Dental implant data monitoring analysis method based on deep learning

Abstract

The application provides a dental implant data monitoring analysis method based on deep learning, which realizes implant failure risk scoring and local causal correlation strength extraction by collecting preoperative images, clinical examination, surgical parameters and other multi-modal structured data, standardizing preprocessing and inputting a deep neural network model with a double-branch structure; based on the model output, a dynamic individualized causal graph is constructed, key clinical intervention variables are used as regulation nodes, and counterfactual disturbance simulation is carried out to quantify variable intervention sensitivity and analyze key causal paths; a priority intervention suggestion list and an auxiliary decision report are generated, and the causal graph can be self-adaptively optimized according to actual follow-up feedback to realize long-term closed-loop evolution; the application improves the accuracy of implant failure risk prediction and the pertinence and operability of intervention schemes, and enhances clinical individualized treatment decision support.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence medical data analysis and causal reasoning, and particularly relates to a dental implant data monitoring analysis method based on deep learning. BACKGROUND

[0002] Current dental implant risk prediction and clinical decision support systems mainly rely on deep learning technology and traditional statistical analysis methods. In recent years, the application of artificial intelligence in the field of dentistry has been continuously developed. The mainstream technical solution takes deep neural network as the core, inputs multi-modal structured data (such as preoperative imaging features, clinical examination results and surgical parameters, etc.), and after model training, can make a certain degree of prediction on the probability of implant failure. Part of the commercialized clinical auxiliary diagnosis platform has realized the quantitative output of the individual implant prognosis risk of patients, which has promoted the preliminary landing of precision medicine in the field of oral implantation;

[0003] The development trend of the prior art emphasizes the complexity of the model structure and the improvement of the prediction accuracy, adopts multi-branch network, feature fusion and end-to-end learning process, and aims to fully excavate the joint discrimination ability of multi-source data such as bone density, alveolar bone height, intraoperative operation index and occlusal force mode. Part of the technical scheme tries to introduce limited attribution mechanism, such as indirectly analyzing the model judgment basis through feature importance sorting, recursive ablation or with the help of part of the explainable algorithm (such as SHAP value, LIME method, etc.) to support the auxiliary decision of doctors in the implant failure risk evaluation. However, the mainstream scheme widely used in the industry at present generally has the problems of insufficient explanation and limited clinical operability;

[0004] In the field of artificial intelligence medical data analysis and causal reasoning, part of the research tries to use causal inference method to study the mechanism behind the implant failure, mainly relying on static causal structure or propensity score for intervention effect evaluation, but subject to the stability of causal structure hypothesis and data availability, the attribution ability and individualized intervention suggestion in actual clinical scene are still weak. The general causal inference technology has not realized the deep integration with deep learning risk prediction, causing the prediction system to be difficult to give specific intervention scheme or clear mechanism explanation. SUMMARY

[0005] The present application provides a dental implant data monitoring analysis method based on deep learning to solve the above technical problems.

[0006] The technical scheme of the present application is implemented as follows: the dental implant data monitoring analysis method based on deep learning comprises:

[0007] S1: Collect preoperative imaging data, clinical examination records and surgical operation parameters of patients, form a multi-modal structured input data set containing bone density, alveolar bone height, implant position, occlusal force distribution and postoperative infection indicators;

[0008] S2: Standardized preprocessing of the multi-modal structured input data set, including missing value interpolation, numerical normalization and classification variable encoding, generating a uniform dimension feature vector to eliminate the bias caused by different measurement scales and data types;

[0009] S3: The uniform dimension feature vector is input into a double-branch deep neural network model, wherein the main branch performs an implant failure risk prediction task, and outputs an initial risk score; the auxiliary branch uses the attention gate mechanism to filter the potential confounding factors strongly related to the anchor point based on the set of interventionable clinical variables, and generates a local causal correlation strength matrix;

[0010] S4: Based on the local causal correlation strength matrix and the initial risk score, an individualized dynamic causal graph is constructed, wherein the nodes represent the clinical variables, the directed edges represent the causal influence direction and strength between the variables, and the weights are initialized by the Bayesian update rule combined with the historical group data, and the interventionable variables are marked as key regulatory nodes;

[0011] S5: In the individualized dynamic causal graph, the counterfactual disturbance simulation is applied to the key regulatory nodes, the predicted output gradient response caused by the change of each interventionable variable under virtual intervention is calculated, and a variable intervention sensitivity sequence is generated as a causal effect quantification index;

[0012] S6: Based on the variable intervention sensitivity sequence, the path in the individualized dynamic causal graph is executed attribution analysis, the key causal propagation path from the interventionable variable to the high-risk output is identified, and the priority intervention suggestion list is generated according to the intervention expected effect ranking;

[0013] S7: Output a comprehensive decision support report containing the initial risk score, the priority intervention suggestion list, the visual causal path graph, and the expected effect estimation and its uncertainty interval of each intervention measure, for reference by the clinician;

[0014] S8: Update the edge weight in the individualized dynamic causal graph according to the actual intervention result of the subsequent follow-up feedback, optimize the long-term adaptability of the model to the causal structure by using the incremental learning mechanism, and realize the closed-loop evolution of the causal reasoning system.

[0015] The dental implant data monitoring and analysis method based on deep learning provided by the application has the following beneficial effects:

[0016] (1) The application realizes the paradigm transition from static probability output to causal mechanism analysis by constructing a dynamic evolving individualized causal graph mechanism, introduces a "intervenable variable guide" mechanism, takes factors with clear clinical operation path such as bone mineral density, bite force distribution and postoperative infection control as anchor points, drives the discovery and weight update of causal correlation, and ensures that the constructed causal graph not only reflects the statistical dependence between variables, but also focuses on the functional pathway with actual intervention significance. Combined with counterfactual disturbance simulation and Bayesian update strategy, the system can simulate the risk change trend under different intervention scenarios in the hidden space, dynamically calibrate the strength and direction of individual causal path, and significantly improve the physiological reasonableness and individual adaptability of the attribution result;

[0017] (2) On the basis of realizing the deep integration of prediction and attribution, the application further designs a double-branch neural network architecture and a path sensitivity sorting mechanism, realizes the integrated closed-loop output of risk warning and intervention suggestion generation, and greatly enhances the practicality and response ability of the clinical decision support system. The main branch focuses on the high-precision risk classification task, and ensures that the prediction performance is not sacrificed due to the demand for explainability; the auxiliary branch uses the attention gate mechanism to automatically screen the key confounding factors affecting the intervenable variables, and preliminarily establishes the local causal correlation, providing structural prior for subsequent refined attribution. Through counterfactual gradient analysis, the system can quantify the influence amplitude of each potential intervention measure on the final prediction result, and generate an intervention suggestion list with priority ranking accordingly, so that doctors can quickly identify the most cost-effective regulation path. At the same time, the output includes a visual causal path diagram and an uncertainty estimate of each intervention effect, which not only improves the model transparency, but also provides an intuitive basis for doctor-patient communication, and effectively enhances the credibility and adoption willingness of artificial intelligence systems in real diagnosis and treatment processes;

[0018] (3) The application discards the dependence on large-scale unlabeled data or strong hypothesis premise, and proposes a lightweight and highly cohesive causal reasoning framework with stronger deployment flexibility and clinical adaptability. The whole attribution process is closely related to the operable variables, avoiding the causal inference bias and computational redundancy caused by the introduction of too many uncontrollable exogenous variables, and significantly reducing the resource consumption in the model training and reasoning stage. In addition, the architecture naturally supports incremental learning and online updating, and as new case data is continuously accumulated, the individual causal graph can be smoothly iteratively optimized through Bayesian rules, forming a closed-loop adaptive intelligent decision support system. The overall scheme can stably run in the environment of a regular medical information system without complex parameter tuning, has good scalability and cross-institution migration potential, and is especially suitable for oral implantation, which is highly dependent on individualized evaluation and precise intervention. The application provides a new technical paradigm for constructing an explainable, intervenable and sustainable evolving medical AI system. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 This is a flowchart of the deep learning-based dental implant data monitoring and analysis method of the present invention;

[0020] Figure 2 This is a sub-flowchart of the deep learning-based dental implant data monitoring and analysis method of the present invention;

[0021] Figure 3 This is another sub-flowchart of the deep learning-based dental implant data monitoring and analysis method of the present invention. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0024] like Figure 1 As shown, this invention provides a method for monitoring and analyzing dental implant data based on deep learning, specifically including:

[0025] S1: Collect patients' preoperative imaging data, clinical examination records and surgical operation parameters to form a multimodal structured input dataset that includes bone density, alveolar bone height, implant position, occlusal force distribution and postoperative infection indicators;

[0026] S2: Perform standardized preprocessing on the multimodal structured input dataset, including missing value imputation, numerical normalization and categorical variable encoding, to generate a unified dimension feature vector to eliminate bias caused by different measurement scales and data types;

[0027] S3: Input the unified dimension feature vector into a dual-branch deep neural network model, where the main branch performs the implant failure risk prediction task and outputs an initial risk score; the auxiliary branch uses the set of interventionable clinical variables as anchors and uses an attention gating mechanism to screen for potential confounding factors that are strongly correlated with them, and generates a local causal association strength matrix.

