A ureterolithotripsy postoperative spasm prevention and control method based on inflammation factor monitoring

Abstract

This invention relates to the field of urological postoperative complication prevention and control technology, and more specifically, to a method for preventing and controlling spasm after ureterolithotomy based on inflammatory factor monitoring. It addresses the problems of existing technologies failing to accurately construct multimodal dynamic features based on physiological mechanisms, being unable to utilize bidirectional LSTM for risk time-series prediction and classification, and hindering the evolution of early warning models from static judgment to dynamic deduction, causal interpretability, and intelligent intervention, thus reducing accuracy, foresight, and clinical usability. This invention constructs multimodal dynamic features based on physiological mechanisms, combines graph neural networks and Granger causality analysis to model cross-system interaction, integrates reinforcement learning to achieve closed-loop control, and utilizes bidirectional LSTM for risk time-series prediction and classification. This promotes the evolution of early warning models from static judgment to dynamic deduction, causal interpretability, and intelligent intervention, significantly improving accuracy, foresight, and clinical usability.

Description

Technical Field

[0001] This invention relates to the field of prevention and control of postoperative complications in urology, and more specifically, to a method for preventing and controlling spasm after ureterolithotomy based on monitoring of inflammatory factors. Background Technology

[0002] The prevention and treatment of post-ureterolithotomy spasticity still faces challenges. Clinical reliance on lagging inflammatory markers makes it difficult to capture inflammatory dynamics in a timely manner. Fixed treatment patterns and lack of individualization lead to unstable efficacy. Existing predictive models often ignore the core mechanism of mucosal edema and increased nerve sensitivity mediated by IL-6 and TNF-α, resulting in limited early warning capabilities and difficulty in effectively blocking the inflammatory cascade response and subsequent spasticity triggered by lithotripsy stimulation.

[0003] Reference patent application CN118658131A discloses a monitoring system and method for infection prevention and control in hemodialysis. This system employs computer vision-based image processing and analysis technology to perform keyframe sampling, optical flow feature extraction, and enhancement on monitoring videos of nurses performing dialysis procedures. Based on the semantic association features of the enhanced optical flow features of each nurse's dialysis procedure across the entire time domain, it automatically obtains monitoring results indicating whether any abnormalities exist in the dialysis operation. It can accurately capture subtle changes in movement during the operation, supports real-time monitoring and rapid feedback. Once a violation or high-risk behavior is detected, the system can immediately issue an alert and trigger an intervention mechanism, significantly reducing response delay and effectively improving the timeliness and automation level of infection prevention and control. However, the aforementioned reference patents, based on computer vision, perform keyframe sampling and optical flow feature enhancement on operation videos. They achieve automatic identification and real-time early warning of abnormal behavior through full-time semantic association analysis, improving the accuracy, timeliness, and automation level of infection control. However, they cannot accurately construct multimodal dynamic features based on physiological mechanisms, cannot use bidirectional LSTM to complete risk time-series prediction and level classification, and are unable to promote the evolution of early warning models from static judgment to dynamic inference, causal interpretability, and intelligent intervention, thus reducing accuracy, foresight, and clinical usability. At the same time, they cannot integrate patient gene and phenotypic data to construct joint decoding models, cannot use cross-patient contrastive learning to improve the generalization ability of feature representations, and cannot combine reinforcement learning to generate individualized drug intervention pathways with the goal of reducing spasticity risk and improving muscle tone, making it difficult to achieve closed-loop precise control from perception to decision-making to execution.

[0004] To address these issues, we propose a method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring. Summary of the Invention

[0005] The purpose of this invention is to provide a method for preventing and controlling postoperative spasticity after ureterolithotomy based on inflammatory factor monitoring. This method addresses the problems of existing technologies, which cannot accurately construct multimodal dynamic features based on physiological mechanisms, cannot utilize bidirectional LSTM to complete risk time-series prediction and grade classification, and cannot promote the evolution of early warning models from static judgment to dynamic inference, causal interpretability, and intelligent intervention, thus reducing accuracy, prospectivity, and clinical usability. Furthermore, these technologies cannot integrate patient genetic and phenotypic data to construct a joint decoding model, cannot utilize cross-patient comparative learning to improve the generalization ability of feature representations, and cannot combine reinforcement learning to generate individualized drug intervention pathways with the goal of reducing spasticity risk and improving muscle tone, making it difficult to achieve closed-loop precise control from perception to decision-making to execution.

[0006] The objective of this invention is achieved through the following technical solution: A method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring includes the following steps: Step 1: Collect postoperative core physiological monitoring parameters from blood, body surface and urination process, and perform time alignment, anomaly identification, signal denoising, data completion and data normalization on the collected parameters in sequence; Step 2: Based on the pre-processed postoperative core physiological monitoring parameters, extract the dynamic characteristics of inflammatory factors, heart rate variability, ureteral pressure dynamic patterns and local electromyographic activity characteristics, model their interaction relationships through graph neural networks, and combine causal time maps, latent variable recursive models and reinforcement learning agents to output the time-series prediction results and risk level classification of spasticity risk. Step 3: Based on the temporal prediction results and risk level classification of spasticity risk, a visualization panel is used to display the dynamics of inflammatory factors, trends of physiological signs, treatment recommendations and risk alerts. Combined with the model interpretation mechanism, it provides decision-making basis prompts and abnormal event reminders. Step 4: When the spasticity risk prediction results reach the preset threshold, a gene-phenotype decoding model is constructed based on the patient's gene chip data and phenotypic characteristics. Combined with the adaptive intervention mechanism of reinforcement learning and the cross-patient comparative learning algorithm, an individualized drug intervention path is generated and the dosage and intensity parameters are dynamically updated.

