Digestive system department prediction analysis system based on big data

Abstract

The invention discloses a big data-based digestive system department prediction analysis system. The system comprises a data acquisition module, a knowledge graph fusion module, a disease course evolution prediction module and an intelligent diagnosis and treatment decision module. The invention relates to the technical field of medical health information, in particular to a big-data-based predictive analysis system for the department of gastroenterology. According to the scheme, a physiological signal is reconstructed by using a multi-scale mask auto-encoder, and the signal and a text are aligned; mapping knowledge graph regularization is introduced, the joint code is mapped to a digestive system knowledge graph space, and medical semantic representation is obtained; constructing a time sequence attenuation dynamic graph through the dynamic memory graph, and adding graph convolution gating; an embedded differential equation module simulates inflammation and tumor marker evolution to realize accurate prediction; patient state vectors are constructed based on multi-modal coding, medicines are rapidly selected by using a scoring function, a treatment process is modeled as a Markov decision, and a strategy network, a value network and PPO optimization are combined to realize balance of curative effect and income.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical and health information technology, in particular to a digestive internal medicine prediction analysis system based on big data. BACKGROUND

[0002] With the aging of the population and the change of lifestyle, the incidence of digestive internal medicine diseases such as gastritis, ulcer, inflammatory bowel disease and digestive tract tumor is rising, and the etiology is complex and the disease course is variable. The traditional diagnosis and treatment relies on endoscopic imaging, laboratory indexes and doctor's experience, and often there are problems such as delayed diagnosis, lack of individualized plan and high risk of recurrence. At the same time, multi-source heterogeneous data such as electronic medical records, laboratory tests and clinical texts accumulate rapidly in medical institutions, but existing systems focus on a single data type, and lack of collaborative analysis of multi-modal information such as physiological signals, medication records, endoscopic reports and symptom scales.

[0003] In order to meet the actual needs of precision medicine and intelligent decision-making, it is urgent to build a digestive internal medicine prediction analysis system based on big data, to evaluate and predict diseases based on big data, machine learning and statistical modeling technology; and intelligently provide diagnosis and treatment plan, help clinical team to realize early intervention, precise medication and resource optimization allocation, improve clinical efficiency and reduce risk. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides a large data-based prediction analysis system for a digestive internal medicine department, aiming at the problems that the multi-modal representation method in the traditional prediction analysis scheme for the digestive internal medicine department lacks prior semantic constraints in the medical field and is difficult to realize the deep alignment of physiological signals and endoscopic text, the present scheme performs coarse and fine granularity reconstruction on physiological signals through a multi-scale mask autoencoder, realizes cross-modal short-term prediction alignment of signals and text by using contrastive prediction coding, and introduces a knowledge graph regularization loss to jointly map the joint coding vector and the digestive system knowledge graph entity to the same space, so as to obtain a deep representation that is both signal-sensitive and rich in medical semantics; aiming at the problems that the time series prediction model in the traditional prediction analysis scheme for the digestive internal medicine department is difficult to capture the dynamic correlation between multiple examination events of a patient and the spatiotemporal evolution of biological fields such as inflammation and tumor markers, the present scheme nodes the physiological signals and text coding of each examination by a dynamic memory graph construction unit and constructs a dynamic graph according to the examination type and time decay, then parallelly introduces a graph convolution gating unit into the self-attention mechanism to fuse the structure information, and finally embeds a differentiable PDE module to simulate the evolution of the inflammation intensity field and the tumor marker concentration; aiming at the problems that the decision support in the traditional prediction analysis scheme for the digestive internal medicine department relies too much on experience rules or only pursues short-term optimization and lacks long-term efficacy optimization, the present scheme first constructs a patient state vector by using multi-modal joint coding and splicing medication records, and quickly selects the optimal medication scheme in the stage through a scoring function; then models the treatment process as a Markov decision process, combines a policy network, a value network and a near-end policy optimization algorithm, and continuously updates the strategy based on the discount cumulative return and the advantage function.

