Primary lung cancer death risk prediction method based on multi-source data

Abstract

The invention provides a primary lung cancer death risk prediction method based on multi-source data, and relates to the field of data processing, and the method comprises the steps: obtaining and associating multi-source heterogeneous data such as hospital treatment, laboratory inspection, image examination and disease control center death registration, and building a training database after cleaning and structuring; screening key predictive factors by adopting LASSO survival regression, constructing a multivariable survival regression model, and generating a score table; converting the baseline features of the patient into an initial death risk score based on a score table; carrying out distribution self-adaptive correction, feature coupling adjustment and risk self-feedback regulation and control in sequence so as to dynamically optimize the score; and finally outputting the adjusted death risk score and providing the adjusted death risk score for clinical use. According to the method, effective fusion of multi-source data and objective screening of key factors are realized, the accuracy, adaptability and interpretability of prediction are improved through a dynamic regulation and control mechanism, and visual and quantitative death risk assessment support can be provided for clinicians.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically to a method for predicting the risk of death from primary lung cancer based on multi-source data. Background Technology

[0002] Primary lung cancer is the leading cause of cancer-related deaths worldwide, and its prognostic assessment is crucial for clinical decision-making. With the popularization of the concept of precision medicine, prediction of mortality risk based on multi-dimensional clinical data of patients has become a research hotspot. The combination of medical big data and artificial intelligence technology provides new possibilities for developing more accurate prediction tools.

[0003] Currently, the technical solutions in this field are mainly divided into two categories. One category is the clinical assessment system that relies on traditional anatomical or histological staging, with relatively fixed assessment dimensions. The other category is models based on statistical learning or machine learning, which construct predictive functions by integrating various clinical indicators of patients. These methods have achieved risk quantification to a certain extent, but each has its own inherent implementation path and characteristics.

[0004] The shortcomings of existing technologies are: the models are not adaptable enough to changes in data sources and distribution, making it difficult to ensure stable performance in different scenarios; at the same time, most methods view the role of features statically and fail to fully consider the complex impact of possible synergistic or antagonistic effects between different clinical indicators on the final risk; in addition, once the model is established, it is difficult to effectively update and adjust itself based on new patient information, which may lead to a decline in predictive efficacy in long-term applications. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method for predicting the risk of death from primary lung cancer based on multi-source data. This method addresses the lack of adaptability to changes in relative data sources and distributions, which makes it difficult to guarantee stable performance in different scenarios. Furthermore, most methods view the effects of features statically and fail to fully consider the complex impact of potential synergistic or antagonistic effects between different clinical indicators on the final risk.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides a method for predicting the risk of death from primary lung cancer based on multi-source data, comprising the following steps: S1. Acquire multi-source heterogeneous data, including hospital visit information, laboratory test results, imaging examination information, and death registration information from the Center for Disease Control and Prevention. Use data encryption and association technology to associate the data according to the patient's unique identity to form the original patient record. S2, perform data cleaning and structured transformation on the original patient records, establish a disease progress data table according to the time series, perform chained multivariate imputation on the missing items in the disease progress data table, perform consistency verification on the imputation results, and generate a structured training database; S3, based on the structured training database, perform feature screening and construct a multivariate survival regression model; use the LASSO survival regression method to screen candidate variables and determine the set of key predictors; construct a multivariate survival regression model based on the set of key predictors; record the output mapping rules of the multivariate survival regression model as a scoring table; S4. Based on the scoring table, the baseline demographic characteristics, initial admission test values, and baseline imaging characteristics of a single patient are converted into scores one by one, and the scores are summed to obtain the initial mortality risk score. S5, the initial mortality risk score is sequentially subjected to distribution adaptive correction, feature coupling adjustment and risk self-feedback control; the distribution adaptive correction and the feature coupling adjustment are performed in every risk prediction, the risk self-feedback control is performed based on subsequent data updates, and the result of the risk self-feedback control affects the distribution adaptive correction and the feature coupling adjustment in the next risk prediction. S6, output the final mortality risk score after risk self-feedback regulation and provide it to clinicians or clinical workstations in the form of the scoring table.

[0007] Preferably, a target patient list is drawn from outpatient and inpatient electronic medical records according to predefined inclusion and exclusion criteria; the inclusion criteria are patients with primary lung cancer coded as ICC10 and aged 18 years or older; the exclusion criteria are patients with incomplete medical records or unclear diagnoses; a unique identifier is assigned to each patient and the initial consultation time and diagnosis time are recorded; hematological and biochemical test values ​​for initial admission and subsequent follow-up are extracted from the laboratory information system; hematological test values ​​include red blood cell count and white blood cell count; biochemical test values ​​include serum potassium and albumin-globulin ratio; imaging examination information is extracted from the imaging archive system and image sequence identifiers are retained; all test values ​​and imaging information retain timestamps and test method information. The process involves: processing unique identifiers using a hash encryption algorithm to generate encrypted identifiers; securely matching the target patient list with death registration information from the Center for Disease Control and Prevention using these encrypted identifiers; recognizing that death registration information includes survival status and survival time; associating successfully matched death outcomes with the corresponding patients; linking all hospital visit information, laboratory test results, imaging examination information, and death registration information to each patient's unique identifier; integrating these to form the original patient record; storing the original patient record in encrypted form in a secure database; initiating a manual review process if identifier conflicts are found during the matching process; prohibiting data association if the identifier cannot be resolved; and generating a data integrity report after association is complete, recording the data source and timestamp.