[0028] S4: Based on the local causal association strength matrix and the initial risk score, construct an individualized dynamic causal graph, where nodes represent clinical variables, directed edges represent the direction and strength of causal influence between variables, their weights are initialized by combining historical population data with Bayesian update rules, and interveneable variables are marked as key regulatory nodes.

[0029] S5: Apply counterfactual perturbation simulation to key control nodes in the individualized dynamic causal graph, calculate the predicted output gradient response caused by changes in each intrusive variable under virtual intervention, and generate a variable intervention sensitivity sequence as a quantitative indicator of causal effect.

[0030] S6: Based on the variable intervention sensitivity sequence, perform attribution analysis on the paths in the individualized dynamic causal graph, identify key causal propagation paths from the modifiable variable to the high-risk output, and generate a priority intervention recommendation list according to the expected intervention effect;

[0031] S7: Output a comprehensive decision support report including initial risk score, priority intervention recommendation list, visualized causal path diagram, and estimated expected effects of each intervention and their uncertainty intervals, for clinicians' reference;

[0032] S8: Update the edge weights in the individualized dynamic causal graph based on the actual intervention results from subsequent follow-up feedback, and use the incremental learning mechanism to optimize the model's long-term adaptability to the causal structure, thereby achieving the closed-loop evolution of the causal reasoning system.

[0033] Step S1: Collect preoperative imaging data, clinical examination records, and surgical parameters of the patient to form a multimodal structured input dataset containing bone mineral density, alveolar bone height, implant position, occlusal force distribution, and postoperative infection indicators. Specifically, this includes:

[0034] S1.1: Obtain the patient's preoperative cone-beam computed tomography (CBCT) image data, extract bone density distribution map and alveolar bone three-dimensional morphological parameters based on voxel intensity threshold segmentation algorithm, and generate a quantitative bone structure feature matrix including cortical bone thickness, mean cancellous bone density and bone mass grade to characterize the biomechanical support capacity of bone tissue in the implant area.

[0035] The input data includes raw three-dimensional alveolar bone image data acquired by a preoperative cone-beam computed tomography (CBCT) scanning system, stored in DICOM format, and containing grayscale intensity information and spatial coordinate metadata for each voxel.

[0036] A voxel intensity threshold segmentation algorithm (parameter: the threshold range is set to 350 to 1500 HU based on the linear mapping curve between gray level and bone density) is used to separate high-density bone tissue regions.

[0037] Furthermore, noise removal of the segmentation results is achieved through three-dimensional connected component analysis (parameter: minimum connected component volume is set to 0.5 mm³), and continuous bone tissue volume data is obtained.

[0038] Furthermore, the Marching Cubes reconstruction algorithm based on the surface of the three-dimensional point cloud (parameter: step size of 1 voxel) is used to realize the surface meshing of the segmented bone region, generate an accurate three-dimensional morphological model of the alveolar bone, and preserve the boundary structure between cortical bone and cancellous bone.

[0039] Furthermore, the cortical bone thickness was quantitatively extracted using a radial thickness calculation method (parameter: sampling point spacing of 0.25 mm), and a cortical bone thickness distribution matrix was generated. The formula for calculating cortical bone thickness is:

[0040] ,

[0041] in, The average cortical bone thickness, For the i-th radial measurement value, This represents the total number of sampling points;

[0042] Furthermore, by calculating the mean grayscale intensity and mapping it to the CT equipment calibration function, the mean trabecular bone mineral density is obtained, and a trabecular bone mineral density distribution matrix is ​​generated. The formula for calculating trabecular bone mineral density is:

[0043] ,

[0044] in, The average cancellous bone mineral density, Let j be the gray value of the j-th sampled voxel. The total number of voxels sampled from cancellous bone;

[0045] Furthermore, by using the volume integral method (parameter: voxel size is 0.2 mm³) and combining it with the bone mass grading standard, the bone mass of the implantation area is determined to be grade 1 to 4, and a bone mass grading identification matrix is ​​generated.

[0046] Through the above algorithm and processing flow, the bone density distribution and three-dimensional morphological parameters of the previous step are transformed into a quantitative bone structure feature matrix that includes cortical bone thickness, mean trabecular bone density and bone mass grading, so as to realize the computable characterization of the biomechanical support capacity of bone tissue in the implantation area.

[0047] For example, in a single acquisition, the CBCT scanner was a Planmeca ProMax 3D Mid with a resolution of 0.15 mm. The DICOM image grayscale range, after calibration, corresponded to a bone density conversion factor of 0.001 g / cm³ per HU. Threshold segmentation was set from 400 to 1600 HU, resulting in a continuous bone tissue volume of approximately 621 mm³. Connected component analysis removed five isolated noise regions, preserving the main bone area. The total number of sampling points N in the cortical bone thickness calculation was 256, and the average radial thickness T was calculated as follows: mm. The total number of voxels M for cancellous bone mineral density sampling was 1024, with a mean gray level of 850 HU, corresponding to a mean cancellous bone mineral density ρ. g / cm³. The volume integral result was 621 mm³, corresponding to a bone mass grade of 3. The results showed that the bone tissue in the implantation area of ​​this patient had a medium to high level of biomechanical support, providing accurate quantitative input for key modifiable clinical variables in subsequent causal modeling;

[0048] S1.2: Collect oral soft tissue status, periodontal probing depth and systemic health indicators (such as history of diabetes, smoking index) from clinical examination records. Use medical ontology mapping rules to transform unstructured text descriptions into standardized classification coding vectors, generate a computable subset of systematic risk factor features, and input them as confounding variable control terms into the subsequent causal modeling process.

[0049] Data on oral soft tissue condition, periodontal probing depth, and overall health indicators obtained through clinical examination are collected. Input data sources include structured scale scores, numerical test records, and scanned copies of doctors' handwritten medical records. The data is then analyzed by a combination of optical character recognition (OCR) and natural language processing (NLP) modules to form a raw text information stream that can be semantically mapped.

[0050] A medical terminology extraction algorithm (parameter: dictionary set = SNOMED CT + ICD-10 extended library) is used to analyze the boundaries and attribute categories of medical entities in the original text information stream based on the conditional random field (CRF) model, so as to achieve accurate identification of oral soft tissue pathological state, periodontal probing value and systemic disease items;

[0051] Furthermore, through ontology mapping rule matching (parameters: mapping threshold = 0.85, semantic similarity calculation method = cosine similarity), the identified medical entities are standardized into unique coded values ​​and filled into a unified clinical feature template structure, thereby realizing the transformation of unstructured descriptions into standardized classification coding vectors;

[0052] Furthermore, a reduction coding compression algorithm (parameters: maximum cardinality 32, optimal coding length constraint) is adopted to compress the high cardinality features in the standardized classification coding vector into binary bit strings to form a computable input vector, ensuring indexing efficiency and storage compactness in the feature space;

[0053] By using a confounding variable screening mechanism (parameter: Pearson correlation coefficient threshold = 0.3), systemic health index variables that may potentially interfere with the causal modeling results are extracted from the above computable input vector, identified as confounding control terms, and a subset of systemic risk factor features is generated to control interference in the subsequent causal modeling process.

[0054] For example, in a patient's clinical examination record, the soft tissue condition is recorded as "mild gingivitis with local redness and swelling" in the doctor's medical record, the periodontal probing depth is 3.2 mm, and the overall health indicators include a history of diabetes (type II) and a smoking index of 15 cigarettes per day for 12 years. After OCR extraction, the text is parsed into three independent medical entities using a medical terminology extraction algorithm, and matched with SNOMED CT codes "235719002", "251004004", and "449868002" respectively. These are then transformed into three classification code vectors using ontology mapping rules. The combination of disease states is represented, and then reduced and compressed to form a 12-bit binary string input vector. The Pearson correlation coefficient between diabetes history and risk score is calculated during the confounding variable screening phase. The correlation coefficient of the smoking index is All exceeded the set threshold. Therefore, both are labeled as confounding control terms. Ultimately, this systematic subset of risk factor features is input into the auxiliary branch of causal modeling to control prediction bias and achieve more accurate risk attribution.

[0055] S1.3: Extract the implant position coordinates, major axis tilt angle and distance parameters to adjacent anatomical structures recorded in the digital surgical planning system, combine them with the actual implantation trajectory deviation data in the intraoperative navigation log, calculate the three-dimensional positioning error tensor of the implant based on the spatial Euclidean distance and angle difference function, and generate operation-related feature variables that reflect the accuracy of surgical execution;

[0056] S1.4: The distribution of occlusal contact points and peak force sequences of opposing teeth collected by the integrated dynamic occlusal force sensing pad are used to identify high stress concentration areas through spatiotemporal clustering algorithm, and the occlusal load asymmetry index is calculated to generate biomechanical input features of the quantitative masticatory function load pattern, which are used to characterize the stability of the postoperative mechanical microenvironment.

[0057] S1.5: Summarize the data items from the three sources of imaging, clinical examination and surgical operation, perform semantic alignment based on a unified medical terminology set (such as SNOMED CT), construct a multimodal structured input dataset containing bone density, alveolar bone height, implant position, occlusal force distribution and postoperative infection indicators, and output a standardized case master index association table for subsequent preprocessing modules to call.

