Cancer pain patient intelligent portraying and precise follow-up visit system based on dynamic mapping knowledge domain

Abstract

The invention discloses a cancer pain patient intelligent portraying and accurate follow-up visit system based on a dynamic knowledge graph, and belongs to the technical field of medical informatization, artificial intelligence, knowledge graphs and accurate medical crossing, and the method comprises the following steps: 1, constructing a double-knowledge graph architecture comprising a medical knowledge graph and a follow-up visit question knowledge graph, wherein the medical knowledge graph covers entities such as cancer pain related diseases, drugs, symptoms and treatment schemes and association relationships, and a mapping relationship between questions and medical entities is established by the follow-up question knowledge graph; 2, accessing HIS / LIS system data, wearable device monitoring data, follow-up feedback data and other multi-source heterogeneous data, and performing cleaning, standardization and desensitization processing to complete multi-source data fusion; and 3, constructing a patient intelligent portrait containing six dimensions of pathology, pain, treatment, psychology, society and dynamic characteristics based on the fusion data, and storing and supporting real-time dynamic updating in a JSON format.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of medical informatics, artificial intelligence, knowledge graphs, and precision medicine, specifically to an intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs. Background Technology

[0002] Cancer pain is one of the most common and unbearable symptoms among cancer patients, severely impacting their quality of life, sleep, and mood, and may even lead to suicidal tendencies. According to the World Health Organization (WHO), the incidence of pain in newly diagnosed cancer patients is approximately 25%, while this figure rises to 70%-80% in patients with advanced cancer, with about one-third of these patients suffering from severe pain. Although the WHO's three-step analgesic ladder has been promoted for many years, inadequate control of cancer pain remains common in clinical practice, especially breakthrough cancer pain (BTcP), which occurs in 40%-80% of cases. BTcP is characterized by its sudden onset, short duration (usually less than 30 minutes), and high intensity, making it difficult to manage with traditional scheduled medication regimens.

[0003] Current cancer pain management and follow-up systems have many core technological deficiencies:

[0004] Pain assessment is too simplistic: The existing assessment system mainly relies on single-dimensional indicators such as the Numerical Rating Scale (NRS), which only focuses on the physical intensity of pain and ignores the comprehensive influence of multiple factors such as the nature of pain, psychological state, and social support on pain perception, making it difficult to implement "holistic treatment of mind and body".

[0005] Standardized follow-up strategies (one-size-fits-all): Most follow-up systems use fixed time intervals and standardized questionnaire templates, applying the same follow-up frequency to patients in the stable and drastic fluctuation phases of their condition. This wastes medical resources, fails to intervene in the deterioration of high-risk patients' conditions in a timely manner, and lacks a personalized dynamic adjustment mechanism.

[0006] Knowledge fragmentation and the island effect: Clinical guidelines, drug instructions, literature research and expert experience related to cancer pain treatment are scattered across different carriers. The existing electronic medical record system (EMR) only serves as a data recording tool and has failed to achieve the systematic integration and structured association of fragmented knowledge. It is difficult for doctors to quickly obtain comprehensive evidence when formulating treatment plans.

[0007] Lack of predictive ability: Existing systems focus on historical data recording and review, lacking predictive ability based on time-series data. They cannot effectively predict the pattern of breakthrough pain attacks, the risk trend of pain worsening, and the probability of drug side effects, leaving treatment in a "passive firefighting" state.

[0008] Low communication efficiency and poor adherence: Follow-up mainly relies on manual telephone, which is inefficient and has a narrow coverage. Patients lack effective communication channels outside the hospital, and medication adherence is only 30%-40%. Patients often reduce the dosage or stop taking the medication on their own, which leads to repeated pain attacks.

[0009] Difficulties in multidisciplinary team (MDT) collaboration: Cancer pain management requires collaboration among multiple disciplines such as oncology, pain management, and psychology, but existing systems lack cross-departmental intelligent triggering and collaborative workflows, and data cannot be effectively shared, resulting in fragmented treatment plans;

[0010] Outpatient setting management has prominent pain points: Pain relief clinics face problems such as high patient mobility, traditional follow-up coverage of only 9.55%, and blind spots in outpatient management leading to a patient loss rate as high as 65%, which further exacerbates the difficulty of cancer pain management.

[0011] From a technological application perspective, traditional electronic medical record systems primarily use unstructured text data, making in-depth analysis difficult. Existing mobile healthcare follow-up apps mostly offer simple questionnaires and patient education functions, lacking powerful intelligent decision-making engines. While knowledge graph technology is applied in the medical field, it largely focuses on static knowledge base construction (such as disease-symptom queries), lacking dynamic knowledge graphs specifically for cancer pain that include temporal evolutionary relationships, thus failing to simulate the dynamic process of disease progression and treatment response. These issues collectively hinder the achievement of precise, personalized, and full-cycle coverage in cancer pain management, urgently requiring technological innovation to overcome existing bottlenecks. Summary of the Invention

[0012] The purpose of this invention is to provide an intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs, in order to solve the problems mentioned in the background art.

