A method for processing nonlinear relationship data of stroke prognosis
By constructing a statistical set of stroke prognosis data, performing variable stability screening and nonlinear relationship identification, and using resampling verification to adjust the sensitive interval, the problem of instability in identifying nonlinear relationships in stroke prognosis was solved, and the stability and repeatability of the identification were improved.
Patent Information
- Application Number
- CN202611356378.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to accurately identify and stably characterize the nonlinear relationships of continuous prognostic variables in long-term prognostic assessment of stroke patients, and resampling processes are insufficient to correct the search range of nonlinear relationships, resulting in unstable results and poor reproducibility.
By constructing a statistical dataset of stroke prognosis, we perform variable stability screening, nonlinear candidate identification, progressive refinement of sensitive intervals, and resampling stability verification to identify and characterize nonlinear relationships. We also use the resampling results to adjust the candidate sensitive intervals to improve stability.
It achieves adaptive identification and stable characterization of nonlinear relationships among stroke prognostic variables, improves the stability and repeatability of identification, and reduces the impact of sample distribution and local data fluctuations.
Smart Images

Figure CN123072731A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of statistical data processing technology, and more specifically, to a method for processing nonlinear relationship data of stroke prognosis. Background Technology
[0002] In the context of long-term prognostic assessment and follow-up data analysis for stroke patients, age, gender, inflammatory markers, nutritional indicators, and survival outcomes can reflect the patient's long-term risk status from different dimensions. Existing protocols typically use Cox proportional hazards analysis, correlation analysis, and variable screening methods based on previous case and follow-up data to identify prognostic variables related to survival outcomes. Furthermore, restricted cubic splines are used to fit the nonlinear relationship between some continuous variables and mortality risk. The prognostic model is evaluated through model discrimination, calibration capability, and resampling validation results to obtain risk statistics and risk stratification results at different follow-up time points.
[0003] However, processing nonlinear relationships of continuous prognostic variables typically relies on pre-specified variables and direct overall fitting, lacking a mechanism to automatically identify nonlinear regions based on risk changes corresponding to different variable value ranges. This makes it difficult to accurately determine the local locations where risk changes transition from continuous to gradual or where a directional shift occurs. Furthermore, fixed fitting structures are easily affected by sample distribution and local data fluctuations, causing shifts in nonlinear inflection points and sensitive intervals obtained under different sample conditions. Existing resampling processes are mostly used for overall model performance evaluation, making it difficult to use the positional shifts generated by resampling to correct the nonlinear relationship search range. This leads to a separation between the identification, fitting, and stability verification of local nonlinear relationships, affecting the stability and repeatability of stroke prognostic nonlinear relationship data processing results. Summary of the Invention
[0004] To overcome the aforementioned deficiencies of the prior art, the following solution is proposed to address the problem of unstable identification of nonlinear relationships among stroke prognostic variables in the aforementioned background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A method for processing nonlinear relationship data of stroke prognosis includes the following steps:
[0007] S1. Obtain demographic data, inflammatory marker data, nutritional marker data, and survival outcome data of stroke subjects, associate them according to the subject identifier, and form a set of stroke prognostic statistics based on the completeness of the data;
[0008] S2. Perform statistical association analysis, variable correlation analysis and repeated shrinkage screening on the statistical data set of stroke prognosis. Determine the stability of variable retention based on the retention of each variable in the repeated screening and generate a set of candidate prognostic variables.
[0009] S3. Divide the candidate prognostic variables into continuous data intervals according to the order of variable values, determine the interval risk association results between each continuous data interval and the survival outcome data, and determine the nonlinear candidate variables and candidate sensitive intervals based on the direction and magnitude of change of the risk association results of adjacent intervals.
[0010] S4. The candidate sensitive intervals are progressively refined. Based on the stability of the risk change position in the adjacent refinement results, the nonlinear turning point and nonlinear sensitive interval are determined. Based on this, the nonlinear fitting structure is configured to generate nonlinear relationship characterization data.
[0011] S5. Resample the statistical data set of stroke prognosis, perform stability check on the nonlinear turning point and nonlinear sensitive interval corresponding to different resampled data, and adjust the candidate sensitive interval according to the offset of the turning point and refine it again. When the check passes, determine the target nonlinear relationship characterization data and generate the stroke prognosis statistical processing results.
[0012] Furthermore, the formation of the stroke prognostic statistical data set specifically includes:
[0013] We obtained age and sex as demographic data for stroke patients, neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, monocyte-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, systemic immune inflammation index, systemic inflammatory response index, and systemic inflammatory aggregation index as inflammatory marker data, body mass index, waist circumference, albumin, hemoglobin, and total protein as nutritional marker data, and follow-up time and all-cause mortality results as survival outcome data.
[0014] The population data, inflammatory marker data, nutritional marker data, and survival outcome data are correlated according to the stroke subject identifier. The completeness of the variables and survival outcome data required for statistical processing in each correlated record is checked. Based on the completeness, corresponding data availability tags are generated. The correlated records that meet the statistical processing conditions are organized according to a unified field structure to form a set of stroke prognostic statistical data.
[0015] Furthermore, the generation of the candidate prognostic variable set specifically includes:
[0016] Survival risk statistical analysis was performed on each variable and survival outcome data in the stroke prognosis statistical data set, and the first candidate variable set was formed based on the obtained statistical correlation results.
[0017] Perform correlation analysis on the first candidate variable set, and mark the variables with correlation to form a set of variables to be narrowed down for screening;
[0018] Based on the statistical data set of stroke prognosis, multiple screening data subsets are repeatedly formed according to a preset sampling method. Shrinkage screening is performed on each screening data subset, and the corresponding shrinkage screening parameters are determined through cross-validation. The retention results of each candidate variable in different screening data subsets are recorded.
[0019] The retention stability of each candidate variable is determined based on its retention frequency and the consistency of retention results in the entire selected data subset. Variables whose retention stability meets the preset retention conditions are identified as candidate prognostic variables, thus forming a set of candidate prognostic variables.
[0020] Furthermore, determining the nonlinear candidate variable and the candidate sensitive interval specifically includes:
[0021] For candidate prognostic variables, the corresponding stroke prognostic statistics are arranged in order of variable value from low to high, and then divided into multiple continuous data intervals in such a way that the number of samples contained in each interval is approximately the same.
[0022] Continuous data intervals are used as interval grouping identifiers. Based on the survival time and all-cause mortality results of stroke subjects within each interval, the corresponding interval risk association results are determined, and an interval risk change sequence is formed according to the arrangement order of continuous data intervals.
[0023] By comparing the risk correlation results of adjacent continuous data intervals in turn, the corresponding risk change direction and risk change magnitude are determined. When the risk change direction changes, the risk change magnitude increases or decreases continuously, or the risk change changes from continuous change to gradual change, the corresponding candidate prognostic variable is determined as a nonlinear candidate variable.
[0024] The initial range of change is determined based on the continuous data interval where the risk change occurs and its adjacent intervals before and after it, and the initial range of change is determined as the candidate sensitive interval of the nonlinear candidate variable.
[0025] Furthermore, the progressive refinement of the candidate sensitive regions specifically includes:
[0026] The candidate sensitive interval is divided into multiple continuous refined data segments, and refined risk association results corresponding to each refined data segment are generated according to the method for determining the interval risk association results.
[0027] The results of the risk association between adjacent data segments are compared according to the order of the detailed data segments. When the direction of risk change changes or the trend of the magnitude of adjacent risk change changes changes, the position between the corresponding data segments is determined as the position of risk change in this round.
[0028] The data segment where the risk change location is located in this round and its adjacent data segments before and after it are used to form the next round of refinement intervals. The width of the data segment in the next round of refinement intervals is narrowed and the risk change location is re-determined.
[0029] Compare the risk change positions obtained in the current round with those in the previous round. When the offset between the risk change positions in consecutive preset rounds does not exceed the preset position tolerance, stop the progressive refinement. The corresponding risk change position is determined as the non-linear turning point, and the continuous data segment covering the non-linear turning point when refinement stops is determined as the non-linear sensitive interval.