[0005] The technical scheme adopted by the present application is as follows: the large data-based prediction analysis system for a digestive internal medicine department provided by the present application comprises a data acquisition module, a knowledge graph fusion module, a disease course evolution prediction module and an intelligent diagnosis and treatment decision module;

[0006] The data acquisition module acquires demographic data, blood index data, vital sign data, medication record data, endoscopic report data and the number of symptoms of the digestive internal medicine department;

[0007] The knowledge graph fusion module first performs multi-scale mask autoencoding on the physiological signal sequence; then uses contrastive prediction coding to simultaneously perform short-term cross-modal prediction alignment of the physiological signal coding and the endoscopic text coding on the time axis; finally, the obtained joint coding vector and the digestive system knowledge graph are jointly mapped to the same vector space, and the relationship is constrained by graph Laplacian regularization loss;

[0008] The disease evolution prediction module first encodes the physiological signals and text of each examination as nodes, dynamically constructs and updates the memory graph according to the examination type and time decay coefficient, then introduces a graph convolution gate based on a normalized adjacency matrix into the multi-head cross attention in parallel, fuses with the standard self-attention to capture the timing and structural information, and finally uses a differentiable PDE module to numerically solve the spatiotemporal evolution of the inflammation intensity field and the tumor marker concentration, embeds the solver into the network, and learns the diffusion coefficient and the reaction term;

[0009] The intelligent diagnosis and treatment decision module first splices the latest multi-modal joint encoding vector of the patient and the medication record vector to form a state vector, scores all candidate medication schemes through a multi-layer perception machine scoring function, and selects the optimal scheme in the stage; subsequently, the treatment process is modeled as a Markov decision process, a strategy network and a state value network are defined, and the strategy parameters and the value parameters are alternately updated according to the policy gradient method by using the discount cumulative return and the advantage function; the discount return and the advantage are calculated by collecting the state, medication and treatment feedback at each step, and then the joint loss of the minimum strategy ratio expectation and the value network regression loss is minimized to perform a stochastic gradient descent until the discount cumulative return reaches a preset threshold, and finally the optimal medication strategy is output.

[0010] Further, the data acquisition module acquires demographic data, blood index data, vital sign data, medication record data, endoscopic report data and the number of symptoms of the department of gastroenterology, the demographic data specifically includes age, gender, BMI and family history of major digestive system diseases; the blood index data includes hemoglobin, white blood cell count, platelet count, C-reactive protein, tumor marker CA72-4, albumin and glutamic-pyruvic transaminase; the vital sign data includes heart rate, blood pressure and body temperature; the medication record data includes medication type, medication dosage and medication time; the endoscopic report data refers to the examination report text of a gastroscope or a colonoscope; and the number of symptoms of the department of gastroenterology refers to the number of symptoms such as difficulty swallowing, heartburn, stomach distension, nausea and vomiting, abdominal pain, liver pain, jaundice and biliary colic.

[0011] Further, the knowledge graph fusion module, on the physiological signals and endoscopic texts, jointly uses a multi-scale autoencoder and self-supervised prediction, and injects a knowledge graph constraint to obtain a deep vector, specifically including the following units:

[0012] The multi-scale mask autoencoding unit adopts multi-scale masks, that is, both coarse-grained large masks and fine-grained small masks, and a self-encoder with cross-scale skip connection to generate a physiological signal sequence matrix; two kinds of masks are constructed, and then reconstructed output is performed;

[0013] By comparing the predictive coding unit and using the hidden vector output by the encoder, short-term prediction is performed on the time axis. This incorporates the multi-scale coding of physiological signals and endoscopic text features into the comparison, enabling cross-modal prediction.

[0014] The knowledge graph regularization fusion unit maps physiological and text encoding vectors to the same space as knowledge graph nodes; it samples the adjacency matrix of the subgraph from the knowledge graph, constructs the Laplacian matrix, and obtains the graph regularization loss.

