A method and system for monitoring the risk of postoperative recurrence of lung nodules

Abstract

The application discloses a kind of lung nodule postoperative recurrence risk monitoring method and system, it is related to postoperative risk monitoring technical field, obtain each time postoperative follow-up data sorted by time, secondly according to the time interval and recurrence risk characteristic identification disease turning point event of each follow-up relative to previous follow-up benchmark, then time interval and disease turning point event combination generates course stage weight.Due to course stage weight simultaneously characterizes follow-up span and feature change state, this method makes the relative role of same follow-up in the whole postoperative course can be described by unified data amount, so as to reduce the time sequence deviation generated when direct aggregation of unequal interval follow-up data, and make the risk change corresponding to disease turning point point can obtain the weight adapted to its course state in subsequent processing.

Description

Technical Field

[0001] This invention relates to the field of postoperative risk monitoring technology, specifically to a method and system for monitoring the risk of recurrence after pulmonary nodule surgery. Background Technology

[0002] Postoperative recurrence risk monitoring of pulmonary nodules usually requires continuous analysis of changes in the condition by combining imaging and clinical examination data collected at different follow-up time points.

[0003] Chinese invention patent publication number CN108898588A discloses a time-series-based method for evaluating treatment efficacy. By acquiring lung CT image data at multiple time points and analyzing the changes in lung nodules based on time series, the method can evaluate treatment efficacy and predict early recurrence.

[0004] The above-mentioned approach can reflect changes in pulmonary nodules using data from different time points. However, the actual postoperative follow-up intervals are not fixed, and significant changes in nodule size, solid component, or other risk characteristics may occur at different follow-up time points. For such unequal-interval follow-up data, there is still room for improvement in determining the contribution of each follow-up data point to overall risk monitoring by combining the follow-up time span and the turning point of the disease. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring the risk of recurrence after pulmonary nodule surgery, so as to solve the technical problems existing in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the risk of recurrence after pulmonary nodule surgery, comprising: Obtain postoperative follow-up data for each patient, sorted by follow-up time. Each postoperative follow-up data includes the time interval between each follow-up and the previous follow-up baseline, as well as the recurrence risk characteristics for each follow-up. Disease turning points are identified based on the time interval between each follow-up visit relative to the previous follow-up baseline and the risk characteristics of each relapse. Disease stage weights are generated based on the time interval between each follow-up visit relative to the previous follow-up baseline and the disease turning points. Based on the risk characteristics of each relapse, a follow-up feature vector is generated, the attention matching score of each follow-up feature vector is calculated, the attention matching score is corrected using the weight of the disease stage to obtain the correction result, and the visit-level attention weight corresponding to each follow-up is determined based on the correction result. The feature vectors of each follow-up visit are aggregated according to the attention weight of each visit, and the postoperative global health characterization data is generated. The recurrence risk monitoring results are output based on the postoperative global health characterization data.

[0007] As one possible design, disease stage weights are generated based on the time interval between each follow-up visit relative to the previous follow-up baseline and disease turning events, including: Determine the changes in nodule size, percentage of solid component, tumor markers, and clinical symptoms relative to the baseline values ​​at the t-th follow-up. Relapse risk features that meet the corresponding change conditions are marked as candidate turning point features, and disease turning point event identifiers are generated based on the type of candidate turning point features. The time interval between the t-th follow-up and the previous follow-up is normalized. The normalization result is combined with the correction amount corresponding to the disease turning point event to obtain the disease stage weight of the t-th follow-up.

[0008] As one possible design, the attention weights for each follow-up visit are determined based on the correction results, including: Each follow-up feature vector is used to obtain the first mapping feature and the second mapping feature through the first feature mapping matrix and the second feature mapping matrix, respectively; An attention matching score is generated based on the matching relationship between the first mapping feature of any follow-up and the second mapping feature of each follow-up. The attention matching score is scaled according to the disease stage weight of the target follow-up, and all scaling results are normalized to obtain the visit-level attention weight for each follow-up.

[0009] As one possible design, follow-up feature vectors are generated based on the risk characteristics of each recurrence, including: Extract the observation values ​​of each quantifiable observation dimension from each follow-up data, standardize and encode the observation values, and generate corresponding code-level aggregated features; The code-level aggregated features of each follow-up are arranged in chronological order to form a follow-up feature matrix, and the feature data corresponding to each follow-up in the follow-up feature matrix are determined as the corresponding follow-up feature vector.

[0010] As one possible design, the feature vectors of each follow-up visit are aggregated according to the attention weights corresponding to each visit, generating postoperative global health representation data. Based on this data, recurrence risk monitoring results are output, including: The attention weights at each follow-up visit are weighted with the corresponding follow-up feature vectors, and the resulting weighted features are aggregated to generate postoperative global health representation data. Determine the similarity of postoperative global health characteristics data with the average characteristics of historically high-risk patient groups and the average characteristics of historically low-risk patient groups, respectively. The recurrence risk assessment value is determined by combining the total recurrence risk score for the current period, and the recurrence risk monitoring results are output based on the recurrence risk assessment value.

[0011] As one possible design, the method also includes: Establish a database of indicators for monitoring the risk of postoperative recurrence of pulmonary nodules. The database includes multiple categories of monitoring indicators, and each category includes at least one quantifiable observation dimension. Based on the patient's preoperative baseline examination data, intraoperative pathological data, and postoperative first follow-up data, the pathological risk parameters and surgical procedure influence parameters corresponding to each monitoring indicator category are determined. The combined results of the pathological risk parameters and surgical procedure influence parameters are normalized to obtain the initial real-time monitoring weights for each monitoring indicator category. The initial real-time monitoring weights are calibrated based on the proportion of real components and the matching relationship between indicator categories.

[0012] As a possible design, when adding new monitoring dimensions to the monitoring indicator library, the method also includes: Determine the correlation between the newly added monitoring dimensions and the original observation dimensions in the monitoring indicator category to which they belong; Based on the correlation, the relationship between the new monitoring dimension and the recurrence risk, and the real-time monitoring weights of each monitoring indicator category before adjustment, the updated weights of the monitoring indicator categories to which the new monitoring dimension belongs are determined. The real-time monitoring weights of the remaining monitoring indicator categories are adjusted synchronously based on the updated weights, so that the sum of the real-time monitoring weights of each monitoring indicator category remains unchanged after the adjustment.

[0013] As one possible design, after outputting relapse risk monitoring results, the method also includes: The core monitoring targets for the current period are determined based on the real-time monitoring weights of each monitoring indicator category, and monitoring tasks corresponding to the core monitoring targets for the current period are pushed out. The real-time monitoring weights and monitoring intervals of the monitoring indicator categories to which the core monitoring targets belong for the current period are updated based on the recurrence risk monitoring results, and the real-time monitoring weights of the other monitoring indicator categories are adjusted simultaneously.

[0014] As a possible design, updating real-time monitoring weights and monitoring intervals includes: When the recurrence risk assessment value is within the safe range and the cumulative number of times the current core monitoring target is selected reaches the corresponding threshold, the real-time monitoring weight of the monitoring indicator category to which the current core monitoring target belongs is reduced, the corresponding monitoring interval is extended, and the cumulative number of selections is cleared. When the relapse risk assessment value exceeds the safe range, the real-time monitoring weight of the corresponding monitoring indicator category is increased, the corresponding monitoring interval is shortened, and joint follow-up information including the relapse risk assessment value, historical monitoring data and recent re-examination data is sent to the multidisciplinary follow-up terminal.

[0015] On the other hand, the present invention also provides a system for monitoring the risk of recurrence after pulmonary nodule surgery, comprising: The data acquisition module is used to acquire postoperative follow-up data of patients sorted by follow-up time. Each postoperative follow-up data includes the time interval between each follow-up and the previous follow-up baseline and the recurrence risk characteristics of each follow-up. The stage weight generation module is used to identify disease turning points based on the time interval between each follow-up visit relative to the previous follow-up baseline and the risk characteristics of each relapse, and to generate disease stage weights based on the time interval between each follow-up visit relative to the previous follow-up baseline and the disease turning points. The attention correction module is used to generate follow-up feature vectors based on the risk characteristics of each relapse, calculate the attention matching score of each follow-up feature vector, correct the attention matching score using the weight of the disease stage, and determine the corresponding consultation-level attention weight based on the correction result. The risk output module is used to aggregate the feature vectors of each follow-up visit according to the attention weight of each visit, generate postoperative global health characterization data, and output the recurrence risk monitoring results based on the postoperative global health characterization data.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. Compared to methods that aggregate postoperative data solely based on follow-up sequence, this invention first acquires postoperative follow-up data ordered by time. Then, it identifies disease turning points based on the time intervals between each follow-up and the previous follow-up baseline, along with recurrence risk characteristics. Finally, it combines the time intervals with these turning points to generate disease stage weights. Since these stage weights simultaneously represent the follow-up duration and the state of characteristic changes, the relative role of the same follow-up in the entire postoperative disease course can be described by a uniform amount of data. This reduces temporal bias caused by direct aggregation of unequal-interval follow-up data and ensures that the risk changes corresponding to disease turning points receive weights appropriate to their disease progression in subsequent processing.

