Non-muscle invasive bladder cancer follow-up management and recurrence risk early warning system and method

Abstract

The invention discloses a non-muscle invasive bladder cancer follow-up management and recurrence risk early warning system and method, and belongs to the technical field of medical information technologies and data analysis systems. The system comprises a patient end module, a nurse end module, a doctor end module and a cloud platform intelligent background, and multi-end cooperation of patient symptom reporting, nursing follow-up visit execution and clinical data management is achieved. A cloud platform intelligent background fuses multi-source data, dynamically predicts the recurrence risk probability of a patient based on a machine learning model, and automatically generates a personalized follow-up visit plan according to clinical guidelines and individual features. The system supports risk visual display, structured task distribution and closed-loop early warning processing, effectively breaks data islands, and improves the accuracy, collaboration and patient participation degree of follow-up visit management, thereby optimizing medical resource configuration and improving prognosis.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology and data analysis system technology, specifically relating to a follow-up management and recurrence risk warning system and method for non-muscle-invasive bladder cancer. Background Technology

[0002] Bladder cancer is a common malignant tumor of the urinary system worldwide, with non-muscle-invasive bladder cancer (NMIBC) accounting for more than 70% of initial cases. Although the prognosis for NMIBC patients after transurethral resection of bladder tumor is relatively good, their high recurrence rate and certain risk of progression necessitate long-term, regular postoperative follow-up. Under the current medical model, healthcare professionals play a central role in this long-term follow-up process; however, traditional follow-up management methods face numerous systemic challenges, with significant bottlenecks in both efficiency and effectiveness. 1. Challenges faced by existing follow-up management models For patients: Current follow-up plans are usually explained verbally by doctors or provided with a paper checklist, which is complex and difficult to remember, making it easy for patients to miss key follow-up appointments and affecting the monitoring of treatment effectiveness. At the same time, patients lack a direct and continuous channel for understanding their condition's dynamics and the risk of relapse, easily leading to unnecessary anxiety. Furthermore, limited channels of doctor-patient communication mean that patients cannot obtain timely professional feedback when symptoms or questions arise, further exacerbating their psychological burden.

[0003] For physicians: Managing a large number of NMIBC postoperative patients presents challenges in manually and efficiently tracking the follow-up progress of each patient. Current risk assessments rely heavily on isolated clinical data from single visits, lacking the integration and trend analysis of long-term, dynamic, multi-dimensional patient data. This information lag and incompleteness can lead to delayed identification of high-risk patients, thus delaying necessary clinical interventions.

[0004] For nurses: Nursing follow-up work is often fragmented and lacks a structured process linked to medical plans. The health education content provided by nurses is mostly generic and fails to be personalized based on the patient's real-time risk level, symptom changes, and psychological state, resulting in limited educational effectiveness and failing to fully realize the professional value of nursing work.

[0005] 2. Limitations of existing technological systems Currently, some patient follow-up management systems or nursing information systems exist in clinical practice, but their functions are significantly insufficient: (1) Most systems are essentially simple reminder tools for review time or simple extensions of electronic medical record functions, focusing on information recording rather than proactive management.

[0006] (2) Data silo problem: These systems have failed to effectively achieve two-way, structured data flow among doctors, nurses, and patients. Clinical data, nursing assessment data, and patient-reported outcome data are isolated from each other and cannot be integrated and utilized.

[0007] (3) Lack of intelligent analysis and early warning capabilities: Existing systems generally lack proactive risk assessment models based on multi-source data fusion and do not have the function of dynamically predicting recurrence risk.

[0008] (4) Lack of process management: The best practices of clinical guidelines have not been solidified into standard, automatically scheduled follow-up pathways, nor has the automatic generation and task distribution of personalized follow-up plans based on risk assessment results been achieved.

[0009] In summary, existing technologies lack a comprehensive NMIBC patient disease follow-up management system that can deeply integrate clinical, nursing, and patient data to achieve dynamic risk warning, personalized follow-up plan management, and team collaborative workflow. Therefore, there is an urgent need to develop a more intelligent, integrated, and efficient solution to overcome these shortcomings and improve medical quality and patient outcomes. Summary of the Invention

[0010] [Technical Issues] The technical problem to be solved by this invention is: how to overcome the problem of data silos among doctors, nurses and patients in the existing bladder cancer follow-up system, and the resulting lack of dynamic risk warning capabilities and personalized collaborative follow-up management capabilities.

[0011] [Technical Solution] To address the above problems, this invention provides a system and method for follow-up management and recurrence risk warning of non-muscle-invasive bladder cancer.

[0012] In a first aspect, the present invention provides a follow-up management and recurrence risk early warning system for non-muscle-invasive bladder cancer, comprising: The system includes a patient-side module, a nurse-side module, a doctor-side module, and a cloud platform intelligent backend. The patient-side module is used for patients to report symptom data, receive health education content, view personalized follow-up plans and personal relapse risk information; The nurse-side module is used by nurses to manage patient follow-up queues, perform structured follow-ups, receive and process relapse warnings, and push health education materials to patients. The doctor-side module is used to allow doctors to view the patient queue, track individual patient risks, receive alerts, and intervene. The cloud platform's intelligent backend is used to store and integrate multi-source data from patients, nurses, and doctors. Based on the integrated data, it runs a machine learning prediction model to calculate the patient's recurrence risk probability. Based on the recurrence risk probability and patient data, it generates a personalized follow-up plan for the patient through a follow-up plan generation engine.

[0013] Optionally, the patient-side module includes: The symptom diary unit is used for patients to report symptom information and quality of life scores through a structured scale, as well as to report hematuria through a self-made questionnaire; The health education receiving unit is used to receive and display personalized health education materials pushed by nurses. The follow-up calendar unit is used to visually display the personalized follow-up plan and provide reminder functionality; The risk index display unit is used to display, in chart form, the degree and probability of individual relapse risk and its changing trend, calculated by the intelligent backend of the cloud platform.

[0014] Optionally, the nurse-side module includes: The patient queue management unit is used to display the list of patients managed by nurses in the form of cards. The cards contain patient risk tags and support filtering by task and warning status. The structured follow-up form unit contains standardized forms that include disease information, psychosocial assessments, and quality of life scores, which nurses can fill out during follow-ups and which automatically generate nursing records. The recurrence early warning monitoring unit is used to receive early warning information triggered by the intelligent backend of the cloud platform. The early warning is generated based on the patient's objective clinical data and subjective PROs data, and includes standardized processing suggestions. The health education toolkit unit has an embedded health education knowledge base, allowing nurses to select content and push it to target patients.

[0015] Optionally, the doctor-side module includes: The patient cohort overview unit is used to centrally display all key information about managed patients and provides sorting and filtering functions by risk level and follow-up time. The individual risk tracking unit is used to display the clinical timeline of a single patient, integrating historical change curves of examination results, reported symptoms, and risk indices. The early warning and intervention center unit is used to receive information on high-risk or highest-risk patients automatically marked by the system, as well as patients who have not been followed up within the stipulated time, and supports doctors to send messages to patients or adjust follow-up plans through the platform.