[0058] Step S2: Standardize the multimodal structured input dataset, including missing value imputation, numerical normalization, and categorical variable encoding, to generate a uniform-dimensional feature vector to eliminate biases caused by different measurement scales and data types. Specifically, this includes:

[0059] S2.1: Obtain a multimodal structured input dataset containing bone density, alveolar bone height, implant location, occlusal force distribution, and postoperative infection indicators. Identify missing data items in each field based on the missing value detection algorithm. Use the K-nearest neighbor imputation method to fill in the missing values ​​according to the clinical characteristics distribution of similar patient groups to generate a clinical data matrix with enhanced integrity.

[0060] S2.2: Based on the clinical data matrix with enhanced integrity, Z-score standardization is performed on the continuous variable fields, that is, the mean of the corresponding variable on the training set is subtracted and divided by its standard deviation, so as to eliminate the dimensional differences between different physiological parameters and obtain a normalized numerical feature subset.

[0061] S2.3: Perform one-hot encoding transformation on categorical variable fields (such as implant brand, surgeon level, infection level), generate corresponding binary encoding vectors based on the cardinality of categorical variables, and concatenate them to the normalized numerical feature subset to form a structured feature representation in a high-dimensional sparse feature space;

[0062] For the clinical data matrix after missing value imputation and Z-score normalization, the One-Hot Encoding algorithm (parameters: cardinality of categorical variables, length of output vector) is used to convert the discrete symbolic representation of categorical variable fields into computable binary vectors.

[0063] Furthermore, through the encoding matrix generation module (parameters: variable category label list, mapping strategy), a mapping relationship between the values ​​of categorical variables and the positions of binary vectors is established, and according to this mapping relationship, the categorical variable value corresponding to each sample data is mapped to a unique "1" bit, and the remaining positions are set to "0", thus obtaining the one-hot encoding representation at the sample level;

[0064] Furthermore, a sparse matrix generation method (parameter: encoding vector density threshold) is used to concatenate the set of one-hot encoding vectors generated for all samples column by column to form a sparse structured feature matrix for the categorical variable domain. The variable labels and position information of each column are recorded in the matrix metadata to ensure semantic consistency in the subsequent feature fusion process.

[0065] Furthermore, by using the feature fusion operator (parameters: splicing order rules, index offset), the sparse structured feature matrix and the previously obtained normalized numerical feature subset are spliced ​​together without overlap according to the feature dimension to form a unified high-dimensional feature matrix containing numerical features and classification features, so that different types of input data are in a consistent computational space.

[0066] Through the above one-hot encoding and high-dimensional matrix concatenation process, the normalized numerical feature subset of the previous step is transformed into a high-dimensional sparse feature representation covering all clinical variables, thereby achieving semantic alignment and computability improvement of various variables in a unified feature space.

[0067] For example, in a clinical implant dataset, there are three categorical variables: implant brand (number of categories = 4), surgeon level (number of categories = 3), and infection level (number of categories = 5). After category label statistics, the output one-hot encoded vector has a length of 12, where the first 4 bits correspond to the brand category, the next 3 bits to the surgeon level category, and the last 5 bits to the infection level category. For a record where the implant brand is category 2, the surgeon level is category 1, and the infection level is category 4, the 2nd, 5th, and 11th bits of the one-hot encoded output matrix are "1", and the remaining bits are "0". When the numerical feature subset includes three continuous variables—bone density (Z-score normalized mean = 0, standard deviation = 1), alveolar bone height, and occlusal force distribution—these 3-dimensional numerical features are concatenated column-wise with the 12-dimensional categorical features to obtain a 15-dimensional high-dimensional feature matrix. When this matrix is ​​input into the subsequent principal component analysis module, it can significantly improve the semantic consistency and computational operability of the feature space while preserving the original semantic information.

[0068] S2.4: Input the spliced ​​high-dimensional sparse feature representation into the feature dimension alignment module, extract the principal components with a cumulative variance contribution rate of more than 95% through the principal component analysis algorithm, compress the feature space dimension and remove redundant information to generate a compact unified dimension feature vector.

[0069] S2.5: Output a compact, uniform-dimensional feature vector to the next processing stage, while recording standardized parameters (mean, standard deviation, encoding mapping table, PCA projection matrix) to ensure that new samples can reproduce the same preprocessing process during subsequent inference, thus ensuring model input consistency.

[0070] like Figure 2As shown, step S3 involves inputting the unified-dimensional feature vector into a dual-branch deep neural network model. The main branch performs the implant failure risk prediction task and outputs an initial risk score. The auxiliary branch uses an attention gating mechanism to filter for potential confounding factors that are strongly correlated with the set of operable clinical variables as anchor points, generating a local causal association strength matrix. Specifically, this includes:

[0071] S3.1: Obtain the standardized preprocessed uniform dimension feature vector and use it as the input data for the dual-branch deep neural network model; based on this input data, design a shared representation layer at the front end of the model, use a multi-layer fully connected network to extract high-order nonlinear features, and generate shared latent feature representations to preserve the joint distribution characteristics of multimodal clinical information and provide a common semantic basis for subsequent main and auxiliary tasks;

[0072] The standardized preprocessed feature vectors of uniform dimension are input into the front-end shared representation layer building module of the dual-branch deep neural network as the basic input data of the model.

[0073] A high-dimensional feature embedding mapping method is adopted (parameters: input dimension = length of compressed feature vector, activation function = ReLU, weight initialization strategy = Xavier uniform distribution) to realize the mapping from the original uniform dimension feature space to a high-order nonlinear feature space and establish a joint feature representation basis for multimodal information;

[0074] Furthermore, through a multi-layer fully connected network structure (parameters: number of layers = 3, number of hidden layer nodes = 256, 128, 64 respectively, activation function = ReLU, regularization method = Dropout rate 0.3), we can model the nonlinear combination and interaction relationship of input features and obtain the implicit vector representation that encodes the statistical correlation and temporal dependence of multimodal clinical data;

[0075] Furthermore, a batch normalization method (parameter: cross-batch mean-variance statistical update momentum = 0.9) is adopted to normalize the mean-variance of the output of each fully connected layer to eliminate the distribution differences between batches, accelerate the convergence process of feature extraction, and improve the stability of the model under different data batch conditions.

[0076] Furthermore, through the residual connection mechanism (parameter: skip connection span = 2), the feature representations extracted from the shallow and deep layers are directly added together, preserving the fine-grained information of low-order features and improving the effectiveness of gradient backpropagation, thereby generating a shared latent feature representation tensor with multi-scale and cross-modal comprehensive representation capabilities.

[0077] Through the above-mentioned multi-layer nonlinear mapping and feature fusion method, the unified dimension feature vector of the previous step is transformed into a shared feature representation that can simultaneously support risk prediction and causal relationship analysis, thus achieving the expected technical effect of providing a common semantic basis for subsequent main and auxiliary tasks.

[0078] For example, in a dental implant failure risk prediction scenario, the compact, uniform-dimensional feature vector (length = 64) output from step S2.5 is input into the shared representation layer. The number of fully connected nodes in the first layer is set to 256, the weights are initialized using a uniform Xavier distribution, the activation function is ReLU, and the output tensor shape is ( , 256). The number of fully connected nodes in the second layer is set to 128, and batch normalization (momentum = 0.9) is performed at the output followed by ReLU, resulting in a shape of ( The intermediate representation of (, 128). The number of fully connected nodes in the third layer is set to 64, and the Dropout rate is set to 0.3 to suppress overfitting. It is then added element-wise to the output of the first layer through residual connections to form the final shape ( The shared latent feature representation tensor of , 64). With a batch sample size of 32, after the above processing, the convergence speed of the model on the validation set is significantly improved, the distribution of hidden layer activation values ​​is more stable, and the main branch achieves higher discrimination accuracy when estimating the probability of implant failure risk in S3.2. At the same time, it provides a more complete multimodal information foundation for the auxiliary branch to guide the feature selection of interventionable variables in S3.3.

[0079] S3.2: Input the shared hidden feature representation into the main branch network, which is composed of stacked residual blocks to capture deep risk discrimination patterns; perform nonlinear mapping transformation based on the forward propagation process of the main branch, calculate the sigmoid activation value of the final output layer, and generate an initial risk score in scalar form, representing the probability estimate of implant failure for the current patient in a specific future period;

[0080] The shared latent feature representation is input into the residual block structure of the main branch network. The residual connection mechanism and multi-layer convolution operation are used (parameters: convolution kernel size 3×3, stride 1, padding method keeps the size unchanged) to realize the nonlinear transfer and gradient stable flow between multi-layer feature maps, and capture cross-modal risk discrimination patterns in the latent space.

[0081] Furthermore, through batch normalization (parameters: momentum 0.9, ...), Mean and variance normalization is performed at the output of each convolutional layer to eliminate the distribution drift effect between different batches of samples and improve the training convergence performance of the main branch network.

[0082] Furthermore, by applying a nonlinear transformation after batch normalization using the ReLU activation function, negative features are suppressed to zero while maintaining the proportional propagation of positive features, thereby enhancing the model's ability to express nonlinear decision boundaries.