[0015] Furthermore, the disease progression prediction module specifically includes the following units:

[0016] The dynamic memory graph construction unit uses the statistical features of physiological signals from each patient's examination and the semantics of the report text as graph nodes. The node representation is obtained by fusion through a multilayer perceptron, and connections are established in the graph according to the examination type. The connection weights between nodes are calculated by the examination time difference and a preset decay rate. The influence on old nodes decays exponentially over time, thereby dynamically updating the memory graph.

[0017] The graph convolutional gating unit introduces the structural information of the dynamic graph into the multi-head cross attention. After normalizing the adjacency information of the graph, it calculates the graph convolutional gating path and self-attention path in parallel, and uses the gating mechanism to fuse the output at each node. The attention weights are dynamically adjusted according to the graph structure.

[0018] The inflammation and biomarker simulation unit treats the inflammation intensity field and tumor biomarker concentration as continuously changing spatiotemporal fields. It simulates diffusion and reaction processes in the network through a differentiable PDE module. The model automatically learns the diffusion coefficient and reaction term, and embeds the solver into the network for backpropagation, capturing the changes in inflammation diffusion and biomarker concentration on a spatiotemporal scale.

[0019] Furthermore, the intelligent diagnosis and treatment decision-making module specifically includes the following units:

[0020] The decision support unit first constructs a patient state vector using the patient's latest multimodal joint encoding vector and medication record vector; then it designs a scoring function and uses a deep neural network to evaluate and score the expected efficacy of each regimen in different states, assesses the expected benefit of the medication regimen in different states, and finally automatically selects the optimal regimen as the recommended medication decision for this stage.

[0021] The decision optimization unit models the entire treatment process as a dynamic decision system. It evaluates medication regimens through policy networks and value networks. First, it collects treatment feedback based on the patient's status at each time point and the current medication regimen, and calculates the cumulative discount reward by accumulating the treatment feedback. Then, it uses the policy gradient method and advantage function to measure the gain of each medication choice relative to the current policy.

[0022] The final decision-making unit first sets a cumulative discount reward threshold and executes multiple interaction sequences in a clinical setting using the current policy network, collecting the state, selected medication, and treatment feedback at each time step. Then, it calculates the cumulative discount reward for each sequence based on a discount factor and combines this with the current value network to calculate the advantage function. Next, it constructs a joint loss function based on the collected data and optimizes it using proximal policies. The loss function consists of the sum of the expected minimization term of the policy ratio and the mean squared error regression term of the value network. Subsequently, it uses a stochastic gradient descent optimization algorithm to alternately update the policy parameters and value network parameters using gradients. Finally, this process is repeated until the cumulative discount reward reaches the threshold, resulting in the optimal policy network. The final medication regimen is then output through the optimal policy network.

[0023] The beneficial effects achieved by the present invention using the above solution are as follows:

[0024] (1) In view of the problem that the multimodal representation method in the traditional gastroenterology predictive analysis scheme lacks the prior semantic constraints of the medical field and is difficult to achieve deep alignment between physiological signals and endoscopic text, this scheme reconstructs physiological signals with coarse and fine granularity through multi-scale mask autoencoder, realizes cross-modal short-term prediction alignment between signals and text by using contrastive predictive coding, and introduces knowledge graph regularization loss to map the joint coding vector and digestive system knowledge graph entities to the same space, thereby obtaining a deep representation that is both signal sensitive and rich in medical semantics.

[0025] (2) In view of the difficulty of capturing the dynamic correlation between multiple examination events and the spatiotemporal evolution of biofields such as inflammation and tumor markers in the time-series prediction analysis scheme of traditional gastroenterology, this scheme uses a dynamic memory graph construction unit to node the physiological signals and text codes of each examination and construct a dynamic graph based on the examination type and time decay. Then, a graph convolutional gating unit is introduced in parallel in the self-attention mechanism to fuse structural information. Finally, a differentiable PDE module is embedded to simulate the evolution of inflammation intensity field and tumor marker concentration, thereby achieving accurate prediction of complex disease course evolution.

