Implant state simulation method and system based on machine learning

Abstract

The invention relates to the field of implant state simulation, and discloses an implant state simulation method and system based on machine learning, and the method comprises the following steps: data collection: collecting multi-modal data related to the implant state through an implantable sensor and an external monitoring device; processing fusion: performing feature extraction and cross-modal fusion on the multi-modal data through tensor decomposition and a graph convolutional network to obtain fusion data; and model training: inputting the fusion data into a deep learning model fusing the space-time double-flow convolutional neural network and the capsule network, and setting a variational reasoning module behind an output layer of the model. A deep learning model fusing a space-time double-flow convolutional neural network and a capsule network is constructed, tensor decomposition and a graph convolutional network are combined to realize high-order feature extraction and cross-modal fusion of multi-modal data, and a variational reasoning module is introduced to quantify model parameter uncertainty.

Description

Technical Field

[0001] The present invention relates to the technical field of implant state simulation, and in particular to an implant state simulation method and system based on machine learning. Background Art

[0002] With the widespread application of medical implants (such as orthopedic prostheses, pacemakers, neurostimulators, etc.) in clinical treatment, accurately simulating the interaction between implants and human tissues has become the key to evaluating implant safety and optimizing treatment plans.

[0003] Traditional implant state simulation methods often use finite element analysis (FEA). Finite element analysis discretizes the geometric structure of the implant and surrounding tissue, simulating the physical field distribution based on physical equations such as elasticity and fluid mechanics. However, this requires pre-setting material properties and boundary conditions, and relies on idealized assumptions (such as uniform continuous media and static loads). This makes it difficult to truly depict the dynamic complexity of the physiological environment (such as the time-varying loads caused by muscle movement and the nonlinear mechanical response of bone tissue during healing), resulting in significant simulation errors in complex scenarios. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides an implant state simulation method and system based on machine learning, which solves the problem that traditional implant state simulation methods rely on idealized assumptions, are difficult to truly portray the dynamic complexity in the physiological environment, and lead to significant simulation errors in complex scenarios.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a method for simulating implant status based on machine learning, comprising the following steps:

[0006] Data acquisition: Collect multimodal data related to implant status through implantable sensors and external monitoring devices;

[0007] Processing fusion: Through tensor decomposition and graph convolutional networks, multimodal data is subjected to feature extraction and cross-modal fusion to obtain fused data;

[0008] Model training: The fused data is input into a deep learning model that integrates a spatiotemporal two-stream convolutional neural network and a capsule network. A variational inference module is placed after the output layer of the model to approximate the posterior distribution of the deep learning model parameters. Evidence lower bound optimization is used to estimate the uncertainty of the results produced by the deep learning model at the output layer. After training, an implant state simulation model is obtained.

[0009] Simulation optimization: Use the implant state simulation model to simulate the implant state in real time, and make optimization decisions based on the simulation results based on the reinforcement learning algorithm;

[0010] Result verification: The reliability of simulation results is verified and explained through causal reasoning and interpretable deep learning techniques.

[0011] By adopting the above technical solutions, high-order feature extraction of multimodal data of implant status, joint modeling of spatiotemporal features and cross-modal deep correlation analysis are achieved. The uncertainty of the model prediction results is estimated by introducing a variational inference module, and dynamic optimization decisions of the simulation results are achieved by combining the reinforcement learning algorithm. At the same time, causal reasoning and interpretability techniques are used to improve the reliability and interpretability of the simulation results. By implicitly learning the dynamic laws of multi-field coupling in the physiological environment through machine learning technology, the interaction process between implants and tissues can be more realistically portrayed, effectively solving the problem of significant simulation errors of traditional methods in complex scenarios.

[0012] Preferably, the implantable sensors include pressure sensors, strain sensors, and biopotential sensors; the external monitoring equipment includes wearable smart bracelets, mobile ultrasound devices, and CT / MRI scanners; and the multimodal data includes physical signals of tissues surrounding the implant, physiological indicators of the patient, and medical imaging data.

