A whole-course intelligent management platform for tumor patients

Abstract

The application provides a tumor patient whole course intelligent management platform, integrates multidimensional data of patients in and out of hospital, forms a six-dimensional hyper-dimensional file by aligning the course time axis, and provides a complete and continuous data basis for accurate decision-making. The constructed tumor specialty knowledge graph combines with a personalized graph reasoning algorithm, can dynamically match the optimal evidence-based knowledge according to the real-time characteristics of the patient, and fundamentally solves the problem of intervention content homogeneity. Combined with the rigid specification execution of the Drools rule engine and the forward-looking risk prediction of the XGBoost model, the compliance of diagnosis and treatment is guaranteed, and the pre-position intervention of adverse reactions is realized, which significantly improves the management effect. The whole course closed-loop management connects the whole process before diagnosis, during diagnosis and after diagnosis, realizes task automatic arrangement, distribution and closed-loop tracking through an event-driven task scheduling system, reaches different patients through multiple channels, and greatly improves the patient compliance and the medical management efficiency.

Description

Technical Field

[0001] This invention relates to the field of medical management platform technology, and in particular to an intelligent management platform for the entire course of cancer patients. Background Technology

[0002] Currently, the oncology clinical big data management system widely used in domestic medical institutions is the mainstream existing technology in the field of oncology patient management. The typical architecture of this type of system consists of four layers: storage layer, data management layer, basic application layer, and algorithm library. Its core functions focus on the collection, storage, statistical analysis, and visualization of in-hospital diagnosis and treatment data. Some advanced systems also incorporate basic follow-up reminders and indicator threshold early warning capabilities. Taking the oncology-specific disease data management platform used by a provincial tertiary hospital as an example, it connects to three core business systems—HIS, EMR, and LIS—through ETL tools to achieve structured storage of patient basic information, medical orders, medical records, and test results. It provides functions such as patient retrieval, data statistical reports, and medical record quality control, but only supports manually configured fixed-time follow-up SMS pushes.

[0003] Existing technologies of this kind have four core technical limitations in clinical applications: First, fragmented data integration. The data scope is limited to structured in-hospital diagnostic and treatment data, excluding out-of-hospital patient-reported symptoms, wearable device physiological data, lifestyle data, and psychological assessment data. Furthermore, multimodal data alignment along the disease progression timeline is not achieved, failing to form a continuous, multidimensional patient-centered profile, resulting in an incomplete data foundation for decision support. Second, insufficient depth of knowledge application. Existing systems' knowledge bases are mostly static documents or two-dimensional rule tables, lacking a structured oncology specialty knowledge graph, hindering complex semantic reasoning. Intervention content is generally generic templates, highly homogenized, and difficult to match with different patients' pathological classifications, treatment plans, comorbidities, and individual characteristics. Third, a single and outdated decision-making mechanism. Most systems only use rule-based warnings with fixed thresholds, failing to achieve collaborative decision-making between rule engines and machine learning models. Alarms are only triggered after abnormal indicators appear, unable to proactively predict adverse reaction risks, resulting in a delayed intervention window and an inability to dynamically adapt to the patient's disease progression. Fourth, a broken closed-loop throughout the entire disease progression. Data and processes are fragmented across pre-diagnosis screening, in-diagnosis treatment, and post-diagnosis rehabilitation follow-up. There is a lack of a unified event-driven task scheduling system, limited channels for reaching patients, poor patient compliance management, and a heavy workload for medical staff in manual management, resulting in low management efficiency. Summary of the Invention

[0004] To address the technical problems existing in the prior art, the present invention provides the following technical solution: A smart management platform for the entire course of cancer patients, comprising: The multi-source heterogeneous data fusion and hyperdimensional archive construction subsystem is used to connect to multi-source heterogeneous medical data and construct a time-series full-course hyperdimensional archive based on the patient after standardization and data governance. The oncology specialty dynamic knowledge graph subsystem is used to construct an oncology specialty knowledge graph based on medical ontology, supporting personalized knowledge reasoning based on patient characteristics. The dual-engine intelligent decision-making subsystem adopts a collaborative architecture of rule engine and machine learning model to monitor patient data in real time and output graded early warnings and task-triggered decisions; The personalized full-course intervention generation subsystem is used to generate personalized rehabilitation education, sports nutrition and psychological support intervention content by combining the patient's disease stage, individual characteristics and knowledge reasoning results. The event-driven task scheduling and distribution subsystem is used to intervene in the orchestration, scheduling, multi-channel distribution and full lifecycle closed-loop tracking of tasks; The multi-channel interactive terminal subsystem provides data interaction and content access points for different user roles, including patients, medical staff, and administrators.

