A breast surgery prognosis evaluation system based on clinical data

Abstract

The application discloses a breast surgery prognosis evaluation system based on clinical data and belongs to the technical field of breast surgery prognosis evaluation, and specifically comprises the following steps: extracting time-series clinical follow-up data of patients, and constructing a prognosis training sample set; generating an initial risk label based on first-visit pathological characteristics; extracting concentration sequences to generate actual treatment execution trajectories and treatment response physiological trajectories, and extracting time-series characteristics thereof; performing trend segmentation identification on the actual treatment execution trajectories to generate execution behavior trajectory characteristics; fusing the execution behavior trajectory characteristics and the treatment response physiological time-series characteristics to construct a dynamic correction characteristic set, and generating a risk correction model; acquiring real-time biochemical data of a target patient and inputting the data into the model, and outputting a calibrated current prognosis risk grade. The application realizes dynamic and continuous calibration and accurate evaluation of a prognosis risk by cross-verification of serum active metabolite concentrations and liver enzyme changes to identify real medication behaviors, and dynamically coupling baseline prognosis and actual treatment execution.

Description

Technical Field

[0001] This invention relates to the field of breast surgical prognostic assessment technology, and more specifically to a breast surgical prognostic assessment system based on clinical data. Background Technology

[0002] Breast cancer patients typically require endocrine adjuvant therapy for five years or longer after surgery to reduce the long-term risk of recurrence. In clinical practice, based on the initial pathological characteristics of the patient after surgery, such as tumor size, lymph node metastasis status, histological grade, proliferation index, and hormone receptor expression level, doctors calculate the patient's baseline recurrence risk score using a breast cancer prognostic assessment model. Based on this score, patients are classified into different risk levels, such as low-risk, intermediate-risk, or high-risk, to guide the development of subsequent treatment and follow-up plans.

[0003] However, the above prognostic assessment methods are based on an implicit assumption that patients strictly adhere to their doctor's orders and take medication regularly during endocrine therapy, maintaining drug exposure levels above therapeutically effective concentrations. Some improved approaches have begun to focus on patient medication adherence, monitoring it through patient self-reporting, pill counting, or prescription refill records.

[0004] However, these approaches generally overlook the inherent limitations of subjectively reported data, including patient memory bias, the tendency to overreport due to social expectation effects, and the information distortion caused by prescription and medication records not being equivalent to actual medication adherence. Existing prognostic assessment systems only make a one-time risk assessment based on baseline pathological information at the start of treatment, failing to incorporate the patient's actual treatment adherence in the real world into the dynamic adjustment process of prognostic risk. For patients assessed as low-risk at baseline but who have not adhered to medication regimens for a long period, their true relapse risk may have significantly increased to intermediate-high risk levels. Existing systems cannot respond to this in a timely manner, maintaining the original optimistic assessment conclusion until distant relapse events such as bone metastasis occur, thus missing the optimal intervention window.

[0005] Therefore, there is a lack of existing technologies that can reverse-engineer a patient's true drug exposure by using only objective biochemical test data generated during routine follow-up, without relying on the patient's subjective reports, and thereby dynamically calibrate the static prognostic risk based on baseline pathological characteristics. Summary of the Invention

[0006] The purpose of this invention is to provide a breast surgery prognostic assessment system based on clinical data, addressing the following technical problems: Existing breast surgery prognostic assessment systems only make a one-time risk assessment based on the static baseline pathological characteristics at the time of diagnosis. They cannot detect changes in the actual drug exposure level caused by deviations from the doctor's prescription during long-term endocrine therapy, nor can they dynamically calibrate the initial prognostic risk based on this.

[0007] The objective of this invention can be achieved through the following technical solutions: A clinical data-based breast surgery prognostic assessment system includes: Sample acquisition module: used to acquire time-series clinical follow-up data of historical patient cohorts within a complete follow-up period, extract serum active metabolite concentration detection value sequences and liver enzyme detection value sequences from the time-series clinical follow-up data, and construct a prognostic training sample set; Initial risk module: used to perform baseline risk coding on the initial pathological features of each sample in the prognostic training sample set, and generate an initial prognostic risk label for each sample; Treatment execution module: used to extract the serum active metabolite concentration detection value sequence of each sample from the prognostic training sample set, generate the actual treatment execution time trajectory of each sample, and extract the liver enzyme detection value sequence of each sample from the prognostic training sample set, generate the treatment response physiological index time trajectory of each sample. Behavioral trajectory module: used to extract the temporal features of the treatment response physiological index time-series trajectory, and when performing trend segmentation identification on the actual treatment execution time-series trajectory, use the temporal features to cross-validate the concentration decrease segment, and generate execution behavioral trajectory features carrying behavioral-related decrease segment markers; Risk correction module: used to align the execution behavior trajectory features with the time-series features of the treatment response physiological indicators in the time dimension and fuse the features to construct a dynamic correction feature set. The risk correction model is trained with the initial prognostic risk label and the dynamic correction feature set as input and the long-term follow-up outcome of each sample as the supervision label. Relapse assessment module: used to obtain real-time serum active metabolite concentration and real-time liver enzyme detection values ​​of target patients, extract real-time treatment execution quantitative indicators and real-time treatment response physiological indicators and input them into the risk correction model, and output the calibrated current prognostic risk level.

[0008] As a further aspect of the present invention: the specific process of constructing the prognostic training sample set in the sample acquisition module is as follows: Time-series clinical follow-up data of historical breast cancer patient cohorts within a complete follow-up period were extracted from hospital information systems and laboratory information systems. The time-series clinical follow-up data included serum active metabolite concentrations, liver enzyme values, and long-term follow-up outcomes at each follow-up node. The time-series clinical follow-up data were then subjected to longitudinal data association on a patient-by-patient basis. The serum active metabolite concentrations and liver enzyme values ​​of the same patient at each follow-up node were arranged in chronological order of follow-up time, forming an individualized time-series data chain for each patient. The individualized time-series data chains are cleaned to remove samples with a follow-up node missing rate exceeding a preset threshold and samples with missing key variables. Abnormal values ​​of serum active metabolite concentration and liver enzyme detection values ​​in the remaining samples are identified and marked. The individualized time-series data chains of each patient after cleaning are aligned with the follow-up time window. The detection time of each follow-up node is converted into a relative time coordinate in months, with the patient's surgery date as the baseline zero point. The individualized time-series data chains of each patient after cleaning and time alignment are integrated into a prognostic training sample set.

[0009] As a further aspect of the present invention: the specific process for generating the initial prognostic risk label for each sample in the initial risk module is as follows: The initial pathological features of each sample are extracted from the prognostic training sample set. These initial pathological features include the maximum diameter of the tumor, the number of lymph node metastases, histological grade, proliferation index, estrogen receptor status, and progesterone receptor status. The actual values ​​of each indicator in the initial pathological features of each sample are compared with a preset pathological indicator coding table, and then the actual values ​​of each indicator are converted into corresponding coded values. The coded values ​​of the initial pathological features of each sample after baseline risk coding are input into a preset breast cancer prognostic assessment model to calculate the baseline recurrence risk score of each sample. The comparison results between the baseline recurrence risk score and the preset risk grading threshold determine the initial prognostic risk level of each sample. The initial prognostic risk level of each sample is used as the initial prognostic risk label of that sample and associated with the corresponding sample in the prognostic training sample set for storage.

[0010] As a further aspect of the present invention: the specific generation process of the temporal characteristics of the treatment response physiological index temporal trajectory in the behavioral trajectory module is as follows: The liver enzyme test values ​​corresponding to each relative time coordinate are extracted from the time-series trajectory of the physiological indicators of treatment response. The time-series trajectory of the physiological indicators of treatment response consists of liver enzyme test values ​​of the same sample at each follow-up node arranged in order of relative time coordinates. The time nodes in the time-series trajectory of the physiological indicators of treatment response show obvious turning points in liver enzyme test values. The nodes where the liver enzyme test values ​​change from stable or decreasing to continuously increasing are marked as abnormal starting points. The nodes where the liver enzyme test values ​​change from increasing to continuously decreasing are marked as peak points. The nodes where the liver enzyme test values ​​fall back to the normal range are marked as recovery points. The duration of the abnormal fluctuation interval formed by adjacent abnormal starting points and recovery points, the peak value of the liver enzyme test values ​​within the interval, the fluctuation amplitude of the liver enzyme test values ​​within the interval, and the rate of change of the liver enzyme test values ​​from the abnormal starting point to the peak point are extracted as interval feature parameters of the abnormal fluctuation interval. The interval feature parameters of all abnormal fluctuation intervals of the same sample are combined in time order to generate the time-series features of the physiological indicators of treatment response of the sample.