[0016] Optionally, the cloud platform's intelligent backend includes a unified patient data model for structured storage of patients' static and dynamic data; The static data includes demographic information and immutable clinical pathology data; The dynamic data includes data from previous laboratory tests, imaging reports, and symptom and quality of life scores from patients.

[0017] Optionally, the cloud platform's intelligent backend also includes an early warning rule engine; The early warning rule engine is used to automatically trigger early warning information based on the data in the patient unified data model and / or the recurrence risk probability output by the machine learning prediction model, according to preset clinical early warning rules.

[0018] Optionally, the early warning information triggered by the early warning rule engine is automatically upgraded and pushed from the recurrence early warning monitoring unit of the nurse module to the early warning and intervention center unit of the doctor module when the preset upgrade conditions are met.

[0019] Optionally, the follow-up plan generation engine includes: A clinical guidelines knowledge base for storing structured clinical follow-up rules; The comprehensive risk stratification unit is used to determine the patient's overall risk level based on the patient's clinical data and the recurrence risk probability output by the machine learning prediction model. A follow-up plan template library is used to store standardized follow-up templates corresponding to different comprehensive risk levels; The follow-up plan generation engine is configured to: call the corresponding follow-up template from the follow-up plan template library according to the comprehensive risk level determined by the comprehensive risk stratification unit, and adjust it in combination with the patient's individual characteristics to generate a personalized follow-up plan.

[0020] Optionally, the personalized parameter adjustment is based on preset clinical rules and dynamically adjusts the follow-up template according to the patient's individual attributes; the individual attributes include at least one of the following: real-time symptoms, age, treatment history, complications or comorbidities, and recurrence events.

[0021] Secondly, the present invention provides a method for follow-up management and recurrence risk warning of non-muscle-invasive bladder cancer, the method using the system described above, and the method comprising the following steps: S1: Receive symptom data and quality of life score data reported by patients through the patient-side module; S2: Receive follow-up records and nursing assessment data entered by nurses through the nurse module; receive or synchronize patients' clinical diagnosis and treatment data through the doctor module. S3: In the intelligent backend of the cloud platform, data from the patient, nurse and doctor ends are integrated, and machine learning prediction models are used to calculate the patient's recurrence risk probability. S4: In the intelligent backend of the cloud platform, a personalized follow-up plan is automatically generated based on clinical guidelines and rules, the recurrence risk probability, and individual patient characteristics. S5: Synchronize the personalized follow-up plan to the patient, nurse, and doctor terminals; based on the follow-up plan and real-time data, trigger follow-up tasks and early warning processing on the nurse terminal, and trigger intervention tasks on the doctor terminal.

[0022] [Beneficial Effects] 1. Breaking down data silos and achieving efficient collaboration: The system of this invention has built an integrated collaborative architecture that connects the patient end, nurse end, doctor end and cloud platform intelligent backend. For the first time, it has achieved unified aggregation and intelligent flow of clinical data, nursing assessment and patient report outcomes among multiple roles, providing a reliable foundation for full-cycle and multi-dimensional patient management and significantly improving the collaboration efficiency among doctors, nurses and patients.

[0023] 2. Integration of Dynamic Prediction and Personalized Plan Generation: The cloud platform intelligent backend of this invention utilizes integrated multi-source time-series data and machine learning models to dynamically predict recurrence risk, overcoming the limitations of static assessment. Simultaneously, the embedded follow-up plan generation engine can automatically generate and flexibly adjust personalized follow-up plans based on structured clinical guidelines, individual risk levels, and specific characteristics, achieving a leap from standardized to precision management.

[0024] 3. A closed-loop early warning and collaborative intervention mechanism has been established: The system of this invention continuously monitors real-time patient data and risk changes through an independent early warning rule engine, and automatically triggers early warnings based on preset rules. Early warning information can be automatically transferred and collaboratively processed between the nurse's and doctor's ends according to clear escalation conditions, forming a complete closed loop from risk monitoring and verification to advanced intervention, which greatly reduces the risk of missed diagnosis and loss to follow-up for high-risk patients.

[0025] 4. Comprehensive optimization of experience and efficiency for all roles: For patients, the system provides a structured symptom reporting channel and visualized risk feedback, improving awareness, participation, and follow-up compliance; for nurses, the system promotes a shift towards structured and proactive management through standardized follow-up forms, proactive early warnings, and personalized health education tools; for doctors, the system helps them efficiently manage patient groups and achieve early and precise intervention through visualized cohort overviews and dynamic tracking of individual risks.

[0026] In summary, the system of this invention systematically improves the accuracy and efficiency of follow-up management and enhances patient experience by organically combining intelligent data fusion, dynamic risk prediction, personalized decision support, and multi-role closed-loop collaboration. Ultimately, it achieves the fundamental goals of optimizing medical resource allocation, improving medical quality, and enhancing patient prognosis. Attached Figure Description

[0027] 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.

[0028] Figure 1 This is a schematic diagram of the system architecture and module composition provided by the present invention.

[0029] Figure 2 This is a schematic diagram of the personalized follow-up plan generation and distribution process based on the cloud platform intelligent backend provided by the present invention. Detailed Implementation

[0030] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0031] Example 1 refer to Figure 1 This embodiment provides a follow-up management and recurrence risk warning system for non-muscle-invasive bladder cancer (NMIBC). The system integrates functional modules for patients, nurses, and doctors, along with a unified cloud platform intelligent backend, by constructing a collaborative platform. This enables data fusion, intelligent risk prediction, and personalized follow-up management.

[0032] The system specifically includes a patient-side module, a nurse-side module, a doctor-side module, and a cloud platform intelligent backend.

[0033] 1. Patient-side module The patient-side module is deployed on the patient's smartphone or other mobile terminal, typically in the form of an application (App) or WeChat mini-program. Its main functional units include: (1) Symptom Diary Unit: This unit provides patients with clinically validated electronic, structured scales to quantify their symptoms and quality of life. For example, the Mini-Fatigue Scale can be used for fatigue; for overall quality of life, the Quality of Life Specific Module for Non-Muscle-Invasive Bladder Cancer Patients or the Bladder Cancer Patient Quality of Life Scale can be used; for pain, the Visual Analogue Scale can be used. For core symptoms such as hematuria, dysuria, and difficulty urinating, the system also provides a self-made structured questionnaire, including multiple-choice questions on the occurrence of hematuria (absence, presence: visible light pink, visible bright red, visible blood clots), number of nighttime urinations, and difficulty urinating (absence / presence). Patients can fill out the questionnaire periodically or as needed, and the data is uploaded after standardization.

[0034] (2) Health Education Receiving Unit: In the form of a message center or knowledge base, it receives and displays personalized health education materials pushed by the nurse terminal module. The content of these materials is related to the patient's current treatment stage, risk assessment results, and specific symptoms, such as precautions after bladder instillation for BCG treatment patients and dietary and exercise recommendations for patients at high risk of recurrence.

