Intelligent gastrointestinal postoperative ileus diagnosis and intervention decision-making system and method

Abstract

The present invention relates to the field of medical system technology, and specifically to an intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision-making system and method thereof, comprising: a data preprocessing module, a multimodal data fusion module, a gastrointestinal function recovery model construction module, an adaptive progressive reinforcement learning optimization module AHRL, and a diagnosis and intervention recommendation module. The adaptive progressive reinforcement learning optimization module optimizes the intervention strategy layer by layer through a five-layer progressive algorithm, from the dynamic adjustment of feature weights, recursive state propagation, hierarchical cumulative optimization, symptom similarity matrix generation, to progressive weight reverse feedback and other steps, to achieve deep fusion of multimodal data on key symptoms, and update the intervention plan in real time according to changes in the patient's status. In the intervention decision recommendation module, the system generates personalized diagnostic suggestions based on the optimized strategy, and performs real-time risk assessment to provide patients with continuous and dynamic intervention support.

Description

Technical Field

[0001] The present invention relates to the technical field of medical systems, and in particular to an intelligent gastrointestinal postoperative ileus diagnosis and intervention decision-making system and method thereof. Background Art

[0002] In recent years, the diagnosis and intervention of intestinal obstruction after gastrointestinal surgery has become a key research direction in the medical field. Intestinal obstruction is a common complication after gastrointestinal surgery. Its clinical manifestations are complex and diverse, including symptoms such as abdominal distension, abdominal pain, nausea and vomiting, as well as dynamic changes in imaging manifestations and intestinal function. Traditional diagnostic methods rely on imaging examinations and physical sign analysis, but these methods have significant limitations in the early detection and intervention of intestinal obstruction, which can easily lead to missed diagnosis or misdiagnosis, thereby delaying treatment. Existing diagnostic and intervention systems are mostly based on a single model with fixed weights, lack the ability to adaptively adjust to the patient's real-time condition, and are difficult to meet the needs of accurate diagnosis and intervention under complex conditions.

[0003] Currently, fixed weight models and single intervention strategies are more common in the management of postoperative intestinal obstruction in the gastrointestinal tract. Although this model can provide partial support, it is difficult to accurately capture the dynamic changes of postoperative patients due to its fixed feature weights and lack of dynamic adjustment strategies. Especially when the patient's condition fluctuates, these systems are unable to update diagnostic data in real time, resulting in delayed intervention plans and difficulty in providing effective personalized support for different patients and different symptoms. At the same time, existing technologies have difficulty in integrating and utilizing multimodal data (such as imaging data, electronic medical record data, and clinical symptoms), resulting in key symptom information being ignored or not fully processed. In addition, these systems often perform well in short-term effects, but lack the ability to accumulate and optimize long-term intervention effects. The long-term effects are poor, and it is difficult to form stable and continuous intervention.

[0004] In the existing technology, due to the single model structure, it is difficult to coordinate the fusion processing of different types of data, the adaptability to dynamic adjustment is insufficient, and the correlation between symptom changes and intervention effects is ignored. Specifically, the traditional method cannot solve the following problems: first, the feature weights are solidified, and the feature weights cannot be dynamically adjusted according to the severity of the patient's symptoms, resulting in the neglect of key symptoms; second, there is a lack of effective mechanism for the fusion of multimodal data, making it difficult to obtain true and complete patient information; third, the update of the intervention strategy is delayed, and there is a lack of multi-level progressive dynamic optimization, making it difficult to balance short-term diagnostic needs and the accumulation of long-term intervention effects. In addition, the existing system is slow to respond in symptom identification and intervention plan optimization, and fails to achieve rapid response and real-time adjustment to the patient's status, reducing the accuracy and effectiveness of diagnosis and intervention.

[0005] In summary, the current state of technological development indicates that existing systems for diagnosing and intervening postoperative gastrointestinal obstruction still have deficiencies in terms of dynamic adaptability, multimodal data fusion, feature weight adjustment, and long-term intervention effectiveness. In particular, in practical applications, existing methods struggle to balance real-time diagnosis with the sustainability of intervention. With the advancement of artificial intelligence and data analysis technologies, new systems based on multimodal data fusion, progressive optimization, and adaptive adjustment are urgently needed to address the shortcomings of existing technologies and provide more comprehensive and accurate diagnostic and intervention support for postoperative patients. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this paper proposes an intelligent postoperative gastrointestinal ileus diagnosis and intervention decision-making system and method. This system addresses shortcomings in existing systems in multimodal data fusion, dynamic feature weight adjustment, real-time status tracking, and long-term intervention effect optimization. By introducing a multi-module collaborative architecture and a progressive optimization algorithm, this paper constructs an intelligent system that balances both short-term and long-term intervention effects.