[0011] As a further aspect of the present invention: the specific process of trend segmentation identification of the actual treatment execution time trajectory in the behavior trajectory module is as follows: The actual treatment execution timeline trajectory of each sample is obtained from the prognostic training sample set. The actual treatment execution timeline trajectory is composed of the serum active metabolite concentration detection values ​​of the same sample at each follow-up node arranged in relative time coordinate order. The difference between the serum active metabolite concentration detection values ​​at adjacent time nodes in the actual treatment execution timeline trajectory is calculated. The segments with positive and continuous differences are divided into concentration rising segments, the segments with negative and continuous differences are divided into concentration falling segments, and the segments with absolute differences less than a preset fluctuation threshold and continuous differences are divided into concentration plateau segments. Extract the starting and ending concentration values ​​of each trend segment. The difference between the ending and starting concentration values ​​is taken as the segment change amplitude of the trend segment. The segment change amplitude is divided by the duration of the trend segment to obtain the segment change rate. The segment type, starting time coordinate, ending time coordinate, duration, change amplitude, and change rate of each trend segment are combined according to their arrangement order in the trajectory to form the trend segmentation identification result of the sample.

[0012] As a further aspect of the present invention: the specific process for generating the behavior trajectory features in the behavior trajectory module is as follows: From the trend segmentation identification results, each trend segment of the type of concentration decrease segment is extracted one by one, and the start time coordinate and end time coordinate of the concentration decrease segment are obtained to form the time interval to be verified for the concentration decrease segment. From the time-series characteristics of physiological indicators of treatment response of the same sample, the interval feature parameters of abnormal fluctuation intervals corresponding to the time interval to be verified are extracted. If the liver enzyme detection value in the interval feature parameters shows a continuous downward trend in the time interval to be verified, the concentration decrease segment is marked as a behavior-related decrease segment. If the liver enzyme detection value in the interval feature parameters does not show a continuous downward trend in the time interval to be verified, the concentration decrease segment is marked as a metabolism-related decrease segment. The trend segments in the trend segmentation identification results that are of the concentration rising segment and the concentration plateau segment retain their original segment type and segment attribute parameters. These segments are then combined with the labeled behavior-related falling segments and metabolism-related falling segments according to their original arrangement order in the trajectory to form the execution behavior trajectory features of the sample.

[0013] As a further aspect of the present invention: the specific construction process of the dynamically corrected feature set in the risk correction module is as follows: The time intervals covered by each trend segment in the execution behavior trajectory features are aligned with the time intervals covered by each abnormal fluctuation interval in the treatment response physiological index time sequence features, so that the segment attribute parameters of each trend segment are mapped to the interval feature parameters of the abnormal fluctuation interval within the time interval covered by the trend segment. The segment type identifier, segment attribute parameters and interval feature parameters of the abnormal fluctuation interval mapped by the trend segment are concatenated to generate a joint feature vector of each trend segment. The joint feature vectors of each trend segment are concatenated according to the original arrangement order of each trend segment in the trajectory to generate a dynamic correction feature set for each sample.

[0014] As a further aspect of the present invention: the specific generation process of the risk correction model in the risk correction module is as follows: The actual follow-up durations of each sample in the prognostic training sample set are arranged in ascending order. The follow-up time points where relapse events occurred in each actual follow-up duration are extracted as time points to be calculated. For each time point to be calculated, the number of samples where relapse events occurred at that time point is counted as the number of events. The number of samples that had not yet occurred and had not been censored at the beginning of that time point is counted as the number of risk set samples. The number of events is divided by the number of risk set samples to obtain the instantaneous relapse risk rate at that time point. From the surgery date to each time point to be calculated, the instantaneous recurrence risk rates of each time point to be calculated are multiplied together after taking the complement of each rate, to obtain the cumulative survival probability without recurrence at that time point. The cumulative survival probability without recurrence is then multiplied by taking the complement of the cumulative survival probability without recurrence to obtain the cumulative recurrence probability at that time point. Each time point to be calculated and its corresponding cumulative recurrence probability are combined in chronological order to form a recurrence probability time series, which is used as a supervision signal for model training. Using the initial prognostic risk label and dynamic correction feature set of each sample as joint input, and the recurrence probability time series as supervision signal, the preset prognostic risk correction network is iteratively trained. The network parameters are adjusted until the deviation between the corrected recurrence probability output by the network and the supervision signal converges to a preset threshold. The trained prognostic risk correction network is then stored as a risk correction model.

[0015] As a further aspect of the present invention: the specific process by which the recurrence assessment module outputs the calibrated current prognostic risk level is as follows: The system acquires real-time serum active metabolite concentrations, real-time liver enzyme levels, and initial pathological features of the target patient. It inputs the initial pathological features into a breast cancer prognostic assessment model to determine the initial prognostic risk level of the target patient. The initial prognostic risk level, the real-time serum active metabolite concentrations, and the real-time liver enzyme levels are then input into a trained risk correction model, which outputs a corrected cumulative recurrence probability sequence for the target patient at multiple future time points. The corrected recurrence probability value corresponding to a preset assessment time point is extracted from the corrected cumulative recurrence probability sequence. Based on the comparison between this corrected recurrence probability value and a preset risk grading threshold, the calibrated current prognostic risk level of the target patient is determined and output.

[0016] The beneficial effects of this invention are: 1) This invention extracts serum active metabolite concentration and liver enzyme levels from hospital and laboratory information systems during patient follow-up to construct a prognostic training sample set based on objective biochemical data. It is understood that serum active metabolite concentration directly reflects the true exposure level of a drug after oral absorption into the systemic circulation, and its concentration depends on both the patient's actual drug intake and their metabolic capacity. Liver enzyme levels, as physiological indicators of hepatocyte stress response during drug metabolism, are dynamically correlated with drug exposure over time. This invention arranges the serum active metabolite concentration and liver enzyme levels of the same patient at various follow-up points in chronological order to form an individualized time-series data chain. It establishes a synchronous record of these two types of objective indicators over time, enabling subsequent analysis to trace the temporal correspondence between each change in drug exposure and physiological response from a data perspective. This eliminates the reliance on subjective data such as patient self-reporting in traditional protocols, achieving a technological leap from subjective compliance assessment to objective biological exposure monitoring.

[0017] 2) This invention performs trend segmentation identification on the actual treatment execution timeline, dividing the concentration change segment into an upward segment, a downward segment, and a plateau segment. When the concentration decreases, the trend of liver enzyme test values ​​during the same period is retrieved for cross-validation. It is understood that a continuous decrease in serum active metabolite concentration means that the patient's drug exposure level is decreasing, but the driving factors for this decrease are fundamentally different: if the concentration decreases while the liver enzyme test value simultaneously returns to normal levels, it indicates that the increase in liver enzymes is a physiological response to drug exposure, and the concentration decrease stems from drug exposure withdrawal. The temporal synchronicity of these two factors constitutes an objective chain of evidence for the patient's dose reduction or discontinuation due to liver injury side effects. If the concentration decreases but the liver enzyme test value does not show a synchronous decline, it indicates that there is no temporal coupling relationship between the occurrence of the concentration decrease and the recovery of liver injury. The driving factor for the patient's dose reduction or discontinuation is not a liver injury side effect, but may involve other non-liver injury related causes. This invention introduces liver enzyme detection values ​​as cross-validation signals during the segmented identification process. Based on whether there is a temporal synchronization relationship between the concentration decrease and the liver enzyme decline, the concentration decrease segment is marked as a behavior-related decrease segment or a metabolism-related decrease segment. This enables the system to automatically identify the behavior of concentration decrease events based solely on objective biochemical data, achieving a technological leap from a single judgment of concentration changes to behavior identification analysis based on the temporal coupling relationship of multiple indicators.