[0013] To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps:

[0014] Step 1: Construct a dual knowledge graph architecture that includes a medical knowledge graph and a follow-up question knowledge graph. The medical knowledge graph covers entities and relationships related to cancer pain, such as diseases, drugs, symptoms, and treatment plans. The follow-up question knowledge graph establishes a mapping relationship between questions and medical entities.

[0015] Step 2: Integrate multi-source heterogeneous data, including HIS / LIS system data, wearable device monitoring data, and follow-up feedback data, and complete the multi-source data fusion after cleaning, standardization, and desensitization.

[0016] Step 3: Construct a smart patient profile based on fused data, encompassing six dimensions: pathology, pain, treatment, psychology, social characteristics, and dynamic features. The profile is stored in JSON format and supports real-time dynamic updates.

[0017] Step 4: Use patient intelligent profiles and dual knowledge graphs for matching and reasoning, and use a multi-factor risk scoring algorithm to assess the patient's pain risk level;

[0018] Step 5: Based on the risk level and profile characteristics, dynamically generate a precise follow-up strategy that includes follow-up frequency, follow-up channels, and personalized questionnaire content. The personalized questionnaire is dynamically generated based on a knowledge graph of follow-up questions.

[0019] Step 6: Using the LSTM (Long Short-Term Memory) network model, input time-series data such as pain scores, medication times, and activity levels from the past 7 days to predict the probability of outbreak pain occurring within the next 24-72 hours and trigger yellow, orange, and red level warnings.

[0020] Step 7: Integrate text, voice, and behavioral multimodal data for sentiment analysis and risk identification to achieve intelligent hierarchical processing of ordinary, intermediate, and important responses;

[0021] Step 8: When a patient's psychological score exceeds the threshold or pain control is poor, a multidisciplinary consultation recommendation form is automatically generated and pushed to the relevant department's terminal.

[0022] Step 9: Collect follow-up feedback data and intervention effects to retrain the prediction model, and simultaneously perform incremental updates and optimizations on the dual knowledge graphs to achieve closed-loop iteration;

[0023] Step 10: In home-based palliative care scenarios, combine wearable devices with non-invasive, continuous monitoring data to provide remote nursing guidance and end-of-life care support;

[0024] As a further preferred embodiment of this technical solution: In step one, when constructing the dynamic medical knowledge graph, the BiLSTM-CRF model is used to identify 200+ disease entities, 500+ drug entities, 300+ symptom entities, and 1000+ treatment plan entities from multi-source data such as clinical guidelines, drug instructions, and medical records. Complex relationships such as treatment, causality, contraindications, and time sequence are extracted, and the fusion confidence is calculated through a weighted fusion algorithm. The formula is: Fusion Confidence = α・Guideline Authority + β・Literature Evidence Strength + γ・Real-world Data + δ・Timeliness, where α=0.4, β=0.3, γ=0.2, and δ=0.1.

[0025] As a further preferred embodiment of this technical solution: the multifactor risk scoring algorithm calculation formula in step four is: S=w1・NRS+w2・ΔTrend+w3・Freq_BTcP+w4・Score_Psy+w5・(100-Compliance)+w6・Progression, where NRS is the pain score, ΔTrend is the slope of the pain trend, Freq_BTcP is the frequency of breakthrough pain, Score_Psy is the psychological score, Compliance is the follow-up compliance rate, and Progression is the disease progression. Based on the score S, patients are divided into three levels: 0-30 points (low risk), 31-60 points (medium risk), and 61-100 points (high risk).

[0026] As a further optimization of this technical solution: the specific implementation method of the dynamic follow-up strategy in step five is as follows: low-risk patients are actively followed up once every 2 weeks, focusing on quality of life; medium-risk patients are actively followed up once a week, focusing on the effect of medication adjustment; high-risk patients are followed up once every 3 days or check in daily via the APP, closely monitoring breakthrough pain and drug side effects; the follow-up channels are selected based on decision tree, with priority given to human telephone for elderly or less educated patients, and priority given to APP push and WeChat reminders for middle-aged and young or highly compliant patients, and critical values ​​such as NRS>8 triggering dual alarms via SMS and telephone.

[0027] As a further preferred embodiment of this technical solution: the multimodal sentiment analysis in step seven adopts the PyTorch+Transformers framework, which realizes the judgment of the patient's emotional state through text sentiment recognition, speech feature extraction, and behavioral pattern analysis, and generates personalized response scripts in combination with empathic interaction strategies, while automatically filtering out invalid follow-up issues based on a 90% consistency threshold.

[0028] As a further preferred embodiment of this technical solution, it includes: a data layer module, used to access data from HIS / LIS / PACS systems, wearable devices, and follow-up data, and to perform data cleaning, standardization, and desensitization processing to meet the requirements for medical data security protection.

