An AI-driven intervertebral disc multi-scale biomechanical simulation analysis method for spine health

Abstract

The application discloses an AI-driven intervertebral disc multiscale biomechanical simulation analysis method for spine health, aiming at solving the problems of idealization, low multiscale coupling efficiency and disconnection with clinic of the prior art model. Firstly, the resting and weight-bearing dual-state magnetic resonance images and long-term wearable sensing data of an individual are collected, a personalized intervertebral disc geometric model is constructed, and dynamic microenvironment parameters are initialized; an AI agent model based on a physical information neural network is constructed to map micro-mechanical responses at a microsecond level; the AI agent model is coupled to a macro finite element solver to perform time-domain simulation under real activity data driving, realizing efficient multiscale coupling; finally, simulation data are processed to generate a three-dimensional dynamic biomechanical risk atlas and extract quantitative clinical indicators, forming a personalized health assessment report. The application realizes a closed loop from individual data to precise simulation to clinical decision-making, and provides an efficient and intuitive tool for spine health assessment and intervention.

Description

Technical Field

[0001] This invention relates to the fields of computer-aided engineering and biomechanical simulation technology, and in particular to an AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs for spinal health. Background Technology

[0002] Spinal health assessment and disease prevention are important topics in rehabilitation medicine and biomechanics. Intervertebral discs, as key biomechanical buffer structures of the spine, are prone to degeneration and damage, leading to conditions such as low back pain. Current techniques typically employ biomechanical simulations based on the finite element method for intervertebral disc mechanical analysis. However, these methods have significant limitations, resulting in substantial discrepancies between the analytical results and actual in vivo conditions, making them unsuitable for precise, individualized health assessments.

[0003] The first type of problem with existing technologies lies in the idealization and staticization of simulation models. Traditional finite element models are usually based on standard geometric libraries or single-moment medical images (such as static CT or MRI), and their material property parameters are mostly derived from in vitro experimental data or empirical values ​​from literature. This method ignores two key facts: First, the living intervertebral disc exists in a dynamically changing mechanical and chemical microenvironment. Its internal osmotic pressure, nutrient concentration, and water content change periodically with daily activities (such as weight-bearing and rest), and these changes directly affect the stiffness and creep characteristics of the intervertebral disc. Second, the geometry, annulus fibrosus laminar structure, and cartilaginous endplate pore distribution of individual intervertebral discs are highly specific, and idealized models cannot capture the subtle influence of these irregular geometric features on the internal stress distribution.

[0004] The second type of problem with existing technologies lies in the rigidity and inefficiency of multi-scale mechanical coupling. The biomechanical behavior of intervertebral discs spans multiple scales, including molecular, cellular, tissue, and organ levels. For example, the macroscopic mechanical properties of the annulus fibrosus depend on the interaction and arrangement of collagen fibers and proteoglycans at the microscopic scale. Existing methods attempt to link different scales through top-down cross-scale simulations or bottom-up homogenization methods, but computational efficiency and accuracy are difficult to balance. Performing a complete finite element analysis at the macroscopic scale is already extremely time-consuming; if microscopic scale changes are to be coupled in real time during simulation, computational resources will be insufficient, making it almost impossible to simulate the long-term (e.g., daily activity cycle) evolution of the intervertebral disc.

[0005] The third problem with existing technologies lies in the disconnect between analytical results and clinical assessment. Current simulation outputs are mostly stress and strain contour maps or single mechanical indicators, lacking a direct, dynamic correlation with clinically observable and modifiable parameters (such as intervertebral disc height, nucleus pulposus displacement trend, and local risk areas of the annulus fibrosus). Physicians or health analysts find it difficult to directly translate complex mechanical data into personalized behavioral guidance or health warnings.

[0006] In summary, existing technologies lack a biomechanical simulation analysis method for intervertebral discs that can integrate dynamic information of the individual living microenvironment, efficiently and accurately couple multi-scale mechanics, and output intuitive clinical guidance indicators. Summary of the Invention

[0007] To achieve the above objectives, this invention provides an AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs aimed at spinal health, the method comprising the following steps:

[0008] Step 1: Collect and fuse multimodal medical imaging data and wearable sensor data of the target individual to construct a personalized intervertebral disc geometric model and initialize dynamic microenvironment parameters;

[0009] Step 2: Construct and train an artificial intelligence agent model based on a physical information neural network. The artificial intelligence agent model is used to map the mechanical response of the intervertebral disc micro-components under given strain and chemical concentration.

[0010] Step 3: Perform macroscopic time-domain biomechanical simulation coupled with the artificial intelligence agent model. Load the intervertebral disc geometric model and dynamic microenvironment parameters obtained in Step 1 and the artificial intelligence agent model trained in Step 2 into the macroscopic finite element solver. Drive the simulation based on the wearable sensor data collected in Step 1, and call the artificial intelligence agent model to calculate the material constitutive response within each simulation time step.

[0011] Step 4: Process the full-time simulation data output from Step 3, generate a dynamic biomechanical risk map of the intervertebral disc, extract quantitative clinical indicators, and form a structured health assessment report.

[0012] Preferably, the step 1 of collecting and fusing multimodal medical imaging data and wearable sensor data of the target individual specifically includes:

[0013] Step 1.1: Acquire high-resolution resting-state magnetic resonance imaging (MRI) data and high-resolution load-bearing-state MRI data of the spinal region of the target individual; the resting-state MRI data is T2-weighted image sequence data acquired by the individual in a supine position; the load-bearing-state MRI data is T2-weighted image sequence data acquired by the individual under a simulated upright load device;

[0014] Step 1.2: Process the resting-state and load-bearing-state magnetic resonance imaging data acquired in Step 1.1; perform three-dimensional segmentation on the resting-state and load-bearing-state magnetic resonance imaging data respectively, with the segmentation objects including the superior endplate, inferior endplate, nucleus pulposus region, and annulus fibrosus region; based on the segmentation results of the resting-state magnetic resonance imaging data, reconstruct the three-dimensional geometric model of the target intervertebral disc;

[0015] Step 1.3: Generate an unstructured tetrahedral mesh inside the three-dimensional geometric model reconstructed in Step 1.2; set the mesh density to a first density value in the region where the inner layer of the annulus fibrosus meets the nucleus pulposus and the region where the outer layer of the annulus fibrosus attaches to the endplate; set the mesh density to a second density value in the central region of the nucleus pulposus and the main body region of the annulus fibrosus; the second density value is lower than the first density value.

[0016] Step 1.4: Synchronously collect the spinal activity and load time-series data of the target individual over a continuous 72-hour period; collect the surface pressure distribution time-series data of the lumbar region through a flexible array pressure sensor attached to the skin surface of the individual's waist; collect the three-dimensional motion angle and angular velocity time-series data of the lumbar spine through an inertial measurement unit sensor integrated in the smart waist belt;

[0017] Step 1.5: Based on the data from Steps 1.1 to 1.4, initialize the dynamic microenvironment parameters of the intervertebral disc; convert the T2 signal intensity value of the nucleus pulposus region in the resting-state magnetic resonance imaging data into an initial moisture content value through a pre-calibrated signal intensity-moisture content relationship curve; estimate the initial instantaneous elastic modulus range of the intervertebral disc tissue based on the changes in nucleus pulposus volume and height in the load-bearing magnetic resonance imaging data and the resting-state magnetic resonance imaging data, combined with the standard load value applied by the simulated upright loading device; use the initial 2-hour activity and load time-series data collected in Step 1.4 as the microenvironment driving input sequence.