[0030] Furthermore, the configuration of the nonlinear fitting structure and the generation of nonlinear relationship representation data specifically includes:
[0031] The candidate positions of fitting nodes are determined based on the nonlinear inflection points, nonlinear sensitive intervals, and the value range of candidate prognostic variables. The distribution density of fitting nodes within the nonlinear sensitive interval is higher than that outside the nonlinear sensitive interval, thus forming a combination of multiple candidate fitting nodes.
[0032] Restricted cubic spline fitting is performed on the nonlinear candidate variables according to each candidate fitting node combination, and the obtained nonlinear fitting results are incorporated into the survival risk statistical model to obtain the risk change relationship corresponding to each candidate fitting node combination;
[0033] Determine whether the direction of change of each risk change relationship before and after the nonlinear turning point is consistent with the risk change sequence of the interval, and determine whether the turning point obtained by fitting is located within the corresponding nonlinear sensitive interval. Evaluate the model discrimination ability of candidate fitting node combinations that meet the above conditions.
[0034] Based on the model's ability to distinguish, the target fitting node combination is determined, and the nonlinear turning points, nonlinear sensitive intervals, fitting node positions, and risk change relationships corresponding to the target fitting node combination are organized into nonlinear relationship characterization data.
[0035] Furthermore, the execution stability check specifically includes:
[0036] Multiple resampling with replacement was performed on the stroke prognosis statistics set to form multiple resampled data sets. Candidate sensitive interval refinement and nonlinear fitting were then performed again on each resampled data set to obtain the corresponding resampled nonlinear inflection points and resampled nonlinear sensitive intervals.
[0037] The position offset of each resampled nonlinear transition position relative to the original nonlinear transition position is determined, and the proportion of resampling times where the position offset does not exceed the preset position tolerance is counted to obtain the stability ratio of the transition positions.
[0038] The overlap length of the resampled nonlinear sensitive interval and the original nonlinear sensitive interval and the combined coverage length of the two are determined respectively. The corresponding interval coverage consistency is determined according to the ratio of the overlap length to the combined coverage length.
[0039] When the stability ratio at the turning point reaches the preset stability threshold, and the number of resampled nonlinear sensitive intervals that meet the preset coverage consistency threshold reaches the preset stability ratio, the stability check is deemed to have passed; otherwise, the stability check is deemed to have failed.
[0040] Furthermore, when the stability check fails, the candidate sensitivity interval is adjusted, specifically including:
[0041] Obtain the offset direction of each resampled nonlinear transition position relative to the original nonlinear transition position, and collect the resampled nonlinear transition positions that have experienced positional offset according to the offset direction;
[0042] When the resampling nonlinear transition positions that reach the preset ratio are concentrated on the same side of the candidate sensitive interval, the corresponding side boundary of the candidate sensitive interval is expanded along the corresponding offset direction. When the resampling nonlinear transition positions are distributed on both sides of the candidate sensitive interval and both reach the corresponding preset ratio, the side boundaries of the candidate sensitive interval are expanded at the same time.
[0043] Based on the candidate sensitive interval after boundary adjustment, the progressive refinement, nonlinear fitting and stability check are re-executed until the stability check is passed or the preset maximum number of corrections is reached.
[0044] When the stability check passes, the corresponding nonlinear relationship representation data is determined as the target nonlinear relationship representation data. If the check fails after reaching the preset maximum number of corrections, an unstable nonlinear relationship marker is generated for the corresponding nonlinear candidate variable.
[0045] Furthermore, the step of determining the target nonlinear relationship characterization data and generating stroke prognostic statistical processing results specifically includes:
[0046] The set of candidate prognostic variables includes age, sex, systemic inflammatory response index, body mass index, and albumin, wherein albumin is identified as a nonlinear candidate variable;
[0047] The data interval containing albumin data of the stroke subjects to be treated is determined based on the target nonlinear relationship characterization data, and the nonlinear risk contribution characterization of albumin is determined based on the risk change relationship corresponding to the data interval.
[0048] The variable data corresponding to age, gender, systemic inflammatory response index and body mass index are combined with the nonlinear risk contribution characterization and incorporated into the survival risk statistical model. The survival risk statistical model is used to generate stroke prognostic statistics for stroke subjects to be treated.
[0049] The prognostic statistics of stroke are correlated with the corresponding target nonlinear relationship characterization data to form a record of prognostic statistical data for the stroke subjects to be treated.
[0050] Furthermore, the generation of stroke prognostic statistical processing results also includes:
[0051] Based on the survival risk statistical model and stroke prognostic statistics, the all-cause mortality risk statistics for stroke patients to be treated at the one-year, three-year and five-year follow-up time points were determined respectively.
[0052] Obtain the stroke prognostic statistics of each stroke object in the stroke prognostic statistics set, determine the risk stratification boundary with the median of the stroke prognostic statistics, compare the stroke prognostic statistics of the stroke object to be treated with the risk stratification boundary, and determine the corresponding high-risk level or low-risk level.
[0053] The stroke patient identifier, stroke prognostic statistics, all-cause mortality risk statistics at each follow-up time point, risk level, nonlinear candidate variable identifier, nonlinear inflection point, nonlinear sensitive interval, and nonlinear risk contribution characterization are correlated to form and output the stroke prognostic statistical processing results.
[0054] The technical effects and advantages of the method for processing nonlinear relationship data of stroke prognosis according to the present invention are as follows:
[0055] This invention achieves adaptive identification and stable characterization of nonlinear relationships among stroke prognostic variables by constructing a data processing chain that includes stroke prognostic statistical data set, stable screening of candidate prognostic variables, identification of nonlinear candidates, progressive refinement of sensitive intervals, and resampling stability verification. In the variable screening stage, statistical association analysis, variable correlation analysis, and repeated shrinkage screening are combined to determine the stability of variable retention based on the retention frequency and consistent direction of action of variables in different screening data subsets. This reduces the impact of accidental screening caused by single sample division and variable correlation, and makes the candidate prognostic variables entering subsequent nonlinear analysis more stable.
[0056] In the nonlinear relationship processing stage, continuous data intervals are divided according to variable values. Nonlinear candidate variables and candidate sensitive intervals are identified by the direction and magnitude of changes in risk association results between adjacent intervals. The candidate sensitive intervals are then progressively refined to gradually converge the risk change positions and determine the nonlinear turning points and nonlinear sensitive intervals. Based on this, nonlinear fitting structures with different local node densities are configured to reduce local relationship shifts caused by overall fitting with fixed nodes. Furthermore, the stability of nonlinear turning points and sensitive intervals is checked using resampling results. If the check fails, the candidate sensitive intervals are adjusted in the opposite direction based on the shift direction of the turning point and refined again, so that the sample reconstruction results directly participate in the nonlinear relationship correction. Finally, stable target nonlinear relationship characterization data and stroke prognosis statistical processing results are formed, thereby improving the stability, repeatability, and reliability of statistical results in nonlinear prognostic relationship identification. Attached Figure Description
[0057] Figure 1 This is a flowchart illustrating a method for processing nonlinear relationship data of stroke prognosis according to the present invention. Detailed Implementation
[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] In order to achieve the above objectives, Figure 1 A schematic diagram of the processing method for nonlinear relationship data of stroke prognosis according to the present invention is given, which specifically includes the following steps;
[0060] S1. Obtain demographic data, inflammatory marker data, nutritional marker data, and survival outcome data of stroke subjects, associate them according to the subject identifier, and form a set of stroke prognostic statistics based on the completeness of the data;
[0061] S2. Perform statistical association analysis, variable correlation analysis and repeated shrinkage screening on the statistical data set of stroke prognosis. Determine the stability of variable retention based on the retention of each variable in the repeated screening and generate a set of candidate prognostic variables.