[0017] 2. Compared to the approach of directly determining aggregation weights based on the matching relationships between features from each follow-up visit, this invention first generates feature vectors for each follow-up visit, then calculates the attention matching score between each vector, corrects the matching score using the corresponding disease stage weights, determines the consultation-level attention weights for each follow-up visit, and finally performs weighted aggregation on each follow-up feature vector to output the recurrence risk monitoring results. Because the attention weights are constrained by disease stage information before normalization, feature relevance and disease location jointly contribute to the formation of the postoperative global health representation, thereby improving the consistency of the contribution allocation of information from each follow-up visit and providing structured data for subsequent risk stratification, adjustment of monitoring indicator weights, and adjustment of follow-up intervals. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0019] Figure 1 This is a flowchart of the monitoring method in this invention.

[0020] Figure 2 This is a diagram of the monitoring system architecture in this invention.

[0021] Figure 3 This is a schematic diagram illustrating the generation of disease stage weights in this invention.

[0022] Figure 4 This is a schematic diagram illustrating the determination of the attention weights at the consultation level for each follow-up visit in this invention.

[0023] Figure 5 This is a schematic diagram of recurrence risk monitoring in this invention.

[0024] Figure 6 This is a schematic diagram of the monitoring interval update in this invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0026] like Figure 1 As shown, the monitoring method involved in this embodiment first acquires postoperative follow-up data of the patient sorted by follow-up time. It then identifies disease turning points based on the time interval between each follow-up and the previous follow-up baseline, and the disease recurrence risk characteristics of each follow-up. Based on the time interval between each follow-up and the disease turning points, it generates disease stage weights. Subsequently, it generates follow-up feature vectors based on the recurrence risk characteristics of each follow-up, calculates the attention matching score of each follow-up feature vector, corrects the attention matching score using the disease stage weights, and determines the consultation-level attention weight corresponding to each follow-up based on the correction result. Finally, it aggregates each follow-up feature vector according to the consultation-level attention weight corresponding to each follow-up, generates postoperative global health representation data, and outputs recurrence risk monitoring results based on the postoperative global health representation data.

[0027] Example 1: This example provides a system for monitoring the risk of recurrence after pulmonary nodule surgery, and describes the process of data input, indicator library establishment and initial weight configuration of the system.

[0028] like Figure 2 As shown, the monitoring system includes a data acquisition module, a stage weight generation module, an attention correction module, and a risk output module. The data acquisition module acquires postoperative follow-up data for each patient, sorted by follow-up time. Each postoperative follow-up data includes the time interval between each follow-up and the previous follow-up baseline, as well as the recurrence risk characteristics for each follow-up. The stage weight generation module identifies disease turning points based on the time interval between each follow-up and the previous follow-up baseline and the recurrence risk characteristics, and generates disease stage weights based on these events. The attention correction module generates follow-up feature vectors based on the recurrence risk characteristics, calculates the attention matching score for each feature vector, corrects the attention matching score using the disease stage weights, and determines the corresponding consultation-level attention weight for each follow-up based on the correction results. The risk output module aggregates the follow-up feature vectors according to the consultation-level attention weights corresponding to each follow-up, generates postoperative global health representation data, and outputs the recurrence risk monitoring results based on the postoperative global health representation data.

[0029] In this embodiment, the monitoring system can be deployed on a hospital follow-up server. The patient's end and the follow-up medical staff's end interact with the server via a network. The patient's end receives monitoring tasks and submits symptom records or notifications of completed examinations. The follow-up medical staff's end inputs or retrieves imaging examinations, serological tests, pulmonary function tests, pathology results, and surgical records. The server stores the above data in association according to patient identification and follow-up time. The server can also be implemented by a combination of an in-hospital business server and a data processing server, as long as it can handle data transfer and processing between modules.

[0030] Postoperative follow-up data is a collection of data from the same patient at different follow-up time points. Each follow-up data point includes at least the follow-up time and one recurrence risk characteristic, which can be derived from imaging features, tumor marker features, clinical symptom features, pulmonary function features, and pathological and surgical features. The time interval between each follow-up relative to the previous follow-up baseline is the difference between the current follow-up time and the previous effective follow-up time; the first postoperative follow-up can use the surgery date or discharge date as the previous time baseline to ensure that each follow-up has comparable disease progression information.

[0031] The system establishes a continuous follow-up sequence based on patient identification and arranges the follow-up data from earliest to latest according to the follow-up time. For data from different sources within the same follow-up, the server uses the examination completion time or report confirmation time as the timestamp for the data and merges data within the same follow-up window into a single follow-up record. Thus, heterogeneous data such as imaging, serological, symptom, and pulmonary function data can form a corresponding set of relapse risk features at the same follow-up time point, providing a consistent data foundation for subsequent disease progression identification and feature vector generation.

[0032] Before forming a continuous follow-up sequence, the server performs consistency checks on patient identification, follow-up time, observation units, and report status. When multiple similar reports exist within the same follow-up window, the confirmed final report is used as valid data; data with missing timestamps, incompatible units, or unconfirmed reports are not used to generate the current follow-up feature vector, nor to trigger candidate inflection features. For the same observation dimension, the data dictionary records uniform units, value directions, and valid ranges, and subsequent follow-ups use the same version of the data dictionary.

[0033] Indicator database establishment module: Establishes an indicator database for monitoring the risk of postoperative recurrence of pulmonary nodules. The indicator database includes... Classified monitoring indicators, each class of monitoring indicators is further divided into The specific observation dimensions and the indicator database cover a total of [number] aspects. Each monitoring indicator has a quantifiable observation dimension and a corresponding real-time monitoring weight. When a patient is discharged after lung nodule surgery, an initial real-time monitoring weight value is assigned to each monitoring indicator based on preoperative baseline examination, intraoperative pathological results, and the first postoperative follow-up data. The weight allocation of the monitoring indicators is simultaneously calibrated to ensure that the initial weight system matches the patient's pathological type (such as invasive adenocarcinoma, carcinoma in situ, etc.) and surgical method (wedge resection, lobectomy, etc.). The monitoring indicators cover imaging characteristics (nodule size, density, edge morphology, proportion of solid component, etc.) and clinical symptom characteristics (presence and severity of symptoms such as cough, chest pain, and hemoptysis). Each indicator is further subdivided into multiple dynamically trackable observation dimensions, and the monitoring indicator library is sent to the dynamic update module.

[0034] The indicator database establishment module first needs to complete the multi-dimensional architecture construction of the indicator database for monitoring the risk of postoperative recurrence of pulmonary nodules. This indicator database covers a total of Classified monitoring indicators, each class of monitoring indicators is further divided into Specific observation dimensions, covering the entire database in total. The system provides quantifiable observation dimensions, with each monitoring indicator assigned a corresponding real-time monitoring weight. The monitoring indicator categories primarily include imaging features, which include observation dimensions such as maximum nodule diameter, solid component diameter, solid component volume percentage, average nodule CT value, standard deviation of solid component CT value, boundary clarity, bronchial translucency, ground-glass opacity, solid component area, and maximum solid component length. All of these dimensions can be extracted through quantitative AI analysis of preoperative and postoperative thin-slice CT (slice thickness ≤ 1 mm). Specifically, for stage IA non-small cell lung cancer patients with a solid component diameter < 1.49 cm and a solid component volume percentage < 22.9%, the negative predictive value for recurrence at 2 years postoperatively can reach 0.95–1.00. These parameters can be directly used as the core quantitative observation dimensions for imaging indicators. When different monitoring indicator categories contain different numbers of observation dimensions, This is a general symbol; in specific processing, it is used as... This represents the number of observation dimensions included in the i-th type of monitoring indicator. Equal to each category The sum of these values. The aforementioned diameters and volume percentages of solid components, along with their corresponding values, are used to illustrate the optional boundaries of the data dictionary and do not imply that the same clinical judgment criteria are applied to all pathological types and all patients; within the same system version, the threshold table should be stored in association with its source dataset and the range of applicable patients.

[0035] The second category is tumor marker characteristics, which includes observation dimensions such as CEA (carcinoembryonic antigen), NSE (neuron-specific enolase), CYFRA21-1 (cytokeratin 19 fragment), and SCC (squamous cell carcinoma antigen). All dimensions are quantified based on serological test results, leaving no room for subjective interpretation. The third category is clinical symptom characteristics, which includes observation dimensions such as cough (none / mild / moderate / severe, graded by duration and frequency), chest pain (none / occasional / persistent, quantified by pain score and frequency of occurrence), hemoptysis (none / small amount of blood streaks in sputum / hemoptysis, graded by amount of bleeding and number of occurrences), unexplained weight loss (quantified by percentage of weight loss per month), and dyspnea (quantified by degree of shortness of breath after activity and 6-minute walking distance). All symptom data are converted into quantitative codes according to preset grading rules. Patient complaints are used as one of the data sources, and the corresponding grades are determined by combining symptom duration, frequency of occurrence, pain score, amount of bleeding, or examination results to ensure that consistent coding standards are used at different follow-up time points. The fourth category is pulmonary function characteristics, which includes observation dimensions such as FEV1 (forced expiratory volume in one second), DLCO (diffusion rate of the lungs), and arterial blood oxygen partial pressure. The quantitative values ​​are directly extracted from the results of postoperative pulmonary function tests and blood gas analysis. The fifth category is pathological and surgical characteristics, which includes observation dimensions such as pathological type (carcinoma in situ / microinvasive adenocarcinoma / invasive adenocarcinoma / squamous cell carcinoma, etc., classified according to postoperative pathology reports), surgical method (wedge resection / segmentectomy / lobectomy, classified according to surgical records), resection range (quantified by the proportion of lung tissue volume removed), and lymph node dissection status (quantified by the number of lymph node groups dissected and the number of positive lymph nodes). All data are derived from medical documents such as intraoperative and postoperative pathology reports and surgical records.

