Breast cancer relative survival risk prediction system based on machine learning

Abstract

The invention relates to the technical field of breast cancer risk prediction, and discloses a breast cancer relative survival risk prediction system based on machine learning. The system comprises four core modules, wherein a data acquisition module is used for acquiring multi-dimensional clinical diagnosis and treatment records of a patient from distributed medical nodes and synchronously acquiring dynamic biomarker time sequence information in a tumor evolution process; the feature compensation module performs missing feature compensation on the diagnosis and treatment records in a cross-modal medical data fusion scene to generate clinical feature indexes, and accordingly, individual monitoring frequency is determined; the risk analysis module extracts patient diagnosis and treatment path information in a medical network, analyzes node risk associated data, and determines a survival risk level in combination with dynamic biomarker time sequence information; and the survival early warning module realizes dynamic early warning of the treatment blind area according to the individualized monitoring frequency and the survival risk level. The system can efficiently integrate medical data, accurately assess survival risks, and provide dynamic and individualized support for clinical diagnosis and treatment of breast cancer.

Description

Technical Field

[0001] This invention relates to the field of breast cancer risk prediction technology, specifically a machine learning-based system for predicting the relative survival risk of breast cancer. Background Technology

[0002] In the clinical diagnosis, treatment, and management of breast cancer, accurate assessment and dynamic monitoring of patient survival risk have always been crucial aspects of medical practice. Currently, commonly used methods for assessing breast cancer survival risk rely heavily on static clinical data, such as tumor size, pathological type, and lymph node metastasis at initial diagnosis. These data often only reflect the patient's condition at a specific point in time and fail to comprehensively capture the dynamic changes during tumor progression. With the diversification of medical data collection methods, distributed medical nodes have accumulated a large amount of multidimensional clinical records, covering patient medical history, treatment plans, and examination reports. However, this data is typically stored in the information systems of different medical institutions, lacking an effective integration mechanism, making it difficult to efficiently integrate and utilize cross-institutional and cross-modal medical data.

[0003] In actual clinical data collection, due to differences in patient visit times, chosen examinations, and incomplete data recording, feature loss is a common problem in multidimensional clinical records. Traditional methods for handling missing values ​​often involve simple filling or direct removal of samples with missing values. This not only easily leads to the loss of important clinical information but may also cause biases in risk assessment results, failing to provide patients with accurate individualized treatment references. Furthermore, existing breast cancer survival risk assessment systems do not pay enough attention to the patient's treatment pathway, failing to fully explore the risk associations between different treatment stages and making it difficult to adjust monitoring strategies and treatment plans in a timely manner based on the patient's actual treatment progress and dynamic changes in biomarkers.

[0004] Dynamic biomarkers, as crucial indicators reflecting tumor progression and treatment response, hold significant value in accurately assessing patient survival risk. However, most current risk assessment systems fail to establish effective correlation models between dynamic biomarker time-series information and patient survival risk, and cannot integrate the dynamic trends of biomarkers with risk nodes in the treatment pathway, resulting in a lack of timeliness and specificity in risk assessment results. Regarding early warning of treatment blind spots, the lack of dynamic early warning mechanisms based on individualized monitoring frequency and real-time risk status levels makes it difficult for clinicians to identify potential treatment risks in advance, potentially delaying optimal treatment and impacting patient prognosis. These issues render existing breast cancer survival risk assessment systems significantly inadequate in terms of accuracy, dynamism, and personalization, failing to meet the current clinical demand for high-quality risk assessment services. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning-based system for predicting the relative survival risk of breast cancer, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a machine learning-based system for predicting relative survival risk in breast cancer, the system comprising:

[0007] The data acquisition module is used to collect multidimensional clinical diagnosis and treatment records of breast cancer patients from distributed medical nodes and to obtain dynamic biomarker time-series information in the process of tumor evolution in real time.

[0008] The feature compensation module is used to compensate for missing features in the multidimensional clinical diagnosis and treatment records under the cross-modal medical data fusion mode, generate patient clinical feature indicators, and determine individualized monitoring frequency based on the clinical feature indicators.

[0009] The risk analysis module is used to acquire patient treatment path information in the medical network, analyze the survival risk characteristics of the treatment path information, output risk association data of patient treatment nodes, and determine the survival risk status level based on the risk association data and the time series information of the dynamic biomarkers.

[0010] The survival early warning module is used to provide dynamic early warning for treatment blind spots based on the individualized monitoring frequency and the survival risk status level.

[0011] Preferably, the multidimensional clinical diagnosis and treatment record includes a structured chain set of patient pathological staging, treatment response records, gene expression profiles, and historical data on complications.

[0012] Preferably, the dynamic biomarker time-series information represents time-series data of tumor marker concentration, imaging feature change trends, drug metabolism cycle, and immune microenvironment status.

[0013] Preferably, determining the individualized monitoring frequency based on the clinical characteristic indicators specifically includes:

[0014] A spatiotemporal constrained model for patient survival analysis was constructed based on the aforementioned clinical characteristic indicators;

[0015] Extract the collaborative boundary threshold for dynamic monitoring from the spatiotemporal constraint model;

[0016] The individualized monitoring frequency for the patient follow-up period is determined by the collaborative boundary threshold.

[0017] Preferably, obtaining patient treatment pathway information in the medical network specifically includes:

[0018] Real-time collection of clinical status records across institutional treatment nodes;

[0019] The core diagnosis and treatment workflow is determined based on the aforementioned clinical status records;

[0020] Patient treatment pathway information is generated through the core treatment workflow.

[0021] Preferably, the survival risk feature analysis of the treatment pathway information specifically includes:

[0022] Based on the aforementioned treatment pathway information, a spatiotemporal correlation model between treatment node behavior and historical survival events is constructed.

[0023] The spatiotemporal correlation model outputs the risk coupling strength of the treatment intervention;

[0024] The risk impact characteristics of each diagnosis and treatment node are determined by the aforementioned risk coupling strength;

[0025] Risk association data for patient diagnosis and treatment nodes is generated based on the aforementioned risk impact characteristics.

[0026] Preferably, determining the survival risk status level based on the risk association data and the time-series information of the dynamic biomarkers specifically includes:

[0027] Based on the aforementioned risk correlation data, risk transmission constraints during the treatment process are determined;

[0028] The granularity of abnormal tumor evolution features is determined by the temporal information of the dynamic biomarkers.

[0029] A survival risk profile level is generated based on the risk transmission constraints and the granularity of the abnormal tumor evolution features.