[0062] S3. Divide the candidate prognostic variables into continuous data intervals according to the order of variable values, determine the interval risk association results between each continuous data interval and the survival outcome data, and determine the nonlinear candidate variables and candidate sensitive intervals based on the direction and magnitude of change of the risk association results of adjacent intervals.
[0063] S4. The candidate sensitive intervals are progressively refined. Based on the stability of the risk change position in the adjacent refinement results, the nonlinear turning point and nonlinear sensitive interval are determined. Based on this, the nonlinear fitting structure is configured to generate nonlinear relationship characterization data.
[0064] S5. Resample the statistical data set of stroke prognosis, perform stability check on the nonlinear turning point and nonlinear sensitive interval corresponding to different resampled data, and adjust the candidate sensitive interval according to the offset of the turning point and refine it again. When the check passes, determine the target nonlinear relationship characterization data and generate the stroke prognosis statistical processing results.
[0065] The creation of a statistical dataset on stroke prognosis includes the following specific implementation details:
[0066] This implementation involves classifying, identifying, and standardizing the structure of demographic, inflammatory marker, nutritional, and survival outcome data for stroke patients. This results in a comprehensive set of stroke prognostic statistics that can be directly accessed for subsequent statistical correlation analysis and nonlinear relationship processing. Specifically, each stroke patient is used as a data unit, and demographic, inflammatory marker, nutritional, and survival outcome data are acquired separately. Demographic data includes age and sex; inflammatory marker data includes neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, monocyte-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, systemic immune inflammation index, systemic inflammatory response index, and systemic inflammatory aggregation index; nutritional marker data includes body mass index, waist circumference, albumin, hemoglobin, and total protein; and survival outcome data includes follow-up time and all-cause mortality results.
[0067] For indicators directly derived from test records, read the corresponding test results and test object identifiers; for ratio-type indicators formed by two or more basic test items, first verify whether the basic test items involved in the calculation belong to the same stroke object based on the object identifier, and generate the corresponding indicator values according to the predetermined indicator composition relationship.
[0068] When there are multiple detection records of the same type for the same object, the records participating in this statistical processing are selected according to the predetermined data retrieval rules, and the source identifier of the selected records is retained in the corresponding data field, so as to maintain the correspondence between the indicator data and the original records.
[0069] After acquiring various types of data, consistency checks are performed on the object identifiers in different data records. Demographic data, inflammatory marker data, nutritional marker data, and survival outcome data with consistent object identifiers are grouped into the same object data record.
[0070] When a data record does not have a corresponding object identifier, the object identifier repeatedly points to different objects, or a conflict occurs in the survival outcome attribution after association, the record is set to an affiliation abnormal state and is not included in the current statistical data set.
[0071] After completing the attribution verification, an internal object number is configured for each stroke object, and a correspondence between different data categories is established based on the internal object number, so that each variable in subsequent statistical processing can be traced back to the same stroke object.
[0072] Further, perform field integrity checks on the object data records after the association is completed. Field integrity checks are performed according to the data requirements of subsequent processing tasks. Among them, the records used for statistical association analysis must have at least the variable to be analyzed, the survival time, and the all-cause mortality result.
[0073] Records used for repeated shrinkage screening should have the corresponding variables included in the set of variables to be screened for shrinkage and complete survival outcome data; records used for subsequent continuous data interval analysis should have valid values of the corresponding candidate prognostic variables, survival time, and all-cause mortality results.
[0074] Therefore, the conditions for a record to participate in the corresponding statistical processing are determined separately according to different processing tasks, rather than directly excluding the entire object record due to the absence of a single field that is unrelated to the current processing task.
[0075] Based on the field integrity check results, data availability tags are generated for each object data record. The data availability tags include at least statistical correlation availability tags, shrinkage filter availability tags, and interval analysis availability tags.
[0076] When an object data record meets the necessary fields required by the corresponding processing stage, the corresponding data availability flag is set to a valid state; if the corresponding necessary fields are missing, the corresponding data availability flag is set to an invalid state. An object data record can have multiple valid data availability flags simultaneously to support the retrieval of the same object record in different subsequent processing stages.
[0077] After determining the data availability status, a unified structure is processed for valid fields. The following fields are set according to a preset field order: object number, age, gender, various inflammatory markers, various nutritional markers, follow-up time, all-cause mortality result, and data availability marker. The data types and value representation methods for similar fields are also standardized. Continuous variables retain their numerical representation, categorical variables are converted to preset category identifiers, follow-up times are standardized to the same time unit, and all-cause mortality results are recorded using preset outcome status identifiers.
[0078] For records where numerical anomalies occur before or after field format conversion, categories cannot be matched, or follow-up time does not match the outcome status, a field anomaly flag is generated, and the record is suspended from entering the corresponding subsequent statistical processing.
[0079] After object attribution verification, field integrity detection, data availability tag generation, and unified structure processing, the object data records that have passed the corresponding validity check are organized according to their internal object numbers to form a stroke prognostic statistics set. Each object data record in the stroke prognostic statistics set maintains an object-level correspondence between demographic variables, inflammatory marker variables, nutritional marker variables, and survival outcome data. The data availability tag limits the subsequent processing stages it can participate in, providing a unified data input for the next implementation to perform statistical association analysis, variable correlation analysis, and duplicate shrinkage screening.
[0080] The stable screening of candidate prognostic variables includes the following specific implementation details:
[0081] This implementation takes the stroke prognostic statistical data set generated in the previous implementation as input, and sequentially performs survival risk statistical analysis, variable correlation analysis and repeated shrinkage screening on variables with available labels for statistical association analysis. The variable retention stability is determined based on the repeated retention of variables in different screening data subsets, so as to reduce the random impact of single sample division and the correlation between variables on the screening results, and form a set of candidate prognostic variables for subsequent nonlinear relationship identification.
[0082] First, data records of objects with available markers for statistical association analysis were read from the stroke prognostic statistics dataset. Demographic variables, inflammatory marker variables, and nutritional marker variables were used as variables to be analyzed, and the follow-up time and all-cause mortality results of the corresponding subjects were used as survival outcome inputs. Survival risk statistical analysis was performed on each variable to be analyzed. The survival risk statistical analysis adopted the univariate Cox proportional hazards method to obtain the risk association direction, risk association strength, and statistical significance results of each variable. For continuous variables, the original continuous values were used for statistical processing; for categorical variables, risk association relationships between categories were established with a pre-determined reference category.
[0083] The univariate Cox proportional hazards method outputs the hazard ratio, 95 percent confidence interval and significance probability for each variable to be analyzed, and uses a significance probability of less than 0.05 as the criterion for satisfying the statistical association condition. For subjects who did not have all-cause mortality at the end of the follow-up, their outcome status is recorded as censored, and the actual follow-up time is retained for survival risk statistics.
[0084] Before performing statistical association screening, a unified statistical significance criterion is pre-defined and remains unchanged throughout the variable screening process. When the statistical association between the variable to be analyzed and the survival outcome meets the criterion, the variable is included in the first candidate variable set; variables that do not meet the criterion are not included in subsequent narrowing screenings. This process first excludes variables that lack a valid statistical association with the survival outcome, avoiding the possibility of all original indicators directly participating in repeated narrowing screenings and causing an excessively large variable search range.
[0085] After obtaining the first set of candidate variables, a correlation analysis is performed on the continuous candidate variables. Specifically, Spearman's rank correlation method is used to obtain the degree and direction of correlation between any two candidate variables. Variables whose correlation reaches the preset correlation criteria are set into the same correlation variable group. The correlation criteria are determined before the start of the correlation analysis based on the number of candidate variables and the allowable degree of variable redundancy, and remain consistent throughout the same screening process. The correlation variable group is only used to record possible duplicate representations between variables; no variable is directly deleted in this step. This allows subsequent shrinking screening to jointly determine the actual retention of correlation variables based on survival outcomes. The first set of candidate variables and the corresponding correlation variable groups are then organized together into the variable set to be shrinked for screening.