[0018] 3) This invention aligns and fuses the behavioral trajectory features carrying behaviorally related decline markers with the temporal features of physiological indicators of treatment response in a time dimension, constructing a dynamically corrected feature set. Based on a survival analysis algorithm, the relapse probability time series is extracted from the long-term follow-up outcomes of historical cohorts as a supervisory signal to train the risk correction model. It is understandable that if a low-risk patient at baseline experiences a prolonged behaviorally related decline in the early stages of treatment, their serum active metabolite concentrations continue to decrease, and liver enzymes decline synchronously, indicating that drug exposure has been substantially interrupted. The actual total drug exposure in the body is far lower than the assumptions of the prognostic assessment model, and the true relapse risk has risen from the low-risk range to the intermediate-high-risk level. While the metabolically related decline also manifests as a decrease in serum active metabolite concentrations, liver enzymes do not decline synchronously. There is no temporal coupling relationship between the decrease in their concentrations and the recovery of liver damage. The correction pattern for relapse risk may differ from that of the behaviorally related decline. This invention uses a survival analysis algorithm to convert long-term follow-up outcomes into cumulative relapse probabilities at various time points. This enables the model to automatically learn the correlation between different types of decline characteristics and long-term outcomes based on training data. In clinical applications, real-time biochemical data of the target patient can be obtained to output the corrected relapse probability and calibrate the risk level. This represents a technological leap from static baseline risk assessment to dynamic risk calibration based on multi-indicator temporal coupling characteristics. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a schematic diagram of a breast surgery prognostic assessment system based on clinical data, according to the present invention. Detailed Implementation

[0021] 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.

[0022] Please see Figure 1 As shown, this invention is a breast surgery prognostic assessment system based on clinical data, comprising: Sample acquisition module: used to acquire time-series clinical follow-up data of historical patient cohorts within a complete follow-up period, extract serum active metabolite concentration detection value sequences and liver enzyme detection value sequences from the time-series clinical follow-up data, and construct a prognostic training sample set; First, time-series clinical follow-up data of historical breast cancer patient cohorts throughout the complete follow-up period were extracted from the hospital information system and laboratory information system. This time-series clinical follow-up data includes serum active metabolite concentrations, liver enzyme levels, and long-term follow-up outcomes at each follow-up time point. Serum active metabolite concentrations refer to the concentrations of active metabolites of endocrine therapy drugs in serum, measured by liquid chromatography-tandem mass spectrometry (LC-MS / MS) after venous blood samples were drawn from patients at each follow-up time point. This value directly reflects the true exposure level of the drug after oral absorption into the systemic circulation. Liver enzyme levels refer to the enzyme activities of alanine aminotransferase (ALT) and aspartate aminotransferase (AST) measured by a biochemical analyzer from the same blood sample. This value reflects the physiological stress state of hepatocytes under drug influence. Long-term follow-up outcomes include markers or censoring of recurrence events, and the time interval from the surgery date to the date of the recurrence event or the last follow-up date.

[0023] The time-series clinical follow-up data are subjected to longitudinal data correlation on a patient-by-patient basis. The serum active metabolite concentration and liver enzyme test values ​​of the same patient at each follow-up node are arranged in chronological order of follow-up time, forming an individualized time-series data chain for each patient. For example, if a patient has blood drawn at follow-up times of 0 months, 6 months, 12 months, 18 months, and 24 months post-surgery, the serum active metabolite concentration and liver enzyme test values ​​of the patient at these five time nodes are arranged in chronological order to form an individualized time-series data chain containing five sets of data points.

[0024] The individualized time-series data chain is cleaned to remove samples with a follow-up node missing rate exceeding a preset threshold and samples with missing key variables. The preset threshold can be set according to the actual data quality; for example, a follow-up node missing rate threshold of 30% means that if more than 30% of the follow-up nodes for a patient lack blood test records, that patient's sample is removed. Missing key variables refer to a situation where the serum active metabolite concentration or liver enzyme test value is completely blank at a certain follow-up node; samples with missing key variables are also removed. Outliers in the remaining samples' serum active metabolite concentration and liver enzyme test values ​​are identified and marked. Outliers are test values ​​that exceed a preset physiologically reasonable range, such as negative serum active metabolite concentration values ​​or values ​​exceeding the upper limit of the linear range of the testing instrument. These are marked as outliers for subsequent interpolation or removal processing.

[0025] The cleaned individualized time-series data chains for each patient were aligned with follow-up time windows. Using the patient's surgery date as the baseline zero point, the detection time of each follow-up node was converted into a relative time coordinate in months. For example, if a patient underwent surgery on March 15, 2018, this date would be marked as relative time coordinate zero. The relative time coordinate for the follow-up detection on September 15, 2018 would be six, and the relative time coordinate for the follow-up detection on March 15, 2019 would be twelve. The cleaned and time-aligned individualized time-series data chains for each patient were then integrated into a prognostic training sample set. It can be understood that each sample in this prognostic training sample set contains a complete follow-up record aligned with the time dimension, constructed based on objective biochemical detection data, providing a data foundation for subsequent steps that does not rely on subjective patient reports.

[0026] Initial risk module: used to perform baseline risk coding on the initial pathological features of each sample in the prognostic training sample set, and generate an initial prognostic risk label for each sample; The initial pathological features of each sample were extracted from the prognostic training sample set. These initial pathological features included the maximum diameter of the tumor, the number of lymph node metastases, histological grade, proliferation index, estrogen receptor status, and progesterone receptor status. The maximum diameter of the tumor was recorded in millimeters as the largest diameter of the invasive tumor in the surgically resected specimen; the number of lymph node metastases was recorded as the number of metastatic lymph nodes confirmed by intraoperative sentinel lymph node biopsy or axillary lymph node dissection; the histological grade was classified as grade I, II, or III according to the Nottingham Grading System; the proliferation index was expressed as a percentage of the proportion of tumor cells positive for Ki-67 immunohistochemical staining; and the estrogen receptor status and progesterone receptor status were recorded as positive or negative, respectively.

[0027] The actual values ​​of each indicator in the initial pathological features of each sample are compared with a pre-set pathological indicator coding table. The actual values ​​of each indicator are converted into corresponding coded values ​​to complete the baseline risk coding of the initial pathological features of each sample. For example, the actual value of the maximum tumor diameter "15 mm" corresponds to the code value "-", the number of lymph node metastases "zero" corresponds to the code value "zero", the histological grade "Grade I" corresponds to the code value "-", the proliferation index "10 percent" corresponds to the code value "zero", the estrogen receptor status "positive" corresponds to the code value "-", and the progesterone receptor status "positive" corresponds to the code value "-". The coding rules for each indicator are pre-set according to the coding standards commonly used in the field of breast cancer prognostic assessment.

[0028] The initial pathological feature codes of each sample, after baseline risk coding, are input into a preset breast cancer prognostic assessment model to calculate the baseline recurrence risk score for each sample. The breast cancer prognostic assessment model can be a clinically used CTS5 calculator or other validated prognostic assessment tools. This model calculates a numerical baseline recurrence risk score based on a weighted average of the input pathological feature codes. The initial prognostic risk level for each sample is determined by comparing the baseline recurrence risk score with a preset risk grading threshold. The preset risk grading threshold divides the risk score into low-risk, intermediate-risk, and high-risk intervals. For example, scores below 5% are classified as low-risk, scores between 5% and 10% as intermediate-risk, and scores above 10% as high-risk. The initial prognostic risk level of each sample is used as its initial prognostic risk label and associated with the corresponding sample in the prognostic training sample set. It is understood that the initial prognostic risk label is generated solely based on static pathological information at the time of surgery and does not incorporate any dynamic information during treatment. It represents the theoretical recurrence risk under the assumption that the patient strictly adheres to medical advice and completes the entire treatment course.

[0029] Treatment execution module: used to extract the serum active metabolite concentration detection value sequence of each sample from the prognostic training sample set, generate the actual treatment execution time trajectory of each sample, and extract the liver enzyme detection value sequence of each sample from the prognostic training sample set, generate the treatment response physiological index time trajectory of each sample. Serum active metabolite concentrations for each sample at each follow-up node were extracted from the prognostic training sample set. These concentrations were then arranged chronologically according to relative time coordinates to form the actual treatment execution timeline for that sample. This trajectory visually reflects the change in the patient's true drug exposure level over time. Liver enzyme values ​​for each sample at each follow-up node were also extracted from the prognostic training sample set. These values ​​were then arranged chronologically according to relative time coordinates to form the treatment response physiological index timeline for that sample. This trajectory reflects the change in the physiological stress state of the patient's hepatocytes under drug action over time, providing physiological response data corresponding to the drug exposure trajectory for subsequent steps.

