stimulus parameters through the action output unit, reward calculation unit, and policy update unit: The action output unit is used to input the multimodal state vector into the fine-tuned general basic stimulus coding model and output the adjustment action for the initial spatiotemporal stimulus coding; the adjustment action includes at least one of the following: adjustment amount of stimulation amplitude, adjustment amount of stimulation frequency and adjustment amount of stimulation pulse width for each electrode channel.

[0034] The reward calculation unit is used to preset a comprehensive reward function that includes gait similarity, motion stability, energy efficiency and safety constraints, and, combined with prior knowledge of dynamics, calculates the reward value of the current gait based on real-time motion feedback after performing the adjustment action.

[0035] The strategy update unit is used to update the parameters of the general basic stimulus coding model in real time based on the reward value.

[0036] The embodiment achieves continuous adaptive optimization through a real-time optimization module. The introduction of reinforcement learning algorithms enables the system to adjust its strategy 24 / 7 based on the user's real-time reactions, dynamically track and adapt to changes in neural function, and achieve long-term training effect optimization of "getting smarter the more you use it".

[0037] Based on the same inventive concept as the spinal cord electrical stimulation parameter determination system based on multimodal data fusion provided in the above embodiments, the present invention also provides a method for determining spinal cord electrical stimulation parameters based on multimodal data fusion.

[0038] Example 2: A method for determining spinal cord electrical stimulation parameters based on multimodal data fusion, comprising: The system stores a standard digital human model library and a knowledge base. The standard digital human model library integrates a parametric spinal cord geometric model, a biophysical model, and a neuromuscular skeletal model. The knowledge base stores anatomical rules, neural activation criteria, and prior knowledge of dynamics for spinal cord electrical stimulation. Using the electrical stimulation parameters to be optimized as input, anatomical rules and neural activation criteria are invoked to conduct reinforcement learning-based simulations in a standard digital human model library. By simulating limb movement states, a state-action-response mapping space is established to obtain a general basic stimulus coding model. Acquire users' medical imaging and electrophysiological data, use anatomical rules to perform geometric registration of spinal cord geometric models in the standard digital human model library, and assign conductivity parameters to biophysical models to construct personalized digital twin models; The general basic stimulus coding model is transferred to a personalized digital twin model. Through simulation, the biophysical model and neuromuscular skeletal model are called for iterative fine-tuning to generate initial stimulus parameters for the target motor function. The initial stimulus parameters include the initial stimulus target and the initial spatiotemporal stimulus coding.

[0039] Furthermore, the method also includes: in the real-time optimization phase, acquiring real-time motion images of the user, and extracting kinematic state features representing the real-time position and posture of the limbs based on the skeletal animation data generated from the real-time motion images; synchronously collecting the user's neural state data in real time; fusing the kinematic state features and neural state data to construct a multimodal state vector representing the user's current physiological and motor performance; using the initial spatiotemporal stimulus encoding as the iterative benchmark, and taking the multimodal state vector as the input variable, using the fine-tuned general basic stimulus encoding model for reinforcement learning closed-loop control, and combining prior knowledge of dynamics to evaluate the current motion quality in real time; and dynamically adjusting the initial spatiotemporal stimulus encoding online based on the motion quality evaluation results to obtain optimized personalized stimulus parameters.

[0040] Figure 2 The flowchart of the data-driven, continuously self-optimizing spinal cord stimulation parameter determination method provided in the embodiments is clearly outlined. Its greatest value lies in its closed-loop learning capability: the system no longer relies on fixed expert experience, but automatically adjusts stimulation parameters by comparing the difference between "expected actions" and "actual performance." This "evaluation-optimization-execution" cycle enables the system to continuously adapt to changes in the user's (or patient's) neurological function, providing a theoretical foundation and technical path for achieving long-term, precise, and adaptive training.

[0041] In this embodiment, an existing digital human database (such as a virtual human dataset) can be used to construct a standard digital human model integrating anatomy, biophysics, and kinesiology. Figure 3 As shown, building a standard digital human model library specifically includes: 1) A detailed parametric model of the spinal cord geometry: Using multiple spinal cord instances from a virtual human dataset, a parameterized spinal cord geometric model was established using statistical methods such as principal component analysis (PCA). This model not only represents the average shape but also generates a series of morphologically diverse virtual spinal cords within physiological limits by adjusting key pattern parameters (such as the weights of PCA components) to simulate anatomical differences in the population. This spinal cord geometric model will serve as the anatomical basis for electric field simulations.

[0042] The anatomical rules for spinal cord stimulation stored in the knowledge base include statistical feature parameters obtained through principal component analysis (PCA) of spinal cord instances in the digital human database, which are used to generate virtual spinal cords of different morphologies. The anatomical rules for spinal cord stimulation also include spatial topological relationships, that is, defining the standard anatomical correspondence between the electrode array and spinal cord segments (such as the lumbosacral segment) and nerve roots.