[0086] Subsequently, the data records of the objects with available shrinkage screening markers are repeatedly sampled. Each time, a screening subset is formed from the original stroke prognostic statistics set. To avoid significant deviations in the proportion of all-cause mortality records and records without all-cause mortality at the end of follow-up during the repeated sampling process, the object records can be extracted separately according to the survival outcome status and then merged to form the current screening subset. This ensures that each screening subset maintains an outcome structure that can perform survival risk shrinkage analysis. The number of repeated samplings is predetermined before the screening begins, for example, set to 100 to 500 times. All subsequent candidate variables are then subjected to the same number of times for stability statistics.
[0087] For each selected data subset, the LASSO Cox shrinkage screening method is used to simultaneously process the set of variables to be shrunk. Each shrinkage screening redetermines the penalty parameter corresponding to the current selected data subset, instead of directly using the penalty parameters of other selected data subsets. The penalty parameter is determined through cross-validation. Specifically, the current selected data subset can be divided into ten mutually exclusive data parts. One data part is selected sequentially as validation data, and the remaining data parts are used as parameter estimation data. After all rounds are completed, the target penalty parameter used in this shrinkage screening is determined based on the cross-validation errors corresponding to different candidate penalty parameters. When the validation results of multiple candidate penalty parameters meet the same selection criteria, the candidate penalty parameter that produces fewer retained variables is prioritized to reduce redundant variables entering the subsequent processing.
[0088] After shrinking the current subset of data using the defined target penalty parameters, the retention status of each variable to be shrunk is recorded. If the shrinkage coefficient of a variable is not compressed to zero, the variable is recorded as retained in this screening; if the shrinkage coefficient of a variable is compressed to zero, it is recorded as not retained in this screening. Simultaneously, the risk action direction of the retained variables is recorded to determine whether the statistical action of the same variable remains consistent across different subsets of data.
[0089] After processing all selected data subsets, the number of times each variable to be narrowed down is retained is accumulated, and the proportion of retained times relative to all repeated screening times is used as the retention frequency of that variable. Further comparison is made of the risk action direction corresponding to each retention of the variable, determining the proportion of screening times with consistent risk action direction to all retained times. This proportion is used as the retention consistency criterion. The retention frequency and retention consistency criterion are combined to determine the variable retention stability, avoiding the need to determine whether a variable should enter subsequent analysis based solely on a single narrowing result. In specific judgments, retention frequency thresholds and result consistency thresholds can be set in advance. For example, when a variable is retained in at least 70% of the repeated screening processes, and its risk action direction remains consistent in at least 80% of the retained results, the variable is determined as a stable retained variable. If either condition does not meet the corresponding threshold, the variable is not considered a stable retained variable in this round. The above thresholds are determined before screening begins based on the sample size and the number of repeated screenings, and a uniform judgment condition is applied to all variables to be narrowed down.
[0090] Finally, variables whose statistical association results meet the requirements and whose retention stability meets the preset retention conditions are aggregated, retaining their variable identifiers, variable types, statistical association directions, related variable group labels, frequency of repeated screening and retention, and consistency of retention results, forming a candidate prognostic variable set. This candidate prognostic variable set only indicates that the variables have passed the statistical association and repeated stability screening, and does not predefine that the variables have a linear or non-linear relationship. It serves as the direct input for the next implementation to perform continuous data interval division and interval risk change analysis.
[0091] The determination of nonlinear candidate variables and candidate sensitive intervals includes the following specific implementation details:
[0092] This implementation uses the set of candidate prognostic variables formed in the previous implementation as the processing object. It determines whether the variable relationship has non-linear characteristics by the risk association changes between the value range of the candidate prognostic variables and the survival outcome, and further limits the initial range of risk changes. This provides a search boundary for the next implementation to perform progressive refinement of the candidate sensitive interval. This implementation does not directly perform restricted cubic spline fitting, nor does it pre-specify that a certain candidate prognostic variable must have a non-linear relationship. Instead, it identifies the variable based on the risk change results formed between continuous value ranges.
[0093] For each continuous candidate prognostic variable in the candidate prognostic variable set, the object data records with corresponding interval analysis available markers are read and arranged from low to high according to the variable value of the candidate prognostic variable. During the arrangement process, the object data records corresponding to the same variable value are kept continuous and are not split, so as to avoid creating artificial boundaries after the same variable value is assigned to different continuous data intervals;
[0094] After sorting the variable values, determine the initial number of intervals based on the number of valid object records. The initial number of intervals should ensure that each continuous data interval has a sufficient number of samples and survival outcome records to perform survival risk statistical processing. Specifically, it can be determined in advance based on the number of valid samples for the current candidate prognostic variable, the number of all-cause mortality records, and the preset minimum interval sample requirement. Following the principle that each interval contains approximately the same number of object records, the interval boundaries are determined sequentially from the lower end to the higher end of the variable value range. When multiple identical variable values exist at the predetermined boundary position, the boundary is moved to the outside of the records with the same variable value, so that object records with the same variable value are classified into the same continuous data interval. Differences in the number of samples between intervals resulting from this are allowed to be retained.
[0095] When the number of valid objects or all-cause mortality records in an initial continuous data interval is lower than the minimum number required to perform interval risk statistics, the risk association results for that interval are not generated directly. Instead, the interval is merged with the continuous data intervals adjacent to the variable values, and the validity of the data in the merged interval is re-examined.
[0096] After processing, each continuous data interval is assigned a serial number according to the variable value from low to high, and the corresponding lower boundary of the variable value, upper boundary of the variable value, number of valid objects, number of survival outcomes, and interval grouping identifier are recorded.
[0097] For each defined continuous data interval, the interval grouping identifier is used as the grouping representation of the current candidate prognostic variable. The survival time and all-cause mortality results corresponding to stroke patients are used as outcome inputs. The Cox proportional hazards processing method, consistent with the aforementioned survival risk statistical analysis, is used to determine the interval risk association results corresponding to each continuous data interval. When performing interval risk analysis on the same candidate prognostic variable, a unified reference method and unified statistical processing conditions are used for each continuous data interval to ensure the comparability of risk association results obtained from different continuous data intervals. The interval risk association results at least record the risk association strength and the direction of risk action corresponding to the interval.
[0098] In the interval risk handling of the same candidate prognostic variable, the continuous data interval with the lowest variable value is used as a unified reference interval. The interval risk association strength is determined according to the risk ratio of each of the remaining continuous data intervals relative to the reference interval, and the reference interval remains unchanged in the process of comparing all intervals of the current variable.
[0099] The risk association results for each consecutive interval are arranged according to their sequence numbers, forming a sequence of interval risk changes corresponding to the current candidate prognostic variable. For any two adjacent consecutive data intervals in the sequence, the risk association strength of the latter consecutive data interval is compared with that of the former consecutive data interval to determine the direction and magnitude of the risk change in the current adjacent interval. Specifically, if the risk association strength of the latter consecutive data interval is higher than that of the former consecutive data interval and the difference reaches the minimum effective change requirement, it is determined as an increase in risk; if it is lower than that of the former consecutive data interval and the difference reaches the minimum effective change requirement, it is determined as a decrease in risk; if the difference between the two does not reach the minimum effective change requirement, it is determined as a leveling-off risk.
[0100] The minimum effective change requirement is used to distinguish between actual risk changes and interval statistical fluctuations. It is determined before the interval analysis of the current candidate prognostic variable begins. Specifically, it can be determined based on the distribution range of the risk association strength difference of all adjacent intervals, the number of effective samples in the current interval, and the confidence range of the risk association results. It remains unchanged during the comparison of all adjacent intervals of the same candidate prognostic variable, thereby avoiding temporary changes to the judgment conditions based on the change results of a certain local interval.
[0101] Further analysis of the direction and magnitude of risk changes, following a sequential order of continuous intervals, includes:
[0102] When the previous consecutive interval showed an increase in risk, and the subsequent consecutive interval turned to a decrease in risk, or from a decrease in risk to an increase in risk, a change in the direction of risk change is identified;
[0103] When multiple consecutive adjacent intervals maintain the same direction of risk change, but the corresponding risk change magnitude continuously expands or continuously shrinks, it is determined that a change in the rate of risk change has occurred.