[0007] In a preferred embodiment of the present invention, step two, which involves extracting dynamic features of inflammatory factors, heart rate variability, ureteral pressure dynamic patterns, and local electromyographic activity features based on pre-treated postoperative core physiological monitoring parameters, includes: We obtained time-series concentration data of interleukin-6 and C-reactive protein at a fixed sampling period, constructed a first-order ordinary differential equation model, and calculated the concentration rise time, peak concentration, half-life and area under the curve based on the model fitting curve. The concentration rise time, peak concentration, half-life and area under the curve were used as dynamic features of inflammatory factors. Continuous electrocardiogram (ECG) signals are acquired, the peak position of the R wave is identified by threshold detection, and adjacent R wave interval sequences are generated. The interval sequences are linearly interpolated with the arithmetic mean as the center to form an equal time interval sequence. The equal time interval sequence is subjected to a fast Fourier transform and the power spectral density is calculated. Numerical integration is performed in the preset first frequency band and second frequency band respectively to obtain the power values ​​of the first frequency band and the second frequency band. The ratio of the power value of the first frequency band to the power value of the second frequency band is calculated and used as the heart rate variability. Ureteral pressure signals and local electromyographic activity signals were acquired simultaneously. The local electromyographic activity signals were rectified by full-wave rectification to generate smoothed local electromyographic activity signals. A linear regression model was constructed using the smoothed local electromyographic activity signals as input variables and ureteral pressure signals as output variables. The regression coefficients and bias terms were solved using the regularized least squares method. The regression coefficients were used as the dynamic mode of ureteral pressure, and the bias terms were used as the characteristics of local electromyographic activity.

[0008] In a preferred embodiment of the present invention, step two, which involves modeling multimodal signal interaction relationships using graph neural networks and combining causal time maps, latent variable recursive models, and reinforcement learning agents for joint modeling, includes: The dynamic features of inflammatory factors, heart rate variability index, ureteral pressure response gain and baseline tension features are obtained. The four features are used as graph nodes to construct a graph structure. The initial feature vector of each node is the normalized value of the corresponding feature. Two layers of graph convolution operation are performed to generate a multimodal fusion feature vector. The multimodal fusion feature vector is input into a two-layer fully connected neural network. The network output is added to Gaussian noise with a zero mean vector and a diagonal covariance matrix to obtain a latent state vector. The latent state vector is concatenated with the multimodal fusion feature vector to form a state input vector. The state input vector is fed into a reinforcement learning agent. The agent selects an action based on a preset set of discrete risk control instructions, calculates the reward value based on the weighted sum of the risk level change and the intervention cost, updates the policy network parameters using a proximal policy optimization algorithm, and outputs the selected risk control instruction.

[0009] In a preferred embodiment of the present invention, the process of predicting the temporal sequence of spasticity risk and classifying the risk level based on the extracted multimodal features in step two includes: Receive dynamic characteristics of inflammatory factors, heart rate variability, dynamic patterns of ureteral pressure, and local electromyographic activity characteristics, and perform the following operations: The multimodal features are arranged in chronological order to form an input sequence of length T, which is then fed into a bidirectional long short-term memory network. The network decoder generates a sequence of spasm risk probabilities for the next N time steps. Take the maximum value in the spasm risk probability sequence as the spasm risk rating r(t) for the current period, substitute r(t) into the risk level determination rule, calculate the risk level L(t), and output the spasm risk probability sequence for the next N time steps and the risk level for the current period. .

[0010] In a preferred embodiment of the present invention, the process of generating a visual output based on the spasticity risk time-series prediction results and risk level classification in step three includes: Receive dynamic characteristics of inflammatory factors, heart rate variability, dynamic patterns of ureteral pressure, regional electromyographic activity characteristics, temporal prediction results of spasticity risk, risk level, and treatment recommendations, and perform the following operations: The time-series data of interleukin-6 and C-reactive protein concentrations were sorted by sampling time to generate a dynamic data sequence of inflammatory factors. The time-series data of heart rate variability, arterial systolic blood pressure, arterial diastolic blood pressure, axillary skin temperature, ureteral pressure, and urine flow rate were aligned by time to generate a physiological sign trend dataset. The spasticity risk scores for the next N time steps in the spasticity risk time-series prediction results were merged with the historical risk scores to generate a complete risk trend sequence. The current risk level is mapped to a visual identifier. When the risk level is Level 3, an alarm event is generated. The treatment recommendation text is parsed into structured recommendation items. The dynamic data sequence of inflammatory factors, the physiological sign trend dataset, the risk trend sequence, the risk level visual identifier, the alarm event, and the structured recommendation items are merged into a unified visualization data package. The visualization data package is pushed to the front-end client via the WebSocket protocol, triggering the visualization panel to update.