[0043] 2) Biophysical model: The biophysical model uses a finite element model to simulate the electric field distribution generated by SCS electrodes in spinal cord tissue, and predicts the changes in membrane potential and the onset of action potential directly from the second spatial derivative of the external electric field along the nerve fiber axis through a spatiotemporal activation function.

[0044] The core objective of biophysical models is to accurately predict how the electric field generated by SCS electrodes in the spinal cord of a specific individual will activate target neural structures.

[0045] In this embodiment, the construction of the biophysical model can be broken down into the following steps: a) Finite element model establishment and electric field distribution simulation, including: geometric reconstruction and network generation, material property assignment and electric field equation solving.

[0046] In this embodiment, based on standard digital human or user medical images (such as MRI), an image segmentation algorithm is used to reconstruct a three-dimensional geometric model containing tissues such as the spinal cord (gray matter and white matter), cerebrospinal fluid, dura mater, and vertebrae. Subsequently, high-precision, non-uniform mesh generation is performed using tetrahedral or hexahedral elements, with mesh refinement near the electrodes and in key areas of neural tissue.

[0047] In this embodiment, different biological tissues (such as gray matter, white matter, and cerebrospinal fluid) are assigned corresponding conductivity parameters. Among them, the anisotropic conductivity of white matter (conductivity along the nerve fiber direction is higher than that in the vertical direction) is crucial and can be estimated from the user's DTI image data, which can significantly improve the accuracy of electric field calculation.

[0048] In this embodiment, Maxwell's equations (usually simplified to Poisson's equations) under quasi-static conditions are solved in multiphysics simulation software. By setting the stimulation parameters (amplitude, pulse width) of the electrodes, the potential and electric field distribution throughout the computational domain are calculated.

[0049] b) Set the activation function along the nerve fiber axis, including: setting the activation function.

[0050] The theoretical basis for setting the activation function is that the initiation of action potentials depends primarily on the second-order spatial rate of change of the transmembrane potential on the axon membrane of nerve fibers, rather than the absolute value of the potential.

[0051] In spinal cord electrical stimulation, transient electrical pulses are processed, rather than constant stimuli. Traditional spatial activation functions (such as the Rattay formula) primarily deal with steady-state or instantaneous fields, while the spatiotemporal activation function used in this embodiment, by introducing a time dimension, can more accurately describe how the pulsed electric field induces neural activation. Its formula is as follows: ; Where: AF(x,t) is the spatiotemporal activation function value at position x and time t; V_e(x,t) is the extracellular potential at position x and time t, calculated by the finite element model; It is the second spatial derivative of the extracellular potential along the nerve fiber axis, reflecting the spatial non-uniformity of the electric field, and is the main driving force leading to local depolarization of the fiber membrane; It is the first derivative of the extracellular potential with respect to time, reflecting the rate of change of the electric field over time; k is a proportionality constant (dimensioned in time) related to the properties of the nerve membrane, which determines the relative importance of the time term. It can be derived from the membrane time constant (e.g., k = R_m × C_m, where R_m is the membrane resistance and C_m is the membrane capacitance).

[0052] In this embodiment, the process for determining whether a fiber is activated includes: Finite element method calculation: Calculate the extracellular potential field V_e(x,t) as a function of time under pulse stimulation in simulation software.

[0053] Path extraction: Sample V_e(x,t) along the virtual neural fiber path.

[0054] Numerical calculation: Numerical calculation using the central difference method and And by combining the membrane time constant k, AF(x,t) is obtained.

[0055] Threshold determination: An empirical activation threshold AF_threshold is set to determine whether a nerve fiber generates an action potential. When the peak value of AF(x,t) exceeds this threshold, the fiber is considered activated.

[0056] In this embodiment of the invention, the core significance of determining whether a nerve fiber is activated lies in establishing a digital mapping relationship between electrical parameters and biological effects. Specifically, the system uses a biophysical model to calculate the spatial distribution of the electric field within the spinal cord and, in conjunction with neural activation criteria, determines the proportion of target nerve bundles (such as dorsal root fibers and motor neuron pools) that generate action potentials under specific current intensities and pulse widths, i.e., the effective activation rate.

[0057] In this embodiment, a biophysical model is used to simulate the electric field distribution within the spinal cord and to calculate the effective activation rate of the target nerve fiber under specific stimulation parameters. A state-action-response mapping space is constructed based on the effective activation rate to constrain the search range of the general basic stimulus encoding model.