[0104] When multiple consecutive adjacent intervals initially maintain an increase or decrease in risk, and the risk change amplitude of subsequent adjacent intervals is continuously lower than the minimum effective change requirement, it is determined that the risk change has changed from continuous change to a flat state. The number of consecutive judgments corresponding to the multiple consecutive adjacent intervals is preset before the interval analysis begins, preferably set to no less than two consecutive comparison positions, in order to reduce nonlinear misjudgments caused by fluctuations in a single adjacent interval.
[0105] When a candidate prognostic variable exhibits any of the above-mentioned risk change characteristics, the candidate prognostic variable is marked as a nonlinear candidate variable, and the location of the risk change that triggers the nonlinear determination is determined. For the risk change direction transformation, the continuous data interval involved by the two risk change directions before and after the transformation is taken as the core change interval.
[0106] For continuous changes in the magnitude of risk, the core change location is the position where the magnitude of change begins to show a continuous increasing or decreasing trend; for changes in risk that transition from continuous change to a flat state, the core change location is the position between the last continuous change interval and the beginning interval of the flat state.
[0107] After the core change location is determined, it is not directly used as the final nonlinear turning point. Instead, continuous data intervals adjacent to the core change location are reserved on the side with lower variable values and the side with higher variable values, respectively. The range of variable values covered by the interval corresponding to the core change location and its adjacent intervals before and after it is determined as the candidate sensitive interval.
[0108] When the same nonlinear candidate variable has multiple adjacent or partially overlapping core change positions, its corresponding initial change ranges are continuously merged to form a candidate sensitive interval.
[0109] When there is at least one stable continuous interval that does not trigger nonlinear change determination between different core change locations, it is retained as an independent candidate sensitive interval and enters the subsequent progressive refinement process respectively.
[0110] After the above processing, a nonlinear candidate analysis record is generated for each nonlinear candidate variable, including variable identifier, initial continuous data interval, interval risk change sequence, risk change type, core change location, and candidate sensitive interval. Candidate prognostic variables that do not trigger nonlinear change judgment continue to retain their statistical correlation results, but do not enter the candidate sensitive interval progressive refinement process. Variables that trigger nonlinear change judgment and their candidate sensitive intervals are used as inputs for the next implementation content to further narrow the range of risk changes and determine the nonlinear turning point and nonlinear sensitive interval.
[0111] The implementation details include: progressive refinement of the sensitive interval and nonlinear fitting.
[0112] This implementation takes the nonlinear candidate variables and corresponding candidate sensitive intervals determined in the previous implementation as the processing objects. Without changing the data organization relationship of the external variables of the candidate sensitive intervals, it progressively refines the data range within the candidate sensitive intervals. By continuously comparing the risk change positions obtained in different refinement rounds, it determines the nonlinear turning points and nonlinear sensitive intervals that can stably represent the risk change characteristics. Furthermore, based on the nonlinear turning points and nonlinear sensitive intervals, it configures a nonlinear fitting structure so that the subsequent fitting process can focus on representing the local areas where risk changes are concentrated and generate nonlinear relationship representation data.
[0113] For each nonlinear candidate variable, first, read the lower and upper boundaries of its candidate sensitivity interval, and extract the object data records whose variable values fall within the candidate sensitivity interval and are available for interval analysis. Based on the number of valid objects and all-cause mortality records contained in the candidate sensitivity interval, determine the number of refined data segments for this round. The number of refined data segments should ensure that each segment has sufficient valid data for performing risk association statistics, while also ensuring that the width of the data segments in this round is smaller than the width of the initial continuous data interval used when forming the candidate sensitivity interval.
[0114] In the first round of refinement, the candidate sensitive intervals are divided into multiple consecutive refined data segments according to the order of variable values;
[0115] For candidate sensitive intervals where the variable values are relatively evenly distributed, they can be divided according to the span of similar variables;
[0116] For candidate sensitive intervals where variable values are concentrated or where there are a large number of duplicate values, the boundaries should be determined first by making the number of valid objects contained in each refined data segment close, and the object data records corresponding to the same variable value should be kept in the same refined data.
[0117] If a refined data segment obtained from the division cannot meet the preset minimum number of valid objects or the minimum number of outcome events, then the refined data segment will be merged with the adjacent data segment in the risk analysis direction until the risk correlation analysis requirements of the current round are met.
[0118] For each refined data segment, the statistical processing conditions of the aforementioned interval risk association results will continue to be used, and the refined risk association results will be determined based on the survival time and all-cause mortality results within the corresponding data segment.
[0119] It should be noted that, since the purpose of this implementation is to compare the risk changes at different local locations within the candidate sensitive interval, the same reference conditions are used for each refined data segment in the same round, and the refined risk association results are obtained by arranging the variable values from low to high.
[0120] The risk correlation strength of any two adjacent refined data segments is compared sequentially to determine the direction and magnitude of risk changes at adjacent positions, and then the changes at consecutive positions are compared. When the risk change of adjacent segments changes from rising to falling, or from falling to rising, or when the magnitude of risk change changes from continuously increasing to continuously decreasing, or from continuously decreasing to leveling off during continuous changes in the same direction, the position between the corresponding two refined data segments is determined as the position of the risk change in this round.
[0121] If multiple risk change locations appear in the same round, record each risk change location and its corresponding change type separately, and prioritize retaining the locations that are consistent with the risk change type identified in the previous implementation content for the next round of refinement;
[0122] For multiple risk change locations that are adjacent to each other and correspond to the same change trend, the data segment range covered by them is treated as a unified local change area.
[0123] After identifying the location of risk changes in this round, the next round of refinement interval is determined by the adjacent data segments and the data segments before and after the location of the risk change in this round. The range of variable values in the next round of refinement interval should be smaller than the processing interval corresponding to the current round, thereby achieving the gradual shrinking of the search range for risk change locations.
[0124] For risk changes near the current candidate sensitive interval boundary, only the actual adjacent data segments are retained, and no data range without corresponding risk change basis is artificially added outside the candidate sensitive interval.
[0125] After entering the next round of refinement, the width of the data segment division is reduced, and the risk change position of the current round is re-determined according to the same refinement risk association processing method. The degree of reduction in the width of the data segment is determined by the variable span of the current interval to be refined and the number of available objects. In principle, the width of the data segment in the next round should be smaller than that in the previous round, while ensuring that each data segment has the amount of data that meets the risk statistics requirements.
[0126] When further reducing the data segment width would result in an insufficient number of valid objects in most refined data segments, the segment width will not be further reduced. Instead, the current data granularity that can maintain effective risk statistics will be used as the minimum refinement granularity.
[0127] Starting from the second round of refinement, the location of risk change in the current round is compared with the location of risk change in the previous round. The location comparison is performed using the variable value scale of the nonlinear candidate variable itself, and the distance between the variable values of the two rounds of risk change locations is used as the location offset. The preset location tolerance is determined based on the detection accuracy of the nonlinear candidate variable, the initial variable span, and the minimum refinement granularity. The location tolerance remains unchanged after entering the continuous stability determination.
[0128] When the positional offset between the risk change positions obtained in consecutive preset rounds does not exceed the preset positional tolerance, and the corresponding risk change types remain consistent, it is determined that the risk change position has reached a stable state.
[0129] The consecutive preset rounds are used to exclude premature termination caused by accidental occurrence of the same position in a single refinement, and can be set to two or more consecutive rounds.
[0130] If the current risk change position still exceeds the position tolerance relative to the previous round, then the range to be refined will be re-formed based on the current risk change position and the next round of refinement will continue.
[0131] If the minimum granularity has been reached but the location of the risk change still does not meet the stability condition, the current granularity is retained and a stability check mark is generated. The subsequent resampling process will then further determine whether the local risk change is repeatable.