[0008] Preferably, the process involves standardizing synonymous fields from different data sources in the original patient records using terminology; establishing a unified field mapping table to map field names from different sources to standard field names; standardizing outliers and unstructured text from medical records according to preset medical text parsing rules; these rules include regular expression patterns for identifying and extracting key numerical values ​​and classification information; writing the parsed structured information into a time-sorted progress data table; each record in the progress data table includes a timestamp, a unique patient identifier, and corresponding feature fields and values; performing chained multivariate imputation on the progress data table in chronological order to fill missing items; chained multivariate imputation predicts missing values ​​based on existing observations using an iterative regression model; the imputation process is executed 10 times to generate 10 complete datasets; and the 10 imputation datasets are then processed... Consistency and clinical rationality checks are performed on the results of the data set. The consistency check examines whether the interpolated values ​​are logically contradictory to the observations of the same patient at other time points. The clinical rationality check determines whether the interpolated values ​​are within the physiologically possible range based on the medical reference range. The interpolated values ​​that pass the check are merged to form the final filling result. The fields filled by interpolation are identified in the disease progress data table and the interpolation event is recorded. After completing the interpolation and check, a structured training database is generated. The structured training database contains a complete and missing feature matrix and corresponding survival time and survival status labels. If the check finds that the interpolated values ​​are contradictory or exceed the reasonable range, the interpolation result is discarded and marked as interpolation failure. If the number of interpolation failures exceeds 3, a manual review process is initiated. The structured training database is stored in encrypted form and provides a data foundation for subsequent model training.

[0009] Preferably, candidate variables from the structured training database are extracted according to a preset time window and baseline window and input into the LASSO survival regression model. The time window is defined as all test records within 3 months after the patient's first diagnosis; the baseline window is defined as the test records at the time of first admission. The LASSO survival regression model automatically performs feature selection by applying an L1 norm penalty to the regression coefficients. During model fitting, the coefficients of variables that contribute little to predicting survival outcomes are compressed to zero. 10-fold cross-validation is used to select the optimal penalty parameter lambda. Under the lambda parameter, the model has the minimum prediction error on the validation set. Variables with non-zero regression coefficients are identified as the set of key predictors. The set of key predictors includes age, sex, red blood cell count, serum potassium, albumin-to-globulin ratio, and prothrombin time activity. The set of key predictors is used as the independent variable, and survival time and survival status are used as the dependent variables. The process involves: establishing a multivariate survival regression model; estimating the regression coefficients and hazard ratios of each key predictor and calculating the baseline survival function; generating score mapping rules based on the regression coefficients of each key predictor through linear transformation; mapping the original value range of each factor to a unified score scale of 0 to 100; converting the score mapping rules into a scoring table containing score scales and corresponding survival probabilities; establishing a nonlinear correspondence between the total score scale of the scoring table and the 1-year, 3-year, and 5-year survival probabilities output by the multivariate survival regression model; if model performance is found to be unstable during cross-validation, readjusting the time window definition and repeating the feature selection process; not constructing a multivariate survival regression model if the set of key predictors is empty and data quality must be rechecked; after model construction, storing the scoring table, regression coefficients, and baseline survival function parameters in a structured format for subsequent prediction.

[0010] Preferably, the score mapping rules are obtained by retrieving the scoring table; the static demographic indicators of a single patient are mapped according to the score scale of the scoring table; the static demographic indicators include age and sex; age is located on the age score axis of the scoring table based on its specific value; sex is located on the sex score axis of the scoring table based on its classification; the initial admission test results are mapped according to the score scale of the scoring table; the test results include red blood cell count, serum potassium, albumin-globulin ratio, and prothrombin time activity; each test result is located on the score axis of the corresponding test item of the scoring table based on its specific value; the imaging baseline features are mapped according to the score scale of the scoring table; the imaging baseline features are located on the corresponding imaging feature score axis of the scoring table based on their structured description; each score is recorded. The calculation process involves several steps: first, mapping the source field and timestamp to the item; second, summing all mapped score items to obtain a total score; third, mapping the total score to an initial mortality risk score based on the score-survival probability correspondence established within the scoring table; fourth, writing the initial mortality risk score, along with the total score and details of each score item, into the corresponding patient's record in the disease progress data table; fifth, if a patient characteristic value exceeds the defined range of the scoring table's score axis, the endpoint value of the score axis is used as the score item; sixth, if a patient lacks a characteristic required by the scoring table, the initial mortality risk score is not calculated and is marked as missing data; seventh, after calculation, outputting a structured prediction result containing the initial mortality risk score, total score, and decomposed score items; and finally, using this result as the input basis for subsequent correction and adjustment.