[0018] Preferably, the flexible array pressure sensor in step 1.4 is composed of multiple miniature pressure sensing units arranged in a matrix, and the sampling frequency of each miniature pressure sensing unit is not less than 100 Hz; the inertial measurement unit sensor includes a three-axis accelerometer and a three-axis gyroscope, and the data of the inertial measurement unit sensor and the data of the flexible array pressure sensor are synchronized and aligned through timestamps; the surface pressure distribution time series data and the lumbar spine three-dimensional motion angle and angular velocity time series data together constitute the spinal activity and load time series data.

[0019] Preferably, the process of reconstructing the three-dimensional geometric model of the target intervertebral disc based on the segmentation results of the resting-state magnetic resonance imaging data in step 1.2 is as follows: the resting-state magnetic resonance imaging data is segmented using a level set algorithm or a three-dimensional U-Net neural network to obtain a binary mask of the nucleus pulposus region and the annulus fibrosus region; the binary mask is reconstructed using a three-dimensional Poisson surface and then smoothed using Laplacian to obtain a smooth three-dimensional geometric surface model of the intervertebral disc that maintains the anatomical boundary features; the three-dimensional geometric surface model of the intervertebral disc clearly distinguishes the interface between the nucleus pulposus and the annulus fibrosus.

[0020] Preferably, the pre-calibrated signal intensity-moisture content relationship curve in step 1.5 is obtained through in vitro experiments. The specific process is as follows: obtain in vitro intervertebral disc samples with different degrees of degeneration; measure the T2 signal intensity of each in vitro intervertebral disc sample using an MRI scanner; measure the absolute moisture content of each in vitro intervertebral disc sample using a drying and weighing method; and perform polynomial curve fitting on multiple sets of T2 signal intensity data and absolute moisture content data to obtain the signal intensity-moisture content relationship curve.

[0021] Preferably, the construction and training of the artificial intelligence agent model based on the physical information neural network in step 2 specifically includes:

[0022] Step 2.1: Construct an offline high-fidelity microscopic representative volume element finite element model library; the microscopic representative volume element finite element model library includes a microscopic representative volume element model of annulus fibrosus tissue and a microscopic representative volume element model of nucleus pulposus tissue; the microscopic representative volume element model of annulus fibrosus tissue includes alternating layers of collagen fibers and a proteoglycan matrix, and the fiber orientation of the collagen fibers is set to alternate angles between adjacent layers; the microscopic representative volume element model of nucleus pulposus tissue includes discrete chondrocytes and a hydrophilic proteoglycan network;

[0023] Step 2.2: For each microscopic representative volume element finite element model in Step 2.1, apply multiple sets of random multiaxial strain loads and chemical concentration loads to its boundaries, and run high-fidelity finite element calculations; record the average stress response tensor of the microscopic representative volume element finite element model under each set of loads, as well as the statistical characteristics of the maximum principal strain distribution within the model; the statistical characteristics of the maximum principal strain distribution include the average value, standard deviation, and volume percentage exceeding a set threshold of the maximum principal strain;

[0024] Step 2.3: Using the random multiaxial strain load and chemical concentration load applied in Step 2.2 as input features, and the average stress response tensor and maximum principal strain distribution statistical features calculated in Step 2.2 as output labels, a training dataset is constructed; a physical information neural network is trained using the training dataset; the loss function of the physical information neural network is composed of a weighted sum of data fitting loss and a predefined linear momentum equilibrium partial differential equation residual loss; the trained physical information neural network becomes the artificial intelligence proxy model.

[0025] Preferably, the specific interaction process in step 3, which involves calling the AI ​​proxy model to calculate the material constitutive response within each simulation time step, is as follows:

[0026] Step 3.1: When the macroscopic finite element solver is performing mechanical equilibrium iterative calculations, it pauses the current iteration process when it reaches the material constitutive response at a certain Gaussian integration point.

[0027] Step 3.2: Extract the current strain tensor at the Gaussian integration point and obtain the updated local chemical concentration value at the Gaussian integration point based on the two-phase porous media theory from historical simulation data;

[0028] Step 3.3: Input the current strain tensor and the local chemical concentration value into the artificial intelligence proxy model; the artificial intelligence proxy model outputs a macroscopic stress tensor corresponding to the Gaussian integration point and a microscopic damage risk coefficient;

[0029] Step 3.4: Return the macroscopic stress tensor to the macroscopic finite element solver, which uses the macroscopic stress tensor to continue performing equilibrium iteration calculations; at the same time, store the microscopic damage risk coefficient for subsequent analysis.

[0030] Preferably, in step 4, the process of generating a dynamic biomechanical risk map of the intervertebral disc from the full-time simulation data output in step 3 is described in the following specific method:

[0031] Step 4.1: Extract the micro-damage risk coefficient of all Gaussian integration points at each simulation time step from the full-time simulation data in Step 3;

[0032] Step 4.2: For each Gaussian integral point corresponding to a spatial location in the three-dimensional geometric model, perform time-weighted accumulation of its micro-damage risk coefficients at all simulation time steps to obtain the spatial risk accumulation value; the weight of the time-weighted accumulation is the actual activity duration represented by each simulation time step;

[0033] Step 4.3: Map the calculated spatial risk accumulation value of each Gaussian integral point back to the nodes of the three-dimensional geometric model of the intervertebral disc generated in Step 1 using an interpolation method;

[0034] Step 4.4: Based on the spatial risk accumulation value mapped to the nodes, the three-dimensional geometric model is rendered using color gradient to generate a three-dimensional dynamic biomechanical risk map; in the dynamic biomechanical risk map, high-risk areas are displayed in warm colors and low-risk areas are displayed in cool colors.

[0035] Preferably, the extraction of quantitative clinical indicators in step 4 specifically includes:

[0036] Calculate the minimum and average values ​​of intervertebral disc height throughout the entire simulation time period, as well as the time point at which the minimum value occurs;

[0037] Calculate the maximum displacement and direction of the nucleus pulposus center point in the sagittal and coronal planes;

[0038] Set a cumulative threshold for the microscopic damage risk coefficient, count the volume in the annulus fibrosus region where the cumulative spatial risk value exceeds the cumulative threshold, calculate the percentage of the volume in the annulus fibrosus region to the total volume of the entire annulus fibrosus region, and record it as the high-risk volume percentage.

[0039] The minimum and average values ​​of the intervertebral disc height, the time point at which the minimum value occurs, the maximum displacement and direction of displacement, and the proportion of high-risk volume together constitute the quantitative clinical indicators.

[0040] Preferably, the formation of the structured health assessment report in step 4 specifically includes: integrating the dynamic biomechanical risk map with the quantitative clinical indicators; the structured health assessment report clearly identifies the biomechanically weak spatial location of the intervertebral disc of the target individual within the 72-hour activity cycle collected in step 1; the structured health assessment report quantitatively assesses the spinal load level of the target individual under current activity habits; and the structured health assessment report, based on the simulation data comparison in step 3, provides specific behavioral pattern adjustment suggestions, including adjusting specific movement postures and introducing intermittent rest periods, with the behavioral pattern adjustment suggestions aiming to reduce the cumulative load of the high-risk areas.

