Clinical index-based breast cancer metastasis risk prediction model
By integrating multimodal data and conducting dynamic risk assessments, combined with genomic characteristics and treatment response data, the accuracy and reliability of existing breast cancer metastasis risk prediction methods have been addressed, enabling precise identification and early warning of breast cancer metastasis risk.
Patent Information
- Application Number
- CN202511750764.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-24
AI Technical Summary
Existing methods for predicting the risk of breast cancer metastasis rely on a single data type, which cannot fully capture the biological characteristics of tumors, lacks a systematic analysis of the spatiotemporal evolution of tumors, and suffers from severe data silos, resulting in inaccurate and unreliable prediction results.
Establish a multimodal clinical data stream, collect imaging, pathological, and blood biomarker parameters, integrate the data through patient unique identifiers to form a time-series clinical dataset, conduct dynamic risk assessment and spatiotemporal pattern recognition, and perform multi-dimensional validation by combining genomic features and treatment response data.
It enables accurate identification and early warning of breast cancer metastasis risk, reduces the risk of false positives, provides more reliable clinical decision-making reference, and is suitable for long-term management of breast cancer diagnosis and treatment.
Smart Images

Figure CN121726080A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of breast cancer risk prediction, in particular to a breast cancer metastasis risk prediction model based on clinical indicators. BACKGROUND
[0002] As a malignant tumor with high morbidity and mortality among global female population, the metastatic property of breast cancer is the core factor leading to treatment failure and poor prognosis. Clinical practice shows that early and accurate identification of breast cancer metastasis risk can provide key direction for formulating individualized treatment plan and adjusting intervention strategy, thereby prolonging the survival period and improving the quality of life of patients. However, there are still many technical bottlenecks to be solved in the current field of breast cancer metastasis risk prediction.
[0003] Most of the existing prediction methods rely on a single type of clinical data, such as judging the tumor size and location only by imaging examination results, or evaluating the tumor differentiation degree only according to pathological histological parameters. The limitation of data dimensionality makes it difficult for the prediction model to fully capture the biological characteristics of the tumor. The occurrence and development of tumor is a dynamic evolution process, while the traditional model often uses static data for risk assessment, which cannot reflect the dynamic changes of tumor characteristics in the diagnosis and treatment cycle of patients, and is easy to miss the key risk evolution nodes.
[0004] In terms of risk identification mode, the existing technology lacks systematic analysis of the spatiotemporal evolution law of tumor, and most of them only focus on the abnormality of single time node indicators, which is difficult to identify the high-risk evolution mode with latency and gradualness, leading to lagging or inaccurate early warning signals. At the same time, in the risk verification link, the existing method often ignores the synergistic effect of genomics characteristics and treatment response data, and only makes judgments through limited clinical indicators, which makes the reliability and accuracy of the prediction results insufficient, and easy to produce false positive or false negative warning, which not only may lead to over-treatment and increase the burden of patients, but also may delay the best treatment opportunity due to missed judgment.
[0005] Clinical data is scattered in different diagnosis and treatment systems, and imaging data, pathological data, blood test data and the like lack effective correlation and integration mechanism, and the data island phenomenon is serious, which further limits the performance improvement of the prediction model. Therefore, it is an urgent need in the current clinical tumor field to develop a breast cancer metastasis risk prediction model which can integrate multi-modal dynamic data, accurately identify risk evolution mode and perform multi-dimensional verification. SUMMARY
[0006] The purpose of the present application is to provide a breast cancer metastasis risk prediction model based on clinical indicators to solve the problems raised in the background.
[0007] To achieve the above purpose, the present application provides a breast cancer metastasis risk prediction model based on clinical indicators, which comprises: establishing a multi-modal clinical data stream, collecting imaging feature parameters, pathological histological parameters and blood biomarker parameters in the diagnosis and treatment cycle of patients, correlating and integrating the data through a unique patient identifier to form a time series clinical data set; performing dynamic risk assessment on the time series clinical data set, extracting tumor biological feature indicators at each time node, calculating a tumor aggressiveness index, and generating a dynamic risk change sequence; performing spatiotemporal pattern recognition based on the dynamic risk change sequence, analyzing the evolution law of tumor characteristics between different time nodes, identifying high-risk evolution patterns, and constructing a risk evolution pattern library; performing pattern matching on real-time clinical data using the risk evolution pattern library, detecting clinical data points that match pre-set high-risk patterns, and generating a preliminary risk warning set; performing multi-dimensional verification on the preliminary risk warning set, combining genomic features and treatment response data, and outputting the final breast cancer metastasis risk prediction result.
[0008] Preferably, the step of establishing a multi-modal clinical data stream comprises: obtaining baseline clinical data of breast cancer patients from a hospital information system, including diagnosis age, tumor stage, and histological grading data; collecting imaging examination results of regular follow-up during treatment, recording tumor size, morphological characteristics, and enhancement degree quantitative parameters; collecting blood specimen detection data at each follow-up, including tumor marker levels, circulating tumor cell counts, and inflammation indicator parameters; correlating and integrating clinical data from different sources through patient hospitalization number or social security number, and arranging them in chronological order to form a complete time series clinical data set.
[0009] Preferably, the step of performing dynamic risk assessment on the time series clinical data set comprises: extracting tumor maximum diameter, number of lymph node metastasis, and Ki-67 proliferation index at each follow-up time point; calculating tumor volume growth rate, marker level change rate, and other dynamic change parameters; combining patient age and underlying disease conditions, and calculating tumor aggressiveness index at each time point using a weighted scoring method; arranging tumor aggressiveness indexes at each time point in chronological order to form a dynamic risk change sequence.
[0010] Preferably, the step of performing spatiotemporal pattern recognition based on the dynamic risk change sequence comprises: analyzing the trend characteristics in the dynamic risk change sequence, identifying the risk rapid rise period, plateau period, and decline period; detecting the synergistic change pattern of tumor characteristics between different time nodes, including the synchronous change relationship between size and marker; establishing a risk evolution pattern recognition algorithm to extract high-risk feature patterns, including sustained rapid rise pattern, fluctuation rise pattern, etc.; storing the identified feature patterns and corresponding clinical index change rules into the risk evolution pattern library.
[0011] Preferably, the step of using the risk evolution pattern library to perform pattern matching on real-time clinical data comprises: inputting newly collected patient clinical data and extracting tumor characteristic parameters at the current time point; combining the current data with historical data to form the latest time series segment; performing similarity calculation on the time series segment and the feature patterns in the risk evolution pattern library; screening patterns with similarity exceeding a set threshold and marking the corresponding clinical data points as high-risk candidate points; summarizing all high-risk candidate points to generate a preliminary risk warning set.