[0030] Behavioral trajectory module: used to extract the temporal features of the treatment response physiological index time-series trajectory, and when performing trend segmentation identification on the actual treatment execution time-series trajectory, use the temporal features to cross-validate the concentration decrease segment, and generate execution behavioral trajectory features carrying behavioral-related decrease segment markers; First, temporal features are extracted from the time-series trajectory of the physiological indicators of treatment response. Time nodes showing significant turning points in liver enzyme test values ​​within the time-series trajectory are identified. Nodes where liver enzyme test values ​​change from stable or decreasing to a continuous increase are marked as abnormal starting points; nodes where liver enzyme test values ​​change from increasing to a continuous decrease are marked as peak points; and nodes where liver enzyme test values ​​return to the normal range are marked as recovery points. The duration of the abnormal fluctuation interval formed by adjacent abnormal starting points and recovery points, the peak value of liver enzyme test values ​​within the interval, the fluctuation amplitude of liver enzyme test values ​​within the interval, and the rate of change of liver enzyme test values ​​from the abnormal starting point to the peak point are extracted as interval feature parameters for that abnormal fluctuation interval. The interval feature parameters of all abnormal fluctuation intervals for the same sample are combined in chronological order to generate the temporal features of the physiological indicators of treatment response for that sample.

[0031] Simultaneously, trend segmentation identification is performed on the actual treatment execution timeline. The difference between serum active metabolite concentration detection values ​​at adjacent time nodes in the actual treatment execution timeline is calculated. Segments with consecutive positive differences are classified as concentration rising segments, segments with consecutive negative differences are classified as concentration falling segments, and segments with consecutive differences less than a preset fluctuation threshold are classified as concentration plateau segments. For example, if the preset fluctuation threshold is set to 5% of the concentration detection value, a concentration change of no more than 5% between two adjacent detections is considered stable and classified as a plateau segment. The starting and ending concentration values ​​of each trend segment are extracted. The difference between the ending and starting concentration values ​​is taken as the segment change amplitude of that trend segment, and the segment change amplitude divided by the segment duration is taken as the segment change rate of that trend segment. The segment type, starting and ending time coordinates, duration, change amplitude, and change rate of each trend segment are combined according to their arrangement in the trajectory to form the trend segmentation identification result for that sample.

[0032] After trend segmentation identification, cross-validation is performed on the concentration decrease segments to generate execution behavior trajectory features. Each trend segment of the concentration decrease type is extracted from the trend segmentation identification results, and the start and end time coordinates of the segment are obtained to form the time interval to be verified for that concentration decrease segment. From the time-series features of the physiological indicators of the treatment response of the same sample, interval feature parameters of abnormal fluctuation intervals corresponding to the time interval to be verified are extracted. If the liver enzyme detection value in the interval feature parameters shows a continuous decreasing trend within the time interval to be verified, the concentration decrease segment is marked as a behavior-related decrease segment. If the liver enzyme detection value in the interval feature parameters does not show a continuous decreasing trend within the time interval to be verified, the concentration decrease segment is marked as a metabolism-related decrease segment. The original segment type and segment attribute parameters of each trend segment of the concentration increase segment and concentration plateau segment in the trend segmentation identification results are retained and combined with the marked behavior-related decrease segments and metabolism-related decrease segments according to the original arrangement order of each trend segment in the trajectory to form the execution behavior trajectory features of the sample.

[0033] Understandably, the distinction between behaviorally related and metabolically related declines hinges on the temporal coupling between the decrease in liver concentration and the subsequent drop in liver enzyme levels. If liver enzyme levels return to normal simultaneously with the decrease in concentration, it indicates that the elevated liver enzymes are a physiological response to drug exposure, and the decrease in concentration stems from drug withdrawal. This temporal synchronicity constitutes an objective chain of evidence supporting the patient's decision to reduce or discontinue medication due to liver injury side effects. Conversely, if liver enzyme levels do not show a synchronous decline in concentration, it indicates that there is no temporal coupling between the decrease in concentration and the recovery from liver injury. The driving factor for the patient's dose reduction or discontinuation is not a liver injury side effect but may involve other non-liver injury-related causes. Through this cross-validation mechanism, the system can automatically identify behavioral patterns of concentration decline events, relying entirely on objective biochemical data, without requiring any subjective reporting from the patient.

[0034] Risk correction module: used to align the execution behavior trajectory features with the time-series features of the treatment response physiological indicators in the time dimension and fuse the features to construct a dynamic correction feature set. The risk correction model is trained with the initial prognostic risk label and the dynamic correction feature set as input and the long-term follow-up outcome of each sample as the supervision label. The time intervals covered by each trend segment in the execution behavior trajectory features are aligned temporally with the time intervals covered by each abnormal fluctuation interval in the treatment response physiological index time series features, establishing a mapping relationship between the segment attribute parameters of each trend segment and the interval feature parameters of the abnormal fluctuation intervals within the time interval covered by that trend segment. The segment type identifier, segment attribute parameters, and mapped interval feature parameters of each trend segment are concatenated to generate a joint feature vector for each trend segment. The joint feature vectors of each trend segment are then concatenated according to the original arrangement order of the trend segments in the trajectory to generate a dynamically corrected feature set for each sample.

[0035] After constructing the dynamically corrected feature set, the risk correction model is trained. The actual follow-up durations of each sample in the prognostic training sample set are arranged in ascending order, and the follow-up time points where recurrence events occurred are extracted as the time points to be calculated. For each time point to be calculated, the number of samples with recurrence events at that time point is counted as the event count, and the number of samples that had not yet experienced recurrence events and were not censored at the beginning of that time point is counted as the risk set sample size. The event count is divided by the risk set sample size to obtain the instantaneous recurrence risk rate for that time point. From the surgery date to each time point to be calculated, the complements of the instantaneous recurrence risk rates of each time point are multiplied together to obtain the cumulative survival probability without recurrence at that time point. The complement of the cumulative survival probability without recurrence is then used to obtain the cumulative recurrence probability at that time point. The time points to be calculated and their corresponding cumulative recurrence probabilities are combined in chronological order to form a recurrence probability time series, which is used as a supervision signal for model training.

[0036] Using the initial prognostic risk labels and dynamically corrected feature sets of each sample as joint input, and the recurrence probability time series as a supervision signal, a pre-defined prognostic risk correction network is iteratively trained. During training, the network parameters are continuously adjusted to gradually reduce the deviation between the corrected recurrence probability output by the network and the supervision signal. Training stops when the deviation converges to a preset threshold range, and the trained prognostic risk correction network is stored as a risk correction model.

[0037] Understandably, the risk-adjusted model does not rely on clinical experience to pre-determine adjustment coefficients. Instead, it treats behaviorally related and metabolically related decline segments as two trajectory patterns with different prognostic implications, allowing the model to autonomously summarize their statistical associations with long-term relapse outcomes within a supervised survival analysis framework. These two types of decline segments exhibit distinguishable combinations of features in segment attribute parameters and accompanying liver enzyme changes. Through iterative training, the model gradually fits the differentiated impact of different feature combinations on the cumulative relapse probability, ultimately forming a data-driven, dynamic adjustment weight system embedded in the network parameters. In application, this system can output a relapse probability estimate that matches the patient's actual drug exposure history without manual intervention.

[0038] Relapse assessment module: used to obtain real-time serum active metabolite concentration and real-time liver enzyme detection values ​​of target patients, extract real-time treatment execution quantitative indicators and real-time treatment response physiological indicators and input them into the risk correction model, and output the calibrated current prognostic risk level.

[0039] In clinical application, real-time serum active metabolite concentrations, real-time liver enzyme levels, and initial pathological features of the target patient are obtained. The real-time serum active metabolite concentrations and liver enzyme levels are the latest data extracted from the hospital and laboratory information systems at the target patient's most recent follow-up visit. The initial pathological features are input into a breast cancer prognostic assessment model, and the initial prognostic risk level of the target patient is determined according to the same baseline risk coding and scoring process as the initial risk module. The initial prognostic risk level, the real-time serum active metabolite concentrations, and the real-time liver enzyme levels are input into a trained risk correction model, which outputs the corrected recurrence probability. Based on the comparison between the corrected recurrence probability and a preset risk grading threshold, the calibrated current prognostic risk level of the target patient is determined and output. It is understood that the recurrence assessment module, based on the trained risk correction model, can recalibrate the patient's true recurrence risk based solely on the dynamic changes in routine follow-up biochemical data, even before a clinically detectable recurrence event occurs. By converting information on changes in drug exposure contained in objective biochemical data into real-time adjustments to risk levels, the system enables timely identification of patients with low baseline risk but whose actual risk of relapse increases due to insufficient actual drug exposure, thus gaining a valuable window of opportunity for clinical intervention.