[0035] (3) Follow-up Calendar Unit: The personalized follow-up plan generated by the system for the patient is displayed intuitively in a calendar or timeline view. Each plan clearly lists the follow-up time, location, required examinations, and related preparation requirements. The system integrates an intelligent reminder function, providing multiple reminders before the planned time through application push notifications and SMS messages. The interface uses color coding, such as green for tasks to be completed, blue for confirmed tasks, and red for overdue tasks, to intuitively distinguish the status of each task and help patients easily manage complex follow-up arrangements.

[0036] (4) Risk Index Display Unit: This unit uses user-friendly visual charts to display the patient's individual recurrence risk index and its historical trend, calculated by the cloud platform's intelligent backend. It also provides easy-to-understand explanations of the main reasons for recent changes in risk level, such as: "Due to your recent report of hematuria, your risk level has been adjusted from intermediate to high risk. Close monitoring is recommended, and your next cystoscopy follow-up examination should be completed on time."

[0037] 2. Nurse-side module The nurse-side module typically provides nurses with a web-based management backend or a dedicated tablet application interface, aiming to free them from fragmented record-keeping tasks and enable them to become proactive follow-up managers. Its core components include: (1) Patient Queue Management Unit: Clearly displays all patients under the care of nurses in the form of cards or lists. Each card integrates key information: patient name, the date of the next scheduled follow-up, and the current risk label automatically calculated and marked by the system. Nurses can use the top filter to view today's tasks, patients requiring warnings, or patients who have not responded by the deadline, etc., with one click, so as to quickly focus their work and manage priorities.

[0038] (2) Structured Follow-up Form Unit: The system has a built-in standardized electronic form for telephone or video follow-up. The form content is structured and covers: 1) Disease information: symptom review; 2) Psychosocial assessment: using simplified scales (PHQ-2 / GAD-2) to assess patients' anxiety and depression; 3) Quality of life score: using the Quality of Life Specific Module for Non-Muscle-Invasive Bladder Cancer Patients (EORTC QLQ-NMIBC24) or the Quality of Life Scale for Bladder Cancer Patients (FACT-BLSF-36) to assess patients' quality of life. Nurses can directly check or fill in the form during the follow-up process. After submission, the data is automatically synchronized to the cloud and a standardized structured nursing record is generated, eliminating the need for secondary paper transcription and significantly improving efficiency and data accuracy.

[0039] (3) Recurrence Early Warning Monitoring Unit: This unit monitors and displays in real time various early warning information automatically triggered by the cloud platform's intelligent backend. The generation of early warnings is based on the fusion analysis of two types of data: objective clinical data (pathological grading and surgical procedures synchronized from the Hospital Information System (HIS)) and subjective PROs data (symptom scores reported by patients through symptom diaries). For example, when the system identifies that a patient has reported gross hematuria and is marked as high-grade (G3) in their electronic file, a high-risk early warning will be automatically triggered and displayed at the top of this unit. Each early warning is accompanied by standardized handling suggestions preset by the system, such as: 1. Immediately contact the patient by phone to verify the hematuria and accompanying symptoms; 2. If the situation is true, mark "verified" in the system and click the "upgrade to doctor" button.

[0040] (4) Health Education Toolkit Unit: Embedded with a structured, categorized, and searchable multimedia knowledge base for health education. Nurses can quickly filter and preview relevant content from the knowledge base based on early warning information, follow-up assessment results, or patients' proactive inquiries, such as key points for home observation of intermittent hematuria, common reactions and management after BCG perfusion, etc., and send them to the target patient's mobile terminal (patient-side module) with one click. The system automatically records the content, time, and receiving patient of each educational push, facilitating tracking and effectiveness evaluation.

[0041] 3. Doctor's side module The doctor-side module provides doctors with an efficient web management view, assisting them in macro-level management of patient cohorts and in-depth assessment of individual patients. Its main units include: (1) Patient Cohort Overview Unit: Displays all NMIBC postoperative patients managed by the attending physician in a centralized manner, either in list or dashboard format. Key information such as patient name, latest risk level, next critical follow-up time, and whether the patient is currently in a warning state is readily available. Physicians can use the sorting and filtering functions in the header to quickly sort the patient cohort by risk level from high to low or by follow-up time from near to far, thereby prioritizing and treating the highest-risk, high-risk, or patients whose follow-up is about to expire, thus optimizing the allocation of medical resources.

[0042] (2) Individual Risk Tracking Unit: When a doctor clicks on a patient in the queue, they will enter a 360-degree detailed view of that patient. This view uses a clinical timeline as its framework, vertically integrating and displaying the patient's entire medical history, including pathology reports, cystoscopy images, ultrasound / CT / MRI imaging reports, laboratory test results, and simultaneously visualizing the symptom score curves reported by the patient through a symptom diary and the historical change curves of the recurrence risk index calculated by the system. This integrated view allows doctors to intuitively understand the potential driving factors of risk fluctuations, thereby achieving truly dynamic and continuous assessment.

[0043] (3) Early Warning and Intervention Center Unit: This unit is the hub for doctors to receive emergency or important intervention signals. It displays two types of information: first, medical-grade early warnings that have been processed and verified by nurses; and second, patients who have not completed their follow-up visits on time and whose follow-up visits have been automatically marked by the system according to preset rules. For these pending matters, doctors can directly send text messages to specific patients within the platform to provide guidance and flexibly adjust their follow-up plans online, such as bringing forward the cystoscopy examination originally scheduled for 3 months to 1 month later, or directly giving a clear instruction to go to the outpatient clinic for a follow-up examination as soon as possible. All operations are recorded by the system.

[0044] 4. Cloud Platform Intelligent Backend The cloud platform's intelligent backend is responsible for centralized data storage, fusion analysis, intelligent prediction, and decision generation. This includes: (1) Data Preprocessing Unit: This unit processes the aggregated multi-source heterogeneous data. Its process includes: Cleaning: Abnormal values ​​should be detected within a clinically reasonable range, and simple imputation of missing values ​​for key prognostic variables is prohibited; Standardization: Use systematic medical terminology sets such as SNOMED CT to encode textual information such as symptoms and diagnoses, and standardize units of measurement and date formats; Alignment: Using the date of the patient's first transurethral resection of bladder tumor as the absolute time anchor, all follow-up, treatment, and recurrence events are converted into relative time series; Integration: The processed data is loaded into a unified patient data model through an ETL framework.

[0045] (2) Feature Engineering and Intelligent Fusion Unit: This unit uses preprocessed, multi-source, structured data stored in a unified data model as a foundation to perform feature engineering and deep data fusion for machine learning prediction models and rule engines. Its core lies in generating higher-order composite features with direct clinical decision-making significance based on clinical medical knowledge. It includes: The system uses pre-defined fusion rules based on embedded clinical expert knowledge to weight and combine data pointing to the same clinical phenotype, generating composite features.

[0046] If the system detects that a patient has actively reported a high frequency of gross hematuria events within a similar time window, and the laboratory data from the hospital information system simultaneously shows a strong positive result for occult blood in urine (3+) and a significant increase in urine red blood cell count, the system will trigger the fusion rule.