[0041] The beneficial effects of this invention are:

[0042] 1. This invention constructs a simulation model reflecting the true geometric morphology of an individual and the dynamic biomechanical microenvironment of a living organism by fusing resting and loaded dual-state magnetic resonance imaging with long-term wearable sensor data. This solves the distortion problem caused by the use of idealized geometry and static material properties in existing technologies. Specifically, the loaded-state imaging provides the morphological response under load, while the wearable data provides continuous biomechanical boundary condition drivers. This allows the simulation initialization conditions to leap from static approximation to dynamic personalization, significantly improving the model's accuracy in representing the true state of an individual and laying the foundation for precise evaluation.

[0043] 2. This invention creatively introduces an artificial intelligence surrogate model trained on a physical information neural network, replacing the high-fidelity micro-finite element calculations required in real-time at each macroscopic integration point in traditional multi-scale simulations. This surrogate model incorporates physical law constraints during training, enabling the mapping from microscopic strain and chemical concentration to macroscopic stress and damage risk coefficients within microseconds. This achieves efficient, real-time coupling between macroscopic and microscopic scales, making previously impossible long-term cross-scale dynamic simulations possible and completely solving the computational efficiency bottleneck.

[0044] 3. This invention transforms the massive and complex mechanical data output from full-time, multi-scale simulations into an intuitive three-dimensional dynamic biomechanical risk map using a risk spatiotemporal accumulation algorithm. It also extracts quantitative indicators directly related to clinical concern, such as intervertebral disc height changes, nucleus pulposus displacement, and the proportion of high-risk areas. This approach transforms abstract mechanical analysis results into a visual and measurable health assessment tool, enabling physicians or health analysts to clearly pinpoint biomechanical weaknesses and directly link the assessment results to specific activity patterns. This generates personalized and actionable behavioral intervention recommendations, achieving an effective transformation from simulation data to clinical health management decisions. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0047] Figure 2 This is a flowchart of the steps in step 1 of the method of the present invention for collecting and fusing multimodal medical image data and wearable sensor data of the target individual;

[0048] Figure 3 A flowchart illustrating the steps of constructing and training an artificial intelligence agent model based on a physical information neural network according to the method of this invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0050] Please see Figures 1-3 This invention provides an AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs aimed at spinal health. To make the objectives, technical solutions, and advantages of this invention clearer, the following will provide a complete and detailed description of the invention in conjunction with the technical solutions. Those skilled in the art will understand that the specific embodiments described herein are only for explaining the invention and are not intended to limit the scope of protection of the invention.

[0051] The core process of this invention consists of four main steps, in sequence: collecting and fusing multimodal individual data to construct a personalized intervertebral disc geometry and microenvironment initialization model; constructing an artificial intelligence agent model based on a physical information neural network to map the mechanical response of microscopic components; performing a macroscopic-scale time-domain biomechanical simulation coupled with the artificial intelligence agent model; and generating a dynamic biomechanical risk map and quantitative report for spinal health. Each step is executed sequentially, with the output of the preceding step constituting the input of the subsequent step, forming a complete technical closed loop.

[0052] Each step is explained in detail below.

[0053] Step 1: Collect and fuse multimodal individual data to construct a personalized intervertebral disc geometry and microenvironment initialization model.

[0054] This step aims to acquire individual-specific data and, based on this, construct a digital model of the intervertebral disc that reflects the initial state of the living organism and can be used for simulation. This step integrates medical imaging and biosensing technologies.

[0055] Step 1.1: Acquire high-resolution resting-state and high-resolution weight-bearing-state MRI data of the spinal region of the target individual. Resting-state MRI data refers to MRI images acquired in a supine, relaxed state without external load, typically using a T2-weighted imaging sequence sensitive to water content. Weight-bearing-state MRI data refers to MRI images acquired under the same scanning parameters while the subject is subjected to a specific axial load (to simulate an upright or weight-bearing state), achieved using a dedicated MRI-compatible loading device.

[0056] Step 1.2: Process the resting-state and load-bearing MRI images acquired in Step 1.1. First, image segmentation techniques are used to precisely separate the various anatomical structures of the target intervertebral disc from the resting-state MRI images, including the superior vertebral endplate, the inferior vertebral endplate, the central nucleus pulposus region, and the annulus fibrosus region surrounding the nucleus pulposus. This segmentation process can be completed using a deep learning-based automatic segmentation network or an interactive semi-automatic segmentation algorithm. Based on the 3D mask of the segmented nucleus pulposus and annulus fibrosus regions, a smooth, non-uniform rational B-spline surface representation of the 3D geometric model is generated using a surface reconstruction algorithm. This model accurately preserves the individualized geometric features extracted from the images. Simultaneously, similar segmentation processing is performed on the load-bearing MRI images for subsequent parameter estimation.

[0057] Step 1.3: An unstructured tetrahedral mesh is generated within the 3D geometric model reconstructed in Step 1.2. Meshing is performed using finite element preprocessing software. To ensure a balance between computational accuracy and efficiency, a gradient mesh density is applied. Specifically, in areas of drastic mechanical change or high interest, such as the interface between the inner layer of the annulus fibrosus and the nucleus pulposus, and the attachment point between the outer layer of the annulus fibrosus and the cartilaginous endplate, a denser mesh is used, i.e., a first density value is set. In areas with relatively uniform or gradual mechanical changes, such as the central region of the nucleus pulposus and the main body of the annulus fibrosus, a relatively sparse mesh is used, i.e., a second density value is set, which is larger than the element size represented by the first density value.

[0058] Step 1.4: Synchronously acquire time-series data on spinal activity and load of the target individual over several consecutive days (e.g., 72 hours). This is achieved through a wearable sensing system. Specifically, a flexible array of pressure sensors is attached to the surface of the lumbar vertebrae to acquire the distribution of surface pressure on the lumbar region and its changes over time at high frequency. Simultaneously, a smart wearable device integrating inertial measurement unit sensors (typically including a triaxial accelerometer and a triaxial gyroscope) is used to continuously monitor the motion angles and angular velocities of the lumbar spine in three-dimensional space. The data from the two sets of sensors are synchronized using precise timestamps to jointly constitute time-series data on spinal activity and load reflecting the individual's daily activity patterns.

[0059] Step 1.5: Based on all the data collected in Steps 1.1 to 1.4, initialize the dynamic microenvironment parameters of the intervertebral disc. First, using the T2 signal intensity of the nucleus pulposus region in the resting-state MRI data, the initial water content value of the intervertebral disc nucleus pulposus is calculated through a pre-calibrated signal intensity-water content relationship curve in the laboratory. Second, by comparing the morphological changes of the nucleus pulposus (such as decreased height and volume) in the load-bearing and resting-state MRI images, and combining the known applied load values, a simplified mechanical formula is used to back-estimate the equivalent instantaneous elastic modulus range of the intervertebral disc tissue as the initial estimate of the material parameters. Finally, the spinal activity and load time-series data of the initial time period (e.g., 2 hours) collected in Step 1.4 are extracted as the initial input sequence to drive the subsequent biomechanical simulation.

[0060] Step 2: Construct an artificial intelligence agent model based on physical information neural networks to map the mechanical response of micro-components.

[0061] This step aims to create an artificial intelligence model that can quickly and accurately predict the microscopic mechanical behavior of the intervertebral disc, replacing the traditional, computationally expensive microscale finite element method, and is key to achieving efficient multiscale coupling.