[0036] Upon discharge of the patient after pulmonary nodule surgery, initial real-time monitoring weights were assigned to each monitoring indicator based on preoperative baseline examinations (including preoperative thin-section CT images, serum tumor markers, baseline pulmonary function, etc.), intraoperative pathological results (including nodule pathological type, resection margin, lymph node status, etc.), and the first postoperative follow-up data (including chest X-ray or low-dose CT scan one month postoperatively, first postoperative tumor markers, first postoperative pulmonary function, etc.). The weights were calculated using a weighted formula based on pathological risk stratification. ,in Let be the initial real-time monitoring weight value for the i-th type of monitoring indicator. The pathological risk coefficient corresponding to the i-th monitoring indicator is assigned according to pathological type: 0.2 for carcinoma in situ, 0.4 for minimally invasive adenocarcinoma, 0.6 for lepidic invasive adenocarcinoma, 0.8 for acinar adenocarcinoma, 1.0 for micropapillary adenocarcinoma, 0.7 for squamous cell carcinoma, and 1.0 for small cell carcinoma. The data are from the 5-year recurrence rate statistics of different pathological types. The influence coefficient of surgical method corresponding to the i-th monitoring indicator (0.8 for wedge resection, 0.9 for segmentectomy, and 1.0 for lobectomy, with data sourced from the postoperative residual and recurrence risk statistics of different surgical methods) is calculated by normalizing the product of pathological risk and surgical method influence to achieve the initial weight allocation of each indicator, ensuring that the total weight is always 100%.

[0037] The pathological risk coefficient and the surgical procedure influence coefficient are both read from the coefficient table according to the correspondence between "patient pathological type, surgical procedure, and monitoring indicator category". The subscript i in the formula indicates that the coefficient corresponds to the i-th type of monitoring indicator, rather than copying the same patient-level coefficient indiscriminately to all indicator categories. The coefficient table and the indicator category adaptation coefficient are saved with the same version number to reproduce the calculation process of the initial weights.

[0038] After the initial weight assignment is completed, the weight allocation of the monitoring indicators is simultaneously calibrated. The calibration process uses a correction formula based on the characteristics of real components. ,in For the calibrated first Real-time monitoring of the weight values ​​of the monitoring indicators The correction factor for the proportion of solid components is -0.1 when the volume proportion of solid components is <22.9%, 0 when the volume proportion of solid components is ≤22.9% and ≤50%, and 0.2 when the volume proportion of solid components is >50%. The data comes from the research conclusions of AI prediction of postoperative recurrence risk of stage IA non-small cell lung cancer. The initial weights were calibrated based on the characteristics of the proportion of solid components in the patient's nodules. This ensured that the initial weighting system matched the patient's pathological type (e.g., invasive adenocarcinoma, carcinoma in situ) and surgical method (wedge resection, lobectomy, etc.) and aligned with the risk stratification logic of the core recurrence prediction parameter of the proportion of solid components. The coefficients were 1.0 for imaging features, 0.8 for tumor marker features, 0.6 for clinical symptoms features, 0.5 for pulmonary function features, and 0.9 for pathological and surgical features. The coefficients were positively correlated with the correlation strength between the indicator and the risk of recurrence of the solid component, and the correlation strength was determined based on the correlation analysis results of previous clinical studies.

[0039] The above correction formula yields the corrected intermediate values ​​for each monitoring indicator category. The server completes all [tasks / processes]. After calculating the intermediate values ​​for each category, the intermediate value for each category is divided by the sum of the intermediate values ​​for all categories to obtain the final real-time monitoring weight. This normalization is performed only when the sum of the intermediate values ​​is greater than 0. If the sum is 0 due to an input anomaly, the weights before calibration are retained. After this processing, the real-time monitoring weights for each category are in the range of 0 to 1, and the sum of the real-time monitoring weights for all categories is 1.

[0040] Through the aforementioned indicator library and initial weight configuration process, preoperative baseline, intraoperative pathology, and first postoperative follow-up data are converted into the initial monitoring status of the same patient. Each monitoring indicator category has a corresponding real-time monitoring weight before entering continuous follow-up processing. Quantitative observation dimensions from different sources are organized by category, preserving the original meaning of various medical data while providing a unified data structure for subsequent time-ordered feature vector generation, indicator library expansion, and monitoring task selection.

[0041] Example 2: Based on Example 1, this example explains the process of identifying disease turning point events, generating disease stage weights, and determining the attention weights for each follow-up visit.

[0042] like Figure 3 As shown, the system determines the first The changes in nodule size, percentage of solid component, tumor markers, and clinical symptoms relative to baseline values ​​at the second follow-up were used to identify recurrence risk characteristics that met the corresponding change criteria as candidate turning points. Based on the type of candidate turning points, disease turning point event identifiers were generated. Subsequently, the changes were analyzed at the second follow-up. The time interval between the first and second follow-ups was normalized, and the normalized result was combined with the correction value corresponding to the disease turning point event marker to obtain the result. Weighting of disease stages in each follow-up visit.

[0043] The baseline values ​​are determined according to the feature type. Imaging features can be the corresponding measurements from preoperative baseline examination or the first postoperative follow-up examination; tumor markers can be the preoperative test values, the first stable postoperative test values, or the upper limit of normal values ​​given in the test report; and clinical symptoms can be the symptom level recorded at the first postoperative follow-up. Continuous comparisons are made using the same baseline source for the same observation dimension to avoid inconsistent changes at different follow-up time points due to changes in the reference standard.

[0044] Candidate turning points are used to represent recurrence risk characteristics that show a directional change relative to baseline or previous follow-up. In this embodiment, any of the following conditions can be marked as candidate turning points: an increase of ≥2 mm in the maximum diameter of the postoperative imaging abnormality from baseline, an increase of at least 10% in the proportion of solid components from baseline, a tumor marker exceeding 1.5 times the upper limit of normal, or the presence of hemoptysis or persistent chest pain. The types of candidate turning points include at least imaging change types, serological change types, and symptom change types; when multiple candidate turning points occur in the same follow-up, a disease turning point event identifier containing multiple types can be formed.

[0045] The term "postoperative imaging abnormality" in this instruction manual refers to residual lesions, suspected recurrent lesions, or newly developed abnormalities that can be continuously located during postoperative follow-up. It does not involve direct, continuous size comparisons between resected preoperative nodules and lesions that are no longer present postoperatively. For patients with traceable abnormalities, the imaging measurement value at which the abnormality can be stably located for the first time postoperatively is used as the imaging baseline for subsequent size and solid component percentage. For cases where no abnormality is seen on the first postoperative follow-up but a new abnormality subsequently appears, the state from absence to presence of the abnormality is recorded as the imaging change type, and a continuous measurement baseline for the abnormality is established from that follow-up visit.

[0046] The time interval between each follow-up visit relative to the previous follow-up baseline reflects the disease progression span between two effective follow-ups. The system uses the maximum time interval in the historical follow-up sequence as the normalization benchmark, converting the current time interval into a time interval value within the range of 0 to 1; the correction value corresponding to the disease turning point event identifier is used to indicate whether a change event requiring increased attention in feature aggregation has occurred in the current follow-up. The disease progression stage weight is obtained by combining the time interval value and the correction value, thus the same weight simultaneously carries the follow-up span and the disease progression status.

[0047] The maximum time interval used for normalization is selected from the positive adjacent intervals already formed for the current patient. The first postoperative follow-up uses the surgery date or discharge date as the previous time reference; duplicate records formed on the same day are first merged into one valid follow-up, therefore, 0 is not used as the maximum time interval. When only one valid postoperative follow-up is formed, the interval between that follow-up and the aforementioned previous time reference is used as the normalization reference. The disease stage weights are used to scale subsequent attention matching scores and are not normalized probability weights between each follow-up; their probabilistic processing is completed by subsequent Softmax operations.

[0048] like Figure 4 As shown, when generating each follow-up feature vector, the system extracts the observation values ​​of each quantifiable observation dimension from the follow-up data, performs standardized encoding on the observation values, and generates corresponding code-level aggregated features. The code-level aggregated features of each follow-up are arranged in chronological order to form a follow-up feature matrix, and the row of feature data corresponding to each follow-up in the follow-up feature matrix is ​​determined as the follow-up feature vector for that follow-up. The interval and direction of the standardized encoding remain consistent for the same observation dimension, ensuring that different indicators, such as increased or decreased values ​​representing increased risk, can be transformed into a unified feature representation space.