[0040] It is important to note that the serum active metabolite concentration detection relied upon in this protocol is not based on random sampling at any arbitrary time point. Instead, it utilizes blood samples collected from patients at routine follow-up points, after continuous medication has reached a steady-state condition. While orally administered drugs are indeed gradually metabolized and cleared by the liver after entering systemic circulation, during long-term, regular medication, the daily intake and metabolic clearance of the drug gradually reach a dynamic equilibrium. At this point, the blood drug concentration no longer fluctuates significantly with the absorption and metabolism cycle of a single dose, but stabilizes within a relatively constant range – this state is called steady state. Under steady-state conditions, the concentration values ​​measured from blood samples taken at any time point are representative and comparable, objectively reflecting the patient's overall drug exposure intensity during this treatment phase, rather than an instantaneous peak or trough after a single dose. Furthermore, this protocol targets the active metabolites of the drug, not the parent drug. Active metabolites are the components generated from the parent drug after being catalyzed and converted by hepatic metabolic enzymes, which actually bind to target cell receptors and exert anti-tumor activity. Their serum concentration integrates information from two dimensions: first, whether the patient actually took the drug; and second, whether the taken drug was successfully converted into a therapeutically active form in the body. If the patient is not taking the medication, the concentration of active metabolites is extremely low; if the patient is taking the medication regularly but liver metabolic function is abnormal, leading to impaired metabolism, the concentration of the parent drug may be normal, while the concentration of active metabolites may also be low. By focusing on the absolute value of the concentration of active metabolites and its dynamic changes over multiple follow-ups, and by cross-validating this with the temporal changes in liver enzyme test values, it is possible to distinguish between two different causes: genuine drug discontinuation behavior and metabolic transformation disorders.

[0041] In a preferred embodiment of the present invention, the specific process of constructing the prognostic training sample set in the sample acquisition module is as follows: Time-series clinical follow-up data of historical breast cancer patients throughout the complete follow-up period were extracted from hospital and laboratory information systems. This included serum active metabolite concentrations, liver enzyme levels, and long-term follow-up outcomes at each follow-up node. Serum active metabolite concentrations were measured by liquid chromatography-tandem mass spectrometry (LC-MS / MS) of endocrine therapy drug active metabolites after venous blood sampling. This value directly reflects the true exposure level of the drug after oral absorption into systemic circulation. Liver enzyme levels were measured by alanine aminotransferase (ALT) and aspartate aminotransferase (AST) activities from the same blood sample using a biochemical analyzer. These two enzymes are released from intracellular cells into the bloodstream when hepatocytes respond to drug stimulation, thus objectively reflecting the physiological stress state of hepatocytes. Long-term follow-up outcomes recorded the occurrence time of recurrence or censored events. The above data were longitudinally correlated on a patient-by-patient basis, arranging the two types of test values ​​for the same patient at each follow-up node in chronological order of sampling date to form an individualized time-series data chain, ensuring that concentration values ​​and liver enzyme values ​​were synchronized over time. The individualized time-series datasets were cleaned to remove samples with missing follow-up node rates exceeding a preset threshold and samples with missing key variables. Outliers exceeding the linear range of the testing instruments or the physiologically reasonable range were identified and marked. The cleaned datasets were then aligned to a time window using the patient's surgery date as the baseline, converting the detection times of each follow-up node into relative time coordinates in months. This ensured that the same time-series nodes from different patients corresponded strictly in terms of disease progression. All cleaned and time-aligned individualized time-series datasets were then integrated into a prognostic training sample set.

[0042] This invention directly extracts serum active metabolite concentrations and liver enzyme test values ​​from existing hospital laboratory data. Both indicators are objective biochemical data that are inevitably generated during routine follow-up, requiring no additional examinations or questionnaires from patients, thus ensuring the objectivity and traceability of information from the data source. Concentration values ​​provide biological evidence of whether the drug has actually entered the systemic circulation, while liver enzyme values ​​provide biological evidence of whether the drug has elicited a physiological response in hepatocytes. Together, they constitute a biological evidence system independent of subjective will. This system is the premise for all subsequent modules: the treatment execution module extracts drug exposure trajectories from it; the behavioral trajectory module relies on the synchronous changes in concentration and liver enzymes to cross-validate the causes of drug discontinuation; and the risk correction model learns the statistical regularity between exposure interruption and relapse outcomes from the time-series trajectories of a large number of samples. If the data basis relies on subjective reporting, the reliability of all subsequent analyses is meaningless. Therefore, constructing a prognostic training sample set driven by objective biochemical data is the fundamental guarantee for achieving the technological leap from subjective compliance assessment to objective biological exposure monitoring.

[0043] In another preferred embodiment of the present invention, the specific process of generating the initial prognostic risk label for each sample in the initial risk module is as follows: Initial pathological features of each sample were extracted from the prognostic training sample set, including maximum tumor diameter, number of lymph node metastases, histological grade, proliferation index, estrogen receptor status, and progesterone receptor status. Maximum tumor diameter is the largest diameter of an invasive tumor directly measured by a pathologist on the surgically removed specimen; a larger tumor usually indicates a higher total tumor cell burden. The number of lymph node metastases is the number of lymph nodes pathologically confirmed to contain cancer cells after intraoperative lymph node dissection; a higher number of metastases indicates that the tumor has spread beyond the local breast tissue to the regional lymphatic system. Histological grade is based on the Nottingham Grading System, comprehensively assessed by observing three indicators under a microscope: the proportion of glandular formation, the degree of nuclear atypia, and the mitotic count. A higher grade indicates poorer tumor differentiation and higher proliferative activity. The proliferation index is determined by positive Ki-67 immunohistochemical staining. The percentage of cells is determined by Ki-67 protein. Since Ki-67 protein exists only during the cell proliferation cycle and degrades rapidly after division, only cells in the proliferative state will be stained positive. A higher percentage indicates a larger proportion of actively proliferating tumor cells. Estrogen receptor and progesterone receptor status are determined by immunohistochemical staining. A positive result indicates the presence of hormone receptor proteins on the surface of tumor cells. These proteins, when bound to hormones in the body, activate the proliferation signaling pathway. Only patients with positive receptors are suitable for endocrine therapy, because this signaling can only be blocked by using endocrine drugs to competitively bind to receptors or lower hormone levels in the presence of receptors. The actual values ​​of each indicator are compared with a pre-defined pathological indicator coding table, converting the actual values ​​into corresponding coded values ​​to complete baseline risk coding. The pathological indicator coding table is a mapping rule established based on general standards in the field of breast cancer prognostic assessment. It maps the clinical descriptive values ​​of each indicator to numerical codes. For example, tumor diameter is discretized by intervals, the number of lymph node metastases is directly coded numerically, and histological grade and hormone receptor status are assigned fixed codes based on qualitative results. The reason for coding is that prognostic assessment models require uniform numerical input, but the same indicator is described differently in pathology reports from different hospitals. Coding converts these heterogeneous descriptions into a standardized, computable form.

[0044] The encoded values ​​are input into a pre-defined breast cancer prognostic assessment model to calculate a baseline recurrence risk score. This breast cancer prognostic assessment model is a widely validated and universally applicable prognostic tool in the clinical field, belonging to existing technology in this area. Its core is a weighted regression formula derived from long-term follow-up data from a large-scale prospective cohort through survival analysis and statistical fitting. Each pathological indicator in the formula is assigned a different weight coefficient, reflecting its independent contribution to the risk of long-term recurrence. For example, the weight of the number of lymph node metastases is usually much higher than other indicators because lymph node metastasis signifies that tumor cells have acquired the biological ability to invade and metastasize; the maximum tumor diameter and proliferation index have secondary weights; hormone receptor status has a specific weight direction in endocrine therapy patients because receptor positivity indicates a baseline response to endocrine therapy, which is weakened if medication is not taken correctly. Substituting the encoded pathological features into this weighted formula yields a numerical baseline recurrence risk score, representing the theoretical probability of distant recurrence within a specific time window after surgery, assuming the patient strictly adheres to the prescribed endocrine therapy regimen. The specific calculation formula and internal parameters of the breast cancer prognostic assessment model are existing technologies and will not be elaborated here.