[0011] Preferably, the initial mortality risk score is processed by sequentially performing distribution adaptive correction, feature coupling adjustment, and risk self-feedback regulation. Distribution adaptive correction and feature coupling adjustment are enforced every time a risk prediction is performed for a single patient. Risk self-feedback regulation is not performed at each prediction, but depends on the receipt of subsequent follow-up data or outcome updates for the patient. The results of risk self-feedback regulation will affect the distribution adaptive correction and feature coupling adjustment processes used in the next risk prediction for the same patient or a new patient. Specifically, the process is as follows: first, distribution adaptive correction is performed on the initial mortality risk score to generate a distribution-corrected intermediate score; then, feature coupling adjustment is performed on the distribution-corrected intermediate score to generate a coupling-adjusted intermediate score; if there is no new follow-up data or outcome update, this coupling-adjusted intermediate score is used as the final mortality risk score for this prediction and output; if the system receives subsequent follow-up data for the patient after this prediction... If the data or outcome is updated, risk self-feedback adjustment is triggered. Risk self-feedback adjustment calculates the prediction deviation based on the new data and adjusts the trigger threshold of distribution adaptive correction and the weight adjustment magnitude of feature coupling adjustment accordingly. The adjusted parameters are saved and applied to the next prediction process. In the next prediction, distribution adaptive correction uses the updated trigger threshold, and feature coupling adjustment uses the updated weight adjustment magnitude. Through this mechanism, the prediction model can dynamically adjust its correction and adjustment strategies based on actual feedback. Each correction or adjustment operation generates a detailed audit log, recording the operation type, input, output, parameters, and timestamp. If any step of distribution adaptive correction or feature coupling adjustment fails due to data issues, the entire adjustment process is terminated and the previous valid intermediate score is returned. If the adjustment parameters calculated by risk self-feedback adjustment exceed the preset safety boundary, the current adjustment is discarded and the original parameters are maintained. Finally, the output is the final mortality risk score after the complete adjustment chain.

[0012] Preferably, the distribution adaptive correction is performed through the following steps: First, a distribution bias index is calculated by comparing the statistical distribution of the selected feature set in the structured training database with the empirical distribution of the current input feature set. The selected feature set consists of all or part of the features in the key predictor set. The statistical distribution uses the mean and standard deviation calculated from the feature values ​​of all patients in the structured training database. The empirical distribution uses the mean and standard deviation calculated from the feature values ​​of a single patient or multiple batches of patients in the current input. The distribution bias index is calculated as the standardized distance between the two distributions. Second, it is determined whether the calculated distribution bias index reaches a preset trigger threshold. The preset trigger threshold is set to 0.1. If the distribution bias index is greater than or equal to 0.1, it is determined that the trigger condition has been met. Third, under the trigger condition, a correction operation is performed. The correction operation includes two parallel paths. Path one: Position and scale alignment is performed on the current input feature set according to the statistical distribution of the structured training database. Position alignment shifts the mean of the current input features to match the mean of the training distribution. Scale alignment scales the standard deviation of the current input features to match the standard deviation of the training distribution. The second step involves a directed translation calibration of the initial mortality risk score calculated based on uncorrected features. The direction of the translation calibration is determined by the sign of the distribution bias. The magnitude of the translation is proportional to the absolute value of the distribution bias index. The adjustment amount generated by the calibration operation is recorded as a calibration factor. The calibration factor includes feature translation, scaling factor, or score translation. The fourth step involves writing the calibration factor as an auditable entry into the audit log. The auditable entry includes a timestamp, patient identifier, calibration type, and calibration value. The fifth step involves performing a stationarity test on the calibrated features or scores after calibration. The stationarity test is performed by calculating the coefficient of variation of the feature values ​​or scores before and after calibration. If the coefficient of variation is greater than 0.05, it is considered that abnormal instability has been detected. The sixth step involves automatically reverting to the uncorrected original input features or initial mortality risk score if abnormal instability is detected. The seventh step involves recording the revert event as an auditable entry. The revert event record includes the revert reason, revert time, and calibration factor before reverting. These steps ensure that the distribution adaptive calibration process has the mandatory specification attributes of being triggerable, executable, detectable, and revertible.