[0058] In some embodiments, the scheme for verifying whether the fiber is activated includes: sampling and verifying the prediction results of the activation function method through a high-fidelity MRG cable equation model to calibrate and confirm its accuracy, forming a dual guarantee mechanism of "efficient prediction - accurate verification".

[0059] The neural activation criteria stored in the knowledge base may include parameters related to the spatiotemporal activation function, activation threshold, and conductance prior values.

[0060] 3) Neuromuscular-skeletal model: The core value of the neuromuscular-skeletal model lies in establishing an end-to-end computable physiological dynamic chain from microscopic electrical stimulation to macroscopic behavioral output. It is not just a simple input-output mapping, but rather a dynamic neural state space constructed by simulating the encoding, transmission, and integration of stimulus signals within the nervous system, thereby providing a physiologically meaningful and interpretable modeling foundation for reinforcement learning.

[0061] In this embodiment, the neuromuscular skeletal model consists of two parts: Motor neuron pools and activation dynamics: neural encoding and spatial integration of stimuli; Neuromuscular transmission and muscle activation dynamics: from nerve signals to muscle force commands.

[0062] Motor neuron pooling and activation dynamics aim to establish a complete computational process that transforms the stimulation parameters (input) of the SCS into the dynamic activation states (output) of the motor neuron pool, i.e., the neural state space. This space is a key bridge connecting stimuli and behavior.

[0063] As an example, constructing a neuromuscular skeletal model includes the following: (1) Spatiotemporal encoding of input stimuli.

[0064] The stimulation pattern of SCS is a typical spatiotemporal code, which needs to be accurately described from both spatial and temporal dimensions.

[0065] Spatial coding is determined by the geometric configuration and activation state of the multi-electrode array implanted in the epidural space of the spinal cord, including: i. Electrode Array Topology: The three-dimensional coordinates of each electrode contact on the array constitute the spatial base of the stimulation. The contact spacing and arrangement (longitudinal, transverse, or encircling) determine the precision of the electric field distribution in space.

[0066] ii. Activation Pattern: At any given time, each contact is defined as cathode, anode, or neutral. This pattern constitutes a spatial stimulus vector S_patial. For example, the activation pattern of an 8-contact array can be represented as [-1,0,0,+1,+1,0,0,-1], where -1 is cathode, +1 is anode, and 0 is neutral. Different patterns create different electric field distributions, thereby targeting nerve fibers at different locations.

[0067] The timing encoding is determined by the characteristics of the electrical pulse sequence applied to each active electrode, including: i. Pulse amplitude: The intensity of the stimulating current or voltage, which is the main parameter controlling the number of activated fibers. The greater the amplitude, the more motor units are recruited.

[0068] ii. Pulse width: The duration of a single pulse, affecting the activation threshold and selectivity. Within a certain range, increasing the pulse width can lower the activation threshold and may affect the excitability of fibers of different diameters.

[0069] iii. Pulse frequency: The number of pulses per unit time, a key parameter controlling the force and pattern of muscle contraction. Low frequencies cause muscle spasms, while high frequencies cause tetanic contractions.

[0070] iv. Temporal pattern: For multi-touch arrays, complex spatiotemporal stimulation patterns can be designed, such as each touch point being activated sequentially in a specific order and with a delay, to simulate natural neural firing sequences, thereby inducing more coordinated movements.

[0071] The integrated spatiotemporal coding can be expressed as: Stimulus(t)=f(S_patial,A(t),PW(t),Freq(t)), where A(t), PW(t), and Freq(t) are time functions of pulse amplitude, pulse width, and pulse frequency, respectively.

[0072] (2) Construction of neural state space and solution of stimulus response: i. Using the predefined spatiotemporal stimulation pattern Stimulus(t) and the personalized spinal cord finite element model as input, solve the quasi-static electromagnetic field equations to calculate the time-varying extracellular potential field V_e(x,y,z,t) in the entire computational domain, where x, y, and z are spatial coordinates.

[0073] ii. Sample V_e(x',t) along the path of each fiber, where x' is the direction vector of the sampling point along the fiber path. Then, calculate its spatial second derivative and temporal first derivative using numerical methods. For the calculated AF(x,t), set an activation threshold AF_threshold. For a given fiber, if at a certain position x and time t, AF(x,t) >= AF_threshold, then the fiber is considered successfully activated at that time and place, generating an action potential.

[0074] (3) From fiber activation to neural state space: dynamic response of motor neuron pool.

[0075] The activation of a single filament is a point event, while the neural state space needs to represent the collective activity of a population of neurons. Constructing the neural state vector: i. Dimensional Definition: The neural state space S_neural(t) is a multidimensional vector. Each dimension represents the activation level of a functional unit at time t. Functional units can be divided according to spinal cord segments (such as L2, L3, L4 cisterns) and / or target muscles (such as quadriceps cisterns, tibialis anterior cisterns).