[0013] Preferably, the tensor decomposition includes tensor singular value decomposition and tensor principal component analysis; the feature extraction includes treating medical imaging data as a high-order tensor, using tensor singular value decomposition for feature extraction, constructing sensor time series data as a tensor, reducing data dimension and enhancing feature expression through tensor principal component analysis; the cross-modal fusion is based on semantic association or statistical correlation of different modal features, constructing a cross-modal graph based on feature association, and using a graph convolutional network to fuse different modal features.

[0014] Preferably, the tensor singular value decomposition converts the high-order tensor X∈R corresponding to the medical image data into H×W×D×C Decomposed into X = u * s * v T , where X is the high-order tensor corresponding to the medical imaging data, R is a real number set, H is the height, W is the width, D is the depth, C is the number of channels, u and v are orthogonal tensors, s is the diagonal tensor, and the principal component analysis of the tensor is solved by The projection matrix W is obtained, and Cov(y) is the covariance tensor of the tensor y constructed by the sensor time series data, which is used to measure the correlation between features.

[0015] Preferably, in the deep learning model, the spatial stream network adopts an improved ResNet-50 architecture and introduces void convolution, the temporal stream network is constructed based on a gated recurrent unit and adds an attention mechanism, the feature map output by the spatial stream network is divided into several capsules, the connection weights between different capsules are determined by a dynamic routing algorithm, the spatial features are expressed, and the spatial features output by the capsule network are fused with the temporal stream features.

[0016] Preferably, in the dynamic routing algorithm, the initial prediction vector u ji =W ij s i , where u ji is the initial prediction vector, W ij is the learnable weight matrix of capsule j to i, s i is the feature vector of the input capsule, and the coupling coefficient is updated iteratively Among them, c ij is the normalized coupling coefficient, b ij is the logarithmic prior probability, k is the iteration index, and the output vector is obtained Among them, v j is the feature vector of the output capsule, i is the input capsule index, u j|i is the prediction vector of capsule i for capsule j.

[0017] Preferably, the variational inference module approximates the true posterior distribution p(w|D) by introducing the variational distribution q(w|θ), and uses the evidence lower bound L(θ)=Eq(w|θ)[logp(D|w)]-KL(q(w|θ)||p(w)) for optimization to estimate the uncertainty of the model prediction results, wherein w is the deep learning model parameter, D is the training data, θ is the parameter of the variational distribution, Eq(w|θ)[logp(D|w)] is the expectation of the log-likelihood of the data, which is used to measure the model's ability to fit the data, and KL(q(w|θ)||p(w)) is the KL divergence, which is used to measure the difference between the variational distribution and the prior distribution p(w).

[0018] Preferably, the optimization decision includes decomposing the decision problem into a strategy layer and an execution layer, wherein the strategy layer formulates a long-term treatment strategy, and the execution layer adjusts implant-related parameters according to the long-term treatment strategy, including the intensity of the neurostimulator and the drug release rate. The strategy layer and the execution layer communicate and collaborate through a shared state space and action space to jointly optimize the global decision.

[0019] Preferably, the reliability verification and interpretation is to construct a causal graph model based on domain knowledge or data-driven, analyze the causal effects of different factors on the implant state through counterfactual reasoning, compare the actual and counterfactual simulation results to evaluate the rationality of the prediction, use the SHAP value or LIME algorithm to calculate the contribution of each input feature to the prediction result, and generate a visual heat map or feature importance ranking report.

[0020] A machine learning-based implant state simulation system, used in the above-mentioned machine learning-based implant state simulation method, comprising:

[0021] Data acquisition module: used to collect multimodal data related to the implant status through implantable sensors and external monitoring devices;

[0022] Processing and fusion module: used to extract features and perform cross-modal fusion of multimodal data through tensor decomposition and graph convolutional network to obtain fused data;

[0023] Model training module: This module is used to input the fused data into a deep learning model that integrates a spatiotemporal dual-stream convolutional neural network and a capsule network. A variational inference module is set after the output layer of the model to approximate the posterior distribution of the deep learning model parameters. Evidence lower bound optimization is used to estimate the uncertainty of the results produced by the deep learning model at the output layer. After training, an implant state simulation model is obtained.