[0012] Preferably, the step of performing multi-dimensional verification on the preliminary risk warning set comprises: obtaining complete clinical history data of high-risk candidate points, including past treatment history and complication conditions; supplementing genomic characteristics, including gene mutation status and gene expression profile characteristics; analyzing treatment response data to evaluate the effectiveness and tolerance of past treatment; establishing a multi-factor verification model to comprehensively analyze clinical characteristics, genomic characteristics, and treatment response data; excluding false positive risk points and confirming true metastasis high-risk patients.
[0013] Preferably, the step of calculating the tumor invasiveness index at each time point using a weighted scoring method comprises: constructing a tumor invasiveness evaluation index system, which includes tumor burden dimension, proliferation activity dimension, and systemic reaction dimension; assigning initial weights to specific indicators under each dimension, wherein the maximum diameter of the tumor and the number of lymph node metastasis belong to the tumor burden dimension, the Ki-67 proliferation index belongs to the proliferation activity dimension, and the tumor marker level and inflammation indicators belong to the systemic reaction dimension; based on the correlation of each indicator with the final metastasis outcome in historical patient data, iteratively optimizing the initial weights to obtain the optimal weight combination; According to the optimal weight combination, the index data collected at each time point is weighted and summed to calculate a quantitative tumor invasiveness index.
[0014] Preferably, the step of establishing a risk evolution pattern recognition algorithm comprises: After the dynamic risk change sequence is normalized, a one-dimensional convolutional neural network is input to extract local features and capture risk fluctuation patterns in the short term; The normalized sequence is analyzed by using a long short-term memory network to learn long-term time-dependent relationships and trend evolution laws; The local features output by the one-dimensional convolutional neural network and the sequence context features output by the long short-term memory network are fused to form a comprehensive spatiotemporal feature vector; The spatiotemporal feature vector is input into a fully connected layer and a flexible maximum value classification layer to identify and classify a plurality of predefined high-risk evolution patterns.
[0015] Preferably, after the step of combining the current data with the historical data to form the latest time sequence segment, the step of performing data enhancement on the time sequence segment is further included: The time sequence segment is sampled by a sliding window to generate a plurality of time sub-sequences with overlapping parts; Random noise is applied to each time sub-sequence or data micro-fluctuation is simulated by an interpolation method to expand the training samples; The dynamic time warping distance of each enhanced time sub-sequence and the feature patterns in the risk evolution pattern library is calculated as a similarity measure; The first K patterns with the smallest dynamic time warping distance are selected as the matching results, and the corresponding clinical data points are included in the preliminary risk warning set.
[0016] Preferably, the step of establishing a multi-factor verification model for joint analysis comprises: A multi-source feature input layer is constructed to receive multi-dimensional features from clinical features, genomics features and treatment response data; A feature cross layer is designed to calculate the interaction weights between different source features by an attention mechanism to generate cross features with high-order correlations; The original features and the cross features are spliced and input into a survival analysis layer based on the Cox proportional hazards model to calculate the metastasis risk score and risk ratio of the patient; According to the risk score threshold, the patients in the preliminary risk warning set are reordered and confirmed to output the final risk prediction results with different risk levels.
[0017] Compared with the prior art, the present application has the following advantages: The model breaks through the limitation of traditional models relying on single data type, systematically collects imaging feature parameters, pathological histology parameters and blood biomarker parameters, realizes deep correlation and integration of data from different sources with the help of unique patient identifier, and forms a complete time series clinical data set.
[0018] The introduction of dynamic risk assessment mechanism changes the inherent defects of traditional static assessment. The model is no longer limited to index analysis at a fixed time node, but extracts tumor biological feature indexes at each time node to generate a dynamic risk change sequence through calculation of tumor invasiveness index. This process can capture the dynamic evolution trajectory of the tumor in the treatment and progression process in real time, clearly present the fluctuation rule of risk level, so that the clinician can discover the potential trend of tumor progression in time, rather than intervening after the risk has appeared. This dynamic monitoring mode is especially suitable for the long-term management process of breast cancer diagnosis and treatment, and can provide timely basis for dynamic adjustment of treatment plan, making the intervention measures more targeted.
[0019] The construction of spatiotemporal pattern recognition and risk evolution mode library further improves the forward-looking and accuracy of risk prediction. The model analyzes the evolution rule of tumor characteristics between different time nodes in depth, accurately identifies high-risk evolution modes and systematically summarizes them to form a structured risk evolution mode library. When processing real-time clinical data, the data points that meet the high-risk mode can be quickly located through pattern matching to realize rapid capture of risk signals. This early warning method based on pattern recognition can break through the limitations of traditional index threshold judgment, identify risk features with latency and correlation, greatly reduce the possibility of risk omission, and make the early warning signal more early and reliable.
[0020] The setting of multi-dimensional verification link provides double protection for the reliability of the prediction results. After generating the preliminary risk warning set, the model does not directly output the results, but conducts secondary verification combined with genomics features and treatment response data. Genomics features can reveal the genetic background and invasive potential of tumors at the molecular level, and treatment response data reflects the reaction characteristics of tumors to clinical intervention, which, combined with clinical indicators, forms a more comprehensive verification system. This multi-dimensional verification can effectively filter out false positive warnings caused by single index abnormalities, and makes up for the lack of analysis at the molecular mechanism level of traditional models, so that the final output prediction results are more in line with the actual progression of tumors, providing a more reliable reference for clinical decision-making. BRIEF DESCRIPTION OF DRAWINGS
[0021] Figure 1 The working principle diagram of the breast cancer metastasis risk prediction model based on clinical indicators described in the present application; Figure 2To establish a flow chart of multi-modal clinical data stream; Figure 3 To establish a flow chart of dynamic risk assessment; Figure 4 To establish a breast cancer risk trend phase analysis chart; Figure 5 To establish a risk pattern matching and early warning analysis chart. DETAILED DESCRIPTION
[0022] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0023] Please refer to Figure 1 The present application provides a breast cancer metastasis risk prediction model based on clinical indicators, which comprises the following steps: establishing a multi-modal clinical data stream, collecting imaging feature parameters, pathological histological parameters and blood biomarker parameters in the diagnosis and treatment cycle of a patient, correlating and integrating the data through a unique patient identifier to form a time series clinical data set; performing dynamic risk assessment on the time series clinical data set, extracting tumor biological feature indicators at each time node, calculating a tumor aggressiveness index, and generating a dynamic risk change sequence; performing spatiotemporal pattern recognition based on the dynamic risk change sequence, analyzing the evolution law of tumor characteristics between different time nodes, identifying high-risk evolution patterns, and constructing a risk evolution pattern library; using the risk evolution pattern library to perform pattern matching on real-time clinical data, detecting clinical data points that meet the preset high-risk patterns, and generating a preliminary risk early warning set; performing multi-dimensional verification on the preliminary risk early warning set, combining genomics characteristics and treatment response data, and outputting the final breast cancer metastasis risk prediction result.