[0076] ii. State quantization: S_neural(t) = [FR_L2(t), FR_L3(t), FR_L4(t), ...]; Among them, FR_L2(t) is the proportion of the number of L2 functional units that are activated (i.e. generate discharge potential) to the total number of fibers, and so on.

[0077] iii. Dynamic solution of stimulus-response – differential equation modeling: τ×d(S_neural(t)) / dt+S_neural(t)=F(Stimulus(t)); This differential equation describes the dynamic evolution of the state space of S_neural(t). Here, F(Stimulus(t)) is the driving function. It encapsulates all the aforementioned steps (electric field calculation, activation function, threshold determination), and its output is the theoretical neural activation level instantaneously driven by the stimulus at time t; τ is a time constant, simulating the inertia brought about by physiological processes such as synaptic transmission and membrane capacitance, and determining the system's response speed.

[0078] Solution: In simulation, this equation can be solved numerically using methods such as the Euler method. The result is the dynamic response of the neural state space to stimuli. It clearly demonstrates how a stimulus, through an inertial system, ultimately translates into the activation pattern of motor neurons.

[0079] The prior knowledge of dynamics stored in the knowledge base may include the time constant τ, activation dynamic parameters, and ideal gait templates.

[0080] During the pre-training phase, a state-action-response mapping space is constructed and reinforcement learning pre-training is performed.

[0081] In the above simulation environment, the mapping relationship between stimuli and motion is defined as a reinforcement learning problem, including state, action, reward and weight policy.

[0082] The motion state vector S(t) should be an integration of multimodal information, including: Kinematic state: includes the angles θ(t) and angular velocities ω(t) of the major joints (hip, knee, ankle) of the lower limb at the current time t and the past N time steps (e.g., t-1, t-2), which helps the agent perceive motion trends and phases.

[0083] Dynamic states: 1) Key muscle activation levels: Standardized activation levels a_i(t) of major flexor and extensor muscle groups (such as quadriceps, hamstrings, and tibialis anterior) are extracted from the neuromuscular model. This directly reflects the physiological effects of the stimulus and is a key intermediate variable connecting stimulus and movement. 2) Ground reaction force: In gait simulation, the contact force GRF(t) between the foot and the ground is simulated to provide information on balance and support states.

[0084] Neural state: The total firing rate FR(t) of the spinal motor neuron pools corresponding to each target muscle is extracted. This is the physiological state that most closely resembles the stimulus action and can greatly shorten the learning time of the agent.

[0085] Task target states: 1) Ideal gait template difference: This includes the difference Δθ(t) between the target angle θ_target(t) and the actual angle θ(t) of each joint at the current moment, and the difference Δω(t) between the target angular velocity ω_target(t) and the actual angular velocity ω(t). This provides the agent with a direct error signal. 2) Gait cycle phase: A scalar phase(t) indicates which stage of the gait cycle the agent is in at the current moment t (e.g., 0% heel strike, 60% toe lift). This helps the agent generate phase-related stimulus parameters.

[0086] The final state vector can be represented as: S(t)=[θ(t),θ(t-1),...,ω(t),ω(t-1),...,a_i(t),GRF(t),FR(t),Δθ(t),Δω(t),phase(t)] action: A(t)=[ΔAmp_1,ΔAmp_2,...,ΔAmp_M,ΔFreq,ΔPW] Where: ΔAmp_m (m=1,…,M): M is the total number of electrode channels, and the adjustment amount of the stimulation amplitude of the m-th electrode channel (e.g., ±0.1mA). ΔFreq: Adjustment amount for stimulation frequency (e.g., ±5Hz); ΔPW: Adjustment amount for stimulation pulse width (e.g., ±10μs). All motion outputs are clipped according to the set safety threshold, and the motions are low-pass filtered in time to avoid drastic changes in stimulation parameters that could cause user discomfort or risk.

[0087] The response is evaluated by a comprehensive reward function R(t) that includes gait similarity rewards and safety constraint rewards. As an example, the comprehensive reward function is as follows: R(t)=w_gait×R_gait+w_stab×R_stab+w_eff×R_eff+w_safe×R_safe+w_energy×R_energy; The comprehensive reward consists of the following components: i. Gait similarity reward: R_gait(t) = -DTW(J_trajectory_actual, J_trajectory_target). The difference between the actual joint trajectory and the target trajectory is calculated using a dynamic time warping algorithm. This algorithm effectively handles small scaling and offsets on the time axis and is more robust than simple mean square error. w_gait is the weight of the gait similarity reward.