[0026] (3) In view of the problem that decision support in traditional gastroenterology predictive analysis schemes often relies on empirical rules or only pursues short-term optimization and lacks long-term efficacy optimization, this scheme first uses multimodal joint coding and medication record splicing to construct patient state vectors, and quickly selects the optimal medication scheme for each stage through a scoring function; then the treatment process is modeled as a Markov decision process, and combined with policy network, value network and proximal policy optimization algorithm, the strategy is continuously updated with discount cumulative reward and advantage function to achieve a balance between short-term efficacy and long-term benefits. Attached Figure Description

[0027] Figure 1A schematic diagram of a gastroenterology predictive analysis system based on big data provided by the present invention;

[0028] Figure 2 This is a schematic diagram of the knowledge graph fusion module;

[0029] Figure 3 This is a schematic diagram of the disease progression prediction module;

[0030] Figure 4 This is a schematic diagram of the intelligent diagnosis and treatment decision-making module.

[0031] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0032] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0033] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0034] Example 1, see Figure 1 The present invention provides a gastroenterology predictive analysis system based on big data, including a data acquisition module, a knowledge graph fusion module, a disease course evolution prediction module, and an intelligent diagnosis and treatment decision module.

[0035] The data acquisition module collects demographic data, blood index data, vital sign data, medication record data, endoscopy report data, and the number of gastroenterology symptoms, and sends the data to the knowledge graph fusion module, the disease progression prediction module, and the intelligent diagnosis and treatment decision module.

[0036] The knowledge graph fusion module receives data sent by the data acquisition module, first performs multi-scale masking autoencoding on the physiological signal sequence; then, it uses contrastive predictive encoding to encode the physiological signal and endoscopic text on the time axis, while simultaneously performing short-term cross-modal prediction alignment; finally, it maps the obtained joint encoding vector and the digestive system knowledge graph to the same vector space, performs relation constraints through graph Laplacian regularization loss, and sends the data to the disease progression prediction module.

[0037] The disease progression prediction module receives data from the data acquisition module and the knowledge graph fusion module. It first uses the physiological signal encoding and text encoding of each examination as nodes, dynamically constructs and updates the memory graph according to the examination type and time decay coefficient. Then, it introduces graph convolution gate based on normalized adjacency matrix in parallel in multi-head cross attention, and fuses it with standard self-attention to capture temporal and structural information. Finally, it uses a differentiable PDE module to numerically solve the spatiotemporal evolution of the inflammation intensity field and tumor marker concentration. The solver is embedded in the network to learn the diffusion coefficient and response term, and the data is sent to the intelligent diagnosis and treatment decision module.

[0038] The intelligent diagnosis and treatment decision-making module receives data from the data acquisition module and the disease progression prediction module. First, it concatenates the patient's latest multimodal joint encoding vector with the medication record vector to form a state vector. Then, it scores all candidate medication plans using a multilayer perceptron scoring function to select the optimal plan for the current stage. Subsequently, it models the treatment process as a Markov decision process, defines a policy network and a state-value network, and uses a discount cumulative reward and advantage function to alternately update the policy parameters and value parameters according to the policy gradient method. By collecting state, medication, and treatment feedback at each step, it calculates the discount reward and advantage. Then, it performs stochastic gradient descent by minimizing the joint loss of the policy ratio expectation and the value network regression loss until the discount cumulative reward reaches a preset threshold, and finally outputs the optimal medication strategy.

[0039] Example 2, see Figure 1 This embodiment is based on the above embodiment. The data acquisition module collects demographic data, blood indicator data, vital sign data, medication record data, endoscopy report data, and the number of gastroenterological symptoms. The demographic data specifically includes: age, sex, BMI, and family history of major digestive system diseases; the blood indicator data specifically includes: hemoglobin, white blood cell count, platelet count, C-reactive protein, tumor marker CA72-4, albumin, and alanine aminotransferase; the vital sign data includes heart rate, blood pressure, and body temperature; the medication record data includes the type of medication, dosage, and time of administration; the endoscopy report data refers to the examination report text of gastroscopy or colonoscopy; the number of gastroenterological symptoms refers to the number of the following symptoms: dysphagia, heartburn, bloating, nausea and vomiting, abdominal pain, liver area pain, jaundice, and biliary colic.