[0011] In a preferred embodiment of the present invention, step three, which combines the model interpretation mechanism to provide decision-making basis prompts and abnormal event alerts, includes: Obtain the feature importance weights and corresponding influence directions generated during the spasticity risk prediction process; obtain the raw measurement values ​​of heart rate, systolic blood pressure, diastolic blood pressure, and axillary skin temperature; obtain the lower and upper limits of the normal range for each physiological parameter; and perform the following operations: The importance weights of features are matched according to the type of physiological parameter, and then sorted in descending order according to the weight value. The top three physiological parameters are selected as the main influencing factors. For each main influencing factor, a decision item is generated based on its direction of influence. All decision items are compiled into a decision basis dataset. The original measurements of heart rate, systolic blood pressure, diastolic blood pressure, and axillary skin temperature are checked one by one and compared with the corresponding lower and upper limits of the normal range. When the measured value is lower than the lower limit of the normal range or higher than the upper limit of the normal range, the corresponding abnormal event record is generated. All abnormal event records are compiled into an abnormal event dataset. The decision basis dataset and the abnormal event dataset are written into a dedicated field of the visualization data package.

[0012] In a preferred embodiment of the present invention, the process of constructing the gene-phenotype decoding model and designing an adaptive intervention mechanism based on reinforcement learning in step four includes: Acquire patients' single nucleotide polymorphism (SNP) data, clinical symptom scores, neurological signs, laboratory test indicators, and physiological monitoring parameters. Perform quality control on the SNP data, standardize the remaining data, extract principal component vectors, generate gene principal component representations, perform Z-score standardization on continuous variables in phenotypic data, generate standardized phenotypic vectors, and concatenate the gene principal component representations with the standardized phenotypic vectors to form a joint feature vector. The joint feature vector is input into a multilayer perceptron model consisting of an input layer, multiple fully connected hidden layers and an output layer, and the output is an estimated value of the risk of spasticity. An L1 regularization term is introduced during the model training process to obtain a gene-phenotype decoding model. The principal component representation of genes, the standardized phenotypic vector, and the estimated probability of spasticity risk output from the gene-phenotype decoding model are concatenated to form the state vector of the Markov decision process. The action space is defined as the Cartesian product of the drug type index and the dosage level index. Each action uniquely corresponds to a drug and its dosage level. The reward signal is calculated based on the changes in spasticity risk and key phenotypic indicators before and after intervention. A deep Q-network is used to achieve policy learning. The deep Q-network model after training convergence is determined as the adaptive intervention mechanism to generate individualized drug intervention recommendations.

[0013] In a preferred embodiment of the present invention, step four, which involves generating a personalized drug intervention pathway based on a cross-patient contrastive learning framework and dynamically adjusting the dosage and intensity parameters during execution, includes: Obtain the principal component representation of genes and the standardized phenotypic vector of a multicenter patient cohort. For any two patients i and j, concatenate their principal component representation of genes and the standardized phenotypic vector and input them into a shared encoder to generate a fixed-dimensional feature embedding. The contrastive loss function is defined as the square of the Euclidean distance between the feature embeddings of the two patients. In a multi-center patient cohort, the Adam optimizer was used to update parameters with the contrastive loss function as the objective, and the shared encoder was pre-trained. The gene principal component representation and standardized phenotypic vector of the current patient were obtained. The gene principal component representation and standardized phenotypic vector of the current patient were concatenated and input into the pre-trained shared encoder to generate feature embeddings. The feature embedding is input into a deep Q-network model to obtain the Q value corresponding to each action in the action space. The action with the largest Q value is selected as the recommended intervention. The recommended action is mapped to a specific drug name and dosage value to form an individualized drug intervention plan. At a fixed time point every day, the patient's latest gene principal component representation and standardized phenotypic vector are re-acquired. The feature embedding generation, strategy reasoning and action selection process are repeated to generate a new drug intervention plan.

[0014] Compared with the prior art, the advantages of this invention are: (1) In this invention, multimodal dynamic features are constructed based on physiological mechanisms, graph neural networks and Granger causal analysis are combined to model cross-system interaction, latent variables are introduced to represent latent physiological states, reinforcement learning is integrated to achieve closed-loop control, and bidirectional LSTM is used to complete risk time series prediction and level classification, promoting the evolution of early warning mode from static judgment to dynamic deduction, causal interpretability and intelligent intervention, significantly improving accuracy, foresight and clinical usability; (2) In this invention, a joint decoding model is constructed by integrating patient gene and phenotypic data, and the generalization ability of feature representation is improved by cross-patient contrastive learning. Combined with reinforcement learning, an individualized drug intervention path is generated with the goal of reducing the risk of spasticity and improving muscle tone. The dosage and drug selection are dynamically adjusted daily to achieve closed-loop precise regulation from perception to decision-making to execution. Attached Figure Description

[0015] Figure 1 This is a flowchart of the steps in the method for preventing and controlling spasm after ureterolithotomy in this invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0017] Example 1: As Figure 1 As shown, the present invention proposes a method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring, comprising the following steps: Step 1: Collect postoperative core physiological monitoring parameters from blood, body surface, and urination processes, including interleukin-6, C-reactive protein concentration, continuous electrocardiogram signal, axillary skin temperature, arterial systolic blood pressure, arterial diastolic blood pressure, ureteral pressure, urine flow rate, and local electromyographic activity signals. The collected parameters are then processed sequentially by time alignment, anomaly identification, signal denoising, data completion, and data normalization. By collecting multimodal postoperative physiological parameters covering inflammation, circulation, nervous system and urinary system, and performing time alignment, anomaly identification, signal denoising, data completion and normalization processing, a high-quality, spatiotemporally consistent and structurally unified input dataset was constructed, which significantly improved the accuracy, robustness and clinical interpretability of subsequent modeling, and provided a reliable data foundation for individualized postoperative intervention.