[0024] Simulation optimization module: used to simulate the implant status in real time using the implant status simulation model and make optimization decisions based on the simulation results based on the reinforcement learning algorithm;

[0025] Result Verification Module: Used to verify and explain the reliability of simulation results through causal reasoning and interpretable deep learning techniques.

[0026] The present invention provides a method and system for simulating implant status based on machine learning. It has the following beneficial effects:

[0027] 1. The present invention constructs a deep learning model that integrates spatiotemporal dual-stream convolutional neural networks and capsule networks, combines tensor decomposition and graph convolutional networks to achieve high-order feature extraction and cross-modal fusion of multimodal data, and introduces a variational reasoning module to quantify the uncertainty of model parameters. The spatiotemporal dual-stream network captures the spatial structural characteristics of medical images and the temporal dynamic characteristics of sensor data respectively. The capsule network models the hierarchical relationship of features through a dynamic routing mechanism to implicitly reflect physical constraints. Tensor decomposition and graph convolutional networks enhance the high-order feature expression and cross-modal association modeling of multimodal data. The variational reasoning module further provides uncertainty estimation for the simulation results, which can more accurately adapt to the complex dynamic changes of implants in real physiological environments, and effectively solves the problem of significant simulation errors of traditional methods in complex scenarios.

[0028] 2. The present invention uses tensor decomposition and graph convolutional networks to perform feature extraction and cross-modal fusion of multimodal data, which can effectively capture the high-order spatial features of medical images and the temporal dynamic features of sensor data, realize deep correlation modeling of multi-dimensional data, enable the model to more comprehensively reflect the complex characteristics of the implant status, and enhance the expressive power and fusion depth of data features.

[0029] 3. The present invention extracts the spatial structural features of the implant through the spatial flow network, captures the dynamic timing features through the temporal flow network, and uses the dynamic routing algorithm of the capsule network to realize the modeling of the feature hierarchy. It can perform a spatiotemporal joint analysis of the implant state, implicitly learn the physical field coupling law, and improve the simulation accuracy of the dynamic behavior of the implant in a complex physiological environment.

[0030] 4. The present invention optimizes the simulation results based on the reinforcement learning algorithm, hierarchizes the decision-making problems and introduces a multi-agent collaborative and adversarial training mechanism. It can dynamically adjust implant-related parameters and treatment strategies according to real-time simulation results, thereby improving the clinical applicability of the simulation results and the robustness of the decision-making strategy, realizing an intelligent closed loop from simulation to decision-making, and adapting to the dynamic changes in the implant status.

[0031] 5. The present invention verifies and explains the reliability of simulation results through causal reasoning and interpretable deep learning technology. It can analyze the causal impact of different factors on the implant status, quantify the importance of features and generate visual explanation reports, converting complex model decisions into understandable clinical indicators, thereby improving the trust in simulation results and the operability of clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a method for simulating implant status based on machine learning proposed by the present invention;

[0033] Figure 2 This is a system architecture diagram of the machine learning-based implant state simulation system proposed in the present invention. DETAILED DESCRIPTION

[0034] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0035] Please see the attached Figure 1 , an embodiment of the present invention provides an implant state simulation method based on machine learning, comprising the following steps:

[0036] Data acquisition: Multimodal data related to the implant status is collected through implantable sensors and external monitoring equipment; implantable sensors include pressure sensors, strain sensors, and biopotential sensors, and external monitoring equipment includes wearable smart bracelets, mobile ultrasound machines, and CT / MRI scanners. Multimodal data includes physical signals of tissues around the implant, patient physiological indicators, and medical imaging data.

[0037] Specifically, the data collection process uses diversified equipment to achieve the simultaneous acquisition of multi-dimensional data. Implantable sensors are packaged with biocompatible materials. For example, MEMS pressure sensors with titanium alloy shells can be directly integrated into the surface of artificial joints to monitor the contact pressure distribution during joint movement in real time at a sampling frequency of 100Hz. Flexible strain sensors are attached to the implant-tissue interface with biological glue to capture tiny deformation signals. In terms of external monitoring equipment, wearable smart bracelets use photoelectric sensors to collect physiological indicators such as heart rate variability (HRV) and skin conductance in real time to assess the patient's overall stress response. Mobile ultrasound machines are equipped with high-frequency probes to perform real-time scanning of soft tissues around implants and obtain two-dimensional ultrasound images. Regular three-dimensional tomography scans are performed using 128-row spiral CT to generate image data of the implant-bone fusion area with a resolution of 0.5mm, forming a multimodal data set containing physical signals, physiological indicators and image data.