[0007] The present invention proposes an intelligent gastrointestinal postoperative ileus diagnosis and intervention decision-making system, which includes: a data preprocessing module, a multimodal data fusion module, a gastrointestinal function recovery model construction module, an adaptive progressive reinforcement learning optimization module (AHRL), and a diagnosis and intervention recommendation module, wherein:

[0008] The data preprocessing module receives and standardizes the multimodal data of the patient to form a consistent data format;

[0009] The multimodal data fusion module extracts features based on the output of the data preprocessing module and aggregates multi-dimensional features to obtain a fused feature vector;

[0010] The gastrointestinal function recovery model construction module constructs a long short-term memory neural network model based on the fused feature vector to track the patient's postoperative intestinal function recovery status;

[0011] The adaptive progressive reinforcement learning optimization module includes a dynamic weight adaptive function generation algorithm, a recursive state propagation algorithm, a hierarchical cumulative effect decision optimization algorithm, a multi-dimensional cross-similarity matrix generation algorithm, and a progressive weight reverse optimization feedback algorithm, which is used to adaptively optimize long-term intervention strategies based on real-time patient data;

[0012] The diagnosis and intervention recommendation module generates personalized diagnosis and intervention decisions based on the optimization strategy output by the adaptive progressive reinforcement learning optimization module.

[0013] Preferably, the dynamic weight adaptive function generation algorithm DWAFG in the adaptive progressive reinforcement learning optimization module generates adaptive weights based on the feature importance of multimodal data and the real-time changes in the patient's status, wherein DWAFG uses the following formula to determine the weights:

[0014]

[0015] Among them, w i is the dynamic weight of the i-th feature, δ i is the criticality factor of the i-th feature, α is the adaptive attenuation factor, x i is the modulus of the i-th feature vector, n is the total number of features, and the adaptive weight of each feature is obtained through the DWAFG and input into the next progressive algorithm.

[0016] Preferably, the adaptive recursive state propagation algorithm ARSP in the adaptive progressive reinforcement learning optimization module recursively propagates symptom states based on the adaptive weights generated by the DWAFG to achieve dynamic updates over long time series. ARSP uses the following formula to implement state propagation:

[0017]

[0018] Among them, s t is the current state vector, s t-1 is the state of the previous time step, λ is the state update smoothing factor, w i is the weight of the i-th feature, x i,t is the value of the i-th feature at time step t. The ARSP further captures the dynamic changes in the state of the patient's symptoms and passes the state to the next progressive algorithm.

[0019] Preferably, the hierarchical cumulative effect decision optimization algorithm HCEDO in the adaptive progressive reinforcement learning optimization module is based on the state vector of the ARSP, hierarchically accumulating the effects of short-term and long-term intervention strategies. HCEDO uses the following formula to optimize the decision of cumulative effect:

[0020]

[0021] Among them, q t is the current cumulative effect score, q t-1 is the cumulative effect score of the previous moment, η is the cumulative weight of the short-term effect, β i is the importance weight of the i-th layer effect, For the characteristic state of the i-th layer in the state vector, HCEDO realizes the cumulative effect of the short-term effects of each layer in the long-term intervention strategy, providing hierarchical effect input to the next progressive algorithm.

[0022] Preferably, the multidimensional cross-similarity matrix generation algorithm MCSMG in the adaptive progressive reinforcement learning optimization module generates a multidimensional cross-similarity matrix based on the HCEDO cumulative effect score to evaluate the correlation between each symptom and the intervention effect. The MCSMG generates the similarity matrix using the following formula:

[0023]

[0024] Among them, C ij is the similarity score between the i-th and j-th features, s i and s j are the state vectors of the i-th and j-th features, q i and q j is the corresponding cumulative effect score, MCSMG is the last algorithm of the progressive algorithm, and provides input to the progressive weight reverse optimization feedback.

[0025] Preferably, the progressive weight reverse optimization feedback mechanism algorithm PWROF in the adaptive progressive reinforcement learning optimization module performs reverse feedback adjustment on the feature weights based on the multidimensional cross-similarity matrix generated by the MCSMG. PWROF performs weight optimization feedback through the following formula:

[0026]

[0027] Where Δw i is the adjustment amount of the i-th feature weight, γ is the feedback adjustment coefficient, C ij Similarity matrix value generated by MCSMG, The partial derivative of the cumulative effect with respect to the weight is used to obtain weight optimization feedback through PWROF to ensure continuous optimization of the identification of key symptoms during the intervention process.

[0028] Preferably, the data preprocessing module standardizes the patient's multimodal data and synchronizes it to the multimodal data fusion module to make the data structure consistent. The multimodal data fusion module obtains a fusion feature vector through feature extraction and aggregation to provide input for the gastrointestinal function recovery model construction module.

[0029] Preferably, the gastrointestinal function recovery model construction module uses a long short-term memory neural network LSTM to model the patient's intestinal function recovery status, and the intestinal function status value output by the model is input into the adaptive progressive reinforcement learning optimization module. The diagnosis and intervention recommendation module generates personalized intervention decisions based on the final output strategy of the adaptive progressive reinforcement learning optimization module, including real-time risk assessment and intervention recommendations.

[0030] Preferably, the natural language processing module extracts the patient's symptoms and medical history data from the electronic medical record, and synchronizes the extracted information to the data preprocessing module to enrich the multimodal data source and improve the accuracy of diagnosis and intervention decisions.