[0040] Example 3, see Figure 1 and Figure 2 This embodiment, based on the above embodiment, further includes the following units: The knowledge graph fusion module combines multi-scale autoencoders and self-supervised predictions on physiological signals and endoscopic text, and injects knowledge graph constraints to obtain a depth vector.

[0041] The multi-scale masked autoencoder unit employs multi-scale masking: it includes both coarse-grained large-block masks and fine-grained small-block masks. The autoencoder incorporates cross-scale skip connections, making the physiological signal sequence matrix... ,in, For time steps, For the channel dimension; construct two types of masks: , Then reconstruct the output:

[0042] ;

[0043] in, This represents the original physiological signal sequence matrix, which is specifically composed of the patient's demographic data, blood index data, and vital sign data; and These represent the coarse-grained and fine-grained mask matrices, respectively, with 0 indicating that the mask is obscured. and These represent the encoder and decoder of an autoencoder, respectively. and These are learnable parameters; and Represents the cross-scale residual weights, satisfying ; This indicates bitwise multiplication. Denotes the Frobenius norm. This represents the loss from autoencoding reconstruction.

[0044] By contrasting the predictive coding unit and utilizing the hidden vectors output by the encoder, short-term predictions are performed on the time axis. This involves simultaneously incorporating multi-scale encoding of physiological signals and endoscopic text features for cross-modal prediction, as shown below:

[0045] ;

[0046] in, This represents the physiological signal encoding vector at time t. This represents the text encoding vector of the endoscopy report at time t. This indicates the original input of the endoscopy report text. This represents a pre-trained text encoder. This represents the temporal context of the combined physiological signals and text. Indicates parameters Gated loop unit; Indicates parameters Multi-step prediction function, input context With step size k, output the predicted vector. Indicates the first The predicted vector output by step; K represents the number of prediction steps, and N represents the total number of comparison samples. For temperature coefficient, This indicates taking the dot product of vectors. This represents the multi-step prediction loss value;

[0047] The knowledge graph regularization fusion unit maps physiological and textual encoding vectors to the same space as knowledge graph nodes, resulting in the following joint encoding vector sequence: ,in and These represent the physiological signal encoding vector and text encoding vector at the maximum moment, respectively; from the knowledge graph Mid-sampled subgraph adjacency matrix Construct the Laplace matrix , The graph regularization loss is Where Z represents the joint encoding matrix concatenated over time. This represents a knowledge graph of the digestive system, containing V entity nodes. Represents the adjacency matrix of a subgraph. Degree matrix, Represents the graph Laplace matrix. This represents the matrix trace operation.

[0048] By performing the above operations, this solution addresses the problem that traditional gastroenterology predictive analysis methods lack prior semantic constraints in the medical field and struggle to achieve deep alignment between physiological signals and endoscopic text. This solution uses a multi-scale masked autoencoder to reconstruct physiological signals at both coarse and fine granular levels, utilizes contrastive predictive coding to achieve cross-modal short-term prediction alignment between signals and text, and introduces knowledge graph regularization loss to map joint encoded vectors and digestive system knowledge graph entities to the same space, thereby obtaining a deep representation that is both signal-sensitive and rich in medical semantics.

[0049] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment, and further, the disease progression prediction module specifically includes the following units:

[0050] The dynamic memory graph construction unit uses the patient's physiological signal statistical characteristics and report semantics as graph nodes. The examination type determines the dynamic edges between nodes. The memory node features are updated over time, and the influence of old nodes is adaptively decayed according to the weights of new and old nodes. The time frame is... At that time, the node features are ,in, and These represent the patient's time period. The text encoding vector and physiological signal encoding vector of the endoscopy report, Represent the multilayer perceptron; then construct the time decay coefficient. ,in An index representing a point in time. This represents the decay rate; finally, a dynamic graph adjacency matrix is ​​defined, where the element in the i-th row and j-th column of the adjacency matrix... The value is defined as follows: when the time is and time When the check types are the same, the value is taken as Otherwise, the value is 0;