[0047] This rule integrates and logically judges the aforementioned discrete and heterogeneous features to generate a composite index for the risk of active bleeding and severe hematuria. This index comprehensively reflects the severity and frequency of the patient's subjective symptoms and the strength of objective evidence. Its value or grade can more sensitively and specifically indicate the high-risk clinical condition of suspected acute active bleeding.

[0048] Each time a risk prediction is performed, the system constructs a dynamically updated feature vector for each patient. This vector not only contains static and dynamic features from a unified data model, but more importantly, it integrates composite features generated in real time through the aforementioned fusion process. This structured, semantically rich feature vector serves as the direct input for the machine learning prediction model to calculate real-time risk probabilities, and also as a high-level basis for the rule engine to determine whether an alert has been triggered and how to categorize it.

[0049] Through the aforementioned feature engineering and intelligent integration process, the system achieves a leap from multi-source data association to intelligent information extraction for clinical decision-making. This enables subsequent predictions and early warnings to be based on a deep understanding of complex clinical patterns such as the coordinated deterioration of patients' subjective feelings and objective test results over time, thereby achieving more accurate risk assessment and intervention.

[0050] (3) Unified Patient Data Model: A unified, structured digital profile is constructed for each patient in a cloud database. This model stores two types of data: 1) Static data, including demographic information and immutable clinicopathological data. Demographic information includes age and gender; immutable clinicopathological data includes T stage of initial diagnosis, pathological grade, presence of carcinoma in situ, number and diameter of tumors, bladder instillation treatment plan, etc. 2) Dynamic data, including laboratory test results over time, imaging reports and key images, as well as symptom and quality of life data from the patient.

[0051] (4) Machine learning prediction model: (4.1) Model training objectives and task definition Machine learning prediction models define multi-timescale prediction tasks for different clinical decision-making scenarios, and their training objectives include: Short-term recurrence prediction: This method aims to predict whether a patient will relapse within the next 3 to 6 months. It is used to identify patients in the acute phase of recurrence and to help decide whether to perform a cystoscopy immediately or in advance.

[0052] Mid- to long-term recurrence or progression prediction: This predicts whether a patient will relapse or progress in the next 1 to 2 years. It is used to assess long-term treatment strategies and assist in decision-making regarding whether an intensive treatment plan is needed or to evaluate the necessity of radical surgery.

[0053] Long-term survival analysis: Using survival analysis models (such as the Cox proportional hazards model), the probability of relapse-free survival of patients in the next 3 to 4 years is predicted, and the survival curve over time is output to assess the patient's overall disease trajectory and long-term prognosis.

[0054] (4.2) Model performance evaluation, validation and acceptable threshold To ensure the reliability, safety, and clinical value of the model's predictions, multi-level validation and reaching preset performance thresholds are required before and during clinical use and subsequent iterations. Core performance evaluation metrics include: Discrimination power: The area under the receiver operating characteristic curve (AUC) was used as the core indicator to measure the model's ability to distinguish between relapsed and non-relapsed patients. An AUC > 0.75 was considered to have basic discrimination power, an AUC > 0.8 was good, and an AUC > 0.85 was excellent.

[0055] Calibration: This is evaluated using calibration curves to verify the accuracy of the model's predicted probabilities and ensure that the predicted risks are consistent with the actual observed risks. Poorly calibrated models can seriously mislead clinical decision-making.

[0056] Clinical utility: Using decision curve analysis (DCA), the net clinical benefit of the model to the patient population under different risk decision thresholds was quantified compared to the default strategy of intervening in all patients or not intervening in all patients.

[0057] In addition, the system employs a multi-stage verification process and performance thresholds: Internal validation: On the model development dataset, techniques such as cross-validation are used for preliminary evaluation. The goal is to achieve an AUC > 0.75 and good calibration on internal data.

[0058] External validation: Validation is performed on completely independent datasets from patient cohorts at different medical institutions or at different times. This stage is crucial for verifying the model's generalization ability and preventing overfitting, requiring that model performance, especially calibration, does not significantly degrade.

[0059] Clinical utility validation: Decision curve analysis demonstrates that using this model to guide clinical decision-making can bring higher net clinical benefits compared to existing standard follow-up strategies, meaning that the benefits significantly outweigh the unnecessary intervention costs and risks.

[0060] (4.3) Construction and processing of model training data The training of machine learning predictive models relies on structured retrospective cohort data containing definitive outcomes extracted from historical electronic medical records and follow-up records. The construction and processing of this dataset follows these core principles and technical steps: Basic structure and label definition of training data: a. Feature fields: The training dataset covers static data, dynamic data, and patient-reported outcome data.

[0061] The static characteristics include demographic features and immutable clinicopathological features. Specific variables include: age (continuous variable), T stage at initial diagnosis (categorical variable, such as Ta, T1), pathological grade (categorical variable, such as low grade, high grade), presence of carcinoma in situ (categorical variable), number of tumors (categorical variable, such as single, multiple), maximum tumor diameter (categorical variable), and bladder instillation treatment regimen (categorical variable). The dynamic data are monitoring data that change over time. Specific variables include: results of previous laboratory tests (such as urine red blood cell count, which is a continuous variable), imaging findings, and patient-reported symptom data, such as the occurrence of hematuria (categorical variable), type of hematuria (categorical variable, such as no hematuria, microscopic hematuria, gross hematuria), frequency of hematuria (continuous variable), and duration of hematuria (continuous variable). Patient-reported outcome data include symptom scores and quality of life scores that patients regularly report using structured scales.

[0062] b. Outcome label: Each training sample must include a clear outcome label, including at least the relapse status (yes / no) and the corresponding relapse time (for relapsed patients) or the last follow-up time (for non-relapsed patients).

[0063] Data standardization and alignment consistent with the online system: To ensure that the training data is consistent with the distribution of online real-time data, historical data is processed in the same way as the data preprocessing unit, including: standardizing terminology using a unified medical terminology coding system such as SNOMED CT and LOINC; aligning the timelines of all events with the date of the first transurethral bladder tumor resection as the absolute time anchor; and using the same logical calculations to derive features.

[0064] Specific data processing and sample construction strategies during the training phase: To prevent data leakage during model training and ensure its ability to effectively predict future unknown data, the following key processes must be performed before model training: a. Sample Definition and Time Window Segmentation: When constructing training samples for each patient, it is necessary to clearly define the observation window and the prediction window. For example, all data collected within 12 months after the patient's first surgery can be used as the observation window feature to predict whether recurrence will occur within the subsequent 6-month prediction window. Each patient can be segmented into multiple independent training samples based on different observation time points throughout their follow-up period.

[0065] b. Handling Censored Data: For patients who did not relapse at the follow-up cutoff, the relapse time is unknown. Model training requires explicitly recording the last follow-up time as the censoring point. When using survival analysis models (such as the Cox model), this information will be directly used by the model to calculate the partial likelihood function.

[0066] c. Addressing Class Imbalance: Relapsed events typically represent a minority of all cases. During the training phase, technical measures are needed to mitigate class imbalance, such as oversampling the minority class, undersampling the majority class, or adjusting class weights at the algorithmic level to prevent the model from overly favoring the prediction of the majority class.