[0031] The method for intelligent diagnosis and intervention decision-making of intestinal obstruction after gastrointestinal surgery based on the system includes the following steps:

[0032] Obtaining multimodal data of the patient through data preprocessing, wherein the multimodal data includes imaging data, clinical symptom data, medical history data and / or electronic medical record information, forming a consistent format of the multimodal data based on standardization processing, and further transmitting the data to the multimodal data fusion step;

[0033] In the multimodal data fusion step, feature extraction and aggregation are performed based on the standardized data obtained in the data preprocessing step to generate a fused feature vector, wherein the fused feature vector is input to the subsequent model building step;

[0034] Based on the fused feature vector, a gastrointestinal function recovery model is constructed, and a long short-term memory neural network (LSTM) is used to model the patient's postoperative intestinal function recovery state, and an intestinal function state value is obtained, and the state value is used as input for an adaptive progressive reinforcement learning optimization step;

[0035] In the adaptive progressive reinforcement learning optimization step, the diagnosis and intervention strategies are optimized based on the adaptive progressive reinforcement learning algorithm AHRL.

[0036] Based on the optimization strategy obtained from the adaptive progressive reinforcement learning optimization step, generating personalized diagnosis and intervention suggestions, including real-time intestinal obstruction risk assessment and intervention recommendations;

[0037] Extract key information from electronic medical records, including symptom descriptions and medical history records, and supplement the data preprocessing step based on the key information to further enrich the multimodal data source to improve the accuracy of diagnosis and intervention decisions.

[0038] The beneficial technical effects brought about by the technical solution of the present invention are:

[0039] The system of the present invention realizes accurate assessment of the risk of intestinal obstruction in postoperative patients and personalized intervention decisions through the organic combination of a data preprocessing module, a multimodal data fusion module, a gastrointestinal function recovery model construction module, an adaptive progressive reinforcement learning optimization module and a diagnosis and intervention recommendation module. The data preprocessing module improves the integration and consistency of data through standardized processing, and the multimodal data fusion module integrates multidimensional information such as images, symptoms, and medical history into a unified feature vector, thereby providing comprehensive data support for subsequent modules. The gastrointestinal function recovery model construction module tracks the changes in the patient's postoperative intestinal function status through a long-short-term memory network, captures key symptoms and recovery trends, and provides high-precision state input for the adaptive progressive reinforcement learning optimization module.

[0040] The core adaptive progressive reinforcement learning optimization module uses a five-layer progressive algorithm to optimize intervention strategies layer by layer, from dynamic adjustment of feature weights, recursive state propagation, hierarchical cumulative optimization, symptom similarity matrix generation, to progressive weight reverse feedback. This allows for deep integration of multimodal data on key symptoms and allows for real-time updates of intervention plans based on changes in patient status. In the intervention decision recommendation module, the system generates personalized diagnostic recommendations based on the optimized strategy and conducts real-time risk assessments, providing patients with continuous and dynamic intervention support.

[0041] Through the close coordination and step-by-step optimization of the above modules, the present invention solves many problems in the prior art and realizes the real-time diagnosis, flexibility of intervention and sustainability of effect. First, the dynamic weight adaptive generation mechanism of the present invention enables the system to assign different feature weights according to the patient's current state, ensuring that the recognition rate of key symptoms is significantly improved. Secondly, the multimodal data fusion module organically combines information from different data sources to ensure that the system can obtain complete and accurate patient information, thereby improving the overall diagnostic accuracy at the data input level. Thirdly, through the five-layer progressive algorithm structure of the progressive reinforcement learning optimization module, the system achieves real-time adaptability in the dynamic processing of data, avoiding the problem of slow response of traditional fixed weight models when symptoms change. In addition, the present invention has carried out in-depth optimization on the sustainability of the intervention effect. Through the layered cumulative effect optimization and progressive feedback mechanism, the system shows stability in the long-term intervention effect, greatly reducing the problems of postoperative recurrence and intervention lag.

[0042] Overall, this invention, through a multi-layered, progressive dynamic optimization structure, addresses key issues in the existing technologies for the diagnosis and intervention of postoperative intestinal obstruction, from the macro to the micro level, thereby forming a more comprehensive intelligent diagnosis and intervention decision-making system. This not only significantly improves the system's real-time performance, accuracy, and adaptability, but also provides personalized and continuous intervention plans for postoperative patients, making it highly valuable and practical for clinical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is the system logic block diagram of the present invention.

[0044] Figure 2 This is a logical block diagram of the adaptive progressive reinforcement learning optimization module of the present invention.

[0045] Figure 3 This is a logic block diagram of the method of the present invention.