[0062] Step 2.1: Construct an offline high-fidelity microscopic representative volume element finite element model library. A microscopic representative volume element is the smallest volume unit that can represent the statistical characteristics of a material's microstructure. Models need to be constructed separately for different components of the intervertebral disc. For the annulus fibrosus tissue, a microscopic representative volume element finite element model is established, containing periodically arranged collagen fiber layers and a proteoglycan matrix. The orientation of the collagen fibers simulates the layered structure of the natural annulus fibrosus, with adjacent fiber layers exhibiting a certain staggered angle. For the nucleus pulposus tissue, a microscopic representative volume element finite element model is established, containing randomly distributed chondrocytes and a proteoglycan-rich gel-like matrix.

[0063] Step 2.2: Perform large-scale offline finite element calculations on each microscopic representative volume element finite element model constructed in Step 2.1. Apply thousands of sets of randomly combined multiaxial mechanical strain loads and chemical concentration loads simulating changes in the extracellular fluid environment to the boundaries of the microscopic representative volume elements. For each set of loads, perform a complete nonlinear finite element analysis. After the calculation, extract the results at two levels: first, the overall average stress response of the microscopic representative volume elements, i.e., the macroscopic stress tensor; second, the statistical characteristics of the damage risk within the microscopic representative volume elements, such as the average value and standard deviation of the maximum principal strain at all integration points, and the volume proportion of material exceeding a certain preset damage threshold strain (e.g., 0.15).

[0064] Step 2.3: Using the load combination applied in Step 2.2 as input features, and the calculated macroscopic stress tensor and microscopic damage risk statistical features as output labels, a dataset is constructed to train the artificial intelligence model. This dataset is used to train a physical information neural network. This neural network not only learns the mapping relationship between input and output, but its loss function is also additionally constrained by physical constraints, such as the residuals of the law of conservation of linear momentum. The trained network is the artificial intelligence agent model, which can simultaneously output predicted macroscopic stress and microscopic damage risk indicators within milliseconds after receiving new strain and chemical concentration inputs.

[0065] Step 3: Perform macroscale time-domain biomechanical simulation of the coupled artificial intelligence agent model.

[0066] This step integrates the personalized model built in the previous steps with the artificial intelligence agent model, and simulates the biomechanical response of the intervertebral disc under real active loads in the time dimension. It is the core computational process of the entire method.

[0067] Step 3.1: Load the dynamic microenvironment parameters of the intervertebral disc initialized in Step 1.5, the geometric mesh model of the intervertebral disc generated in Step 1.3, and the artificial intelligence agent model trained in Step 2.3 into the macro-scale finite element solver. Set the total simulation duration to cover the complete activity monitoring cycle collected in Step 1.4 (e.g., 72 hours), and determine an appropriate time step.

[0068] Step 3.2: Within each simulation time step, execute the coupled calculation process. First, based on the surface pressure data and lumbar spine kinematic data collected in Step 1.4 corresponding to the current time step, calculate the equivalent mechanical loads and displacement constraints acting on the upper and lower endplates of the target intervertebral disc at that moment using an inverse dynamic model. Second, the macroscopic finite element solver begins iterative solution of mechanical equilibrium for that time step based on these boundary conditions.

[0069] Step 3.3: When the macroscopic finite element solver needs to calculate the material constitutive response (i.e., stress) at a certain Gaussian integration point (i.e., a spatially discrete point) during the iteration process, the solution process will pause at this point. The system will extract the current strain tensor at that Gaussian integration point and obtain the current chemical concentration value at that point from the mass transport field updated in the previous time step. Subsequently, the artificial intelligence proxy model is invoked, and this set of strain and chemical concentration values ​​are input. The artificial intelligence proxy model quickly outputs two results: one is the macroscopic stress tensor corresponding to the Gaussian integration point, and the other is the local microscopic damage risk coefficient at that point.

[0070] Step 3.4: The macroscopic finite element solver receives the macroscopic stress tensor returned from the AI ​​proxy model and substitutes it into the equilibrium equation to continue iterating until the mechanical equilibrium converges at that time step. Simultaneously, the microscopic damage risk coefficient calculated at this Gaussian integration point is stored along with the spatial location and temporal information.

[0071] Step 3.5: After the mechanical equilibrium calculation is completed in one time step, based on the stress field and pore pressure field obtained in the current step, and according to the two-phase porous media theory in biomechanics, update the flow of water and the diffusion and concentration distribution of solutes (such as ions) inside the intervertebral disc. The updated chemical field will be used as the input to the artificial intelligence agent model in the next time step, Step 3.3.

[0072] Step 3.6: Repeat steps 3.2 to 3.5 until all time steps of the set simulation duration (e.g., 72 hours) are completed. Throughout the process, all complex, microstructure-related nonlinear material response calculations are performed by an efficient artificial intelligence proxy model, thereby achieving efficient dynamic coupling between macro and micro scales.

[0073] Step 4: Generate a dynamic biomechanical risk map and quantitative report for spinal health.

[0074] This step involves post-processing and visualizing the massive amounts of data generated from the simulation, extracting indicators with clinical guidance significance, and generating the final user report.

[0075] Step 4.1: Process the full-time simulation data output from Step 3. The key operation is to extract and integrate the microscopic damage risk coefficients. Specifically, for each Gaussian integral point in the intervertebral disc 3D model, read the sequence of microscopic damage risk coefficients calculated over all simulation time steps. Considering the differences in activity intensities, perform time-weighted summation on this sequence. The weighting coefficient is determined based on the activity load level at that moment (e.g., 1.0 for resting and 2.5 for lifting heavy objects), thus obtaining the cumulative risk value for each spatial location.

[0076] Step 4.2: Generate a dynamic biomechanical risk map of the intervertebral disc. The cumulative risk value of each Gaussian integral point calculated in Step 4.1 is mapped back to the surface nodes of the three-dimensional geometric model of the intervertebral disc using an interpolation algorithm. In a three-dimensional visualization environment, a continuous color gradient from blue (low risk) to red (high risk) is used for rendering based on the magnitude of the node risk value, ultimately generating a three-dimensional color risk map. This map can visually display which parts of the intervertebral disc, especially the annulus fibrosus region, experienced the highest cumulative mechanical load during monitoring.

[0077] Step 4.3: Quantitatively extract clinical indicators from the full-time data. Key indicators include: the dynamic change curve of intervertebral disc height during the simulation period, from which the minimum height value, average height value, and the time of occurrence of the minimum height can be read; the maximum displacement and direction of the nucleus pulposus center point in the sagittal and coronal planes; and, after defining a risk threshold, calculating the percentage of the annulus fibrosus tissue with a cumulative risk value exceeding that threshold, i.e., the high-risk volume percentage.

[0078] Step 4.4: Combining the risk map generated in Step 4.2 and the quantitative indicators extracted in Step 4.3, a structured health assessment report is generated. This report not only presents intuitive visualizations and specific data, but also analyzes the correlation between the formation of high-risk areas and specific activity patterns (such as prolonged sitting and bending) in conjunction with the simulation process. Finally, it provides personalized spinal health guidance suggestions aimed at reducing the cumulative load on target areas, such as adjusting work posture, increasing specific rest intervals, and conducting targeted rehabilitation exercises.

[0079] Through the detailed implementation of the above four steps, this invention completes the entire process from individual data collection to personalized health report output, realizing a precise, efficient, and clinically significant assessment of the biomechanical state of the intervertebral disc.