[0067] d. Divide the dataset according to the time-series principle: To simulate real-world scenarios where the model predicts future patients, it is essential to ensure that all patients in the test set used for final performance evaluation have follow-up periods that are significantly later than those in the training and validation sets. This time-based partitioning effectively avoids spurious overestimation of model performance due to data changes over time, and is crucial for verifying the model's generalization ability and practicality.

[0068] (4.4) Deployment, update and iteration mechanism of the model The trained model is integrated into the cloud platform's intelligent backend as a service, providing a real-time risk prediction interface. The model employs a strategy of fixed parameters and periodic full retraining for updates. Service model: The deployed model has fixed parameters, providing stable and consistent prediction services for patient data.

[0069] Update Cycle and Process: The system initiates the model iteration process automatically or upon administrator triggering, either at a preset cycle or after accumulating sufficient new follow-up data. This process includes: a. Retraining: Retrain the model using all historical data, including old data and newly accumulated data.

[0070] b. Rigorous evaluation: The retrained model shall be evaluated in strict accordance with the multi-stage verification process described in (4.2).

[0071] c. Clinical approval and deployment: After the validated new model version is approved through necessary clinical review, it replaces the old model in the online service through a smooth switch or blue-green deployment, completing a full model iteration.

[0072] d. Performance monitoring: The system continuously monitors the prediction performance of the online model. When the performance is detected to have degraded to below a preset threshold, a temporary model retraining and update process can be triggered.

[0073] In a specific implementation, a gradient boosting mechanism can be used to construct a supervised learning model; in this embodiment, the XGBoost model is preferred. Taking a binary classification task predicting whether a patient will relapse within the next 6 months as an example: the model is trained using historical large-scale patient data to learn the complex nonlinear relationship between features and relapse outcomes. After the model is trained, for patients in follow-up, inputting their latest fused data will output their relapse probability value for a specific time period in the future, such as 6 months or 1 year.

[0074] In another specific implementation, a time-series survival analysis model (such as the Cox proportional hazards model) can be used. This model can handle censored data and output a risk function that changes over time. As shown in Table 1, the model obtains the risk coefficient (β) for each clinical characteristic by fitting historical data, and then calculates the hazard ratio relative to baseline risk and the cumulative relapse probability at a specific time point for new patients.

[0075] Table 1: Examples of Characteristic Coefficients in the Cox Proportional Risk Model

[0076] (5) Follow-up plan generation engine: such as Figure 2 As shown, the engine is the core logic module for automatically generating personalized follow-up plans. It consists of three basic components and an intelligent execution logic.

[0077] (5.1) Clinical Guideline Knowledge Base: The clinical guideline knowledge base is a core component of the follow-up plan generation engine. Its role is to transform unstructured medical text guidelines into structured rules that can be automatically parsed and executed by computers. The construction and operation of this knowledge base depends on two working components: a configurable parameter table and a rule engine.

[0078] (a) Configurable parameter table This table stores variable numerical or optional parameters from clinical guidelines in key-value pairs, enabling the system to flexibly adapt to guideline updates without modifying the core program logic. As shown in Table 2, the parameter table defines differentiated follow-up intervals and total durations for patients with different risk levels.

[0079] Table 2 Parameter Table

[0080] (b) Rule Engine The rule engine is responsible for executing structured "IF-THEN" logical judgments. It calls specific values ​​from the parameter table, transforming the decision path of clinical guidelines into executable code logic. For example, a follow-up rule for low-risk patients can be expressed as: IF Patient. Risk Level == "Low Risk" THEN Scheduled task.Type = "Cystoscopy"; Scheduled task.Due date = Surgery date + Get parameter("LOW_RISK_FU_INTERVAL_1"); END IF Through this mechanism, complex textual guidelines are transformed into a series of precise, automatically scheduled clinical decision flows.

[0081] (c) Knowledge base construction and updating mechanism The initial construction and subsequent updates of the clinical guideline knowledge base were achieved through a semi-automated process combining automated information extraction and manual review and confirmation, specifically including the following steps: Information Extraction and Structuring: The system imports unstructured clinical guideline texts, typically in PDF or WORD format, and extracts key information using an integrated medical natural language processing module. This module, based on a pre-trained medical NLP model, can identify key medical entities in the text, such as diseases, examinations, drugs, numerical values, and time frequencies, as well as the conditional, parallel, and causal logical relationships between entities.

[0082] Medical ontology mapping and standardization: Extracted medical entities are mapped and aligned with a standard medical terminology system. For example, the textual description of high-grade urothelial carcinoma is mapped to the standard code SNOMED CT: 399510009, and cystoscopy is mapped to LOINC: LP30773-5. This step ensures semantic consistency and unambiguity of multi-source guideline terms within the system.

[0083] Logical Modeling and Rule Generation: Utilizing a business rule model, this embodiment preferably employs a decision tree model to formally model the standardized guideline logic. The system automatically or semi-automatically converts text descriptions into structured "IF-THEN" rules (see Table 4 for follow-up rule descriptions) and extracts variable parameters such as thresholds and intervals into a configurable parameter table (see Table 2).

[0084] Review, Release, and Version Management: The draft rules generated through the above process must be submitted to the system review interface for clinical experts to confirm, revise, or reject before they can be officially released and take effect. When a new version of the clinical guidelines is released, the system administrator can re-import the document to trigger the above process and generate an updated draft. The system supports version management of the knowledge base to ensure the traceability of changes to follow-up logic and the smooth migration of patient follow-up plans.

[0085] Through the above process, this embodiment realizes the dynamic transformation of unstructured, text-based clinical guidelines into a structured knowledge base that can be automatically executed by computers and updated synchronously with the progress of medical evidence, providing an accurate and reliable decision-making basis for the generation of personalized follow-up plans.

[0086] (5.2) Comprehensive Risk Stratification Unit: This model receives the recurrence risk probability value output by the machine learning prediction model, and combines it with the patient's key, static clinical characteristics for fine-tuning and comprehensive judgment to form the final risk probability used to guide clinical follow-up decisions. Comprehensive stratification is based not only on predicted probabilities, but also fully considers clinicopathological characteristics.

[0087] (5.3) Follow-up Plan Template Library: Standardized, time-series-based follow-up plan templates are pre-set for each risk level. These templates specify the recommended examinations and their frequencies at different postoperative time points. For example, the template corresponding to high-risk NMIBC is defined in a JSON structured format, and its content example is as follows: { "risk_level": "high", “followup_plan”: [ { “month”: 3, “examinations”: [“cystoscopy”, “urine cytology”], “critical”:true}, { “month”: 6, “examinations”: [“cystoscopy”], “critical”: true}, { “month”: 9, “examinations”: [“cystoscopy”, “urine cytology”], “critical”:false} / / ... More time points ] } The follow-up plan generation engine's generation and adjustment logic is as follows: First, the engine retrieves the corresponding basic follow-up template from the template library based on the patient's risk level. Then, based on the patient's specific individual circumstances, it automatically and personally adjusts the template's parameters. Personalized adjustment is a rule-based engine-driven decision-making process, with adjustment dimensions and rules based on clinical guidelines and expert consensus, including at least the following scenarios: Real-time symptom-based adjustments: If a patient reports gross hematuria, the system will analyze and respond according to preset rules. For example, the system may suggest suspending bladder instillation therapy and automatically generate an emergency follow-up task. The adjustment logic follows relevant clinical consensus. For instance, for patients with hematuria, combined urine cytology and urine tumor marker imaging examinations are required; if any result is abnormal, cystoscopy should be scheduled.