[0030] Preferably, the dynamic early warning of treatment blind spots based on the individualized monitoring frequency and the survival risk status level specifically includes:

[0031] The intensity of elasticity monitoring within the treatment cycle is determined based on the individualized monitoring frequency.

[0032] The risk exposure index of the treatment blind spot is determined based on the aforementioned survival risk status level;

[0033] The elasticity monitoring intensity is matched with the risk exposure index using early warning rules to generate a dynamically adjustable treatment blind spot early warning map.

[0034] Preferably, the missing feature compensation for the multidimensional clinical diagnosis and treatment records specifically includes:

[0035] In the cross-modal medical data fusion model, the characteristic transmission path of the diagnosis and treatment process is determined based on the multidimensional clinical diagnosis and treatment records;

[0036] The missing feature nodes in the multidimensional clinical diagnosis and treatment records are probabilistically compensated to generate a complete medical event sequence.

[0037] The complete medical event sequence is subjected to feature aggregation, and the risk entropy value information during the treatment process is output.

[0038] The patient's clinical characteristic indicators are generated from the risk entropy value information.

[0039] Preferably, the system further includes a survival decision module:

[0040] This data is used to integrate the individualized monitoring frequency, the survival risk status level, and the treatment blind spot early warning map to generate a dynamic prediction report on the relative survival risk of breast cancer.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] By setting up a data acquisition module, multidimensional clinical records of breast cancer patients can be collected from distributed medical nodes, while simultaneously acquiring real-time time-series information on dynamic biomarkers during tumor progression. This design breaks through the limitations of traditional decentralized storage of medical data, achieving efficient aggregation of cross-institutional and cross-modal medical data. This allows the system to acquire more comprehensive and richer patient clinical information, providing a sufficient data foundation for accurate survival risk assessment. Compared to traditional assessment methods that rely solely on static data, the introduction of dynamic biomarker time-series information more realistically reflects the evolution of the tumor at different stages, enabling risk assessment to closely follow changes in the patient's condition and avoiding assessment biases caused by data lag or incompleteness.

[0043] The feature compensation module compensates for missing features in multidimensional clinical records under a cross-modal medical data fusion model, generating patient clinical characteristic indicators and determining individualized monitoring frequencies accordingly. Addressing the common problem of missing features in clinical data, this module abandons traditional methods of simply filling or removing samples. Through scientific feature compensation methods, it maximizes the retention of effective information in clinical data, ensuring that the generated patient clinical characteristic indicators accurately reflect the patient's actual condition. The individualized monitoring frequencies determined based on this foundation allow for the creation of personalized monitoring plans for each patient, taking into account factors such as the severity of their condition, data completeness, and the rate of tumor progression. This avoids the problems associated with traditional uniform monitoring frequencies, where over-monitoring of some patients leads to wasted medical resources, or under-monitoring of others results in missed risks, making monitoring more targeted and rational.

[0044] The risk analysis module acquires patient treatment pathway information from the medical network, analyzes the survival risk characteristics of this information, outputs risk association data for patient treatment nodes, and determines the survival risk profile level by combining this data with dynamic biomarker time-series information. This process fully explores the potential risk associations between various nodes in the treatment pathway, treating the patient's treatment process as an organic whole rather than isolated events, and clearly identifies potential risk nodes in the treatment process. Simultaneously, by combining risk association data with dynamic biomarker time-series information, a more comprehensive risk assessment model is constructed. This ensures that the determination of the survival risk profile level not only considers objective risk factors during the treatment process but also incorporates biomarker information reflecting the real-time status of the tumor, making risk level classification more scientific and accurate, and helping clinicians clearly understand the patient's risk level at different stages of treatment.

[0045] The survival early warning module provides dynamic warnings for treatment blind spots based on individualized monitoring frequency and survival risk level. This module links individualized monitoring frequency with real-time risk level, forming a dynamic early warning mechanism. When a patient's survival risk level changes or reaches a set individualized monitoring time point, the system can promptly issue an early warning signal, prompting clinicians to pay attention to potential treatment blind spots. This dynamic early warning method allows clinicians to perceive potential treatment risks in advance, adjust treatment plans and monitoring strategies in a timely manner, and avoid delays in treatment due to failure to identify treatment blind spots. Furthermore, the dynamic early warning mechanism can also assist clinicians in more rationally allocating medical resources, focusing their efforts on high-risk patients and key treatment points, improving overall medical service efficiency, and enhancing the patient's treatment experience and prognosis. Attached Figure Description

[0046] Figure 1 This is a time series diagram of the machine learning-based system for predicting relative survival risk of breast cancer as described in this invention.

[0047] Figure 2 A flowchart for analyzing survival risk characteristics;

[0048] Figure 3 A flowchart for dynamic early warning of treatment blind spots;

[0049] Figure 4 This is a flowchart for compensating for missing features. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 This invention provides a machine learning-based system for predicting relative survival risk in breast cancer, the system comprising:

[0052] By integrating clinical data and dynamic biomarker information from distributed medical nodes, the system achieves personalized survival risk prediction and dynamic early warning. First, the data acquisition module obtains multidimensional clinical records of breast cancer patients from multiple medical nodes, including structured data such as pathological staging and treatment response records. Simultaneously, it collects real-time time-series information on dynamic biomarkers during tumor progression, such as tumor marker concentrations and trends in imaging characteristics. The feature compensation module, under a cross-modal medical data fusion mode, performs probability compensation for missing features, generating complete clinical characteristic indicators for patients and determining personalized monitoring frequencies based on a spatiotemporal constraint model. The risk analysis module analyzes treatment pathway information in the medical network, constructs a spatiotemporal correlation model between treatment nodes and survival events, outputs risk-related data, and determines the survival risk status level by combining dynamic biomarker time-series information. The survival early warning module dynamically matches early warning rules based on personalized monitoring frequencies and risk status levels, generating a treatment blind spot early warning map. Finally, the survival decision module integrates all data to generate a dynamic prediction report.