[0051] The graph convolutional gating unit is added to the multi-head cross-attention process. It utilizes adjacency information to modulate the attention weights, making the node feature matrix as follows: For dynamic graph adjacency matrix Perform Laplace normalization to obtain the normalized adjacency matrix: , The degree matrix is ​​represented; parallel computation of mutually gated graph convolutions and self-attention is performed, as shown below:

[0052] ;

[0053] in, , and These represent the query, key, and value in the attention mechanism, respectively. , and This represents the projection matrix in a Transformer. Indicates the output of convolution gate. and These represent the weights and biases of the graph convolution gate, respectively. This represents the sigmoid activation function. Represents the attention key dimension. This represents standard scaled dot product attention. Indicates gated fusion output;

[0054] The inflammation and biomarker simulation unit will simulate the intensity of inflammation. With tumor marker concentration Treating it as a continuous field, the spatiotemporal evolution is simulated through a differentiable PDE module. The PDE solution is embedded into the network backpropagation, allowing the model to directly learn the diffusion coefficient and reaction term, as shown below:

[0055] ;

[0056] in, Representing spatially continuous variables, For the simulated spatial range, express On the one-dimensional real number line; This represents the inflammation intensity field at location x in space and time t. This indicates the concentration of the tumor marker CA72-4 in the blood. Indicates the inflammation diffusion coefficient. and Based on node features And continuous fields, and response term functions trained by a three-layer multilayer perceptron; Indicates the marker clearance rate, Indicates the location of the checkpoint. and These represent the predicted values ​​of inflammation intensity and tumor marker concentration by the PDE solver, respectively.

[0057] By performing the above operations, this solution addresses the challenge of traditional gastroenterology predictive analysis schemes where time-series prediction models struggle to capture the dynamic correlations between multiple patient examinations and the spatiotemporal evolution of biofields such as inflammation and tumor markers. Instead, it addresses this by using a dynamic memory graph construction unit to node the physiological signals and textual codes from each examination and constructing a dynamic graph based on examination type and time decay. Furthermore, it introduces a graph convolutional gating unit in parallel within a self-attention mechanism to fuse structural information. Finally, it embeds a differentiable PDE module to simulate the evolution of the inflammatory intensity field and tumor marker concentration, achieving accurate prediction of complex disease progression.

[0058] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. Further, the intelligent diagnosis and treatment decision module specifically includes the following units:

[0059] The decision support unit specifically includes the following:

[0060] First, the patient's latest multimodal joint encoding vector is used. and medication record vector Construct the patient state vector: Then design the scoring function. To evaluate and determine the medication regimen In state Expected benefits:

[0061] ;

[0062] in, Indicates parameters The multilayer perceptron outputs a scalar score. This represents the set of all medication regimens, specifically composed of medication type, dosage, and administration time.

[0063] Find the optimal medication regimen for the stage: ;

[0064] The decision optimization unit specifically includes the following:

[0065] Viewing the treatment process as a Markov decision process, defining strategies and state value function Let the discount factor be The reinforcement learning objective is to maximize cumulative reward, as shown below:

[0066] ;

[0067] in, This indicates the cumulative return on the discount. This represents the treatment feedback at time t, specifically the change in the number of gastroenterology symptoms after medication administration. This indicates taking the average. Indicates the strategy parameters;

[0068] Then, a policy gradient update is performed, as shown below:

[0069] ;

[0070] in, This represents the value function of the medication regimen. and These represent the state and medication regimen at the initial time t=0, respectively; Represents the state value function. Represents the parameters of the value network. The parameter is The strategy in the state Below is the probability distribution formed by all possible medication regimens (act). Indicates medication regimen from Mid-sampling; The dominant function represents the state. Next, implement the medication regimen. Compared to the additional expected return that the average execution strategy could bring in this state; when This indicates that the state is... Next medication regimen This is better than the average level achieved by following the strategy, and the medication regimen needs to be improved. The probability when; Explain the medication regimen The performance was poor and needs to be reduced. The sampling probability;