[0038] Processing fusion: Multimodal data is subjected to feature extraction and cross-modal fusion through tensor decomposition and graph convolutional networks to obtain fused data; tensor decomposition includes tensor singular value decomposition and tensor principal component analysis; feature extraction includes treating medical imaging data as high-order tensors, using tensor singular value decomposition for feature extraction, constructing sensor time series data as tensors, reducing data dimensions and enhancing feature expression through tensor principal component analysis, cross-modal fusion is based on semantic association or statistical correlation of different modal features, constructing a cross-modal graph based on feature association, and using graph convolutional networks to fuse different modal features.

[0039] Tensor singular value decomposition converts the high-order tensor X∈R corresponding to the medical imaging data H×W×D×C Decomposed into X = u * s * v T , where X is the high-order tensor corresponding to the medical imaging data, R is a set of real numbers, H is the height, W is the width, D is the depth, C is the number of channels, u and v are orthogonal tensors, s is the diagonal tensor, and the principal component analysis of the tensor is solved by The projection matrix W is obtained, and Cov(y) is the covariance tensor of the tensor y constructed by the sensor time series data, which is used to measure the correlation between features.

[0040] Specifically, the processing and fusion of multimodal data consists of two stages: feature extraction and cross-modal correlation modeling. For CT / MRI image data, the region of interest (ROI) of the implant and surrounding tissue is first extracted using threshold segmentation and region growing algorithms. This ROI is then converted into a four-dimensional tensor X (spatial dimensions H × W × D, with channel dimensions C corresponding to CT values ​​or MRI signal intensity). This is then decomposed into a core tensor s and orthogonal basis tensors u and v using tensor singular value decomposition (T-SVD). The principal elements of the diagonal tensor s correspond to the characteristic energy at different spatial frequencies. Retaining the top 20% of singular values ​​captures over 90% of the image information, achieving data compression and feature enhancement. For sensor time series data (such as pressure, strain, and HRV), a three-dimensional tensor (time dimension T × feature dimension F × number of channels 1) is constructed. The covariance tensor Cov(y) is calculated using tensor principal component analysis (T-PCA). The projection matrix W is then solved using iterative singular value decomposition (ISVD) to ensure dimensionality reduction of the real-time data, compressing the feature dimensions while preserving information entropy.

[0041] During the cross-modal fusion phase, a cross-modal graph is constructed based on the Pearson correlation coefficient between imaging features (such as bone density and implant surface roughness) and sensor features (such as peak stress and strain rate). Nodes are the principal component eigenvectors of each modality, and edge weights are set to the absolute values ​​of the correlation coefficients. A two-layer graph convolutional network (GCN) is used for feature fusion. The first layer has a 3×3 convolution kernel size and 64 output channels, while the second layer has 32 output channels. A nonlinear transformation is introduced using the ReLU activation function, ultimately generating a fused feature vector that contains spatiotemporal correlation information.

[0042] Model training: The fused data is input into a deep learning model that integrates a spatiotemporal dual-stream convolutional neural network and a capsule network. A variational inference module is set after the output layer of the model to approximate the posterior distribution of the deep learning model parameters. The evidence lower bound optimization is used to estimate the uncertainty of the results produced by the deep learning model in the output layer. After training, an implant state simulation model is obtained; in the deep learning model, the spatial stream network adopts an improved ResNet-50 architecture and introduces void convolution. The temporal stream network is constructed based on a gated recurrent unit and adds an attention mechanism. The feature map output by the spatial stream network is divided into several capsules. The connection weights between different capsules are determined by a dynamic routing algorithm, the spatial features are expressed, and the spatial features output by the capsule network are fused with the temporal stream features.