[0029] The dual knowledge graph construction module uses Neo4j database for storage and constructs medical knowledge graph and follow-up question knowledge graph through self-developed graph construction algorithm. It supports real-time individual graph updates and batch general graph optimization.

[0030] The patient profile building module generates a digital patient profile containing six dimensions, which is stored in a PostgreSQL database and linked to historical data and profile tags.

[0031] The follow-up strategy generation module dynamically outputs follow-up frequency, channels, and personalized questionnaire plans based on a risk scoring algorithm, and the follow-up records are stored in a MongoDB database.

[0032] The prediction and early warning module uses an LSTM neural network model to predict outbreak pain and is equipped with a three-level early warning mechanism.

[0033] The multimodal AI analysis module integrates text, voice, and behavioral data to perform sentiment analysis and risk classification.

[0034] The multidisciplinary collaboration module automatically initiates the MDT process and pushes a consultation recommendation form when the trigger conditions are met; the closed-loop optimization module feeds back the follow-up results to the system, driving model iteration and dual knowledge graph evolution.

[0035] The interaction layer module provides multiple interfaces, including a patient-side mini-program, a doctor-side web workbench, a case manager-side management backend, and a nurse's workbench.

[0036] As a further preferred embodiment of this technical solution: the medical knowledge graph in the dual knowledge graph construction module covers entity relationships such as cancer type, pain mechanism, treatment plan, drug effect, and complications; the follow-up question knowledge graph is based on the key nodes of the medical knowledge graph to build a targeted follow-up question library, establish a mapping relationship between questions and medical entities, and incorporate the disease development path prediction dimension to build a temporal relationship model.

[0037] As a further preferred embodiment of this technical solution: the LSTM model of the prediction and early warning module includes an input layer, two hidden layers and an output layer. The input features include at least the pain score sequence of the past 7 days, medication time points, activity levels and weather and air pressure data. The output is the probability distribution of outbreak pain in the next 24 to 72 hours, with a target prediction accuracy of >75%.

[0038] As a further preferred embodiment of this technical solution: the multidisciplinary collaboration module supports cross-departmental collaboration among oncology, pain management, psychology and other departments. When it is detected that the patient's PHQ-9 is greater than 15 or that the pain from bone metastases is not well controlled by medication, a consultation request form containing a summary of the patient's key profile is automatically generated and pushed to the relevant department's doctor workstation.

[0039] As a further preferred embodiment of this technical solution: the patient-side mini-program of the interaction layer module supports pain score reporting, medication records, follow-up feedback and patient education interaction functions; the nurse workbench has risk classification display, emergency treatment triggering and manual intervention recording functions; the management backend can realize data analysis, system configuration and effect monitoring; data from all terminals is synchronized in real time and supports traceability query.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] 1. Improve pain control: By constructing a six-dimensional intelligent patient profile covering pathological, pain, treatment, psychological, social and dynamic characteristics, and combining it with precise matching and reasoning of dynamic knowledge graphs, the rate of good pain control (NRS≤3) is significantly improved, the frequency of breakthrough pain is effectively reduced, and the transformation from "passive pain relief" to "active prevention" is achieved.

[0042] 2. Optimize the accuracy of follow-up strategies: Based on the multi-factor risk scoring algorithm, patients are stratified by risk. The frequency, channels and questionnaire content of follow-up are dynamically adjusted for low, medium and high-risk patients. Low-risk patients are followed up for a long period of time, while high-risk patients are monitored frequently or in real time. This ensures that high-risk patients receive timely intervention, avoids waste of medical resources, and improves the rationality of medical resource allocation.

[0043] 3. Break down knowledge and data silos: Construct a dynamic medical knowledge graph that includes diseases, drugs, symptoms, treatment plans, and time-series evolution relationships, integrate fragmented medical knowledge, and achieve deep integration of medical knowledge and follow-up scenarios through a dual knowledge graph architecture (medical knowledge graph + follow-up question knowledge graph), providing doctors with comprehensive and structured decision-making basis and supporting the generation of personalized follow-up questions.

[0044] 4. Achieve accurate prediction and graded early warning: Based on the LSTM long short-term memory network model, it integrates multimodal time series data to predict the probability of outbreak pain, with a prediction accuracy of over 75%. Combined with the yellow, orange, and red three-level early warning mechanism, it can provide early warning of pain worsening and drug side effects, providing a time window for preventive intervention. High-risk adverse reactions can be warned within 24 hours.

[0045] 5. Improve communication efficiency and patient compliance: Through multiple follow-up channels such as patient-side APP / mini-program, intelligent voice robot, and WeChat reminders, combined with the emotion recognition and personalized dialogue strategy of the empathy interaction module, the doctor-patient communication experience is improved, significantly increasing the patient follow-up completion rate (from the traditional 9.55% to over 85%) and medication compliance, with patient satisfaction reaching over 95 points.