[0049] For continuous observation dimensions, the system transforms observation values ​​to the 0-1 interval according to the lower and upper bounds recorded in the data dictionary, truncating values ​​exceeding the upper or lower bounds as 1 or 0 respectively. For observation dimensions where increased values ​​indicate decreased risk, reverse encoding is used. For categorical observation dimensions, values ​​are mapped to the 0-1 interval according to a pre-determined rank order. Under the same data dictionary version, each follow-up uses the same boundaries, units, and encoding direction, ensuring that each column of the follow-up feature matrix has a fixed and traceable meaning.

[0050] The system obtains first-mapped features and second-mapped features by passing each follow-up feature vector through a first feature mapping matrix and a second feature mapping matrix, respectively. An attention matching score is generated based on the matching relationship between the first-mapped feature of any follow-up and the second-mapped features of all follow-ups. The first and second feature mapping matrices are pre-trained from historical follow-up data and are used to map code-level aggregated features of different categories to a feature space of the same dimension.

[0051] Historical follow-up samples were generated using the same data dictionary and time-ordering method as the patients to be monitored, resulting in follow-up feature matrices. High-risk and low-risk samples were then categorized based on completed follow-up outcomes. During training, the first and second feature mapping matrices were initially established with preset values, and then iteratively updated based on the error between the model's output recurrence risk assessment value and the sample outcome label. Updates stopped and the matrix version was saved when the error on the validation samples no longer decreased. In actual monitoring, the validated matrix version was consistently used, and the matrix parameters were not changed during the online assessment of individual patients.

[0052] The term "patient-level" in this manual refers to a hierarchy where a single effective postoperative follow-up visit is considered an independent data processing unit. Each effective postoperative follow-up visit corresponds to a single follow-up record grouped within the same follow-up window, and generates a corresponding follow-up feature vector. Therefore, "patient-level attention weight" refers to the attention weight assigned to the follow-up feature vector corresponding to each effective postoperative follow-up visit within a follow-up sequence comprised of multiple effective follow-up visits. This weight is used to characterize the contribution of a single follow-up visit relative to the target follow-up visit in generating postoperative global health representation data.

[0053] Here, "visit-level" and "follow-up-level" (which refers to each follow-up visit as the processing object) represent the same data processing granularity, both using a single effective postoperative follow-up visit as the basic unit. This differs from item-level or feature-level processing within a single follow-up visit, which uses quantifiable observable dimensions or corresponding code-level aggregated features such as imaging indicators, tumor markers, clinical symptoms, and pulmonary function indicators within that follow-up visit as the processing object. In other words, the "visit-level attention weight" mentioned in this article refers to the attention weight corresponding to each follow-up visit, not the attention weight determined separately for each examination item within a single follow-up visit.

[0054] For target follow-ups, the system scales the attention matching score using the disease stage weights corresponding to the target follow-up, and then normalizes all scaled results to obtain the visit-level attention weights for each follow-up. Thus, the attention matching score is used to represent the content association between each follow-up feature, and the disease stage weights are used to represent the follow-up span and turning point status. The two types of information are coupled before normalization to avoid the disease turning point information being weakened in subsequent aggregation as an independent additional feature.

[0055] To clarify the calculation object of attention matching score, a follow-up feature matrix is ​​defined. Include lines and Column, number OK Represents the follow-up feature vector at the current evaluation point; the first feature mapping matrix. Second feature mapping matrix All shapes are Regarding the first The second follow-up was conducted by... and The product of the two vectors yields the query vector. Then by the first Follow-up vector and The product of the two is used to obtain the key vector. ;Will and dot product divided by The square root of the first digit is obtained. The second follow-up was relative to the first The baseline matching score at the second follow-up, and the score at the third follow-up. The weight of the disease stage corresponding to each follow-up visit is used to scale the basic matching score.

[0056] right From 1 to All scaling scores were Softmax normalized to obtain the consultation-level attention weights for each follow-up visit. And all of the follow-up corresponding to the same goal The sum is 1; the postoperative global health characteristics at the current assessment point. By each Corresponding The product is summed. The scaling dot product attention formula and aggregation formula in the manual are matrix-simplified versions of the above successive calculation process.

[0057] The dynamic update module randomly selects a monitoring indicator from the monitoring indicator library according to its real-time monitoring weight as the core monitoring target for the current period. It then pushes the corresponding monitoring task (such as completing a lung function test) during the patient's follow-up period. This monitoring task information is used to notify the patient and follow-up medical staff to complete the corresponding examinations and data collection according to the core monitoring target for the current period. The real-time monitoring weight of each monitoring indicator is updated based on a dynamic adjustment process, which includes: The system determines the core monitoring objectives for the current period based on real-time monitoring weights. This is used to prioritize adjustable monitoring tasks without canceling basic follow-up examinations that patients are required to complete according to the established follow-up plan. For imaging, laboratory, or pulmonary function tests that are clearly defined as mandatory by the follow-up plan, the system retains their task status. The weighted core monitoring objectives are used to determine the priority of additional data collection items, reminder order, or the next adjustable task.

[0058] When a new monitoring dimension is added to the indicator library, an initial real-time monitoring weight value is assigned, and the real-time monitoring weights of other monitoring indicators in the indicator library are adjusted simultaneously. After the current core monitoring objectives are completed, a risk assessment of the postoperative recovery status of the patients using the current monitoring plan is conducted according to a preset evaluation system. The preset evaluation system includes: Established based on the monitoring indicator database The system constructs a total recurrence risk score for the current core monitoring target across several evaluation dimensions. For each follow-up, based on whether the follow-up is a turning point in the disease (such as the first follow-up showing nodule enlargement) and the follow-up time interval, the system obtains the disease stage weight for that follow-up. Based on the disease stage weight for that follow-up and a feature matrix composed of code-level aggregated features from all follow-ups, the system generates corresponding consultation-level attention weights for each follow-up. Based on the consultation-level attention weights for each follow-up, the system performs weighted aggregation of the code-level aggregated features to generate the current patient's postoperative global health representation. This representation, combined with the total score, is used as a preset indicator to assess the postoperative recovery status of the patient using the current monitoring plan. The risk assessment results are then sent to the adjustment module.

[0059] The core operating logic of the dynamic update module begins with a weighted random selection from the established database of postoperative recurrence risk monitoring indicators for pulmonary nodules. This selection is based entirely on the probability distribution of the weighted values, with no subjective human intervention. Once the core monitoring target is determined, the module automatically matches the corresponding monitoring tasks and pushes them to the patient and follow-up healthcare personnel. For example, if the core monitoring target is a pulmonary function characteristic, a pulmonary function test task is pushed, requiring the patient to complete postoperative pulmonary function testing and upload quantitative data such as FEV1, DLCO, and arterial oxygen partial pressure. If the core monitoring target is an imaging characteristic, a postoperative thin-slice CT scan task is pushed, requiring the patient to complete a chest CT scan with a slice thickness ≤1mm and upload the image data. All tasks clearly define the corresponding monitoring dimensions and data collection standards to ensure the consistency and comparability of subsequent evaluation data. The real-time monitoring weights for each type of monitoring indicator are not fixed, but are continuously updated based on a dynamic adjustment process throughout the entire workflow. This adjustment process first covers the dynamic expansion of the indicator library: when a new monitoring dimension is added to the indicator library, the module automatically assigns an initial real-time monitoring weight value to the monitoring indicators of the category to which the new dimension belongs. The initial weight assignment uses a calculation method based on indicator correlation, with the formula as follows: ,in The newly added dimension belongs to the first The initial real-time monitoring weight values ​​after the update of the monitoring indicators are as follows: For the new dimension and the first The correlation coefficient of the original observation dimensions under the monitoring indicators (calculated by cosine similarity, with a value range of 0 to 1. The higher the correlation, the closer the value is to 1. It is determined based on the correlation analysis results of radiomics features or serological indicators). To adjust the real-time monitoring weight values ​​of the j-th type of monitoring indicators, This represents the total number of categories of monitoring indicators before adjustment. To adjust the sum of the correlation coefficients between all observation dimensions and recurrence risk under the k-th monitoring indicator, when adding a new dimension, the real-time monitoring weights of all other monitoring indicators in the indicator library are adjusted simultaneously through normalization calculation of correlation weight, ensuring that the total weight of the entire library is always 100%, and that the weight allocation of the new dimension is completely matched with its clinical value.

[0060] When a new monitoring dimension is added, the system first calculates the target weight of the category to which the new dimension belongs based on the correlation coefficient. Then, it allocates the remaining weights according to the proportion of the weights of the other categories before adjustment to the total weights of the other categories. If the new dimension is added to an existing category, the weight of that category is updated first, and then the weights of the other categories are scaled up by the same proportion. After all categories are updated, normalization is performed again, and a 0-to-1 boundary check is performed on each weight. When the sum of the correlation coefficients is 0, the weight reallocation is not performed, and the new dimension is simply saved as a state to be calibrated to avoid a denominator of 0.

