A fracture treatment scheme operation scheme simulation system based on big data analysis

Abstract

The application relates to the technical field of big data processing, in particular to a fracture treatment scheme operation scheme simulation system based on big data analysis, which comprises a data preprocessing module, an adversarial network module and a model training and optimization module; a plurality of heterogeneous teacher models are used to pretrain a general large-scale fracture treatment data set, and fracture treatment decision knowledge learned by the plurality of heterogeneous teacher models is transferred to a student model for transfer learning. The application overcomes the contradiction between the small number of comminuted fracture cases (sample sparsity) of a single hospital and the high individualization of medical resources and capabilities (resource difference) by fusing the strategies of generative adversarial network (GAN) sample expansion and heterogeneous teacher model knowledge distillation transfer learning.

Description

Technical Field

[0001] This application relates to the field of big data processing technology, and more specifically, to a simulation system for surgical plans of fracture treatment based on big data analysis. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] In fracture treatment, comminuted fractures are among the most complex injuries with the highest prognosis uncertainty due to their severe bone fragmentation, extreme instability of the fracture ends, and the difficulty of reduction and fixation. Choosing a treatment plan is extremely challenging, requiring comprehensive consideration of numerous individual factors such as the specific location of the fracture, the degree of comminutedness, the direction of displacement, soft tissue damage, the patient's age and physical condition, and underlying diseases. In clinical practice, it is often necessary to weigh the risks and benefits among a complex combination of various surgical approaches (such as open reduction and internal fixation, and minimally invasive percutaneous fixation), internal fixation devices (such as bone plates, intramedullary nails, and external fixators), and auxiliary techniques (such as bone grafting) to select the best treatment method for the patient with the best prognosis (such as better functional recovery and fewer complications).

[0004] To assist physicians in decision-making, surgical simulation systems based on big data analytics are now widely used. These systems typically employ complex statistical or machine learning models (such as deep learning, decision trees, and random forests) to deeply mine and analyze massive amounts of historical case data. Based on the input of new patient characteristics (such as fracture type and patient indicators), the system retrieves or learns from the database to generate the treatment plan with the best prognosis, and can simulate and model the potential final treatment effects of this plan (such as bone healing process, functional score prediction, and complication probability estimation). Finally, the simulation results are presented to physicians and patients in an intuitive way (such as 3D visualization simulation, data charts, and risk-benefit analysis reports) as an important auxiliary decision-making reference.

[0005] While existing big data-driven surgical simulation systems can uncover general patterns in historical data and recommend treatment plans with good prognoses in a "macro" sense, the optimal plans they recommend are usually "population-based optimal" derived from the overall statistical results of their training dataset, rather than "localized optimal" tailored to the specific medical resource allocation and capabilities of a particular hospital. Different hospitals vary significantly in terms of the precision of surgical equipment models, the types of instruments in stock, the availability of consumables, and, most importantly, the proficiency and preferences of their surgical teams in specific techniques (such as minimally invasive techniques and specific instrument operation). A theoretically optimal plan may be ineffective due to a lack of corresponding resources in the target hospital or a lack of relevant technical experience among doctors, potentially increasing risks (resource and capability mismatch risk). However, because the number of comminuted fracture case data samples that a single hospital can accumulate to meet the requirements of high-quality modeling (sample sparsity), its own data scale is usually insufficient to support the training of a dedicated big data model that can accurately capture its unique resource and capability constraints and recommend the most suitable localized treatment plan based on those conditions. The contradiction between "sample sparsity" and "resource variability" results in low feasibility of the simulation results of the recommended schemes and low accuracy of the final effect prediction when the existing simulation system is applied to different hospitals. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this application propose a surgical simulation system for fracture treatment based on big data analysis to address the technical problems mentioned in the background section above.

[0008] As a first aspect of this application, some embodiments of this application provide a surgical simulation system for fracture treatment based on big data analysis, including:

[0009] The data preprocessing module is configured to collect local hospital case datasets that include preliminary examination information, intermediate treatment information, and prognostic information.

[0010] The adversarial network module receives a local hospital case dataset and uses at least one adversarial generative neural network model to train and augment the local hospital case dataset to generate a simulated case dataset.

[0011] The simulated case dataset matches the local medical resources and capabilities represented by the hospital's local case dataset in terms of statistical distribution;

[0012] The model training and optimization module is configured as follows:

[0013] We pre-trained a general large-scale fracture treatment dataset using multiple heterogeneous teacher models, and then transferred the fracture treatment decision-making knowledge learned by the multiple heterogeneous teacher models to a single student model.

[0014] During transfer learning, the student model is trained a second time using a simulated case dataset generated by the adversarial network module.

[0015] The simulation module integrates a student model as the final localized simulation model, which is configured as follows:

[0016] The process involves inputting preliminary examination information of the target patient, and simulating the output of intermediate treatment information and corresponding prognostic prediction results for the potential fracture surgical plan developed for the target patient based on the input preliminary examination information.