[0083] Furthermore, the mean of each channel is calculated in the spatial dimension through a global average pooling layer to obtain a compact feature vector at the channel level, reducing the number of parameters and retaining key discriminative information. The convolution output is then mapped to a one-dimensional vector to drive subsequent fully connected layer processing.

[0084] Furthermore, a mapping relationship is established between the feature vector and the output node by performing linear combination operations and weight parameter matrix multiplication through fully connected layers. The final linear output is then fed into the Sigmoid activation function for probabilistic processing to obtain a normalized score for implant failure risk. ,in This is a linear combination of fully connected layers;

[0085] By constructing convolutional residuals, normalizing and processing nonlinear mapping as described above, the shared latent feature representation is transformed into an initial risk score reflecting the current patient's implant failure in a specific future period, thus achieving accurate output of the risk prediction task.

[0086] For example, in the clinical dataset, the uniform-dimensional feature vector length is 512. The feature tensor size extracted by the shared representation layer is (64 channels, height=8, width=8). The main branch uses 3 residual blocks, each containing two 3×3 convolutional kernels, batch normalization, and ReLU activation. The convolutional kernel weights are initialized using He Normal to ensure gradient stability. After global average pooling, the feature vector size is reduced to 64 dimensions and mapped to a single scalar z through a fully connected layer. z is input to the Sigmoid function to calculate the probability P, resulting in an initial risk score of 0.72, indicating a high risk of implant failure in the future observation period. This output is further analyzed by the downstream causal inference module in conjunction with modifiable variables, ultimately generating clinical intervention suggestions for bone density adjustment and occlusal force distribution optimization, significantly improving the model's operability and interpretability in actual diagnosis and treatment.

[0087] S3.3: Define a set of interventionable clinical variables, including bone mineral density, occlusal force distribution, implant placement angle, postoperative infection control level, and soft tissue healing status; based on the modifiable properties of each variable in this set in clinical practice, use them as guiding anchors for auxiliary branches, and extract the embedding vectors of the corresponding variables from the shared latent feature representation to form an anchor-guided feature subspace to focus on factors with clinical intervention significance.

[0088] Based on the standardized preprocessed unified dimension feature vector, the complete set of clinical variables and their structured coding information are obtained under the input conditions of the auxiliary branch. The specific set of variables that are directly controllable in dental implant clinical practice is extracted through variable attribute screening rules, thereby realizing the definition of the set of interventionable clinical variables.

[0089] Using rule matching and medical knowledge base retrieval methods (parameters: SNOMED CT code set, adjustable marker bits), the system can locate and label modifiable variables such as bone density, occlusal force distribution, implant placement angle, postoperative infection control level, and soft tissue healing status, and generate a set description structure containing unique variable identifiers, numerical ranges, and adjustable attribute labels.

[0090] Furthermore, through an embedding mapping algorithm (parameters: shared latent feature representation matrix, variable index list), each variable in the set of influencing variables is mapped to its feature embedding vector position in the shared latent feature representation space, and an embedding matrix is ​​output with variables as rows and feature dimensions as columns. The generation of this embedding matrix depends on the weight parameters of the shared representation layer, ensuring that the embedding vectors retain the semantic consistency of multimodal information.

[0091] Furthermore, by constructing an anchor-guided feature subspace (parameters: embedding matrix, feature selection threshold), we can extract feature dimensions that are highly related to the embedding vector of the manipulable variable from the shared latent feature representation and remove irrelevant feature terms to form an anchor-guided feature subspace with concentrated dimensions and semantic focus, which is used for the subsequent calculation of relevance weights in the attention gating mechanism.

[0092] Through the above-mentioned embedding extraction and subspace construction processing, the shared latent feature representation of the previous step is transformed into a highly correlated feature subspace centered on clinically controllable factors, so that the causal reasoning process focuses on the variable group with direct intervention significance, providing an accurate feature basis for generating a local causal association strength matrix.

[0093] For example, in a structured case dataset containing 200 patients, a unified dimensional feature vector with a dimension of 128 was generated after preprocessing. The entire variable set contained 45 structured clinical variables, among which five modifiable variables were selected based on the dental implant adjustability rules: bone mineral density (variable ID: V01, range: 0.6~1.2g / cm³), occlusal force distribution (variable ID: V02, range: 100~350N), implant placement angle (variable ID: V03, range: 0°~30°), postoperative infection control level (variable ID: V04, level code: 0~3), and soft tissue healing status (variable ID: V05, level code: 0~4). The variable IDs were mapped to feature index positions in the shared latent feature representation matrix; for example, V01 corresponds to index [12,45,78], V02 corresponds to index [8,34,56], etc., and the corresponding embedding vectors (length 128) were extracted to form a 5×128 embedding matrix. An anchor-guided feature subspace construction algorithm with a Pearson correlation coefficient threshold of 0.85 was used to select 42 feature dimensions with correlation coefficients higher than 0.85 with the embedding matrix from 128 dimensions of shared latent feature representations, forming a 42-dimensional anchor-guided feature subspace. This subspace significantly improved the signal-to-noise ratio of highly correlated terms in the subsequent attention weight calculation, ensuring that the clinical variable path for causal inference localization is concentrated within the range of factors with high intervention value.

[0094] S3.4: An attention gating mechanism is introduced in the auxiliary branch. This mechanism is based on a scaled dot product attention structure to calculate the correlation weight between the anchor-guided feature subspace and the globally shared latent features. The correlation weight is subjected to softmax normalization to obtain the attention allocation probability of each potential confounding factor relative to the manipulable variable. The feature response is then aggregated according to the probability to generate a local causal association strength matrix, the elements of which represent the influence strength of a specific confounding factor on a certain manipulable variable.

[0095] S3.5: The local causal association strength matrix is ​​sparsified, and L1 regularization is used to filter significant association terms and remove weak correlation noise interference. The optimized matrix is ​​used as the auxiliary branch output and is passed to the next stage together with the initial risk score of the main branch output. It is used to initialize the node connection weights in the individualized dynamic causal graph to ensure that the causal reasoning process focuses on the operational and strongly associated clinical variable path.

[0096] like Figure 3As shown, step S4 involves constructing an individualized dynamic causal graph based on the local causal association strength matrix and the initial risk score. Nodes represent clinical variables, and directed edges represent the direction and strength of causal influence between variables. The weights are initialized using a Bayesian update rule combined with historical population data, and modifiable variables are marked as key regulatory nodes. Specifically, this includes:

[0097] S4.1: Obtain the local causal association strength matrix output by the auxiliary branch and the initial risk score generated by the main branch. The local causal association strength matrix contains potential confounding factors screened with interventionable clinical variables as anchors and their attention-weighted association values ​​with each prognostic-related variable. Based on this matrix, perform variable node initialization processing to determine the set of nodes in the individualized dynamic causal atlas. Each node corresponds to a structured clinical feature or imaging parameter, forming a set of variable nodes with semantic labels as the input basis for atlas construction.

[0098] Based on the local causal association strength matrix output by the auxiliary branch and the initial risk score generated by the main branch, a semantic parsing algorithm (parameters: medical terminology set index table, feature field mapping rules) is used to realize the semantic recognition function of the anchor points of the interventionable clinical variables in the matrix and their corresponding potential confounding factors.

[0099] Furthermore, by using the feature decomposition method (parameters: matrix row and column index mapping, attention weighting threshold setting), the matrix elements are decomposed into "variable-variable" association pairs, and their clinical attribute categories are determined by combining the variable category labels, resulting in a set of association pairs with semantic labels;

[0100] Furthermore, a set merging algorithm (parameters: feature vector hash encoding, duplicate item removal rules) is used to uniquely process the duplicate variable nodes appearing in the association set and generate an array of variable nodes without duplicates.

[0101] Furthermore, by using a node attribute generation method (parameters: structured clinical feature data table, imaging parameter index), a corresponding structured feature description is assigned to each unique variable node, forming a node metadata table with readable attributes;

[0102] Furthermore, a node set construction algorithm (parameters: node metadata table, node unique identifier) ​​is adopted to transform the above metadata table into a graph construction input set and output a set of variable nodes with complete semantic identification.

[0103] By using semantic parsing and node set construction, the local causal association strength matrix and initial risk score results from the previous step are transformed into node set data that can be directly used for causal graph construction, thus achieving the effect of clinical variable node initialization technology in the stage of individualized dynamic causal graph construction.

[0104] For example, in a dental implant risk prediction scenario, the local causal association strength matrix has a dimension of 20×20, where anchor variables include bone density, occlusal force distribution, implantation angle, and infection control level, and matrix elements are normalized attention weights in the range of [0,1] floating-point numbers. The semantic parsing algorithm calls the SNOMEDCT medical terminology database to match the matrix anchor variable "bone density" to term ID 264392003, and labels it with the clinical category "bone tissue features". The feature decomposition method reads the first row of the matrix corresponding to "bone density" and the seventh column corresponding to "cortical bone thickness," with an association value of 0.82, greater than the set threshold of 0.7, and stores it in the association pair set as <bone density, cortical bone thickness>. The set merging algorithm uses hash encoding to uniquely process the repeated occurrences of the "bone density" variable, resulting in a variable array of length 14. The node attribute generation method retrieves the structured description field of the variable "cortical bone thickness" from the standardized case database, including the numerical range [2.4,3.6] mm and parameters collected from CBCT voxel segmentation. The node set construction algorithm encapsulates all node attributes with unique variable identifiers (UUID format), generating a set of 14 nodes for subsequent causal graph construction. The results show that the node set fully covers the modifiable variable and its associated confounding factors; all node attribute fields are populated and conform to medical semantic standards, providing high-quality input for generating causal connections.