[0071] The final decision-making unit first sets a cumulative discount reward threshold based on the current strategy network. Execute multiple interactive sequences in a clinical setting and collect the state at each time step. Selected medication and treatment feedback Then, for each sequence, apply a discount factor. Calculate the cumulative discounted return from time t. and in conjunction with the current value network Calculate the advantage function Next, a joint loss function is constructed based on the collected data. Through proximal policy optimization, the loss function consists of the sum of the expected minimization term of the policy ratio and the mean squared error regression term of the value network. Subsequently, a stochastic gradient descent optimization algorithm is used to optimize the policy parameters. and value network parameters Gradient updates are performed alternately; finally, this process is repeated until the cumulative discount reward reaches a threshold, resulting in the optimal policy network. The final medication plan is then output through the optimal policy network.

[0072] By performing the above operations, this solution addresses the problem that traditional gastroenterology predictive analysis schemes often rely on empirical rules for decision support or only pursue short-term optimization while lacking long-term efficacy optimization. First, it constructs a patient state vector using multimodal joint encoding and medication record concatenation, and quickly selects the optimal medication plan for each stage through a scoring function. Then, it models the treatment process as a Markov decision process, combining policy networks, value networks, and proximal policy optimization algorithms. The strategy is continuously updated using discount cumulative rewards and advantage functions to achieve a balance between short-term efficacy and long-term benefits.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0075] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A gastroenterology predictive analysis system based on big data, characterized in that: It includes a data acquisition module, a knowledge graph fusion module, a disease progression prediction module, and an intelligent diagnosis and treatment decision-making module; The data acquisition module collects demographic data, blood index data, vital sign data, medication record data, endoscopy report data, and the number of gastroenterology symptoms. The knowledge graph fusion module first performs multi-scale masking autoencoding on the physiological signal sequence; then, it uses contrastive predictive encoding to encode the physiological signal and the endoscopic text on the time axis, while simultaneously performing short-term cross-modal prediction alignment; finally, it maps the resulting joint encoding vector and the digestive system knowledge graph to the same vector space, and performs relation constraints through graph Laplacian regularization loss. The disease progression prediction module first uses the physiological signal encoding and text encoding of each examination as nodes, dynamically constructs and updates the memory graph according to the examination type and time decay coefficient, then introduces graph convolution gating based on normalized adjacency matrix in parallel in multi-head cross attention, and fuses it with standard self attention to capture temporal and structural information. Finally, it uses a differentiable PDE module to numerically solve the spatiotemporal evolution of the inflammation intensity field and tumor marker concentration, embedding the solver into the network to learn the diffusion coefficient and response term. The intelligent diagnosis and treatment decision-making module first concatenates the patient's latest multimodal joint encoding vector with the medication record vector to form a state vector. Then, it scores all candidate medication plans using a multilayer perceptron scoring function to select the optimal plan for the current stage. Subsequently, it models the treatment process as a Markov decision process, defines a policy network and a state-value network, and uses a discount cumulative reward and advantage function to alternately update the policy parameters and value parameters according to the policy gradient method. By collecting state, medication, and treatment feedback at each step, it calculates the discount cumulative reward and advantage. Then, it performs stochastic gradient descent by minimizing the joint loss of the policy ratio expectation and the value network regression loss until the discount cumulative reward reaches a preset threshold, and finally outputs the optimal medication strategy.

2. The gastroenterology predictive analysis system based on big data according to claim 1, characterized in that: The knowledge graph fusion module, based on physiological signals and endoscopic text, combines multi-scale autoencoders and self-supervised prediction, and injects knowledge graph constraints to obtain a depth vector, specifically including the following units: The multi-scale masking autoencoder unit employs multi-scale masks: both coarse-grained large-block masks and fine-grained small-block masks. The autoencoder has built-in cross-scale skip connections to generate a physiological signal sequence matrix. Two types of masks are constructed, and then the output is reconstructed. By comparing the predictive coding unit and using the hidden vector output by the encoder, short-term prediction is performed on the time axis. This incorporates the multi-scale coding of physiological signals and endoscopic text features into the comparison, enabling cross-modal prediction. The knowledge graph regularization fusion unit maps physiological and text encoding vectors to the same space as knowledge graph nodes; it samples the adjacency matrix of the subgraph from the knowledge graph, constructs the Laplacian matrix, and obtains the graph regularization loss.