[0017] This application overcomes the contradiction between the scarcity of comminuted fracture cases in a single hospital (sample sparsity) and the highly personalized nature of medical resources and capabilities (resource variability) by integrating a strategy of sample augmentation using Generative Adversarial Networks (GANs) and knowledge distillation transfer learning from heterogeneous teacher models. The core effect is that the simulated case dataset generated by the system not only significantly expands the sample size for localized modeling, but more importantly, it accurately matches the unique constraints of the target hospital (such as equipment models, instrument inventory, and doctors' specific skill levels). Combined with rich fracture treatment knowledge and pattern recognition capabilities transferred from general large-scale data, the resulting localized simulation student model can not only recommend feasible treatment plans under the existing resources and capabilities of the target hospital, but also reliably simulate and predict the execution process (intermediate treatment information) and actual prognostic effects (such as bone healing progress, functional scores, and complication probabilities) of these plans. This significantly improves the adaptability, feasibility, and accuracy of prognostic predictions of the recommended plans in the actual clinical environment, thereby bridging the gap between the macro-optimal plan and the localized optimal execution effect.

[0018] Furthermore, the data preprocessing module is configured as follows:

[0019] Using a structural feature extraction network, non-numerical or unstructured data in the collected preliminary examination information, intermediate treatment information, and prognostic information are converted into structured numerical feature vectors according to a preset standard encoding table.

[0020] Using an image feature extraction network, medical image data from the collected preliminary examination information, intermediate treatment information, and prognostic information are processed to extract structured visual feature vectors related to fracture morphology, degree of comminutedness, and surrounding soft tissue condition.

[0021] By utilizing feature organization networks, preliminary examination information, intermediate treatment information, and prognostic information are transformed into structured numerical feature vectors and structured visual feature vectors, respectively, to obtain digital examination information, digital treatment information, and digital prognostic information.

[0022] This data preprocessing module, through deep feature engineering, leverages the synergistic effects of structural feature extraction networks, image feature extraction networks, and feature organization networks to achieve standardized, deeply structured transformation, and efficient fusion of multi-source heterogeneous medical information (unstructured text / numerical data, multimodal images). It accurately transforms raw, messy, and difficult-to-model unnumerical / unstructured data (such as diagnostic descriptions and surgical procedure selections) into structured numerical feature vectors according to standard encoding. Simultaneously, using an image feature extraction network specifically designed for fracture scenarios, it precisely extracts quantitative visual feature vectors of fracture morphology (such as the number of fragments, displacement angle, and soft tissue injury grade) from raw medical images. The feature organization network unifies these key information from different dimensions into structured, high-quality digital examination, treatment, and prognostic information.

[0023] Furthermore, the adversarial network module includes:

[0024] The noise source generation module determines the noise source generation range based on the local medical conditions of the hospital, and generates noise signals within the noise source generation range;

[0025] The generator generates simulated samples based on noise signals. The simulated samples include simulated digital examination information, digital treatment information, and digital prognostic information.

[0026] The discriminator, forming an adversarial network with the generator, distinguishes the authenticity of input simulated samples from real samples;

[0027] The simulation data generator has a built-in pre-trained generator that generates new simulation samples based on new noise sequences, and uses these new simulation samples to generate a simulated case dataset.

[0028] This adversarial network module, through its unique "constrained noise source generation" and bidirectional adversarial training mechanism, effectively generates a high-quality simulated case dataset that statistically closely matches the unique medical resource capacity boundaries of the target hospital (such as specific equipment types, limited instrument inventory, and physician skill levels). The noise source generation module determines the noise generation range based on the hospital's specific conditions (such as equipment model list, range of operable surgical procedures, and physician team experience levels), constraining the generation space of simulated cases from the source and ensuring that new samples are aligned with local reality in the dimension of "resource supply capacity." Under this constraint, the generator and discriminator dynamically optimize through adversarial game, ultimately driving the generator to produce "pseudo-real" case samples that contain reasonable medical logic (such as matching fracture types with treatment plans) and naturally incorporate local resource constraint characteristics (such as prioritizing the use of inventory instruments and avoiding high-risk minimally invasive procedures). This mechanism significantly overcomes the problem of sample scarcity and accurately captures the key local factors that affect the feasibility and effectiveness of the plan. It provides sufficient, reliable, and truly localized training data that reflects "what can be implemented in this hospital and what kind of effect it has had in this hospital" for the subsequent training of the localized decision-making model (student model).

[0029] Furthermore, the noise signal is generated as follows:

[0030] S1: Load the medical resources of the local hospital, generate the original encoding table from the local medical resources, map the original encoding table to the standard encoding table, and determine the vertical dimension of the noise signal based on the dimension of the original encoding table.

[0031] Load the maximum dimension of the feature vector of real cases from the local hospital and use it as the horizontal dimension of the noise signal;

[0032] The vertical dimension is used to constrain the numerical fluctuation range of the noise signal, while the horizontal dimension is used to constrain the length of the noise signal.

[0033] S2: Predefine an entropy source seed signal source and collect random signals generated by the entropy source seed signal source;

[0034] S3: Based on the proportional information, each element in the random signal is sequentially mapped to the element at the corresponding position in the noise signal to generate the noise signal.