[0046] Figure 4 Flowchart of the adaptive progressive reinforcement learning optimization module of the present invention. DETAILED DESCRIPTION

[0047] To further illustrate the technical means and effects employed by the present invention to achieve its intended objectives, the following detailed description, along with the accompanying drawings and preferred embodiments, includes a detailed description of the specific implementations, structures, features, and effects thereof. In the following description, references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0048] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0049] See Figure 1-4 , the present invention provides an intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision-making system and method thereof, the system includes a data preprocessing module 1, a multimodal data fusion module 2, a gastrointestinal function recovery model construction module 3, an adaptive hierarchical reinforcement learning optimization module 4 (Adaptive Hierarchical Reinforcement Learning, AHRL) and a diagnosis and intervention recommendation module 5, the functions of each module are smoothly connected, and together realize the intelligent diagnosis and intervention decision-making of postoperative intestinal obstruction. The system of the present invention first obtains the patient's multimodal data, including imaging, medical history, clinical symptom data, etc. through the data preprocessing module 1, obtains a unified data format through standardization processing, and then transmits the standardized data to the multimodal data fusion module 2 for further feature extraction and aggregation processing. The multimodal data fusion module 2 extracts the key features of the data, performs dimensional aggregation, and generates a fusion feature vector, which is then used as the input of the gastrointestinal function recovery model construction module 3.

[0050] In the gastrointestinal function recovery model construction module 3, the patient's postoperative intestinal function recovery state is modeled through a long short-term memory neural network (LSTM), thereby obtaining the intestinal function state value. This state value is further transmitted to the adaptive progressive reinforcement learning optimization module 4, and the diagnosis and intervention strategy is optimized layer by layer through a progressive algorithm. The adaptive progressive reinforcement learning optimization module 4 of the present invention solves the current dynamic adjustment problem in the postoperative diagnostic strategy through five progressive algorithms, from adaptive adjustment of feature weights to recursive state propagation, hierarchical cumulative decision optimization, multi-dimensional cross-similarity generation, and then to the optimization of the progressive feedback mechanism.

[0051] Finally, the Diagnosis and Intervention Recommendation Module 5 generates personalized diagnostic recommendations and intervention plans based on the output strategy of the Adaptive Progressive Reinforcement Learning Optimization Module 4, and performs real-time assessments of the patient's postoperative risk. In practice, after data processing, the system can accurately predict the risk of postoperative intestinal obstruction and generate adaptive intervention recommendations based on the patient's specific clinical data and monitoring status.

[0052] The present invention further proposes a Dynamic Weighted Adaptive Function Generation (DWAFG) algorithm as the starting algorithm for the adaptive progressive reinforcement learning optimization module 4. This algorithm aims to achieve adaptive adjustment of feature weights, giving priority to key features in multimodal data, thereby optimizing the initial steps of diagnostic decision-making. In this process, the following formula is used:

[0053]

[0054] Among them, w i is the dynamic weight of the i-th feature, δ i is the criticality factor of the i-th feature, which is used to measure the importance of each feature to the diagnosis result, α is the adaptive attenuation factor, x i is the modulus of the i-th feature vector, n is the total number of features, and the adaptive weight of each feature is obtained through the DWAFG and input into the next progressive algorithm.

[0055] Preferably, δ can be determined according to the severity of the symptoms i For example, for symptoms such as abdominal distension or persistent vomiting, you can set δ i The empirical value of α is between 0.7 and 1, indicating high priority. The algorithm also introduces an adaptive attenuation factor, α, to control the adjustment range of each feature weight. Empirically, the value is generally between 0.01 and 0.1 to ensure smooth adjustment of feature weights. The algorithm adaptively adjusts the weights of each feature based on the patient's condition, providing optimized weighted input for subsequent recursive state propagation.

[0056] In the adaptive progressive reinforcement learning optimization module 4, the adaptive recursive state propagation algorithm (ARSP) recursively propagates the patient's symptom state value through the output weights of the above-mentioned DWAFG to achieve adaptive updating of multi-dimensional features over a long time series. Specifically, ARSP uses the following formula:

[0057]

[0058] Among them, s t is the current state vector, preferably, it can represent the comprehensive state of the patient's postoperative intestinal function; t -1 is the state of the previous time step, and λ is the state update smoothing factor, which is usually set between 0.8 and 0.95 to maintain smooth state propagation. The value of λ can be dynamically adjusted according to the patient's recovery progress. For example, it can be set to 0.9 within 7 days after surgery and adjusted to 0.85 after 7 days to more accurately adapt to the patient's recovery changes. This algorithm captures the dynamic changes of patients' symptoms by smoothing the state vector, providing high-precision input for subsequent cumulative effect decision optimization. i is the weight of the i-th feature, x i,t is the value of the i-th feature at time step t. The ARSP further captures the dynamic changes in the state of the patient's symptoms and passes the state to the next progressive algorithm.