[0024] Embodiment 1: Please refer to Figure 2In specific implementations, the hospital information system serves as the data source, providing baseline clinical data of breast cancer patients, including diagnostic age, tumor stage, and histological grade data. These data are extracted from electronic health records, batch exported through standardized data interfaces, ensuring information integrity and format consistency. Diagnostic age records the age of the patient at the time of first diagnosis, tumor stage is coded according to the TNM staging system of the International Union Against Cancer, and histological grade is divided into I to III levels according to the World Health Organization standard. All data items are reviewed and confirmed by clinicians to eliminate input errors or ambiguities. In some embodiments, the hospital information system integrates multiple subsystems, such as the pathology report system and the radiology information system, automatically matching related records through the patient's unique identifier, avoiding delays or omissions caused by manual retrieval. The data extraction process uses a timing task or trigger mechanism that automatically triggers the collection process when a new patient is admitted or existing data is updated, ensuring real-time and continuous data flow. The collected baseline data is stored in a specific table in the central database, with a timestamp and source identifier.
[0025] In specific implementations, the results of periodic follow-up imaging during treatment are continuously collected, focusing on recording quantitative parameters such as tumor size, morphological characteristics, and enhancement degree. These parameters are automatically or semi-automatically extracted from mammography, ultrasound, or magnetic resonance imaging through professional image processing software. Tumor size measurement is based on the maximum diameter in the image, morphological feature evaluation includes tumor boundary sharpness, internal echo uniformity, and other indicators, and enhancement degree is calculated by early enhancement rate and peak enhancement value through dynamic contrast-enhanced magnetic resonance time-signal intensity curve. Imaging results are collected every 3 to 6 months, corresponding to the patient's standardized follow-up plan, and a structured report is generated for each examination. Parameter values are stored in numerical or categorical form for subsequent calculation and analysis. It can be understood that image parameter extraction needs to consider device differences and operator variability, so calibration protocols and unified measurement guidelines are adopted, such as regular quality control tests for all imaging devices, measurements are completed independently by two or more radiologists, and significant differences are arbitrated by experienced doctors. The data collection module is integrated into the hospital workflow, and the results are automatically pushed to the risk prediction system after the examination is completed, reducing manual intervention and improving efficiency and data accuracy.
[0026] In specific implementations, blood specimen testing data at each follow-up visit are synchronously collected, covering tumor marker levels, circulating tumor cell counts, and inflammatory indicator parameters, which are obtained through standardized testing procedures of clinical laboratories. Tumor markers such as CA15-3 or CEA are determined using chemiluminescent immunoassay, circulating tumor cell counts are enumerated by microfluidic chip technology based on epithelial cell adhesion molecules, and inflammatory indicators such as C-reactive protein or interleukin-6 are detected by immunoturbidimetry or enzyme-linked immunosorbent assay. Blood specimens are collected on the day of follow-up visit, and processing and analysis follow the laboratory operation specifications. The results are recorded in numerical form, with reference ranges and units, and abnormal values are automatically marked for review. In some embodiments, blood testing data are arranged synchronously with imaging examinations to ensure time node alignment and reduce bias introduced by collection time differences. Data are automatically imported through laboratory information system interfaces and merged with baseline data. Optionally, for special patient groups such as individuals with high-risk family history, circulating tumor DNA testing is added as a supplementary parameter, but the core data set still mainly consists of the above standard parameters to maintain the universality and comparability of the model.
[0027] In specific implementations, clinical data from different sources are associated and integrated through patient hospitalization numbers or social security numbers, which serve as unique identifiers to establish association relationships at the database level. The data integration process includes data cleaning, conversion, and loading steps. The cleaning stage handles missing values, abnormal values, or format inconsistencies, such as missing tumor sizes estimated by interpolation or marked as to be supplemented, and abnormal values such as marker levels exceeding physiological ranges are corrected or removed after clinical confirmation. The conversion stage unifies data from different sources into the same time granularity and units, such as aligning all time points to monthly intervals and standardizing parameter units to the International System of Units. The loading stage arranges the processed data in chronological order to form a complete time series clinical data set. It can be understood that the use of unique identifiers is crucial to ensure that multiple follow-up data of the same patient are correctly linked and avoid misidentification. The integrated data set is stored in the form of a relational database table, with each row representing the multi-modal data at a time point and each column corresponding to a specific parameter, with time stamps and patient identifiers as indexes. Optionally, the data integration module supports incremental updates, automatically appending new follow-up data to the existing data set and triggering consistency checks such as time sequence verification or numerical logic verification to ensure the time sequence integrity and internal consistency of the data set. In specific implementations, the final form of the time series clinical data set is a structured table containing patient identifiers, time point sequences, and corresponding multi-modal parameter values. This data set serves as input for subsequent dynamic risk assessment. Data storage uses columnar databases to optimize query performance, supporting fast retrieval and time series analysis, while access permissions and encryption measures are set to protect patient privacy.
[0028] Embodiment 2: see Figure 3In specific implementations, key indicators such as tumor maximum diameter, number of lymph node metastasis, and Ki-67 proliferation index at each follow-up time point are extracted. The tumor maximum diameter is directly read from the structured report of imaging examination, with the unit unified as millimeters. The number of lymph node metastasis is derived from the pathology report or clinical examination record, recorded as the number of positive lymph nodes. The Ki-67 proliferation index is evaluated by immunohistochemical staining, expressed as the percentage of positive cells. The extraction of these key indicators is automatically completed by analyzing the corresponding fields in the time series clinical data set. The extraction process includes data validity check, such as tumor maximum diameter value greater than zero and within a reasonable medical range, number of lymph node metastasis as a non-negative integer, and Ki-67 proliferation index percentage between 0 and 100. Any abnormal value or missing value will be marked and trigger a review process to ensure the quality and reliability of the input data. It can be understood that the selection of key indicators is based on their known strong correlation with tumor biological behavior, and the extraction operation is performed by system script in batches to avoid errors and inconsistencies introduced by manual operation.