[0035] This noise signal generation mechanism limits the numerical range of noise based on the "vertical dimension" of the actual medical resources of the local hospital (mapped to standard codes) and defines the structural length of noise based on the "horizontal dimension" of the maximum feature complexity of real cases. This fundamentally ensures that the generated noise signal strictly adheres to the capability boundaries of the target hospital (such as available instrument types, equipment parameter limitations, and surgical procedure range) and can carry all the key features of local cases. At the same time, it uses high-entropy true random source seeds and generates initial random signals through proportional mapping rules. Under the premise of meeting the strict dual-dimensional constraints, it effectively ensures the diversity and unpredictability of the noise signal. This provides the Generative Adversarial Network with a high-quality "seed" that can accurately reflect local resource constraints and drive the generation of rich and reasonable medical logic simulation cases. It is the cornerstone for the subsequent generation of high-fidelity simulation case datasets whose statistical distribution closely matches the unique conditions of the target hospital.

[0036] Furthermore, the noise signal is , ; express The j-th element in the noise signal, where j represents the index of the element in the noise signal;

[0037] ;

[0038] in, This represents the value of the j-th element of the random signal. This represents the maximum value of the random signal. This represents the maximum value of the noise signal at the j-th bit. This indicates rounding down to the nearest integer.

[0039] The precise quantification calculation method for noise signal elements combines normalization proportional control with local resource upper limits to achieve deterministic and discretized transformation of high-entropy random source signals into noise elements that strictly conform to the boundary conditions of the target hospital's local resources and capabilities. The core effect is that it effectively applies the mathematically expressed "maximum values" of hospital resources and capabilities (such as the upper limit of a specific equipment model, the threshold of concurrent operation complexity allowed in the operating room, or the upper limit of a certain skill proficiency level of the doctor team) directly as a hard ceiling for noise element generation. By rounding down, it ensures that the value of each generated noise element is an integer and falls within the actual executable range of the target hospital in that feature dimension. Thus, the fundamental constraints of local resources are forcibly embedded at the most basic unit level of the noise space, laying a mathematical foundation for the subsequent generator to produce simulated cases that absolutely conform to the hospital's actual operating conditions (such as only using equipment models available in the inventory, and surgical complexity that can only be handled within the doctor's experience threshold).

[0040] Furthermore, the generator includes:

[0041] The generator input layer is used to receive the noise vector;

[0042] The feature transformation layer, based on fully connected layers and residual networks, extracts hidden features from noise vectors;

[0043] The generator output layer outputs simulated samples based on the hidden features of the input.

[0044] This generator network accurately receives noise signals under localized constraints through the input layer, and deeply analyzes the local resource capability boundary information and potential medical logic correlations contained in the noise through a feature transformation layer composed of fully connected and residual networks. The residual structure in the feature transformation layer effectively overcomes the gradient vanishing problem in deep networks, stably capturing the complex and nonlinear deep structures and correlations in fracture diagnosis and treatment (such as the implicit relationship between specific imaging features and complication risks, and the differential impact of different surgical procedures on the prognostic models of patients with specific physical conditions), ensuring that the generated samples still possess a high degree of medical logic authenticity and pattern diversity under strict resource constraints.

[0045] Furthermore, the discriminator includes:

[0046] The feature pyramid backbone network is used to take simulated or real samples as input to generate case features;

[0047] The true / false discrimination head, based on the Sigmoid function, regresses case features to the probability that the sample is true.

[0048] This discriminator uses a feature pyramid backbone network to deeply analyze simulated and real case samples (digital examination, treatment, and prognostic information). It utilizes its multi-level and multi-scale perception capabilities to accurately capture the complex, cross-dimensional medical logical relationships contained in the case data (such as the matching degree between fracture comminuted morphology and internal fixation device selection, and the implicit mapping between specific intraoperative operation details and postoperative complication risks) as well as the statistical patterns of key local resource constraints (such as the frequency of use of preferred device models and traces of specific team operating habits). Its authenticity discrimination head then efficiently focuses these high-dimensional and complex fusion features into the final sample authenticity probability judgment.

[0049] Furthermore, the loss functions for the generator and discriminator are:

[0050] ;

[0051] ;

[0052] ;

[0053] Let E represent the discriminator loss, and let E represent the mathematical expectation operator. Let x represent a real sample drawn from the real data distribution. This represents the penalty sample generated through interpolation. Let x represent a mixed random number, and let x represent a real sample. This represents the generated samples produced by the generator; Let z represent the noise vector sampled from the noise distribution. This represents the discriminator's score for the real sample. This indicates that the discriminator evaluates the generated samples. The rating, The weights of the gradient penalty term are represented by the coefficients. This represents a sample drawn from a penalty distribution. This indicates that the discriminator output is related to the input. The gradient; Let l represent the feature matching loss, and l represent the hierarchical index of the feature pyramid. This represents the feature map output by the discriminator for the real sample x at the l-th layer. This indicates that the discriminator generates samples for each pair. The feature map output at layer l. This represents the square of the L2 norm.

