Lung cancer postoperative management method and system based on dynamic symptom network and reinforcement learning

Abstract

The invention discloses a dynamic symptom network and reinforcement learning-based lung cancer postoperative management method, which comprises the following steps: S1, acquiring postoperative electronic medical records of a patient, physiological time sequence data of wearable equipment and report outcome and environmental data of the patient, and carrying out standardization and feature engineering processing; and S2, constructing an initial symptom network by taking medical symptoms as nodes and taking a time sequence causal relationship determined by cross-lag analysis among the symptoms as a directed edge, performing embedded representation on the nodes by adopting a graph attention network, generating a dynamic symptom network model, and identifying core symptom nodes and bridge symptom nodes based on a network centrality index of the model. The cancer postoperative symptom management method solves the technical problems that in the prior art, a cancer postoperative symptom management method generally depends on static evaluation and isolated intervention, the dynamic evolution law and the time sequence causal relationship among multiple symptoms are difficult to capture, and the self-adaptive optimization capability based on individual response is lacked, so that intervention lags, target spots are inaccurate, and the effect is limited.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare and digital therapy technology, specifically to a method and system for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning. Background Technology

[0002] Postoperative management of lung cancer is a crucial aspect affecting patients' postoperative recovery and long-term prognosis. During this period, patients often suffer from a variety of symptoms such as pain, fatigue, dyspnea, anxiety, and sleep disorders. Current clinical practices and technical solutions have significant limitations: First, symptom management largely relies on nurse-led intermittent assessments and experience-based interventions, lacking sufficient understanding of the complex interactions and dynamic evolution patterns between symptoms, resulting in inaccurate intervention targets. Second, existing electronic health record systems or mobile health applications mostly only achieve digital recording and simple reminders of symptoms, lacking the ability to predict symptom risks in real time and prospectively. Finally, although some studies have attempted to introduce artificial intelligence for decision-making assistance, their models are mostly static and cannot adaptively optimize based on the individualized responses of patients, making it difficult to achieve truly personalized management. Summary of the Invention

[0003] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method and system for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning. This addresses the technical problems in existing cancer postoperative symptom management methods, which generally rely on static assessment and isolated intervention, making it difficult to capture the dynamic evolution patterns and temporal causal relationships among multiple symptoms, and lacking adaptive optimization capabilities based on individual responses, resulting in delayed intervention, inaccurate targeting, and limited effectiveness.

[0004] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a method for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning, comprising:

[0006] S1. Collect postoperative electronic medical records, wearable device physiological time-series data, patient-reported outcomes and environmental data, and perform standardization and feature engineering processing.

[0007] S2. Using medical symptoms as nodes and the temporal causal relationships between symptoms determined by cross-lag analysis as directed edges, an initial symptom network is constructed. A graph attention network is used to embed the nodes to generate a dynamic symptom network model. Based on the network centrality index of this model, core symptom nodes and bridge symptom nodes are identified.

[0008] S3. Integrate the topological features of the dynamic symptom network with the patient's real-time physiological data, score the severity of symptoms using the XGBoost algorithm, and output green, yellow, and red three-color graded warning signals;

[0009] S4. The key node information, risk stratification structure and individual patient characteristics are combined to form the state space of the reinforcement learning model. The preset intervention measure library is used as the action space, and the improvement of patient symptom scores and quality of life are used as reward signals. The optimal intervention strategy is trained and iteratively updated in real time through the proximal strategy optimization algorithm.

[0010] S5. Implement personalized intervention strategies through patient terminals and medical staff terminals, collect patient compliance data and physiological feedback data, and transmit the data back to drive online adaptive optimization of the strategy.

[0011] In some embodiments, the temporal causal relationship in S2 is determined by analyzing the cross-lag coefficients between symptom variables in panel data and ensuring statistical significance.

[0012] In some embodiments, in S2, core symptom nodes and bridge symptom nodes are identified to calculate the degree centrality and number centrality of the network center nodes. Nodes with a degree centrality higher than a first threshold are identified as core symptoms, and nodes with a number centrality higher than a second threshold are identified as bridge symptoms.

[0013] In some embodiments, the reward signal in S4 is R, and the specific calculation formula is as follows: ,in The value represents the decrease in symptom severity, among which The improvement value for the quality of life score, among which For intervention compliance scores, the aforementioned , , These are the preset weighting coefficients.