[0059] Preferably, in the adaptive progressive reinforcement learning optimization module 4, the Hierarchical Cumulative Effect Decision Optimization (HCEDO) algorithm further implements the cumulative effect decision of the short-term and long-term intervention strategies based on the state vector obtained by the ARSP. The following formula is used:

[0060]

[0061] Among them, q t is the current cumulative effect score, which is used to evaluate the patient's current condition. t-1 is the cumulative effect score of the previous moment, η is the cumulative weight of the short-term effect, which is used to control the weight of the short-term effect in the cumulative score. It is usually set between 0.5 and 0.8 to ensure the timeliness of the short-term strategy; β i is the importance weight of the effect at layer i. It is preferably set to 1 / n in the initial stage to balance the effects of each layer, and then gradually adjusted according to symptom improvement. This algorithm can ensure that short-term effects play a positive role in long-term intervention and provide accurate cumulative effect input for long-term intervention strategy optimization. For the characteristic state of the i-th layer in the state vector, HCEDO realizes the cumulative effect of the short-term effects of each layer in the long-term intervention strategy, providing hierarchical effect input to the next progressive algorithm.

[0062] Based on the cumulative effect score, the present invention proposes a multi-dimensional cross-similarity matrix generation algorithm (MCSMG) to generate a multi-dimensional cross-similarity matrix to evaluate the correlation between different symptoms and intervention effects. The following formula is used:

[0063]

[0064] Among them, C ij is the similarity score between the i-th and j-th features. Preferably, a threshold value |q i -q j |Between 0.1 and 0.5 to balance the influence weights of different features; i and s j are the state vectors of the i-th and j-th features, representing the eigenvalues of different dimensions of multimodal data. This algorithm ensures that the interactive correlation between symptoms and intervention effects can be quantified, providing efficient data support for the subsequent progressive feedback mechanism. i and q j is the corresponding cumulative effect score, MCSMG is the last algorithm of the progressive algorithm, and provides input to the progressive weight reverse optimization feedback.

[0065] Based on the above, the Progressive Weight Reverse Optimization Feedback (PWROF) algorithm of the present invention performs reverse feedback adjustment on feature weights based on the multidimensional cross-similarity matrix generated by MCSMG. PWROF performs weight optimization feedback using the following formula:

[0066]

[0067] Where Δw i is the adjustment amount of the i-th feature weight, γ is the feedback adjustment coefficient, preferably, the value of the feedback adjustment coefficient γ can be set to 0.01-0.05 to prevent the model from being unstable due to excessive fluctuations in the feature weight; C ij Similarity matrix value generated by MCSMG, The partial derivative of the cumulative effect with respect to the weight is used to obtain weight optimization feedback through PW ROF. This mechanism can adjust the weight according to the similarity between each feature to ensure that the identification of key symptoms is continuously optimized during the intervention process.

[0068] Furthermore, the present invention standardizes the patient's multimodal data in the data preprocessing module 1 and synchronizes it to the multimodal data fusion module 2. Preferably, data standardization employs a normalization method to normalize the imaging data and clinical data to the range of 0-1, facilitating subsequent feature extraction. After data preprocessing is complete, the multimodal data fusion module 2 extracts and aggregates features from the standardized data using a feature extraction algorithm, thereby generating a fused feature vector that serves as input to the gastrointestinal function recovery model construction module 3.

[0069] In the gastrointestinal function recovery model construction module 3, a long short-term memory neural network (LSTM) is used to model the patient's intestinal function recovery status, and the output intestinal function status value is used as the input of the adaptive progressive reinforcement learning optimization module 4. Preferably, the number of hidden layer units used in the LSTM network can be set to 64-128 units based on the amount of patient recovery data to ensure data prediction accuracy.

[0070] In the diagnosis and intervention recommendation module 5, a personalized intervention decision is generated based on the final output strategy of the adaptive progressive reinforcement learning optimization module 4. Preferably, the intervention decision includes real-time risk assessment and personalized intervention recommendations. The risk assessment is based on the dynamic range of the output value set to 0-1, where a value exceeding 0.8 is considered high risk.

[0071] Preferably, in the electronic medical record information extraction module, key information is extracted from the patient's electronic medical record, including symptom descriptions, medical history records, etc., to further enrich the multimodal data source and improve the accuracy of diagnosis and intervention decisions.

[0072] This paper also proposes an intelligent method for diagnosing and deciding interventions for postoperative gastrointestinal ileus. This method aims to achieve intelligent diagnosis and personalized interventions for postoperative ileus through a multi-module collaborative and progressively optimized algorithm. This method comprises a series of sequential, interconnected steps to fully capture changes in postoperative patient symptoms and accurately inform risk assessments and intervention decisions.

[0073] First, the intelligent postoperative gastrointestinal ileus diagnostic method of the present invention receives and processes the patient's multimodal data, including imaging, clinical symptoms, medical history information, and electronic medical record content. Through the data preprocessing module 1, various input data are standardized to obtain a consistent format. This process ensures that the multimodal data can be further integrated and used for subsequent feature extraction. Preferably, the standardization process normalizes all data to the range of 0-1 so that each module can process data from different sources while maintaining a consistent numerical range.