[0054] Furthermore, multiple heterogeneous teacher models include intermediate treatment information generation models and prognostic information generation models;

[0055] Among them, the intermediate treatment information generation model is trained based on a pre-prepared first sample set;

[0056] The prognostic information generation model is trained based on a pre-prepared second sample set;

[0057] The intermediate treatment information generation model generates optimal intermediate treatment information based on the input preliminary examination information.

[0058] The prognostic information generation model generates prognostic information based on the input preliminary examination information and intermediate treatment information.

[0059] This heterogeneous teacher model design decomposes treatment plan decisions into a model focused on generating optimal "intermediate treatment information" and a model focused on predicting "prognostic information." Both models are deeply pre-trained using large-scale professional sample sets, achieving a strong cohesion and abstraction of highly specialized knowledge inherent in the core decision-making chain of fracture treatment ("determining the treatment plan based on the injury" and "evaluating the effectiveness of the plan"). This division of labor enables the two teacher models to achieve expert-level accuracy and depth of understanding in optimal matching reasoning of treatment plans (inputting preliminary examinations and outputting treatment information) and accurate simulation of treatment plan effects (inputting preliminary examinations and treatment information and outputting prognostic information). This lays a solid foundation for the subsequent complete, independent, and faithful injection of these universal, big data-validated deep decision-making knowledge—especially the logic of generating high-value treatment plans and complex prognostic causal mapping relationships—into the local student model through transfer learning. This effectively improves the student model's ability to formulate feasible treatment plans with good prognostic prospects under local constraints and enhances the accuracy of its prognostic predictions.

[0060] Furthermore, the student model includes:

[0061] The feature input layer is used to input digital inspection information and extract feature information from the digital inspection information;

[0062] The treatment information generation layer outputs corresponding digital treatment information based on the input feature information.

[0063] The prognostic information generation layer generates digital prognostic information based on the digital treatment information and digital examination information output by the treatment information generation layer.

[0064] Furthermore, the intermediate treatment information generation model is signal-connected to the treatment information generation layer, and the error between the output of the intermediate treatment information generation model and the output of the treatment information generation layer is used as the first loss term.

[0065] The prognostic information generation model is connected to the prognostic information generation layer signal, and the output of the prognostic information generation model and the error between the prognostic information generation layer are used as the second loss term.

[0066] The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer;

[0067] The second loss term and the total loss of the student model are used to update the model parameters of the prognostic information generation layer.

[0068] This student model, through a modular, hierarchical design (treatment information generation layer and prognostic information generation layer), accurately inherits the chain-like decision-making capability (treatment plan generation → effect prediction) of the heterogeneous teacher model. It innovatively employs a dual-channel differentiated knowledge transfer mechanism: calculating the first loss term based on the output of the treatment information teacher model to guide the optimization of treatment layer parameters, and calculating the second loss term based on the output of the prognostic information teacher model to guide the updating of prognostic parameters. While strengthening localized learning (the total loss covers local cases generated by the adversarial network), it achieves two core values: 1) the treatment decision layer accurately absorbs the generation logic of universally optimal solutions (such as the deep mapping between fracture characteristics and instrument selection); 2) the prognostic prediction layer finely internalizes the universal laws of the solution-effect causal chain (such as the correlation between specific instruments and bone healing rate). Ultimately, it forms a decision-making closed loop that combines a global big data perspective (teacher knowledge) with local resource adaptability (simulated data training). This enables the system to perform highly reliable localized simulation predictions of the actual effects (prognostic information) of treatment plans implemented under target hospital conditions, completely opening up the clinical decision-making chain from "feasible solutions under resource constraints" to "precise prognostic assessment." Attached Figure Description

[0069] Figure 1 This is a flowchart of a simulation system for fracture treatment and surgical procedures based on big data analysis.

[0070] Figure 2 This is a training graph for the heterogeneous teacher and student models. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0072] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0073] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0074] refer to Figure 1 Example 1: A simulation system for fracture treatment surgical plans based on big data analysis includes: a data preprocessing module, an adversarial network module, a model training and optimization module, and a simulation module. The data preprocessing module, adversarial network module, model training and optimization module, and simulation module are connected sequentially, with the data preprocessing module and simulation module being signal-connected.

[0075] The data preprocessing module transforms the raw medical data (including text, records, and images such as CT scans) from the hospital into structured, numerical feature vectors, providing a unified and processable digital input for subsequent modules.

[0076] The data preprocessing module uses a structural feature extraction network (such as an encoder-transformer) to convert non-numerical examination information, treatment plans, and prognostic records into structured numerical feature vectors based on a standard medical coding table. Then, a specialized image feature extraction network (such as a convolutional neural network) is used to analyze medical images and extract structured vectors describing key visual features such as fracture morphology and comminuted details. Finally, a feature organization network integrates these two parts of information to form complete digital examination information, digital treatment information, and digital prognostic information that represent the patient's condition, thereby achieving standardized and structured expression of multi-source heterogeneous medical data.

[0077] Specifically, the data preprocessing module is configured to collect local hospital case datasets that include preliminary examination information, intermediate treatment information, and prognostic information;

[0078] The data preprocessing module is configured as follows:

[0079] Using a structural feature extraction network, non-numerical or unstructured data in the collected preliminary examination information, intermediate treatment information, and prognostic information are converted into structured numerical feature vectors according to a preset standard encoding table.