[0043] In the dynamic routing algorithm, the initial prediction vector u ji =W ij s i , where u ji is the initial prediction vector, W ij is the learnable weight matrix of capsule j to i, si is the feature vector of the input capsule, and the coupling coefficient is updated iteratively Among them, c ij is the normalized coupling coefficient, b ij is the logarithmic prior probability, k is the iteration index, and the output vector is obtained Among them, v j is the feature vector of the output capsule, i is the input capsule index, u j|i is the prediction vector of capsule i for capsule j.

[0044] The variational inference module approximates the true posterior distribution p(w|D) by introducing the variational distribution q(w|θ) and uses the evidence lower bound L(θ) = Eq(w|θ)[logp(D|w)]-KL(q(w|θ)||p(w)) for optimization to estimate the uncertainty of the model prediction results, where w is the deep learning model parameter, D is the training data, θ is the parameter of the variational distribution, Eq(w|θ)[logp(D|w)] is the expectation of the log-likelihood of the data, which is used to measure the model's ability to fit the data, and KL(q(w|θ)||p(w)) is the KL divergence, which is used to measure the difference between the variational distribution and the prior distribution p(w).

[0045] Specifically, in the model training step, the present invention achieves high-precision modeling and uncertainty quantification of implant status by fusing the spatiotemporal dual-stream convolutional neural network (ST-DCNN) and the capsule network (CapsNet) and integrating the variational inference module.

[0046] First, the processed and fused multimodal data is fed into the spatiotemporal dual-stream network of the deep learning model. The spatial stream network uses a modified ResNet-50 architecture. Its input is the ROI region of interest (ROI) of a 3D medical image (e.g., the implant-bone fusion region in a CT scan). Multiple convolutional layers (with dilated convolutions of 2 inserted in the first three layers) extract spatial features and output a 512×32×32 feature map containing detailed information such as the implant geometry and tissue boundaries. The temporal stream network, based on a bidirectional gated recurrent unit (Bi-GRU), takes sensor time series data (e.g., dynamic load sequences acquired by a pressure sensor) as input. Two Bi-GRU layers (hidden dimension 128) capture feature associations along the temporal dimension and output a feature vector containing temporal dependencies.

[0047] Subsequently, the feature map output by the spatial stream is divided into 8×8 capsules (each capsule corresponds to a 4×4 region in the 32×32 feature map), and the input of each capsule is the feature vector s of the region i (i is the input capsule index). The dynamic routing algorithm determines the connection weights between capsules through iterative calculation at this stage: the initial prediction vector u jiBy input capsule feature s i With the learnable weight matrix W ij (characterizing the connection strength from capsule i to j) is multiplied to obtain ij Optimized by back propagation algorithm. By logarithmic prior probability b ij (Initial value is 0) Iteratively update the coupling coefficient c ij , the formula is Where k is the iteration index, usually 3 iterations are performed. The final output capsule feature vector v j The initial prediction vector is obtained by weighted summation and normalization, that is, Through dynamic routing, the model can automatically identify the hierarchical structure of spatial features (such as the association between implant surface texture and stress concentration areas), enhancing the ability to express complex spatial relationships.

[0048] In the dual-stream network fusion stage, the temporal feature vector output by the time stream and the spatial feature vector output by the capsule network are fused through a splicing operation to form a composite feature representation containing spatiotemporal information, which is input to the fully connected layer for preliminary state prediction (such as stress distribution and loosening probability).

[0049] To quantify the uncertainty of model prediction, a variational inference module is set after the output layer. This module introduces the variational distribution q(w|θ) to approximate the model parameters w (including convolution kernel weights, GRU parameters, capsule connection weights W ij The true posterior distribution p(w|D) of the prior distribution (e.g., D is the training dataset and θ is the parameter of the variational distribution (e.g., mean μ and variance σ of a Gaussian distribution). θ is optimized using the evidence lower bound (ELBO) L(θ) = Eq(w|θ)[logp(D|w)] - KL(q(w|θ)||p(w)), where Eq(w|θ)[logp(D|w)] measures the model's ability to fit the training data, and KL(q(w|θ)||p(w)) measures the difference between the variational distribution and the prior distribution p(w). The stochastic gradient estimate is converted into a differentiable operation through reparameterization techniques, and the parameters are iteratively updated using the Adam optimizer. Ultimately, the model can output predictions that include uncertainty estimates (e.g., confidence intervals for stress values ​​and probability distributions for loosening probabilities), providing a reliable basis for clinical decision-making.