[0080] Example

[0081] The application scenario of this embodiment is: to conduct a biomechanical health assessment of the intervertebral disc (L4 / L5) between the fourth and fifth lumbar vertebrae of a male patient (Mr. Zhang, 45 years old) who works at a desk for a long time and has intermittent lower back pain. The aim is to quantify the cumulative mechanical load on the intervertebral disc caused by his daily activities, identify potential risk areas, and provide a scientific basis for lifestyle adjustments.

[0082] Step 1: Collect and fuse multimodal individual data to construct a personalized intervertebral disc geometry and microenvironment initialization model.

[0083] Step 1.1: Acquire high-resolution resting-state magnetic resonance imaging data and high-resolution load-bearing-state magnetic resonance imaging data of the spinal region of the target individual.

[0084] First, resting-state T2-weighted images of Mr. Zhang's lumbar spine region were acquired using a fast spin-echo sequence on a 1.5 T magnetic resonance imaging (MRI) scanner. Scanning parameters were set as follows: repetition time (TR) 3000 ms, echo time (TE) 100 ms, slice thickness 3 mm, no slice spacing, matrix size 512×512, and field of view (FOV) 280 mm×280 mm. The patient underwent the scan in a supine position to ensure the spine was in a naturally relaxed state; the acquired image data constitutes the resting-state MRI data.

[0085] Subsequently, a dedicated MRI-compatible spinal load simulation device was used. This device, through a pulley and weight-bearing belt system, applies a constant 300 N tension along the spinal axis to the patient's shoulders while the patient lies supine on the scanning table, simulating the static load on the upper body when upright. Under this load-bearing state, a T2-weighted scan was performed again using the same sequence and parameters as the resting-state scan, and the acquired data became the load-bearing MRI image data.

[0086] Step 1.2: Process the magnetic resonance imaging data acquired in Step 1.1.

[0087] Resting-state and load-bearing-state MRI data were imported into medical image processing software. An automatic segmentation algorithm based on a 3D U-Net convolutional neural network was used to segment the L4 / L5 intervertebral disc. This neural network was pre-trained using thousands of labeled lumbar spine MRI images. The segmentation process automatically identified and output four independent 3D masks: the L4 inferior endplate mask, the L5 superior endplate mask, the L4 / L5 intervertebral disc nucleus pulposus region mask, and the L4 / L5 intervertebral disc annulus fibrosus region mask. After segmentation, necessary fine-tuning was performed manually to ensure anatomical accuracy. Subsequently, using the nucleus pulposus and annulus fibrosus masks obtained from resting-state image segmentation, a smooth, non-uniform rational B-spline surface representation of a 3D geometric model was generated using a Poisson surface reconstruction algorithm. This model accurately preserves the irregular natural boundary between the nucleus pulposus and annulus fibrosus, as well as the complex geometry of the connection between the outer layer of the annulus fibrosus and the endplate.

[0088] Step 1.3: Generate an unstructured tetrahedral mesh inside the 3D geometric model reconstructed in Step 1.2.

[0089] The aforementioned 3D geometric model was imported into the finite element preprocessing software. An adaptive tetrahedral mesh generation algorithm was used for discretization. The mesh size was gradient-divided according to the importance of the regions. Specifically, in the transition region between the inner layer of the annulus fibrosus and the nucleus pulposus (approximately 2 mm thick), and in the region where the outer layer of the annulus fibrosus attaches to the upper and lower endplates (approximately 1.5 mm thick), the target element edge length was set to 0.4 mm, which is the first density value. In the central region of the nucleus pulposus and the main body of the annulus fibrosus, the target element edge length was set to 1.2 mm, which is the second density value. The final generated mesh contains approximately 185,000 tetrahedral elements and approximately 38,000 nodes. The software checked the mesh quality to ensure that the minimum Jacobian coefficient of all elements was greater than 0.6, meeting the quality requirements of the finite element calculation.

[0090] Step 1.4: Synchronously collect spinal activity and load time-series data of the target individual over a continuous 72-hour period.

[0091] Mr. Zhang was fitted with an integrated sensor system. A flexible array of pressure sensor pads was attached to the skin surface around the L3 to S1 spinous processes. This pad contained 16×8 miniature pressure sensing units, each with an area of ​​5 mm × 5 mm, and a sampling frequency of 200 Hz, continuously recording the two-dimensional distribution of pressure on the lumbar surface and its changes over time. Simultaneously, Mr. Zhang wore a smart belt with a built-in inertial measurement unit (IMU) sensor located at the surface projection point behind the L4 / L5 intervertebral disc. The IMU included a three-axis accelerometer and a three-axis gyroscope, with a sampling frequency of 100 Hz, used to continuously measure the flexion-extension, lateral bending, and rotation angles and angular velocities of the lumbar spine. Both devices were connected to the same data logger via Bluetooth. The data logger used a GPS clock signal as a reference to timestamp all data streams, achieving microsecond-level synchronization.

[0092] Step 1.5: Based on the data from Steps 1.1 to 1.4, initialize the dynamic microenvironment parameters of the intervertebral disc.

[0093] First, the average T2 signal intensity of the L4 / L5 nucleus pulposus region was measured to be 125 (in arbitrary units) on resting-state magnetic resonance imaging. The calibration curve obtained from previous in vitro experiments was reviewed. The calibration process was as follows: Fifteen human lumbar intervertebral disc samples of different ages and degeneration levels were collected, and the T2 signal intensity of their nuclei was measured. Subsequently, the moisture content was accurately measured using a freeze-drying method (dry weight method). A second-order polynomial fitting was performed on the data points to obtain the calibration curve, the function expression of which is:

[0094]

[0095] Where W is the percentage of moisture content (in %), no further percentage conversion is needed, and S represents the T2 signal strength value. Substituting S=125 into the formula, the initial moisture content W is calculated to be approximately 67.2%.

[0096] Secondly, the images under resting and loading conditions were compared. At rest, the sagittal height of the nucleus pulposus was measured to be 10.2 mm; under a 300 N load, this height decreased to 8.9 mm, representing a compression of approximately 12.7%. Combining the applied 300 N standard load with the approximate cross-sectional area of ​​the intervertebral disc, the initial instantaneous elastic modulus of the tissue was estimated to be between 2.5 MPa and 4.0 MPa using a simplified form of Hooke's Law. This range will serve as prior knowledge for subsequent material parameter optimization.

[0097] Finally, the sensor data (surface pressure time series and kinematic time series) collected in step 1.4 within the first 2 hours are used as the driving input sequence for starting the simulation. This sequence includes various initial activities of the patient, such as getting up from bed rest, walking, and sitting.

[0098] Step 2: Construct an artificial intelligence agent model based on a physical information neural network to map the mechanical response of micro-components.

[0099] Step 2.1: Construct an offline high-fidelity microscopic representative volumetric finite element model library.

[0100] For the annulus fibrosus, a cubic volumetric model with a side length of 0.5 mm was constructed. The model contains 10 alternating layers of composite material, each 0.05 mm thick. In odd-numbered layers, the collagen fibers are oriented at an angle of +30 degrees to the long axis of the cube, while in even-numbered layers, the fiber orientation is -30 degrees, thus achieving an alternating angle arrangement. The fibers are embedded in a matrix composed of proteoglycans. The fiber material is defined as linearly elastic with an elastic modulus of 500 MPa; the matrix material is defined as a hyperelastic Ogden model. For the nucleus pulposus, another cubic volumetric model with a side length of 1.0 mm was constructed, in which approximately 1% of the volume of spherical chondrocytes are randomly distributed as inclusions, surrounded by a dominant hydrophilic proteoglycan network. This network material model employs a biphasic porous hyperelastic constitutive model capable of describing large deformations and osmotic pressure effects.