[0005] As an optional implementation of this application, the multi-source heterogeneous data fusion and hyperdimensional archive construction subsystem includes: The data access module supports two modes: ETL batch synchronization and API real-time access. It connects to in-hospital data sources such as HIS, EMR, image archiving, laboratory systems, and follow-up systems, as well as out-of-hospital patient self-reporting and wearable device data sources. The data standardization module performs terminology mapping and data format conversion based on the HL7 FHIR standard and the UMLS Unified Medical Language System; The data governance module achieves data consistency and accuracy governance through the construction of the patient master index and data quality verification rules; The hyperdimensional profile construction module uses the patient's unique ID as the primary key and aligns data from six dimensions—demographics, clinical diagnosis and treatment, laboratory imaging, treatment response, lifestyle behavior, and psychological state—on a timeline to construct a full-time hyperdimensional profile.

[0006] As an optional implementation of this application, the oncology specialty dynamic knowledge graph subsystem is based on the UMLS semantic network as its ontology, and includes eight categories of entity nodes: oncology diseases, treatment plans, drugs, adverse reactions, test indicators, rehabilitation intervention, nutritional support, and psychological intervention, as well as sixteen categories of semantic relationship edges: indications, contraindications, induced adverse reactions, efficacy correlation, disease stage adaptation, and side effect relief. The knowledge graph is stored in the Neo4j graph database, and each relation edge is associated with a weight based on the level of evidence-based proof. The subsystem employs an improved Personalized PageRank algorithm to achieve personalized knowledge ranking. It uses entities that match the patient's characteristics as initial preference nodes and outputs the ranking result of knowledge entities that best fit the patient.

[0007] As an optional implementation of this application, the improved Personalized PageRank algorithm formula is as follows:

[0008] in: For target entity node Personalized sorting score, The damping coefficient is... Pointing to a node The set of predecessor nodes, For nodes To the node The edge weights are determined by the level of evidence-based knowledge. For nodes The sum of out-degree weights, For patient-specific preference vectors.

[0009] As an optional implementation of this application, the dual-engine intelligent decision-making subsystem includes: The Drools rule engine module is based on the Rete-OO algorithm and has four built-in rule bases for follow-up, abnormal indicator warning, adverse reaction management, and diagnosis and treatment quality control. It performs pattern matching on real-time incoming patient data and outputs deterministic decision instructions. The machine learning risk prediction module constructs a risk prediction model for severe adverse reactions to chemotherapy based on the XGBoost algorithm, outputs the patient's adverse reaction risk level, and outputs predictive pre-intervention decision instructions.

[0010] As an optional implementation of this application, the objective function of the XGBoost model is:

[0011] in The logarithmic loss function measures the deviation between the predicted value and the true label. For the first The regularization term of the decision tree includes regularization constraints on the number of leaf nodes and the weights of the leaf nodes, which are used to control model complexity and prevent overfitting.

[0012] As an optional implementation of this application, the personalized full-course intervention generation subsystem uses a three-level matching mechanism to generate intervention content: The first level is stage framework matching, which automatically identifies the patient's pre-diagnosis screening period, neoadjuvant therapy period, perioperative period, adjuvant radiotherapy and chemotherapy period, follow-up observation period, and palliative treatment period, and matches the corresponding intervention content framework. The second level is feature matching, which uses the cosine similarity algorithm to calculate the matching degree between the patient's feature vector and the content item, and filters out content items with a fit degree higher than the threshold. The third level is knowledge reasoning supplementation, which combines the reasoning results of the knowledge graph to supplement taboo hints and precautions, and finally generates a complete personalized intervention plan.

[0013] As an optional implementation of this application, the event-driven task scheduling and distribution subsystem is built on an event-driven architecture and RabbitMQ message queue, and includes an event bus, task orchestrator, timer scheduler, channel distributor, and task tracker. The event bus is used for the routing, distribution, and priority scheduling of event messages; The task orchestrator is used for orchestrating dependencies in complex multi-step tasks; The timer scheduler is used for the timed triggering of periodic tasks; The channel distributor is used to select the optimal push channel based on the task type and patient preferences; The task tracker is used to record the task execution status throughout the entire process and automatically upgrade and push channels for unresponsive tasks.

[0014] As an optional implementation of this application, the multi-channel interactive terminal subsystem is divided into: The patient-facing platform supports three formats: WeChat official account, mini-program, and APP, providing functions such as health record inquiry, reminder reception, random questionnaire filling, and adverse reaction reporting. The healthcare client includes PC workstations and mobile devices, providing functions such as viewing the patient's entire medical record, handling early warnings, reviewing intervention plans, and managing batch tasks; The management interface provides functions such as system configuration, rule base management, knowledge graph maintenance, content library management, data statistics, and access control.

[0015] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: First, comprehensive data fusion is used to construct a time-series hyperdimensional profile. This breaks through the limitations of existing systems that only integrate structured data from within the hospital, and incorporates multi-dimensional data from patients outside the hospital, including self-reported data, wearable devices, lifestyle behaviors, and psychological data. This data is aligned with the timeline of the disease course to form a six-dimensional hyperdimensional profile, providing a complete and continuous data foundation for accurate decision-making.

[0016] Second, a dynamic knowledge graph enables deep evidence-based reasoning. Unlike the simple query mode of existing static knowledge bases, the oncology-specific knowledge graph constructed in this solution, combined with personalized graph reasoning algorithms, can dynamically match the optimal evidence-based knowledge based on the patient's real-time characteristics, fundamentally solving the problem of homogenized intervention content.