[0050] After the above process, the model training phase uses a dynamic routing algorithm to enhance the hierarchical expression of spatiotemporal features, combines the variational inference module to achieve uncertainty quantification, and finally generates an implant state simulation model that can accurately simulate the implant state and provide reliability assessment.

[0051] Simulation optimization: The implant state simulation model is used to simulate the implant state in real time, and the simulation results are optimized based on the reinforcement learning algorithm. The optimization decision-making includes decomposing the decision-making problem into a strategy layer and an execution layer. The strategy layer formulates a long-term treatment strategy, and the execution layer adjusts the implant-related parameters according to the long-term treatment strategy, including the intensity of the neurostimulator and the drug release rate. The strategy layer and the execution layer communicate and collaborate through a shared state space and action space to jointly optimize the global decision.

[0052] Specifically, the simulation optimization module dynamically generates decision-making strategies based on a hierarchical reinforcement learning (HRL) framework. The upper-level PolicyAgent is responsible for setting long-term treatment goals. For example, in an orthopedic implant scenario, the goal is to achieve a 15% increase in bone density at the bone-implant interface within six months. The Temporal Difference Learning (TD-Learning) algorithm generates monthly stress stimulation threshold intervals. The lower-level ExecutionAgent adjusts specific parameters within the policy layer's defined range based on real-time simulation results. For example, it dynamically adjusts the knee prosthesis's load feedback threshold from 300N to 250N. It also pushes personalized exercise plans (such as 10-minute knee flexion and extension exercises three times a day) to patients via wearable devices.

[0053] Multi-agent collaboration is achieved through a shared experience pool (ExperienceReplayBuffer). The action-state pairs of the strategy layer and the execution layer are represented by tuples (s, a p ,a e ,r) form, where a p is the policy-level action (such as target threshold adjustment), a e To perform layer actions (such as parameter fine-tuning), the adversarial training mechanism introduces an interference agent (AdversaryAgent), which tests the robustness of the decision agent by generating adversarial samples (such as simulated sensor noise and image artifacts). When the prediction error exceeds the preset threshold (such as the loose probability error > 10%), the model online update mechanism is triggered, and the feedforward network is quickly fine-tuned for 5 rounds (learning rate 1e -3 ), ensuring reliability under abnormal working conditions.

[0054] Results Verification: The simulation results are validated and explained using causal reasoning and interpretable deep learning techniques. This involves constructing a causal graph model based on domain knowledge or data-driven analysis, analyzing the causal effects of different factors on the implant state through counterfactual reasoning, and comparing actual and counterfactual simulation results to assess the rationality of the predictions. The SHAP value or LIME algorithm is used to calculate the contribution of each input feature to the prediction, generating a visual heat map or feature importance ranking report.

[0055] Specifically, the reliability verification link combines causal reasoning and explainability technology to provide dual guarantees for clinical decision-making. Causal analysis constructs a causal graph model based on the Do-calculus algorithm. For example, when analyzing the risk of implant loosening, "interface stress", "bone metabolism index" and "patient activity level" are determined as direct causal variables. Intervention operations (Do (interface stress>400N)) are used to simulate the loosening probability under high-stress scenarios. The actual data is compared with the counterfactual results. If the relative risk increases by more than 20%, the model prediction is determined to be valid. Interpretability analysis uses SHAP values ​​to calculate the marginal contribution of each input feature. For example, in CT image features, the SHAP value of the implant-bone contact area accounts for 35%, indicating that this feature is a key factor affecting stability. A heat map is generated and superimposed on the original image to intuitively display high-risk areas.