[0132] When the risk change position reaches a stable condition, the concentrated position covered by the risk change position corresponding to the continuous stable rounds is determined as the nonlinear turning point position, and the range of variable values corresponding to the continuous data segment covering the nonlinear turning point position and the adjacent data segment directly involved in the turning point determination when the refinement stops is determined as the nonlinear sensitive interval.
[0133] After determining the nonlinear inflection points and nonlinear sensitivity intervals, a nonlinear fitting structure is configured for the nonlinear candidate variables, including:
[0134] First, the candidate positions of the fitting nodes are set according to the overall effective range of the variable, and the node distribution is adjusted according to the nonlinear sensitive interval. In particular, relatively dense candidate positions of fitting nodes are set inside the nonlinear sensitive interval so that the fitting structure can provide a more detailed representation of risk changes near the turning point.
[0135] Outside the nonlinear sensitive region, the distribution density of fitting nodes is reduced according to the range of variable values and the stability of risk changes, so as to avoid configuring too many nodes in the region with weak risk changes, which would cause local fluctuations to be overfitted.
[0136] Candidate fitting node positions should not all be concentrated on one side of the nonlinear inflection point. Node positions that can constrain the overall trend of change should be reserved before, inside, and after the nonlinear sensitive interval. Based on this, multiple candidate fitting node combinations are formed by changing the number of fitting nodes, node spacing, and the positions of nodes adjacent to the inflection point within the nonlinear sensitive interval. Each candidate fitting node combination uses the same input object record and the same survival outcome data to ensure that different fitting structures can be directly compared.
[0137] Candidate fitting node combinations should retain fitting nodes at least before, within, and after the nonlinear sensitive interval, with at least two candidate nodes configured within the nonlinear sensitive interval. Cox proportional hazards model parameters should be estimated for each candidate fitting node combination using the same modeling data, and the Harrell consistency index should be used to evaluate the model's discriminative ability. Candidate fitting node combinations with higher consistency indices should be retained first.
[0138] For each candidate fitting node combination, a restricted cubic spline fitting is performed on the current nonlinear candidate variable, and the resulting nonlinear fitting term is incorporated into the survival risk statistical model along with other candidate prognostic variables to form the risk change relationship of the corresponding candidate fitting node combination. After the fitting is completed, the fitting structure is not directly selected based on the evaluation result of a single model, but rather the risk change relationship obtained by fitting is first verified to maintain the actual change characteristics revealed by the aforementioned interval analysis results.
[0139] Specifically, the direction of change of the fitted risk change relationship before and after the nonlinear turning point is extracted, and it is compared with the interval risk change sequence formed by the third implementation and the refined risk change result formed by this implementation. When the direction of change of the fitted relationship before and after the turning point is significantly opposite to the local risk analysis result, the corresponding candidate fitted node combination is excluded. When the fitted relationship can maintain the identified risk change direction, the position where the change trend in the fitted relationship changes significantly is determined, and it is judged whether the position falls into the determined nonlinear sensitive interval. When the fitted change position is outside the nonlinear sensitive interval, it indicates that the position offset generated by the current node configuration exceeds the support range of the local statistical results, and the corresponding candidate fitted node combination is not used as the target structure.
[0140] The model discrimination ability of candidate fitting node combinations that pass the above-mentioned risk direction consistency and turning point range verification is evaluated. Under the same object data and the same other prognostic variables, the discrimination degree of each candidate fitting structure corresponding to the model for different survival outcomes is compared, and the target fitting node combination is determined according to the preset model selection rules.
[0141] When the discriminative power of multiple candidate fitting node combinations meets the same selection range, priority is given to selecting candidate fitting node combinations with fewer fitting nodes and smaller deviations between the turning points and the progressive refinement results, in order to reduce the complexity of the nonlinear structure.
[0142] Once the target fitting node combination is determined, its corresponding nonlinear candidate variable identifier, nonlinear inflection point, nonlinear sensitive interval, position of each fitting node, risk change direction before and after the inflection point, and the resulting risk change relationship are organized according to a unified data structure to generate nonlinear relationship representation data corresponding to the nonlinear candidate variable.
[0143] It should be noted that the target fitting structure determined in this implementation is still a nonlinear relationship result obtained based on the current set of stroke prognosis statistics. It is not directly identified as the final stable relationship in this stage. The nonlinear relationship characterization data formed, as well as the corresponding nonlinear inflection points and nonlinear sensitive intervals, are used as the benchmark data for the resampling stability verification in the next implementation. Subsequent processing will further determine whether the identified nonlinear relationship remains stable under the sample reconstruction conditions.
[0144] The resampling stability check and result generation process includes the following specific implementation details:
[0145] This implementation uses the nonlinear relationship characterization data obtained in the previous implementation as a benchmark, and performs resampling stability verification on the nonlinear inflection points and nonlinear sensitive intervals recorded therein. When the nonlinear inflection points or nonlinear sensitive intervals obtained under different resampling data conditions shift significantly, the candidate sensitive intervals are corrected according to the shift direction and local refinement and nonlinear fitting are re-executed.
[0146] When the stability meets the preset requirements, the target nonlinear relationship characterization data is determined, and the stroke prognosis statistical processing results are generated based on the stable nonlinear relationship. Specifically, multiple resampling is performed from the stroke prognosis statistical data set with replacement. The number of object records obtained in each resampling is consistent with the number of object records participating in the current nonlinear analysis in the original stroke prognosis statistical data set, thereby forming multiple resampled data sets.
[0147] The same stroke subject can appear repeatedly in a single resampled dataset or may not be sampled in a partial resampled dataset. The number of resampling times is predetermined before the stability check begins and remains unchanged throughout the entire stability check process of the current nonlinear candidate variable. In practice, the number of resampling times can be determined based on the number of valid subjects participating in the nonlinear analysis. When the number of valid subjects can meet the requirements of repeated statistics, no less than 500 resampling times can be set to reduce the impact of a small number of random sampling results on the stability determination.
[0148] For each resampled dataset, the data processing rules, risk change judgment conditions, and fitting structure selection rules determined in the previous implementation remain unchanged. The risk change position is re-determined for the current nonlinear candidate variable, and the corresponding resampled nonlinear inflection position and resampled nonlinear sensitive interval are obtained. During the resampling process, the original nonlinear inflection position is not directly copied as a fixed node, but it is allowed to be repositioned according to the risk change results in the current resampled data. This allows us to test whether the original nonlinear inflection position depends on a specific sample combination.
[0149] After all resampling is completed, the nonlinear inflection point obtained from the original stroke prognosis statistical data set is used as the reference point. The position offset of each resampled nonlinear inflection point relative to the reference point is determined. The position offset is expressed using the variable value scale of the corresponding nonlinear candidate variable itself, so that different resampling results can be compared in a unified variable space.
[0150] The preset position tolerance is determined based on the minimum refinement granularity adopted in the previous implementation. The position tolerance is not less than the minimum data span that the current variable can be effectively distinguished, and is not greater than the range of the nonlinear sensitive interval when progressive refinement stops. When the position offset between a certain resampled nonlinear turning point and the reference position does not exceed the preset position tolerance, the resample is recorded as a turning point stable. The proportion of all turning point stable resamples to all effective resamples is calculated to obtain the turning point stable ratio.
[0151] In addition to the turning points, the range consistency between each resampled nonlinear sensitive interval and the original nonlinear sensitive interval is compared. For each resampling result, the range of variable values jointly covered by the resampled nonlinear sensitive interval and the original nonlinear sensitive interval, as well as the total coverage range after merging the two, are determined. The interval coverage consistency is determined according to the degree of the common coverage range to the total coverage range. The higher the interval coverage consistency, the closer the sensitive interval obtained after resampling is to the original sensitive interval.
[0152] When the two do not share a common coverage area, the resampling is determined to be an interval coverage inconsistency.