[0046] 6. Promote efficient multidisciplinary collaboration: Set up a multidisciplinary collaboration intelligent trigger mechanism. When the patient's psychological score exceeds the threshold or pain control is poor, the system will automatically generate a consultation suggestion form and push it to the relevant department terminal, breaking down departmental barriers and realizing cross-disciplinary data sharing and collaborative diagnosis and treatment.

[0047] 7. Empowering full-cycle and home-based palliative care: Through non-invasive wearable device monitoring and remote nursing guidance, we build an integrated full-cycle management system that combines in-hospital and out-of-hospital care. In particular, we provide continuous monitoring, emergency intervention, and end-of-life care support for patients with advanced cancer at home, ensuring patient safety and comfort and improving dignity and quality of life in the final stages of life.

[0048] 8. Enhance management efficiency and research value: Improve the work efficiency of nursing staff by 30%, reduce the medical dispute rate by 30%, build a high-quality cancer pain management database, form standardized disease guidelines, promote academic research and industrial application of related technologies, and provide replicable solutions for similar hospitals.

[0049] 9. Achieve system self-evolution and continuous optimization: Through a closed-loop optimization mechanism, follow-up feedback data and real-world data (RWD) are used for model retraining and incremental updates of the knowledge graph. Combined with an automatic problem-suppression mechanism with a 90% consistency threshold, the system continuously improves its accuracy and applicability as it is used over time, thus optimizing the quality of medical services in the long term. Attached Figure Description

[0050] Figure 1 This is the overall architecture diagram of the intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs of this invention.

[0051] Figure 2 This is a flowchart illustrating the overall technical processing of the intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs, as described in this invention.

[0052] Figure 3 This is a flowchart illustrating the business process of the intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs, as described in this invention. Detailed Implementation

[0053] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Example

[0055] Please see Figures 1-3 This is a schematic diagram of some embodiments of the intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs, as described in this application.

[0056] In this embodiment, the present invention is applicable to multiple medical scenarios such as oncology, pain management, palliative care, and pain clinics. It focuses on the full life-cycle management needs of cancer pain patients, integrates dynamic knowledge graphs, multimodal AI analysis, and intelligent decision-making technologies, and creates a full-process intelligent management system from patient data access to follow-up closed-loop optimization, effectively solving various pain points in existing cancer pain management.

[0057] It should be noted that during the system deployment and environment setup phase, a server cluster meeting the needs of medical data processing must be configured, encompassing data storage servers, algorithm computing servers, and application service servers to ensure stable 24 / 7 system operation. Simultaneously, low-invasive wearable devices, such as heart rate monitoring bracelets and sleep monitors, should be provided for patients receiving home-based palliative care to enable real-time collection and transmission of physiological data. Regarding the software environment, system software compliant with medical data security requirements will be deployed. The data storage layer adopts a PostgreSQL + MongoDB + Neo4j architecture, where Neo4j stores medical knowledge graphs and follow-up question knowledge graphs, PostgreSQL stores patient profile data, and MongoDB stores follow-up records and interaction logs. The AI ​​framework uses PyTorch + Transformers, and the knowledge graph engine employs Neo4j combined with a self-developed graph construction algorithm to ensure the system possesses efficient data processing and intelligent analysis capabilities. The data interface development needs to achieve seamless integration with the hospital's HIS / LIS / PACS system, enabling smooth access to multi-source data such as patient basic information, diagnostic data, treatment plans, test results, and imaging reports. At the same time, it needs to develop data transmission interfaces for patient-side apps / mini-programs and wearable devices to support real-time uploading of various types of data from outside the hospital.

[0058] Specifically, in the initialization phase of the dynamic knowledge graph, a medical knowledge graph is first constructed. Using a BiLSTM-CRF deep learning model, over 200 disease entities, over 500 drug entities, over 300 symptom entities, and over 1000 treatment plan entities are accurately identified from national clinical guidelines, drug instructions, RCT research literature, and clinical medical records. Complex relationships such as treatment, causality, contraindications, and temporal sequence are meticulously extracted. A weighted fusion algorithm is used to calculate the fusion confidence score. The formula for fusion confidence score is: Fusion Confidence Score = α・Guideline Authority + β・Literature Evidence Strength + γ・Real-world Data + δ・Timeliness, where α = 0.4, β = 0.3, γ = 0.2, and δ = 0.1. This completes the systematic integration and structured association of multi-source knowledge. Based on this, a follow-up question knowledge graph is built. A targeted follow-up question database is constructed based on key nodes in the medical knowledge graph, establishing a mapping relationship between questions and medical entities. Simultaneously, a disease progression path prediction dimension is incorporated to construct a temporal relationship model of patient disease evolution, providing support for prospective intervention. Temporal relationship modeling introduces the time dimension into the knowledge graph, constructing a "disease progression map" and a "therapeutic efficacy evolution chain," which records in detail the disease progression trajectory, drug efficacy metabolism curve, and the time window of side effects. For example, it marks the temporal characteristics of "oxycodone extended-release tablets" reaching peak blood drug concentration 1 hour after administration and being metabolized completely 12 hours later, providing temporal data support for treatment plan optimization and risk warning.