[0045] In another preferred embodiment of the present invention, the specific generation process of the temporal characteristics of the treatment response physiological index temporal trajectory in the behavior trajectory module is as follows: The liver enzyme test values ​​corresponding to each relative time coordinate are extracted from the time-series trajectory of the physiological indicators of treatment response. The time-series trajectory of the physiological indicators of treatment response consists of liver enzyme test values ​​of the same sample at each follow-up node arranged in order of relative time coordinates. The time nodes in the time-series trajectory of the physiological indicators of treatment response show obvious turning points in liver enzyme test values. The nodes where the liver enzyme test values ​​change from stable or decreasing to continuously increasing are marked as abnormal starting points. The nodes where the liver enzyme test values ​​change from increasing to continuously decreasing are marked as peak points. The nodes where the liver enzyme test values ​​fall back to the normal range are marked as recovery points. The duration of the abnormal fluctuation interval formed by adjacent abnormal starting points and recovery points, the peak value of the liver enzyme test values ​​within the interval, the fluctuation amplitude of the liver enzyme test values ​​within the interval, and the rate of change of the liver enzyme test values ​​from the abnormal starting point to the peak point are extracted as interval feature parameters of the abnormal fluctuation interval. The interval feature parameters of all abnormal fluctuation intervals of the same sample are combined in time order to generate the time-series features of the physiological indicators of treatment response of the sample.

[0046] In another preferred embodiment of the present invention, the specific process of trend segmentation identification of the actual treatment execution time trajectory in the behavior trajectory module is as follows: The liver enzyme test values ​​corresponding to each relative time coordinate were extracted from the time-series trajectory of physiological indicators of treatment response. This trajectory is composed of liver enzyme test values ​​of the same sample at each follow-up node arranged in order of relative time coordinate. The liver enzyme test values ​​are the alanine aminotransferase and aspartate aminotransferase activities measured at each follow-up blood draw. These two types of enzymes are mainly distributed in the cytoplasm and mitochondria of hepatocytes. When the hepatocyte membrane is intact, only a small amount is released into the blood. When hepatocytes are stimulated by drugs and undergo a stress response, the cell membrane permeability increases, and a large amount of enzyme proteins are released from the cell into the blood, and the serum enzyme activity increases significantly. Therefore, liver enzyme values ​​can objectively reflect the real-time physiological stress intensity of hepatocytes under drug exposure. The identification trajectory identifies the time points where liver enzyme test values ​​show significant turning points: the turning point where liver enzyme values ​​change from stable or decreasing to continuously increasing is marked as the abnormal start point, signifying the beginning of an abnormal liver function event; the turning point where values ​​change from increasing to continuously decreasing is marked as the peak point, representing the most severe moment of hepatocyte stress and also the turning point where the body's repair mechanism begins to reverse the damage; when liver enzyme values ​​continue to decrease and return to the normal range, it is marked as the recovery point, meaning that hepatocyte membrane permeability has basically returned to normal, and the abnormal event has ended. Interval characteristic parameters were extracted from the abnormal fluctuation intervals formed by adjacent abnormal onset points and recovery points: the interval duration is the time span from the abnormal onset point to the recovery point, reflecting the complete cycle of liver injury from initiation to resolution; the peak value of liver enzymes within the interval is the maximum reading of liver enzymes within the interval, reflecting the strongest degree of hepatocellular stress; the fluctuation amplitude is the difference between the peak value and the liver enzyme value at the abnormal onset point, reflecting the absolute increase of liver enzymes from baseline to peak value, eliminating the influence of individual patient baseline differences; the rate of change from the abnormal onset point to the peak point is the fluctuation amplitude divided by the duration of the rising segment, reflecting the sensitivity of hepatocellular cells to drug stimulation. The interval characteristic parameters of all abnormal fluctuation intervals of the same sample were combined in chronological order to generate the time-series characteristics of physiological indicators of treatment response.

[0047] While the original sequence contains all the information about changes in liver function, it is contaminated with detection noise and normal physiological fluctuations, making it susceptible to interference from irrelevant signals when used directly for cross-validation. By identifying inflection points, the complete trajectory is segmented into a modular structure with abnormal fluctuation intervals as the basic units. Each interval corresponds to an independent abnormal liver function event, with clear temporal boundaries and well-defined event attributes. Four feature parameters comprehensively characterize each event from four dimensions: time, intensity, relative change, and rate of change. This allows subsequent cross-validation to accurately locate abnormal fluctuation intervals that match a specific concentration decrease time interval. By comparing the start and end times of the two intervals, it is determined whether there is temporal coupling between the concentration decrease and the decline in liver enzymes, providing a structured basis for the automatic classification of behaviorally related and metabolically related decline segments.

[0048] In another preferred embodiment of the present invention, the specific process of generating behavioral trajectory features in the behavioral trajectory module is as follows: From the trend segmentation identification results, each trend segment of the concentration decrease segment was extracted, and the start and end time coordinates of the segment were obtained to form the time interval to be verified. The trend segmentation identification results have divided the actual treatment execution time trajectory into rising, falling, and plateau segments. The concentration decrease segment corresponds to the time interval in which the serum active metabolite concentration continues to decline, which may be caused by dose reduction, drug discontinuation, or metabolic abnormalities, and the cause needs to be identified segment by segment. From the time sequence characteristics of the physiological indicators of the treatment response of the same sample, the interval feature parameters of the abnormal fluctuation interval corresponding to the time interval to be verified were extracted. If the liver enzyme detection value in the interval feature parameters shows a continuous downward trend within the time interval to be verified, the concentration decrease segment is marked as a behavior-related decrease segment. The continuous decrease of liver enzymes means that the hepatocyte stress level is falling back and the cell membrane permeability is recovering at the same time as the concentration decrease. This is consistent with the physiological law of the liver entering the repair process after drug exposure withdrawal. The time synchronization relationship between the concentration decrease and the liver enzyme decline constitutes an objective biological evidence chain for the patient's dose reduction or drug discontinuation due to liver damage side effects. If the liver enzyme detection values ​​in the interval feature parameters do not show a continuous downward trend within the time interval to be verified, including situations where liver enzymes remain at a high level, remain within the normal range, or even continue to rise, then this concentration decrease segment is marked as a metabolism-related decrease segment. This indicates that there is no temporal coupling relationship between the occurrence of concentration decrease and liver injury recovery, and the driving factor for dose reduction or discontinuation is not a side effect of liver injury, but may involve other reasons such as amnesia, cognitive bias, or economic factors. The trend segments in the trend segmentation identification results that are of the concentration increase segment and concentration plateau segment retain their original segment type and segment attribute parameters. These are then combined with the marked behavior-related decrease segments and metabolism-related decrease segments according to the original arrangement order of each trend segment in the trajectory to form the execution behavior trajectory features of the sample.

[0049] While a decrease in drug concentration is understandable as an objective biochemical phenomenon, its underlying clinical causes are quite different: discontinuing medication due to liver damage side effects forces patients to compromise between side effects and the risk of relapse, while discontinuation due to forgetfulness or cognitive factors can be corrected through enhanced medication education. By categorizing the concentration decrease into behavioral and metabolic-related types, the risk correction model can learn the differentiated associations between the two types of decrease patterns and long-term relapse outcomes during training, rather than treating all concentration decrease events equally. This allows for more accurate dynamic risk calibration results in application.

[0050] In another preferred embodiment of the present invention, the specific construction process of the dynamically corrected feature set in the risk correction module is as follows: The time intervals covered by each trend segment in the execution behavior trajectory features are aligned with the time intervals covered by each abnormal fluctuation interval in the treatment response physiological index time sequence features, so that the segment attribute parameters of each trend segment are mapped to the interval feature parameters of the abnormal fluctuation interval within the time interval covered by the trend segment. The segment type identifier, segment attribute parameters and interval feature parameters of the abnormal fluctuation interval mapped by the trend segment are concatenated to generate a joint feature vector of each trend segment. The joint feature vectors of each trend segment are concatenated according to the original arrangement order of each trend segment in the trajectory to generate a dynamic correction feature set for each sample.