[0014] In some embodiments, it includes a multi-source data fusion processing module, a time-series symptom network analysis module, a comprehensive risk assessment and early warning module, a reinforcement learning decision-making and optimization module, and a human-computer collaborative interaction and feedback module;

[0015] The multi-source data fusion processing module is used to collect patient data and perform standardization and feature engineering processing.

[0016] The temporal symptom network analysis module is used to construct an initial symptom network and identify key nodes;

[0017] The comprehensive risk assessment and early warning module is used to integrate the information output by the multi-source data fusion processing module and the time-series symptom network analysis module, combine it with the patient's real-time physiological data, use the XGBoost algorithm to calculate the symptom severity score, and issue a graded early warning signal accordingly.

[0018] The reinforcement learning decision and optimization module is used to find the optimal intervention measures through reinforcement learning methods.

[0019] The human-machine collaborative interaction and feedback module is responsible for executing personalized intervention measures and collecting feedback information after execution, so as to facilitate the system's self-adjustment and optimization.

[0020] In some embodiments, the human-machine collaborative interaction and feedback module includes a patient terminal and a medical staff terminal; the patient terminal is used to receive early warning information, execute intervention strategies, and fill in patient report outcomes; the medical staff terminal is used to receive high-level early warnings, approve and approve access intervention strategies, and conduct remote medical interventions.

[0021] In some embodiments, the system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can implement a method for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning.

[0022] In some embodiments, a computer program is stored thereon that, when executed by a processor, implements a dynamic symptom network and reinforcement learning method for postoperative management of lung cancer.

[0023] Compared with existing technologies, this invention provides a dynamic symptom network and reinforcement learning-based method and system for postoperative lung cancer management. By introducing cross-lag analysis and time-series panel data, this invention can identify causal driving relationships between symptoms across time points, such as "sleep disturbance leading to increased anxiety," and construct a directed dynamic symptom network accordingly. Compared with traditional symptom lists that rely solely on contemporaneous correlation or fixed rules, this method can truly reflect the evolution mechanism of postoperative symptom clusters, providing a scientific basis for prospective intervention. By embedding symptom nodes using graph attention networks and combining network topology indicators such as degree centrality and betweenness centrality, the system can automatically identify two types of key nodes: core symptoms (those with a wide impact and dominating overall deterioration) and bridge symptoms (those connecting multiple symptom subgroups and easily triggering chain reactions). Prioritizing intervention for these nodes can effectively block the vicious cycle of symptoms and avoid inefficient management that is "all-encompassing but lacks focus." The structural characteristics of the dynamic symptom network, such as the activity of key nodes and changes in network density, are combined with real-time physiological indicators collected by wearable devices, such as heart rate variability and blood oxygen saturation, and input into XG. The Boost model generates a quantitative symptom severity score and outputs green, yellow, and red warning signals. This mechanism not only improves the sensitivity and specificity of risk identification but also facilitates the on-demand allocation of medical resources and enables tiered responses. This invention innovatively introduces reinforcement learning into postoperative symptom management: using the patient's current state, including key symptoms, risk level, and individual characteristics, as the state space; using a library of non-pharmacological digital therapies, such as breathing training, cognitive behavioral intervention, and health education, as the action space; and using symptom improvement, quality of life improvement, and compliance as composite reward signals, the system continuously iterates the optimal strategy through the proximal strategy optimization (PPO) algorithm. During use, the system continuously learns individual response patterns, making intervention strategies "more accurate with use," overcoming the shortcomings of existing models that are "one-time deployment and cannot evolve." Personalized interventions are executed through the patient terminal, and compliance and physiological feedback data are automatically collected. At the same time, it supports manual intervention and strategy approval for high-risk events by medical staff, ensuring the executability and safety of digital therapies while providing high-quality online learning signals for the reinforcement learning model, driving the entire system to evolve towards better decision-making. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the process of the lung cancer postoperative management method using dynamic symptom networks and reinforcement learning provided in an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the framework of the dynamic symptom network and reinforcement learning-based lung cancer postoperative management system provided in an embodiment of the present invention;

[0026] Figure 3 This is a flowchart illustrating the temporal symptom network analysis module of the dynamic symptom network and reinforcement learning-based lung cancer postoperative management method provided in this embodiment of the invention.