[0105] S4.2: Based on the set of variable nodes, the attention weights in the local causal association strength matrix are used as the basis for the initial connection strength. A directed edge generation operation is performed, and a threshold truncation method is used to filter weak associations. Weighted relationships that are significantly greater than the preset causal threshold are retained to generate a preliminary causal connection structure. Each directed edge represents the potential causal influence direction from the source node to the target node. Its initial weight reflects the significance of the association between the two variables in the current sample context. A preliminary causal topology diagram is output as the skeleton of the dynamic graph.

[0106] S4.3: Input the preliminary causal topology graph into the Bayesian prior fusion module. Use the stable causal pattern at the population level obtained from training on historical population data as the prior distribution. Perform Bayesian update processing on the initial weight of each directed edge and calculate the posterior edge weight = (prior weight × population confidence + current sample weight × individual fitness coefficient) / normalization factor to balance the statistical regularity of the population and the specific bias of the individual. Generate a set of dynamic causal edge weights that has been corrected by the prior, and improve the robustness and generalization ability of the graph under small sample conditions.

[0107] The initial causal topology graph output from step S4.2 is used as input data. The Bayesian prior fusion module is called (parameter setting: the prior distribution is a stable causal pattern trained from historical group data) to fuse the initial edge weights of the current sample with the group prior weights, ensuring that the two types of causal information are comparable under the same probabilistic framework.

[0108] A probability-weighted combination formula is used to calculate the posterior weight of each directed edge as a weighted sum of the prior weight of the population and the weight of the current sample, and a normalization factor is used to ensure the consistency of the total weight. The formula is as follows:

[0109] ,

[0110] in, Let be the posterior edge weight from node i to node j. For prior weights, This is the group confidence coefficient. The initial weights are calculated for the current sample. For individual fitness coefficient, This is a normalization factor to ensure that all edge weights maintain consistent dimensions after fusion;

[0111] Furthermore, by using the Bayesian update rule (based on the product of the prior distribution and the likelihood function to construct the posterior distribution), the above fusion formula is applied to the weight set of all edges to obtain a posterior edge weight matrix that combines the statistical regularity of the population with the specificity of the individual.

[0112] Furthermore, using the Monte Carlo sampling method (with a sampling number of 1000 times), random weight samples are generated between the prior distribution and the individual sample distribution. The mean and variance of the posterior distribution are calculated to estimate the uncertainty of the fusion weights and provide a confidence reference for subsequent graph structure optimization.

[0113] By using confidence interval filtering, edge weights below a specified significance level are marked as weakly correlated and temporarily stored in the removal queue, while high-confidence edges that have been corrected a priori are retained, forming a dynamic causal edge weight set optimized by a priori fusion. This achieves the technical effect of improving robustness and generalization ability under small sample conditions.

[0114] For example, in the topology of a certain patient, assume the initial sample weights between node A and node B. for Prior weights of historical groups for Group confidence coefficient Set as Individual fitness coefficient Set as Normalization factor Set as Substitute the parameters into the formula:

[0115] ,

[0116] The posterior edge weights were calculated to be: Subsequently, 1000 weighted samples were generated using Monte Carlo sampling to obtain the mean. ,variance The confidence interval range meets the set significance level threshold. In this embodiment, the edge weights fused a priori remain stable across multiple sampling results, significantly improving the interpretability and generalization adaptability of the causal graph in this patient sample;

[0117] S4.4: Based on the dynamically causal edge weight set that has been corrected a priori, a graph structure optimization operation is performed. A sparse constraint regularization method is used to remove redundant reverse connections and loop paths to ensure that the causal graph satisfies the directed acyclic graph (DAG) constraint. At the same time, a feedback adjustment mechanism is introduced to allow postoperative follow-up variables to form an inverse weight decay correction for preoperative nodes, generating a quasi-DAG structure that conforms to clinical temporal logic and outputting a structurally compliant individualized dynamic causal graph.

[0118] S4.5: In the structurally compliant individualized dynamic causal graph, based on the predefined set of interventionable clinical variables (including bone density, occlusal force distribution, implant placement angle, postoperative infection control level, etc.), the labeling mechanism performs node attribute annotation processing, uniformly marking nodes belonging to this set as 'key regulatory nodes' and assigning them interventionable flags for selective intervention operations in the subsequent counterfactual perturbation simulation stage, ensuring that the causal reasoning process focuses on clinically controllable factors and enhancing the operability and practicality of the model output recommendations.

[0119] Step S5: Apply counterfactual perturbation simulation to key control nodes in the individualized dynamic causal graph, calculate the predicted output gradient response caused by changes in each modifiable variable under virtual intervention, and generate a variable intervention sensitivity sequence as a causal effect quantification index. Specifically, this includes:

[0120] S5.1: Based on the set of key regulatory nodes in the individualized dynamic causal graph, obtain the current observation sequence of the interventionable clinical variables as the benchmark input condition for counterfactual perturbation; use the historical distribution characteristics of the interventionable variables to determine the perturbation amplitude range, and generate a virtual intervention sample set containing positive and negative small perturbations to construct the counterfactual input space;

[0121] S5.2: Perform latent space mapping processing on the virtual intervention sample set, input it into the main branch model of the dual-branch deep neural network for forward propagation, and calculate the implant failure risk score output sequence corresponding to each perturbed sample; based on the risk score of the original unperturbed sample as a benchmark, obtain the risk change sequence under each virtual intervention as a preliminary effect response indicator;

[0122] Based on the virtual intervention sample set generated by S5.1, the input object is determined to be a multidimensional array of interventionable clinical variables containing small positive and negative perturbations and a unified dimension feature vector of the original unperturbed sample;

[0123] The latent space mapping method (parameters: mapping network structure and shared representation layer weights) is adopted to input the feature vector of each virtual intervention sample into the main branch shared representation layer, realizing the nonlinear transformation of multimodal clinical features in the high-order feature space and obtaining the perturbation latent feature representation matrix;

[0124] Furthermore, the forward propagation algorithm of the main branch deep residual network (parameters: number of residual blocks, activation function type ReLU, output layer type sigmoid) is used to perform risk discrimination mapping on the perturbation latent feature representation matrix sample by sample, generating the corresponding implant failure risk score output sequence;

[0125] Furthermore, based on the risk score values ​​of the original unperturbed samples, the risk score elements of each perturbed sample are subtracted to obtain the risk change sequence R, where each element represents the difference between the predicted output and the baseline risk under the current virtual intervention conditions.

[0126] The risk change for a single virtual intervention sample is calculated using the following formula. :

[0127] ,

[0128] in, This represents the feature vector of the original, undisturbed sample. The feature vector after applying virtual intervention. Main branch risk scoring function;

[0129] By integrating the risk change series, a preliminary effect response index matrix is ​​formed, providing a quantitative basis for subsequent sensitivity calculations;

[0130] Through the above latent space mapping and risk change calculation, the impact of virtual intervention is transformed into quantifiable initial effect response data, thereby characterizing the predictive output response capability of each intrusive variable under different disturbance intensities.

[0131] For example, in a bivariate virtual intervention sample set containing bone mineral density perturbations and occlusal force distribution perturbations, the original observed bone mineral density value was 1.45 g / cm³, and the virtual intervention amplitude was set to +0.10 g / cm³ and -0.10 g / cm³; the original observed peak occlusal force was 120 N, and the virtual intervention amplitude was set to ±15 N. The above four sets of perturbation samples were input into a shared representation layer latent space mapping network (hidden layer units 256, activation function ReLU), obtaining a latent space feature matrix with a dimension of [4×128]. After forward propagation through the main branch residual network, the perturbation risk score sequence [0.38, 0.35, 0.42, 0.40] was obtained, and the original unperturbed risk score was 0.37. The formula was used... Calculate the change; bone mineral density +0.10 g / cm³ corresponds to ΔR = 0.38. 0.37 = 0.01, bone mineral density 0.10 g / cm³ corresponds to ΔR = 0.35 0.37 = 0.02, bite force +15N corresponds to ΔR = 0.42 0.37 = 0.05, bite force 15N corresponds to ΔR = 0.40 0.37 = 0.03. This effect response matrix shows that the gradient of change in the predicted output due to occlusal force perturbation is significantly higher than that due to bone mineral density perturbation, providing a clear quantitative indication for the subsequent generation of sensitivity sequences;

[0132] S5.3: Based on the functional relationship between the risk change sequence and the corresponding virtual intervention intensity, the local gradient response value of each intrusive variable relative to the main branch output is calculated using automatic differentiation technology; the gradient response values ​​are arranged according to the variable dimension to form an initial variable intervention sensitivity vector to characterize the direction and intensity of the influence of each variable on the prediction result;