[0013] Preferably, feature coupling adjustment is performed through the following steps: First, a scan is performed within the key predictor set based on a predefined set of conditions to identify feature pairs or feature groups that meet the coupling triggering conditions. The predefined set of conditions is pre-set based on clinical knowledge; the conditions include biological associations or pathophysiological synergies between features. For example, a feature group is defined as a serum potassium to albumin / globulin ratio simultaneously being below its clinical reference lower limit. Second, predefined coupling rules corresponding to the identified feature pairs or feature groups are retrieved. The coupling rules are stored in the form of weighted adjustment coefficients. The rules specify a weighted gain operation or a weighted attenuation operation to be performed on the feature pairs or feature groups. The gain operation multiplies the feature's contribution mapping weight in the scoring table by a coefficient greater than 1; the attenuation operation multiplies the contribution mapping weight by a coefficient less than 1. Third, adjustment is triggered when a feature pair or feature group that meets the conditions is identified. After triggering, the specified gain or attenuation operation is performed on the feature pair or feature group according to the retrieved coupling rules. The operation directly adjusts the contribution mapping relationship of the feature in the scoring table. The contribution mapping is reflected in the change of the scale value of the corresponding score axis on the scoring table. Fourth, the contribution mapping before and after adjustment is recorded. The steps are as follows: 1. **Difference Values:** Difference values ​​are stored as weighted adjustments and associated with the current patient identifier. 2. **Validation Testing:** The new contribution mapping after adjustment undergoes validation testing. Validation testing recalculates the predictive performance metrics using historical validation set data. The predictive performance metrics are expressed as the area under the time-dependent curve. 3. **Performance Degradation:** If validation testing finds that the adjustment causes a decrease in the model's predictive performance metrics exceeding 5%, it is considered a performance degradation. 4. **Automatic Rollback Rules:** When a performance degradation determination is established, rollback rules are automatically initiated based on set priorities. Rollback rules prioritize restoring the most recently adjusted contribution mapping. 5. **Restored Contribution Mapping:** A restoration operation is performed to restore the contribution mapping to its original state before adjustment. 6. **Restored Events:** The restore event record is recorded as an auditable entry. The restore event record includes the restore reason, restore time, and involved feature identifiers. 7. **Further Steps:** After adjustment or restoration, the final contribution mapping result is locked and applied to the current patient's risk score calculation. If no feature pairs or feature groups meeting predefined conditions are identified during the scan, no adjustment operation is triggered, and the process proceeds directly to the next step. These steps ensure that the feature coupling adjustment process possesses mandatory specification attributes of conditional triggering, rule execution, performance validation, and safe rollback.

[0014] Preferably, risk self-feedback regulation is implemented through the following steps: First, the system receives subsequent follow-up data from the patient, including new laboratory test results, imaging reports, or outcome updates, and triggers the regulation process upon confirmation of death status from the Center for Disease Control and Prevention. Second, the risk score predicted based on the patient's historical data is compared with the actual observed survival outcome, and the prediction bias is calculated, defined as the absolute difference between the predicted survival probability and the actual binary outcome. Third, deviation events are marked in the disease progress data table, with the marking information including patient identification, prediction time, actual outcome time, and the calculated prediction bias. Fourth, based on the calculated prediction bias, parameters are adjusted according to preset convergence rules: if the average of three consecutive prediction biases is greater than 0.1, the trigger threshold for distribution adaptive correction is increased, such as from 0.1 to 0.15; if the standard deviation of the prediction bias is less than 0.05, the weight adjustment amplitude of feature coupling regulation is reduced, specifically to half of the original amplitude. Fifth, the adjusted trigger threshold and weight adjustment amplitude are compared... As new adjustment parameters, these are sent to the distributed adaptive correction module and the feature coupling adjustment module. These new parameters influence the subsequent correction and adjustment process when predicting risk for the same patient or a new patient. The sixth step involves writing the prediction deviation, adjustment parameters, and adjustment results into the structured training database in the form of an audit log. The audit log includes a timestamp, trigger conditions, parameter values ​​before adjustment, and parameter values ​​after adjustment. The seventh step involves applying the adjusted parameters in real-time during subsequent risk prediction runs: the distributed adaptive correction module uses the new trigger threshold to determine whether to initiate correction, and the feature coupling adjustment module uses the new weight adjustment magnitude to perform gain or attenuation operations. The eighth step involves not initiating the adjustment process and recording data anomalies if the received subsequent data is incomplete or cannot match historical prediction records. The ninth step involves using the boundary value as the final parameter if the calculated adjustment parameters exceed the preset safety boundary, whereby the trigger threshold is set to be no more than 0.2 and the weight adjustment magnitude is no less than one-quarter of the original magnitude. The tenth step involves closing the current adjustment process after completing parameter adjustment and log recording. These steps ensure that the risk self-feedback adjustment process possesses mandatory regulatory attributes of data triggering, deviation calculation, parameter adjustment, audit recording, and real-time application.