[0059] In this embodiment, during the patient multimodal data acquisition and preprocessing process, in-hospital data is automatically captured through the HIS / LIS / PACS system interface, including structured and unstructured data such as patient pathology data, treatment data, test results, and imaging reports. Out-of-hospital data is collected by patients via mobile apps / mini-programs for pain assessment, psychological status assessment, medication feedback, and quality of life self-assessment. Wearable devices automatically collect real-time data such as patient heart rate, sleep status, activity level, and weather / barometric pressure. For elderly or less educated patients, data is collected via intelligent voice robot telephone follow-ups or human telephone follow-ups, with medical staff or family members assisting in data entry into the system. Subsequently, ETL technology is used to clean, standardize, and de-identify the multi-source heterogeneous data, removing redundant data and outliers. Unstructured text data is converted into structured data using a Natural Language Processing (NLP) engine to ensure data quality meets subsequent analysis requirements.

[0060] It should be noted that, based on preprocessed multimodal data, a digital twin profile of the patient is constructed, encompassing six dimensions: pathology, pain, treatment, psychological, social, and dynamic characteristics. Stored in JSON format, the profile includes the patient's unique identifier, pathology-related information such as tumor type, stage, and metastatic sites; pain-related characteristics such as pain type, intensity, and attack patterns; treatment-related records such as medication use, surgical history, and chemotherapy regimens; psychological data such as anxiety scores, depression scores, and sleep quality; social data such as family support levels, economic status, and adherence; and dynamic characteristics such as changing pain trends, drug responses, and follow-up adherence rates. The system supports dynamic profile updates, continuously updating the patient's dynamic characteristic dimensions through real-time collected follow-up data, wearable device data, and intervention effect feedback, ensuring the profile accurately reflects the patient's "current" state.

[0061] As a preferred implementation method, the risk stratification assessment process uses a multi-factor risk scoring algorithm to calculate the patient's risk score. The scoring formula involves factors such as pain score, pain trend slope, frequency of breakthrough pain, psychological score, follow-up compliance rate, and disease progression. Based on the scoring results, patients are divided into three levels: low risk, medium risk, and high risk. Low risk corresponds to patients with stable conditions and good pain control, medium risk corresponds to patients with fluctuating pain and mild side effects, and high risk corresponds to patients with frequent breakthrough pain, poor compliance, or severe psychological distress. Based on this, a personalized follow-up strategy is generated. The follow-up frequency is dynamically adjusted according to the risk level. Low-risk patients are followed up proactively every 2 weeks, focusing on quality of life. Medium-risk patients are followed up proactively every week, focusing on the effect of medication adjustment. High-risk patients are followed up every 3 days or check in daily via the APP, closely monitoring breakthrough pain and drug side effects. The follow-up channels are selected based on decision trees. Elderly / low-education patients are given priority for human telephone calls, while young and middle-aged / high-compliance patients are given priority for APP push notifications and WeChat reminders. Critical values, such as a pain score greater than 8, trigger a double alarm via SMS and telephone. The questionnaire content is generated dynamically based on patient profile characteristics using knowledge graph reasoning. If the patient is using strong opioids, constipation assessment and nausea and vomiting assessment questions are automatically added. If the profile shows neuropathic pain, numbness / burning sensation assessment questions are automatically added. If the patient has psychological high-risk factors, psychological status tracking questions are added.

[0062] Specifically, the pain prediction and graded early warning system relies on a breakthrough pain prediction model built using a Long Short-Term Memory (LSTM) network. The input layer includes hourly pain scores, medication times, activity levels, and weather / barometric pressure data from the past 7 days. The hidden layer is a two-layer LSTM structure used to capture the time-dependent and periodic characteristics of pain. The output layer uses Softmax classification to output the probability of breakthrough pain occurring within the next 24 / 48 / 72 hours, aiming for a prediction accuracy exceeding 75%. The model is fine-tuned and optimized using historical patient data to improve prediction accuracy. A multi-level early warning mechanism triggers different levels of warnings based on prediction results and monitoring data. A yellow warning corresponds to a slow increase in pain, and the system automatically pushes reminders to medical staff, suggesting shorter follow-up intervals and increased monitoring. An orange warning indicates a predicted probability of breakthrough pain exceeding 80% within the next 24 hours; the system pushes preventative suggestions to the patient's app and sends an "orange warning" notification to the case manager. A red warning triggers an emergency alarm immediately for detected severe side effects such as a respiratory rate less than 8 breaths / min or a pain crisis, notifying medical staff to intervene and simultaneously sending emergency treatment guidance to family members.