[0027] Figure 4 This is a schematic diagram of the reinforcement learning decision-making and optimization module of the dynamic symptom network and reinforcement learning-based lung cancer postoperative management method provided in this embodiment of the invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0029] Please see Figure 1 , Figure 1 This is a schematic diagram of a lung cancer postoperative management method using dynamic symptom networks and reinforcement learning in one embodiment of the present invention. The method includes:

[0030] S1. The system obtains basic patient information, surgical records, pathological stages, and medication history from the hospital's electronic medical record (EMR) system. It continuously collects physiological time-series data such as heart rate, blood oxygen saturation, heart rate variability (HRV), including indicators like SDNN and RMSSD, and activity levels through wearable devices such as smart bracelets and chest-strap respiratory monitors. Patients submit daily Patient Reported Outcomes (PROs) via a mobile app, including scores for pain, fatigue, dyspnea, anxiety, and sleep quality. Simultaneously, it accesses environmental sensors or third-party platforms to obtain data on indoor temperature and humidity, PM2.5 concentration, light intensity, and noise levels. It also assesses social support levels using questionnaires. All raw data undergoes standardization: missing values ​​are filled using multiple imputation (MICE) or a time-series repair strategy based on generative adversarial networks (GANs); different dimensional features are normalized, and derived features such as symptom volatility and cumulative burden index are extracted to form a structured input vector.

[0031] S2. Based on the panel data above, cross-lag analysis is used to calculate the lagged regression coefficients between symptom variables at different time points, such as day t and day t+1. Statistically significant (p<0.05) causal directions are retained as directed edges. Then, symptoms are used as nodes and causal relationships as edges to construct an initial symptom network. A graph attention network (GAT) is introduced to embed this network, where GAT automatically learns the importance weights between symptom nodes, effectively mitigating the instability problem of high-dimensional sparse networks with small sample sizes. The two work together to ensure the interpretability of causal directions and improve the model's generalization ability. Furthermore, a time decay factor, such as exponential decay, is introduced into the network edge weights to make recent... Symptom relationships have a greater impact on the current state. The system dynamically reconstructs the symptom network daily or when a new symptom is detected for the first time or when a symptom score suddenly exceeds a preset threshold. Based on the updated network, the degree centrality and betweenness centrality of each node are calculated: nodes with a degree centrality higher than the first threshold, such as 0.8, are identified as core symptoms, such as "pain", and nodes with a betweenness centrality higher than the second threshold, such as 0.7, are identified as bridging symptoms, such as "sleep disorder". It should be noted that this method is not only applicable to patients after lung cancer surgery, but can also be extended to postoperative scenarios of solid tumors such as breast cancer, colorectal cancer, and gastric cancer, because its model does not rely on prior knowledge of cancer type, but learns the symptom evolution law based on individual data.

[0032] S3. The topological features of the dynamic symptom network, such as the activity of core symptoms, the connection strength of bridge symptoms, and the rate of change of network density, are input together with real-time physiological indicators, such as blood oxygen <92% and significant decrease in HRV, into the XGBoost classifier. The output is a symptom severity score of 0-100, which is divided into three levels according to the threshold: green is less than 40 points, which is low risk; yellow is between 40 and 70 points, which is medium risk; and red is greater than 70 points, which is high risk. There are three levels of warning signals.

[0033] S4. The state space for reinforcement learning is constructed using key node information, risk level, and individual patient characteristics, including age, comorbidities, and adherence history. The action space comes from a pre-set intervention library, including digital therapies such as pushing diaphragmatic breathing training videos, progressive muscle relaxation audio, and guided mindfulness meditation; behavioral guidance such as generating goal-oriented activity plans, such as "Starting from the 3rd day post-surgery, sit in a chair 3 times a day with the nurse's assistance, 5-10 minutes each time"; psychological support such as providing cognitive restructuring exercises, anonymous recovery patient stories, and emotional self-assessment feedback; and medical collaboration such as triggering follow-up visit reminders or remote consultation appointments.

[0034] S5. Personalized intervention plans are implemented via patient terminals and smartphone apps. Patient completion status and physiological feedback are automatically transmitted back. Medical staff terminals receive yellow / red alerts and approve or remotely intervene in high-risk intervention strategies, such as medication adjustments. All data is stored in the hospital's private cloud or a SaaS platform that complies with medical safety standards after being de-identified and encrypted. It supports integration with HIS / EMR systems. Some lightweight alert models can be deployed on terminal devices to achieve offline response and implement effect verification. In the simulated cohort, compared with traditional static management, this method reduces the average identification time of high-risk events, such as acute anxiety attacks, by 2.3 days and shortens the core symptom relief cycle by 35%. The dynamic network identifies the "shortness of breath-fatigue-sadness" symptom cluster.