[0088] ii. Motion stability reward: R_stab(t) = -var(trunk_tilt) - |L_GRF - R_GRF|. This penalizes the variance of the trunk tilt angle (encouraging upright posture) and the difference in ground reaction forces between the left and right feet (encouraging symmetrical support). w_stab is the weight of the motion stability reward.

[0089] iii. Energy Efficiency Reward: R_eff(t) = -Σ(a_i(t)^2). This penalizes the sum of squares of activation across all monitored muscles. Based on biological principles, muscle strength and activation level have a non-linear relationship; the sum of squares effectively encourages movements to be completed at lower activation levels, promoting efficient movement patterns. w_eff is the weight of the energy efficiency reward.

[0090] iv. Safety Constraint Reward: R_safe(t) = -Σ(I(joint_angle>limit)). This is a sparse reward. When any joint angle exceeds the physiological limit, a very large negative reward (e.g., -10) is given, and the current training round is terminated early, forcing the agent to learn the safety boundary. w_safe is the weight of the safety constraint reward.

[0091] v. Stimulus Energy Reward: R_energy(t) = -(Σ(Amp_m^2) × Freq × PW). This penalizes the total energy expenditure of the stimulus, encouraging the use of minimal energy to achieve the target effect, which helps extend the battery life of the implanted pulse generator. w_energy is the stimulus energy reward weight, Freq is the stimulus frequency, and PW is the stimulus pulse width.

[0092] Weighting strategy: Using a course-based learning approach, initially (w_safe, w_gait) has a high weight to ensure safety and basic functionality; later, the weight of (w_eff, w_energy) is gradually increased to optimize exercise quality and efficiency.

[0093] As an example, the course includes the following stages: Phase 1: Basic Activation. Task: Induce simple movements of a single joint (e.g., ankle dorsiflexion). The reward function focuses on R_gait and R_safe.

[0094] Phase Two: Static Posture. Task: Maintain a standing posture with support. The reward function focuses on R_stab.

[0095] Phase 3: Simple Gait. Task: Walk on flat ground with treadmill assistance. Reward function fully activated.

[0096] Phase Four: Complex Challenges. Task: To improve the robustness and adaptability of stimulus parameters in response to disturbances such as varying walking speeds and inclines / declines.

[0097] The workflow of this phase yields the following deliverables: a high-performance, transferable, general basic stimulus coding model π(A|S); the general basic stimulus coding model can include information such as physiological structure, target, parameters, and expected kinematic output; and a validated, multi-scale standard digital human simulation environment, which can serve as a platform for subsequent research and development.

[0098] Figure 3 The flowchart shown is for the standard digital human pre-learning stage. Figure 3 This invention reveals a key step in some embodiments that breaks through the traditional trial-and-error model. Its core value lies in establishing a safe and efficient "virtual laboratory." Through millions of unrestricted explorations in a simulation environment, the reinforcement learning agent can discover complex stimulus patterns that are difficult for human experts to intuitively design, thereby obtaining a high-quality initial model rich in "knowledge." This lays a high starting point for subsequent personalized adaptation, significantly shortens the debugging time on real users, and reduces risks.

[0099] The second stage, the personalized adaptation stage, involves personalized modeling and initialization. First, data acquisition is performed, obtaining high-resolution MRI and diffusion tensor imaging (DTI) data of the user's cervical and thoracic segments to clearly display the individual anatomical structure of the spinal cord and the pathways of neural pathways such as the corticospinal tract.

[0100] Figure 4 A schematic diagram for building a personalized digital twin model. Figure 4 This paper elucidates the core steps to achieving precision medicine. Its value lies in shifting training from a "group average" to an "individualized" approach. By integrating a user's real anatomical and physiological data, the system can predict the precise distribution and activation effect of SCS current within that user's body, thereby achieving "tailor-made" target localization and parameter initialization. This significantly improves the accuracy, safety, and effectiveness of individualized training.

[0101] To align the standard digital human model library with the user's personalization, this process is not a simple data replacement, but a multi-layered model transfer learning and optimization process based on physical and physiological constraints, including: 1) Anatomical alignment: Geometric personalization based on non-rigid registration.

[0102] In this embodiment, a high-precision non-rigid registration algorithm (such as B-spline or Demons algorithm) is used to match the target population's general basic stimulation coding model of the spinal cord with the individual spinal cord geometry obtained by user segmentation in the first stage. The physiological structure data in the general basic stimulation coding model is used as a template to register and align the user (or patient) image reconstruction data with the standard digital human physiological structure.

[0103] As an example, non-rigid registration methods include: A complex spatial transformation field is calculated that can "push" or "pull" each point of the standard model to its corresponding position in the user's anatomical space.