[0053] Example 1: A structured chain of multidimensional clinical treatment records is constructed by integrating pathological staging, treatment response records, gene expression profiles, and historical complication data of breast cancer patients. Pathological staging data is sourced from the medical institution's pathology reporting system and is stored in a structured manner using the internationally recognized TNM staging standard, including key indicators such as primary tumor size, regional lymph node involvement, and the presence of distant metastases. This data is automatically extracted from the hospital information system through a standardized data interface and timestamped to record changes in the patient's pathological status at different stages of treatment. Treatment response records cover detailed implementation of interventions such as surgery, chemotherapy, radiotherapy, and endocrine therapy. Surgical records include surgical procedures, resection extent, and number of lymph nodes dissected; chemotherapy regimens record drug types, dosages, and number of cycles; radiotherapy data includes the target area, total dose, and fractionation pattern; and endocrine therapy records drug types and duration. This data is collected through structured fields in an electronic medical record system to ensure the integrity and traceability of treatment information. Gene expression profiling data were acquired using a high-throughput sequencing platform, detecting germline mutation status of the BRCA1 / BRCA2 genes, HER2 gene amplification levels, and estrogen and progesterone receptor expression. Sequencing results, after bioinformatics analysis, were stored in a standardized gene variant format in a dedicated database and correlated with the patient's clinical information. Historical complication data recorded various adverse events occurring during treatment, including chemotherapy-induced myelosuppression, radiotherapy-induced skin reactions, surgery-related infections, and endocrine therapy-induced bone and joint symptoms. This data was collected through the medical institution's adverse event reporting system and sorted by severity and occurrence time to form a complete complication time series.

[0054] The acquisition and processing of dynamic biomarker time-series information involves various detection technologies and analytical methods. Tumor marker concentration data is obtained by periodically collecting serum samples from patients; commonly used markers include CA15-3, CEA, and CA125. Detection results are transmitted to a central database through a laboratory information management system, generating concentration change curves according to sampling time. Imaging feature trend data is obtained from imaging examinations such as breast MRI, ultrasound, and mammography. Morphological features of the tumor are automatically extracted using deep learning algorithms, including parameters such as maximum diameter changes, tumor volume reduction rate, and lesion margin clarity. These feature parameters are arranged according to examination time points, forming a dynamic sequence of tumor morphological evolution. Drug metabolism cycle data is obtained through periodic blood drug concentration monitoring during treatment, recording the metabolic kinetics of chemotherapy drugs such as paclitaxel and doxorubicin in patients, including peak blood concentration, half-life, and clearance rate. This data is acquired using detection technologies such as high-performance liquid chromatography and correlated with the dosing regimen. Immune microenvironment status data were obtained by detecting immune cell infiltration in tumor tissue using flow cytometry or immunohistochemistry, including parameters such as CD8+ T cell density, regulatory T cell proportion, and PD-L1 expression level. These immune characteristic parameters were compiled into a time series based on biopsy time points, reflecting the dynamic changes in the tumor immune microenvironment during treatment.

[0055] In the data integration phase, multidimensional clinical records and dynamic biomarker time-series information are linked through patient-unique identifiers. Pathological staging data and imaging feature trends corroborate each other to assess the local progression of the tumor; gene expression profile data combined with immune microenvironment parameters reveal the potential association between tumor molecular characteristics and immunotherapy response; treatment response records and drug metabolism cycle data are cross-validated to assess the actual effectiveness of the treatment regimen. This multidimensional data fusion analysis comprehensively reflects the patient's disease status and treatment response.

[0056] Data quality control measures are implemented throughout the entire acquisition and processing process. Pathological staging data must be independently reviewed by at least two pathologists to ensure consistency of diagnostic results. Treatment response records are protected for integrity through mandatory field entry and logical validation rules in the electronic medical record system. Gene expression profile data undergoes standardized quality control processes, including sequencing depth assessment and variant filtering. Dynamic biomarker detection results must meet laboratory quality control standards, and outliers must be manually verified. This rigorous quality control mechanism ensures the reliability of the data upon which the analysis relies.

[0057] The storage and management of time-series data employs a specially designed database architecture. Clinical diagnosis and treatment records are organized around the patient, using a document-based database to store structured diagnosis and treatment events. Dynamic biomarker data is managed using a time-series database, optimizing the storage and retrieval efficiency of time-series data. Imaging feature data uses an object storage system to store the original image files and their derived feature data. This hybrid storage scheme takes into account the characteristics and access requirements of different types of data.

[0058] In the data preprocessing stage, the raw data are standardized and normalized. Pathological staging data are converted into numerical TNM staging codes; drug names in treatment response records are mapped to a standardized drug coding system; gene variation data are annotated according to the naming rules of the Human Genome Variation Association; and dynamic biomarker data undergo baseline correction and unit standardization. This standardized preprocessing enables unified analysis of data from different medical institutions and testing platforms.

[0059] The timeliness of data is maintained through a real-time update mechanism. New medical records trigger the data collection process immediately upon entry into the electronic medical record system; dynamic biomarker test results are automatically pushed to the central database after laboratory review and confirmation; and imaging examination reports automatically extract feature parameters via a DICOM gateway after issuance. This near real-time data update mechanism ensures that the analytical model can perform risk assessments based on the latest patient status. Data security and privacy are protected through multiple measures. Patient identification information is anonymized immediately after collection; data transmission uses encrypted channels; database access is subject to strict access control; and data use follows the principle of minimum necessity. These measures comply with relevant regulations on medical data privacy protection, ensuring the secure use of patient information.

[0060] The integrated analysis of multidimensional clinical records and dynamic biomarker time-series information provides a comprehensive data foundation for risk assessment. Pathological staging and treatment response records reflect the macroscopic progression of the disease; gene expression profiles reveal the molecular characteristics of the tumor; historical complication data indicate treatment tolerance; and dynamic biomarker time-series information provides continuous monitoring indicators of disease evolution. This multidimensional, multi-temporal data fusion enables the system to capture the disease characteristics and treatment responses of breast cancer patients from different levels, creating conditions for accurate risk prediction.

[0061] Example 2: See Figure 4When constructing a spatiotemporal constrained model for patient survival analysis based on clinical characteristic indicators, the collected multidimensional clinical data is first transformed into numerical feature representations suitable for machine learning model processing. Pathological staging data is converted through digital encoding using the TNM staging system, mapping primary tumor stages from T1 to T4 to continuous values ​​from 1 to 4. Lymph node metastasis statuses from N0 to N3 are also numerically processed, while distant metastasis statuses M0 and M1 correspond to binary variables of 0 and 1, respectively. Surgical procedures in treatment response records are assigned values ​​according to the degree of radicality: breast-conserving surgery is recorded as 1, modified radical mastectomy as 2, and extended radical mastectomy as 3. Chemotherapy regimens are graded according to the intensity of drug combinations: single-drug regimens are recorded as 1, dual-drug regimens as 2, and triple-drug regimens as 3. Radiotherapy doses are normalized according to the ratio of actual irradiation dose to standard dose. Endocrine therapy duration is directly recorded in months. Gene expression profile data is standardized, converting raw sequencing reads into log2-transformed relative expression levels. The mutation status of key genes is represented by 0 for wild type, 1 for heterozygous mutation, and 2 for homozygous mutation. Historical data on complications were graded by severity, with grade 1 being mild, grade 2 being moderate, and grade 3 being severe. A comprehensive complication index was calculated by weighting the severity based on the time of occurrence.