[0015] Preferably, the system receives the final mortality risk score after risk self-feedback adjustment, integrates this probability value along with the corresponding predicted time point into a structured output message. This message includes the patient's unique identifier, predicted date, final mortality risk score, and survival probability at each time point. The structured output message format is converted to conform to the clinical workstation data interface specification, including field renaming, numerical unit standardization, and timestamp formatting. The formatted output message is sent to the target clinical workstation via a secure data transmission protocol using an encrypted socket layer protocol. The clinical workstation decrypts and verifies the integrity of the message at the receiving end. Integrity verification is achieved by comparing message digests. After successful verification, the clinical workstation parses the message content and renders it into a visual scoring table, displayed using graphical user interface elements. This includes a patient basic information column, the original values ​​of each key predictor factor and their corresponding scores, a total score scale, and a final survival probability graph. The survival probability graph clearly indicates the predicted survival rates for one year, three years, and five years. Clinicians can directly view the scoring table on the clinical workstation interface, including scrolling, zooming, and printing operations. If data verification fails during transmission, a retransmission mechanism is triggered, attempting up to three times. If all three retransmissions fail, an offline report file is generated in portable document format and sent to the designated clinician's email address via an internal secure email system. Simultaneously, the entire output operation is recorded in the system audit log, including the output time, receiving workstation identifier, transmission status, and final presentation format. Through these steps, the final mortality risk score is delivered to the clinical decision-making process in a secure, reliable, and intuitive scoring table format.

[0016] Preferably, the preset parameters and rules upon which the entire prediction method depends are set and managed through the following mechanism: the preset parameters include the trigger threshold for distribution adaptive correction, with an initial value set to 0.1, and the weight adjustment magnitude for feature coupling adjustment, with an initial value set to a proportional factor calculated based on the regression coefficient; the preset rules include a predefined set of conditions for feature coupling adjustment, pre-compiled based on clinical pathophysiological knowledge, with knowledge sources including medical textbooks and clinical guidelines, stored in list form, each condition clearly specifying the specific feature name and its numerical relationship constituting the feature pair or feature group, as well as the corresponding coupling rule, which is bound to each condition one by one, clearly specifying whether to perform gain operation or attenuation operation and the specific adjustment coefficient. Preset parameters and rules are loaded from the configuration file during system initialization, using an extensible markup language format. During method execution, the distribution adaptive correction module reads the trigger threshold to determine whether to initiate correction, and the feature coupling adjustment module reads the predefined condition set to perform feature scanning and retrieve the corresponding coupling rules. The risk self-feedback control module can dynamically update the trigger threshold and weight adjustment range based on convergence rules. The updated parameters are written back to the runtime memory cache and asynchronously persisted to the configuration file to ensure that the latest parameters are loaded the next time the system restarts. Preset rules are read-only during runtime, and revisions must be performed by authorized administrators through the offline management interface. The management interface provides CRUD functions for condition sets and coupling rules. Any modification must be confirmed twice and a version snapshot is automatically generated. The version snapshot records the modification time, operator, and modified content. The modified new rules will automatically take effect in the next prediction cycle. The system performs the following upon each load: providing a consistency check function for parameters and rules; if the check finds a conflict or anomaly, it automatically rolls back to the previous valid version and issues an alarm, sending the information to the system administrator; through the above settings, storage and management mechanisms, it ensures that the preset foundation supporting the operation of the method has mandatory specification attributes that are configurable, updatable, auditable and rollbackable.

[0017] (III) Beneficial Effects This invention provides a method for predicting the risk of death from primary lung cancer based on multi-source data. It has the following beneficial effects: 1. This invention significantly improves the accuracy and stability of prediction through a triple mechanism of distribution adaptive correction, feature coupling adjustment, and risk self-feedback regulation. Distribution adaptive correction ensures that the distribution of input features is consistent with that of training data, reducing bias caused by data drift. Feature coupling adjustment optimizes the interaction weights between features based on clinical knowledge, making the model more consistent with pathophysiological laws. Risk self-feedback regulation dynamically adjusts the correction parameters according to the actual outcome, enabling the model to have continuous learning and adaptive capabilities, thereby maintaining robust and reliable predictive performance when facing new data or different patient groups.