[0017] Third, the dual-engine collaborative decision-making approach balances certainty and predictability. Existing systems mostly use a single fixed threshold for early warning. This solution combines the rigid enforcement of rules by the Drools rule engine with the forward-looking risk prediction of the XGBoost model, ensuring both compliance with treatment guidelines and enabling early intervention for adverse reactions, thus significantly improving management effectiveness.

[0018] Fourth, closed-loop management throughout the entire disease process, with precise outreach through multiple channels. The system connects the pre-diagnosis, diagnosis, and post-diagnosis stages, using an event-driven task scheduling system to automatically arrange, distribute, and track tasks in a closed loop. This multi-channel approach caters to the needs of different patients, significantly improving patient compliance and the efficiency of medical and nursing management.

[0019] Fifth, a decoupled and scalable architecture supports continuous iteration. Adopting a layered and decoupled cloud-native architecture, rules, knowledge, models, and content can all be independently iterated and updated without modifying core business logic, allowing for rapid adaptation to new treatment guidelines and disease expansions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the platform structure of the present invention; Figure 2 This is a schematic diagram of the composition structure of the dual-engine intelligent decision-making subsystem of the present invention; Figure 3 This is a diagram of the task scheduling and distribution mechanism of the present invention; Figure 4 This is a schematic diagram of a platform localization deployment architecture provided by an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0023] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0024] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0025] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0026] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0027] like Figure 1 As shown, the intelligent management platform for the entire course of cancer patients proposed in this invention includes the following six core subsystems: multi-source heterogeneous data fusion and hyperdimensional archive construction subsystem, oncology specialty dynamic knowledge graph subsystem, dual-engine intelligent decision-making subsystem, personalized full-course intervention generation subsystem, event-driven task scheduling and distribution subsystem, and multi-channel interactive terminal subsystem.

[0028] The overall architecture achieves decoupling and data communication between modules through a unified API gateway and message middleware, constructing a closed-loop management system encompassing "data integration - knowledge support - intelligent decision-making - intervention generation - feedback," covering the entire disease cycle of cancer patients before, during, and after diagnosis. Detailed descriptions of each subsystem are as follows: (I) Multi-source heterogeneous data fusion and hyperdimensional archive construction subsystem The multi-source heterogeneous data fusion and hyperdimensional archive construction subsystem serves as the platform's data foundation. It is responsible for connecting multiple business systems within the hospital with external data sources across all scenarios, completing the entire process of data extraction, cleaning, standardization, and fusion, and constructing a patient-centric, time-series, complete, hyperdimensional medical records to provide comprehensive, standardized, and high-quality data support for upper-level decision-making and applications.

[0029] The content of multi-source heterogeneous data fusion and hyperdimensional archive construction is as follows: 1. Multi-source data access A dual-mode access mechanism of "batch ETL + real-time API" is adopted: For existing business systems such as HIS, EMR, LIS, PACS, and in-hospital follow-up systems, a CDC (Change Data Capture) + incremental ETL process is used to achieve T+1 hour-level data synchronization; for data from external wearable devices, patient-reported data, and third-party follow-up platforms, RESTful API and MQTT protocol are used to achieve second-level real-time access. Before data access, identity authentication and data format verification are uniformly completed through the API gateway.

[0030] 2. Data Standardization Governance A standardized resource model is constructed based on the HL7 FHIR R4 international medical data exchange standard. Using the UMLS Unified Medical Language System as the terminology benchmark, it maps and converts disparate terms, codes, and formats across different systems, resolving the issues of "different names for the same disease" and "different codes for the same test." Simultaneously, a patient master index (EMPI) is built, using weighted matching of multiple fields such as name, ID number, mobile phone number, and medical insurance number to achieve unique patient identification across systems.

[0031] Data quality governance employs a dual mechanism of "rule verification + anomaly detection," which verifies integrity, consistency, and legality through rules and identifies outliers using the isolated forest algorithm, automatically triggering the data correction process.

[0032] 3. Construction of a comprehensive, multi-dimensional medical record throughout the entire disease course Using the patient's unique EMPI as the primary key, and integrating all data along the disease course timeline, a hyperdimensional profile was constructed encompassing six dimensions: ① Demographic dimension (age, gender, allergy history, family history, comorbidities); ② Clinical treatment dimension (diagnosis, surgery, chemotherapy, radiotherapy, medication orders); ③ Laboratory and imaging dimension (complete blood count, liver and kidney function, tumor markers, imaging report conclusions); ④ Treatment response dimension (efficacy assessment, adverse reaction grading, recurrence and metastasis); ⑤ Lifestyle and behavioral dimension (diet, exercise, sleep, smoking and drinking history); ⑥ Psychological state dimension (anxiety and depression scores, psychological intervention records).

[0033] Each dimension has multiple levels of time series labels, supporting backtracking and trend analysis by time node.

[0034] The system ultimately outputs a standardized, full-course, multi-dimensional medical record dataset, which is stored in a distributed relational database and object storage, providing a unified data service interface to upper-level subsystems.