[0063] Specifically, multimodal AI analysis and intelligent grading processing integrate text sentiment analysis, voice feature analysis, and behavioral pattern analysis to determine the patient's emotional state and establish a patient personality profile. Text sentiment analysis identifies keywords and emotional tendencies in patient feedback; voice feature analysis identifies the patient's degree of weakness and emotional state through voiceprint features; and behavioral pattern analysis combines historical data to identify abnormal behaviors. Intelligent grading response is divided into three categories based on the patient's condition: ordinary, moderate, and critical. In the ordinary category, the patient has only mild discomfort and no obvious risk factors; AI automatically provides nursing suggestions and medication guidance without human intervention. In the moderate category, the patient exhibits moderate symptoms; AI guides the patient to take corresponding measures and records feedback. In the critical category, the patient exhibits severe symptoms; the system automatically generates an emergency work order, requiring nurses to respond within 30 minutes, conduct manual intervention, and record the results.

[0064] It should be noted that in the multidisciplinary collaborative diagnosis and treatment process, when the system detects that a patient meets the collaborative triggering conditions, it automatically initiates the multidisciplinary team (MDT) process. If the patient's psychological score continues to deteriorate, such as PHQ-9 being greater than 15, a consultation recommendation from the Department of Psychiatry is automatically triggered, generating a consultation request form containing a summary of the patient's key profile and pushing it to the psychiatrist's workstation. If the patient has bone metastasis pain that is not well controlled by medication, a consultation recommendation from the Department of Radiation Oncology is triggered, and a consultation request is pushed to the Department of Radiation Oncology. If the patient develops complex complications or refractory pain, consultations from multiple departments, such as oncology, pain management, and psychology, are triggered simultaneously, achieving cross-departmental data sharing and collaborative decision-making, and avoiding fragmented treatment plans.

[0065] Prioritizing closed-loop optimization and knowledge graph updates, follow-up results, intervention effects, and user feedback data are collected. After cleaning and processing, this data is used as real-world data to retrain pain prediction models, risk stratification algorithms, and emotion recognition models, continuously optimizing algorithm parameters and improving system decision-making accuracy. If a patient's response to a specific drug combination is found to be inconsistent with existing knowledge in the knowledge graph, the system automatically marks it and requests expert review. Upon approval, new knowledge nodes are generated, expanding the medical knowledge graph and the follow-up question knowledge graph. Simultaneously, the general knowledge graph is optimized weekly based on full data in batches, automatically eliminating invalid follow-up questions by passing a 90% consistency threshold to ensure the timeliness and effectiveness of the knowledge graph.

[0066] Specifically, for advanced cancer patients choosing home-based palliative care, a specialized management process is implemented. A personalized profile is created based on the patient's condition and needs, and a nursing plan focusing on analgesia and symptomatic treatment is developed. Patients wear monitoring wristbands, and family members install a care app on their phones. A passive monitoring strategy combined with daily symptom reports completed by family members is employed, minimizing the need for frequent patient intervention. When the wristband detects abnormal data, such as a heart rate consistently exceeding 110 beats per minute or frequent body movement, the system determines that agitation or pain crisis may be occurring and issues a red alert. Upon receiving a pop-up message, the on-duty nurse immediately contacts the family and guides them through emergency procedures, avoiding unnecessary emergency room visits. The system continuously assesses the patient's KPS score. When the predicted survival is less than two weeks, it automatically sends content on end-of-life symptom identification and care, as well as grief counseling, to the family, providing comprehensive support for both the patient and their family.

[0067] It should be noted that the system's daily maintenance establishes a 24 / 7 system monitoring mechanism to promptly detect and handle issues such as server failures and data transmission anomalies. Regular data backups are performed, and a comprehensive disaster recovery plan is employed to ensure medical data security. Algorithm optimization is based on the RLHF (Human Feedback Reinforcement Learning) mechanism, collecting feedback from medical staff on AI decision-making suggestions to continuously optimize model parameters. Algorithm performance evaluations are conducted quarterly, and model structure or training strategies are adjusted based on the evaluation results. Knowledge updates track the latest clinical guidelines, research literature, and expert experience, regularly updating the medical knowledge graph. New disease-drug-symptom association patterns discovered in clinical practice are incorporated to expand the knowledge graph nodes and relationships, improving the system's knowledge coverage. Simultaneously, system operation training is provided to medical staff and case managers, including patient profile interpretation, follow-up strategy adjustment, early warning handling, and multidisciplinary collaborative processes. Patients and their families are provided with APP / mini-program usage guidance to improve follow-up adherence.