[0029] In specific implementations, dynamic change parameters such as tumor volume growth rate and marker level change rate are calculated. The tumor volume growth rate is estimated based on the tumor maximum diameter at consecutive time points. Assuming the tumor shape is approximately spherical, the volume is derived from the diameter using the spherical volume formula, and then the volume change percentage between adjacent follow-up time points is calculated and divided by the time interval. The marker level change rate is calculated for tumor markers such as CA15-3, and the difference in concentration between adjacent time points is divided by the time interval to obtain the change amount per unit time, with the time interval standardized in months. The calculation of dynamic change parameters requires at least two consecutive time points of data. The calculation script automatically identifies valid time series segments. For irregular intervals, linear interpolation is used for preprocessing to standardize the time scale. The calculation results are added as new derived fields to the data set. In some embodiments, for cases with multiple lesions, the tumor volume growth rate is calculated based on the main lesion or the largest lesion to ensure consistency of the evaluation focus. Dynamic change parameters provide a quantitative perspective of the evolution of indicators over time, making up for the static limitations of single time point indicators.
[0030] In specific implementations, the tumor aggressiveness index at each time point is calculated using a weighted scoring method, which is implemented in conjunction with patient age and underlying disease conditions. The weighted scoring method is implemented by constructing a tumor aggressiveness assessment index system. The tumor aggressiveness assessment index system includes three dimensions: a tumor burden dimension, a proliferation activity dimension, and a systemic response dimension. The tumor burden dimension encompasses the maximum tumor diameter and the number of lymph node metastases. The proliferation activity dimension includes the Ki-67 proliferation index. The systemic response dimension involves tumor marker levels and inflammatory indicators such as C-reactive protein. Each specific indicator of the tumor aggressiveness assessment index system is assigned an initial weight. The initial weight is based on clinical consensus or commonly used values in the literature. For example, the maximum tumor diameter, the number of lymph node metastases, the Ki-67 proliferation index, tumor marker levels, and inflammatory indicators can be assigned initial scores that reflect their relative importance. Based on the correlation of each indicator with the final metastasis outcome in historical patient data, the initial weights are iteratively optimized to obtain an optimal weight combination. Historical patient data refers to a labeled data set containing complete time series data for a large number of past patients and the final metastasis outcome. The iterative optimization process uses a logistic regression or machine learning algorithm, with the metastasis outcome as the dependent variable and the indicator values at each time point as the independent variables. The weights are repeatedly adjusted using optimization algorithms such as gradient descent until the model prediction performance is optimal, thereby determining the optimal weight combination. It can be understood that the determination of the optimal weight combination is a data-driven process that relies on high-quality historical data sets. Cross-validation is required during the weight optimization process to prevent overfitting and ensure the robustness and generalizability of the optimal weight combination.
[0031] In specific implementations, according to the optimal weight combination, the indicator data collected at each time point is weighted and summed to calculate the quantitative tumor aggressiveness index. The weighted sum operation is performed on all valid indicators at a single time point. Each indicator value is multiplied by the corresponding optimal weight and then added up to obtain the original score at that time point. Subsequently, the original score is normalized, for example, using the min-max normalization method to linearly transform it to the interval of 0 to 1, where 0 represents the lowest aggressiveness and 1 represents the highest aggressiveness. The calculation of the tumor aggressiveness index is fully automated. The system reads the time point data, loads the optimal weight combination, performs the calculation, and outputs the results. The calculation process records logs for auditing. In some embodiments, if a specific indicator at a certain time point is missing, the weighted sum can use the interpolated value or estimate based on the existing indicators to maintain the continuity of the time series, but the data point will be marked as an estimated value. Optionally, after the calculation of the tumor aggressiveness index is completed, the original indicator values are stored together, and a data quality report is generated, listing the number of time points for which the calculation is successful, the missing value situation, and any abnormalities in the calculation process.
[0032] In specific implementations, the tumor aggressiveness indices at each time point are arranged in chronological order to form a dynamic risk change sequence. The dynamic risk change sequence is an ordered list or array, with each element containing a timestamp and the corresponding tumor aggressiveness index value, strictly following the chronological order of the follow-up time points. The dynamic risk change sequence is stored and managed on a patient-by-patient basis, usually using a time series database or special fields in a relational database to support efficient time range queries and sequence analysis operations. The formation of the dynamic risk change sequence marks the transition from static index evaluation to dynamic trend analysis, and this sequence directly displays the fluctuations in tumor aggressiveness of the patient throughout the entire diagnosis and treatment cycle, providing core input data for subsequent spatiotemporal pattern recognition. It can be understood that the accuracy and completeness of the dynamic risk change sequence directly depend on the quality of the previous data extraction, parameter calculation, and index weighting, and any errors in any step will accumulate in the sequence, so quality control at each step is crucial. Optionally, the system can provide a dynamic risk change sequence visualization function to generate a line chart of the tumor aggressiveness index over time, assisting clinicians in intuitively grasping the risk trend.
[0033] In specific implementations, the trend characteristics in the dynamic risk change sequence are analyzed to identify the rapid risk rise period, plateau period, and decline period, and the trend characteristic analysis uses a sliding window average method combined with linear fitting. The size of the sliding window is dynamically set according to the follow-up interval, for example, for quarterly follow-up data, the window size is set to 3 time points, and the moving average of the tumor aggressiveness index within the window is calculated to smooth short-term fluctuations. In each window, linear regression is performed on the tumor aggressiveness index sequence, and the slope of the fitted straight line is used as a quantitative indicator of the trend. The rapid risk rise period is defined as a continuous time period where the slope value continuously exceeds a pre-set positive threshold (such as +0.05 / month), the plateau period is defined as an interval where the absolute value of the slope is less than a smaller threshold (such as ±0.01 / month), and the decline period corresponds to a phase where the slope continuously falls below a negative threshold (such as -0.03 / month). The identified periods are labeled with their start and end time points, and the duration and average slope are recorded, which together constitute a description of the macro trend of the sequence. In some embodiments, in addition to linear fitting, quadratic fitting or more complex models are also used to capture the trend changes of the curve, but the core judgment is still based on the high-low interval division of the change rate.