[0051] In another preferred embodiment of the present invention, the specific generation process of the risk correction model in the risk correction module is as follows: The actual follow-up durations of each sample in the prognostic training sample set were arranged in ascending order, and the follow-up time points where recurrence events occurred were extracted as the time points to be calculated. The actual follow-up duration is the interval from the patient's surgery date to the date of the recurrence event or the last follow-up date. The time point of recurrence event occurrence was specifically extracted after sorting, rather than the censored time point, because the basic principle of survival analysis is to use the risk set at the time of recurrence event occurrence to estimate the instantaneous recurrence risk. Only the time point of recurrence event occurrence contains information about the recurrence risk of the population at that time. Censorship does not produce recurrence events; it only affects the number of people in the risk set after that time point. For each time point to be calculated, the number of samples with recurrence events at that time point was counted as the event count, and the number of samples that had not yet had a recurrence event and were not censored at the beginning of that time point was counted as the risk set sample size. The event count was divided by the risk set sample size to obtain the instantaneous recurrence risk rate at that time point. The statistical logic behind this ratio is that at a given moment, all individuals still under observation and not yet experiencing recurrence constitute the risk set for that moment. The frequency of recurrence events observed within this risk set reflects the intensity of the instantaneous recurrence risk at that moment. Dividing the number of events by the sample size of the risk set provides the most direct and unbiased estimate of this risk intensity. From the surgery date to each time point to be calculated, the complement of the instantaneous recurrence risk rates at each time point is multiplied together to obtain the cumulative survival probability without recurrence at that time point. The complement of this cumulative survival probability without recurrence is then used to obtain the cumulative recurrence probability at that time point. The complement is represented by subtracting the instantaneous recurrence risk rate from one to represent the probability of successfully avoiding recurrence at that moment. The multiplication operation reflects the progressive nature of the survival process; only by successfully avoiding recurrence at each preceding time point can one maintain a recurrence-free state at the target time point. Therefore, the product of the probabilities at each time point is the cumulative survival probability without recurrence from the starting point to that moment. Taking the complement again yields the cumulative recurrence probability, and the sum of the two is 100%. The time points to be calculated and their corresponding cumulative recurrence probabilities are combined in chronological order to form a recurrence probability time series, which is then used as a supervision signal for model training.

[0052] The initial prognostic risk label and dynamically adjusted feature set of each sample are used as joint input, and the recurrence probability time series is used as the supervision signal to iteratively train the pre-defined prognostic risk correction network. At the start of training, the network parameters are randomly initialized. The network is then fed the initial risk label and dynamically adjusted feature set of a single sample. The network calculates and outputs the predicted recurrence probability after correction layer by layer. The difference between this predicted value and the supervision signal is the bias. During training, the bias is propagated backward, and the parameters of each layer are fine-tuned in the direction that reduces the bias, completing one update. This cycle is repeated for all samples, and the bias gradually decreases with each iteration. Training stops when the bias converges to a preset threshold, and the network parameters at this point are stored, thus forming the risk correction model.

[0053] Traditional prognostic assessments only output static risk values, failing to reflect risk changes during treatment. Cumulative relapse probability curves, however, record the cumulative relapse levels of a historical population at different time points, containing risk information over time. Using this curve as a supervisory signal allows the model to learn not only whether a relapse will occur, but also the trajectory of relapse risk changes throughout the entire follow-up period. During training, the model automatically learns the impact of different types and degrees of drug exposure interruption on relapse risk by comparing the cumulative relapse probabilities of samples with the same initial risk label but different dynamic correction features. The behaviorally related decline segment, due to objective evidence of a substantial interruption in drug exposure—accompanied by a synchronous decline in liver enzymes—has a mapping relationship with long-term relapse probability that is distinct from the metabolically related decline segment through parameter fitting. After training, the model possesses the ability to output individually corrected relapse probabilities based solely on initial risk levels and biochemical dynamic data. This allows for the timely identification of patients whose true risk is elevated due to insufficient actual drug exposure, hidden beneath a static low-risk label, thus providing a window of opportunity for clinical intervention.

[0054] In another preferred embodiment of the present invention, the specific process of outputting the calibrated current prognostic risk level in the recurrence assessment module is as follows: The system acquires real-time serum active metabolite concentrations, real-time liver enzyme values, and initial pathological features of the target patient. Real-time concentration and liver enzyme values ​​are the latest data extracted from the hospital information system at the most recent follow-up visit. Initial pathological features include the maximum tumor diameter, number of lymph node metastases, histological grade, proliferation index, and hormone receptor status, all of which were entered into the system at the time of the patient's postoperative diagnosis. These initial pathological features are input into a breast cancer prognostic assessment model. Following the same coding and scoring process as the initial risk module, the initial prognostic risk level of the target patient is determined. The initial prognostic risk level, real-time concentration values, and real-time liver enzyme values ​​are then input into a trained risk correction model. The initial risk level provides a theoretical recurrence risk benchmark under ideal medication conditions. Real-time concentration values ​​reflect the current actual drug exposure level in the body, and real-time liver enzyme values ​​reflect the current physiological stress state of hepatocytes. Based on the mapping relationship between drug exposure trajectory and recurrence probability learned during the training phase, the model automatically calculates and outputs the corrected cumulative recurrence probability sequence of the target patient at multiple future time points, such as the predicted cumulative recurrence probability values ​​for the third, fifth, and tenth years post-surgery. The system extracts the corrected recurrence probability value corresponding to the preset assessment time point from the corrected cumulative recurrence probability sequence. The preset assessment time point is usually the long-term prognostic assessment time window recommended by clinical guidelines. The extracted probability value is compared with the preset risk grading threshold. If it is lower than the low-risk threshold, it is judged as low-risk; if it is between the low-risk and high-risk thresholds, it is judged as intermediate-risk; and if it is higher than the high-risk threshold, it is judged as high-risk. The comparison result is the calibrated current prognostic risk level, which is output by the system and marked in the electronic medical record.

[0055] It is understandable that the initial risk label only reflects the theoretical risk under the ideal medication assumption, while real-time concentrations and liver enzymes carry all the actual exposure and response information of the patient from the start of treatment to the present. The model integrates both to correct the static baseline to a dynamic relapse probability that matches the actual treatment implementation. Outputting a sequence of relapse probabilities for multiple future time points instead of a single value is because the distribution of relapse risk varies across different patients over time. The sequence format fully preserves this temporal information, allowing clinicians to judge the trend and urgency of risk changes. The probability values ​​for preset assessment time points are extracted and compared with grading thresholds, transforming the model's probability output into a clinically usable grading decision-making basis. This allows patients whose actual risk is rising due to insufficient actual drug exposure, hidden beneath the static low-risk label, to be identified and their ratings upgraded in a timely manner, gaining a valuable time window for adjusting treatment plans and increasing follow-up frequency.

[0056] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A breast surgery prognostic assessment system based on clinical data, characterized in that, include: Sample acquisition module: used to acquire time-series clinical follow-up data of historical patient cohorts within a complete follow-up period, extract serum active metabolite concentration detection value sequences and liver enzyme detection value sequences from the time-series clinical follow-up data, and construct a prognostic training sample set; Initial risk module: used to perform baseline risk coding on the initial pathological features of each sample in the prognostic training sample set, and generate an initial prognostic risk label for each sample; Treatment execution module: used to extract the serum active metabolite concentration detection value sequence of each sample from the prognostic training sample set, generate the actual treatment execution time trajectory of each sample, and extract the liver enzyme detection value sequence of each sample from the prognostic training sample set, generate the treatment response physiological index time trajectory of each sample. Behavioral trajectory module: used to extract the temporal features of the treatment response physiological index time-series trajectory, and when performing trend segmentation identification on the actual treatment execution time-series trajectory, use the temporal features to cross-validate the concentration decrease segment, and generate execution behavioral trajectory features carrying behavioral-related decrease segment markers; Risk correction module: used to align the execution behavior trajectory features with the time-series features of the treatment response physiological indicators in the time dimension and fuse the features to construct a dynamic correction feature set. The risk correction model is trained with the initial prognostic risk label and the dynamic correction feature set as input and the long-term follow-up outcome of each sample as the supervision label. Relapse assessment module: used to obtain real-time serum active metabolite concentration and real-time liver enzyme detection values ​​of target patients, extract real-time treatment execution quantitative indicators and real-time treatment response physiological indicators and input them into the risk correction model, and output the calibrated current prognostic risk level.

2. The breast surgery prognostic assessment system based on clinical data according to claim 1, characterized in that, The specific process of constructing the prognostic training sample set in the sample acquisition module is as follows: Time-series clinical follow-up data of historical breast cancer patient cohorts within a complete follow-up period were extracted from hospital information systems and laboratory information systems. The time-series clinical follow-up data included serum active metabolite concentrations, liver enzyme values, and long-term follow-up outcomes at each follow-up node. The time-series clinical follow-up data were then subjected to longitudinal data association on a patient-by-patient basis. The serum active metabolite concentrations and liver enzyme values ​​of the same patient at each follow-up node were arranged in chronological order of follow-up time, forming an individualized time-series data chain for each patient. The individualized time-series data chains are cleaned to remove samples with a follow-up node missing rate exceeding a preset threshold and samples with missing key variables. Abnormal values ​​of serum active metabolite concentration and liver enzyme detection values ​​in the remaining samples are identified and marked. The individualized time-series data chains of each patient after cleaning are aligned with the follow-up time window. The detection time of each follow-up node is converted into a relative time coordinate in months, with the patient's surgery date as the baseline zero point. The individualized time-series data chains of each patient after cleaning and time alignment are integrated into a prognostic training sample set.