[0056] By constructing a deep learning model that integrates a spatiotemporal dual-stream convolutional neural network and a capsule network, combined with tensor decomposition and graph convolutional networks to extract high-order features from multimodal data and integrate them across modalities, a variational inference module is introduced to quantify model parameter uncertainty. This enables joint modeling of spatiotemporal features of implant states, deep cross-modal correlation analysis, and reliability assessment of prediction results. This addresses the problem of traditional implant state simulation methods relying on idealized physical assumptions and failing to accurately depict the dynamic complexity of multi-field coupling in physiological environments. This effectively improves the model's simulation accuracy and clinical applicability for implant-tissue interaction states. The spatiotemporal dual-stream network captures the spatial structural characteristics of medical images and the temporal dynamic characteristics of sensor data, respectively. The capsule network implicitly reflects physical constraints by modeling hierarchical relationships through a dynamic routing mechanism. Tensor decomposition and graph convolutional networks enhance the high-order feature representation of multimodal data and model cross-modal correlations. The variational inference module further provides uncertainty estimates for the simulation results, enabling the model to more accurately adapt to the complex dynamic changes of implants in real physiological environments, providing a reliable basis for clinical evaluation and treatment optimization.

[0057] Please see the attached Figure 2 A machine learning-based implant state simulation system, used for the above-mentioned machine learning-based implant state simulation method, comprising:

[0058] Data acquisition module: used to collect multimodal data related to the implant status through implantable sensors and external monitoring devices;

[0059] Processing and fusion module: used to extract features and perform cross-modal fusion of multimodal data through tensor decomposition and graph convolutional network to obtain fused data;

[0060] Model training module: This module is used to input the fused data into a deep learning model that integrates a spatiotemporal dual-stream convolutional neural network and a capsule network. A variational inference module is set after the output layer of the model to approximate the posterior distribution of the deep learning model parameters. Evidence lower bound optimization is used to estimate the uncertainty of the results produced by the deep learning model at the output layer. After training, an implant state simulation model is obtained.

[0061] Simulation optimization module: used to simulate the implant status in real time using the implant status simulation model and make optimization decisions based on the simulation results based on the reinforcement learning algorithm;

[0062] Result Verification Module: Used to verify and explain the reliability of simulation results through causal reasoning and interpretable deep learning techniques.

[0063] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for simulating implant status based on machine learning, characterized in that: The following steps are involved: Data acquisition: Collect multimodal data related to implant status through implantable sensors and external monitoring devices; Processing fusion: Through tensor decomposition and graph convolutional networks, multimodal data is subjected to feature extraction and cross-modal fusion to obtain fused data; Model training: The fused data is input into a deep learning model that integrates a spatiotemporal two-stream convolutional neural network and a capsule network. A variational inference module is placed after the output layer of the model to approximate the posterior distribution of the deep learning model parameters. Evidence lower bound optimization is used to estimate the uncertainty of the results produced by the deep learning model at the output layer. After training, an implant state simulation model is obtained. Simulation optimization: Use the implant state simulation model to simulate the implant state in real time, and make optimization decisions based on the simulation results based on the reinforcement learning algorithm; Result verification: The reliability of simulation results is verified and explained through causal reasoning and interpretable deep learning techniques.

2. The implant state simulation method based on machine learning according to claim 1, characterized in that: The implantable sensors include pressure sensors, strain sensors, and biopotential sensors; the external monitoring devices include wearable smart bracelets, mobile ultrasound devices, and CT / MRI scanners; and the multimodal data include physical signals of tissues surrounding the implant, physiological indicators of the patient, and medical imaging data.

3. The implant state simulation method based on machine learning according to claim 1, characterized in that: The tensor decomposition includes tensor singular value decomposition and tensor principal component analysis. The feature extraction includes treating medical imaging data as a high-order tensor, using tensor singular value decomposition for feature extraction, constructing sensor time series data as a tensor, reducing data dimensions and enhancing feature expression through tensor principal component analysis. The cross-modal fusion is based on semantic association or statistical correlation of different modal features, constructing a cross-modal graph based on feature association, and using a graph convolutional network to fuse different modal features.