[0062] These preprocessed clinical feature indicators are organized into a spatiotemporal feature matrix, where each row represents the patient's state at a given time point, and each column represents a clinical feature dimension. The spatiotemporal constrained model is constructed using a graph neural network architecture, treating feature vectors from different time points as nodes in the graph, with edges between nodes representing state transitions over time. A time decay mechanism is introduced during model training, ensuring that recent features contribute more to the prediction results than longer-term features. For any two time points t... i and t j Time decay weight w ij The calculation uses an exponential decay function:

[0063]

[0064] Where: λ is the decay rate hyperparameter, controlling the decay rate influenced by historical characteristics, |t i -t j | indicates the time interval between two points in time. This time decay mechanism allows the model to adaptively focus on more timely changes in clinical features while retaining the reference value of historical features.

[0065] When extracting the collaborative boundary threshold for dynamic monitoring from the spatiotemporal constraint model, the gradient of change of each clinical feature dimension over time is first calculated. For feature f k The gradient value g at time t k(t) reflects the rate of change of this feature at the current moment. By analyzing the changing patterns of these gradient values ​​through a sliding time window, the sensitive intervals of feature change are identified. Cooperative boundary threshold θ k The threshold for a feature is determined using a feature-specific percentile method, where the 75th percentile of the historical gradient distribution for each feature is used as the monitoring threshold. When the current gradient value of a feature exceeds its corresponding threshold, it indicates that the feature is undergoing a significant change and requires enhanced monitoring.

[0066] The individualized monitoring frequency during patient follow-up is dynamically adjusted by a collaborative boundary threshold. The system monitors the changes in the gradient of each clinical feature in real time, and counts the number and extent of features exceeding the threshold. The monitoring frequency adjustment algorithm comprehensively considers three factors: the number of features exceeding the threshold, the magnitude of exceeding the threshold, and the importance weight of these features in the survival prediction model. High-risk feature combinations trigger high-frequency monitoring, while low-risk feature combinations maintain routine monitoring. In specific implementation, the system maintains a monitoring intensity index I. m Its value ranges from 0 to 1, corresponding to different monitoring frequencies from quarterly to weekly follow-up. This index is calculated by weighted summation of feature gradients exceeding the threshold and then normalized using the sigmoid function.

[0067] When compensating for missing features in a cross-modal medical data fusion model, the system first establishes a feature transmission path graph for the diagnosis and treatment process. Nodes in this directed graph model represent clinical examination items or diagnostic and treatment operations, and edges represent the sequential dependencies of data generation. For example, the pathological biopsy node points to pathological staging features, and the chemotherapy administration node points to drug metabolism features. Based on this transmission path graph, the system can identify the upstream data sources of missing features, providing a basis for probabilistic compensation. For a certain missing feature f... m Its compensation value f m The following is obtained through conditional probability estimation:

[0068]

[0069] Where: S represents f m The upstream feature set, P(f) m |s) represents f under the condition that the upstream feature s is known. m The conditional probability, v s These are the current observations of the upstream feature s. These conditional probabilities are obtained through statistical learning from historical data and are updated periodically to reflect the latest treatment patterns.

[0070] After generating a complete sequence of medical events, the system employs a multi-head attention mechanism for feature aggregation. Each clinical feature is mapped to a query vector, a key vector, and a value vector. By calculating the attention weights between features, the system identifies the most critical feature combinations for the current risk assessment. This attention mechanism adaptively captures dominant risk factors at different stages of diagnosis and treatment, such as focusing on complication features in the early postoperative period and on drug metabolism features during chemotherapy. The output of feature aggregation is a fixed-dimensional risk representation vector that comprehensively reflects the patient's current overall condition.

[0071] The risk entropy value during treatment is obtained by calculating the distribution uncertainty of the feature sequence. For each clinical feature dimension, its value changes at the three most recent time points are statistically analyzed, and the information entropy of that feature is calculated. The comprehensive risk entropy value H is obtained by weighted summation of all feature entropy values. r The weights are determined by the importance of each feature in survival prediction. A high-risk entropy value indicates that the patient's clinical status is undergoing complex and variable changes, requiring closer monitoring and intervention. The final generation process of the patient's clinical feature indicators integrates the original feature values, the compensated complete feature sequence, and the risk entropy value information, and integrates this information from different sources into a unified feature representation through a fully connected neural network.

[0072] The spatiotemporally constrained model is trained using right-censored survival data, with the optimization objective of accurately predicting patients' survival probabilities within specific future time windows. The model input is a spatiotemporal feature matrix, and the output is a survival probability curve. A partial likelihood loss function is used during training to account for differences in follow-up time among patients. The model is periodically incrementally trained using newly collected patient data to maintain the timeliness of predictive performance. The collaborative boundary threshold for dynamic monitoring is also adjusted with model updates to ensure that the monitoring strategy is synchronized with the latest clinical practices.

[0073] The implementation of individualized monitoring frequencies is presented to healthcare professionals through a clinical decision support interface. The system automatically generates a monitoring plan, listing suggested follow-up time points and key examination items. Suggestions for adjusting the monitoring frequency are accompanied by detailed explanations, highlighting key clinical changes that trigger adjustments. Healthcare professionals can accept the system's suggestions or make appropriate adjustments based on their actual situation; these human decision-making feedbacks are then recorded by the system for model optimization.

[0074] The quality assessment of missing feature compensation is conducted through simulated missing feature experiments. Features from a portion of the complete records are randomly selected and artificially set to be missing, and the differences between the compensated values ​​and the true values ​​are compared. Compensation error analysis helps identify conditional probability estimates that need improvement, especially for critical but frequently missing features. The update cycle of the compensation model is synchronized with the spatiotemporal constraint model to ensure the consistency of all components in the system.

[0075] The visualization of risk entropy information uses a heatmap format, displaying entropy changes for various clinical characteristics along a timeline. This intuitive presentation helps healthcare professionals quickly identify periods of instability in a patient's condition. Entropy peaks often correspond to periods of treatment changes or the occurrence of complications, requiring special attention. The system links these high-risk periods with adjustments to monitoring frequency, forming a closed-loop risk management process.