[0034] In specific implementation, the synergistic change pattern of tumor characteristics between different time nodes is detected, including the synchronous change relationship of size and marker, and the analysis of synergistic change pattern mainly relies on the calculation of correlation coefficient and time lag correlation between different index sequences. The synchronous change relationship is evaluated by calculating the Pearson correlation coefficient between the maximum diameter of tumor and the level of tumor marker at the same time point or within a small time difference, and a strong positive correlation coefficient indicates that the two change in the same direction. Time lag correlation analysis is used to explore whether the change of one index leads or lags behind the change of another index, for example, the correlation coefficient of the value of tumor maximum diameter at time t and the value of tumor marker level at time t+Δt is calculated, and the time lag point of maximum correlation is found by changing Δt. This analysis helps to find patterns with clinical significance such as "tumor enlargement followed by marker elevation" or "marker elevation precedes visible tumor enlargement in imaging". It can be understood that the detection of synergistic change pattern requires long enough and uniformly sampled time series data, and for cases with irregular follow-up intervals, data resampling may be needed to align the time points before correlation calculation.
[0035] In specific implementation, a risk evolution pattern recognition algorithm is established to extract high-risk feature patterns, including sustained rapid rise pattern, fluctuation rise pattern, etc. The core of the risk evolution pattern recognition algorithm is a deep learning architecture combining one-dimensional convolutional neural network and long short-term memory network. The first step of the algorithm is to normalize the dynamic risk change sequence, usually using Z-score standardization to make the sequence mean 0 and standard deviation 1, to eliminate the influence of dimension and accelerate model convergence. The normalized sequence is input into two branches in parallel for processing, the first branch is a one-dimensional convolutional neural network for local feature extraction, the one-dimensional convolutional neural network uses multiple convolution kernels of different widths to perform sliding convolution operation on the sequence, and the activation function uses ReLU to capture short-term risk fluctuation patterns such as sudden rise, sudden drop or short-term plateau. The convolution operation can be represented as: wherein: is the feature map output at the th position of the th layer, is the weight of the th convolution kernel at the th position of the th layer, is the convolution kernel size, is the input at the th position of the previous layer, The bias term of the layer. The second branch is a long short-term memory network, which is used to analyze the normalized sequence, learn long-term temporal dependencies and trend evolution laws. The long short-term memory network controls the flow of information and the update of memory through its internal gating mechanism (input gate, forget gate, output gate), thereby effectively capturing long-term context information in the sequence. The local feature maps output by the one-dimensional convolutional neural network are flattened and fused with the hidden state features of the last time step of the long short-term memory network to form a comprehensive spatio-temporal feature vector. The fused spatio-temporal feature vector is then input to a fully connected layer for non-linear transformation, and finally a flexible maximum value classification layer is used to output the probability of belonging to each predefined high-risk evolution mode. The predefined modes include but are not limited to the sustained rapid rise mode (sustained high positive slope) and the fluctuating rise mode (overall rise with frequent fluctuations), each of which has its corresponding clinical significance. In some embodiments, the risk evolution pattern recognition algorithm is trained on a large historical patient data set using a cross-entropy loss function and an Adam optimizer, and the network parameters are updated through a backpropagation algorithm. The training goal is to enable the model to accurately classify the input dynamic risk change sequence into the correct risk evolution mode.
[0036] In specific implementations, the identified feature patterns and corresponding clinical indicator change laws are stored in a risk evolution pattern library, which is usually constructed using a relational database. Each record in the database corresponds to an identified feature pattern, and the record fields include but are not limited to: a unique pattern identifier, a pattern name, a pattern description, typical morphological features of the corresponding dynamic risk change sequence, clinical characteristic statistical information of the patient cohort in which the pattern is identified, and significant clinical indicator change laws associated with the pattern. The clinical indicator change law refers to the trend of coordinated changes that other indicators such as tumor size and marker levels usually exhibit when the pattern appears, which is statistically derived from the training data. The construction of the risk evolution pattern library is a continuous process, and as new data accumulates and the model is retrained, new patterns can be continuously added to the library or existing patterns can be updated. It can be understood that the richness and accuracy of the risk evolution pattern library directly determine the effectiveness of the subsequent pattern matching step, so the patterns in the library need to be verified and optimized regularly. Optionally, the risk evolution pattern library can provide a query interface to allow users to search and browse based on pattern characteristics or clinical indicators, and visualize the typical sequence morphology of the pattern. In some embodiments, for each pattern stored in the library, the frequency of occurrence and predictive value in different patient subgroups are also recorded to provide a reference for precision medicine.
[0037] In specific implementation, the entire process of spatiotemporal pattern recognition is driven by an automated workflow, starting from reading the dynamic risk change sequence, to finally storing the new pattern, the intermediate results and model decision logic generated during the process are recorded to ensure the traceability and reproducibility of the process. The performance of the risk evolution pattern recognition algorithm is evaluated by accuracy, recall rate and other indicators on an independent test set to ensure its generalization ability.
[0038] Referring to Figure 4 , the results of the trend feature analysis method based on dynamic risk change sequence are shown. The chart conducts in-depth analysis on the time series data of tumor aggressiveness index through the method of sliding window average combined with linear fitting, and identifies key trend stages such as rapid risk rising period, plateau period and declining period. The chart contains the complete tumor aggressiveness index change curve and the trend slope line calculated by linear regression. The background areas of different colors clearly identify the risk stages, where the red area represents the rapid risk rising period, the green area represents the relatively stable plateau period, and the blue area represents the risk declining period. The chart also shows the preset threshold limits, including the rising threshold, the stable threshold and the declining threshold, which are key judgment criteria set based on clinical experience. The data points in the chart reflect the tumor aggressiveness state of the patient at different follow-up time points, and the change of the trend line reveals the dynamic law of risk evolution. Through this analysis method, clinicians can clearly understand the evolution trajectory of the patient's condition and timely detect the early warning signal of rapid risk rising, providing an important basis for developing personalized treatment plans.
[0039] In specific implementation, input the newly collected patient clinical data, extract the tumor feature parameters at the current time point, the newly collected data comes from the real-time output of hospital information system, imaging examination and laboratory detection, the data format and standard are consistent with the historical data used when constructing the risk evolution pattern library. The extracted tumor feature parameters include tumor maximum diameter, number of lymph node metastasis, Ki-67 proliferation index, tumor marker level, etc., the extraction process is automatically completed through pre-defined data parsing rules, such as reading the value of tumor maximum diameter from the image report, extracting the percentage of Ki-67 proliferation index from the pathology report. After extraction, the system will perform data verification to check whether the parameter value is within a reasonable physiological range, and mark the missing items, only the parameters that pass the verification will be used in the subsequent pattern matching process, ensuring that the quality of the input data is the basis for the reliability of the subsequent analysis. It can be understood that the automation of real-time data extraction is the key to efficient risk early warning, it reduces manual intervention, improves response speed, and enables clinicians to obtain risk assessment results in a timely manner.