[0018] Step 2: Based on the pre-processed postoperative core physiological monitoring parameters, extract the dynamic characteristics of inflammatory factors, heart rate variability, ureteral pressure dynamic patterns and local electromyographic activity characteristics, model their interaction relationships through graph neural networks, and combine causal time maps, latent variable recursive models and reinforcement learning agents to output the time-series prediction results and risk level classification of spasticity risk. Step two involves extracting dynamic characteristics of inflammatory factors, heart rate variability, ureteral pressure dynamic patterns, and local electromyographic activity characteristics based on postoperative core physiological monitoring parameters after pretreatment. This process includes: We obtained time-series concentration data of interleukin-6 and C-reactive protein at a fixed sampling period, constructed a first-order ordinary differential equation model, estimated the model parameters using the gradient descent method, and calculated the concentration rise time, peak concentration, half-life, and area under the curve based on the model fitting curve. The concentration rise time, peak concentration, half-life, and area under the curve were used as dynamic features of inflammatory factors. Continuous electrocardiogram (ECG) signals are acquired, the peak position of the R wave is identified by threshold detection, and adjacent R wave interval sequences are generated. The interval sequences are linearly interpolated with the arithmetic mean as the center to form an equal time interval sequence. The equal time interval sequence is subjected to a fast Fourier transform and the power spectral density is calculated. Numerical integration is performed in the preset first frequency band and second frequency band respectively to obtain the power values ​​of the first frequency band and the second frequency band. The ratio of the power value of the first frequency band to the power value of the second frequency band is calculated and used as the heart rate variability. The ureteral pressure signal and local electromyographic activity signal were acquired simultaneously. The local electromyographic activity signal was rectified by full wave and passed through a third-order Butterworth low-pass filter with a cutoff frequency of the system preset value to generate a smoothed local electromyographic activity signal. The smoothed local electromyographic activity signal was used as the input variable and the ureteral pressure signal was used as the output variable to construct a linear regression model. The regression coefficient and bias term were solved by regularized least squares method. The regression coefficient was used as the dynamic mode of ureteral pressure and the bias term was used as the feature of local electromyographic activity. Step two involves modeling multimodal signal interaction relationships using graph neural networks and combining causal time maps, latent variable recursive models, and reinforcement learning agents for joint modeling. The dynamic features of inflammatory factors, heart rate variability index, ureteral pressure response gain and baseline tension features are obtained. The four features are used as graph nodes to construct a graph structure. The connection relationship between nodes is determined by Granger causality analysis. The initial feature vector of each node is the normalized value of the corresponding feature. Two layers of graph convolution are performed. Each layer uses a learnable weight matrix to perform a weighted summation of the features of neighboring nodes. The summation result is transformed by the ReLU activation function and then updated to update the node representation. The four node feature vectors output by the second layer are concatenated in order to generate a multimodal fusion feature vector. The multimodal fusion feature vector is input into a two-layer fully connected neural network. The network output is added to Gaussian noise with a mean of zero and a covariance matrix of diagonal to obtain the latent state vector, which is expressed by the following formula: ,in This represents the latent state vector at time t. This represents the multimodal fusion feature vector at time t. This represents a mapping function implemented by a two-layer fully connected neural network, with ReLU as the activation function and a parameter set of... , This represents independent and identically distributed Gaussian noise with a zero mean vector and a diagonal covariance matrix. ; The latent state vector is concatenated with the multimodal fusion feature vector to form a state input vector. The state input vector is fed into the reinforcement learning agent. The agent selects an action according to the preset set of discrete risk control instructions, calculates the reward value based on the weighted sum of the risk level change and the intervention cost, updates the policy network parameters using the proximal policy optimization algorithm, and outputs the selected risk control instruction. Step two involves the process of predicting the temporal risk of spasticity and classifying its risk level based on the extracted multimodal features. Receive dynamic characteristics of inflammatory factors, heart rate variability, dynamic patterns of ureteral pressure, and local electromyographic activity characteristics, and perform the following operations: The multimodal features are arranged in chronological order to form an input sequence of length T, which is then fed into a bidirectional long short-term memory network. The network decoder generates a sequence of spasm risk probabilities for the next N time steps, with each time step outputting a value between 0 and 1. The maximum value in the spasticity risk probability sequence is taken as the spasticity risk assessment r(t) for the current period. Substituting r(t) into the risk level determination rule, the risk level L(t) is calculated. The determination rule is as follows: ,in This indicates the risk score for spasms in the current cycle. This indicates the first risk threshold. This represents the second risk threshold, and , Indicates the risk level for the current period; Output the probability sequence of spasm risk for the next N time steps and the risk level for the current period. The risk level is updated every N time steps, and the update cycle is consistent with the system's preset time interval. By constructing a multimodal dynamic feature system based on physiological mechanisms, combining graph neural networks and Granger causal analysis to model cross-system interaction relationships, introducing a latent variable recursive model to characterize unobserved physiological states, and integrating reinforcement learning to optimize closed-loop risk control strategies; at the same time, using a bidirectional long short-term memory network for temporal prediction and hierarchical classification of spasticity risk, a leap from "static judgment" to "dynamic inference, causal explanation, and intelligent intervention" has been achieved, significantly improving the accuracy, foresight, and clinical usability of postoperative spasticity risk warning.