[0101] Step 2.2: Perform extensive offline calculations on the microscopic representative volume element model.

[0102] The two representative volume element models were calculated using the commercial finite element software Abaqus. Periodic boundary conditions were applied to the boundaries of the representative volume element models, and 5000 different sets of multiaxial strain loads (including combinations of tension, compression, and shear) and chemical concentration loads (simulating the variation of extracellular sodium ion concentration in the range of 0.12 mol / L to 0.16 mol / L) were randomly generated. For each set of loads, a nonlinear finite element analysis was performed. After the calculation was completed, the volume-mean Cauchy stress tensor of the representative volume element was extracted. Simultaneously, the maximum principal strain values ​​at all integration points within the entire representative volume element are scanned, and their average values ​​are calculated. Standard deviation The percentage of unit volumes with a maximum principal strain value exceeding 0.15 (this threshold is based on the strain level at which micro-damage begins to appear in the collagen fibers of the fibrous ring, as defined in the literature) was also calculated. The stress tensor and these three statistical features together constitute the output label.

[0103] Step 2.3: Train the physical information neural network.

[0104] Using 4000 sets of the 5000 load data sets from step 2.2 as the training set, 500 sets as the validation set, and 500 sets as the test set, a physical information neural network is constructed. Its input layer has 7 nodes, corresponding to 6 independent strain tensor components and 1 chemical concentration value. The output layer has 9 nodes, corresponding to 6 independent stress tensor components and... , , The network contains five fully connected hidden layers, each with 128 neurons, using the Swish activation function. The unique feature of this network is the definition of its loss function L:

[0105]

[0106] in, It is the sum of the mean square error between the predicted stress and the finite element calculated stress, and the mean square error between the predicted statistic and the finite element statistic. It is the physical constraint loss, calculated by taking the stress tensor predicted by the network. Substituting into the linear momentum balance equation (At the scale representing volume element, physical strength) Set it to zero), and calculate its residual norm. and The weights are 1.0 and 0.1 respectively in this embodiment. The Adam optimizer is used to train the network until the validation set loss no longer decreases. The trained network becomes an artificial intelligence agent model, which can complete a mapping from microscopic strain and chemical environment to macroscopic stress and damage risk indicators within milliseconds.

[0107] Step 3: Perform macroscale time-domain biomechanical simulation of the coupled artificial intelligence agent model.

[0108] Step 3.1: Simulation initialization.

[0109] The L4 / L5 intervertebral disc tetrahedral mesh model generated in step 1.3, the water content and elastic modulus range parameters initialized in step 1.5, and the AI ​​proxy model trained in step 2.3 (encapsulated as a dynamic link library function) were loaded into the self-developed macroscale coupled finite element solver. The total simulation duration was set to 72 hours, and the simulation time step was [not specified]. Setting it to 0.1 s, which matches the frequency of the main activity, requires a total of 2,592,000 time steps to be calculated.

[0110] Step 3.2: Apply boundary conditions at each time step.

[0111] For the i-th time step, the information at the corresponding moment is read from the synchronous sensor data acquired in step 1.4. The surface pressure distribution data is converted into distributed pressure loads acting on the surfaces of the inferior endplate of the L4 vertebra and the superior endplate of the L5 vertebra through an inverse mechanics model (which considers the force transmission characteristics of skin and soft tissue). Simultaneously, the lumbar vertebra angle changes measured by the inertial measurement unit are converted into displacement constraints of the L4 vertebra relative to the L5 vertebra.

[0112] In one example, the method for constructing the inverse mechanics model is as follows: Based on the three-dimensional geometric model reconstructed in step 1.2 and its surrounding soft tissue envelope, a simplified finite element model is constructed, including the L4 vertebral body, L5 vertebral body, intervertebral disc, ligaments, and the surface soft tissue layer. The surface pressure distribution measured by the flexible array pressure sensor is applied as a surface force load to the corresponding area of ​​the soft tissue layer on the surface of the model; the lumbar spine angle measured by the inertial measurement unit is applied as a displacement boundary condition to the L4 vertebral body. Through a single linear or quasi-static finite element solution, the equivalent nodal forces, i.e., the distributed pressure loads, transmitted from the surface soft tissue to the lower endplate of L4 and the upper endplate of L5 can be calculated.

[0113] Step 3.3: Call the artificial intelligence agent model to calculate the material response.

[0114] During the nonlinear iterative solution process, the macroscopic finite element solver pauses when it needs to calculate the material stress at a specific Gaussian integration point (e.g., a point located on the outer rear side of a fiber ring). The solver then provides the current strain tensor at that integration point. And the current local sodium ion concentration updated based on the pore pressure and ion diffusion results calculated at this point in the previous time step. Then, the AI ​​Model(ε_c, C_c) function is called. This function returns two results within approximately 0.5 ms: one is the macroscopic stress tensor at that point. Secondly, it integrates... , , The microscopic damage risk coefficient R obtained from information calculation is calculated using the following formula:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] Step 3.4: Update the internal mass transport field.

[0123] After mechanical equilibrium is achieved at a macroscopic time step, the flow velocity of water and the concentration distribution of chemical ions within the intervertebral disc are updated based on the calculated new solid matrix stress and pore fluid pressure, according to two-phase porous media theories (such as Darcy's law and Fick's law). The updated concentration field will then serve as the input to the artificial intelligence proxy model in the next time step, step 3.3. .

[0124] Step 3.5: Execute repeatedly.

[0125] Repeat steps 3.2 to 3.4 until the full 72-hour simulation is completed. Because the artificial intelligence agent model greatly replaces the traditional time-consuming process of iteratively solving complex constitutive equations at each integration point, this simulation spanning 72 hours of physiological time was completed within approximately 36 hours of computing cluster time.

[0126] Step 4: Generate a dynamic biomechanical risk map and quantitative report for spinal health.

[0127] Step 4.1: Post-process the full-time risk data.

[0128] From the stored data, the micro-damage risk coefficient R(t) for all Gaussian integration points over all 2,592,000 time steps is extracted. For each fixed spatial integration point j, its cumulative spatial risk value over the entire simulation period is calculated. :

[0129]

[0130] in, This is the weight of the actual activity duration represented by that time step. For example, the weight of sitting and slow walking is 1.0, while the weight of the moment of bending over to lift a heavy object may be set to 2.0 to emphasize the contribution of high-intensity activities.

[0131] Step 4.2: Generate a three-dimensional dynamic biomechanical risk map.

[0132] All integration points The values ​​are mapped to the nodes of the intervertebral disc mesh model generated in step 1.3 using linear interpolation. In the 3D visualization software, color mapping is performed based on the node's ARC value: the top 5% of ARC values ​​are set to red, the bottom 50% to dark blue, and the middle portion is smoothly transitioned according to the color spectrum. This generates a 3D color atlas that can be arbitrarily rotated, scaled, and sectioned for viewing. The atlas clearly shows that in Mr. Zhang's L4 / L5 intervertebral disc model, a significant red high-risk plaque appears in the posterolateral right region of the annulus fibrosus.

[0133] Step 4.3: Extract quantitative clinical indicators.