[0040] In specific implementations, the current data is combined with historical data to form the latest time series segment, the historical data refers to the multi-modal clinical data collected at all previous follow-up time points of this patient and stored in the time series clinical data set. The combination operation retrieves the complete historical record of this patient from the database based on the patient's unique identifier, and then appends the newly collected data at the current time point as a new data point to the end of the historical sequence in chronological order. The formed time series segment contains all data from the first follow-up to the current latest time point, which fully reflects the dynamic evolution of the patient's condition and is the direct input object for pattern matching. The system automatically maintains the time series segment for each patient to ensure its continuity and integrity, and dynamically updates the time series segment when new follow-up data is added to support continuous risk monitoring. In some embodiments, the length of the time series segment may be affected by the length of the patient's follow-up time, for newly diagnosed patients, the segment may be shorter, and the system will set a minimum length threshold, only when the segment length reaches the threshold, the pattern matching will be started, to avoid false positives due to insufficient data.
[0041] In specific implementations, the time series segment is subjected to data augmentation, which aims to artificially increase the diversity of the time series segment and improve the robustness of the pattern matching algorithm. The specific method of data augmentation includes sliding window sampling of the time series segment to generate multiple time sub-sequences with overlapping parts, the width of the sliding window is set according to the needs of pattern recognition, for example, it is set to contain the last 6 follow-up time points, and the sliding step is set to 1 time point, thereby generating a series of continuous and overlapping sub-sequences. For each time sub-sequence obtained by sliding window, further random noise or interpolation method is applied to simulate small data fluctuations, to expand the training sample, the random noise usually follows a Gaussian distribution with a mean of zero and a small variance, and the amount of added noise is controlled to simulate the small errors or physiological fluctuations that may exist in the data collection process in the real world, without masking the true trend. The interpolation method is used to insert new simulated data points between existing time points, thereby generating denser time series, this method is particularly suitable for cases where the follow-up interval is long and the data points are sparse. The data augmentation process is automated, the system will process each patient's time series segment according to the preset parameters to generate a set of augmented time sub-sequences, which together constitute the input set for pattern matching. It can be understood that the data augmentation technique effectively amplifies the amount of available data, especially for small sample or sparse data patients, it helps the model to learn more stable and generalizable feature representations, reducing the risk of overfitting to a single data sequence.
[0042] In specific implementations, the enhanced time series data is compared with the feature patterns in the risk evolution pattern library for similarity, with dynamic time warping distance as the core measurement index. Dynamic time warping distance is an algorithm for measuring the similarity between two time series, which can handle the stretching and twisting of time series on the time axis, that is, it allows the matching of two sequences that are not completely aligned in time. The calculation process is to find a warping path that minimizes the sum of the distances of the two time series along this path. The smaller the distance value, the more similar the two sequences. The system calculates the dynamic time warping distance between each time subsequence generated after data enhancement and each predefined high-risk feature pattern in the risk evolution pattern library. This process involves a large number of distance calculations, usually using an optimized dynamic programming algorithm to complete efficiently. The calculated dynamic time warping distance value is recorded and associated with the corresponding pattern name and subsequence identifier, forming a detailed similarity score list to provide data support for subsequent screening decisions. Optionally, in addition to dynamic time warping distance, the system may also calculate other similarity indicators such as Euclidean distance or Pearson correlation coefficient as a reference, but dynamic time warping distance is the main basis for decision-making.
[0043] In specific implementations, patterns with similarity exceeding a set threshold are selected, and the corresponding clinical data points are marked as high-risk candidate points. The similarity threshold is a pre-set upper limit value of the dynamic time warping distance, which is determined by experiments on labeled historical data sets to balance the recall rate and precision rate. The screening process traverses all calculated dynamic time warping distance values. For those distance values less than or equal to the set threshold, it is considered that the time subsequence is highly similar to the corresponding risk evolution pattern. Each time subsequence that is determined to be highly similar, its corresponding clinical data point will be automatically marked by the system as a high-risk candidate point. The marking operation can add a label field in the database or create a record in a special risk warning list, recording information including patient identification, time point triggering the warning, matched high-risk pattern name, and calculated dynamic time warping distance value. It can be understood that threshold setting is a key link. If the threshold is too loose, it will result in too many false positive warnings, increasing the clinical workload; if the threshold is too tight, it may result in missing real high-risk patients, which needs to be calibrated repeatedly through historical data to find the best balance point. Optionally, the system can support multiple threshold settings, such as dividing the risk level into "high", "medium", and "low" according to the size of the dynamic time warping distance, providing more detailed reference for clinical decision-making.
[0044] In a specific implementation, all high-risk candidate points are aggregated to generate a preliminary risk alert set, and the aggregation operation is performed on all patient data processed in the current batch. The system collects all records marked as high-risk candidate points, removes duplicates, and then sorts them according to certain rules, such as sorting by dynamic time warping distance from small to large, or sorting by the risk level of the matched pattern itself. The final form of the preliminary risk alert set can be a structured list or report, as shown in Table 1, which shows a preliminary risk alert set.
[0045] Table 1: Preliminary risk alert set After the preliminary risk alert set is generated, it is passed to the next multi-dimensional verification link for further confirmation, and at the same time, it will remind the relevant clinicians to pay attention to the patients in the list in an appropriate form, so as to achieve early detection and prompt of risk.
[0046] Referring to Figure 5 , which shows the results of risk evolution pattern recognition and matching method. The chart calculates the similarity between real-time collected patient data and feature patterns in the risk evolution pattern library through the dynamic time warping distance algorithm, achieving early identification and early warning of high-risk patients. The chart shows the matching of time series data of multiple patients and standard risk patterns. Each curve represents the change trajectory of the tumor aggressiveness index of a patient, and the dashed line represents the standard risk pattern in the pattern library. The chart clearly shows the most similar risk pattern and its distance value for each patient through color coding and legends. The smaller the distance value, the higher the matching degree. It is particularly noteworthy that the chart uses red stars to mark the candidate points identified as high-risk, which correspond to patient data highly matching high-risk patterns such as sustained rapid rising pattern and fluctuating rising pattern. This visual warning mechanism enables clinicians to quickly locate patients who need to be paid attention to and take timely intervention measures. Through this pattern matching method, the system can automatically identify patient groups with specific risk evolution characteristics from a large amount of patient data, greatly improving the efficiency and accuracy of risk warning, and providing strong support for precise prediction of breast cancer metastasis risk.