[0104] Through this transformation, statistical prior knowledge (such as common shape variation patterns) from the standard digital human model library is effectively transferred to the user's personalized digital twin model. Simultaneously, predefined virtual neural fiber bundles and electrode locations from the standard digital human model library are also mapped to the user's individual space through the same transformation, completing preliminary target mapping.

[0105] Ultimately, a personalized spinal cord geometric model is obtained that is highly consistent with the user's actual anatomical structure in terms of geometric morphology, namely a personalized digital twin model.

[0106] 2) Organizational characteristics alignment: Personalized assignment of conductivity parameters.

[0107] By utilizing the user's DTI data, the conductivity parameters in the finite element model are assigned personalized values.

[0108] In some embodiments, the specific method includes: The diffusion tensor provided by DTI data aligns with the direction of nerve fibers. Based on specific physical models (such as volume conduction theory), the conductivity of white matter in different directions (i.e., anisotropic conductivity) can be estimated from the diffusion tensor. These estimated personalized conductivity parameters are then assigned to the corresponding white matter regions in the finite element model. For isotropic tissues such as gray matter and cerebrospinal fluid, adjustments are made based on their MRI signal intensity or literature values. This significantly improves the accuracy of electric field distribution simulation, enabling the model to predict the true distribution of SCS current in the user's specific neural structure.

[0109] 3) Functional mapping alignment: rapid initialization of stimulus parameters.

[0110] The pre-trained general basic stimulus encoding model π(A|S) from the first stage is transferred to the newly constructed personalized digital twin model. The specific method is as follows: State-space recalibration: Ensure that the physical meaning of the state S (such as joint angle) input to the general basic stimulus encoding model is consistent with that of the standard model in the personalized model.

[0111] Simulation environment switching: The interactive environment of the reinforcement learning agent is switched from "standard digital human simulation" to "personalized user digital twin simulation".

[0112] Rapid fine-tuning: Since standard policy networks already contain rich stimulus encoding knowledge, training from scratch is not required in this embodiment. The agent only needs to be allowed to explore and fine-tune a few times in a personalized environment. The goal is to quickly find stimulus parameters that can induce near-target motion in the user, while ensuring safety. This process can be viewed as a form of meta-learning or rapid adaptation.

[0113] 4) Finally, a personalized optimal initial stimulus target and initial spatiotemporal stimulus code are obtained for the user.

[0114] Target optimization: A series of simulation experiments are run on a personalized spinal cord geometry model. The system evaluates the activation selectivity of different electrode combinations and locations for target neural pathways (such as fiber bundles innervating the lower limbs), as well as the risk of activating non-target structures (such as the dorsal root). Taking into account surgical access constraints (such as feasible electrode implantation trajectories), a multi-objective optimization algorithm determines the optimal initial stimulation target.

[0115] Parameter determination: Using the pre-initialized personalized stimulation parameters, a set of initial stimulation parameters (amplitude, frequency, pulse width) is generated in the simulation for the determined target point. These parameters can stably induce the desired movement (such as ankle dorsiflexion) and all physiological indicators (such as muscle strength and joint angle) are within the safe range.

[0116] The final outputs of this stage are: 1) A personalized digital twin model of a user, including accurate geometric and physical attributes and functional simulation capabilities; 2) A set of personalized initial stimulation parameters, which clearly defines the optimal initial stimulation target coordinates and a set of safe and effective initial stimulation parameters (i.e., initial spatiotemporal stimulation codes); 3) An initialized personalized stimulation parameter, which lays the foundation for subsequent true closed-loop adaptive optimization and realizes the "hot start" of the spinal cord stimulation process.

[0117] In some embodiments, electrode implantation and closed-loop adaptive optimization are achieved in the third stage (real-time optimization stage). In these embodiments, based on the stimulation target determined in the second stage, the SCS electrode array is surgically implanted into the corresponding segment of the user's spinal epidural space.

[0118] like Figure 5As shown, during the session, the user enters a training area equipped with multiple depth cameras. The system applies stimuli using initial parameters, while the cameras capture images of the user attempting to perform stepping movements.

[0119] Online optimization using reinforcement learning includes: State awareness: skeletal animation data generated in real time from motion videos is extracted into feature vectors (such as the angles and angular velocities of the hip, knee, and ankle joints during the gait cycle).

[0120] Decision: The reinforcement learning agent outputs fine-tuning instructions for the stimulation parameters of each channel of SCS based on the current motion state (such as increasing the amplitude of the 3rd channel by 0.1mA).

[0121] Reward calculation: The system calculates the similarity between the current gait and the ideal gait template in multiple indicators such as smoothness, symmetry, and joint range of motion, and feeds the result as a reward to the agent.

[0122] Stimulus parameter updates: The agent continuously updates its personalized stimulus parameters based on the rewards it receives. After multiple sessions, the system can learn a unique, time-adaptive optimal stimulus encoding scheme for the user.