[0019] Step 3: Based on the temporal prediction results and risk level classification of spasticity risk, a visualization panel is used to display the dynamics of inflammatory factors, trends of physiological signs, treatment recommendations and risk alerts. Combined with the model interpretation mechanism, it provides decision-making basis prompts and abnormal event reminders. Step three, which involves generating a visual output based on the temporal prediction results and risk level classification of spasticity risk, includes: Receive dynamic characteristics of inflammatory factors, heart rate variability, dynamic patterns of ureteral pressure, regional electromyographic activity characteristics, temporal prediction results of spasticity risk, risk level, and treatment recommendations, and perform the following operations: The time-series data of interleukin-6 and C-reactive protein concentrations are sorted by sampling time to generate a dynamic data sequence of inflammatory factors. Each data point includes a timestamp and a normalized concentration value. The time-series data of heart rate variability, arterial systolic blood pressure, arterial diastolic blood pressure, axillary skin temperature, ureteral pressure, and urine flow rate are aligned by time to generate a physiological sign trend dataset. The time axis range covers the most recent 24 hours. The spasticity risk scores for the next N time steps in the spasticity risk time-series prediction results are merged with the historical risk scores to generate a complete risk trend sequence. The first risk threshold and the second risk threshold are output as reference line parameters. The current risk level is mapped to a visual identifier, with the following mapping relationship: green for risk level 1, yellow for risk level 2, and red for risk level 3. When the risk level is level 3, an alarm event is generated. The alarm event includes the alarm level, trigger time, patient identifier, and a list of associated parameter tags. The treatment recommendation text is parsed into structured recommendation items, each of which includes the recommendation type, suggested measures, and confidence score. The dynamic data sequence of inflammatory factors, the physiological sign trend dataset, the risk trend sequence, the risk level visual identifier, the alarm event, and the structured recommendation items are merged into a unified visualization data package. The visualization data package is pushed to the front-end client via the WebSocket protocol, triggering the visualization panel to update. Step three, which combines the model's interpretation mechanism to provide decision-making support and alerts to abnormal events, includes: Obtain the feature importance weights and corresponding influence directions generated during the spasticity risk prediction process (the contribution value of each physiological parameter to the current risk prediction result is calculated using the SHAP algorithm to generate feature importance weights, and the influence direction is determined according to the sign of the contribution value; positive numbers indicate positive influence, and negative numbers indicate negative influence. The feature importance weights and influence directions are used as additional outputs and packaged together with the risk score into a prediction result package). Obtain the raw measurement values ​​of heart rate, systolic blood pressure, diastolic blood pressure, and axillary skin temperature. Obtain the lower and upper limits of the normal range for each physiological parameter and perform the following operations: The feature importance weights are matched according to the physiological parameter type, and sorted in descending order according to the weight value. The top three physiological parameters are selected as the main influencing factors. For each main influencing factor, a decision item is generated based on its influence direction. Each item includes the physiological parameter name, influence direction and normalized weight value. All decision items are compiled into a decision basis dataset. The original measurements of heart rate, systolic blood pressure, diastolic blood pressure, and axillary skin temperature are checked one by one and compared with the corresponding lower and upper limits of the normal range. When the measured value is lower than the lower limit of the normal range or higher than the upper limit of the normal range, a corresponding abnormal event record is generated. Each record includes the physiological parameter name, abnormality type, detection time, and measurement value. All abnormal event records are compiled into an abnormal event dataset. The decision basis dataset and the abnormal event dataset are written into a dedicated field of the visualization data packet. The visualization data packet containing the decision basis and abnormal event information is sent to the front-end terminal through the WebSocket communication protocol to trigger the update of the interface content. By constructing a time-aligned multidimensional visualization panel, the spatiotemporal alignment and visualization integration of multimodal physiological data and risk prediction results are achieved. The efficiency of clinical situation awareness is improved by using color-coded risk levels, dynamic trend charts, and structured treatment recommendations. The SHAP model interpretation mechanism is introduced to output key influencing factors and their contribution, enhancing the credibility of AI decision-making. Abnormal event alerts are generated by combining physiological parameter threshold rules, forming a dual early warning system of "model-driven + rule verification". All information is encapsulated into structured data packets and pushed to the front end in real time via WebSocket, supporting low-latency, auditable, and operable closed-loop clinical decision support.