[0047] In specific implementations, the complete clinical history of high-risk candidate points is obtained, including the history of previous treatment and complications. The clinical history is extracted from the hospital electronic medical record system, covering the complete diagnosis and treatment records of the patient from the initial diagnosis to the current time point. The history of previous treatment records the key treatment parameters such as the type of surgery (such as breast-conserving surgery or total mastectomy), chemotherapy regimen (drug type, dose and cycle), radiotherapy range and dose, and the complications include adverse reaction events such as infection, bone marrow suppression, cardiotoxicity and the control status of comorbidities such as diabetes and hypertension. These clinical background information is crucial for understanding the biological behavior of tumors and treatment response patterns. For example, a patient matched to the "persistent rapid rise pattern" has a significantly higher credibility of early warning signals if he has received multi-line treatment and shows drug resistance. It can be understood that a comprehensive assessment of clinical history helps to distinguish between real high-risk patients and false positive warnings caused by insufficient treatment or poor compliance. The system automatically extracts key fields through structured forms, while supporting manual doctor annotations.
[0048] In specific implementations, genomic features including gene mutation status and gene expression profile are detected as additional examination items after the early warning is triggered. Gene mutation status is detected by next-generation sequencing technology to detect somatic mutations of breast cancer-related genes such as BRCA1, BRCA2, PIK3CA, TP53, etc. A targeted gene panel is used to cover known driver mutations and drug resistance mutation sites. Gene expression profile is determined by RNA sequencing, focusing on the activity of signaling pathways related to tumor invasiveness and metastatic potential, such as the expression levels of epithelial-mesenchymal transition markers, angiogenic factors, and cell cycle regulatory genes. The detection results are recorded in the system in the form of structured data, and are analyzed in association with clinical features. For example, patients carrying BRCA1 mutations and negative for ER expression may exhibit metastatic patterns. This molecular-level information provides additional biological evidence for risk verification. In some embodiments, the detection of genomic features may be limited by sample accessibility or detection period. The system will mark early warning cases for which genomic data has not been obtained, and automatically re-evaluate the risk level after the data is updated to ensure the dynamic updating ability of the prediction model.
[0049] In a specific implementation, the treatment response data is analyzed to assess the effectiveness and tolerability of prior treatments. The treatment response data is derived from periodic follow-up imaging evaluations and laboratory test results. The imaging evaluations use the Response Evaluation Criteria in Solid Tumors to classify changes in tumor burden into four categories: complete remission, partial remission, stable disease, and progressive disease. The results of each evaluation are ordered in time to form a treatment response curve. Laboratory tests monitor trends in tumor marker levels, blood counts, liver and kidney function, and other indicators to reflect the biological effects and toxic reactions of treatment. The system automatically calculates key indicators such as progression-free survival, time to treatment failure, and other parameters, and combines them with physician- filled toxicity grading scales to comprehensively evaluate the clinical benefit-to-risk ratio of treatment. For example, a patient who matches the "fluctuating rise pattern" may need to adjust rather than directly confirm a high-risk status if their tumor marker levels fluctuate and rise, but imaging evaluations remain stable disease, and treatment tolerability is good. It can be understood that dynamic analysis of treatment response data can reveal the actual response patterns of tumors to treatment, help to distinguish between different risk mechanisms of inherent invasiveness and acquired drug resistance of tumors, and provide a basis for subsequent treatment decisions.
[0050] In a specific implementation, a multi-factor validation model is established for joint analysis. The multi-factor validation model is a machine learning framework that integrates clinical features, genomic features, and treatment response data. The model constructs a multi-source feature input layer to receive feature vectors from different dimensions. The clinical feature vector includes patient demographic data, tumor pathological features, treatment history, and complications, among other structured data. The genomic feature vector contains mutation profiles, gene expression levels, and other molecular data. The treatment response data vector integrates efficacy evaluation indicators and toxicity scores. A feature cross layer is designed to calculate the interaction weights between different source features through an attention mechanism to generate cross-features with high-order correlations. The attention mechanism uses a query-key-value pair architecture, where the query vector comes from the clinical features, the key vector comes from the genomic features, and the value vector comes from the treatment response data. The attention weights are obtained by calculating the similarity between the query and the key, and then the value vector is weighted and summed. After concatenating the original features and the cross-features, they are input into a survival analysis layer based on the Cox proportional hazards model. The Cox model fits the partial likelihood function to estimate the contribution of each feature to the metastasis risk, and outputs the patient's risk score and risk ratio. The risk score is a comprehensive quantitative indicator that reflects the predicted probability of metastasis within a given time window, and the risk ratio represents the multiplicative effect of a specific feature on the baseline risk. Optionally, the multi-factor validation model supports an online learning mechanism that automatically triggers incremental updates of model parameters when new validation data accumulates to a certain size, maintaining the timeliness of the prediction ability.
[0051] In specific implementations, the patients in the preliminary risk alert set are reordered and confirmed according to a risk score threshold, which is calibrated by survival analysis results of historical cohorts. The system compares the risk score of each patient in the preliminary risk alert set with the preset threshold, and the cases higher than the threshold are confirmed as final high-risk patients, while the cases lower than the threshold are marked for further observation or downgraded processing. The reordering is based on the high and low of the risk score, generating a priority list with the highest scoring patients at the top for the orderly allocation of clinical resources. The confirmation operation also includes excluding false positive risk points, such as manually reviewing those marginal cases with scores close to the threshold, and making the final decision in combination with the doctor's clinical judgment. It can be understood that the threshold setting needs to balance sensitivity and specificity, and under the support of clinical trial data, different thresholds can be set for different subgroups to achieve precise risk stratification. Optionally, the system can generate a visual report showing the comprehensive evaluation results of each high-risk patient, including risk score, main risk factors, supporting evidence and clinical recommendations, to assist multidisciplinary team discussion and decision-making.
[0052] In specific implementations, the final breast cancer metastasis risk prediction results are output, and the prediction results are presented in the form of structured reports and visual dashboards. The structured report contains core content such as patient basic information, risk level, main risk factor list, predicted time window and clinical management recommendations. The visual dashboard displays the risk evolution trend, key indicator change trajectory and risk factor relative contribution through interactive charts, supporting drill-down analysis. The prediction results are pushed to the hospital information system through a secure interface, integrated with the electronic medical record, and an alert flag is set in the doctor's workstation to ensure that high-risk patients receive timely attention. In some embodiments, the system also regularly tracks the accuracy of the prediction results, records the occurrence time and mode of actual metastasis events, and compares them with the prediction results. These feedback data are used for continuous optimization and performance evaluation of the model. The final output of the risk prediction result is not a static conclusion, but a dynamic monitoring and iterative updating process. As new data of the patient accumulates, the system will periodically recalculate the risk score, adjust the risk level, and form a closed-loop management.