[0035] (II) Dynamic Knowledge Graph Subsystem for Oncology This system primarily outputs a structured oncology specialty knowledge graph, personalized knowledge reasoning results, entity association paths, and sources of knowledge evidence. By constructing a structured, reasonable, and updatable knowledge system in the field of oncology, it provides evidence-based medical knowledge support for intelligent decision-making and personalized intervention, achieving an upgrade in capabilities from "static knowledge query" to "dynamic personalized reasoning."

[0036] 1. The knowledge system is constructed based on the UMLS semantic network ontology framework, referencing CSCO and NCCN oncology diagnosis and treatment guidelines, drug instructions, top evidence-based medicine literature, and clinical expert experience to expand the oncology-specific entity and relation system. Eight core entity nodes are defined: tumor disease entity, treatment plan entity, drug entity, adverse reaction entity, laboratory indicator entity, rehabilitation intervention entity, nutritional support entity, and psychological intervention entity. Sixteen core semantic relation edges are defined, including indications, contraindications, induced adverse reactions, positive correlation with efficacy, negative correlation with efficacy, disease stage adaptation, side effect relief, nutritional contraindications, exercise contraindications, and evidence level. Each relation edge is accompanied by a weight value and evidence level (1A to 2B).

[0037] 2. The graph storage and reasoning mechanism utilizes the Neo4j native graph database for knowledge storage and efficient querying, supporting millisecond-level relational queries across tens of thousands of nodes. Knowledge reasoning employs a hybrid approach of "graph path reasoning + personalized ranking": first, indirect relationships between entities are mined through graph path traversal, such as the relationship path "chemotherapy drugs → induce bone marrow suppression → white blood cell boosting drugs can alleviate it"; then, an improved Personalized PageRank algorithm is used to achieve personalized knowledge ranking for specific patients. An example of the algorithm is shown below:

[0038] In the formula: Target entity node The personalized ranking score indicates that the knowledge item is more suitable for the current patient; Damping coefficient, with a value of 0.85, is the classic optimal value for graph sorting algorithms, balancing the weights of global graph structure associations and patient-specific preferences, and taking into account both algorithm convergence speed and sorting accuracy. : Pointer to node The set of all predecessor nodes; :node To the node The edge weights, ranging from 0 to 1, are determined by the level of evidence-based knowledge. Level 1A evidence has a weight of 1, and the weight decreases by 0.15 for each level decrease. :node The sum of all out-degree weights; : Patient personalized preference vector, generated based on the patient's current disease stage, pathological type, allergy history, comorbidities, adverse reaction history, etc., corresponding to the initial preference value of the associated entity, and 0 if there is no association.

[0039] When the algorithm runs, it uses the set of entities that match the patient's features as the initial preference nodes, performs iterative random walks in the knowledge graph, and outputs a sorted list of knowledge entities with the highest fit after convergence, providing accurate knowledge basis for the generation of intervention plans.

[0040] 3. The dynamic update mechanism is set up with a dual mechanism of monthly batch update and real-time incremental update: every month, new knowledge is automatically extracted from the latest guidelines, FDA / NMPA drug approval information and core journal literature, and then added to the database in batches after entity alignment, conflict detection and expert review; clinical experts can add or modify knowledge entries in real time through the management terminal, and the inference cache is automatically refreshed after incremental update to ensure the timeliness of knowledge.

[0041] (III) Dual-engine intelligent decision-making subsystem The intelligent decision-making subsystem, as the core decision-making brain of the platform, adopts a dual-driven collaborative architecture of "rule engine + machine learning model" to balance the execution of deterministic diagnosis and treatment guidelines with forward-looking risk prediction, and realize real-time monitoring of patient data throughout the entire course of the disease, abnormal warnings, and automatic task triggering.

[0042] like Figure 2 As shown, the subsystem is divided into a Drools rule engine module and a machine learning risk prediction module, which run in parallel and merge the results: The Drools rule engine is responsible for strongly evidence-based, deterministic rule validation and task triggering; Machine learning models are responsible for probabilistic and predictive risk assessment and early intervention, ultimately outputting tiered decision-making results.

[0043] The specific working mechanisms of the two entities are as follows: 1. Drools Rule Engine Module The rule engine uses Drools 7.5 Enterprise Edition as its core, employing the Rete-OO algorithm for efficient pattern matching. Single rule matching latency is ≤10ms, supporting millisecond-level validation for tens of thousands of rules. The rule base is built upon tumor diagnosis and treatment guidelines and clinical pathways, using a structured definition of "IF condition set - THEN action set," and is divided into four main rule systems: Follow-up rules: Based on the patient's surgery, chemotherapy, radiotherapy time points and treatment plan, the system automatically calculates the next follow-up time, follow-up items and follow-up period. For example, "IF patient is a lung cancer patient who has completed the second cycle of adjuvant chemotherapy. THEN triggers a blood routine check-up reminder on the 7th day after chemotherapy and a comprehensive check-up reminder on the 14th day."