[0020] Example 2: The technical solution of this embodiment of the invention differs from that of Example 1 in that: like Figure 1 As shown, step four: when the spasticity risk prediction result reaches the preset threshold, a gene-phenotype decoding model is constructed based on the patient's gene chip data and phenotypic characteristics. Combined with the adaptive intervention mechanism of reinforcement learning and the cross-patient comparative learning algorithm, an individualized drug intervention path is generated and the dosage and intensity parameters are dynamically updated. Step four, which involves constructing a gene-phenotype decoding model and designing an adaptive intervention mechanism based on reinforcement learning, includes: We acquire patients' single nucleotide polymorphism (SNP) data, clinical symptom scores, neurological signs, laboratory test indicators, and physiological monitoring parameters. We perform quality control on the SNP data, removing low-quality samples and variant sites. We standardize the remaining data, use principal component analysis to eliminate batch effects and technical noise, extract principal component vectors, retain principal components whose cumulative variance contribution rate reaches a preset proportion, and generate gene principal component representations. We perform Z-score standardization on continuous variables in phenotypic data, perform one-hot encoding on categorical variables, and use multiple imputation to fill missing values, generating standardized phenotypic vectors. We concatenate the gene principal component representations with the standardized phenotypic vectors to form a joint feature vector. The joint feature vector is input into a multilayer perceptron model consisting of an input layer, multiple fully connected hidden layers, and an output layer. The number of neurons in the hidden layers decreases layer by layer in a fixed proportion. Each layer uses the ReLU activation function. The output layer contains one neuron and uses the Sigmoid activation function to output an estimated probability of spasticity. An L1 regularization term is introduced during model training. The regularization coefficient is determined through a hyperparameter optimization process. The model performance is evaluated using a cross-validation method. The model parameter with the highest AUC is selected on the validation set to obtain the gene-phenotype decoding model. The principal component representation of genes, the standardized phenotypic vector, and the estimated probability of spasticity risk output from the gene-phenotype decoding model are concatenated to form the state vector of the Markov decision process. The action space is defined as the Cartesian product of the drug type index and the dosage level index. The drug type set includes anticonvulsant drugs listed in clinical guidelines, and the dosage level is divided into five discrete levels. Each action uniquely corresponds to a drug and its dosage level. The reward signal is calculated based on the changes in spasticity risk and key phenotypic indicators before and after intervention. ,in This represents the immediate reward at time step t. This represents the predicted risk of spasticity output by the gene-phenotype decoding model before intervention. This represents the predicted risk of spasticity output by the gene-phenotype decoding model after intervention. This indicates the numerical value of the key reference phenotypic index before intervention, specifically the muscle tone score. This indicates the value of the key reference phenotypic index after intervention, specifically the muscle tone score. The weighting coefficients representing the risk variation terms are preset non-negative constants. The weighting coefficients for the phenotypic change terms are preset non-negative constants; Deep Q-networks are used to implement policy learning. The Q-function is approximated by a neural network. The input is a state vector, and the output is the Q-value corresponding to each action in the action space. The experience replay buffer stores the historical state transition sequence. During training, samples are randomly drawn from the buffer for gradient updates. The target network synchronizes parameters at fixed step intervals. The discount factor and learning rate are preset constants. The deep Q-network model after training convergence is determined as an adaptive intervention mechanism to generate individualized drug intervention suggestions. Step four, which involves generating a personalized drug intervention pathway based on a cross-patient contrastive learning framework and dynamically adjusting dosage and intensity parameters during execution, includes: Obtain the principal component representations of genes and standardized phenotypic vectors from a multicenter patient cohort. For any two patients i and j, concatenate their principal component representations and standardized phenotypic vectors and input them into a shared encoder to generate fixed-dimensional feature embeddings. The shared encoder consists of an input layer, multiple fully connected hidden layers, and an output layer. Each layer uses the ReLU activation function. The contrastive loss function is defined as the squared Euclidean distance between the feature embeddings of the two patients. ,in This represents the loss value for contrastive learning. This represents the feature embedding generated after mapping the gene principal component vector and the standardized phenotype vector of patient i through a shared encoder. This represents the feature embedding generated after mapping the gene principal component vector and the standardized phenotype vector of patient j through a shared encoder. Let be the principal component vectors of the genes of patient i and patient j, respectively. These are the standardized phenotypic vectors for patient i and patient j, respectively. Positive sample pairs are selected when the cosine similarity of the gene features of two patients is greater than a set threshold and the direction of change of muscle tone scores is the same. Negative sample pairs are selected when the direction of change of muscle tone scores is opposite and the absolute change is greater than a set threshold. In a multi-center patient cohort, the Adam optimizer is used to update parameters with the contrast loss function as the objective and the batch size is a preset value. The pre-training of the shared encoder is completed, the gene principal component representation and standardized phenotypic vector of the current patient are obtained, the gene principal component representation and standardized phenotypic vector of the current patient are concatenated and input into the pre-trained shared encoder to generate feature embeddings. The feature embedding is input into a deep Q-network model to obtain the Q value corresponding to each action in the action space. The action with the largest Q value is selected as the recommended intervention. The recommended action is mapped to a specific drug name and dosage value to form an individualized drug intervention plan. At a fixed time point every day, the patient's latest gene principal component representation and standardized phenotypic vector are re-acquired. The feature embedding generation, strategy reasoning and action selection process are repeated to generate a new drug intervention plan. The new intervention plan is executed in the next 24-hour cycle. By constructing a gene-phenotype decoding model, we can achieve joint risk modeling of genetic background and clinical phenotype; by using cross-patient comparative learning to pre-train a shared encoder, we can improve the generalization ability of individual feature representations; by combining deep Q-networks to design an adaptive intervention mechanism, with risk and phenotype improvement as reward objectives, we can generate interpretable personalized drug intervention pathways; and by supporting daily dynamic updates of dosage and drug selection, we can achieve closed-loop optimization of "perception-decision-execution", which can significantly improve treatment accuracy and long-term adaptability.