[0123] Figure 5 This demonstrates the system's intelligent essence, moving from "static treatment" to "dynamic optimization." Its core value lies in endowing the system with the ability to continuously improve. The system is no longer a static tool configured only once, but an intelligent partner capable of "learning" and "growing" alongside the patient. It can fine-tune the treatment plan in real time based on the patient's different daily or even individual training states, ensuring that treatment remains on the optimal path. This is expected to overcome the "plateau" commonly found in traditional rehabilitation, achieving sustained maximization of rehabilitation effects.

[0124] This invention utilizes a standard digital human model containing large-scale population anatomical and kinematic data for pre-training in a simulation environment. This model learns the complex mapping relationships between different spinal cord segments, different stimulation parameters, and induced movements, forming a high-performance, universal basic stimulus coding model. Medical images (such as MRI / CT) and electrophysiological data of the user are acquired, and the basic standard digital human model is then fine-tuned to generate a personalized spinal cord digital twin model that highly matches the user's actual spinal cord anatomy and neural characteristics. This model is used to accurately predict stimulation targets and initial parameters.

[0125] In some embodiments, visual sensors such as depth cameras are used to capture motion image sequences (such as motion videos) of users in real time under stimulation without contact, and generate 3D skeletal animation data to accurately quantify kinematic indicators such as joint angles and speeds.

[0126] The reinforcement learning agent (decision module) takes skeletal animation data as state input, adjusts stimulus parameters (such as electrode selection, amplitude, and frequency) as actions, and reduces the difference between the actual movement and the ideal target movement as a reward signal. The agent learns the optimal stimulus parameters through continuous interaction with the environment (user).

[0127] In some embodiments, the method further includes setting up a safety closed-loop control interface to ensure that all stimulation parameters are within physiologically safe ranges and to have an abnormal state monitoring and emergency handling mechanism.

[0128] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, or a tablet computer, or any combination of these devices.

[0129] The above provides a detailed description of the spinal cord stimulation parameter determination system and method based on multimodal data fusion provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the concept of this application and should not be construed as limiting the scope of protection of this application.

Claims

1. A spinal cord electrical stimulation parameter determination system based on multimodal data fusion, characterized in that, include: The storage module is used to store a standard digital human model library and a knowledge base. The standard digital human model library integrates a parametric spinal cord geometric model, a biophysical model, and a neuromuscular skeletal model. The knowledge base stores anatomical rules, neural activation criteria, and prior knowledge of dynamics for spinal cord electrical stimulation. The pre-training processing module is used to take the electrical stimulation parameters to be optimized as input, call the anatomical rules and neural activation criteria, and perform reinforcement learning-based simulation in the standard digital human model library. By simulating limb movement states, a state-action-response mapping space is established to obtain a general basic stimulus encoding model. The personalized modeling module is used to acquire the user's medical images and electrophysiological data, perform geometric registration of the spinal cord geometric model in the standard digital human model library using the anatomical rules, and assign conductivity parameters to the biophysical model to construct a personalized digital twin model. The strategy generation module is used to transfer the general basic stimulus encoding model to the personalized digital twin model, and to perform iterative fine-tuning by calling the biophysical model and the neuromuscular skeletal model through simulation to generate initial stimulus parameters for the target motor function. The initial stimulus parameters include initial stimulus target points and initial spatiotemporal stimulus codes.

2. The spinal cord electrical stimulation parameter determination system according to claim 1, characterized in that, The system also includes a real-time optimization module, which includes: The state awareness module is used to acquire the user's real-time motion images during the real-time optimization phase, and extract kinematic state features that characterize the real-time position and posture of the limbs based on the skeletal animation data generated from the real-time motion images. The neural acquisition module is used to collect the user's neural state data in real time. The data fusion module is used to fuse the kinematic state features with the neural state data to construct a multimodal state vector representing the user's current physiological and motor performance. The evaluation and control module is used to evaluate the current motion quality in real time by taking the initial stimulus parameters as the iterative benchmark, the multimodal state vector as the input variable, the fine-tuned general basic stimulus coding model for reinforcement learning closed-loop control, and the dynamic prior knowledge. The parameter fine-tuning module is used to dynamically adjust the initial spatiotemporal stimulus code online based on the motion quality assessment results, so as to obtain optimized personalized stimulus parameters.

3. The spinal cord electrical stimulation parameter determination system according to claim 1, characterized in that, The personalized stimulation parameters include spatial encoding and temporal encoding: the spatial encoding is characterized by the spatial stimulation vector S_patial defined by the electrode array topology and contact activation mode; the temporal encoding is characterized by the pulse amplitude A(t), pulse width PW(t), and pulse frequency Freq(t) that vary with time.