[0088] Age-based adjustments: The system considers patient age when balancing the benefits of examination with the risks of invasiveness. If the patient is over 75 years old, the system can automatically replace cystoscopy with the relatively non-invasive urinary tract ultrasound at non-critical follow-up points. In addition, for elderly (≥60 years) high-risk patients for bladder cancer, the system will enhance urine testing in the follow-up plan.

[0089] Adjustments based on treatment history: If the patient is receiving or has completed BCG bladder instillation therapy, the system will automatically add specific monitoring and assessment points at key post-treatment time points (3 months after treatment) on the basis of the standard follow-up template, such as urine cytology and inflammatory marker monitoring, to assess efficacy and adverse reactions in the early stages.

[0090] Adjustments based on complications / comorbidities: If the patient has renal insufficiency, the system will automatically adjust the suggested CT urography (CTU) to magnetic resonance urography (MRU) when generating a plan involving imaging examinations to avoid the risk of contrast-induced kidney injury. Furthermore, if the patient experiences severe bladder irritation symptoms or high fever during infusion, the system may recommend delaying or discontinuing the infusion treatment.

[0091] Relapse-based reset: The system follows the principle that once a patient experiences a relapse, the follow-up plan must be restarted. When the system records a relapse event, it will re-stratify the patient's risk based on the new pathological results and generate a completely new follow-up plan corresponding to the current risk level, completely replacing the previous plan.

[0092] The specific parameters in the above adjustment logic, such as age thresholds, follow-up intervals, and examination substitution relationships, are all stored in a configurable clinical rule knowledge base and can be updated based on the latest medical evidence. Ultimately, the engine outputs a complete, structured, personalized follow-up plan.

[0093] (6) Early warning rule engine The early warning rule engine is a core component of the cloud platform's intelligent backend, responsible for real-time risk monitoring and automatic alerts. It operates alongside the follow-up plan generation engine, but its function focuses on the continuous scanning of real-time and dynamic patient data, automatically identifying high-risk scenarios and triggering alerts based on pre-stored, configurable clinical rules.

[0094] The engine shares its configurable parameter table (see Table 2) with the clinical guideline knowledge base, where key parameters such as warning trigger thresholds and overdue days can be defined and updated. The warning rule engine integrates multiple types of rule-based triggering conditions based on clinical evidence and expert consensus, mainly including the following four categories: Real-time warning of key symptoms: When a patient reports gross hematuria, the highest level of warning is immediately triggered, which is the most direct sign of tumor recurrence or progression; Dynamic symptom score early warning: Dynamic monitoring of standardized scores regularly reported by patients aims to capture subtle symptom changes that may indicate disease progression; High-risk patient compliance warning: For patients with high or highest risk, if the planned execution date of their key follow-up tasks is delayed by more than a preset number of days, a tiered warning will be triggered to proactively manage the risk of loss to follow-up; Early warning of deteriorating quality of life: Monitor changes in the patient's total score or key dimensions of the quality of life scale. If the total score drops by more than a preset percentage in consecutive assessments, an early warning will be triggered, indicating that the cause of the deterioration needs to be evaluated.

[0095] To illustrate its operating mechanism, two core rules are listed below: Core symptom real-time alert rule: When the system detects a patient reporting gross hematuria coded by SNOMED CT, and the electronic medical record rules out benign causes such as recent urinary tract infection, a red alert is immediately triggered. Alert actions include pushing urgent messages to nurses and doctors, automatically creating urgent examination tasks, and suspending routine follow-up plans.

[0096] High-risk patient compliance warning rules: If a patient with a high or highest risk level has a critical follow-up task that is more than 7 days overdue, a yellow warning will be triggered and sent to the nurse's end; if the overdue period exceeds 30 days, it will be upgraded to a red warning and automatically sent to the doctor's end.

[0097] When the early warning rule engine triggers an early warning, the system message routing logic first pushes it to the recurrence early warning monitoring unit in the nurse module, where nurses perform initial verification and preliminary intervention, forming a nursing-level early warning. If the early warning meets the preset escalation conditions, the early warning and its processing context will be automatically escalated and pushed to the early warning and intervention center unit in the doctor module, thus forming a complete collaborative closed loop from monitoring and verification to advanced intervention.

[0098] (7) Early warning escalation mechanism The system defines clear escalation conditions for alerts to achieve a collaborative closed loop from nursing to medical levels. An alert will be automatically or manually escalated to the physician's alert and intervention center unit when any of the following conditions are met: Manual escalation: After initial assessment in the recurrence warning monitoring unit, if the nurse determines the situation is complex, beyond the scope of nursing intervention, or suspected of being a serious adverse event, they can manually click the "Escalate to Doctor" button. The system will then change the warning status to "Escalated" and push it along with the nurse's assessment notes to the attending physician.

[0099] Automatic rule escalation: This is automatically triggered by the early warning rule engine based on preset high-risk logic, requiring no manual operation from nurses. It mainly includes: Warning level reached: The warning itself is the highest level.

[0100] Patients at extremely high risk: The patients covered by the warning have the highest current system risk level.

[0101] High-risk symptom pattern: Patients who report ≥2 gross hematuria events in a short period of time indicate a high risk of active bleeding.

[0102] Critical treatment period alert: The warning occurs within a special treatment window, such as reporting "high fever (body temperature ≥38.5℃)" within 72 hours after BCG bladder instillation treatment, indicating that a serious adverse reaction may occur.

[0103] Forced System Upgrade: To prevent alerts from being missed due to busy schedules or negligence, the system monitors the processing status of alerts over a timeout period. If an alert is not processed or updated by a nurse within a preset time limit after it is issued, the system will automatically force an upgrade and notify the head nurse or a higher-level manager for supervision.

[0104] Through the aforementioned multi-path and intelligent upgrade mechanism, the system has constructed a complete collaborative closed loop from monitoring and assessment to tiered intervention, ensuring that high-risk clinical signals can be captured and processed in a timely and accurate manner by medical staff in different roles.

[0105] In this embodiment, the patient-side, nurse-side, and doctor-side modules are all developed and implemented as WeChat mini-programs, facilitating quick access for users on different devices. Routine data queries, symptom reporting, and follow-up record submissions between each module and the cloud platform's intelligent backend are all encrypted and transmitted via secure HTTPS. To enable real-time and proactive push of early warning information, a TLS / SSL encrypted WebSocket long-lived connection is established between the cloud platform's intelligent backend and the nurse-side and doctor-side modules, ensuring that high-risk warnings can be received and processed by medical staff immediately, forming a closed-loop response mechanism.