[0061] Once the monitoring tasks corresponding to the current core monitoring objectives are completed, the module will initiate a postoperative recovery status risk assessment process based on a preset evaluation system. This preset evaluation system is first established based on the architecture of the monitoring indicator library. The evaluation dimensions cover all quantifiable observable dimensions under five major categories: imaging features, tumor marker features, clinical symptom features, pulmonary function features, and pathological and surgical features. Based on this, a total recurrence risk score is constructed for the current core monitoring target. The score is calculated using a multi-dimensional weighted summation formula. ,in This represents the total recurrence risk score corresponding to the core monitoring target in the current period. For the first The quantitative observation values ​​of each evaluation dimension (e.g., the volume percentage of the real component is 0 for below 22.9%, 1 for 22.9% to 50%, and 2 for above 50%. All values ​​are determined based on quantitative thresholds, leaving no room for subjective interpretation). For the first The fixed weight coefficients of each evaluation dimension (assigned based on the correlation analysis results between the dimension and the postoperative recurrence risk; the higher the correlation, the larger the weight coefficient; the data comes from the conclusions of a multifactor study on the postoperative recurrence risk of stage IA non-small cell lung cancer) are used to generate the total recurrence risk score of the core monitoring target for the current period by weighted summation of the quantitative values ​​of all evaluation dimensions.

[0062] In risk stratification calculation, the fixed weight coefficients of each evaluation dimension First, normalize within the evaluation dimensions involved in this calculation. For evaluation dimensions using a 0, 1, or 2 scale, divide the scale value by 2 to obtain a quantified value in the range of 0 to 1. Other evaluation dimensions were also converted to the range of 0 to 1 according to the corresponding data dictionary. This resulted in... The range is between 0 and 1. If a certain evaluation dimension lacks valid observations, that dimension is not included in the summation, and the summation is applied to the remaining participating dimensions. Renormalize.

[0063] For each follow-up visit, the disease stage weight is calculated based on whether the visit is a turning point event and the follow-up time interval. A turning point event is defined as a medical event observed during postoperative follow-up, such as nodule enlargement (maximum diameter increased by ≥2mm from baseline), increased proportion of solid components (increased by ≥10% from baseline), elevated tumor markers (exceeding the upper limit of normal by 1.5 times), or new clinical symptoms (e.g., hemoptysis, persistent chest pain). The formula for calculating the disease stage weight is as follows: ,in For the first Weighting of disease course stages in each follow-up visit For the first The time interval between the current follow-up and the previous follow-up (unit: days). The maximum time interval for all historical follow-ups. The correction factor for disease turning point events is 0.5 when a disease turning point event occurs and 0 when no event occurs. For the first The disease turning point event markers for each follow-up visit (1 for occurrence, 0 for absence) quantify the weight of each follow-up visit in the patient's overall disease course based on the length of the follow-up interval and whether a disease turning point event occurs. The longer the interval and the higher the weight when a disease turning point event occurs, which aligns with the dynamic adjustment of postoperative recurrence risk over time and with changes in the patient's condition.

[0064] After calculating the weights for each disease stage, a feature matrix is ​​constructed by extracting the code-level aggregated features from all follow-ups. Code-level aggregated features refer to the standardized coded values ​​of all quantifiable observable dimensions in each follow-up. For example, the maximum nodule diameter, the proportion of solid component, and the CEA value are standardized to coded values ​​between 0 and 1 according to preset thresholds. All features are arranged in chronological order to form the feature matrix. This represents the total number of follow-up visits. This represents the total number of code-level aggregated features for each follow-up. Based on the disease stage weights for each follow-up and the feature matrix, the module generates corresponding patient-level attention weights for each follow-up. The attention weights are calculated using a scaled dot product attention formula based on the disease stage weights. ,in For the first The corresponding consultation-level attention weights for each follow-up visit. For the first Code-level aggregated feature vectors from the next follow-up , It is a learnable feature mapping matrix (pre-trained from historical patient data with clear follow-up outcomes, and the corresponding version is fixed before actual monitoring). The mapped feature dimensions are combined with the weights of disease stages and the correlation between features of all follow-ups to calculate the attention weight of each follow-up in the patient's overall health representation. Follow-ups corresponding to disease turning events and follow-ups with longer time intervals will receive higher attention weights.

[0065] Based on the patient-level attention weights corresponding to each follow-up visit, the code-level aggregated features are weighted and aggregated to generate the current patient's postoperative global health representation. The aggregation formula is as follows: ,in The postoperative global health representation vector is generated by weighting and fusing the features of each follow-up visit through attention weights, highlighting the feature contribution of high-weight follow-up visits and weakening the feature interference of low-weight follow-up visits, thus generating a health representation that can comprehensively reflect the patient's overall postoperative recovery status.

[0066] Using the postoperative global health characteristics combined with the aforementioned total recurrence risk score as preset indicators, a risk assessment was conducted on the postoperative recovery status of patients using the current monitoring protocol. The risk assessment employed a risk stratification formula based on characteristic similarity. ,in This is a recurrence risk assessment value (ranging from 0 to 1, with values ​​closer to 1 indicating higher risk). The cosine similarity between the patient's postoperative global health profile and the average profile of a historically high-risk recurrence patient group is used. The cosine similarity between the patient's postoperative global health profile and the average profile of a historically low-risk recurrence patient group is used. The total recurrence risk score for the current core monitoring target is used to calculate the patient's current risk assessment value by combining the similarity between the patient's health characteristics and the characteristics of different risk groups with the scores of the current core monitoring target.

[0067] In the risk stratification formula and This is the symbolic representation of the similarity function, not a sine function; in this embodiment, it is uniformly denoted as... and The cosine similarity is used for calculation. Since both follow-up features and population average representations use non-negative 0-1 encoding, the cosine similarity is within the range of 0-1; when any representation is a zero vector, the corresponding similarity is recorded as 0. If the sum of the similarities of the two classes is 0, then the normalized similarity is used. This value serves as the risk assessment value for this recurrence; in other cases, it is calculated according to the aforementioned risk stratification formula, with the value limited to the range of 0 to 1.

[0068] This embodiment converts the time intervals of each follow-up relative to the previous follow-up baseline and disease turning events into disease stage weights, and uses these weights to correct the attention matching score. This ensures that the aggregate contribution of each follow-up depends not only on the degree of feature similarity but also on the span and changes of that follow-up within the patient's disease course. For follow-up sequences with uneven acquisition intervals, disease changes formed by longer intervals and follow-ups with turning events can be correspondingly expressed in a unified attention calculation process, thereby maintaining the correlation between postoperative global health representation and continuous disease course information.

[0069] Example 3: Based on Example 2, this example describes the process of generating postoperative global health characteristics, determining recurrence risk monitoring results, and updating monitoring strategies.

[0070] like Figure 5 As shown, the system weights the patient-level attention weights corresponding to each follow-up visit with the corresponding follow-up feature vectors, and aggregates the resulting weighted features to generate postoperative global health representation data. The postoperative global health representation data and the follow-up feature vectors have corresponding feature dimensions. Each dimension is formed by weighting the encoded values ​​of the same observation dimension from each follow-up visit according to the patient-level attention weights corresponding to each visit, representing the patient's overall recovery status up to the current follow-up time.

[0071] The system determines the similarity between postoperative global health representation data and the average representation of historical high-risk and low-risk patient groups. The average representations of high-risk and low-risk patient groups are generated from historical patient data with clear follow-up outcomes using the same feature encoding and aggregation methods to ensure that the representations of the patients to be monitored and the historical group representations are in the same feature space.

[0072] Historically high-risk and historically low-risk patient groups were established using samples with clear follow-up outcomes and within the same observation window, and processed according to the same data dictionary, mapping matrix version, and characterization generation process as the monitored patients. The average characterization of each group was obtained by averaging the postoperative global health characterization of each sample in that group across the corresponding dimensions; when the data dictionary or mapping matrix version was updated, the average characterization of both groups was recalculated synchronously to avoid directly calculating similarity between different feature spaces.

[0073] The total relapse risk score for the current period is obtained by weighting the quantitative observation values ​​of multiple evaluation dimensions established in the monitoring indicator library with their fixed weight coefficients. The system combines the similarity of the two types of representations with the total relapse risk score for the current period to obtain a relapse risk assessment value ranging from 0 to 1, and outputs the relapse risk monitoring results based on this relapse risk assessment value. The monitoring results may include the risk assessment value, the corresponding risk interval, the follow-up time points with high contributions, and the core monitoring targets for the current period, so as to facilitate display on the follow-up terminal and updates for subsequent monitoring tasks.

[0074] The relapse risk assessment value and its risk range are used to indicate the monitoring priority of follow-up data and provide data basis for adjusting the weight of subsequent monitoring indicators, task order, and monitoring interval. This output is not used alone to confirm relapse, nor does it directly generate treatment instructions; when the output meets the conditions for joint follow-up, the follow-up terminal displays the corresponding historical data and the source of abnormalities to relevant medical staff, who then make further judgments based on complete clinical information.

[0075] like Figure 6 As shown, after outputting the recurrence risk monitoring results, the system determines the core monitoring target for the current period based on the real-time monitoring weight of each monitoring indicator category, and pushes the monitoring task corresponding to the core monitoring target. The system updates the real-time monitoring weight and monitoring interval of the monitoring indicator category to which the core monitoring target belongs based on the recurrence risk monitoring results, while adjusting the real-time monitoring weights of the remaining monitoring indicator categories to ensure that the sum of the real-time monitoring weights of all monitoring indicator categories remains at 1.