[0068] Through the above implementation process, this invention achieves intelligent management of cancer pain patients throughout the entire lifecycle, from in-hospital treatment to outpatient follow-up, from pain control to psychological care, and from passive intervention to proactive prevention. It effectively integrates multi-source medical knowledge and data, improves pain control, optimizes follow-up strategies, breaks down knowledge and data silos, achieves accurate prediction and early warning, enhances communication efficiency and patient compliance, promotes multidisciplinary collaboration, and empowers home-based palliative care. Simultaneously, it enables system self-evolution and continuous optimization, providing clinical practice with precise, efficient, and personalized cancer pain management solutions.

[0069] Working principle: This system is based on medical knowledge graphs and follow-up question knowledge graphs, combined with multimodal AI analysis and intelligent decision-making technology, to achieve intelligent management of cancer pain patients throughout the entire life cycle. The core logic is as follows:

[0070] Data fusion and graph construction: Data from hospital HIS / LIS / PACS systems, patient-side APP / wearable device monitoring data, and follow-up feedback data are accessed, cleaned and standardized, and then identified as entities such as diseases, drugs, and symptoms through the BiLSTM-CRF model. Multiple relationships are extracted and time-series features are incorporated to construct a dynamically updated dual knowledge graph that integrates fragmented medical knowledge with the needs of follow-up scenarios.

[0071] Multidimensional profile generation: Based on fused data, construct intelligent patient profiles covering pathological, pain, treatment, psychological, social and dynamic characteristics, and synchronize dynamic information such as changes in patient condition and drug response in real time to accurately depict the individual patient status;

[0072] Risk stratification and strategy generation: Through a multi-factor risk scoring algorithm, combined with profiling and knowledge graph reasoning, patients are classified into low, medium and high risk levels, and personalized follow-up frequency, channels and questionnaire content are dynamically matched to achieve differentiated follow-up management.

[0073] Predictive Warning and Intelligent Response: Utilize LSTM models to analyze time-series data, predict the probability of pain outbreaks in the next 24-72 hours, and trigger yellow, orange, and red three-level warnings; integrate text, voice, and behavioral multimodal data for sentiment recognition and risk assessment, and implement graded responses according to ordinary, medium, and important situations;

[0074] Multidisciplinary collaboration and closed-loop optimization: When the triggering conditions are met, the multidisciplinary consultation process is automatically started to achieve cross-departmental data sharing; follow-up feedback and real-world data are fed back to the system for model iterative training and incremental updates of the knowledge graph, continuously improving the system's accuracy and applicability.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0076] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art, inspired by this description, design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the scope of protection of this invention.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for intelligent profiling and precise follow-up of cancer pain patients based on dynamic knowledge graphs, characterized by: Includes the following steps: Step 1: Construct a dual knowledge graph architecture that includes a medical knowledge graph and a follow-up question knowledge graph. The medical knowledge graph covers entities and relationships related to cancer pain, such as diseases, drugs, symptoms, and treatment plans. The follow-up question knowledge graph establishes a mapping relationship between questions and medical entities. Step 2: Integrate multi-source heterogeneous data, including HIS / LIS system data, wearable device monitoring data, and follow-up feedback data, and complete the multi-source data fusion after cleaning, standardization, and desensitization. Step 3: Construct a smart patient profile based on fused data, encompassing six dimensions: pathology, pain, treatment, psychology, social characteristics, and dynamic features. The profile is stored in JSON format and supports real-time dynamic updates. Step 4: Use patient intelligent profiles and dual knowledge graphs for matching and reasoning, and use a multi-factor risk scoring algorithm to assess the patient's pain risk level; Step 5: Based on the risk level and profile characteristics, dynamically generate a precise follow-up strategy that includes follow-up frequency, follow-up channels, and personalized questionnaire content. The personalized questionnaire is dynamically generated based on a knowledge graph of follow-up questions. Step 6: Using the LSTM (Long Short-Term Memory) network model, input time-series data such as pain scores, medication times, and activity levels from the past 7 days to predict the probability of outbreak pain occurring within the next 24-72 hours and trigger yellow, orange, and red level warnings. Step 7: Integrate text, voice, and behavioral multimodal data for sentiment analysis and risk identification to achieve intelligent hierarchical processing of ordinary, intermediate, and important responses; Step 8: When a patient's psychological score exceeds the threshold or pain control is poor, a multidisciplinary consultation recommendation form is automatically generated and pushed to the relevant department's terminal. Step 9: Collect follow-up feedback data and intervention effects to retrain the prediction model, and simultaneously perform incremental updates and optimizations on the dual knowledge graphs to achieve closed-loop iteration; Step 10: In home-based palliative care scenarios, combine wearable devices with non-invasive, continuous monitoring data to provide remote nursing guidance and end-of-life care support.

2. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 1, characterized in that: In step one, when constructing the dynamic medical knowledge graph, the BiLSTM-CRF model is used to identify 200+ disease entities, 500+ drug entities, 300+ symptom entities, and 1000+ treatment plan entities from multi-source data such as clinical guidelines, drug instructions, and medical records. Complex relationships such as treatment, causality, contraindications, and time sequence are extracted, and the fusion confidence is calculated through a weighted fusion algorithm. The formula is: Fusion Confidence = α・Guideline Authority + β・Literature Evidence Strength + γ・Real-world Data + δ・Timeliness, where α=0.4, β=0.3, γ=0.2, and δ=0.

1.

3. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 2, characterized in that: The multifactor risk scoring algorithm in step four is calculated as follows: S = w1・NRS + w2・ΔTrend + w3・Freq_BTcP + w4・Score_Psy + w5・(100-Compliance) + w6・Progression, where NRS is the pain score, ΔTrend is the slope of the pain trend, Freq_BTcP is the frequency of breakthrough pain, Score_Psy is the psychological score, Compliance is the follow-up compliance rate, and Progression is the disease progression. Based on the score S, patients are divided into three levels: 0-30 points (low risk), 31-60 points (medium risk), and 61-100 points (high risk).

4. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 3, characterized in that: The specific implementation method of the dynamic follow-up strategy in step five is as follows: low-risk patients are actively followed up once every 2 weeks, focusing on quality of life; medium-risk patients are actively followed up once a week, focusing on the effect of medication adjustment; high-risk patients are followed up once every 3 days or check in daily via the APP, closely monitoring breakthrough pain and drug side effects; the follow-up channels are selected based on decision tree, with priority given to human telephone for elderly or less educated patients, and priority given to APP push and WeChat reminders for middle-aged and young or highly compliant patients, and critical values ​​such as NRS>8 triggering dual alarms via SMS and telephone.

5. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 4, characterized in that: The multimodal sentiment analysis in step seven uses the PyTorch+Transformers framework to judge the patient's emotional state through text sentiment recognition, speech feature extraction, and behavioral pattern analysis. It generates personalized response scripts by combining empathic interaction strategies, and automatically filters out invalid follow-up questions based on a 90% consistency threshold.

6. A smart profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs, characterized in that, include: The data layer module is used to access data from HIS / LIS / PACS systems, wearable devices, and follow-up data, and to perform data cleaning, standardization, and desensitization to meet the requirements for medical data security protection. The dual knowledge graph construction module uses Neo4j database for storage and constructs medical knowledge graph and follow-up question knowledge graph through self-developed graph construction algorithm. It supports real-time individual graph updates and batch general graph optimization. The patient profile building module generates a digital patient profile containing six dimensions, which is stored in a PostgreSQL database and linked to historical data and profile tags. The follow-up strategy generation module dynamically outputs follow-up frequency, channels, and personalized questionnaire plans based on a risk scoring algorithm, and the follow-up records are stored in a MongoDB database. The prediction and early warning module uses an LSTM neural network model to predict outbreak pain and is equipped with a three-level early warning mechanism. The multimodal AI analysis module integrates text, voice, and behavioral data to perform sentiment analysis and risk classification. The multidisciplinary collaboration module automatically initiates the MDT process and pushes a consultation recommendation form when the triggering conditions are met. The closed-loop optimization module feeds back the follow-up results to the system, driving model iteration and dual knowledge graph evolution; The interaction layer module provides multiple interfaces, including a patient-side mini-program, a doctor-side web workbench, a case manager-side management backend, and a nurse-side workbench.

7. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 6, characterized in that: The medical knowledge graph in the dual knowledge graph construction module covers entity relationships such as cancer types, pain mechanisms, treatment plans, drug effects, and complications. The follow-up question knowledge graph is based on key nodes of the medical knowledge graph to build a targeted follow-up question database, establishes a mapping relationship between questions and medical entities, and incorporates the disease development path prediction dimension to build a temporal relationship model.

8. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 7, characterized in that: The LSTM model of the prediction and early warning module includes an input layer, two hidden layers and an output layer. The input features include at least the pain score sequence of the past 7 days, medication time points, activity level and weather and air pressure data. The output is the probability distribution of outbreak pain in the next 24 to 72 hours, with a target prediction accuracy of >75%.

9. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 8, characterized in that: The multidisciplinary collaboration module supports cross-departmental collaboration among oncology, pain management, and psychology departments. When a patient's PHQ-9 is detected to be greater than 15 or when bone metastasis pain is poorly controlled by medication, a consultation request form containing a summary of the patient's key profile is automatically generated and pushed to the relevant department's doctor workstation.

10. The intelligent profiling and precise follow-up system for cancer pain patients based on dynamic knowledge graphs according to claim 9, characterized in that: The patient-side mini-program of the interaction layer module supports pain score reporting, medication records, follow-up feedback and patient education interaction functions. The nurse workbench has functions for risk classification display, emergency treatment triggering and manual intervention recording. The management backend can realize data analysis, system configuration and effect monitoring. Data from all terminals is synchronized in real time and supports traceability query.

Images