[0018] 2. This invention achieves a leap from static prediction to dynamic adaptive regulation, greatly enhancing clinical applicability and operability. The model receives subsequent follow-up data and calculates prediction bias, automatically optimizing internal correction and adjustment parameters so that the prediction results can be iteratively updated as the patient's disease progresses. Finally, the risk score is output in an intuitive scoring table format, which is convenient for clinicians to understand and apply quickly. The entire process organically combines multi-source data fusion, intelligent feature screening, and dynamic risk calibration, providing an efficient and reliable quantitative tool for personalized prognostic assessment and dynamic management of lung cancer patients. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is the overall flowchart of the primary lung cancer mortality risk prediction of the present invention; Figure 2 This is a detailed flowchart of the dynamic adjustment mechanism of the present invention; Figure 3 This is a patient baseline feature map of the present invention; Figure 4 This is the feature selection diagram of the present invention; Figure 5 This is a graph showing the results of the multivariate survival regression model of this invention. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a method for predicting the risk of death from primary lung cancer based on multi-source data. Figure 1This document describes the overall flowchart of primary lung cancer mortality risk prediction, including the entire process of its initial implementation in a specific hospital scenario. Implementation begins with hospital IT engineers exporting data from the hospital information system and laboratory information system. The engineers first perform data acquisition steps: querying the hospital information system according to pre-defined inclusion and exclusion criteria. Inclusion criteria are patients with ICC10 code C34 and aged 18 years or older; exclusion criteria are patients with missing key diagnostic information in their medical records or only outpatient records without inpatient records. This yields an initial list of 1308 patients. Each patient on the list is assigned a unique identifier generated from an internal hospital identifier and a hash of the consultation date, while the initial consultation time and the time of diagnosis confirmed by the pathology report are recorded. Subsequently, hematological and biochemical test results from the laboratory information system are extracted in batches from these patients' initial admission and subsequent follow-up. Extracted indicators include 26 items such as red blood cell count, white blood cell count, serum potassium, serum sodium, albumin-globulin ratio, and prothrombin time activity. Each test result is associated with its sampling timestamp and the code of the testing method used by the laboratory department. The initial baseline CT image sequence identifiers of patients are extracted from the image archiving system. Then, a crucial external data integration is performed: engineers generate encrypted identifiers from the anonymized unique identifiers of 1308 patients using secure one-way hash encryption, and upload these encrypted identifiers to the authorized regional disease control and prevention center's data interface. The disease control and prevention center matches these identifiers in its death registration database and returns the matching results, which include the specific survival status (dead or alive) and the specific date of death or last follow-up for each identifier. The hospital's information department integrates the returned death outcome information with internal data based on the unique identifiers to form the original patient records; these original patient records are stored in a secure database with access controls.

[0022] Next, data cleaning and structuring transformation steps are performed. The data processing program first reads the original patient records. The program's built-in terminology standardization module automatically identifies and maps synonymous fields from different sources. For example, the "age" field from the patient record's homepage and the "patient age" field from the demographic table are uniformly mapped to the standard field "age". Next, the unstructured text in the progress notes is parsed. The parsing program identifies and extracts key information according to preset medical text parsing rules. For example, keywords such as "cough," "hemoptysis," and "chest pain" are extracted from the "chief complaint" text and converted into structured categorical variables. All structured data is organized chronologically and written into a progress note table; each row in the progress note table represents a patient's complete record at a specific point in time. The program then checks for missing values ​​in the progress note table. For missing laboratory indicators, the program initiates a chained multivariate imputation algorithm. The algorithm uses other complete variables in the dataset as predictors and iteratively performs regression prediction for each missing variable. The imputation process is repeated 10 times, generating 10 complete datasets. The program performed a consistency check on the 10 imputation results. The check found that the serum potassium values ​​for a certain patient ranged from 3.9 to 4.2 mmol / L across the 10 imputations, a fluctuation within a reasonable range, and was therefore considered consistent. Next, a clinical rationality check was performed. The check rules identified a negative albumin-to-globulin ratio from one of the imputations; the program automatically discarded that result and recalculated. Finally, the imputations that passed the checks were merged and filled back into their original missing positions, and marked "Impacted" in the patient progress data table. After these operations, the final structured training database was generated.

[0023] The next step involves feature selection and model building. Data analysts load the structured training database using statistical software. The software sets patient survival time and survival status as dependent variables and 26 candidate features as independent variables. The software performs LASSO survival regression analysis, using 10-fold cross-validation to optimize the model's penalty parameter. The cross-validation results show that when the penalty parameter lambda equals 0.055, the model is both concise and has good prediction error. Under this parameter, LASSO regression compresses the coefficients of 20 independent variables to zero, ultimately selecting 6 key predictors with non-zero coefficients: age, sex, red blood cell count, serum potassium, albumin-to-globulin ratio, and prothrombin time activity. Data analysts use these 6 key predictors as independent variables to refit a standard multivariate Cox proportional hazards regression model. The model calculates the regression coefficient and hazard ratio for each factor; for example, the regression coefficient for age is 0.046, and the hazard ratio is 1.047. Based on these regression coefficients, the software automatically generates score mapping rules. These rules linearly map the value range of each factor to a score axis of 0 to 100. For example, the score increases by 0.5 points for every year of age. The software further converts the score mapping rules into a visual nomogram scoring table. Below the scoring table, a conversion scale is established between the total score and the survival probability for 1 year, 3 years, and 5 years. This scoring table serves as the core tool for subsequent predictions.