[0074] Next, in the multimodal data fusion step, the method further performs feature extraction and aggregation based on the above-mentioned standardized data. The multimodal data fusion module 2 extracts features from image, symptom, medical history and other data according to the importance of different features to form a unified fusion feature vector. In this process, the image features are preferentially used to extract key features such as dynamic indicators of intestinal function using convolutional neural networks; while clinical symptoms and medical history features are generated using statistical aggregation methods to generate numerical features, such as quantifying indicators such as abdominal distension and pain into values between 0 and 1. The final fusion feature vector is used as input for the gastrointestinal function recovery model construction module 3.

[0075] Based on the fused feature vectors, the gastrointestinal function recovery model construction module 3 uses a long short-term memory neural network (LSTM) to model the patient's postoperative intestinal function recovery status. The LSTM structure can effectively process time series data, thereby generating accurate intestinal function status values as input for subsequent diagnostic optimization. Preferably, the number of hidden units of the LSTM is set to 64-128 to ensure sufficient feature dimensions to capture postoperative recovery dynamics. This state value represents the patient's intestinal recovery trend, which is further passed to the adaptive progressive reinforcement learning optimization module 4 for multi-level optimization analysis.

[0076] In the adaptive progressive reinforcement learning optimization module 4, the present invention introduces a five-layer progressive algorithm, including dynamic weight adaptive function generation, adaptive recursive state propagation, hierarchical cumulative effect decision optimization, multi-dimensional cross-similarity matrix generation, and progressive weight reverse optimization feedback mechanism, to optimize the decision plan layer by layer and output the optimal intervention strategy. Preferably, the present invention gives higher weights to key symptoms through the dynamic weight adaptive function generation algorithm (DWAFG). For example, when a patient has significant abdominal distension and nausea symptoms, they can be given a greater influence in decision optimization through weight setting. The AHRL includes the following progressive steps:

[0077] Based on the importance and real-time status of the patient's multimodal data features, a dynamic weight of each feature is generated to obtain a dynamic weight adaptive function;

[0078] Furthermore, the patient's symptom status is recursively propagated based on the dynamic weight adaptive function to achieve dynamic update of the status over a long time series to obtain a state vector of adaptive recursive state propagation;

[0079] Based on the state vector of the adaptive recursive state propagation, hierarchical cumulative effect decision optimization is achieved, and the effects of short-term and long-term intervention strategies are hierarchically accumulated to obtain a cumulative effect score;

[0080] Based on the cumulative effect scores, a multidimensional cross-similarity matrix is generated to evaluate the correlation between each symptom and the intervention effect, so as to obtain a multidimensional cross-similarity matrix;

[0081] Furthermore, based on the multidimensional cross-similarity matrix, feedback optimization is performed on the feature weights in the dynamic weight adaptive function through progressive weight reverse optimization feedback to achieve continuous optimization of key symptom recognition;

[0082] This mechanism ensures that feature weights are gradually optimized in subsequent adjustments by adjusting the feedback adjustment coefficient γ between 0.01 and 0.05. This mechanism further ensures that intervention recommendations are sensitive to key symptoms, thereby optimizing intervention strategies.

[0083] Based on the final output strategy of the adaptive progressive reinforcement learning optimization module 4, the diagnosis and intervention recommendation module 5 can generate personalized intervention decisions and conduct real-time assessments of the patient's postoperative risk. Preferably, when the final output risk score exceeds 0.8, the diagnosis and intervention module will recommend a more aggressive intervention strategy.

[0084] Finally, in the electronic medical record information extraction module, the present invention extracts relevant symptom descriptions, medical history records and other information from the patient's electronic medical record, and further synchronizes them to the data preprocessing module 1 to increase the richness of multimodal data.

[0085] In order to verify the superiority of the intelligent gastrointestinal postoperative intestinal obstruction diagnosis and intervention decision-making system and its method of the present invention, the present invention uses a real gastrointestinal postoperative patient data set to test the effect of the system in clinical application, and makes a detailed comparison with the traditional comparison ratio based on the fixed weight model. The data set comes from multiple clinical medical centers and covers multimodal data such as patient imaging data, clinical symptoms, and electronic medical records. The total data sample is 500 cases, which includes patient symptom information and follow-up records at different time periods after surgery. In order to ensure the authenticity of the test and the repeatability of the results, 100 patient data with complete follow-up records were specially selected for this test.

[0086] In order to comprehensively evaluate the diagnostic accuracy, intervention accuracy, and adaptability of the system, the present invention selected the following four core indicators as test criteria:

[0087] 1. Diagnosis Accuracy: This refers to the accuracy of the system's prediction of the risk of intestinal obstruction after the patient's data is input. The diagnostic accuracy of each patient is tested by comparing the predicted value output by the system with the actual clinical results. The detection method is to calculate the consistency rate between the predicted result and the actual situation.

[0088] 2. Intervention Adaptability: This evaluates the system's ability to flexibly adjust intervention strategies for patients at different stages of recovery, focusing on how quickly the system can adjust intervention plans when input symptoms change. This is measured by measuring the average response time (i.e., the time it takes for the system to output an adaptive intervention plan after a change in the patient's condition). Shorter response times indicate greater system adaptability.