[0021] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring, characterized in that, Includes the following steps: Step 1: Collect postoperative core physiological monitoring parameters from blood, body surface and urination process, and perform time alignment, anomaly identification, signal denoising, data completion and data normalization on the collected parameters in sequence; Step 2: Based on the pre-processed postoperative core physiological monitoring parameters, extract the dynamic characteristics of inflammatory factors, heart rate variability, ureteral pressure dynamic patterns and local electromyographic activity characteristics, model their interaction relationships through graph neural networks, and combine causal time maps, latent variable recursive models and reinforcement learning agents to output the time-series prediction results and risk level classification of spasticity risk. Step 3: Based on the temporal prediction results and risk level classification of spasticity risk, a visualization panel is used to display the dynamics of inflammatory factors, trends of physiological signs, treatment recommendations and risk alerts. Combined with the model interpretation mechanism, it provides decision-making basis prompts and abnormal event reminders. Step 4: When the spasticity risk prediction results reach the preset threshold, a gene-phenotype decoding model is constructed based on the patient's gene chip data and phenotypic characteristics. Combined with the adaptive intervention mechanism of reinforcement learning and the cross-patient comparative learning algorithm, an individualized drug intervention path is generated and the dosage and intensity parameters are dynamically updated.

2. The method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring according to claim 1, characterized in that, The process of extracting dynamic features of inflammatory factors, heart rate variability, ureteral pressure dynamic patterns, and local electromyographic activity features based on the pre-treated postoperative core physiological monitoring parameters in step two includes: We obtained time-series concentration data of interleukin-6 and C-reactive protein at a fixed sampling period, constructed a first-order ordinary differential equation model, and calculated the concentration rise time, peak concentration, half-life and area under the curve based on the model fitting curve. The concentration rise time, peak concentration, half-life and area under the curve were used as dynamic features of inflammatory factors. Continuous electrocardiogram (ECG) signals are acquired, the peak position of the R wave is identified by threshold detection, and adjacent R wave interval sequences are generated. The interval sequences are linearly interpolated with the arithmetic mean as the center to form an equal time interval sequence. The equal time interval sequence is subjected to a fast Fourier transform and the power spectral density is calculated. Numerical integration is performed in the preset first frequency band and second frequency band respectively to obtain the power values ​​of the first frequency band and the second frequency band. The ratio of the power value of the first frequency band to the power value of the second frequency band is calculated and used as the heart rate variability. Ureteral pressure signals and local electromyographic activity signals were acquired simultaneously. The local electromyographic activity signals were rectified by full-wave rectification to generate smoothed local electromyographic activity signals. A linear regression model was constructed using the smoothed local electromyographic activity signals as input variables and ureteral pressure signals as output variables. The regression coefficients and bias terms were solved using the regularized least squares method. The regression coefficients were used as the dynamic mode of ureteral pressure, and the bias terms were used as the characteristics of local electromyographic activity.

3. The method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring according to claim 2, characterized in that, The second step, which involves modeling multimodal signal interaction relationships using graph neural networks and combining causal time maps, latent variable recursive models, and reinforcement learning agents for joint modeling, includes: The dynamic features of inflammatory factors, heart rate variability index, ureteral pressure response gain and baseline tension features are obtained. The four features are used as graph nodes to construct a graph structure. The initial feature vector of each node is the normalized value of the corresponding feature. Two layers of graph convolution operation are performed to generate a multimodal fusion feature vector. The multimodal fusion feature vector is input into a two-layer fully connected neural network. The network output is added to Gaussian noise with a zero mean vector and a diagonal covariance matrix to obtain a latent state vector. The latent state vector is concatenated with the multimodal fusion feature vector to form a state input vector. The state input vector is fed into a reinforcement learning agent. The agent selects an action based on a preset set of discrete risk control instructions, calculates the reward value based on the weighted sum of the risk level change and the intervention cost, updates the policy network parameters using a proximal policy optimization algorithm, and outputs the selected risk control instruction.

4. The method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring according to claim 3, characterized in that, The process of predicting the temporal risk of spasticity and classifying the risk level based on the extracted multimodal features in step two includes: Receive dynamic characteristics of inflammatory factors, heart rate variability, dynamic patterns of ureteral pressure, and local electromyographic activity characteristics, and perform the following operations: The multimodal features are arranged in chronological order to form an input sequence of length T, which is then fed into a bidirectional long short-term memory network. The network decoder generates a sequence of spasm risk probabilities for the next N time steps. Take the maximum value in the spasm risk probability sequence as the spasm risk rating r(t) for the current period, substitute r(t) into the risk level determination rule, calculate the risk level L(t), and output the spasm risk probability sequence for the next N time steps and the risk level for the current period. .