[0080] Using an image feature extraction network, medical image data from the collected preliminary examination information, intermediate treatment information, and prognostic information are processed to extract structured visual feature vectors related to fracture morphology, degree of comminutedness, and surrounding soft tissue condition.

[0081] By utilizing feature organization networks, preliminary examination information, intermediate treatment information, and prognostic information are transformed into structured numerical feature vectors and structured visual feature vectors, respectively, to obtain digital examination information, digital treatment information, and digital prognostic information.

[0082] Specifically, preliminary examination information, intermediate treatment information, and prognostic information are common information in case reports. If the information is incomplete, data entry is required.

[0083] Preliminary examination information refers to relevant information collected from the patient before surgical treatment, such as name, age, gender, medical history, CT images, and fracture location images.

[0084] Intermediate treatment information refers to the patient's surgical plan, such as plate and screw fixation, intramedullary nailing, closed reduction and external fixation, screw / tension screw fixation, and combined use.

[0085] Prognostic information refers to the patient's prognosis, which is generally the recovery status from surgery to discharge. It usually includes the wound healing status recorded by the nurse and whether there are any complications.

[0086] The preliminary examination information, intermediate treatment information, and prognostic information mentioned above all consist of textual information plus video and image information. This information cannot be directly used for subsequent processing. Therefore, a structural feature extraction network is used to convert the textual information into encoding. An image feature extraction network is used to convert the video and image information into structured visual feature vectors. This allows the preliminary examination information, intermediate treatment information, and prognostic information to be transformed into digital examination information, digital treatment information, and digital prognostic information.

[0087] Specifically, digital examination information, digital treatment information, and digital prognosis information are essentially the digital processing of information.

[0088] Specifically, digitized examination information essentially consists of the patient's written medical records and imaging examination data. The written medical records are converted into corresponding encoding tables according to pre-set encoding rules. For example, 0011 indicates that the patient is male and 22 years old. By simply arranging the data according to the rules and then using an encoding dictionary, the textual information can be converted into codes.

[0089] Imaging data includes various ultrasound, CT, and X-ray examination data. This data is then transformed into specific examination results using an image feature extraction network. For example, ultrasound data might show a grade I comminuted fracture (comminutedness) at a certain location. After being transformed into specific examination results, this data is further converted into coded data, thus obtaining digital examination information.

[0090] Digital treatment information encodes the treatment process. Generally speaking, fracture surgery plans can be roughly divided into:

[0091] Time of surgery (time since fracture);

[0092] Anesthesia method (general anesthesia, local anesthesia);

[0093] Surgical method (open surgery or external fixation);

[0094] Treatment of bone fragments (removal of bone fragments or preservation of bone fragments);

[0095] Fracture fixation methods (intramedullary nailing, plate fixation, and screw fixation);

[0096] Postoperative fixation range (whether it needs to cover the entire limb).

[0097] The above sections can be flexibly adjusted according to actual needs. For each section, corresponding data items can be added or removed and converted into corresponding codes to obtain digital prognostic information.

[0098] Digital prognostic information: This involves providing a phased description of the prognosis, for example:

[0099] Day 1 post-surgery: Awake, no fever;

[0100] Day 5 post-surgery: No wound infection.

[0101] Day 10 post-surgery: No wound infection, etc.

[0102] For each prognostic outcome, appropriate labels can be set according to actual needs, such as good, poor, and unfavorable. These labels can then be converted into codes to obtain digital prognostic information.

[0103] The adversarial network module receives a local hospital case dataset and uses at least one adversarial generative neural network model to train and augment the local hospital case dataset to generate a simulated case dataset.

[0104] The simulated case dataset is statistically matched to the local medical resources and capabilities represented by the hospital's local case dataset.

[0105] The number of cases in each hospital is relatively limited, and training a neural network with a limited number of cases generally yields poor results. Therefore, an adversarial network module was used to augment the data. Specifically:

[0106] The adversarial network module includes:

[0107] The noise source generation module determines the noise source generation range based on the local medical conditions of the hospital, and generates noise signals within the noise source generation range;

[0108] The generator generates simulated samples based on noise signals. The simulated samples include simulated digital examination information, digital treatment information, and digital prognostic information.

[0109] The discriminator, forming an adversarial network with the generator, distinguishes the authenticity of input simulated samples from real samples;

[0110] The simulation data generator has a built-in pre-trained generator that generates new simulation samples based on new noise sequences, and uses these new simulation samples to generate a simulated case dataset.

[0111] The noise signal is generated as follows:

[0112] S1: Load the medical resources of the local hospital, generate the original encoding table from the local medical resources, map the original encoding table to the standard encoding table, and determine the vertical dimension of the noise signal based on the dimension of the original encoding table.

[0113] Load the maximum dimension of the feature vector of real cases from the local hospital and use it as the horizontal dimension of the noise signal;

[0114] The vertical dimension is used to constrain the numerical fluctuation range of the noise signal, while the horizontal dimension is used to constrain the length of the noise signal.