[0153] The location stability threshold and coverage consistency threshold are determined before the stability check begins. The location stability threshold limits the proportion of resampling inflection points that should remain near the baseline position, while the coverage consistency threshold limits the minimum level of coverage consistency required between the sensitive interval of a single resampling and the original sensitive interval. A stability ratio is further set to limit the minimum percentage of resampling results that meet the coverage consistency threshold among all valid resampling results. Once these criteria are determined, they will not be adjusted until the current variable completes this round of stability check.
[0154] When the stability ratio of the turning point reaches the position stability threshold, and the number of resampled nonlinear sensitive intervals that reach the coverage consistency threshold reaches the stability ratio, the current nonlinear relationship is determined to have passed the stability check. At this time, the nonlinear relationship representation data formed by the previous implementation is determined as the target nonlinear relationship representation data, and the corresponding nonlinear turning point, nonlinear sensitive interval, target fitting node combination and risk change relationship are retained.
[0155] When any stability condition is not met, the current nonlinear candidate variable is not directly discarded. Instead, the offset distribution of the resampled nonlinear inflection point relative to the reference position is analyzed. Specifically, the position offset is divided into offset to the side with lower variable value and offset to the side with higher variable value, and the number of resampled results corresponding to the two directions is counted respectively.
[0156] When the resampling nonlinear turning points that reach the preset offset ratio are concentrated on the same side of the candidate sensitive interval, it indicates that the current candidate sensitive interval may not fully cover the risk change area in that direction. In this case, keep the other side boundary of the candidate sensitive interval unchanged, and expand the corresponding side boundary along the concentrated offset direction. The expansion range is determined according to the concentrated distribution range of the resampling turning points in that direction, and prioritizes covering the resampling turning points that reach the preset offset ratio, rather than directly expanding to the entire value range of the current variable.
[0157] When the resampled nonlinear turning points are distributed to both sides of the candidate sensitive interval, and the offset on both sides reaches the corresponding preset proportion, it indicates that the overall range of the original candidate sensitive interval may be too narrow. At this time, the boundaries on both sides of the candidate sensitive interval are adjusted according to the distribution range of the resampled turning points on the low-value side and the high-value side, so that the adjusted candidate sensitive interval can cover the turning points on both sides with repeated occurrence characteristics.
[0158] After the candidate sensitive interval boundary is adjusted, only the progressive refinement and nonlinear fitting are re-executed on the adjusted local data range, and the resampling stability check is performed again. The aforementioned candidate prognostic variable screening process is not re-executed.
[0159] It should be noted that, to prevent repeated expansion of the candidate sensitive interval due to the absence of stable nonlinear transitions in the sample, a maximum number of corrections is preset. The maximum number of corrections is determined based on the overall value range of the nonlinear candidate variable, the initial candidate sensitive interval range, and the allowable expansion degree. If the stability check passes before reaching the maximum number of corrections, the result of the last successful check is used as the target nonlinear relationship representation data. If the stability condition is still not met after reaching the maximum number of corrections, an unstable nonlinear relationship marker is generated for the nonlinear candidate variable, and the generation of stable nonlinear risk contribution representations based on the variable is stopped, in order to avoid outputting non-repeatable local risk changes as deterministic relationships.
[0160] For nonlinear candidate variables that have passed the stability check, the data interval of the variable value corresponding to the stroke subject to be treated is determined according to the target nonlinear relationship characterization data, and the risk change relationship corresponding to the data interval is read to generate the corresponding nonlinear risk contribution characterization. In this embodiment, the prognostic variables formed after the candidate variable stability screening include age, gender, systemic inflammatory response index, body mass index and albumin. Among them, albumin is processed according to the determined target nonlinear relationship, and age, gender, systemic inflammatory response index and body mass index participate in the subsequent joint statistics according to the corresponding survival risk relationship.
[0161] Specifically, the albumin value of the stroke subject to be treated is substituted into the restricted cubic spline structure corresponding to the target fitting node combination to obtain the corresponding spline transformation result. Then, the model parameters corresponding to each spline transformation term in the survival risk statistical model are combined, and the result is used as the nonlinear risk contribution characterization of the albumin variable of the stroke subject.
[0162] The nonlinear risk contribution of age, sex, systemic inflammatory response index, body mass index, and albumin of the stroke subjects to be treated is input into the survival risk statistical model that has completed the target nonlinear structure configuration to obtain the stroke prognostic statistics corresponding to the stroke subjects. The stroke prognostic statistics represent the relative prognostic risk level of multiple candidate prognostic variables under the current survival risk statistical model, and maintain the correspondence with the target nonlinear relationship representation data.
[0163] Based on the correspondence between follow-up time and cumulative risk in the survival risk statistical model, the stroke prognostic statistics were mapped to the one-year, three-year and five-year follow-up time points to obtain the all-cause mortality risk statistics at the corresponding time points. The same nonlinear relationship was used to represent the data at each follow-up time point, and the albumin nonlinear inflection point was not re-determined for different follow-up time points, so as to ensure that the risk statistical results of the same stroke subject at different follow-up time points have a consistent variable processing basis.
[0164] After the survival risk statistical model completes parameter estimation, it simultaneously determines the baseline survival relationship corresponding to the modeling sample and extracts the baseline survival values corresponding to the one-year, three-year and five-year follow-up time points. The stroke prognostic statistics of the stroke subjects to be treated are combined with the baseline survival values of the corresponding follow-up time points to obtain the all-cause mortality risk statistics for one year, three years and five years respectively.
[0165] When classifying risk levels, the stroke prognostic statistics of stroke subjects used to establish the current survival risk statistical model are obtained and arranged from low to high. The risk stratification boundary is determined by the median position of the stroke prognostic statistics. When the stroke prognostic statistics of the stroke subject to be treated reach or exceed the risk stratification boundary, it is determined as a high-risk level; when it is below the risk stratification boundary, it is determined as a low-risk level. The risk stratification boundary is determined by the statistical value distribution of the current modeling data and remains fixed under the same model version, without being recalculated based on a single stroke subject to be treated.
[0166] Finally, the stroke subject identifier, stroke prognostic statistics, one-year all-cause mortality risk statistics, three-year all-cause mortality risk statistics, five-year all-cause mortality risk statistics, risk level, nonlinear candidate variable identifier, nonlinear inflection point, nonlinear sensitive interval, and nonlinear risk contribution characterization are associated according to the subject identifier to form the stroke prognostic statistical processing results.
[0167] Through the above processing, nonlinear relationships are only included in the final prognostic statistics when the turning point and sensitive interval remain stable under different sample reconstruction conditions. At the same time, when the stability is insufficient, the candidate sensitive interval is adjusted in the opposite direction of the turning point offset, so as to realize continuous data processing between nonlinear relationship discovery, local correction, stability confirmation and result output.
[0168] A specific example of progressive refinement of the albumin nonlinear sensitive region is as follows:
[0169] This embodiment uses albumin in the candidate prognostic variable set as the object of nonlinear relationship processing to illustrate the process of progressive refinement of candidate sensitive intervals and determination of nonlinear fitting structure.
[0170] Extract object data records with valid albumin values, survival time, and all-cause mortality results. Arrange them in order of albumin value from low to high. Under the condition that each interval has a number of valid objects and outcome events that meet the risk statistics requirements, divide the data into multiple initial continuous data intervals. Determine the interval risk association results corresponding to each continuous data interval. Compare the risk change direction and risk change magnitude of adjacent intervals according to the albumin value order.
[0171] When albumin is in a low range, the risk association results of the interval show a continuous downward trend as albumin increases; as albumin continues to increase, the risk decrease between adjacent intervals gradually decreases and then enters a flat state in subsequent intervals. Based on the position where the risk decrease changes from continuous change to flat state, albumin is identified as a nonlinear candidate variable, and the variable range covered by the corresponding change interval and its adjacent intervals before and after it is identified as the candidate sensitive interval.
[0172] The object data is re-extracted from the candidate sensitive interval, and the data is divided into multiple refined data segments at a data granularity smaller than the initial interval. The refined risk association results corresponding to each refined data segment are determined, and the risk changes of adjacent refined data segments are compared. The position where the risk decrease rate changes significantly is determined as the risk change position in this round.