4. The spinal cord electrical stimulation parameter determination system according to claim 1, characterized in that, The spinal cord geometric model includes: using spinal cord instances from a digital human database, and employing principal component analysis to establish a geometric model capable of generating virtual spinal cords of different morphologies by adjusting mode parameters.

5. The spinal cord electrical stimulation parameter determination system according to claim 1, characterized in that, The biophysical model predicts neural activation using a spatiotemporal activation function, the expression of which is: ； Where AF(x,t) is the spatiotemporal activation function value at position x and time t; V_e(x,t) is the extracellular potential at position x and time t, calculated by the finite element model; It is the second spatial derivative of the extracellular potential along the nerve fiber axis; It is the first derivative of the extracellular potential with respect to time; k is a proportionality constant related to the properties of the nerve membrane, with the dimension of time.

6. The spinal cord electrical stimulation parameter determination system according to claim 1, characterized in that, The neuromuscular skeletal model simulates the physiological dynamics chain from microscopic electrical stimulation to macroscopic behavior in the following way: The spatial and temporal encoding of stimuli is transformed into a spatial response of neural states that characterizes the collective activity of a neuronal population; the spatial response of neural states is then transformed into muscle activation levels and muscle force commands.

7. The spinal cord electrical stimulation parameter determination system according to claim 6, characterized in that, The neural state-space response is dynamically solved using the following differential equation: τ×d(S_neural(t)) / dt+S_neural(t)=F(Stimulus(t)); Where S_neural(t) is the neural state vector representing the activation ratio of different functional units, F(Stimulus(t)) is the stimulus driving function for electric field calculation and activation function determination, and its input is represented by both spatial and temporal encoding; t is time; τ is the time constant for simulating synaptic transmission and membrane capacitance physiological inertia.

8. The spinal cord electrical stimulation parameter determination system according to claim 2, characterized in that, The parameter fine-tuning module obtains optimized personalized stimulation parameters through the following units: An action output unit is used to input the multimodal state vector into the fine-tuned general basic stimulus coding model and output an adjustment action for the initial spatiotemporal stimulus coding; the adjustment action includes at least one of the following: adjustment of the stimulation amplitude, adjustment of the stimulation frequency, and adjustment of the stimulation pulse width for each electrode channel. The reward calculation unit is used to preset a comprehensive reward function that includes gait similarity, motion stability, energy efficiency and safety constraints, and, in combination with the aforementioned prior knowledge of dynamics, calculate the reward value of the current gait based on the real-time motion feedback after performing the adjustment action; The strategy update unit is used to update the parameters of the general basic stimulus encoding model in real time according to the reward value.

9. A method for determining spinal cord electrical stimulation parameters based on multimodal data fusion, characterized in that, include: The system stores a standard digital human model library and a knowledge base. The standard digital human model library integrates a parametric spinal cord geometric model, a biophysical model, and a neuromuscular skeletal model. The knowledge base stores anatomical rules, neural activation criteria, and prior knowledge of dynamics for spinal cord electrical stimulation. Using the electrical stimulation parameters to be optimized as input, the anatomical rules and neural activation criteria are invoked, and a reinforcement learning-based simulation is performed in the standard digital human model library. By simulating limb movement states, a state-action-response mapping space is established to obtain a general basic stimulation encoding model. The user's medical imaging and electrophysiological data are acquired, and the spinal cord geometric model in the standard digital human model library is geometrically registered using the anatomical rules. The conductivity parameter of the biophysical model is assigned to the model to construct a personalized digital twin model. The general basic stimulus encoding model is transferred to the personalized digital twin model. The biophysical model and neuromuscular skeletal model are iteratively fine-tuned through simulation to generate initial stimulus parameters for the target motor function. The initial stimulus parameters include initial stimulus target points and initial spatiotemporal stimulus encoding.

10. The method for determining spinal cord electrical stimulation parameters based on multimodal data fusion according to claim 9, characterized in that, The method further includes: During the real-time optimization phase, real-time motion images of the user are acquired, and kinematic state features representing the real-time position and posture of the limbs are extracted based on the skeletal animation data generated from the real-time motion images. Real-time synchronous collection of users' neural state data; The kinematic state features are fused with the neural state data to construct a multimodal state vector representing the user's current physiological and motor performance; Using the initial spatiotemporal stimulus encoding as the iterative benchmark, the multimodal state vector is used as the input variable. The fine-tuned general basic stimulus encoding model is used for reinforcement learning closed-loop control. Combined with the prior knowledge of dynamics, the current motion quality is evaluated in real time. Based on the results of the exercise quality assessment, the initial spatiotemporal stimulus code is dynamically adjusted online to obtain optimized personalized stimulus parameters.

Images