3. The breast surgical prognostic assessment system based on clinical data according to claim 1, characterized in that, The specific process for generating the initial prognostic risk label for each sample in the initial risk module is as follows: The initial pathological features of each sample were extracted from the prognostic training sample set. The initial pathological features included the maximum diameter of the tumor, the number of lymph node metastases, histological grade, proliferation index, estrogen receptor status, and progesterone receptor status. The actual values ​​of each indicator in the initial pathological features of each sample are compared with the preset pathological indicator coding table, and then the actual values ​​of each indicator are converted into corresponding coding values. The initial pathological feature coding values ​​of each sample after baseline risk coding are input into the preset breast cancer prognostic assessment model to calculate the baseline recurrence risk score of each sample. The comparison results of the baseline recurrence risk score and the preset risk grading threshold determine the initial prognostic risk level of each sample. The initial prognostic risk level of each sample is used as the initial prognostic risk label for that sample, and then associated and stored with the corresponding sample in the prognostic training sample set.

4. The breast surgical prognostic assessment system based on clinical data according to claim 1, characterized in that, In the behavioral trajectory module, the specific generation process of the temporal characteristics of the treatment response physiological index time-series trajectory is as follows: The liver enzyme test values ​​corresponding to each relative time coordinate are extracted from the time-series trajectory of the physiological indicators of treatment response. The time-series trajectory of the physiological indicators of treatment response consists of liver enzyme test values ​​of the same sample at each follow-up node arranged in order of relative time coordinates. The time nodes in the time-series trajectory of the physiological indicators of treatment response show obvious turning points in liver enzyme test values. The nodes where the liver enzyme test values ​​change from stable or decreasing to continuously increasing are marked as abnormal starting points. The nodes where the liver enzyme test values ​​change from increasing to continuously decreasing are marked as peak points. The nodes where the liver enzyme test values ​​fall back to the normal range are marked as recovery points. The duration of the abnormal fluctuation interval formed by adjacent abnormal starting points and recovery points, the peak value of the liver enzyme test values ​​within the interval, the fluctuation amplitude of the liver enzyme test values ​​within the interval, and the rate of change of the liver enzyme test values ​​from the abnormal starting point to the peak point are extracted as interval feature parameters of the abnormal fluctuation interval. The interval feature parameters of all abnormal fluctuation intervals of the same sample are combined in time order to generate the time-series features of the physiological indicators of treatment response of the sample.

5. The breast surgical prognostic assessment system based on clinical data according to claim 1, characterized in that, In the behavior trajectory module, the specific process of trend segmentation identification of the actual treatment execution time trajectory is as follows: The actual treatment execution timeline trajectory of each sample is obtained from the prognostic training sample set. The actual treatment execution timeline trajectory is composed of the serum active metabolite concentration detection values ​​of the same sample at each follow-up node arranged in relative time coordinate order. The difference between the serum active metabolite concentration detection values ​​at adjacent time nodes in the actual treatment execution timeline trajectory is calculated. The segments with positive and continuous differences are divided into concentration rising segments, the segments with negative and continuous differences are divided into concentration falling segments, and the segments with absolute differences less than a preset fluctuation threshold and continuous differences are divided into concentration plateau segments. Extract the starting and ending concentration values ​​of each trend segment. The difference between the ending and starting concentration values ​​is taken as the segment change amplitude of the trend segment. The segment change amplitude is divided by the duration of the trend segment to obtain the segment change rate. The segment type, starting time coordinate, ending time coordinate, duration, change amplitude, and change rate of each trend segment are combined according to their arrangement order in the trajectory to form the trend segmentation identification result of the sample.

6. The breast surgical prognostic assessment system based on clinical data according to claim 5, characterized in that, In the behavior trajectory module, the specific process for generating behavior trajectory features is as follows: From the trend segmentation identification results, each trend segment of the type of concentration decrease segment is extracted one by one, and the start time coordinate and end time coordinate of the concentration decrease segment are obtained to form the time interval to be verified for the concentration decrease segment. From the time-series characteristics of physiological indicators of treatment response of the same sample, the interval feature parameters of abnormal fluctuation intervals corresponding to the time interval to be verified are extracted. If the liver enzyme detection value in the interval feature parameters shows a continuous downward trend in the time interval to be verified, the concentration decrease segment is marked as a behavior-related decrease segment. If the liver enzyme detection value in the interval feature parameters does not show a continuous downward trend in the time interval to be verified, the concentration decrease segment is marked as a metabolism-related decrease segment. The trend segments in the trend segmentation identification results that are of the concentration rising segment and the concentration plateau segment retain their original segment type and segment attribute parameters. These segments are then combined with the labeled behavior-related falling segments and metabolism-related falling segments according to their original arrangement order in the trajectory to form the execution behavior trajectory features of the sample.

7. The breast surgical prognostic assessment system based on clinical data according to claim 1, characterized in that, The specific construction process of the dynamically corrected feature set in the risk correction module is as follows: The time intervals covered by each trend segment in the execution behavior trajectory features are aligned with the time intervals covered by each abnormal fluctuation interval in the treatment response physiological index time sequence features, so that the segment attribute parameters of each trend segment are mapped to the interval feature parameters of the abnormal fluctuation interval within the time interval covered by the trend segment. The segment type identifier, segment attribute parameters and interval feature parameters of the abnormal fluctuation interval mapped by the trend segment are concatenated to generate a joint feature vector of each trend segment. The joint feature vectors of each trend segment are concatenated according to the original arrangement order of each trend segment in the trajectory to generate a dynamic correction feature set for each sample.

8. The breast surgical prognostic assessment system based on clinical data according to claim 1, characterized in that, The specific generation process of the risk correction model in the risk correction module is as follows: Arrange the actual follow-up durations of each sample in the prognostic training sample set in ascending order, and extract the follow-up time points where recurrence events occurred in each actual follow-up duration as the time points to be calculated. For each time point to be calculated, the number of samples that recurred at that time point is counted as the number of events, and the number of samples that have not yet recurred and have not been deleted at the beginning of that time point is counted as the number of risk set samples. The number of events is divided by the number of risk set samples to obtain the instantaneous recurrence risk rate at that time point. From the surgery date to each time point to be calculated, the instantaneous recurrence risk rates of each time point to be calculated are multiplied together after taking the complement of each rate, to obtain the cumulative survival probability without recurrence at that time point. The cumulative survival probability without recurrence is then multiplied by taking the complement of the cumulative survival probability without recurrence to obtain the cumulative recurrence probability at that time point. Each time point to be calculated and its corresponding cumulative recurrence probability are combined in chronological order to form a recurrence probability time series, which is used as a supervision signal for model training. Using the initial prognostic risk label and dynamic correction feature set of each sample as joint input, and the recurrence probability time series as supervision signal, the preset prognostic risk correction network is iteratively trained. The network parameters are adjusted until the deviation between the corrected recurrence probability output by the network and the supervision signal converges to a preset threshold. The trained prognostic risk correction network is then stored as a risk correction model.

9. A breast surgical prognostic assessment system based on clinical data according to claim 1, characterized in that, The specific process for outputting the calibrated current prognostic risk level in the recurrence assessment module is as follows: The system acquires real-time serum active metabolite concentrations, real-time liver enzyme levels, and initial pathological features of the target patient. It inputs the initial pathological features into a breast cancer prognostic assessment model to determine the initial prognostic risk level of the target patient. The initial prognostic risk level, the real-time serum active metabolite concentrations, and the real-time liver enzyme levels are then input into a trained risk correction model, which outputs a corrected cumulative recurrence probability sequence for the target patient at multiple future time points. The corrected recurrence probability value corresponding to a preset assessment time point is extracted from the corrected cumulative recurrence probability sequence. Based on the comparison between this corrected recurrence probability value and a preset risk grading threshold, the system determines and outputs the calibrated current prognostic risk level of the target patient.

Images