[0115] S2: Predefine an entropy source seed signal source and collect random signals generated by the entropy source seed signal source;

[0116] S3: Based on the proportional information, each element in the random signal is sequentially mapped to the element at the corresponding position in the noise signal to generate the noise signal.

[0117] The noise signal is , ; express The j-th element in the noise signal, where j represents the index of the element in the noise signal;

[0118] ;

[0119] in, This represents the value of the j-th element of the random signal. This represents the maximum value of the random signal. This represents the maximum value of the noise signal at the j-th bit. This indicates rounding down to the nearest integer.

[0120] The generator includes:

[0121] The generator input layer is used to receive the noise vector;

[0122] The feature transformation layer, based on fully connected layers and residual networks, extracts hidden features from noise vectors;

[0123] The generator output layer outputs simulated samples based on the hidden features of the input.

[0124] The discriminator includes:

[0125] The feature pyramid backbone network is used to take simulated or real samples as input to generate case features;

[0126] The true / false discrimination head, based on the Sigmoid function, regresses case features to the probability that the sample is true.

[0127] The loss functions for the generator and discriminator are:

[0128] ;

[0129] ;

[0130] ;

[0131] Let E represent the discriminator loss, and let E represent the mathematical expectation operator. Let x represent a real sample drawn from the real data distribution. This represents the penalty sample generated through interpolation. Let x represent a mixed random number, and let x represent a real sample. This represents the generated samples produced by the generator; Let z represent the noise vector sampled from the noise distribution. This represents the discriminator's score for the real sample. This indicates that the discriminator evaluates the generated samples. The rating, The weights of the gradient penalty term are represented by the coefficients. This represents a sample drawn from a penalty distribution. This indicates that the discriminator output is related to the input. The gradient; Let l represent the feature matching loss, and l represent the hierarchical index of the feature pyramid. This represents the feature map output by the discriminator for the real sample x at the l-th layer. This indicates that the discriminator generates samples for each pair. The feature map output at layer l. This represents the square of the L2 norm.

[0132] Thus, by jointly training the generator and discriminator, a highly accurate generator can be obtained. By inputting random noise into the generator, the required number of simulated samples can be obtained. After obtaining a sufficient number of simulated samples, the student model can be trained.

[0133] The model training and optimization module is configured as follows:

[0134] We pre-trained a general large-scale fracture treatment dataset using multiple heterogeneous teacher models, and then transferred the fracture treatment decision-making knowledge learned by the multiple heterogeneous teacher models to a single student model.

[0135] During transfer learning, the student model is trained a second time using a simulated case dataset generated by the adversarial network module.

[0136] Multiple heterogeneous teacher models include intermediate treatment information generation models and prognostic information generation models;

[0137] Among them, the intermediate treatment information generation model is trained based on a pre-prepared first sample set;

[0138] The prognostic information generation model is trained based on a pre-prepared second sample set;

[0139] The intermediate treatment information generation model generates optimal intermediate treatment information based on the input preliminary examination information.

[0140] The prognostic information generation model generates prognostic information based on the input preliminary examination information and intermediate treatment information.

[0141] The preliminary examination information, intermediate treatment information, and prognostic information are in the form of the aforementioned digital examination information, digital treatment information, and digital prognostic information. The intermediate treatment information generation model is a pre-trained decision-making model based on a deep residual neural network architecture. Its input is standardized and structured digital examination information of the patient, and its output is digital treatment information adapted to the input injury. The model is trained under supervision on a first sample set composed of high-quality treatment cases from multiple centers worldwide. It learns the high-dimensional mapping logic between fracture type, patient physiological parameters, and optimal treatment plan, forming a treatment strategy generator with general medical decision-making capabilities.

[0142] The prognostic information generation model is a spatiotemporal predictive neural network employing a multimodal feature fusion mechanism. Its input includes digital examination information (the patient's initial injury) and digital treatment information (the implemented treatment plan), and its output is structured digital prognostic information. Model training relies on a second sample set with complete follow-up records (covering prognostic trajectories over 6-24 months under different treatment plans). By modeling a dynamic causal chain of "injury-treatment-outcome," it achieves refined simulation prediction of the effectiveness of specific treatment plans.

[0143] The intermediate treatment information generation model and the prognostic information generation model are large-scale neural network models. Their specific structures are based on existing technologies. Given the model inputs and outputs, deep residual neural networks and spatiotemporal prediction neural networks can be used to construct the required intermediate treatment information generation model and prognostic information generation model, respectively. In practice, it is necessary to construct intermediate treatment information generation models and prognostic information generation models with complex hierarchical structures and high depth to increase their data processing and learning capabilities.

[0144] The key to this application is to use a general, large-scale intermediate treatment information generation model and a prognostic information generation model to distill and learn from the local student model, thereby increasing the applicability of the student model locally.

[0145] refer to Figure 2 The student model includes:

[0146] The feature input layer is used to input digital inspection information and extract feature information from the digital inspection information.

[0147] The feature input layer is a fully connected feature fusion network used to extract feature information from digital inspection information and then concatenate the feature information.

[0148] The treatment information generation layer outputs corresponding digital treatment information based on the input feature information.