[0133] S5.4: Standardize and normalize the initial variable intervention sensitivity vector to eliminate comparison bias caused by different variable scales; combine the prior weight information of the corresponding edge in the individualized dynamic causal graph, use the Bayesian weighted fusion rule to optimize the sensitivity estimation, and generate a calibrated variable intervention sensitivity sequence as the quantitative output of causal effect;

[0134] S5.5: Perform a sorting operation based on the calibrated variable intervention sensitivity sequence to identify the preceding highly sensitive interventionable variables that have the greatest impact on the predicted output; mark the highly sensitive variables as priority intervention targets, and output the variable intervention sensitivity sequence and its sorting results to support subsequent path attribution analysis and intervention suggestion generation;

[0135] After obtaining the variable intervention sensitivity sequence calibrated by Bayesian weighted fusion rules as input conditions, the sensitivity sequence is globally sorted using a quick sorting algorithm (parameter: comparison function based on sensitivity value), generating a variable index mapping table arranged in descending order of sensitivity value, thereby realizing the ordered expression of the influence of interventionable variables;

[0136] Furthermore, by using a threshold truncation method (parameter: the significance threshold is twice the standard deviation of the sensitivity sequence), a set of variables with sensitivity values ​​greater than or equal to the threshold is selected, and this set is intersected with the set of key regulatory nodes in the original individualized dynamic causal graph to obtain a subset of interventionable variables that actually have a high intervention effect in the current prediction context;

[0137] Furthermore, a mapping annotation mechanism (parameter: annotation field type is Boolean) is used to write the "priority intervention target" flag into the node attribute set of the individualized dynamic causal graph of the above high intervention effect variables, so as to realize the linkage between the graph structure and the intervention priority;

[0138] Furthermore, by using the matrix indexing and reorganization method (parameter: based on the sorted sensitivity index table), the sensitivity values ​​of high intervention effect variables are rearranged into a priority sequence vector, and their position numbers in the sorting are appended to form the final priority-labeled sensitivity output matrix, which is used to drive the path screening weight setting in subsequent path attribution analysis.

[0139] Through the sorting and labeling process described above, the calibration sensitivity results from the previous step are transformed into structured variable priority information, enabling rapid identification and targeted labeling of the variables that have the greatest impact on the prediction output.

[0140] For example, in a set of calibrated sensitivity sequences containing bone mineral density (sensitivity value 0.82), occlusal force distribution (sensitivity value 0.76), implant placement angle (sensitivity value 0.65), postoperative infection control level (sensitivity value 0.91), and soft tissue healing status (sensitivity value 0.58), the sorting algorithm outputs a sequential index mapping table of infection control level, bone mineral density, occlusal force distribution, implantation angle, and soft tissue healing status. By using a threshold truncation method and taking twice the standard deviation as 0.70, three high-sensitivity variables, infection control level (0.91), bone mineral density (0.82), and occlusal force distribution (0.76), are selected. These variables are marked as priority intervention targets in the individualized dynamic causal atlas node attributes, and an output matrix with a priority sequence of {infection control level: rank 1, bone mineral density: rank 2, occlusal force distribution: rank 3} is generated. In terms of application effectiveness, this priority matrix can assign initial weights to each path during the path attribution analysis stage, which significantly increases the ranking priority of high-sensitivity variable paths in the generation of intervention recommendations, thereby enhancing the targeting and clinical guidance value of intervention strategies.

[0141] Step S6: Based on the variable intervention sensitivity sequence, perform attribution analysis on the paths in the individualized dynamic causal graph to identify key causal propagation paths from modifiable variables to high-risk outputs, and generate a priority intervention recommendation list based on the expected intervention effects. Specifically, this includes:

[0142] S6.1: Based on the variable intervention sensitivity sequence output by the previous steps, obtain the set of all potential causal paths in the individualized dynamic causal graph that start with the interventionable variable and end with the high-risk prediction output. Each path consists of a series of ordered connected nodes and directed edges. The nodes represent clinical variables, and the directed edges represent the direction and intensity of causal influence after Bayesian update, forming an initial path candidate pool.

[0143] Based on the variable intervention sensitivity sequence input conditions generated by the S5 step, the individualized dynamic causal graph optimized by structural compliance constraints is loaded as the path retrieval space, and it is clarified that both the graph node set and the directed edge set have the weight attribute of Bayesian prior update.

[0144] An anchor point starting point matching algorithm (parameters: set of influencing variable labels, graph node index table) is used to retrieve all directed paths that satisfy the condition that the starting node belongs to the set of influencing variables and the ending node is a high-risk prediction output node, and generate a path candidate index list to limit the scope of attribution analysis.

[0145] Furthermore, by using a depth-first traversal algorithm (parameters: path length upper limit = maximum graph depth, loop path detection threshold = 0), the individualized dynamic causal graph is recursively traversed from the starting node along the directed edges to construct each node-edge sequence that meets the conditions, ensuring that each node in the path is unique and satisfies directional consistency, and recording the corresponding edge weight in each path.

[0146] Furthermore, by using a path structured coding method (parameters: node semantic labels, edge weight matrix), the node set of each path is transformed into a structured sequence of clinical variables, and the directed edge set is transformed into an array of causal influence weights arranged in sequence, which are used for subsequent effect calculation and visualization output.

[0147] Furthermore, a path validity determination algorithm (parameters: directed acyclic graph constraint, node type mapping table) is used to verify whether each generated path conforms to the causal structure logic and clinical temporal constraints, and to eliminate candidates with reverse time flow or clinically unrealizable paths to form an initial path candidate pool.

[0148] Through the above path retrieval and legality filtering, the directed connection relationship between the high-sensitivity starting variables in the variable intervention sensitivity sequence and the individual causal graph is directly mapped into a complete potential causal propagation link, realizing the construction of a set of causal paths from the interventionable variables to the high-risk output, laying a data foundation for subsequent path effect quantification;

[0149] For example, in an individualized dynamic causal graph containing bone mineral density (node ​​A), occlusal force distribution (node ​​B), postoperative infection indicators (node ​​C), and failure risk output node (node ​​Y), the input variable intervention sensitivity sequence is used, where nodes A and B are highly sensitive variables. An anchor point origin matching algorithm is used to retrieve paths starting from A or B and ending at Y, resulting in two candidate paths: A→C→Y and B→Y. A depth-first traversal is performed on the A→C→Y path, recording the node sequence [A,C,Y] and edge weight array [w(A,C)=0.65, w(C,Y)=0.72]; for the B→Y path, the node sequence [B,Y] and edge weight [w(B,Y)=0.88] are recorded. The paths were structured and encoded as follows: A→C→Y was encoded as {"nodes":["bone density","postoperative infection","failure risk"], "weights":[0.65,0.72]}, and B→Y was encoded as {"nodes":["occlusal force distribution","failure risk"], "weights":[0.88]}. The path validity algorithm was used to verify that both paths satisfied the requirements of a directed acyclic graph and clinical temporal logic, ultimately forming an initial path candidate pool including the two structured paths mentioned above, which were used for subsequent effect integral calculation and intervention suggestion generation.

[0150] S6.2: Perform path effect integral calculation for each path in the set of potential causal paths, and generate a path-level causal effect quantification value by combining the weighted product of the weights along each edge of the path with the counterfactual perturbation response gradient, which serves as an evaluation index of the overall intervention influence of the path for subsequent sorting and screening.

[0151] S6.3: Based on the path-level causal effect quantification value, a threshold filtering mechanism is set to select key causal propagation paths that are higher than the preset significance threshold, eliminate weakly influential paths, retain the main causal links that have a substantial driving effect on the predicted output, and generate a simplified key path subset.

[0152] Using the path-level causal effect quantification matrix as input data, a significance threshold setting algorithm (parameters: the causal effect unit is the risk score change ΔR, and the significance level α is set according to the historical sample statistical distribution) is adopted to calculate the significance judgment value corresponding to each path, so as to achieve the preliminary screening of path influence.

[0153] Furthermore, the path-level causal effect quantification value is scaled uniformly by a normalization processing algorithm (parameter: minimum-maximum normalization range [0,1]) to ensure the comparability of path effect indicators of different dimensions in threshold judgment and to obtain a normalized causal effect vector.

[0154] Furthermore, a binary mask generation algorithm (parameter: threshold T is determined by the percentile of the distribution corresponding to α) is used to perform element-by-element threshold comparison operation on the normalized causal effect vector to generate a path-preserving mask matrix, and output the filtered path index set based on the items with a mask value of 1.

[0155] Furthermore, a directed graph structure constraint algorithm (parameter: DAG constraint, feedback loops are not allowed) is used to perform topological verification on the filtered path index set, eliminating paths that do not meet the causal directed acyclic condition, thus ensuring the clinical interpretability and reasoning legitimacy of the final output path set in terms of structure.

[0156] By using a pruning optimization algorithm (parameter: L1 sparse regularization coefficient λ), the influence ranking of the structure-verified path set is performed, and edge connections with weights below the preset sparse threshold are removed to generate a simplified path set containing only high-weight causal edges, thereby extracting the main causal links that have a substantial driving effect on the prediction output.

[0157] Through the above processing method, the path-level causal effect quantification value sequence calculated in the previous step is transformed into a simplified critical path subset, achieving the expected technical effect of noise reduction and strengthening of the backbone link.