[0044] Abnormal indicator warning rules: Three-level warning thresholds are set for test indicators and vital signs, corresponding to different intervention intensities. For example, "IF absolute neutrophil count < 1.0×10^9 / L AND patient is in chemotherapy cycle THEN triggers level IV bone marrow suppression warning, pushes white blood cell boosting guidance and emergency medical treatment prompts."

[0045] Adverse reaction management rules: Match the adverse reaction symptoms and grades reported by patients, automatically associate them with the management plan in the knowledge graph, and trigger the corresponding intervention task.

[0046] Treatment quality control rules: Verify whether treatment behaviors comply with guidelines and standards, and trigger quality control reminders on the medical staff side.

[0047] The rules engine receives patient data change events pushed by the data platform in real time, inserts the patient fact object into the working memory, performs multi-rule pattern matching through the Rete network, and executes the corresponding rule action after priority sorting by the conflict resolution strategy, and outputs decision instructions.

[0048] 2. Machine Learning Risk Prediction Module A risk prediction model for severe adverse reactions to chemotherapy based on XGBoost was constructed to achieve prospective prediction of adverse reactions, shifting the intervention window from "after the occurrence of abnormalities" to "when the risk increases." The model's objective function is as follows:

[0049] Formula definition: Logarithmic loss function, used to measure predicted values. With real labels The deviation is that the true label is defined as "whether grade III or higher adverse reactions to chemotherapy occurred"; : No. The regularization term for the decision tree is specifically expressed as follows: ,in The number of leaf nodes. The weight of the leaf node. and The regularization coefficient was set to 0.1 and 0.5 based on calibration experiments. This value was determined through 5-fold cross-validation on a dataset of 1200 clinical chemotherapy cases, which effectively controls model complexity and avoids overfitting while ensuring prediction accuracy. The total number of decision trees is set to 100, and the maximum depth of a single tree is limited to 6.

[0050] The reasoning process is as follows: Feature engineering: Extract 32-dimensional input features from the hyperdimensional archive, covering 7 major categories: demographic characteristics, ECOG score, chemotherapy regimen and cycle, baseline test indicators, liver and kidney function, comorbidities, and history of adverse reactions; Model training: The model is trained based on labeled data of patients undergoing chemotherapy in the hospital in the past. The SMOTE algorithm is used to handle sample imbalance (after model validation, the final model test set AUC reaches above 0.85 before being put into application; if it does not meet the standard, the training set is optimized). Real-time reasoning: Before each chemotherapy session and when the test data is updated, the latest characteristic data is automatically extracted and input into the model to output the probability of severe adverse reactions, which is divided into three levels: low risk (<30%), medium risk (30%-70%), and high risk (>70%), corresponding to different intensities of pre-intervention programs.

[0051] (iv) Personalized full-course intervention generation subsystem Based on the patient's current disease stage, efficacy assessment, risk level, and individual characteristics, this system automatically generates personalized rehabilitation education, sports nutrition, and psychological support intervention content, combined with knowledge graph reasoning results, to achieve precise intervention tailored to each individual.

[0052] The system adopts a three-level matching generation mechanism of "stage framework + feature matching + knowledge supplementation": 1. Disease Stage Identification: By matching the patient's diagnosis and treatment events and time nodes through a rule engine, the system automatically identifies the patient's six core disease stages: pre-diagnosis screening, neoadjuvant therapy, perioperative period, adjuvant radiotherapy and chemotherapy, follow-up observation, and palliative care. Each stage corresponds to a standardized intervention content framework and core intervention goals.

[0053] 2. Feature Matching and Screening: Patient characteristics such as pathological classification, treatment plan, adverse reaction risk, dietary preferences, exercise capacity, and comorbidities are converted into feature vectors. The cosine similarity algorithm is used to calculate the matching degree between patient features and entries in the content library, and suitable entries are screened. The matching degree calculation formula is as follows:

[0054] in: For the patient's first The weight values ​​of the features, For the content entry number The fit value of the feature. The total number of feature dimensions is used, and the matching threshold is set to, for example, 0.7. Content items with a matching degree higher than the threshold are included in the candidate set. 3. Knowledge Reasoning Supplement: The personalized reasoning interface of the knowledge graph subsystem is called to supplement the contraindications, precautions, and adverse reaction management plans for this patient, and finally integrate them to generate a complete personalized intervention plan.

[0055] 4. Dynamic Iterative Adjustment: Based on patient feedback data, changes in efficacy, and the occurrence of adverse reactions, the focus and frequency of intervention content are dynamically adjusted to achieve continuous optimization of the intervention plan.

[0056] (v) Event-driven task scheduling and distribution subsystem like Figure 3 As shown, the task scheduling and distribution subsystem adopts an event-driven architecture (EDA) and uses RabbitMQ message queues to achieve asynchronous decoupling of tasks. Its core components include: Event Bus: Receives event messages generated by all subsystems, uses a publish-subscribe pattern for routing and distribution, supports event priority configuration, and prioritizes warning events.