[0024] Subsequently, risk prediction was performed on Zhang San, a newly diagnosed patient with primary lung cancer. The doctor entered Zhang San's unique identifier into the clinical workstation, and the system automatically retrieved Zhang San's baseline data, including: Zhang San's age 65, male, and initial admission test results, such as a red blood cell count of 4.1 × 10⁻⁶. 12 The serum potassium level was 4.0 mmol / L, the albumin-to-globulin ratio was 1.5, and the prothrombin time activity was 105%. The system retrieved the scoring table, mapping Zhang San's age (65 years old) to the age score axis, resulting in 32 points; his gender (male) was mapped to 10 points; his red blood cell count to 8 points; his serum potassium level to 15 points; his albumin-to-globulin ratio to 20 points; and his prothrombin time activity to 18 points. The system added these six scores together, resulting in a total score of 103 points. Based on the scale at the bottom of the scoring table, the system mapped these 103 points to specific survival probabilities: Zhang San's 1-year survival probability was 85%, his 3-year survival probability was 75%, and his 5-year survival probability was 60%. This result was recorded as the initial mortality risk score.

[0025] After calculating the initial score, the system immediately performs distribution adaptive correction. The system compares the six feature values ​​of the current input Zhang San with the distribution of these six features in the structured training database, calculating a distribution bias index of 0.08. Since 0.08 is less than the preset trigger threshold of 0.1, position scale alignment or score translation calibration is not triggered. Next, the system performs feature coupling adjustment. The system scans Zhang San's six key predictor values ​​and, based on a predefined set of conditions, finds that Zhang San's serum potassium is 4.0 mmol / L and the albumin-to-globulin ratio is 1.5, both near the lower limit of the normal range. This feature combination satisfies a predefined coupling trigger condition, and the system retrieves the corresponding rule requiring a slight attenuation operation on this combination. According to the rules, the system multiplied the contribution weight of serum potassium to albumin-globulin ratio in this prediction by a coefficient of 0.9. This weight adjustment caused the scores for these two factors to change from 15 and 20 to 13.5 and 18 respectively. Zhang San's total score thus changed from 103 to 98.5. After remapping, Zhang San's one-year survival probability was updated to 87%. The system recorded this weight adjustment. Since this was the first prediction and there was no follow-up data, it did not trigger risk self-feedback regulation. The final mortality risk score obtained after correction and adjustment was locked.

[0026] Finally, the system outputs the final result. The system sends an encrypted structured message containing patient Zhang San's identifier, prediction date, final score, and 1-year probability of 87%, 3-year probability of 77%, and 5-year probability of 62% to the clinical workstation used by the doctor. The clinical workstation receives and decrypts the message, rendering Zhang San's personalized score sheet on the screen. The score sheet clearly displays the scores corresponding to each indicator value and the final survival probability curve in a graphical interface. The doctor reviews this score sheet and uses it as a reference for communicating the prognosis with the patient's family and developing a treatment plan. All data operations and model calls implemented in this process generate audit logs, which are stored in the database for future verification. Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the risk of death from primary lung cancer based on multi-source data, characterized in that, Includes the following steps: S1. Acquire multi-source heterogeneous data, including hospital visit information, laboratory test results, imaging examination information, and death registration information from the Center for Disease Control and Prevention. Use data encryption and association technology to associate the data according to the patient's unique identity to form the original patient record. S2, perform data cleaning and structured transformation on the original patient records, establish a disease progress data table according to the time series, perform chained multivariate imputation on the missing items in the disease progress data table, perform consistency verification on the imputation results, and generate a structured training database; S3, based on the structured training database, perform feature screening and construct a multivariate survival regression model; use the LASSO survival regression method to screen candidate variables and determine the set of key predictors; construct a multivariate survival regression model based on the set of key predictors; record the output mapping rules of the multivariate survival regression model as a scoring table; S4. Based on the scoring table, the baseline demographic characteristics, initial admission test values, and baseline imaging characteristics of a single patient are converted into scores one by one, and the scores are summed to obtain the initial mortality risk score. S5, the initial mortality risk score is sequentially subjected to distribution adaptive correction, feature coupling adjustment and risk self-feedback control; the distribution adaptive correction and feature coupling adjustment are performed in every risk prediction, the risk self-feedback control is performed based on subsequent data updates, and the result of the risk self-feedback control affects the distribution adaptive correction and feature coupling adjustment in the next risk prediction. S6, output the final mortality risk score after risk self-feedback regulation and provide it to clinicians or clinical workstations in the form of the scoring table.

2. The method for predicting the risk of death from primary lung cancer based on multi-source data according to claim 1, characterized in that, S1 includes: S11. Select a list of target patients from outpatient and inpatient electronic medical records according to predefined inclusion and exclusion criteria, assign a unique identifier to each patient, and record the start time of consultation and the time of diagnosis. S12, extract hematological and biochemical test values ​​from the laboratory information system and image archiving system for initial admission and subsequent follow-up, and retain the timestamps, test method information and image sequence identifiers of each test; S13, using the unique identifier to perform secure matching with a hash-encrypted identifier, associate the target patient list with the death registration information of the disease prevention and control center, and save the matching result in the form of an encrypted identifier to the original patient record.