[0106] Example 2 This embodiment describes a method for follow-up management and recurrence risk warning of non-muscle-invasive bladder cancer using the above-described system. For example... Figure 2 As shown, this method is closely integrated with the system functions and includes the following steps: S1: Patient Data Reporting. Patients use the patient-side module on their mobile devices, based on system prompts or their own needs, to periodically or event-triggeredly report their symptom severity and quality of life scores using a built-in structured electronic questionnaire. This subjective report data is transmitted securely and in real-time to the cloud platform's intelligent backend via the internet.

[0107] S2: Medical and Nursing Data Collection. During routine telephone, video, or outpatient follow-ups, nurses use structured electronic forms within the nurse's end module to record their communications with patients, generating standardized nursing assessment data, including symptom review and psychological scores. Simultaneously, data generated by doctors during the diagnosis and treatment process, such as clinical decisions, newly prescribed examination plans, and confirmed postoperative pathology conclusions, are manually entered through the doctor's end module or automatically synchronized from the Hospital Information System (HIS) and Electronic Medical Records (EMR) to the cloud platform's intelligent backend via system interfaces. This step completes the collection of objective clinical data and nursing assessment data.

[0108] S3: Data Fusion and Dynamic Risk Prediction. The cloud platform's intelligent backend receives multi-source heterogeneous data from S1 and S2, cleans, standardizes, and aligns it, and integrates and stores it into the patient's unified data model. Subsequently, it calls a pre-trained machine learning prediction model (the XGBoost model or Cox model described in Example 1), using the patient's latest fused time-series data as input, to calculate the recurrence probability value within a specific future time period in real time.

[0109] S4: Personalized Follow-up Plan Generation. The follow-up plan generation engine in the cloud platform's intelligent backend is triggered. The engine first matches the risk level based on the clinical characteristics collected in S2, combined with the clinical guideline knowledge base, and then, based on the risk probability obtained in step S3, determines the overall risk level used to guide follow-up through a comprehensive risk stratification unit. Table 3. Risk grouping of patients with non-muscle-invasive bladder cancer

[0110] Next, the engine matches the corresponding rules from the clinical guideline knowledge base and retrieves the basic follow-up template corresponding to the risk level from the follow-up plan template library: Table 4. Description of Follow-up Rules

[0111] Then, the engine automatically adjusts the basic template according to preset rules based on the patient's specific individual attributes. Finally, it generates a structured draft follow-up plan that includes multiple future follow-up appointment times, specific examination items, reasons for examination, and personalized explanations.

[0112] S5: Plan Distribution, Execution, and Collaborative Intervention. The generated draft follow-up plan is first pushed to the corresponding interface of the doctor's module for online review, confirmation, or manual fine-tuning by the attending physician. After the doctor's confirmation, the plan officially takes effect and is simultaneously distributed to the patient's and nurse's modules. The system enters a state of automated execution and monitoring based on the plan: On the patient's side, the system sends reminders according to the plan; on the nurse's side, the system generates a list of tasks to be done based on the plan, and when the patient's real-time reported data triggers the early warning rules, an early warning is generated in the recurrence early warning monitoring unit to guide the nurse to execute standard procedures; on the doctor's side, the early warning and intervention center unit receives upgraded early warnings from the nurse's side or extremely high-risk signals automatically identified by the system. Throughout the entire process, the information flow, task flow, and early warning flow flow are orderly and closed-loop collaboratively between the patient, nurse, and doctor's ends, jointly forming a dynamic, proactive, and precise follow-up management ecosystem.

[0113] Example 3 To illustrate the working mechanism and advantages of the system of the present invention more specifically and completely, this embodiment uses the complete follow-up process of a fictional patient with non-muscle-invasive bladder cancer—Mr. Zhang—for explanation. See below for details: Mr. Zhang was diagnosed with low-risk NMIBC on January 1, 2025, and underwent transurethral resection of bladder tumor. After obtaining his surgery date and risk level, the system immediately triggered the follow-up plan generation engine. Based on the low-risk patient follow-up rules in the structured clinical guidelines, the system queried the parameter "LOW_RISK_FU_INTERVAL_1" (value 3) from the configurable parameter table and automatically generated the first follow-up task: cystoscopy, scheduled for April 1, 2025. This task was simultaneously pushed to the follow-up calendar unit in the patient-side module, the patient queue management unit in the nurse-side module, and the patient queue overview unit in the doctor-side module.

[0114] Step 1: Routine review and plan update: On April 1, 2025, Mr. Zhang underwent his first cystoscopy, which returned a negative result. Upon receiving this structured examination result, the system triggered the corresponding rules in the rule engine. The rule engine queried the parameter table, retrieving the parameters "LOW_RISK_FU_INTERVAL_2" (value 12) and "LOW_RISK_FU_INTERVAL_SUBSEQ" (value 12), and generated a new follow-up plan: the second cystoscopy is scheduled for April 1, 2026, 12 months after the initial negative result, and a routine cystoscopy cycle is set for once a year starting April 1, 2027, with the total follow-up duration following the parameter "LOW_RISK_FU_DURATION" (5 years).

[0115] Step 2: Real-time early warning and planned intervention triggered by sudden symptoms: On February 15, 2026, Mr. Zhang reported gross hematuria through the symptom diary unit of the patient-side module. The system standardized the identification of this symptom using the SNOMED CT code (17170007) and immediately triggered the preset warning rules. The system performed the following automated operations: 1. Generate a high-level warning: A red warning is generated in the warning and intervention center unit of the doctor's module and the recurrence warning monitoring unit of the nurse's module, with the title "Gross hematuria occurs during active monitoring, indicating that an urgent assessment is required."

[0116] 2. Dynamically adjust follow-up plans: Automatically pause all scheduled routine cystoscopy tasks for this patient, including the examination originally scheduled for April 1, 2026.

[0117] 3. Create an urgent task: Generate a highest priority urgent cystoscopy task that must be completed within 7 days, i.e., by February 22, 2026, and specify the reason as "symptom trigger: gross hematuria".

[0118] Step 3: Closed-loop processing and personalized reset: After receiving the alert through the early warning and intervention center unit, the attending physician arranged for Mr. Zhang to undergo an emergency cystoscopy on February 20, 2026. Based on the results of this emergency examination, the system entered a critical logic branch: Branch A (Negative Examination): The system uses this emergency examination date, February 20, 2026, as the new time anchor point to reset and restart the annual follow-up cycle. This means the next routine cystoscopy will be scheduled for February 20, 2027, instead of the original April 1, 2026. This demonstrates the system's ability to dynamically adjust the follow-up schedule based on real-time clinical events.

[0119] Branch B (positive test, i.e., recurrence): The system will trigger a pre-defined set of rules for handling recurrence. The core logic of this set of rules is to generate a completely new and more targeted treatment and follow-up plan based on the time of the recurrence event, the pathological characteristics of the recurrent tumor, and the patient's previous treatment history, completely replacing the previous plan.