[0057] Task orchestrator: Performs dependency orchestration for complex multi-step intervention processes. For example, a review reminder task needs to be executed according to the process of "first push 7 days in advance → second push 1 day in advance → push notices after review → trigger warning for abnormal review results", automatically handling the sequential dependencies between tasks.

[0058] Scheduler: Implements scheduled task management based on the Quartz framework, supports triggering by day, week, month and custom period, and supports calendar exclusion and time window configuration.

[0059] Channel Distributor: Selects the optimal push channel based on task type, urgency, and patient channel preference, including mobile APP, WeChat official account, SMS, and outbound calls from human agents. High-priority alert tasks are pushed in parallel through multiple channels.

[0060] Task tracker: Records the distribution status, reach status, and user reading / response status of tasks throughout the entire process. It automatically upgrades and pushes tasks that have not been responded to within 24 hours to other channels, and triggers manual intervention by medical staff when high-risk warnings are not responded to.

[0061] It receives trigger commands from the decision engine (such as task trigger commands, timed task configurations, and patient channel preference settings), completes task orchestration, scheduling, multi-channel distribution, and full lifecycle tracking, and realizes automated closed-loop management of intervention tasks, decoupling decision logic from reach logic.

[0062] (vi) Multi-channel interactive terminal subsystem It provides interactive entry points for three user roles: patients, medical staff, and system administrators, enabling functions such as data collection, content display, message delivery, manual intervention, and system configuration. It serves as the carrier for interaction between the platform and users.

[0063] 1. Patient side: Supports three forms: WeChat official account, mini program and APP. The core functions include querying the health record of the whole disease course, receiving follow-up reminders, filling out follow-up questionnaires, reporting adverse reactions with one click, learning rehabilitation content, and online medical consultation, which reduces the threshold for patients to use.

[0064] 2. Healthcare Client: Includes PC workstation and mobile healthcare client version. Core functions include viewing the entire patient medical record, processing early warning messages, reviewing and adjusting intervention plans, batch patient management, follow-up task processing, management data statistics, and integration into clinical workflows.

[0065] 3. Management side: Provides functions such as system parameter configuration, rule base management, knowledge graph maintenance, content library management, data statistical analysis, and user permission management, supporting flexible iteration and operation of the platform.

[0066] When patients generate new diagnostic data, test data, or self-reported data, the data platform automatically publishes data update events, triggering the decision-making process: the two engines run in parallel, the rule engine completes deterministic rule verification, and if a match is found, it directly triggers the corresponding task; the machine learning model simultaneously completes risk prediction, outputs the risk level, and triggers pre-intervention tasks; the intervention generation subsystem receives decision instructions, combines patient characteristics with knowledge graph reasoning results, and generates personalized intervention content; the task scheduling subsystem arranges and distributes tasks according to task priority and channel preference, and tracks the execution status.

[0067] Patient response data, behavioral data, and subsequent treatment results are fed back to the data platform, where the system automatically evaluates the intervention effect and continuously optimizes model parameters, rule configurations, and content libraries.

[0068] During system operation, data changes in the hospital's business systems are captured by the CDC and pushed to the data platform via a Kafka message queue. Patient events generated after data standardization updates are pushed to the dual-engine decision-making subsystem via the RabbitMQ event bus. Decision results are synchronized to the intervention generation subsystem and task scheduling subsystem, and finally distributed to the terminals. The end-to-end latency is ≤5 seconds, meeting the real-time requirements for abnormal early warning.

[0069] Non-real-time tasks such as historical data backtracking, image file synchronization, statistical report generation, batch model training, and batch knowledge graph updates can be executed in batches during the early morning hours of each day using ETL tools, thus avoiding the occupation of resources during business hours.

[0070] All synchronous service calls between subsystems uniformly adopt RESTful APIs, and unified authentication, rate limiting, and log auditing are performed through API gateways. The interface data format adopts JSON, and medical terminology uniformly follows the UMLS and HL7 FHIR standards to ensure consistency of cross-system interaction.

[0071] Example 1 like Figure 4 As shown, a localized deployment solution was adopted in the oncology department of a tertiary hospital, with the following configuration: Data middleware server: 2 servers (primary and backup for high availability), deployed with MySQL 8.0 distributed relational database, MinIO object storage, and DataX ETL tool, to carry out data storage and data computing tasks.

[0072] Graph database server: 1 server, deployed with Neo4j 4.4 Enterprise Edition, to host the storage and inference computation of the oncology specialty knowledge graph.

[0073] Application and Decision Servers: 2 servers (load-balanced deployment), deploying Drools rule engine, XGBoost model inference service, business application service, and RabbitMQ message middleware.

[0074] On the terminal side: the medical staff terminal uses the hospital's existing PC workstations, which are compatible with browser access; the patient terminal adopts the form of WeChat official account + mini program, without requiring patients to install an additional APP.

[0075] Software environment: The server operating system is CentOS 7.9, and the application is deployed using Docker+Kubernetes containerization, supporting elastic scaling.