3. The gastroenterology predictive analysis system based on big data according to claim 1, characterized in that: The disease progression prediction module specifically includes the following units: The dynamic memory graph construction unit uses the statistical features of physiological signals from each patient's examination and the semantics of the report text as graph nodes. The node representation is obtained by fusion through a multilayer perceptron, and connections are established in the graph according to the examination type. The connection weights between nodes are calculated by the examination time difference and a preset decay rate. The influence on old nodes decays exponentially over time, thereby dynamically updating the memory graph. The graph convolutional gating unit introduces the structural information of the dynamic graph into the multi-head cross attention. After normalizing the adjacency information of the graph, it calculates the graph convolutional gating path and self-attention path in parallel, and uses the gating mechanism to fuse the output at each node. The attention weights are dynamically adjusted according to the graph structure. The inflammation and biomarker simulation unit treats the inflammation intensity field and tumor biomarker concentration as continuously changing spatiotemporal fields. It simulates their diffusion and reaction processes in the network through a differentiable PDE module. The model automatically learns the diffusion coefficient and reaction term, and embeds the solver into the network for backpropagation, capturing the diffusion of inflammation and changes in biomarker concentration on a spatiotemporal scale.

4. The gastroenterology predictive analysis system based on big data according to claim 1, characterized in that: The intelligent diagnosis and treatment decision-making module specifically includes the following units: The decision support unit first constructs a patient state vector using the patient's latest multimodal joint encoding vector and medication record vector; then it designs a scoring function and uses a deep neural network to evaluate and score the expected efficacy of each regimen in different states, assesses the expected benefit of the medication regimen in different states, and finally automatically selects the optimal regimen as the recommended medication decision for this stage. The decision optimization unit models the entire treatment process as a dynamic decision system. It evaluates medication regimens through policy networks and value networks. First, it collects treatment feedback based on the patient's status at each time point and the current medication regimen, and calculates the cumulative discount reward by accumulating the treatment feedback. Then, it uses the policy gradient method and advantage function to measure the gain of each medication choice relative to the current policy. The final decision-making unit first sets a cumulative discount reward threshold and executes multiple interaction sequences in a clinical setting using the current policy network, collecting the state, selected medication, and treatment feedback at each time step. Then, it calculates the cumulative discount reward for each sequence based on a discount factor and combines this with the current value network to calculate the advantage function. Next, it constructs a joint loss function based on the collected data and optimizes it using proximal policies. The loss function consists of the sum of the expected minimization term of the policy ratio and the mean squared error regression term of the value network. Subsequently, it uses a stochastic gradient descent optimization algorithm to alternately update the policy parameters and value network parameters using gradients. Finally, this process is repeated until the cumulative discount reward reaches the threshold, resulting in the optimal policy network. The final medication regimen is then output through the optimal policy network.

5. The gastroenterology predictive analysis system based on big data according to claim 1, characterized in that: The data acquisition module collects demographic data, blood indicator data, vital sign data, medication record data, endoscopy report data, and the number of gastroenterological symptoms. The demographic data specifically includes age, sex, BMI, and family history of major digestive system diseases. The blood indicator data includes hemoglobin, white blood cell count, platelet count, C-reactive protein, tumor marker CA72-4, albumin, and alanine aminotransferase (ALT). The vital sign data includes heart rate, blood pressure, and body temperature. The medication record data includes the type of medication, dosage, and time of administration. The endoscopy report data refers to the text of the gastroscopy or colonoscopy examination report. The number of gastroenterological symptoms refers to the number of the following symptoms: dysphagia, heartburn, bloating, nausea and vomiting, abdominal pain, liver area pain, jaundice, and biliary colic.

Images