[0158] For example, in a dental implant failure risk prediction case, the input path-level causal effect quantification values ​​are: bone density → occlusal force distribution → risk output path effect 0.86, bone density → infection control → risk output path effect 0.34, implant angle → bone density → risk output path effect 0.78, and soft tissue healing → risk output path effect 0.22. The significance threshold T is set to 0.5 based on historical samples. Max-min normalization is performed on the initial effect value sequence, where the normalized value for the bone density → occlusal force distribution path is 1.0, the normalized value for the bone density → infection control path is 0.395, the normalized value for the implant angle → bone density path is 0.91, and the normalized value for the soft tissue healing path is 0.0. A binarization mask is used for judgment; paths with a normalization value ≥ 0.5 are marked as retained, resulting in two path indices: bone density → occlusal force distribution and implant angle → bone density. DAG structure verification is performed on these two paths, and the existence of causal loops is checked. The verification results show that both paths conform to a causal directed acyclic structure. Through L1 sparse regularization pruning, the significance of the bone mineral density → occlusal force distribution path remained at 0.84 after weight decay, and the significance of the implant angle → bone mineral density path remained at 0.75 after weight decay. The final simplified critical path subset consists of the above two main chains. In subsequent intervention suggestion generation, the system will use these two paths as a basis to push priority intervention measures such as bone mineral density optimization and implant angle adjustment, significantly improving the targeting and expected effect of the suggestions.

[0159] S6.4: For each path in the critical path subset, extract the associated initiating variable that can be intervened, and estimate the expected effect of the corresponding intervention measure by combining its path-level causal effect quantification value, and generate an effect estimation tuple containing the initiating variable identifier, the expected risk reduction amount, and the uncertainty interval.

[0160] S6.5: Sort all effect estimation tuples in descending order of expected risk reduction to generate a priority intervention recommendation list. Each item includes the name of the modifiable variable, the recommended intervention intensity level, the expected effect and its scientific basis (i.e. the key causal pathway), for clinicians to refer to in developing personalized treatment plans.

[0161] Step S7: Output a comprehensive decision support report containing the initial risk score, a list of priority intervention recommendations, a visualized causal path diagram, and estimates of the expected effects and uncertainty intervals of each intervention measure, for clinicians' reference. Specifically, this includes:

[0162] S7.1: Obtain the initial risk score output by the main branch network and the local causal association strength matrix generated by the auxiliary branch. Perform semantic consistency verification based on the bimodal result alignment mechanism, calculate the correlation coefficient between the two at key variable nodes, and ensure that the prediction results and causal structure remain logically consistent in the context of individual patients. Generate the fusion decision tensor after consistency verification.

[0163] S7.2: Based on the consistent verification of the fusion decision tensor and the directed edge weights in the individualized dynamic causal graph, the Dijkstra shortest path algorithm combined with the causal effect accumulation rule is used to identify the key causal propagation paths from each intrusive variable to the high-risk output node, and generate a set of candidate attribution paths containing multiple potential influence chains.

[0164] S7.3: Perform intervention sensitivity weighted integral operation on each path in the candidate attribution path set, assign path importance weights based on the variable intervention sensitivity sequence obtained from counterfactual perturbation simulation, calculate the comprehensive attribution score of each path, and generate a priority-ranked list of dominant causal paths according to the scores in descending order.

[0165] Based on the candidate attribution path set and the fusion decision tensor verified by consistency as input conditions, the intervention sensitivity weighted integral algorithm (parameters: path node set, edge weight matrix, intervention sensitivity sequence) is used to quantify the comprehensive importance of each path.

[0166] Furthermore, by performing edge-by-edge effect accumulation operation (parameters: weight product sequence, perturbation gradient vector) on the directed edge weights of path nodes and their causal connections, the overall effect of the path is quantified, and the weighted integral value is obtained as a path-level quantification index.

[0167] Furthermore, based on the variable intervention sensitivity sequence obtained from counterfactual perturbation simulation, a weight allocation function is used to map the sensitivity of each node to path-level weights, and then element-wise multiplication is performed with the effect quantification index to achieve a comprehensive attribution score. The calculation is performed using the following formula:

[0168] ,

[0169] in, For the first in the path The intervention sensitivity weighting coefficient of each node, Its corresponding cumulative causal effect value;

[0170] Furthermore, the comprehensive attribution scores of all paths are sorted, and a priority path index table is generated by sorting in descending order. Paths with attribution scores below a preset threshold are removed, and the set of highly important paths is retained to form a priority-ranked list of dominant causal paths.

[0171] By using intervention sensitivity weighted integral and ranking algorithms, the candidate attribution path set from the previous step is transformed into a structured priority-driven causal path index, thereby achieving more accurate path selection and ranking, and providing highly reliable input for the construction of subsequent intervention suggestions.

[0172] For example, in a patient case, the candidate attribution path set contains 15 paths with a total number of nodes ranging from 3 to 6. The weight allocation function of the intervention sensitivity weighted integral algorithm is configured as a linear normalized mapping, and the sensitivity sequence values ​​range from 0.05 to 0.42. For a path containing 5 nodes, the weights of each node are... to The cumulative causal effects at each node are 0.42, 0.35, 0.28, 0.15, and 0.05, respectively. to The scores were 0.8, 0.6, 0.55, 0.3, and 0.1, respectively. The overall attribution score was calculated using the formula. The result was 1.018. The same calculation was performed on all 15 paths, and the scores ranged from 0.15 to 1.25. After sorting, the 6 paths with scores ≥0.85 were retained as the list of dominant causal paths and provided to S7.4 to construct multi-dimensional intervention suggestion knowledge tuples, which significantly improved the prediction of intervention effect.

[0173] S7.4: Based on the modifiable variables corresponding to each path in the list of dominant causal paths, and combined with the estimated expected effect under virtual intervention and the Bayes confidence interval, construct a multi-dimensional intervention suggestion knowledge tuple. Each knowledge tuple includes the variable name, recommended operation direction, effect improvement magnitude and uncertainty range, and generate a structured priority intervention suggestion list.

[0174] S7.5: The initial risk score, the structured priority intervention recommendation list, the list of dominant causal pathways and their corresponding visual causal pathway diagram elements are encapsulated in a multi-channel information format, and a standardized comprehensive decision support report document is generated using Extensible Markup Language (XML) format. The report is then output to the clinical terminal through a human-computer interaction interface for doctors to refer to and implement.

[0175] Step S8: Update the edge weights in the individualized dynamic causal graph based on the actual intervention results from subsequent follow-up feedback, and optimize the model's long-term adaptability to the causal structure using an incremental learning mechanism to achieve closed-loop evolution of the causal inference system. Specifically, this includes:

[0176] S8.1: Obtain data on the implementation of actual intervention measures and changes in implant status from the patient's regular postoperative follow-up records, including osseointegration progress, occlusal load adjustment records, infection control effectiveness, and outcome indicators such as whether loosening or dislodgement occurs, as real feedback input; based on this feedback input, identify the deviation signal between the initial predicted risk and the actual clinical outcome, and label the corresponding key interventionable variables and their observed value change sequences to generate an incremental learning sample set for model calibration;

[0177] S8.2: Input the incremental learning sample set into the constructed individualized dynamic causal graph, locate the key regulatory nodes (i.e., intervention variables) involved and the local causal path segments formed by their direct successor variables; based on the Bayesian update rule, calculate the posterior causal strength probability distribution of each directed edge under the new evidence, so as to quantify the degree of support of the current group data for the original causal weights, and generate a preliminary updated edge weight parameter set;

[0178] S8.3: Perform gradient consistency test on the initially updated edge weight parameter set, and compare the consistency between the intervention sensitivity sequence obtained by the counterfactual perturbation simulation and the actual intervention response trend; if the two directions are inconsistent, introduce sparse regularization constraints to attenuate the abnormal edge weights, suppress the false causal association enhancement caused by confounding bias or measurement noise, and generate the robust optimized edge weight update amount.

[0179] S8.4: Based on the robustness-optimized edge weight update amount, perform online incremental update operation on the edge weights in the individualized dynamic causal graph, retaining the prior of historical causal structure while integrating the latest clinical feedback information; during the update process, a sliding time window weighting strategy is adopted to give higher confidence weight to recent follow-up data to reflect the time-varying characteristics of the treatment environment and patient group characteristics, and generate a time-adaptive updated version of the individualized dynamic causal graph.

[0180] S8.5: Synchronize the updated individualized dynamic causal graph to the auxiliary branch of the dual-branch deep neural network model, reinitialize the causal prior matrix on which the attention gating mechanism depends, and trigger a round of lightweight fine-tuning training to align the prediction performance of the main branch with the attribution logic of the auxiliary branch; after completing the model synchronization, store the updated model version and the corresponding causal graph snapshot to form an iterative instance of a causal inference system with closed-loop evolution capability.