[0076] The dynamic updating mechanism of clinical characteristic indicators ensures that risk assessment is always based on the latest patient data. Each time a new examination result or medical record is entered, the system automatically triggers the characteristic update process, recalculating the risk entropy value and monitoring frequency. This real-time responsiveness enables the system to promptly detect subtle changes in clinical status, providing data support for precision medicine. The entire implementation process strictly adheres to medical data privacy protection standards, and all data processing is conducted in a secure computing environment.

[0077] Example 3: See Figure 2 The process of collecting clinical status records across different medical institutions is achieved through standardized medical data exchange interfaces. Medical institutions' information systems output medical data according to predefined format specifications, including structured information such as outpatient records, inpatient medical records, laboratory reports, and imaging results. Outpatient records contain key fields such as chief complaint, physical examination, preliminary diagnosis, and treatment recommendations, while inpatient medical records cover complete medical process documents such as admission records, progress notes, surgical records, and discharge summaries. Laboratory reports are output through the standard interface of the laboratory information system, containing core data elements such as test item name, result value, reference range, and test time. Imaging results are stored in DICOM format, along with structured report text recording examination findings and diagnostic opinions. These heterogeneous data sources are uniformly mapped to a common semantic model through the conversion layer of the medical information exchange platform, eliminating terminology differences between different institutions. The data collection process employs an incremental update mechanism, synchronizing only records added or modified since the last collection, reducing network transmission burden. The collected clinical status records are organized in chronological order. Duplicate records of the same patient from multiple institutions are merged and deduplicated to form a continuous and complete diagnosis and treatment timeline.

[0078] The identification process of the core treatment pathway employs a state transition-based analysis method. The system first defines key state nodes in the breast cancer treatment process, including standard stages such as initial assessment, pathological confirmation, multidisciplinary discussion, surgical decision-making, neoadjuvant therapy, surgical treatment, adjuvant therapy, and follow-up monitoring. Each state node is associated with a set of clinical characteristic conditions; when a patient's medical records meet these conditions, they are considered to have entered the corresponding state. The transition relationships between states constitute the basic framework of the treatment pathway, and specific path instances are obtained by analyzing the patient's actual medical record sequence. Hidden Markov Models (HMMs) are used to handle uncertainties in the treatment process, inferring the most likely state transition sequence by observing the patient's examination results and treatment records. Model training uses historical treatment data to learn the transition probabilities and observation probability distributions between different states. For each patient, the most likely treatment pathway is decoded using the Viterbi algorithm, serving as the patient's individualized treatment pathway. Each node in the pathway is labeled with a specific time point, executing agency, and clinical decision content, forming a structured description of the treatment process.

[0079] The process of generating treatment pathway information links the identified core workflow paths with detailed clinical records. The system constructs a time-axis-based treatment event graph, anchoring specific treatment events such as pathology reports, imaging examinations, and surgical records to corresponding nodes along the path. This association allows abstract path nodes to be mapped to specific treatment behaviors, providing substantial content for risk analysis. Treatment pathway information is stored in a graph data structure, where nodes represent treatment states or specific operations, edges represent state transitions or operation sequences, and edges are appended with attributes such as time intervals and transition conditions. This graph representation method facilitates path similarity calculations and pattern mining, supporting the analysis of patterns and variations in treatment behaviors from a population perspective.

[0080] The spatiotemporal correlation model between treatment node behavior and historical survival events employs a case-based reasoning approach. It maintains a knowledge base containing complete treatment paths and survival outcomes for historical patients. Treatment nodes for new patients are associated with historical cases through similarity matching. For each treatment node, the temporal distribution characteristics of its appearance in historical cases and the subsequent survival outcome distribution are calculated to establish a statistical correlation between node behavior and survival risk. The risk impact of nodes is quantified through two dimensions: the temporal dimension focuses on the impact of the timing of node appearance on prognosis, and the spatial dimension focuses on the node's positional relationship within the overall path. The spatiotemporal correlation model uses graph embedding technology to map nodes in the treatment path graph to a low-dimensional vector space, ensuring that similar node behaviors are close in distance within the vector space. This representation learning method can capture complex correlation patterns between treatment behaviors, going beyond simple statistical correlation.

[0081] Risk coupling strength analysis of treatment interventions focuses on key decision points in the treatment pathway. The system identifies important nodes influencing subsequent treatment outcomes, such as surgical approach selection, chemotherapy regimen formulation, and radiotherapy timing, and analyzes the differentiated impact of different choices at these nodes on survival risk. The calculation of risk coupling strength comprehensively considers direct and indirect effects. Direct effects refer to the inherent risk of the intervention itself, while indirect effects refer to the limitation or promotion of subsequent treatment choices by the intervention. For example, choosing breast-conserving surgery instead of total mastectomy may increase the risk of local recurrence, but may improve the patient's quality of life, thus indirectly affecting long-term survival. By analyzing outcome differences among patients with similar pathways in historical data, the system quantifies the risk coupling strength of each key decision point, providing a reference for current patients' treatment choices.

[0082] The risk impact feature extraction process for each treatment node employs a gradient-based analysis method. A differentiable survival risk prediction model is constructed, taking the patient's treatment pathway information as input and outputting a predicted survival risk score. By calculating the gradient of the model output with respect to each pathway node, the node features with the greatest impact on risk prediction are identified. These features include the node's temporal attributes, such as the interval between diagnosis and surgery; the node's content attributes, such as the thoroughness of surgery or the dosage intensity of chemotherapy; and the node's contextual attributes, such as the effectiveness of pre-treatment or the continuity of subsequent treatments. The extraction of risk impact features not only focuses on the static attributes of individual nodes but also emphasizes the dynamic role of nodes within the overall pathway, capturing the synergistic or antagonistic effects between treatment behaviors.

[0083] The risk association data generation for patient treatment nodes employs a multi-dimensional feature encoding method. Each treatment node is represented as a fixed-dimensional feature vector, encompassing multiple aspects such as basic attributes, temporal features, and risk impact features. Basic attributes record fundamental information such as the node's type, implementing agency, and operational content; temporal features describe the node's temporal position and interval characteristics within the overall path; and risk impact features summarize the node's various impact measures on survival risk. These feature vectors are fused and compressed using a neural network to generate the final risk association representation. The risk association data is organized in matrix form, with rows corresponding to nodes in the treatment path, columns corresponding to different risk-related features, and cell values ​​representing the quantitative assessment of a specific node on a specific feature dimension. This structured risk representation supports complex query and analysis operations, such as filtering key nodes by risk features or comparing risk distribution patterns across different paths.