[0053] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.
[0054] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.
Claims
1. A breast cancer metastasis risk prediction model based on clinical indicators, characterized in that, The method includes the following steps: Establish a multimodal clinical data stream, collect imaging feature parameters, pathological histological parameters and blood biomarker parameters during the patient's diagnosis and treatment cycle, and integrate the data through the patient's unique identifier to form a time series clinical dataset; Dynamic risk assessment is performed on the time-series clinical dataset, tumor biological characteristic indicators are extracted for each time point, tumor invasiveness index is calculated, and dynamic risk change sequence is generated. Based on the dynamic risk change sequence, spatiotemporal pattern recognition is performed to analyze the evolution of tumor characteristics at different time points, identify high-risk evolution patterns, and construct a risk evolution pattern library. The risk evolution pattern library is used to perform pattern matching on real-time clinical data, detect clinical data points that conform to preset high-risk patterns, and generate a preliminary risk warning set. The preliminary risk warning set is validated in multiple dimensions, and the final breast cancer metastasis risk prediction result is output by combining genomic features and treatment response data.
2. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 1, characterized in that, The steps for establishing a multimodal clinical data stream include: Baseline clinical data of breast cancer patients were obtained from the hospital information system, including age at diagnosis, tumor stage, and histological grade data. Collect imaging examination results during regular follow-up of treatment, and record quantitative parameters such as tumor size, morphological characteristics, and enhancement degree; Collect blood sample test data at each follow-up visit, including tumor marker levels, circulating tumor cell count, and inflammatory marker parameters; By linking and integrating clinical data from different sources using patient hospitalization numbers or social security numbers, and arranging them in chronological order, a complete time-series clinical dataset is formed.
3. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 2, characterized in that, The step of performing dynamic risk assessment on the time-series clinical dataset includes: Key indicators such as the maximum diameter of the tumor, the number of lymph node metastases, and the Ki-67 proliferation index were extracted at each follow-up time point. Calculate dynamic parameters such as tumor volume growth rate and biomarker level change rate; The tumor invasiveness index at each time point was calculated using a weighted scoring method, taking into account the patient's age and underlying disease status. The tumor invasiveness index at each time point is arranged in chronological order to form a dynamic risk change sequence.
4. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 3, characterized in that, The step of performing spatiotemporal pattern recognition based on the dynamic risk change sequence includes: Analyze the trend characteristics in the dynamic risk change sequence to identify the periods of rapid risk increase, plateau, and decline; Detecting the coordinated change patterns of tumor features across different time points, including the synchronous changes in size and biomarkers; Establish a risk evolution pattern recognition algorithm to extract high-risk characteristic patterns, including continuous rapid rise patterns and volatile rise patterns; The identified feature patterns and corresponding clinical indicator change patterns are stored in the risk evolution pattern library.
5. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 4, characterized in that, The step of performing pattern matching on real-time clinical data using the risk evolution pattern library includes: Input newly collected patient clinical data and extract tumor characteristic parameters at the current time point; Combine current data with historical data to form the latest time series segment; The similarity between the time series segments and the feature patterns in the risk evolution pattern library is calculated; Patterns with similarity exceeding a set threshold are filtered out, and the corresponding clinical data points are marked as high-risk candidate points; All high-risk candidate points are aggregated to generate a preliminary risk warning set.
6. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 5, characterized in that, The step of performing multi-dimensional verification on the preliminary risk warning set includes: Obtain complete clinical history of high-risk candidate sites, including previous treatments and complications; Supplement the detection of genomic characteristics, including gene mutation status and gene expression profile characteristics; Analyze treatment response data to assess the effectiveness and tolerability of previous treatments; Establish a multi-factor validation model to conduct joint analysis of clinical characteristics, genomic characteristics, and treatment response data; To eliminate the risk of false positives and identify patients who are truly at high risk of metastasis.
7. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 3, characterized in that, The step of calculating the tumor invasiveness index at each time point using a weighted scoring method includes: A tumor invasiveness assessment index system was constructed, which includes tumor burden, proliferative activity, and systemic response dimensions. Initial weights were assigned to specific indicators under each dimension, with the maximum tumor diameter and number of lymph node metastases belonging to the tumor burden dimension, the Ki-67 proliferation index belonging to the proliferative activity dimension, and tumor marker levels and inflammatory indicators belonging to the systemic response dimension. Based on the correlation between various indicators and the final metastasis outcome in historical patient data, the initial weights are iteratively optimized to obtain the optimal weight combination; Based on the optimal weight combination, the index data collected at each time point are weighted and summed to calculate a quantified tumor invasiveness index.
8. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 4, characterized in that, The steps for establishing the risk evolution pattern recognition algorithm include: After normalizing the dynamic risk change sequence, it is input into a one-dimensional convolutional neural network for local feature extraction to capture the risk fluctuation pattern in the short term. By using long short-term memory networks to analyze normalized sequences, we can learn long-term time dependencies and trend evolution patterns. By fusing the local features output by a one-dimensional convolutional neural network with the sequence context features output by a long short-term memory network, a comprehensive spatiotemporal feature vector is formed. The spatiotemporal feature vectors are input into a fully connected layer and a flexible maximum value classification layer to identify and classify a variety of predefined high-risk evolution patterns.
9. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 5, characterized in that, Following the step of combining current data with historical data to form the latest time series segment, the method further includes a step of data augmentation for the time series segment: Sliding window sampling is performed on the time series segments to generate multiple time subsequences with overlapping portions; Random noise is added to each time subsequence or small fluctuations in the data are simulated by interpolation to expand the training samples; Calculate the dynamic temporal regularization distance between each enhanced time subsequence and the feature patterns in the risk evolution pattern library, and use it as a similarity measure; The top K patterns with the smallest dynamic time-normalized distances are selected as matching results, and their corresponding clinical data points are included in the preliminary risk warning set.
10. The breast cancer metastasis risk prediction model based on clinical indicators as described in claim 6, characterized in that, The steps for establishing a multi-factor validation model and performing joint analysis include: A multi-source feature input layer is constructed to receive multi-dimensional features from clinical features, genomic features, and treatment response data; Design a feature cross layer, and use an attention mechanism to calculate the interaction weights between features from different sources to generate cross features with high-order correlation; The original features and cross features are concatenated and input into the survival analysis layer based on the Cox proportional hazards model to calculate the patient's metastasis risk score and hazard ratio; Based on the risk score, a threshold is set, and patients in the initial risk warning set are reordered and confirmed to output the final risk prediction results with different risk levels.