5. The method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring according to claim 1, characterized in that, Step three, which involves generating a visual output based on the temporal prediction results and risk level classification of spasticity risk, includes: Receive dynamic characteristics of inflammatory factors, heart rate variability, dynamic patterns of ureteral pressure, regional electromyographic activity characteristics, temporal prediction results of spasticity risk, risk level, and treatment recommendations, and perform the following operations: The time-series data of interleukin-6 and C-reactive protein concentrations were sorted by sampling time to generate a dynamic data sequence of inflammatory factors. The time-series data of heart rate variability, arterial systolic blood pressure, arterial diastolic blood pressure, axillary skin temperature, ureteral pressure, and urine flow rate were aligned by time to generate a physiological sign trend dataset. The spasticity risk scores for the next N time steps in the spasticity risk time-series prediction results were merged with the historical risk scores to generate a complete risk trend sequence. The current risk level is mapped to a visual identifier. When the risk level is Level 3, an alarm event is generated. The treatment recommendation text is parsed into structured recommendation items. The dynamic data sequence of inflammatory factors, the physiological sign trend dataset, the risk trend sequence, the risk level visual identifier, the alarm event, and the structured recommendation items are merged into a unified visualization data package. The visualization data package is pushed to the front-end client via the WebSocket protocol, triggering the visualization panel to update.

6. The method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring according to claim 5, characterized in that, Step three, which combines the model interpretation mechanism to provide decision-making basis prompts and abnormal event alerts, includes the following: Obtain the feature importance weights and corresponding influence directions generated during the spasticity risk prediction process; obtain the raw measurement values ​​of heart rate, systolic blood pressure, diastolic blood pressure, and axillary skin temperature; obtain the lower and upper limits of the normal range for each physiological parameter; and perform the following operations: The importance weights of features are matched according to the type of physiological parameter, and then sorted in descending order according to the weight value. The top three physiological parameters are selected as the main influencing factors. For each main influencing factor, a decision item is generated based on its direction of influence. All decision items are compiled into a decision basis dataset. The original measurements of heart rate, systolic blood pressure, diastolic blood pressure, and axillary skin temperature are checked one by one and compared with the corresponding lower and upper limits of the normal range. When the measured value is lower than the lower limit of the normal range or higher than the upper limit of the normal range, the corresponding abnormal event record is generated. All abnormal event records are compiled into an abnormal event dataset. The decision basis dataset and the abnormal event dataset are written into a dedicated field of the visualization data package.

7. The method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring according to claim 1, characterized in that, The process of constructing the gene-phenotype decoding model and designing an adaptive intervention mechanism based on reinforcement learning in step four includes: Acquire patients' single nucleotide polymorphism (SNP) data, clinical symptom scores, neurological signs, laboratory test indicators, and physiological monitoring parameters. Perform quality control on the SNP data, standardize the remaining data, extract principal component vectors, generate gene principal component representations, perform Z-score standardization on continuous variables in phenotypic data, generate standardized phenotypic vectors, and concatenate the gene principal component representations with the standardized phenotypic vectors to form a joint feature vector. The joint feature vector is input into a multilayer perceptron model consisting of an input layer, multiple fully connected hidden layers and an output layer, and the output is an estimated value of the risk of spasticity. An L1 regularization term is introduced during the model training process to obtain a gene-phenotype decoding model. The principal component representation of genes, the standardized phenotypic vector, and the estimated probability of spasticity risk output from the gene-phenotype decoding model are concatenated to form the state vector of the Markov decision process. The action space is defined as the Cartesian product of the drug type index and the dosage level index. Each action uniquely corresponds to a drug and its dosage level. The reward signal is calculated based on the changes in spasticity risk and key phenotypic indicators before and after intervention. A deep Q-network is used to achieve policy learning. The deep Q-network model after training convergence is determined as the adaptive intervention mechanism to generate individualized drug intervention recommendations.

8. A method for preventing and controlling postoperative spasm after ureterolithotomy based on inflammatory factor monitoring according to claim 7, characterized in that, Step four, which involves generating a personalized drug intervention pathway based on a cross-patient contrastive learning framework and dynamically adjusting the dosage and intensity parameters during execution, includes: Obtain the principal component representation of genes and the standardized phenotypic vector of a multicenter patient cohort. For any two patients i and j, concatenate their principal component representation of genes and the standardized phenotypic vector and input them into a shared encoder to generate a fixed-dimensional feature embedding. The contrastive loss function is defined as the square of the Euclidean distance between the feature embeddings of the two patients. In a multi-center patient cohort, the Adam optimizer was used to update parameters with the contrastive loss function as the objective, and the shared encoder was pre-trained. The gene principal component representation and standardized phenotypic vector of the current patient were obtained. The gene principal component representation and standardized phenotypic vector of the current patient were concatenated and input into the pre-trained shared encoder to generate feature embeddings. The feature embedding is input into a deep Q-network model to obtain the Q value corresponding to each action in the action space. The action with the largest Q value is selected as the recommended intervention. The recommended action is mapped to a specific drug name and dosage value to form an individualized drug intervention plan. At a fixed time point every day, the patient's latest gene principal component representation and standardized phenotypic vector are re-acquired. The feature embedding generation, strategy reasoning and action selection process are repeated to generate a new drug intervention plan.

Images