[0084] The visualization of treatment pathway information uses a timeline combined with a Gantt chart. The patient's main treatment stages are arranged along the timeline, with the duration and key procedures of each stage displayed as color-coded bars, different colors representing different risk levels. High-risk nodes in the pathway are marked with prominent warning signs; clicking on them reveals detailed risk correlation data and analysis. This intuitive presentation helps healthcare professionals quickly grasp the overall picture of a patient's treatment and risk distribution, supporting the formulation and adjustment of treatment decisions. The system also provides a pathway comparison function, allowing for a visual comparison of the current patient's treatment pathway with typical pathways of patients with similar medical histories, highlighting key deviations from the norm and aiding in the identification of potential risk factors. All visualizations support interactive exploration, allowing healthcare professionals to delve into detailed clinical data and risk analysis results for specific nodes.

[0085] Example 4: See Figure 3 The determination of risk transmission constraints is based on the analysis of risk propagation patterns between nodes in the treatment pathway. The system constructs a directed acyclic graph model, where nodes represent treatment stages and edge weights reflect the intensity of risk transmission. This model is built by analyzing the correlation patterns between treatment event sequences and survival outcomes in historical data. For common treatment pathway patterns, such as "neoadjuvant chemotherapy → surgery → adjuvant chemotherapy" or "direct surgery → adjuvant radiotherapy → endocrine therapy," the system quantifies the risk changes during each stage transition. The intensity of risk transmission is calculated statistically, reflecting the degree of influence of risk factors from the previous stage on subsequent stages. For example, inadequate tumor regression after neoadjuvant chemotherapy may increase the risk of local recurrence after surgery; this risk transmission relationship is reflected in the model through edge weights. The final representation of risk transmission constraints is a set of rules and thresholds, defining under what conditions the risk of a certain stage will significantly affect the prognosis of subsequent stages.

[0086] The identification of anomalous features in tumor evolution relies on multi-scale analysis of time-series data of dynamic biomarkers. Cell-level anomalies focus on changes in microscopic indicators such as the number of circulating tumor cells and the frequency of specific gene mutations; tissue-level anomalies analyze the evolution of imaging features such as tumor density and boundary clarity; and organ-level anomalies monitor the appearance and growth of metastases. The system employs a time-series segmentation algorithm to divide continuous biomarker monitoring data into relatively stable stages and significant turning points. The granularity level of each anomalous event is determined based on its scope of influence and detection method. For example, de novo mutations in the BRCA gene are classified as cellular-level anomalies, blurred lesion margins on breast MRI are classified as tissue-level anomalies, and metastases detected by bone scans are classified as organ-level anomalies. These anomalous features of different granularities are organized chronologically to form a hierarchical event sequence of tumor evolution.

[0087] Table 1 shows the evaluation dimensions and grading standards for the survival risk status level. This evaluation system integrates information from two aspects: risk transmission constraints and the granularity of abnormal tumor evolution characteristics, and determines the final risk level through multidimensional decision rules.

[0088] Table 1: Dimensions for Evaluating Survival Risk Situation Level

[0089]

[0090] The flexible monitoring intensity adjustment during the treatment cycle is based on a dynamic balance between risk profile level and individualized monitoring frequency. The system maintains a monitoring intensity matrix, defining the recommended frequency range for each monitoring item under different risk levels. For example, in a very high-risk state, tumor marker testing is recommended weekly, and imaging examinations every two weeks; while in a low-risk state, the same tests may only need to be performed every two months or quarterly. The flexible adjustment mechanism allows for fine-tuning of the monitoring plan within the standard range according to the patient's specific situation, taking into account factors such as abnormalities in previous examinations, tolerance to treatment side effects, and patient compliance. The flexible range setting of monitoring intensity avoids overly rigid implementation standards, adapting to individual differences and special circumstances in clinical practice.

[0091] The risk exposure index for treatment blind spots is calculated using a multi-indicator weighted integral method, systematically identifying key time windows in the treatment process, such as the first two weeks after surgery, the intervals between chemotherapy cycles, and the three months after radiotherapy—periods traditionally under-covered by monitoring. Within each time window, a risk exposure score is calculated based on the activity level of risk transmission constraints and the accumulation of abnormal features. The score integrates three aspects of information: the intensity of historical risk factors transmitted to that time period, the number and severity of abnormal features actually observed during that time period, and the critical weight of that time period in the overall treatment plan. The calculation of the risk exposure index employs standardization, ensuring comparability of scores across different time periods and for different patients.

[0092] The early warning rule matching process employs a case-based reasoning method. The system's knowledge base stores the correlation patterns between monitoring intensity, risk exposure index, and final outcome of historical cases. The current patient's parameter combination is compared with historical cases through similarity matching to identify the closest cases, and early warning decisions are generated based on their outcomes. Early warning levels are divided into four levels: observation, alert, warning, and emergency, each corresponding to different response strategies. The early warning rules consider the cumulative effect over time; short-term high-risk exposure may trigger an observation-level warning, while long-term moderate-risk exposure may escalate to a warning-level warning. This dynamic matching mechanism can capture subtle differences in risk patterns, avoiding false alarms or missed alarms caused by simple threshold divisions.

[0093] The treatment blind spot early warning map is generated using spatiotemporal visualization technology. The timeline displays the entire treatment cycle, marking the standard monitoring plan and actual implementation status for each stage. A risk heatmap is overlaid on the timeline, using color intensity to represent the risk exposure index at different times. Early warning markers are indicated at the time points when the warning threshold is reached, with shape and color distinguishing the warning level. The map supports interactive exploration; clicking on a specific time period allows viewing detailed risk analysis and recommended countermeasures. The early warning map is updated dynamically; as new monitoring data is input and the risk situation evolves, the map adjusts the display of warning markers in real time.

[0094] The clinical application of risk profile grading is reflected in dynamic adjustments to treatment plans. The system generates recommendations for adjusting treatment intensity based on the current risk level, treatment guidelines, and experience from similar cases. For example, for patients persistently at extremely high risk, it may be recommended to intensify treatment or consider novel therapies in clinical trials; for patients with significantly reduced risk, it may suggest appropriately reducing treatment intensity to minimize side effects. These recommendations are accompanied by detailed supporting evidence, including risk factor analysis, historical case references, and treatment guideline guidelines, for clinicians to consider when making decisions.