[0089] 3. Sensitivity to Key Symptoms: This function is primarily used to assess the system's sensitivity in detecting and distinguishing different patient symptoms. By comparing the system's recognition rate of key symptoms in multimodal data, the system's recognition rate for symptoms such as abdominal distension, pain, and vomiting is tested.

[0090] 4. Sustainability of Intervention Effectiveness: This refers to the stability of the system's recovery effect on patients during short-term and long-term interventions. The detection method is to record the recovery progress after the intervention and compare it with the traditional model.

[0091] The above indicators are recorded and repeatedly verified through data, and all tests are carried out in the same test environment to ensure the fairness of experimental data and the reliability of results.

[0092] Example (system of the present invention)

[0093] In the system test of the present invention, the system was fully tested and evaluated on 100 patient data through the combined application of progressive algorithms such as dynamic weight adaptive function generation, recursive state propagation, and hierarchical cumulative effect decision optimization. During the test, the system's recognition rate for key symptoms in multimodal data reached 96.5%, and the diagnostic accuracy rate reached 92.8%. In particular, when the patient's symptoms changed, the system was able to generate adaptive intervention suggestions within an average of 1.2 seconds, reflecting strong adaptability. At the same time, the system's intervention effect on patients showed good sustainability. In the follow-up data 3 months after surgery, the stability of the intervention effect was as high as 90.3%.

[0094] Comparative proportion (traditional fixed weight model)

[0095] The comparative test used a traditional fixed-weight model, whose feature weights were initially fixed to set values and lacked support for a progressive algorithm. Using data from 100 identical patients, the model achieved a diagnostic accuracy of only 75.4%, with a key symptom recognition rate of approximately 73.2%. Furthermore, due to the lack of a real-time update mechanism, the model took an average of 3.5 seconds to adjust the intervention plan when a patient's symptoms changed, and the persistence of the intervention plan during postoperative follow-up was only 63.7%.

[0096] The test results comparison table is as follows:

[0097] index Example Comparative Example Diagnostic accuracy 92.8％ 75.4％ Adaptability of intervention decisions (reaction time) 1.2 seconds 3.5 seconds Key symptom recognition rate 96.5％ 73.2％ Durability of intervention effects 90.3％ 63.7％

[0098] The system of the present invention is superior to the traditional fixed-weight model in all indicators, especially in terms of diagnostic accuracy and key symptom recognition rate. The diagnostic accuracy rate is 92.8%, which shows that the system can efficiently use multimodal data to accurately diagnose patients, significantly improving the reliability of diagnosis. At the same time, the key symptom recognition rate reaches 96.5%, indicating that the system can accurately capture symptoms such as abdominal distension and pain that are crucial to patient recovery, greatly reducing misdiagnosis or missed diagnosis. In response to changes in patient status, the adaptability of intervention decisions of this system is significantly improved, and the response time is only 1.2 seconds, which greatly enhances the real-time and flexibility of intervention plans. This is especially applicable to situations where symptoms of patients fluctuate repeatedly after surgery. However, the traditional fixed-weight model has no dynamic weight update capability, resulting in a long response time. It takes 3.5 seconds to complete the intervention adjustment, which reduces efficiency in practical applications.

[0099] In addition, the test results of the persistence of the intervention effect showed that the persistence of the effect of this system in the 3-month follow-up was as high as 90.3%, while the traditional fixed weight model of the control group was only 63.7%. This shows that the intervention strategy generated by the system of the present invention is more stable in long-term rehabilitation and helps to accelerate patient recovery.