[0181] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0182] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring and analyzing dental implant data based on deep learning, characterized in that, Includes the following steps: S1: Collect patients' preoperative imaging data, clinical examination records, and surgical operation parameters to form a multimodal structured input dataset; S2: Perform standardized preprocessing on the multimodal structured input dataset to generate a unified dimension feature vector; S3: Input the unified dimension feature vector into a dual-branch deep neural network model, where the main branch performs the implant failure risk prediction task and outputs an initial risk score; the auxiliary branch uses the set of interventionable clinical variables as anchor points and uses an attention gating mechanism to screen for potential confounding factors that are strongly correlated with them, and generates a local causal association strength matrix. S4: Based on the local causal association strength matrix and the initial risk score, construct an individualized dynamic causal graph, where nodes represent clinical variables, directed edges represent the direction and strength of causal influence between variables, their weights are initialized by combining historical population data with Bayesian update rules, and interveneable variables are marked as key regulatory nodes. S5: Apply counterfactual perturbation simulation to the key control nodes in the individualized dynamic causal graph, calculate the predicted output gradient response caused by the change of each intrusive variable under virtual intervention, and generate a variable intervention sensitivity sequence; S6: Based on the variable intervention sensitivity sequence, perform attribution analysis on the paths in the individualized dynamic causal graph to identify key causal propagation paths from the modifiable variable to the high-risk output, and generate a priority intervention recommendation list according to the expected intervention effect.

2. The method for monitoring and analyzing dental implant data based on deep learning according to claim 1, characterized in that, Following step S6, the following is also included: S7: Output a comprehensive decision support report including initial risk score, priority intervention recommendation list, visualized causal path diagram, and estimated expected effects of each intervention and their uncertainty intervals, for clinicians' reference; S8: Update the edge weights in the individualized dynamic causal graph based on the actual intervention results from subsequent follow-up feedback, and use an incremental learning mechanism to optimize the model's long-term adaptability to the causal structure.

3. The method for monitoring and analyzing dental implant data based on deep learning according to claim 1, characterized in that, Step S1 specifically includes: Preoperative cone-beam CT images of patients were acquired, and bone density distribution maps and three-dimensional morphological parameters of alveolar bone were extracted based on voxel intensity threshold segmentation algorithm to generate a quantitative bone structure feature matrix. The oral soft tissue condition, periodontal probing depth, and overall health indicators were collected from clinical examination records. Unstructured text descriptions were transformed into standardized classification coding vectors using medical ontology mapping rules to generate a subset of systemic risk factor features. Extract the implant position coordinates, major axis tilt angle and distance parameters to adjacent anatomical structures recorded in the digital surgical planning system, combine them with the actual implantation trajectory deviation data in the intraoperative navigation log, calculate the three-dimensional positioning error tensor of the implant based on the spatial Euclidean distance and angle difference function, and generate operation-related feature variables; The distribution of occlusal contact points and peak force sequences of opposing teeth are collected by an integrated dynamic occlusal force sensing pad. High stress concentration areas are identified through a spatiotemporal clustering algorithm, and the occlusal load asymmetry index is calculated to generate biomechanical input features. The quantitative bone structure feature matrix, the systemic risk factor feature subset, the operation-related feature variables, and the biomechanical input features are summarized, and semantic alignment is performed based on a unified medical terminology set to construct a multimodal structured input dataset, and a standardized case master index association table is output.

4. The method for monitoring and analyzing dental implant data based on deep learning according to claim 3, characterized in that, The quantitative bone structure feature matrix includes cortical bone thickness, mean cancellous bone mineral density, and bone mass grading.

5. The method for monitoring and analyzing dental implant data based on deep learning according to claim 1, characterized in that, Step S2 specifically includes: A multimodal structured input dataset is obtained, and missing data items in each field are identified based on a missing value detection algorithm. The missing values ​​are then filled in using the K-nearest neighbor imputation method according to the clinical characteristic distribution of similar patient groups, generating a clinical data matrix with enhanced integrity. Based on the clinical data matrix with enhanced integrity, standardization is performed on the continuous variable fields to obtain a normalized subset of numerical features. One-hot encoding transformation is performed on the categorical variable field to generate a corresponding binary encoding vector based on the cardinality of the categorical variable, and then concatenated to the normalized numerical feature subset to form a high-dimensional sparse feature representation. The high-dimensional sparse feature representation is input into the feature dimension alignment module, and a compact, uniform-dimensional feature vector is generated through principal component analysis algorithm.

6. The method for monitoring and analyzing dental implant data based on deep learning according to claim 1, characterized in that, Step S3 specifically includes: Obtain a standardized, preprocessed, uniform-dimensional feature vector and use it as input data for a dual-branch deep neural network model. Based on the input data, design a shared representation layer at the front end of the model and use a multi-layer fully connected network to extract high-order nonlinear features and generate a shared latent feature representation. The shared latent feature representation is input into the main branch network. A nonlinear mapping transformation is performed based on the forward propagation process of the main branch network. The sigmoid activation value of the final output layer is calculated to generate an initial risk score in scalar form. Define a set of interventionable clinical variables. Based on the controllable properties of each variable in the set of interventionable clinical variables in clinical practice, use them as guiding anchors for auxiliary branches. Extract the embedding vectors of the corresponding variables from the shared latent feature representation to form an anchor-guided feature subspace. An attention gating mechanism is introduced in the auxiliary branch to calculate the correlation weight between the anchor-guided feature subspace and the globally shared latent features. The correlation weight is then subjected to softmax normalization to obtain the attention allocation probability of each potential confounding factor relative to the intrusive variable. Based on the attention allocation probability, the feature response is weighted and aggregated to generate a local causal association strength matrix. The local causal correlation strength matrix is ​​sparsified, and L1 regularization is used to filter significant correlation terms and remove weak correlation noise interference to generate an optimized local causal correlation strength matrix.

7. The method for monitoring and analyzing dental implant data based on deep learning according to claim 6, characterized in that, The set of modifiable clinical variables includes bone mineral density, occlusal force distribution, implant placement angle, postoperative infection control level, and soft tissue healing status.

8. The method for monitoring and analyzing dental implant data based on deep learning according to claim 1, characterized in that, Step S4 specifically includes: Obtain the local causal association strength matrix output by the auxiliary branch and the initial risk score generated by the main branch. Perform variable node initialization processing based on the local causal association strength matrix to determine the node set in the individualized dynamic causal graph. Each node corresponds to a structured clinical feature or imaging parameter, forming a variable node set. Based on the set of variable nodes, the attention weights in the local causal association strength matrix are used as the basis for the initial connection strength. A directed edge generation operation is performed, and a threshold truncation method is used to filter weak associations to generate a preliminary causal connection structure and output a preliminary causal topology graph. The preliminary causal topology graph is input into the Bayesian prior fusion module. The group-level stable causal pattern trained based on historical group data is used as the prior distribution. The initial weight of each directed edge is updated using Bayesian methods. The posterior edge weight is calculated to generate a set of dynamic causal edge weights that have been corrected by prior. Based on the prior-corrected set of dynamic causal edge weights, a graph structure optimization operation is performed. A sparse constraint regularization method is used to remove redundant reverse connections and loop paths. At the same time, a feedback adjustment mechanism is introduced to allow postoperative follow-up variables to form a reverse weight decay correction for preoperative nodes, generating a quasi-DAG structure that conforms to clinical temporal logic and outputting a structured and compliant individualized dynamic causal graph. In the structured and compliant individualized dynamic causal graph, based on a predefined set of interventionable clinical variables, the labeling mechanism performs node attribute annotation processing, uniformly marking nodes belonging to the set of interventionable clinical variables as key regulatory nodes and assigning them interventionable flag bits.

9. The method for monitoring and analyzing dental implant data based on deep learning according to claim 8, characterized in that, In the preliminary causal connection structure, each directed edge represents the potential causal influence direction from the source node to the target node, and its initial weight reflects the significance of the association between the two variables in the current sample context.

10. The method for monitoring and analyzing dental implant data based on deep learning according to claim 1, characterized in that, Step S5 specifically includes: Based on the set of key regulatory nodes in the individualized dynamic causal graph, the current observation sequence of the modifiable clinical variables is obtained, and the perturbation range is determined by utilizing the historical distribution characteristics of the modifiable clinical variables to generate a virtual intervention sample set; The virtual intervention sample set is subjected to latent space mapping processing and input into the main branch model of a dual-branch deep neural network for forward propagation. The implant failure risk score output sequence corresponding to each perturbed sample is calculated. Based on the risk score of the original unperturbed sample, the risk change sequence under each virtual intervention is obtained. Based on the functional relationship between the risk change sequence and the corresponding virtual intervention intensity, the automatic differentiation technique is used to calculate the local gradient response value of each intrusive variable relative to the output of the main branch. The local gradient response values ​​are arranged according to the variable dimension to form an initial variable intervention sensitivity vector. The initial variable intervention sensitivity vector is standardized and normalized, and combined with the prior weight information of the corresponding edge in the individualized dynamic causal graph, the sensitivity estimation is optimized using the Bayesian weighted fusion rule to generate a calibrated variable intervention sensitivity sequence. Based on the calibrated variable intervention sensitivity sequence, a sorting operation is performed to identify the preceding highly sensitive interventionable variables that have the greatest impact on the predicted output. These preceding highly sensitive interventionable variables are marked as priority intervention targets, and the variable intervention sensitivity sequence and its sorting results are output.

Images