4. The implant state simulation method based on machine learning according to claim 3, characterized in that: The tensor singular value decomposition converts the high-order tensor X∈R corresponding to the medical imaging data H×W×D×C Decomposed into X = u * s * v T , where X is the high-order tensor corresponding to the medical imaging data, R is a real number set, H is the height, W is the width, D is the depth, C is the number of channels, u and v are orthogonal tensors, s is the diagonal tensor, and the principal component analysis of the tensor is solved by The projection matrix W is obtained, and Cov(y) is the covariance tensor of the tensor y constructed by the sensor time series data, which is used to measure the correlation between features.

5. The implant state simulation method based on machine learning according to claim 1, characterized in that: In the deep learning model, the spatial stream network adopts an improved ResNet-50 architecture and introduces void convolution. The temporal stream network is built based on the gated recurrent unit and adds an attention mechanism. The feature map output by the spatial stream network is divided into several capsules. The connection weights between different capsules are determined by a dynamic routing algorithm to express spatial features. The spatial features output by the capsule network are then fused with the temporal stream features.

6. The implant state simulation method based on machine learning according to claim 5, characterized in that: In the dynamic routing algorithm, the initial prediction vector u ji =W ij s i , where u ji is the initial prediction vector, W ij is the learnable weight matrix of capsule j to i, s i is the characteristic vector of the input capsule, and the coupling coefficient c is updated iteratively ij : Among them, c ij is the normalized coupling coefficient, b ij is the logarithmic prior probability, k is the iteration index, and the output vector is obtained Among them, v j is the feature vector of the output capsule, i is the input capsule index, u j|i is the prediction vector of capsule i for capsule j.

7. The implant state simulation method based on machine learning according to claim 1, characterized in that: The variational inference module approximates the true posterior distribution p(w|D) by introducing the variational distribution q(w|θ), and uses the evidence lower bound L(θ)=Eq(w|θ)[logp(D|w)]-KL(q(w|θ)||p(w)) for optimization to estimate the uncertainty of the model prediction results, where w is the deep learning model parameter, D is the training data, θ is the parameter of the variational distribution, Eq(w|θ)[logp(D|w)] is the expectation of the log-likelihood of the data, which is used to measure the model's ability to fit the data, and KL(q(w|θ)||p(w)) is the KL divergence, which is used to measure the difference between the variational distribution and the prior distribution p(w).

8. The implant state simulation method based on machine learning according to claim 1, characterized in that: The optimization decision-making includes decomposing the decision-making problem into a strategy layer and an execution layer. The strategy layer formulates a long-term treatment strategy, and the execution layer adjusts implant-related parameters according to the long-term treatment strategy, including the intensity of the neurostimulator and the drug release rate. The strategy layer and the execution layer communicate and collaborate through a shared state space and action space to jointly optimize the global decision.

9. The implant state simulation method based on machine learning according to claim 1, characterized in that: The reliability verification and interpretation described above is to construct a causal graph model based on domain knowledge or data-driven, analyze the causal effects of different factors on the implant status through counterfactual reasoning, compare the actual and counterfactual simulation results to evaluate the rationality of the prediction, use the SHAP value or LIME algorithm to calculate the contribution of each input feature to the prediction result, and generate a visual heat map or feature importance ranking report.

10. An implant state simulation system based on machine learning, characterized by: A method for simulating an implant state based on machine learning according to any one of claims 1 to 9, comprising: Data acquisition module: used to collect multimodal data related to the implant status through implantable sensors and external monitoring devices; Processing and fusion module: used to extract features and perform cross-modal fusion of multimodal data through tensor decomposition and graph convolutional network to obtain fused data; Model training module: This module is used to input the fused data into a deep learning model that integrates a spatiotemporal dual-stream convolutional neural network and a capsule network. A variational inference module is set after the output layer of the model to approximate the posterior distribution of the deep learning model parameters. Evidence lower bound optimization is used to estimate the uncertainty of the results produced by the deep learning model at the output layer. After training, an implant state simulation model is obtained. Simulation optimization module: used to simulate the implant status in real time using the implant status simulation model and make optimization decisions based on the simulation results based on the reinforcement learning algorithm; Result Verification Module: Used to verify and explain the reliability of simulation results through causal reasoning and interpretable deep learning techniques.

Images