[0100] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. Intelligent gastrointestinal postoperative ileus diagnosis and intervention decision-making system, characterized by: include: Data preprocessing module, multimodal data fusion module, gastrointestinal function recovery model construction module, adaptive progressive reinforcement learning optimization module AHRL and diagnosis and intervention recommendation module, including: The data preprocessing module receives and standardizes the multimodal data of the patient to form a consistent data format; The multimodal data fusion module extracts features based on the output of the data preprocessing module and aggregates multi-dimensional features to obtain a fused feature vector; The gastrointestinal function recovery model construction module constructs a long short-term memory neural network model based on the fused feature vector to track the patient's postoperative intestinal function recovery status; The adaptive progressive reinforcement learning optimization module includes a dynamic weight adaptive function generation algorithm, a recursive state propagation algorithm, a hierarchical cumulative effect decision optimization algorithm, a multi-dimensional cross-similarity matrix generation algorithm, and a progressive weight reverse optimization feedback algorithm, which is used to adaptively optimize long-term intervention strategies based on real-time patient data; The diagnosis and intervention recommendation module generates personalized diagnosis and intervention decisions based on the optimization strategy output by the adaptive progressive reinforcement learning optimization module; The dynamic weight adaptive function generation algorithm DWAFG in the adaptive progressive reinforcement learning optimization module generates adaptive weights based on the feature importance of multimodal data and the real-time changes in the patient's status. DWAFG uses the following formula to determine the weights: ， in, For the The dynamic weight of each feature, For the The criticality factor of a feature, is the adaptive attenuation factor, For the The modulus of the eigenvectors, is the total number of features, and the adaptive weight of each feature is obtained through the DWAFG and input into the next progressive algorithm; The recursive state propagation algorithm ARSP in the adaptive progressive reinforcement learning optimization module recursively propagates symptom states based on the adaptive weights generated by the DWAFG to achieve dynamic updates over long time series. ARSP uses the following formula to implement state propagation: ， in, is the current state vector, is the state of the previous time step, is the state update smoothing factor, For the The weight of the feature, For the Features at time step The ARSP further captures the dynamic changes of the patient's symptom status and transmits the status to the next progressive algorithm; The hierarchical cumulative effect decision optimization algorithm HCEDO in the adaptive progressive reinforcement learning optimization module is based on the state vector of the ARSP and hierarchically accumulates the effects of short-term and long-term intervention strategies. HCEDO uses the following formula to optimize the decision of cumulative effect: ， in, is the current cumulative effect score, is the cumulative effect score of the previous moment, is the cumulative weight of short-term effects, For the The importance weight of the layer effect, is the first The characteristic state of each layer,HCEDO realizes the cumulative effect of each layer’s short-term effects in the long-term intervention strategy, and provides hierarchical effect input to the next progressive algorithm; The multidimensional cross-similarity matrix generation algorithm MCSMG in the adaptive progressive reinforcement learning optimization module generates a multidimensional cross-similarity matrix based on the HCEDO cumulative effect score to evaluate the correlation between each symptom and the intervention effect. The MCSMG generates the similarity matrix using the following formula: ， in, For the The first and The similarity score of the features, and Respectively Hedi The state vector of features, and is the corresponding cumulative effect score, MCSMG is the final algorithm of the progressive algorithm, and provides input to the progressive weight reverse optimization feedback; The progressive weight reverse optimization feedback mechanism algorithm PWROF in the adaptive progressive reinforcement learning optimization module performs reverse feedback adjustment on the feature weights based on the multidimensional cross-similarity matrix generated by the MCSMG. PWROF performs weight optimization feedback using the following formula: ， in, For the The adjustment amount of feature weights, is the feedback adjustment coefficient, Similarity matrix value generated by MCSMG, The partial derivative of the cumulative effect with respect to the weight is used to obtain weight optimization feedback through PWROF to ensure continuous optimization of the identification of key symptoms during the intervention process.

2. The intelligent post-operative gastrointestinal ileus diagnosis and intervention decision-making system according to claim 1 is characterized in that: The data preprocessing module standardizes the patient's multimodal data and synchronizes it to the multimodal data fusion module to make the data structure consistent. The multimodal data fusion module obtains a fusion feature vector through feature extraction and aggregation to provide input for the gastrointestinal function recovery model construction module.

3. The intelligent post-operative gastrointestinal ileus diagnosis and intervention decision-making system according to claim 2 is characterized in that: The gastrointestinal function recovery model construction module uses a long short-term memory neural network (LSTM) to model the patient's intestinal function recovery status. The intestinal function status value output by the model is input into the adaptive progressive reinforcement learning optimization module. The diagnosis and intervention recommendation module generates personalized intervention decisions based on the final output strategy of the adaptive progressive reinforcement learning optimization module, including real-time risk assessment and intervention recommendations.

4. The intelligent post-operative gastrointestinal ileus diagnosis and intervention decision-making system according to claim 3 is characterized in that: It also includes a natural language processing module, which extracts the patient's symptoms and medical history data from the electronic medical record and synchronizes the extracted information to the data preprocessing module to enrich the multimodal data source and improve the accuracy of diagnosis and intervention decisions.

5. A method for intelligent diagnosis and intervention decision-making of postoperative gastrointestinal ileus based on the system according to any one of claims 1 to 4, characterized in that: The following steps are involved: Obtaining multimodal data of the patient through data preprocessing, wherein the multimodal data includes imaging data, clinical symptom data, medical history data and / or electronic medical record information, forming a consistent format of the multimodal data based on standardization processing, and further transmitting the data to the multimodal data fusion step; In the multimodal data fusion step, feature extraction and aggregation are performed based on the standardized data obtained in the data preprocessing step to generate a fused feature vector, wherein the fused feature vector is input to the subsequent model building step; Based on the fused feature vector, a gastrointestinal function recovery model is constructed, and a long short-term memory neural network (LSTM) is used to model the patient's postoperative intestinal function recovery state, and an intestinal function state value is obtained, and the state value is used as input for an adaptive progressive reinforcement learning optimization step; In the adaptive progressive reinforcement learning optimization step, the diagnosis and intervention strategies are optimized based on the adaptive progressive reinforcement learning algorithm AHRL. Based on the optimization strategy obtained from the adaptive progressive reinforcement learning optimization step, generating personalized diagnosis and intervention suggestions, including real-time intestinal obstruction risk assessment and intervention recommendations; Extract key information from electronic medical records, including symptom descriptions and medical history records, and supplement the data preprocessing step based on the key information to further enrich the multimodal data source to improve the accuracy of diagnosis and intervention decisions.

Images