[0035] In this embodiment, by introducing cross-lag analysis and time-series panel data, the present invention can identify the causal driving relationship between symptoms across time points, such as "sleep disorder leading to increased anxiety," and construct a directed dynamic symptom network based on this. Compared with traditional symptom lists that rely solely on contemporaneous correlation or fixed rules, this method can truly reflect the evolution mechanism of postoperative symptom clusters, providing a scientific basis for prospective intervention. By embedding symptom nodes using graph attention networks and combining network topology indicators such as degree centrality and betweenness centrality, the system can automatically identify two types of key nodes: core symptoms, which have a wide impact and dominate overall deterioration, and bridge symptoms, which connect multiple symptom subgroups and are prone to triggering chain reactions. Prioritizing intervention for these nodes can effectively block the vicious cycle of symptoms and avoid inefficient management that is "all-encompassing but lacks focus." The structural characteristics of the dynamic symptom network, such as the activity of key nodes and changes in network density, along with real-time physiological indicators collected by wearable devices, such as heart rate variability and blood oxygen saturation, are input into the XGBoost model to generate a quantitative symptom severity. The system scores patients and outputs green / yellow / red warning signals. This mechanism not only improves the sensitivity and specificity of risk identification but also facilitates the on-demand allocation of medical resources and enables tiered responses. This invention innovatively introduces reinforcement learning into postoperative symptom management. The state space is the patient's current state, including key symptoms, risk level, and individual characteristics. The action space is a library of non-pharmacological digital therapies, such as breathing training, cognitive behavioral intervention, and health education. The composite reward signals are symptom improvement, quality of life improvement, and compliance. The system continuously iterates the optimal strategy through the proximal strategy optimization (PPO) algorithm. During use, the system continuously learns individual response patterns, making the intervention strategy "more accurate with use." This overcomes the shortcomings of existing models that are "one-time deployment and cannot evolve." Personalized interventions are executed through the patient terminal, and compliance and physiological feedback data are automatically collected. At the same time, it supports manual intervention and strategy approval for high-risk events by medical staff. This ensures the executability and safety of digital therapies and provides high-quality online learning signals for the reinforcement learning model, driving the entire system to evolve towards better decision-making.

[0036] Example 1: A lung cancer postoperative patient initially presented with "pain" (core symptom) and "anxiety." Dynamic network analysis revealed that "sleep disorder" was the bridging symptom connecting the two, and the system risk assessment was "yellow." After comprehensively analyzing this network state, the reinforcement learning engine did not recommend a strong analgesic, but instead generated a combined strategy, including "progressive muscle relaxation" (targeting pain and anxiety) and "sleep hygiene education" (a precise intervention bridge). After the patient implemented the strategy, sleep improved, which in turn led to a decrease in pain and anxiety scores. This positive feedback, as a high-reward signal, reinforced the strategy, and the system thus accumulated better decision-making experience for patients with similar conditions.

[0037] Example 2: Precise intervention targeting the vicious cycle of postoperative pain-fear activity

[0038] Scenario: Mr. Zhao, on the first day after surgery, the system identified pain (the node with the highest intensity) as the most critical symptom through dynamic symptom network analysis. Further analysis revealed that the bridge strength of fear of activity (i.e., being afraid to move due to fear of pain) was extremely high. As a key bridge, it closely connected the "pain" symptom group with the symptom group of delayed respiratory function recovery, such as "shallow and rapid breathing, weak coughing," forming the core contradiction hindering rehabilitation.

[0039] System Action: The risk assessment is "yellow." Based on the principle of "breaking the vicious cycle and precisely intervening in the bridge," the reinforcement learning engine generates a combination of strategies: 1. Targeting the core: In addition to drug analgesia, add pain education, explain the postoperative pain pattern to the patient, and establish reasonable expectations; 2. Targeting the bridge: Develop a "goal-level activity plan," breaking down "early ambulation" into specific, actionable, and pain-free goals such as "getting out of bed and sitting in a chair 3 times a day with the assistance of a nurse, 5-10 minutes each time," and setting up a reward mechanism for completion on the patient's terminal.