[0076] This section uses the complete management process of a patient with right lung adenocarcinoma undergoing adjuvant chemotherapy after surgery as an example to explain the system operation process in detail: 1. Basic patient information Ms. Zhang, female, 56 years old, ECOG score 1, postoperative pT2N0M0 stage IIA adenocarcinoma of the right lung, no history of drug allergy, complicated with type 2 diabetes, daily habit of taking walks, is undergoing adjuvant chemotherapy with pemetrexed + carboplatin for a total of 4 cycles, and is currently on the 3rd day after the 2nd cycle of chemotherapy.

[0077] 2. Data Update and Event Triggering: Three days after discharge from the hospital following the second cycle of chemotherapy, the patient underwent a follow-up blood routine test at the outpatient clinic. The LIS system returned the test results: white blood cell count 2.8 × 10^9 / L, absolute neutrophil count 1.2 × 10^9 / L. The data platform captured this test data in real time through the CDC, and after terminology standardization and quality verification, updated it to the patient's full-course medical record. Simultaneously, a "Patient Test Data Update" event was published to the event bus.

[0078] 3. Dual-engine collaborative decision-making (1) Rule Engine Validation: After receiving the event, the Drools rule engine loads the patient's current disease characteristics and test data, and matches the "chemotherapy-induced myelosuppression warning" rule set: Rule conditions: Within 7 days after chemotherapy AND absolute neutrophil count is between 1.0 and 2.0 × 10^9 / L AND the patient is undergoing adjuvant chemotherapy after lung cancer surgery. Rule Action: Trigger a Level II bone marrow suppression warning, generate dietary guidance for increasing white blood cell count, and remind users to have a follow-up blood test every other day. (2) Machine learning risk prediction: The patient's 32-dimensional features were simultaneously extracted and input into the XGBoost model. The probability of the patient developing severe myelosuppression was calculated to be 62%, which was judged as intermediate risk, triggering the pre-intervention task of "strengthening disease monitoring and increasing the intensity of nutritional intervention". After the results of the two engines were fused, a secondary warning instruction and a corresponding task list were output.

[0079] 4. Personalized intervention generation The intervention generation subsystem receives decision instructions, first identifying the patient as being in the "adjuvant chemotherapy period - bone marrow suppression risk sub-stage," and matching the corresponding intervention framework; then, considering the patient's characteristics of having diabetes, no seafood allergy, and regular walks, it matches content library entries using cosine similarity; finally, it invokes knowledge graph reasoning to supplement dietary restrictions for diabetes and precautions for bone marrow suppression, generating the final intervention plan: Rehabilitation education: Key points of bone marrow suppression nursing after chemotherapy, precautions for infection prevention, and emergency treatment procedures for fever; Nutritional recommendations: Low-sugar, high-protein diet to promote white blood cell count (lean meat, eggs, mushrooms), avoid high-sugar foods that promote white blood cell count, and daily meal recommendations; Exercise recommendations: Suspend outdoor walks and focus on light activities in bed. Avoid strenuous activity and crowded places. Follow-up reminder: Have a blood routine test the following day. If fever or worsening fatigue occurs, seek medical attention immediately. All content items have a match score of ≥0.75, meeting the push notification criteria.

[0080] 5. Task distribution and closed-loop tracking The task scheduling subsystem receives the intervention plan and, based on the patient's preference for WeChat, pushes the warning message and intervention content through the WeChat official account. Simultaneously, a scheduled task is set: a follow-up reminder is pushed again at 8:00 AM the following morning. The system tracks the task status in real time. If the patient clicks to read and confirm within 10 minutes, the task status updates to "reached and responded to," completing the closed loop.

[0081] The patient underwent a follow-up blood test the next day, and the absolute neutrophil count recovered to 1.8 × 10^9 / L, without progressing to severe myelosuppression. After the test data was fed back, the system automatically marked the intervention as effective and included the case in the model training sample library to continuously optimize the model's predictive accuracy.

[0082] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An intelligent management platform for the entire course of cancer patient disease, characterized in that, include: The multi-source heterogeneous data fusion and hyperdimensional archive construction subsystem is used to connect to multi-source heterogeneous medical data and construct a time-series full-course hyperdimensional archive based on the patient after standardization and data governance. The oncology specialty dynamic knowledge graph subsystem is used to construct an oncology specialty knowledge graph based on medical ontology, supporting personalized knowledge reasoning based on patient characteristics. The dual-engine intelligent decision-making subsystem adopts a collaborative architecture of rule engine and machine learning model to monitor patient data in real time and output graded early warnings and task-triggered decisions; The personalized full-course intervention generation subsystem is used to generate personalized rehabilitation education, sports nutrition and psychological support intervention content by combining the patient's disease stage, individual characteristics and knowledge reasoning results. The event-driven task scheduling and distribution subsystem is used to intervene in the orchestration, scheduling, multi-channel distribution and full lifecycle closed-loop tracking of tasks; The multi-channel interactive terminal subsystem provides data interaction and content access points for different user roles, including patients, medical staff, and administrators.