[0120] For high-risk patients who have received BCG bladder instillation therapy, the management logic after relapse differs significantly from the initial follow-up rules. The system will execute the following specific rules: Table 5 Examples of rules for managing recurrence after BCG treatment

[0121] Through the aforementioned rules, the system can dynamically and individually adjust treatment strategies and follow-up intensity based on the critical event of relapse, achieving a closed-loop, adaptive disease management system covering the entire lifecycle from initial treatment, monitoring, and early warning to post-relapse decision-making. The plan generated after relapse will be simultaneously updated to both the patient and nurse's devices after doctor confirmation.

[0122] This embodiment clearly demonstrates, through a complete clinical scenario simulation, how the system deeply integrates structured guideline knowledge, real-time patient report data, automated early warning rules, and dynamic follow-up plan generation to construct an intelligent closed-loop management system from planning, execution, monitoring, early warning to intervention adjustment, achieving truly personalized and precise bladder cancer follow-up management.

[0123] Furthermore, the system's data processing capabilities also support the accurate operation of machine learning prediction models. Taking the historical data of another patient, A, as an example, the original discrete events, after being aligned by the system, are all converted into relative time series with the date of the first surgery as the origin, as shown in the table below: Table 6 Relative Time Series

[0124] When the model needs to predict patient A's future risk over the next 120 days, the system extracts a snapshot of the fused data at that moment, including both static and dynamic features as input, and outputs the recurrence probability over the next 6 months. As time progresses to the next 365 days, the dynamic features are updated, and the model makes predictions again based on the updated historical data. This demonstrates how the system transforms multi-source, heterogeneous time-series data into standardized features usable by the model, enabling dynamic and continuous risk assessment.

[0125] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A follow-up management and recurrence risk early warning system for non-muscle-invasive bladder cancer, characterized in that, include: The system includes a patient-side module, a nurse-side module, a doctor-side module, and a cloud platform intelligent backend. The patient-side module is used for patients to report symptom data, receive health education content, view personalized follow-up plans and personal relapse risk information; The nurse-side module is used by nurses to manage patient follow-up queues, perform structured follow-ups, receive and process relapse warnings, and push health education materials to patients. The doctor-side module is used to allow doctors to view the patient queue, track individual patient risks, receive alerts, and intervene. The cloud platform's intelligent backend is used to store and integrate multi-source data from patients, nurses, and doctors. Based on the integrated data, it runs a machine learning prediction model to calculate the patient's recurrence risk probability. Based on the recurrence risk probability and patient data, it generates a personalized follow-up plan for the patient through a follow-up plan generation engine.

2. The system according to claim 1, characterized in that, The patient-side module includes: The symptom diary unit is used for patients to report symptom information and quality of life scores through a structured scale, as well as to report hematuria through a self-made questionnaire; The health education receiving unit is used to receive and display personalized health education materials pushed by nurses. The follow-up calendar unit is used to visually display the personalized follow-up plan and provide reminder functionality; The risk index display unit is used to display, in chart form, the degree and probability of individual relapse risk and its changing trend, calculated by the intelligent backend of the cloud platform.

3. The system according to claim 1, characterized in that, The nurse-side module includes: The patient queue management unit is used to display the list of patients managed by nurses in the form of cards. The cards contain patient risk tags and support filtering by task and warning status. The structured follow-up form unit contains standardized forms that include disease information, psychosocial assessments, and quality of life scores, which nurses can fill out during follow-ups and which automatically generate nursing records. The recurrence early warning monitoring unit is used to receive early warning information triggered by the intelligent backend of the cloud platform. The early warning is generated based on the patient's objective clinical data and subjective PROs data, and includes standardized processing suggestions. The health education toolkit unit has an embedded health education knowledge base, allowing nurses to select content and push it to target patients.

4. The system according to claim 1, characterized in that, The doctor-side module includes: The patient cohort overview unit is used to centrally display all key information about managed patients and provides sorting and filtering functions by risk level and follow-up time. The individual risk tracking unit is used to display the clinical timeline of a single patient, integrating historical change curves of examination results, reported symptoms, and risk indices. The early warning and intervention center unit is used to receive information on high-risk or highest-risk patients automatically marked by the system, as well as patients who have not been followed up within the stipulated time, and supports doctors to send messages to patients or adjust follow-up plans through the platform.

5. The system according to claim 1, characterized in that, The cloud platform's intelligent backend includes a unified patient data model, used for structured storage of patients' static and dynamic data; The static data includes demographic information and immutable clinical pathology data; The dynamic data includes data from previous laboratory tests, imaging reports, and symptom and quality of life scores from patients.

6. The system according to claim 5, characterized in that, The cloud platform's intelligent backend also includes an early warning rule engine; The early warning rule engine is used to automatically trigger early warning information based on the data in the patient unified data model and / or the recurrence risk probability output by the machine learning prediction model, according to preset clinical early warning rules.

7. The system according to claim 6, characterized in that, When the pre-set upgrade conditions are met, the early warning information triggered by the early warning rule engine is automatically upgraded and pushed from the recurrence early warning monitoring unit of the nurse module to the early warning and intervention center unit of the doctor module.

8. The system according to claim 1, characterized in that, The follow-up plan generation engine includes: A clinical guidelines knowledge base for storing structured clinical follow-up rules; The comprehensive risk stratification unit is used to determine the patient's overall risk level based on the patient's clinical data and the recurrence risk probability output by the machine learning prediction model. A follow-up plan template library is used to store standardized follow-up templates corresponding to different comprehensive risk levels; The follow-up plan generation engine is configured to: call the corresponding follow-up template from the follow-up plan template library according to the comprehensive risk level determined by the comprehensive risk stratification unit, and adjust it in combination with the patient's individual characteristics to generate a personalized follow-up plan.

9. The system according to claim 8, characterized in that, The personalized parameter adjustment is based on preset clinical rules and dynamically adjusts the follow-up template according to the patient's individual attributes; the individual attributes include at least one of the following: real-time symptoms, age, treatment history, complications or comorbidities, and recurrence events.

10. A method for follow-up management and recurrence risk warning of non-muscle-invasive bladder cancer, characterized in that, The method uses the system as described in any one of claims 1 to 9, and the method includes the following steps: S1: Receive symptom data and quality of life score data reported by patients through the patient-side module; S2: Receive follow-up records and nursing assessment data entered by nurses through the nurse module; receive or synchronize patients' clinical diagnosis and treatment data through the doctor module. S3: In the intelligent backend of the cloud platform, data from the patient, nurse and doctor ends are integrated, and machine learning prediction models are used to calculate the patient's recurrence risk probability. S4: In the intelligent backend of the cloud platform, a personalized follow-up plan is automatically generated based on clinical guidelines and rules, the recurrence risk probability, and individual patient characteristics. S5: Synchronize the personalized follow-up plan to the patient, nurse, and doctor terminals; based on the follow-up plan and real-time data, trigger follow-up tasks and early warning processing on the nurse terminal, and trigger intervention tasks on the doctor terminal.

Images