[0134] Intervertebral disc height variation: The variation curve of the intervertebral disc centerline height in the simulation was calculated, and the minimum height within 72 hours was 8.7 mm, and the average height was 9.5 mm. The minimum height occurred at 41 hours, corresponding to sensor data showing that it occurred after a long period of driving.

[0135] Nucleus pulposus displacement: By tracing the movement trajectory of the centroid of the nucleus pulposus region, the maximum posterior displacement in the sagittal plane was calculated to be 1.8 mm, and the maximum rightward displacement in the coronal plane was 1.2 mm.

[0136] High-risk volume percentage: The cumulative threshold for ARC is set as the upper quartile of the ARC values ​​at all fiber ring integration points. The total volume of cells in the fiber ring region with ARC values ​​exceeding this threshold is statistically analyzed, and it is calculated that this high-risk volume accounts for approximately 18.5% of the total fiber ring volume.

[0137] Step 4.4: Generate a structured health assessment report.

[0138] The report integrates the aforementioned atlases and indicators. The conclusion section clearly states: "Under the current activity pattern, the patient's L4 / L5 intervertebral disc exhibits a clear biomechanical load concentration zone on the posterolateral aspect of the annulus fibrosus (right), with a high-risk tissue volume accounting for 18.5%. Prolonged sitting (especially while driving) results in the greatest loss of disc height, accompanied by a tendency for posterior displacement of the nucleus pulposus." The recommendations include: "1. It is recommended to perform 5 minutes of standing and lumbar extension exercises every 45 minutes of continuous sitting; 2. Strengthen core muscle training, especially transverse abdominis activation training; 3. Avoid lifting heavy objects requiring significant lumbar rotation when fatigued."

[0139] Comparison of proportions

[0140] To demonstrate the superiority of the method of this invention, a comparative example is established. The comparative example adopts the traditional method: a homogenized finite element model is established using only resting-state magnetic resonance imaging. The annulus fibrosus and nucleus pulposus are assigned fixed material parameters from the literature, without coupling the chemical microenvironment or using an artificial intelligence proxy model. Only a single static load (simulating lifting a 20 kg weight) simulation is performed. The results are compared in Table 1.

[0141] Table 1. Comparison of Results

[0142] Model personalization Based on individual dual-state magnetic resonance imaging and long-term sensor data, the geometry and microenvironment are highly personalized. Based solely on static images, the data is geometrically approximated; material parameters are population averages, with no information on individual microenvironments. The model of this invention more realistically reflects the unique anatomical structure and in vivo environment of an individual. Simulation content and efficiency The dynamic simulation of a real-world activity-driven coupled chemical field was completed in approximately 36 hours. The simulation of just one static working condition took about 1 hour to complete. This invention enables long-term dynamic analysis within an acceptable timeframe, whereas traditional methods cannot simulate the cumulative effects over time. Richness of output results The output includes a full-time three-dimensional dynamic risk map, mechanical status at multiple time points, and clinical quantitative indicators such as nucleus pulposus displacement and the proportion of high-risk volumes. The output is mainly stress-strain contour plots under a single loading event, lacking a time dimension and intuitive health indicators. The results of this invention can be directly used for health assessment and guidance, whereas traditional methods have difficulties in interpreting results and are disconnected from clinical practice. Ability to identify local risks High-risk plaques (accounting for 18.5%) were clearly located in specific areas on the outer side of the annulus fibrosus. It shows areas of high overall stress, but cannot accurately locate and quantify the "volume of high-risk tissue". This invention can provide more accurate and quantifiable local risk assessment, which is conducive to targeted intervention. Evaluation of activity patterns It can correlate specific activities (such as long-distance driving) with the occurrence of minimum intervertebral disc height and provide specific behavioral suggestions. It is impossible to assess the impact of different activity modes; only a single extreme condition can be evaluated. This invention enables personalized health management based on an individual's actual lifestyle.

[0143] As can be seen from the above embodiments and comparisons, the method provided by the present invention achieves a more realistic, comprehensive, and clinically valuable in-depth analysis of the biomechanical health status of individual intervertebral discs through multi-source data fusion and multi-scale dynamic coupling simulation driven by an artificial intelligence agent model.

[0144] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0145] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs aimed at spinal health, characterized in that, The method includes the following steps: Step 1: Collect and fuse multimodal medical imaging data and wearable sensor data of the target individual to construct a personalized intervertebral disc geometric model and initialize dynamic microenvironment parameters; Step 2: Construct and train an artificial intelligence agent model based on a physical information neural network. The artificial intelligence agent model is used to map the mechanical response of the intervertebral disc micro-components under given strain and chemical concentration. Step 3: Perform macroscopic time-domain biomechanical simulation coupled with the artificial intelligence agent model. Load the intervertebral disc geometric model and dynamic microenvironment parameters obtained in Step 1 and the artificial intelligence agent model trained in Step 2 into the macroscopic finite element solver. Drive the simulation based on the wearable sensor data collected in Step 1, and call the artificial intelligence agent model to calculate the material constitutive response within each simulation time step. Step 4: Process the full-time simulation data output from Step 3, generate a dynamic biomechanical risk map of the intervertebral disc, extract quantitative clinical indicators, and form a structured health assessment report.

2. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 1, characterized in that, Step 1, which involves collecting and fusing multimodal medical imaging data and wearable sensor data of the target individual, specifically includes: Step 1.1: Acquire high-resolution resting-state magnetic resonance imaging (MRI) data and high-resolution load-bearing-state MRI data of the spinal region of the target individual; the resting-state MRI data is T2-weighted image sequence data acquired by the individual in a supine position; the load-bearing-state MRI data is T2-weighted image sequence data acquired by the individual under a simulated upright load device; Step 1.2: Process the resting-state and load-bearing-state magnetic resonance imaging data acquired in Step 1.1; perform three-dimensional segmentation on the resting-state and load-bearing-state magnetic resonance imaging data respectively, with the segmentation objects including the superior endplate, inferior endplate, nucleus pulposus region, and annulus fibrosus region; based on the segmentation results of the resting-state magnetic resonance imaging data, reconstruct the three-dimensional geometric model of the target intervertebral disc; Step 1.3: Generate an unstructured tetrahedral mesh inside the three-dimensional geometric model reconstructed in Step 1.2; set the mesh density to a first density value in the region where the inner layer of the annulus fibrosus meets the nucleus pulposus and the region where the outer layer of the annulus fibrosus attaches to the endplate; set the mesh density to a second density value in the central region of the nucleus pulposus and the main body region of the annulus fibrosus; the second density value is lower than the first density value. Step 1.4: Synchronously collect the spinal activity and load time-series data of the target individual over a continuous 72-hour period; collect the surface pressure distribution time-series data of the lumbar region through a flexible array pressure sensor attached to the skin surface of the individual's waist; collect the three-dimensional motion angle and angular velocity time-series data of the lumbar spine through an inertial measurement unit sensor integrated in the smart waist belt; Step 1.5: Based on the data from Steps 1.1 to 1.4, initialize the dynamic microenvironment parameters of the intervertebral disc; convert the T2 signal intensity value of the nucleus pulposus region in the resting-state magnetic resonance imaging data into an initial moisture content value through a pre-calibrated signal intensity-moisture content relationship curve; estimate the initial instantaneous elastic modulus range of the intervertebral disc tissue based on the changes in nucleus pulposus volume and height in the load-bearing magnetic resonance imaging data and the resting-state magnetic resonance imaging data, combined with the standard load value applied by the simulated upright loading device; use the initial 2-hour activity and load time-series data collected in Step 1.4 as the microenvironment driving input sequence.

3. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 2, characterized in that, The flexible array pressure sensor described in step 1.4 consists of multiple miniature pressure sensing units arranged in a matrix, with each miniature pressure sensing unit having a sampling frequency of not less than 100 Hz; the inertial measurement unit sensor includes a three-axis accelerometer and a three-axis gyroscope, and the data from the inertial measurement unit sensor and the data from the flexible array pressure sensor are synchronized and aligned using timestamps; the time-series data of surface pressure distribution and the time-series data of the three-dimensional motion angle and angular velocity of the lumbar spine together constitute the time-series data of spinal activity and load.

4. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 2, characterized in that, The specific process for reconstructing the three-dimensional geometric model of the target intervertebral disc based on the segmentation results of the resting-state magnetic resonance imaging data in step 1.2 is as follows: the resting-state magnetic resonance imaging data is segmented using a level set algorithm or a three-dimensional U-Net neural network to obtain a binary mask of the nucleus pulposus region and the annulus fibrosus region; the binary mask is reconstructed using a three-dimensional Poisson surface and smoothed using Laplacian to obtain a smooth three-dimensional geometric surface model of the intervertebral disc that maintains the anatomical boundary features; the three-dimensional geometric surface model of the intervertebral disc clearly distinguishes the interface between the nucleus pulposus and the annulus fibrosus.

5. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 2, characterized in that, The pre-calibrated signal intensity-moisture content relationship curve mentioned in step 1.5 is obtained through in vitro experiments. The specific process is as follows: obtain in vitro intervertebral disc samples with different degrees of degeneration; measure the T2 signal intensity of each in vitro intervertebral disc sample using an MRI scanner; measure the absolute moisture content of each in vitro intervertebral disc sample using the drying and weighing method; and perform polynomial curve fitting on multiple sets of T2 signal intensity data and absolute moisture content data to obtain the signal intensity-moisture content relationship curve.

6. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 1, characterized in that, Step 2, which involves constructing and training an AI agent model based on a physical information neural network, specifically includes: Step 2.1: Construct an offline high-fidelity microscopic representative volume element finite element model library; the microscopic representative volume element finite element model library includes a microscopic representative volume element model of annulus fibrosus tissue and a microscopic representative volume element model of nucleus pulposus tissue; the microscopic representative volume element model of annulus fibrosus tissue includes alternating layers of collagen fibers and a proteoglycan matrix, and the fiber orientation of the collagen fibers is set to alternate angles between adjacent layers; the microscopic representative volume element model of nucleus pulposus tissue includes discrete chondrocytes and a hydrophilic proteoglycan network; Step 2.2: For each microscopic representative volume element finite element model in Step 2.1, apply multiple sets of random multiaxial strain loads and chemical concentration loads to its boundaries, and run high-fidelity finite element calculations; record the average stress response tensor of the microscopic representative volume element finite element model under each set of loads, as well as the statistical characteristics of the maximum principal strain distribution within the model; the statistical characteristics of the maximum principal strain distribution include the average value, standard deviation, and volume percentage exceeding a set threshold of the maximum principal strain; Step 2.3: Using the random multiaxial strain load and chemical concentration load applied in Step 2.2 as input features, and the average stress response tensor and maximum principal strain distribution statistical features calculated in Step 2.2 as output labels, a training dataset is constructed; a physical information neural network is trained using the training dataset; the loss function of the physical information neural network is composed of a weighted sum of data fitting loss and a predefined linear momentum equilibrium partial differential equation residual loss; the trained physical information neural network becomes the artificial intelligence proxy model.

7. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 1, characterized in that, The specific interaction process described in step 3, which involves calling the AI ​​proxy model to calculate the material constitutive response within each simulation time step, is as follows: Step 3.1: When the macroscopic finite element solver is performing mechanical equilibrium iterative calculations, it pauses the current iteration process when it reaches the material constitutive response at a certain Gaussian integration point. Step 3.2: Extract the current strain tensor at the Gaussian integration point and obtain the updated local chemical concentration value at the Gaussian integration point based on the two-phase porous media theory from historical simulation data; Step 3.3: Input the current strain tensor and the local chemical concentration value into the artificial intelligence proxy model; the artificial intelligence proxy model outputs a macroscopic stress tensor corresponding to the Gaussian integration point and a microscopic damage risk coefficient; Step 3.4: Return the macroscopic stress tensor to the macroscopic finite element solver, which uses the macroscopic stress tensor to continue performing equilibrium iteration calculations; at the same time, store the microscopic damage risk coefficient for subsequent analysis.

8. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 1, characterized in that, Step 4 involves processing the full-time simulation data output from step 3 to generate a dynamic biomechanical risk map of the intervertebral disc. The specific method is as follows: Step 4.1: Extract the micro-damage risk coefficient of all Gaussian integration points at each simulation time step from the full-time simulation data in Step 3; Step 4.2: For each Gaussian integral point corresponding to a spatial location in the three-dimensional geometric model, perform time-weighted accumulation of its micro-damage risk coefficients at all simulation time steps to obtain the spatial risk accumulation value; the weight of the time-weighted accumulation is the actual activity duration represented by each simulation time step; Step 4.3: Map the calculated spatial risk accumulation value of each Gaussian integral point back to the nodes of the three-dimensional geometric model of the intervertebral disc generated in Step 1 using an interpolation method; Step 4.4: Based on the spatial risk accumulation value mapped to the nodes, the three-dimensional geometric model is rendered using color gradient to generate a three-dimensional dynamic biomechanical risk map; in the dynamic biomechanical risk map, high-risk areas are displayed in warm colors and low-risk areas are displayed in cool colors.

9. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 1, characterized in that, Step 4, which involves extracting quantitative clinical indicators, specifically includes: Calculate the minimum and average values ​​of intervertebral disc height throughout the entire simulation time period, as well as the time point at which the minimum value occurs; Calculate the maximum displacement and direction of the nucleus pulposus center point in the sagittal and coronal planes; Set a cumulative threshold for the microscopic damage risk coefficient, count the volume in the annulus fibrosus region where the cumulative spatial risk value exceeds the cumulative threshold, calculate the percentage of the volume in the annulus fibrosus region to the total volume of the entire annulus fibrosus region, and record it as the high-risk volume percentage. The minimum and average values ​​of the intervertebral disc height, the time point at which the minimum value occurs, the maximum displacement and direction of displacement, and the proportion of high-risk volume together constitute the quantitative clinical indicators.

10. The AI-driven multi-scale biomechanical simulation analysis method for intervertebral discs oriented towards spinal health according to claim 8, characterized in that, The specific content of forming a structured health assessment report in step 4 is as follows: the dynamic biomechanical risk map is correlated and integrated with the quantitative clinical indicators; the structured health assessment report clearly indicates the location of the biomechanical weakness of the intervertebral disc of the target individual within the 72-hour activity cycle collected in step 1. The structured health assessment report quantifies the spinal load level of the target individual under current activity habits; based on the simulation data comparison in step 3, the structured health assessment report provides specific behavioral pattern adjustment suggestions, including adjusting specific movement postures and introducing intermittent rest periods, with the behavioral pattern adjustment suggestions aiming to reduce the cumulative load of the high-risk areas.

Images