[0095] A quality feedback mechanism for monitoring data ensures the reliability of risk analysis. The system tracks the execution of each monitoring check and the timeliness of the results report, and annotates the potential data gaps caused by delayed or missing checks. When key monitoring items are missing, the system adjusts the uncertainty range of the risk analysis and indicates the confidence level of the analysis results in the warning message. This transparent quality annotation helps healthcare professionals correctly understand the limitations of risk assessment and make reasonable judgments even with incomplete data.

[0096] The risk transmission constraint model is updated through a continuous learning mechanism. Newly accumulated patient diagnosis and treatment data and outcome information are periodically used to reassess the strength and correlation of risk transmission pathways. The model update process employs incremental learning, incorporating new evidence while maintaining the stability of existing knowledge. This dynamic update mechanism enables risk analysis to adapt to the evolution of clinical practice and the introduction of new treatment modalities, maintaining the timeliness and relevance of assessment results. After each model update, the system analyzes the differences in risk prediction between the old and new versions, identifies key changes in treatment modalities, and provides clues for improving clinical practice.

[0097] The evaluation of the treatment blind spot early warning system employs a prospective follow-up approach. The system records the trigger time, level, and response status of each early warning, tracks subsequent changes in patient status, and establishes a long-term evaluation mechanism for early warning effectiveness. By analyzing the accuracy, timeliness, and clinical response patterns of early warnings, the early warning rules and parameter settings are continuously optimized. The evaluation of early warning effectiveness not only focuses on the prediction of short-term risk events but also emphasizes the improvement of long-term survival outcomes, reflecting the long-term nature of breast cancer treatment. The evaluation results are used to guide iterative improvements to the system, forming a virtuous cycle where clinical practice and risk early warning mutually reinforce each other.

[0098] Example 5: The integration process of the survival decision-making module is achieved through multi-source data fusion technology. Individualized monitoring frequency data comes from the spatiotemporal constraint model output of the feature compensation module, including the time node arrangement of patient follow-up cycles and a list of key monitoring items. These monitoring plan data are organized in chronological order, annotating the expected examination content and clinical significance of each monitoring point. Survival risk status level data is generated by the risk analysis module, recording the current five-level risk assessment results, with detailed explanations of the classification basis, including risk transmission path analysis, tumor evolution characteristic description, and abnormal biomarker patterns. The treatment blind spot early warning map is received from the survival early warning module, including risk hotspot markers on the time axis, early warning level annotations, and a summary of response suggestions. These three types of data are aligned through a unified time benchmark to ensure that the information output by different modules remains synchronized in the time dimension.

[0099] The dynamic prediction report on relative survival risk for breast cancer is generated using a template-based document construction method. The report document is divided into six standard chapters: Patient Basic Information Summary, Current Risk Status Overview, Monitoring Plan Details, Risk Factor Analysis, Treatment Recommendations, and Long-Term Prognosis Prediction. The Patient Basic Information chapter integrates core data from the electronic medical record, including demographic characteristics, disease stage, molecular subtype, and a summary of treatment history. The Current Risk Status Overview section describes the risk level and key risk factors in concise language, using color-coded visual cues to highlight the risk level. The Monitoring Plan Details chapter lists the specific monitoring arrangements for the next three months, including the timeframe, program, implementing agency, and clinical significance, with special markings for key monitoring programs. The Risk Factor Analysis section provides an in-depth analysis of the factors contributing to the current risk level, presented in order of importance, with each factor accompanied by a time evolution curve and a comparison with the normal reference range.

[0100] The treatment recommendations section generates individualized intervention plans based on risk profile levels and early warning maps. Recommendations are divided into three categories: enhanced treatment measures for high-risk patients, including adjustments to treatment plans, multidisciplinary consultation recommendations, and strengthened supportive care; optimized monitoring strategies for intermediate-risk patients, focusing on improving follow-up plans and combinations of examinations; and lifestyle guidance for low-risk patients, emphasizing long-term health management and relapse prevention. Each recommendation is marked with its level of evidence, distinguishing between different levels of clinical evidence, such as explicit guideline recommendations, expert consensus support, and exploratory approaches. The long-term prognosis prediction section displays time-series curves of the probability of survival in the next year, accompanied by actual survival data from patients with similar medical histories for reference, while also explaining the range of uncertainty and key influencing factors of the predicted results.

[0101] The output format of the dynamic prediction report is designed to balance clinical practicality and system interaction needs. The PDF version adopts professional medical document formatting standards, including standardized headers and footers, chapter navigation, and medical institution identification, ensuring readability for both printing and electronic viewing. The structured data output in JSON format includes a machine-readable version of the complete report, employing standardized data fields and coding systems, and supports direct import into hospital information systems or clinical decision support platforms. Both output formats share the same content logic and timestamps, ensuring consistency between the different formats. Strict data validation is performed during report generation to ensure consistent units for numerical data, logical order of time data, and correct coding of categorical data.

[0102] The report transmission and storage mechanism adheres to medical information security standards. Generated predictive reports are transmitted via an encrypted channel to the attending physician's authorized terminal device, with a complete operation log recorded throughout the transmission process. A copy of the report is simultaneously stored in a designated area of ​​the patient's electronic health record, with access permissions controlled according to the medical institution's role-based management system. Each report update retains historical versions, creating a report revision trajectory and supporting retrospective analysis of the risk assessment's evolution. Upon report transmission, the system automatically triggers a notification mechanism, alerting relevant medical staff to review the new report via the medical institution's internal communication platform.

[0103] Clinical decision support features are embedded in the dynamic predictive reports. These reports include interactive elements such as clickable risk factor charts, expandable treatment recommendation details, and adjustable monitoring plan options. Physicians can perform simple interactive operations directly on a PDF reader or the hospital information system interface while reviewing the reports, such as confirming receipt, providing feedback, or adjusting the monitoring plan. This feedback is transmitted back to the system in real time for personalized adjustments to subsequent reports and system optimization. For key treatment recommendation options, the system provides a quick consultation channel, allowing physicians to initiate multidisciplinary consultation requests or consult relevant treatment guidelines directly from the report interface.

[0104] The report update mechanism is synchronized with the data updates of each module. When new monitoring data is input into the system, triggering recalculation in the feature compensation and risk analysis modules, the survival decision-making module automatically initiates the report update process. The update process employs an incremental revision strategy, modifying only the report sections affected by the new data to maintain overall report stability. When there are significant changes in risk levels, the system generates a report update description, highlighting the changes and their clinical significance. Minor adjustments resulting from routine monitoring data updates are indicated in a less conspicuous manner to avoid unduly disrupting clinical workflows.