2. The intelligent management platform for the entire course of cancer patient treatment according to claim 1, characterized in that, The multi-source heterogeneous data fusion and hyperdimensional archive construction subsystem includes: The data access module supports two modes: ETL batch synchronization and API real-time access. It connects to in-hospital data sources such as HIS, EMR, image archiving, laboratory systems, and follow-up systems, as well as out-of-hospital patient self-reporting and wearable device data sources. The data standardization module performs terminology mapping and data format conversion based on the HL7 FHIR standard and the UMLS Unified Medical Language System; The data governance module achieves data consistency and accuracy governance through the construction of the patient master index and data quality verification rules; The hyperdimensional profile construction module uses the patient's unique ID as the primary key and aligns data from six dimensions—demographics, clinical diagnosis and treatment, laboratory imaging, treatment response, lifestyle behavior, and psychological state—on a timeline to construct a full-time hyperdimensional profile.

3. The intelligent management platform for the entire course of cancer patient treatment according to claim 1, characterized in that, The oncology specialty dynamic knowledge graph subsystem is based on the UMLS semantic network and includes eight categories of entity nodes: oncology diseases, treatment plans, drugs, adverse reactions, test indicators, rehabilitation intervention, nutritional support, and psychological intervention, as well as sixteen categories of semantic relationship edges: indications, contraindications, induced adverse reactions, efficacy correlation, disease stage adaptation, and side effect relief. The knowledge graph is stored in the Neo4j graph database, and each relation edge is associated with a weight based on the level of evidence-based evidence. The subsystem employs an improved Personalized PageRank algorithm to achieve personalized knowledge ranking. It uses entities that match the patient's characteristics as initial preference nodes and outputs the ranking result of knowledge entities that best fit the patient.

4. The intelligent management platform for the entire course of cancer patient treatment according to claim 3, characterized in that, The improved Personalized PageRank algorithm formula is as follows: in: For target entity node Personalized sorting score, The damping coefficient is... Pointing to a node The set of predecessor nodes, For nodes To the node The edge weights are determined by the level of evidence-based knowledge. For nodes The sum of out-degree weights, For patient-specific preference vectors.

5. The intelligent management platform for the entire course of cancer patient treatment according to claim 1, characterized in that, The dual-engine intelligent decision-making subsystem includes: The Drools rule engine module is based on the Rete-OO algorithm and has four built-in rule bases for follow-up, abnormal indicator warning, adverse reaction management, and diagnosis and treatment quality control. It performs pattern matching on real-time incoming patient data and outputs deterministic decision instructions. The machine learning risk prediction module constructs a risk prediction model for severe adverse reactions to chemotherapy based on the XGBoost algorithm, outputs the patient's adverse reaction risk level, and outputs predictive pre-intervention decision instructions.

6. The intelligent management platform for the entire course of cancer patient treatment according to claim 5, characterized in that, The objective function of the XGBoost model is: in The logarithmic loss function measures the deviation between the predicted value and the true label. For the first The regularization term of the decision tree includes regularization constraints on the number of leaf nodes and the weights of the leaf nodes, which are used to control model complexity and prevent overfitting.

7. The intelligent management platform for the entire course of cancer patients according to claim 1, characterized in that, The personalized full-course intervention generation subsystem uses a three-level matching mechanism to generate intervention content: The first level is stage framework matching, which automatically identifies the patient's pre-diagnosis screening period, neoadjuvant therapy period, perioperative period, adjuvant radiotherapy and chemotherapy period, follow-up observation period, and palliative treatment period, and matches the corresponding intervention content framework. The second level is feature matching, which uses the cosine similarity algorithm to calculate the matching degree between the patient's feature vector and the content item, and filters out content items with a fit degree higher than the threshold. The third level is knowledge reasoning supplementation, which combines the reasoning results of the knowledge graph to supplement taboos and precautions, and finally generates a complete personalized intervention plan.

8. The intelligent management platform for the entire course of cancer patient treatment according to claim 1, characterized in that, The event-driven task scheduling and distribution subsystem is built on an event-driven architecture and RabbitMQ message queue, and includes an event bus, task orchestrator, timer scheduler, channel distributor, and task tracker. The event bus is used for the routing, distribution, and priority scheduling of event messages; The task orchestrator is used for orchestrating dependencies in complex multi-step tasks; The timer scheduler is used for the timed triggering of periodic tasks; The channel distributor is used to select the optimal push channel based on the task type and patient preferences; The task tracker is used to record the task execution status throughout the entire process and automatically upgrade and push channels for unresponsive tasks.

9. The intelligent management platform for the entire course of cancer patient treatment according to claim 1, characterized in that, The multi-channel interactive terminal subsystem is divided into: The patient-facing platform supports three formats: WeChat official account, mini-program, and APP, providing functions such as health record inquiry, reminder reception, random questionnaire filling, and adverse reaction reporting. The healthcare client includes PC workstations and mobile devices, providing functions such as viewing the patient's entire medical record, handling early warnings, reviewing intervention plans, and managing batch tasks; The management interface provides functions such as system configuration, rule base management, knowledge graph maintenance, content library management, data statistics, and access control.

Images