[0040] Results and Feedback: On the 5th postoperative day, the patient's "fear activity" score decreased significantly due to successful completion of activity goals, and the willingness to cough up phlegm and take deep breaths increased, with accelerated recovery of lung function. Subsequently, due to the improvement in activity ability, the pain score also decreased from 7 to 4. The system recorded high reward signals, which strengthened the effectiveness of the combined strategy of "analgesia + cognitive behavioral intervention + goal decomposition" for this type of "pain-fear activity" network structure, and accumulated key decision-making experience for the management of pain dominated by psychological factors.

[0041] Example 3: Intervention targeting the postoperative mid-term symptom cluster of "shortness of breath-fatigue-emotional state".

[0042] Scenario: Ms. Li, a 60-year-old patient, underwent thoracoscopic lobectomy on the 5th day post-surgery. Systematic dynamic symptom network analysis showed that her core symptoms in the early postoperative period had evolved from "cough" to shortness of breath (intensity centrality = 1.19) and fatigue (intensity centrality = 1.04). The network model further revealed that sadness, as a key bridging symptom (bridge strength = 0.87), established a strong connection between the somatic symptom group (shortness of breath-fatigue) and the emotional symptom group (distress-sleep disturbance), forming a vicious cycle of "shortness of breath aggravating fatigue, fatigue inducing sadness, and sadness amplifying the feeling of shortness of breath."

[0043] System Action: Based on this complex network structure, the comprehensive risk assessment module classifies the risk level as "yellow (high)". The reinforcement learning decision engine then generates multi-dimensional intervention strategies: 1. Targeting core symptom clusters: Develop individualized progressive pulmonary rehabilitation plans, including three daily quantitative pursed-lip breathing exercises (to improve shortness of breath) and bedside stepping activities performed at different times (to combat fatigue). The intensity of all activities is dynamically adjusted based on real-time blood oxygen saturation data. 2. Precise intervention bridge: Daily scheduled push of mindfulness-based stress reduction audio based on cognitive behavioral therapy, and access to anonymous success stories of fellow rehabilitation patients through the patient terminal to build a peer support network. 3. Collaborative intervention: Play relaxation music before rehabilitation training to reduce anticipatory anxiety and improve treatment adherence.

[0044] Results and Feedback: After 72 hours of continuous intervention, dynamic monitoring data showed that patients' 6-minute walking distance increased by 25 meters, their shortness of breath score decreased by 2 points, and their fatigue severity scale score improved by 30%. At the same time, the patient-reported outcome score of the emotional state subscale increased by 40%. After capturing this series of positive changes, the system generated a high-value reward signal. This not only verified the effectiveness of shifting the focus of intervention from respiratory symptoms to the comprehensive management of "shortness of breath-fatigue-emotion" in the mid-postoperative period, but more importantly, it strengthened the optimal decision-making path of "simultaneous physiological rehabilitation and psychological support" for specific symptom network topology, namely, for emotional symptoms connected to somatic symptom groups with high bridging strength. This provides a reusable clinical paradigm for the precise management of this patient group.

[0045] Example 4: Adaptive intervention for the "fatigue-cognitive" symptom cluster in elderly patients

[0046] Scenario: Mr. Zhou, 78 years old, underwent thoracoscopic segmentectomy for early-stage lung cancer on the 4th day post-surgery. Systemic dynamic symptom network analysis showed that his symptom network structure was significantly different from that of young and middle-aged patients: fatigue (node ​​strength = 1.25) and forgetfulness (bridge strength = 0.81) and drowsiness formed a tightly connected independent symptom cluster. This cluster was weakly connected to the typical respiratory symptom cluster (cough, shortness of breath), exhibiting a unique "fatigue-cognitive dominant" network characteristic. The network model further revealed that "forgetfulness" played a key bridging role in this cluster, connecting symptoms such as "fatigue" and "drowsiness".

[0047] System Action: The risk assessment module, combining patient age and network characteristics, identified a "yellow alert." The reinforcement learning engine recognized that traditional high-intensity pulmonary rehabilitation programs might be counterproductive under this network structure, and thus generated adaptive management strategies: 1. Fatigue Management: Implement a "segmented energy conservation" program, breaking down daily rehabilitation training into four 10-minute sessions, with mandatory rest periods during activity breaks; 2. Cognitive Activation: Schedule 10-minute cognitive training sessions twice daily, such as number sequence memory and object classification games, to specifically improve "forgetfulness" symptoms; 3. Sleep Regulation: Establish a regular sleep-wake cycle, limiting daytime bed rest time to no more than 2 hours to improve "drowsiness" symptoms; 4. Nutritional Support: Develop a high-protein, easily digestible nutritional plan (protein intake ≥1.5g / kg / day), and synchronize care points through family terminals;