3. The method for predicting the risk of death from primary lung cancer based on multi-source data according to claim 1, characterized in that, S2 includes: S21, standardize the terminology for synonymous fields from different data sources and establish a unified field mapping table; S22, outliers and unstructured text from medical records are normalized according to preset medical text parsing rules, and the parsing results are written into a time-sorted medical record data table. S23, perform chained multivariate imputation on the disease course data table in chronological order to fill missing items, perform consistency and clinical rationality checks on the imputation results, and mark the imputation events.

4. The method for predicting the risk of death from primary lung cancer based on multi-source data according to claim 1, characterized in that, S3 includes: S31, the candidate variables in the structured training database are input into the LASSO survival regression model according to the time window and the baseline window for screening to obtain a set of key predictive factors; S32, establish a multivariate survival regression model using the set of key predictors, estimate the regression coefficients and baseline survival function, and generate score mapping rules; S33, convert the score mapping rule into a rating table, and establish the correspondence between the score scale of the rating table and the survival probability output by the multivariate survival regression model.

5. The method for predicting the risk of death from primary lung cancer based on multi-source data according to claim 1, characterized in that, S4 includes: S41, map the static demographic indicators, initial admission test results and imaging baseline characteristics of a single patient to score items according to the score scale of the scoring table, and record the source field and timestamp of each score item; S42, sum the scores to obtain a total score, map the total score to the initial mortality risk score through the score-probability correspondence of the scoring table, and write the initial mortality risk score into the disease course data table.

6. The method for predicting the risk of death from primary lung cancer based on multi-source data according to claim 1, characterized in that, S4 includes: S51, in the distribution adaptive correction, the following operations are performed: comparing the statistical distribution of the selected feature set in the structured training database with the empirical distribution of the current input feature set, and calculating the distribution deviation index; when the distribution deviation index reaches a preset trigger threshold, performing position and scale alignment on the current input feature or performing translation calibration on the initial mortality risk score, and recording the correction factor; The distribution adaptive correction includes: S511, by comparing the statistical distribution of the structured training database on the selected feature set with the empirical distribution of the current input features, calculate the deviation index and determine whether the deviation index reaches the preset trigger threshold; S512, upon triggering, align the current input features with the training distribution in terms of position and scale or perform a directed translation calibration on the initial mortality risk score, and record the correction factor as an auditable entry; S513 performs stationarity testing on the corrected features or scores after correction is completed. When abnormal instability is detected, it automatically reverts to the uncorrected input and records the revert event.

7. The method for predicting the risk of death from primary lung cancer based on multi-source data according to claim 1, characterized in that, S5 further includes: S52, in the feature coupling adjustment, the following operations are performed: in the set of key predictive factors, feature pairs or feature groups that satisfy specific combination conditions predefined based on clinical knowledge are identified; according to the predefined coupling rules corresponding to the specific combination conditions, the contribution weights of the identified feature pairs or feature groups in the scoring table are subjected to gain or attenuation operations, and the weight adjustment amount is recorded; The feature coupling adjustment includes: Within the set of key predictors, feature pairs or feature groups that satisfy the coupling triggering conditions are identified according to a predefined set of conditions, and the corresponding coupling rules are retrieved. When triggered, a weighted gain or attenuation operation is performed on the feature pair or feature group according to the coupling rule to adjust its contribution mapping in the scoring table, and the difference in contribution mapping before and after adjustment is recorded. Verification tests were conducted on the adjusted contribution mapping. If the adjustment caused a decrease in model prediction performance, a rollback rule was initiated according to the set priority to restore the original contribution mapping, and the restoration event was recorded.

8. The method for predicting the risk of death from primary lung cancer based on multi-source data according to claim 1, characterized in that, In S5: S53, in the risk self-feedback regulation, the following operations are performed: receiving subsequent follow-up data or outcome updates of patients, comparing the risk score predicted based on historical data with the actual outcome, and calculating the prediction deviation; based on the prediction deviation, adjusting the trigger threshold of the distribution adaptive correction and the weight adjustment magnitude of the feature coupling regulation according to the preset convergence rule. The adjusted trigger threshold, weight adjustment range, and prediction deviation are written back to the structured training database; The aforementioned risk self-feedback control includes: When receiving follow-up data or outcome updates, the actual outcome is compared with the previous prediction and the deviation is calculated. The deviation event is marked in the disease progress data table. Based on the deviation, the trigger threshold for the distribution adaptive correction and the weight adjustment magnitude for the feature coupling adjustment are adjusted according to the preset convergence rules, and the adjustment parameters are sent out to affect the subsequent prediction correction process. The deviation, adjustment parameters, and adjustment results are written into an audit log, recording the timestamp and triggering conditions, and the adjustment parameters are applied in real time during subsequent operation.

Images