[0105] The personalized report feature allows healthcare professionals with different specialties to customize how information is presented. Surgeons can highlight surgery-related risk factors and monitoring recommendations, oncologists can prioritize viewing assessments and adjustments to systemic treatment effectiveness, and radiation oncologists can focus on analyzing local control effects. This role-adaptive report view optimizes information retrieval efficiency for different specialties without altering the core content. Customized preferences are saved in user profiles and applied to all subsequently generated reports.

[0106] Quality control of the dynamic prediction report is achieved through a multi-layered verification mechanism. Content integrity checks ensure all standard sections and data elements are correctly included; logical consistency verification identifies potential contradictions between different sections; and numerical reasonableness audits identify possible outliers or data points outside the normal range. After report generation, it undergoes sampling review by medical professionals to confirm the accuracy of clinical interpretations and the applicability of recommendations. Issues discovered during quality control are fed back to the corresponding module's algorithm optimization process, forming a closed loop of continuous improvement.

[0107] The maintenance and upgrades of the reporting system follow strict version control guidelines. Each algorithm adjustment or template modification is logged in detail, explaining the content, reasons, and expected impact of the changes. Compatibility testing is performed before major updates to ensure a smooth transition between old and new reporting versions. The system retains the reporting generation logic of historical versions, supporting the retrospective review of evaluation standards for specific periods as needed. Version update information is marked in appropriate locations within the reports to help healthcare professionals understand potential differences in evaluation criteria across different periods.

[0108] The adaptive design for clinical applications enables the dynamic predictive report to be integrated into the workflows of various healthcare institutions. In large general hospitals, the report can serve as the basis for multidisciplinary discussions; in specialized oncology centers, it can be integrated into clinical pathway management systems; and in community healthcare institutions, a concise version of the report can serve as a referral reference. The system provides a report summary generation function, automatically extracting the most critical information to form a one-page summary, meeting the information delivery needs of different scenarios. Report usage is collected and analyzed anonymously to assess the system's clinical impact and identify opportunities for improvement.

[0109] The integration mechanism of long-term follow-up data enables the prediction reports to continuously calibrate their accuracy. The system periodically retrieves patients' subsequent medical records and survival status, comparing actual outcomes with historical predictions. These comparisons are used to calculate the calibration index of the prediction model and identify potential systematic biases. The results of the calibration analysis are fed back into subsequent reports in an appropriate format, helping healthcare professionals understand the uncertainties of the predictions and make more comprehensive clinical judgments. Over time, the accumulated follow-up data also becomes an important resource for model retraining, continuously improving the accuracy of predictions.

[0110] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine learning-based system for predicting relative survival risk in breast cancer, characterized in that, include: The data acquisition module is used to collect multidimensional clinical diagnosis and treatment records of breast cancer patients from distributed medical nodes and to obtain dynamic biomarker time-series information in the process of tumor evolution in real time. The feature compensation module is used to compensate for missing features in the multidimensional clinical diagnosis and treatment records under the cross-modal medical data fusion mode, generate patient clinical feature indicators, and determine individualized monitoring frequency based on the clinical feature indicators. The risk analysis module is used to acquire patient treatment path information in the medical network, analyze the survival risk characteristics of the treatment path information, output risk association data of patient treatment nodes, and determine the survival risk status level based on the risk association data and the time series information of the dynamic biomarkers. The survival early warning module is used to provide dynamic early warning for treatment blind spots based on the individualized monitoring frequency and the survival risk status level.

2. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that: The multidimensional clinical diagnosis and treatment records include a structured chain set of patient pathological staging, treatment response records, gene expression profiles, and historical data on complications.

3. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that: The dynamic biomarker time-series information represents time-series data on tumor marker concentrations, trends in imaging features, drug metabolism cycles, and the state of the immune microenvironment.

4. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that, Determining the individualized monitoring frequency based on the aforementioned clinical characteristic indicators specifically includes: A spatiotemporal constrained model for patient survival analysis was constructed based on the aforementioned clinical characteristic indicators; Extract the collaborative boundary threshold for dynamic monitoring from the spatiotemporal constraint model; The individualized monitoring frequency for the patient follow-up period is determined by the collaborative boundary threshold.

5. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that, Obtaining patient treatment pathway information from the healthcare network specifically includes: Real-time collection of clinical status records across institutional treatment nodes; The core diagnosis and treatment workflow is determined based on the aforementioned clinical status records; Patient treatment pathway information is generated through the core treatment workflow.

6. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that, The specific steps of analyzing the survival risk characteristics of the aforementioned treatment pathway information include: Based on the aforementioned treatment pathway information, a spatiotemporal correlation model between treatment node behavior and historical survival events is constructed. The spatiotemporal correlation model outputs the risk coupling strength of the treatment intervention; The risk impact characteristics of each diagnosis and treatment node are determined by the aforementioned risk coupling strength; Risk association data for patient diagnosis and treatment nodes is generated based on the aforementioned risk impact characteristics.

7. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that, Determining the survival risk profile level based on the aforementioned risk association data and the time-series information of the dynamic biomarkers specifically includes: Based on the aforementioned risk correlation data, risk transmission constraints during the treatment process are determined; The granularity of abnormal tumor evolution features is determined by the temporal information of the dynamic biomarkers. A survival risk profile level is generated based on the risk transmission constraints and the granularity of the abnormal tumor evolution features.

8. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that, Dynamic early warning of treatment blind spots based on the individualized monitoring frequency and the survival risk status level specifically includes: The intensity of elasticity monitoring within the treatment cycle is determined based on the individualized monitoring frequency. The risk exposure index of the treatment blind spot is determined based on the aforementioned survival risk status level; The elasticity monitoring intensity is matched with the risk exposure index using early warning rules to generate a dynamically adjustable treatment blind spot early warning map.

9. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that, The missing feature compensation for the aforementioned multidimensional clinical diagnosis and treatment records specifically includes: In the cross-modal medical data fusion model, the characteristic transmission path of the diagnosis and treatment process is determined based on the multidimensional clinical diagnosis and treatment records; The missing feature nodes in the multidimensional clinical diagnosis and treatment records are probabilistically compensated to generate a complete medical event sequence. The complete medical event sequence is subjected to feature aggregation, and the risk entropy value information during the treatment process is output. The patient's clinical characteristic indicators are generated from the risk entropy value information.

10. The machine learning-based system for predicting relative survival risk of breast cancer according to claim 1, characterized in that, The system also includes a survival decision module: This data is used to integrate the individualized monitoring frequency, the survival risk status level, and the treatment blind spot early warning map to generate a dynamic prediction report on the relative survival risk of breast cancer.

Images