[0048] Results and Feedback: Five days after implementation, multidimensional assessment showed that the patient's fatigue severity score improved by 35%, the Mini-Mental State Examination score improved by 4 points, and daytime awake time increased by 2 hours. Family members reported that the patient's "mental state improved significantly and cooperation improved." The system learned from this case that, for elderly patients with a "fatigue-cognition" dominant network structure, a comprehensive intervention strategy of "segmented activities + cognitive activation + sleep regulation" is more effective in improving the overall symptom burden than simple pulmonary rehabilitation training. This experience has been integrated into the system's decision-making module specifically for elderly patients, providing evidence-based support for managing elderly patients with similar network characteristics.

[0049] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning, characterized in that, include: S1. Collect postoperative electronic medical records, wearable device physiological time-series data, patient-reported outcomes and environmental data, and perform standardization and feature engineering processing. S2. Using medical symptoms as nodes and the temporal causal relationships between symptoms determined by cross-lag analysis as directed edges, an initial symptom network is constructed. A graph attention network is used to embed the nodes to generate a dynamic symptom network model. Based on the network centrality index of this model, core symptom nodes and bridge symptom nodes are identified. S3. Integrate the topological features of the dynamic symptom network with the patient's real-time physiological data, score the severity of symptoms using the XGBoost algorithm, and output green, yellow, and red three-color graded warning signals; S4. The key node information, risk stratification structure and individual patient characteristics are combined to form the state space of the reinforcement learning model. The preset intervention measure library is used as the action space, and the improvement of patient symptom scores and quality of life are used as reward signals. The optimal intervention strategy is trained and iteratively updated in real time through the proximal strategy optimization algorithm. S5. Implement personalized intervention strategies through patient terminals and medical staff terminals, collect patient compliance data and physiological feedback data, and transmit the data back to drive online adaptive optimization of the strategy.

2. The method for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning according to claim 1, characterized in that: The temporal causal relationship in S2 is determined by analyzing the cross-lag coefficients between symptom variables in the panel data and ensuring statistical significance.

3. The method for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning according to claim 1, characterized in that: In step S2, core symptom nodes and bridge symptom nodes are identified to calculate the degree centrality and number centrality of the network center nodes. Nodes with a degree centrality higher than a first threshold are identified as core symptoms, and nodes with a number centrality higher than a second threshold are identified as bridge symptoms.

4. The method for postoperative management of lung cancer using dynamic symptom networks and reinforcement learning according to claim 1, characterized in that: The reward signal in S4 is R, and the specific calculation formula is as follows: ,in The value represents the decrease in symptom severity, among which The improvement value for the quality of life score, among which For intervention compliance scores, the aforementioned , , These are the preset weighting coefficients.

5. A postoperative management system for lung cancer based on dynamic symptom networks and reinforcement learning, characterized in that: It includes a multi-source data fusion processing module, a time-series symptom network analysis module, a comprehensive risk assessment and early warning module, a reinforcement learning decision-making and optimization module, and a human-computer collaborative interaction and feedback module; The multi-source data fusion processing module is used to collect patient data and perform standardization and feature engineering processing. The temporal symptom network analysis module is used to construct an initial symptom network and identify key nodes; The comprehensive risk assessment and early warning module is used to integrate the information output by the multi-source data fusion processing module and the time-series symptom network analysis module, combine it with the patient's real-time physiological data, use the XGBoost algorithm to calculate the symptom severity score, and issue a graded early warning signal accordingly. The reinforcement learning decision and optimization module is used to find the optimal intervention measures through reinforcement learning methods. The human-machine collaborative interaction and feedback module is responsible for executing personalized intervention measures and collecting feedback information after execution, so as to facilitate the system's self-adjustment and optimization.

6. The postoperative management system for lung cancer based on dynamic symptom networks and reinforcement learning according to claim 5, characterized in that: The human-computer collaborative interaction and feedback module includes a patient terminal and a medical staff terminal; the patient terminal is used to receive early warning information, execute intervention strategies, and fill in patient report outcomes; the medical staff terminal is used to receive high-level early warnings, approve and approve admission intervention strategies, and conduct remote medical interventions.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 4.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 4.

Images