[0076] When the relapse risk assessment value is within the safe range and the cumulative number of selections of the current core monitoring target reaches the corresponding threshold, the system reduces the real-time monitoring weight of the monitoring indicator category to which the current core monitoring target belongs, extends the corresponding monitoring interval, and clears the cumulative number of selections. When the relapse risk assessment value exceeds the safe range, the system increases the real-time monitoring weight of the corresponding monitoring indicator category, shortens the corresponding monitoring interval, and sends joint follow-up information containing the relapse risk assessment value, historical monitoring data, and recent re-examination data to the multidisciplinary follow-up terminal.

[0077] Adjustment Module: Based on the risk assessment results, the real-time monitoring weights of the corresponding monitoring indicators are adjusted, and the real-time monitoring weights of the remaining monitoring indicators in the monitoring indicator library are adjusted simultaneously. The real-time monitoring weights of all monitoring indicators range from [0, 1]. The logic for adjusting the real-time monitoring weights based on the risk assessment results is as follows: When the risk assessment results meet the preset safety threshold and the cumulative number of times such monitoring indicators are selected as core monitoring targets reaches the preset value, the monitoring interval of such monitoring indicators in subsequent follow-ups is increased, and the cumulative number of times such monitoring indicators are selected as core monitoring targets is reset to zero; when the risk assessment results do not meet the preset safety threshold, the monitoring interval of such monitoring indicators is reduced, and a multidisciplinary joint follow-up mechanism is triggered.

[0078] The adjustment module iteratively updates the real-time monitoring weights of various monitoring indicators in the monitoring indicator library, and simultaneously adjusts the real-time monitoring weights of the remaining monitoring indicators in the entire library. The real-time monitoring weights of all monitoring indicators are strictly limited to the range of [0, 1] to ensure that the sum of the weights is 1, which conforms to the calculation logic of probability normalization. When the risk assessment result meets the preset safety threshold, the module first determines that the postoperative recovery status of the patient corresponding to this type of monitoring indicator is in the low-risk range. The preset safety threshold is defined as the recurrence risk assessment value. ( The postoperative recovery status risk assessment value output by the dynamic update module ranges from 0 to 1 (the closer to 1, the higher the risk), and the cumulative number of times this type of monitoring indicator has been selected as a core monitoring target reaches a preset value. ( The preset threshold for the cumulative number of selections is set according to the pathology type: For carcinoma in situ, 3 times were taken; for minimally invasive adenocarcinoma, 4 times were taken; and for invasive adenocarcinoma, 5 times were taken (data derived from the decreasing trend of postoperative recurrence risk over time for different pathological types). At this point, the module will perform weight reduction and monitoring interval adjustment operations. The weight adjustment uses a decay formula based on the safety status. ,in For the adjusted number Real-time monitoring weight values ​​for class monitoring indicators To adjust the previous number Real-time monitoring weight values ​​for class monitoring indicators The weight decay coefficient for the safe state is set at 0.2, which is determined based on the decrease in the contribution of this type of indicator to the recurrence risk under low-risk conditions. The cumulative number of times this type of monitoring indicator has been selected as a core monitoring target. A preset cumulative selection threshold is set for this type of monitoring indicator. Based on the ratio of the cumulative selection count to the threshold, the weight ratio of this type of monitoring indicator is reduced.

[0079] The weights of all other monitoring indicators are adjusted simultaneously using the following formula: ,in For the adjusted number Real-time monitoring weight values ​​for class (j≠i) monitoring indicators. To adjust the first The real-time monitoring weight value of the category monitoring indicator, where N is the total number of categories of monitoring indicators, ensuring that the total weight of the entire database is still 1.

[0080] Under both safe and risky conditions, the weights calculated by the formula are used as intermediate update values. (In the attenuation calculation...) and The ratio is set to a value no greater than 1, and the target category weight after risk gain does not exceed 1. Equal allocation is only used when it does not make the weight of any other category less than 0 or greater than 1; if it exceeds the limit, the weight is released or the deduction is borne according to the proportion of the weight of the other categories before adjustment, and then all categories are normalized again so that each weight is kept in the range of 0 to 1 and the sum is 1.

[0081] After the weight adjustment is completed, the monitoring interval for this type of monitoring indicator in subsequent follow-ups will be increased. The formula for adjusting the monitoring interval is: ,in The monitoring interval (in days) for the adjusted i-th type of monitoring indicator. The monitoring interval for the i-th type of monitoring indicator was adjusted (the default interval is 90 days after the first follow-up examination one month after surgery). The monitoring interval extension coefficient (set to 0.5, determined based on clinical guidelines recommending that the frequency of re-examination can be appropriately reduced under low-risk conditions) is used to extend the monitoring interval of this type of indicator according to the ratio of the cumulative number of selections to the threshold, thereby reducing unnecessary examination burden. At the same time, the cumulative number of times this type of monitoring indicator is selected as a core monitoring target is reset to zero, and the counting benchmark is reset for the next round of weight adjustment.

[0082] When the risk assessment result does not meet the preset safety threshold, i.e., the recurrence risk assessment value The module determines that the postoperative recovery status of patients corresponding to this type of monitoring indicator is in the medium-to-high risk range. At this time, it will perform weight adjustment and monitoring interval shortening operations. The weight adjustment adopts a gain formula based on risk status: ,in To adjust the real-time monitoring weight value of the i-th type of monitoring indicator, To adjust the real-time monitoring weight values ​​of the i-th type of monitoring indicator, The risk status weight gain coefficient is set to 0.3, which is determined based on the increase in the contribution of this type of indicator to the recurrence risk under medium-to-high risk conditions. Based on the current recurrence risk assessment value, the weight of this type of monitoring indicator is increased according to the magnitude of the risk assessment value. At the same time, the weights of all other monitoring indicators are adjusted. The adjustment formula is the same as the weight update formula of other indicators under the safe state, ensuring that the total weight of the entire database is always 1.

[0083] After the weight adjustment is completed, the monitoring interval for this type of monitoring indicator will be reduced. The formula for adjusting the monitoring interval is: ,in For the adjusted number Monitoring intervals (unit: days) for various monitoring indicators. To adjust the first Monitoring intervals for various monitoring indicators The monitoring interval shortening coefficient (set to 0.4, determined based on clinical guidelines recommending shorter follow-up intervals under medium- to high-risk conditions) is used to shorten the monitoring interval for this type of indicator according to the magnitude of the risk assessment value, thereby increasing the monitoring frequency and timely detection of changes in the condition.

[0084] The calculation result of the monitoring interval formula is used as the candidate interval and rounded down to the nearest day. The system saves the shortest and longest allowed intervals for each monitoring indicator category in the task configuration; the shortest interval is used when the candidate interval is lower than the shortest interval, and the longest interval is used when the candidate interval is higher than the longest interval, to avoid unexecutable interval values ​​after multiple iterations. The interval adjustment only changes the subsequent task suggestions and does not modify the time records of completed follow-ups.

[0085] The mechanism for automatically triggering multidisciplinary joint follow-up is based on risk threshold determination. when At that time, the system automatically sends a joint follow-up notification to multidisciplinary teams such as thoracic surgery, oncology, radiology, and pathology. The notification includes the current risk assessment value, historical monitoring data of this type of monitoring indicator, and recent follow-up imaging and serological results. The joint follow-up notification is only used to indicate the priority of follow-up and summarize relevant data. The multidisciplinary team completes the assessment and decides on the subsequent monitoring plan based on the complete medical records, imaging, test results, and clinical procedures. The recurrence risk assessment value output by the system does not replace the diagnosis or treatment decision made by the physician.

[0086] By comparing the postoperative global health profile with the average profiles of historical high- and low-risk groups, and combining this with the current recurrence risk score, a current risk assessment value can be generated while retaining the contribution of each follow-up timeline. This assessment value is further used to adjust the real-time monitoring weights and monitoring intervals of the monitoring indicator categories, ensuring a continuous connection between risk output and the next stage of data collection. Simultaneously, the weights of other indicators are normalized to avoid drift in the total weight of the entire database caused by changes in the weight of a single indicator.

[0087] Example 4: This example provides a complete case of postoperative follow-up management of pulmonary nodules to illustrate the operation of the above method in continuous follow-up data.

[0088] Establish a database of indicators for monitoring the risk of postoperative recurrence of pulmonary nodules. The database includes... Classified monitoring indicators, each class of monitoring indicators is further divided into The specific observation dimensions and the indicator database cover a total of [number] aspects. Each monitoring indicator has a quantifiable observation dimension and a corresponding real-time monitoring weight. When a patient is discharged after lung nodule surgery, an initial real-time monitoring weight value is assigned to each monitoring indicator based on preoperative baseline examination, intraoperative pathological results, and the first postoperative follow-up data. The weight allocation of the monitoring indicators is simultaneously calibrated to ensure that the initial weight system matches the patient's pathological type (such as invasive adenocarcinoma, carcinoma in situ, etc.) and surgical method (wedge resection, lobectomy, etc.). The monitoring indicators cover imaging characteristics (nodule size, density, edge morphology, proportion of solid component, etc.) and clinical symptom characteristics (presence and severity of symptoms such as cough, chest pain, and hemoptysis). Each indicator is further subdivided into multiple dynamically trackable observation dimensions.