[0149] The treatment information generation layer is a Transformer-based sequential decision-making network that simulates the progressive decision-making process of clinical treatment pathways. It treats treatment plan generation as a multi-step sequential decision-making task: logically decomposing it into specific operational steps (e.g., surgical plan → rehabilitation plan). Each step dynamically selects the optimal operation from a pre-defined set of feasible operations using a mini-classifier.

[0150] The prognostic information generation layer generates digital prognostic information based on the digital treatment information and digital examination information output by the treatment information generation layer.

[0151] The prognostic information generation layer is a gated feature interaction network that constructs a mapping relationship between "patient baseline condition + treatment plan → prognostic outcome". Treatment plan vectors are decoded into computable embedding representations (e.g., different surgical methods correspond to different numerical effects) and dynamically weighted in fusion with patient examination features. The gating unit automatically adjusts the contribution weights of "patient baseline condition" and "treatment plan" to the prognosis, and finally, the residual network performs multi-level nonlinear transformations to output the corresponding digital prognostic information.

[0152] The specific functions and structures of the feature input layer, treatment information generation layer, and prognostic information generation layer have been described. This application will not further describe the specific network structures, input data formats, or data output formats of these layers. The key aspect of this application lies in the corresponding training design that sets different loss terms for each functional part of the student model.

[0153] The intermediate treatment information generation model is connected to the treatment information generation layer signal, and the error between the output of the intermediate treatment information generation model and the output of the treatment information generation layer is used as the first loss term.

[0154] The prognostic information generation model is connected to the prognostic information generation layer signal, and the output of the prognostic information generation model and the error between the prognostic information generation layer are used as the second loss term.

[0155] The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer;

[0156] The second loss term and the total loss of the student model are used to update the model parameters of the prognostic information generation layer.

[0157] The total loss of the student model is the difference between the final output obtained by the student model when inputting training samples and the labels in the training samples. The total loss describes the difference between the student model's predicted data and the labeled data, and is used to adjust the network parameters of the student model in reverse based on the difference between the predicted data and the labeled data.

[0158] The student model is trained through a two-stage knowledge transfer and local data fine-tuning. After obtaining a training sample, the training sample is input into both the student model and the teacher model. The output of the intermediate treatment information generation model in the teacher model is combined with the output of the treatment information generation layer to calculate the first loss term.

[0159] Then, the output of the prognostic information generation model in the teacher model and the output of the updated prognostic information generation layer are used to calculate the second loss term;

[0160] Finally, the total loss of the student model is calculated by comparing its output with the labels of the training samples.

[0161] Then, based on the weighted parameters, the total loss and the first loss term are fused into a first loss function value, which is then used to correct the model parameters of the treatment information generation layer.

[0162] The total loss and the second loss term are merged into a second loss function value based on the weighted parameters, and the second loss function value is used to correct the model parameters of the treatment information generation layer.

[0163] Thus, compared to traditional neural network model training methods, this approach replaces the previous method of using a single total loss to guide the updating of parameters within the entire model with using two different loss statements to replace the model parameters of two different network layers. This ensures the training accuracy of the student model.

[0164] How to update model parameters based on loss function values ​​is an existing technique, such as using the Adam tool to update model parameters.

[0165] The prognostic information generation layer integrates examination information with its own generated treatment information, outputs a prognostic prediction result, and compares it with the prediction effect of the prognostic information generation model to generate a second loss term, driving the update of prognostic parameters. In this process, both teacher loss terms are jointly backpropagated with the supervision loss (total loss) of local real cases. Through a weighted fusion mechanism, global knowledge inheritance and local feature adaptation are coordinated, ultimately enabling the model to simultaneously improve the localization feasibility of treatment plans and the clinical reliability of prognostic assessment in continuous iteration.

[0166] The above outlines the specific structure of the student model, the construction method of the teacher model, and the training principle. The specific training process will not be described here. The general logic is as follows: During training, each training sample is input into both the student and teacher models, resulting in a first loss term, a second loss term, and a total loss term. The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer, and the second loss term and the total loss of the student model are used to update the model parameters of the prognostic information generation layer, until the training result of the student model reaches the expected level. Then, the student model that has reached the expected level is placed into the simulation module.

[0167] The simulation module integrates a student model as the final localized simulation model, which is configured as follows:

[0168] The process involves inputting preliminary examination information of the target patient, and simulating the output of intermediate treatment information and corresponding prognostic prediction results for the potential fracture surgical plan developed for the target patient based on the input preliminary examination information.

[0169] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A surgical simulation system for fracture treatment based on big data analysis, characterized in that, include: The data preprocessing module is configured to collect local hospital case datasets that include preliminary examination information, intermediate treatment information, and prognostic information. The adversarial network module receives a local hospital case dataset and uses at least one adversarial generative neural network model to train and augment the local hospital case dataset to generate a simulated case dataset. The simulated case dataset matches the local medical resources and capabilities represented by the hospital's local case dataset in terms of statistical distribution; The model training and optimization module is configured as follows: We pre-trained a general large-scale fracture treatment dataset using multiple heterogeneous teacher models, and then transferred the fracture treatment decision-making knowledge learned by the multiple heterogeneous teacher models to a single student model. During transfer learning, the student model is trained a second time using a simulated case dataset generated by the adversarial network module. The simulation module integrates a student model as the final localized simulation model, which is configured as follows: Input the preliminary examination information of the target patient; and simulate the intermediate treatment information process and corresponding prognostic information prediction results of the potential fracture surgical plan for the target patient based on the input preliminary examination information. The data preprocessing module is configured as follows: Using a structural feature extraction network, non-numerical or unstructured data in the collected preliminary examination information, intermediate treatment information, and prognostic information are converted into structured numerical feature vectors according to a preset standard encoding table. Using an image feature extraction network, medical image data from the collected preliminary examination information, intermediate treatment information, and prognostic information are processed to extract structured visual feature vectors related to fracture morphology, degree of comminutedness, and surrounding soft tissue condition. By using feature organization networks, the preliminary examination information, intermediate treatment information and prognostic information are transformed into structured numerical feature vectors and structured visual feature vectors, respectively, to obtain digital examination information, digital treatment information and digital prognostic information; The adversarial network module includes: The noise source generation module determines the noise source generation range based on the local medical conditions of the hospital, and generates noise signals within the noise source generation range; The generator generates simulated samples based on noise signals. The simulated samples include simulated digital examination information, digital treatment information, and digital prognostic information. The discriminator, forming an adversarial network with the generator, distinguishes the authenticity of input simulated samples from real samples; The simulation data generator has a built-in pre-trained generator that generates new simulation samples based on new noise sequences, and uses these new simulation samples to generate a simulated case dataset. The noise signal is generated as follows: S1: Load the medical resources of the local hospital, generate the original encoding table from the local medical resources, map the original encoding table to the standard encoding table, and determine the vertical dimension of the noise signal based on the dimension of the original encoding table. Load the maximum dimension of the feature vector of real cases from the local hospital and use it as the horizontal dimension of the noise signal; The vertical dimension is used to constrain the numerical fluctuation range of the noise signal, while the horizontal dimension is used to constrain the length of the noise signal. S2: Predefine an entropy source seed signal source and collect random signals generated by the entropy source seed signal source; S3: Based on the proportional information, each element in the random signal is sequentially mapped to the element at the corresponding position in the noise signal to generate the noise signal; The noise signal is , ; express The j-th element in the noise signal, where j represents the index of the element in the noise signal; ； in, This represents the value of the j-th element of the random signal. This represents the maximum value of the random signal. This represents the maximum value of the noise signal at the j-th bit. This indicates rounding down to the nearest integer.

2. The fracture treatment surgical simulation system based on big data analysis according to claim 1, characterized in that, The generator includes: The generator input layer is used to receive the noise vector; The feature transformation layer, based on fully connected layers and residual networks, extracts hidden features from noise vectors; The generator output layer outputs simulated samples based on the hidden features of the input.

3. The fracture treatment surgical simulation system based on big data analysis according to claim 2, characterized in that, The discriminator includes: The feature pyramid backbone network is used to take simulated or real samples as input to generate case features; The true / false discrimination head, based on the Sigmoid function, regresses case features to the probability that the sample is true.

4. The fracture treatment surgical simulation system based on big data analysis according to claim 3, characterized in that, The loss functions for the generator and discriminator are: ； ； ； Let E represent the discriminator loss, and let E represent the mathematical expectation operator. Let x represent a real sample drawn from the real data distribution. This represents the penalty sample generated through interpolation. Let x represent a mixed random number, and let x represent a real sample. This represents the generated samples produced by the generator; Let z represent the noise vector sampled from the noise distribution. This represents the discriminator's score for the real sample. This indicates that the discriminator evaluates the generated samples. The rating, The weights of the gradient penalty term are represented by the coefficients. This represents a sample drawn from a penalty distribution. This indicates that the discriminator output is related to the input. The gradient; Let l represent the feature matching loss, and l represent the hierarchical index of the feature pyramid. This represents the feature map output by the discriminator for the real sample x at the l-th layer. This indicates that the discriminator generates samples for each pair. The feature map output at layer l. This represents the square of the L2 norm.

5. The fracture treatment surgical simulation system based on big data analysis according to claim 1, characterized in that, The student model includes: The feature input layer is used to input digital inspection information and extract feature information from the digital inspection information; The treatment information generation layer outputs corresponding digital treatment information based on the input feature information. The prognostic information generation layer generates digital prognostic information based on the digital treatment information and digital examination information output by the treatment information generation layer.

6. The fracture treatment surgical simulation system based on big data analysis according to claim 5, characterized in that, The intermediate treatment information generation model is connected to the treatment information generation layer signal, and the error between the output of the intermediate treatment information generation model and the output of the treatment information generation layer is used as the first loss term. The prognostic information generation model is connected to the prognostic information generation layer signal, and the output of the prognostic information generation model and the error between the prognostic information generation layer are used as the second loss term. The first loss term and the total loss of the student model are used to update the model parameters of the treatment information generation layer; The second loss term and the total loss of the student model are used to update the model parameters of the prognostic information generation layer.

Images