[0089] A monitoring indicator is randomly selected from the monitoring indicator database according to its real-time monitoring weight as the core monitoring target for the current period. The corresponding monitoring task (such as completing a lung function test) is then pushed to the patient during the follow-up period. This monitoring task information is used to notify the patient and follow-up medical staff to complete the corresponding examinations and data collection according to the core monitoring target for the current period. The real-time monitoring weight of each monitoring indicator is updated based on a dynamic adjustment process, which includes: When a new monitoring dimension is added to the indicator library, an initial real-time monitoring weight value is assigned, and the real-time monitoring weights of other monitoring indicators in the indicator library are adjusted simultaneously. After the current core monitoring objectives are completed, a risk assessment of the postoperative recovery status of the patients using the current monitoring plan is conducted according to a preset evaluation system. The preset evaluation system includes: Established based on the monitoring indicator database The system employs several evaluation dimensions to construct a total recurrence risk score for the current core monitoring target. For each follow-up visit, based on whether the visit is a turning point in the disease (e.g., the first follow-up examination reveals nodule enlargement) and the follow-up time interval, the system obtains the disease stage weight for that visit. Based on the disease stage weight for that visit and a feature matrix composed of code-level aggregated features from all follow-ups, the system generates corresponding consultation-level attention weights for each visit. Based on the consultation-level attention weights for each follow-up visit, the system performs weighted aggregation of the code-level aggregated features to generate the current patient's postoperative global health profile. This profile, combined with the total score, is used as a preset indicator to assess the risk of postoperative recovery status under the current monitoring protocol.

[0090] The real-time monitoring weights of the corresponding monitoring indicators are adjusted based on the risk assessment results, and the real-time monitoring weights of the remaining monitoring indicators in the monitoring indicator library are adjusted simultaneously. The real-time monitoring weights of all monitoring indicators range from [0, 1]. The logic for adjusting the real-time monitoring weights based on the risk assessment results is as follows: When the risk assessment results meet the preset safety threshold and the cumulative number of times such monitoring indicators are selected as core monitoring targets reaches the preset value, the monitoring interval of such monitoring indicators in subsequent follow-ups is increased, and the cumulative number of times such monitoring indicators are selected as core monitoring targets is reset to zero; when the risk assessment results do not meet the preset safety threshold, the monitoring interval of such monitoring indicators is reduced, and a multidisciplinary joint follow-up mechanism is triggered.

[0091] This embodiment uses a patient who has undergone pulmonary nodule resection surgery and entered the continuous follow-up phase as an example. Upon discharge, the system reads the preoperative thin-section CT images, preoperative CEA and CYFRA21-1 test values, baseline lung function, intraoperative pathological type, resection margin status, lymph node status, and surgical procedure. After the first postoperative follow-up examination, the system writes the images, tumor markers, and lung function data into the monitoring indicator database. The system configures the initial real-time monitoring weights for each monitoring indicator category according to pathological risk parameters and surgical procedure influence parameters, and then calibrates these initial real-time monitoring weights based on the proportion of solid components.

[0092] For example, patients completed four effective follow-up visits on postoperative days 30, 120, 240, and 330, with the time intervals between each follow-up visit relative to the previous follow-up baseline being 30 days, 90 days, 120 days, and 90 days, respectively. The system standardized and encoded the imaging measurements, tumor marker test values, symptom grades, and lung function values ​​from the four follow-up visits using the same data dictionary, forming four follow-up feature vectors, which were then arranged into a follow-up feature matrix according to the follow-up time.

[0093] During the 240-day follow-up, if a tumor marker is detected to exceed 1.5 times the upper limit of normal, and the proportion of solid components increases by at least 10% compared to the corresponding baseline, the system will record both the serological change type and the imaging change type as disease turning point events for that follow-up. Since the time interval corresponding to this follow-up is the maximum among the four follow-ups, its normalized time interval is relatively high. Combined with the correction value corresponding to the disease turning point event, this follow-up receives a higher weight for disease stage than follow-ups without a turning point event.

[0094] Subsequently, the system inputs the four follow-up feature vectors into the first and second feature mapping matrices, respectively, calculates the attention matching score between any target follow-up and all follow-ups, and scales and normalizes the score using the weights corresponding to the disease stage. For example, the normalized attention weights for the four follow-ups at the consultation level can be 0.12, 0.18, 0.50, and 0.20, respectively, with the 240-day follow-up having a higher aggregation contribution due to both a longer disease duration and a disease turning point. The above values ​​are only used to illustrate the calculation order; the actual weights are determined jointly by the patient's follow-up features and the pre-trained mapping matrix.

[0095] The system weights and aggregates the four follow-up feature vectors using the patient-level attention weights corresponding to each follow-up visit, obtaining the patient's postoperative global health representation data up to day 330. It then calculates the cosine similarity between this representation and the average representation of historically high-risk and historically low-risk patient groups, and combines this with the total recurrence risk score to obtain a recurrence risk assessment value. If the recurrence risk assessment value is greater than 0.3, the system increases the real-time monitoring weight of monitoring indicator categories related to the current abnormal change, shortens the corresponding monitoring interval, and sends joint follow-up information to the follow-up terminals of thoracic surgery, oncology, radiology, and pathology departments.

[0096] If the recurrence risk assessment values ​​in subsequent follow-ups are all within the safe range, and the cumulative number of times a certain monitoring indicator category is selected as the core monitoring target reaches the threshold corresponding to the patient's pathological type, the system reduces the real-time monitoring weight of that category and extends the corresponding monitoring interval, while simultaneously resetting the cumulative selection count to zero. Through the above processing, the monitoring plan for the same patient can be updated sequentially based on continuous follow-up data without changing the historical feature vector and patient-level attention calculation records corresponding to completed follow-ups.

[0097] This embodiment fully presents the data flow from indicator library initialization, follow-up data acquisition, disease turning point identification, disease stage weight generation, attention score correction, postoperative global health characterization generation to monitoring task updates. Each processing step uses the data output from the previous step as input, enabling risk monitoring results to be traced back to the corresponding follow-up time point, recurrence risk characteristics, and monitoring indicator categories.

[0098] Example 5: This example provides a deployment and module collaboration method for a postoperative recurrence risk monitoring system for pulmonary nodules.

[0099] The system comprises a data acquisition module, a stage weight generation module, an attention correction module, and a risk output module. The data acquisition module acquires postoperative follow-up data for each patient, sorted by follow-up time. The stage weight generation module identifies disease turning points based on the time interval between each follow-up and the previous follow-up baseline, and generates disease stage weights. The attention correction module generates feature vectors for each follow-up, calculates attention matching scores, corrects these scores using disease stage weights, and determines the corresponding consultation-level attention weight for each follow-up. The risk output module aggregates the follow-up feature vectors, generates postoperative global health representation data, and outputs recurrence risk monitoring results.

[0100] The data acquisition module can be configured with separate interfaces for image data, laboratory data, medical record data, and terminal interaction. The image data interface receives quantitative results from thin-section CT scans, including the maximum diameter of nodules, diameter of solid components, volume percentage of solid components, average CT value of nodules, and boundary morphology. The laboratory data interface receives serological test results such as CEA, NSE, CYFRA21-1, and SCC. The medical record data interface receives records of pathological type, surgical procedure, surgical margins, lymph node status, lung function, and symptoms. The terminal interaction interface sends monitoring tasks and risk monitoring results to the patient and follow-up medical staff.

[0101] The stage weight generation module stores the baseline source, change conditions, and event type mapping relationships for each observation dimension. Upon receiving new follow-up data, this module first calculates the time interval between the current follow-up and the previous effective follow-up. Then, it generates disease turning point event identifiers based on changes in nodule size, proportion of solid components, tumor markers, and clinical symptoms. Finally, it combines the normalized time interval and event correction amount into disease stage weights and stores these weights in association with the current follow-up identifier.

[0102] The attention correction module stores the data dictionary used for standardized coding, as well as the first and second feature mapping matrices. This module converts quantifiable observation dimensions from each follow-up visit into code-level aggregated features according to the data dictionary, forming follow-up feature vectors and follow-up feature matrices. It then calculates the matching relationships between the mapped features, scales the attention matching scores by disease stage weights, and normalizes the scaling results. When updating the first and second feature mapping matrices, they can be retrained using historical relapse patient data and version-labeled, ensuring that the same version of the mapping matrix is ​​used for the same risk assessment.

[0103] The risk output module stores the average characteristics of historical high-risk and low-risk patient groups, and aggregates the follow-up feature vectors based on the patient attention weights corresponding to each follow-up visit. The risk output module writes postoperative global health characteristic data, characteristic similarity, current recurrence risk score, and recurrence risk assessment value into the same risk record, and sends this risk record to the monitoring task management unit. The monitoring task management unit updates the real-time monitoring weights, monitoring intervals, and joint follow-up information based on the risk assessment value and the cumulative number of selections.

[0104] The risk record simultaneously stores the patient identifier, target follow-up identifier, data dictionary version, mapping matrix version, population average representation version, disease stage weight, visit-level attention weight for each follow-up, and recurrence risk assessment value. The server locks the above versions after an assessment begins until the assessment ends; if any base version is updated, it is recalculated from the standardized coding steps to avoid using different versions of data processing parameters for the same risk record.

[0105] In one implementation, the aforementioned modules are executed by a processor in a server, running programs stored in memory. The monitoring indicator library, follow-up feature matrix, mapping matrix, and historical population average representation are stored in a data storage system connected to this server. In another implementation, the data acquisition module is located on an in-hospital business server, while the stage weight generation module, attention correction module, and risk output module are located on a data processing server. Each server exchanges corresponding data using patient identifiers and follow-up identifiers. These deployment methods do not alter the input-output relationships between the modules.

[0106] For example, in low-risk continuous follow-up scenarios, the risk output module obtains a recurrence risk assessment value of no more than 0.3, and the cumulative number of times imaging features are selected as the core monitoring target reaches the corresponding threshold. The monitoring task management unit reduces the real-time monitoring weight of imaging features and extends its monitoring interval, distributing the released weight to other monitoring indicator categories according to the weight ratio of each category before adjustment. In medium-to-high-risk scenarios, the monitoring task management unit increases the real-time monitoring weight of the category to which the abnormal indicator belongs and shortens the monitoring interval, while simultaneously sending joint follow-up information.

[0107] This embodiment clarifies the processing path of continuous follow-up data from access, joint representation of time and events, attention correction, global representation generation to monitoring task updates through a modular approach. Each module can be deployed on the same server or distributed across multiple servers. As long as the stage weights, follow-up feature vectors, the consultation-level attention weights corresponding to each follow-up, and postoperative global health representation data are passed according to the aforementioned relationships, the same recurrence risk monitoring process can be completed.

[0108] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0109] The embodiments disclosed above are used to help illustrate the present invention and do not represent a complete list of implementation details, nor do they limit the present invention to the specific embodiments described. Based on the technical content disclosed in this specification, those skilled in the art can make corresponding adjustments and changes; without departing from the technical concept of the present invention, the above adjustments and changes can all be used as other implementations of the present invention.

Claims

1. A method for monitoring the risk of recurrence after pulmonary nodule surgery, characterized in that: The method includes: Obtain postoperative follow-up data of patients sorted by follow-up time. Each postoperative follow-up data includes the time interval between each follow-up and the previous follow-up baseline and the recurrence risk characteristics of each follow-up. Disease turning points are identified based on the time intervals between each follow-up visit relative to the previous follow-up baseline and the risk characteristics of each relapse. Disease stage weights are generated based on the time intervals between each follow-up visit relative to the previous follow-up baseline and the disease turning points. Based on the recurrence risk characteristics of each follow-up, a follow-up feature vector is generated, the attention matching score of each follow-up feature vector is calculated, the attention matching score is corrected using the disease stage weight to obtain the correction result, and the visit-level attention weight corresponding to each follow-up is determined based on the correction result. The feature vectors of each follow-up visit are aggregated according to the attention weights of each visit, and postoperative global health characterization data is generated. The recurrence risk monitoring results are output based on the postoperative global health characterization data.

2. The method for monitoring the risk of postoperative recurrence of pulmonary nodules according to claim 1, characterized in that, The generation of disease stage weights based on the time intervals between each follow-up visit relative to the previous follow-up baseline and disease turning events includes: Determine the changes in nodule size, percentage of solid component, tumor markers, and clinical symptoms relative to the baseline values ​​at the t-th follow-up. Relapse risk features that meet the corresponding change conditions are marked as candidate turning point features, and disease turning point event identifiers are generated based on the type of candidate turning point features. The time interval between the t-th follow-up and the previous follow-up is normalized. The normalization result is combined with the correction amount corresponding to the disease turning point event to obtain the disease stage weight of the t-th follow-up.

3. The method for monitoring the risk of postoperative recurrence of pulmonary nodules according to claim 2, characterized in that, The determination of the visit-level attention weights corresponding to each follow-up based on the correction results includes: Each follow-up feature vector is used to obtain the first mapping feature and the second mapping feature through the first feature mapping matrix and the second feature mapping matrix, respectively; An attention matching score is generated based on the matching relationship between the first mapping feature of any follow-up and the second mapping feature of each follow-up. The attention matching score is scaled according to the disease stage weight of the target follow-up, and all scaling results are normalized to obtain the consultation-level attention weight corresponding to each follow-up.

4. The method for monitoring the risk of postoperative recurrence of pulmonary nodules according to claim 3, characterized in that, The generation of each follow-up feature vector based on the recurrence risk characteristics includes: Extract the observation values ​​of each quantifiable observation dimension from each follow-up data, standardize and encode the observation values, and generate corresponding code-level aggregated features; The code-level aggregated features of each follow-up are arranged in chronological order to form a follow-up feature matrix, and the feature data corresponding to each follow-up in the follow-up feature matrix are determined as the corresponding follow-up feature vector.

5. The method for monitoring the risk of postoperative recurrence of pulmonary nodules according to claim 4, characterized in that, The process involves aggregating each follow-up feature vector according to the patient attention weight corresponding to each follow-up visit, generating postoperative global health representation data, and outputting recurrence risk monitoring results based on the postoperative global health representation data, including: The attention weights at each follow-up visit are weighted with the corresponding follow-up feature vectors, and the resulting weighted features are aggregated to generate postoperative global health representation data. Determine the similarity between the postoperative global health characterization data and the average characterization of historical high-risk patient groups and historical low-risk patient groups, respectively. The recurrence risk assessment value is determined by combining the total recurrence risk score for the current period, and the recurrence risk monitoring results are output based on the recurrence risk assessment value.

6. A method for monitoring the risk of postoperative recurrence of pulmonary nodules according to any one of claims 1-5, characterized in that, Before acquiring the patient's postoperative follow-up data, the method further includes: Establish a monitoring index library for the risk of postoperative recurrence of pulmonary nodules. The monitoring index library includes multiple monitoring index categories, and each monitoring index category includes at least one quantifiable observation dimension. Based on the patient's preoperative baseline examination data, intraoperative pathological data, and postoperative first follow-up data, the pathological risk parameters and surgical method influence parameters corresponding to each monitoring indicator category are determined. The combined results of the pathological risk parameters and surgical method influence parameters are normalized to obtain the initial real-time monitoring weights for each monitoring indicator category. The initial real-time monitoring weights are calibrated based on the proportion of real components and the matching relationship between indicator categories.

7. The method for monitoring the risk of postoperative recurrence of pulmonary nodules according to claim 6, characterized in that, When adding a new monitoring dimension to the monitoring indicator library, the method further includes: Determine the correlation between the newly added monitoring dimension and the original observation dimensions in its respective monitoring indicator category; Based on the correlation, the relationship between the newly added monitoring dimension and the recurrence risk, and the real-time monitoring weights of each monitoring indicator category before adjustment, the updated weights of the monitoring indicator categories to which the newly added monitoring dimension belongs are determined. The real-time monitoring weights of the remaining monitoring indicator categories are adjusted synchronously according to the updated weights, so that the sum of the real-time monitoring weights of each monitoring indicator category remains unchanged after the adjustment.

8. The method for monitoring the risk of postoperative recurrence of pulmonary nodules according to claim 7, characterized in that, After outputting the recurrence risk monitoring results, the method further includes: The current core monitoring target is determined based on the real-time monitoring weight of each monitoring indicator category, and the monitoring task corresponding to the current core monitoring target is pushed. The real-time monitoring weight and monitoring interval of the monitoring indicator category to which the current core monitoring target belongs are updated based on the recurrence risk monitoring results, and the real-time monitoring weight of the other monitoring indicator categories are adjusted simultaneously.

9. The method for monitoring the risk of postoperative recurrence of pulmonary nodules according to claim 8, characterized in that, The updated real-time monitoring weights and monitoring intervals include: When the recurrence risk assessment value is within a safe range and the cumulative number of times the current core monitoring target is selected reaches the corresponding threshold, the real-time monitoring weight of the monitoring indicator category to which the current core monitoring target belongs is reduced, the corresponding monitoring interval is extended, and the cumulative number of selections is cleared. When the relapse risk assessment value exceeds the safe range, the real-time monitoring weight of the corresponding monitoring indicator category is increased, the corresponding monitoring interval is shortened, and joint follow-up information including the relapse risk assessment value, historical monitoring data and recent re-examination data is sent to the multidisciplinary follow-up terminal.

10. A system for monitoring the risk of recurrence after pulmonary nodule surgery, characterized in that, include: The data acquisition module is used to acquire postoperative follow-up data of patients sorted by follow-up time. The postoperative follow-up data includes the time interval between each follow-up and the previous follow-up baseline and the recurrence risk characteristics of each follow-up. The stage weight generation module is used to identify disease turning points based on the time interval between each follow-up visit and the previous follow-up baseline and the relapse risk characteristics of each visit, and to generate disease stage weights based on the time interval between each follow-up visit and the disease turning point. The attention correction module is used to generate each follow-up feature vector based on the recurrence risk characteristics, calculate the attention matching score of each follow-up feature vector, correct the attention matching score using the disease stage weight, and determine the consultation-level attention weight corresponding to each follow-up based on the correction result. The risk output module is used to aggregate each follow-up feature vector according to the consultation-level attention weight corresponding to each follow-up visit, generate postoperative global health characterization data, and output recurrence risk monitoring results based on the postoperative global health characterization data.

Images