[0173] The data segment corresponding to the risk change position in this round and its adjacent data segments are used to form the next round of refinement intervals. The width of the data segment is further reduced, and the risk change position is determined again. The progressive refinement is performed continuously in the above manner, and the variable value offset between the risk change positions obtained in adjacent rounds is compared.
[0174] When the risk change position offset of consecutive preset rounds does not exceed the preset position tolerance, and the corresponding risk change type remains from a significant decrease to a gradual change, the progressive refinement stops, the stable risk change position is determined as the albumin nonlinear transition position, and the continuous data segment covering the nonlinear transition position when refinement stops is determined as the nonlinear sensitive interval.
[0175] After determining the nonlinear sensitive interval, candidate positions of fitting nodes are set according to the overall range of albumin values, so that the distribution density of fitting nodes inside the nonlinear sensitive interval is higher than that outside the region, and multiple combinations of candidate fitting nodes are formed by changing the number and position of nodes in the sensitive interval.
[0176] Restricted cubic spline fitting was performed on albumin according to each candidate fitting node combination. The resulting albumin nonlinear fitting results were then incorporated into the survival risk statistical model along with age, gender, systemic inflammatory response index, and body mass index. First, it was determined whether the direction of risk change obtained from the fitting was consistent with the risk change results of the aforementioned interval. Then, it was determined whether the position of the change obtained from the fitting was within the nonlinear sensitive interval.
[0177] For candidate fitting node combinations that simultaneously meet the requirements of consistent risk direction and range of change location, the discrimination ability of the corresponding survival risk statistical model is further compared, and the combination with fewer fitting nodes and fitting change location close to the nonlinear turning point is selected as the target fitting node combination.
[0178] Finally, the albumin variable identifier, nonlinear inflection points, nonlinear sensitive intervals, target fitting node positions, and corresponding risk change relationships are organized into nonlinear relationship characterization data, which are then used as input for subsequent resampling stability verification.
[0179] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.
[0180] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0181] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.
[0182] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
[0183] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for processing nonlinear relationship data of stroke prognosis, characterized in that: The specific steps include: S1. Obtain demographic data, inflammatory marker data, nutritional marker data, and survival outcome data of stroke subjects, associate them according to the subject identifier, and form a set of stroke prognostic statistics based on the completeness of the data; S2. Perform statistical association analysis, variable correlation analysis and repeated shrinkage screening on the statistical data set of stroke prognosis. Determine the stability of variable retention based on the retention of each variable in the repeated screening and generate a set of candidate prognostic variables. S3. Divide the candidate prognostic variables into continuous data intervals according to the order of variable values, determine the interval risk association results between each continuous data interval and the survival outcome data, and determine the nonlinear candidate variables and candidate sensitive intervals based on the direction and magnitude of change of the risk association results of adjacent intervals. S4. The candidate sensitive intervals are progressively refined. Based on the stability of the risk change position in the adjacent refinement results, the nonlinear turning point and nonlinear sensitive interval are determined. Based on this, the nonlinear fitting structure is configured to generate nonlinear relationship characterization data. S5. Resample the statistical data set of stroke prognosis, perform stability check on the nonlinear turning point and nonlinear sensitive interval corresponding to different resampled data, and adjust the candidate sensitive interval according to the offset of the turning point and refine it again. When the check passes, determine the target nonlinear relationship characterization data and generate the stroke prognosis statistical processing results.
2. The method for processing nonlinear relationship data of stroke prognosis according to claim 1, characterized in that: The set of statistical data on stroke prognosis specifically includes: We obtained age and sex as demographic data for stroke patients, neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, monocyte-to-lymphocyte ratio, lymphocyte-to-monocyte ratio, systemic immune inflammation index, systemic inflammatory response index, and systemic inflammatory aggregation index as inflammatory marker data, body mass index, waist circumference, albumin, hemoglobin, and total protein as nutritional marker data, and follow-up time and all-cause mortality results as survival outcome data. The population data, inflammatory marker data, nutritional marker data, and survival outcome data are correlated according to the stroke subject identifier. The completeness of the variables and survival outcome data required for statistical processing in each correlated record is checked. Based on the completeness, corresponding data availability tags are generated. The correlated records that meet the statistical processing conditions are organized according to a unified field structure to form a set of stroke prognostic statistical data.
3. The method for processing nonlinear relationship data of stroke prognosis according to claim 2, characterized in that: The generated candidate prognostic variable set specifically includes: Survival risk statistical analysis was performed on each variable and survival outcome data in the stroke prognosis statistical data set, and the first candidate variable set was formed based on the obtained statistical correlation results. Perform correlation analysis on the first candidate variable set, and mark the variables with correlation to form a set of variables to be narrowed down for screening; Based on the statistical data set of stroke prognosis, multiple screening data subsets are repeatedly formed according to a preset sampling method. Shrinkage screening is performed on each screening data subset, and the corresponding shrinkage screening parameters are determined through cross-validation. The retention results of each candidate variable in different screening data subsets are recorded. The retention stability of each candidate variable is determined based on its retention frequency and the consistency of retention results in the entire selected data subset. Variables whose retention stability meets the preset retention conditions are identified as candidate prognostic variables, thus forming a set of candidate prognostic variables.
4. The method for processing nonlinear relationship data of stroke prognosis according to claim 3, characterized in that: The determination of nonlinear candidate variables and candidate sensitive intervals specifically includes: For candidate prognostic variables, the corresponding stroke prognostic statistics are arranged in order of variable value from low to high, and the data are divided into multiple continuous data intervals in such a way that the number of samples contained in each interval is approximately the same. Continuous data intervals are used as interval grouping identifiers. Based on the survival time and all-cause mortality results of stroke subjects within each interval, the corresponding interval risk association results are determined, and an interval risk change sequence is formed according to the arrangement order of continuous data intervals. By comparing the risk correlation results of adjacent continuous data intervals in turn, the corresponding risk change direction and risk change magnitude are determined. When the risk change direction changes, the risk change magnitude increases or decreases continuously, or the risk change changes from continuous change to gradual change, the corresponding candidate prognostic variable is determined as a nonlinear candidate variable. The initial range of change is determined based on the continuous data interval where the risk change occurs and its adjacent intervals before and after it, and the initial range of change is determined as the candidate sensitive interval of the nonlinear candidate variable.
5. The method for processing nonlinear relationship data of stroke prognosis according to claim 4, characterized in that: The progressive refinement of the candidate sensitive regions specifically includes: The candidate sensitive interval is divided into multiple continuous refined data segments, and refined risk association results corresponding to each refined data segment are generated according to the method for determining the interval risk association results. The results of the risk association between adjacent data segments are compared according to the order of the detailed data segments. When the direction of risk change changes or the trend of the magnitude of adjacent risk change changes changes, the position between the corresponding data segments is determined as the position of risk change in this round. The data segment where the risk change location is located in this round and its adjacent data segments before and after it are used to form the next round of refinement intervals. The width of the data segment in the next round of refinement intervals is narrowed and the risk change location is re-determined. Compare the risk change positions obtained in the current round with those in the previous round. When the offset between the risk change positions in consecutive preset rounds does not exceed the preset position tolerance, stop the progressive refinement. The corresponding risk change position is determined as the non-linear turning point, and the continuous data segment covering the non-linear turning point when refinement stops is determined as the non-linear sensitive interval.
6. The method for processing nonlinear relationship data of stroke prognosis according to claim 5, characterized in that: The configuration of the nonlinear fitting structure and the generation of nonlinear relationship representation data specifically include: The candidate positions of fitting nodes are determined based on the nonlinear inflection points, nonlinear sensitive intervals, and the value range of candidate prognostic variables. The distribution density of fitting nodes within the nonlinear sensitive interval is higher than that outside the nonlinear sensitive interval, thus forming a combination of multiple candidate fitting nodes. Restricted cubic spline fitting is performed on the nonlinear candidate variables according to each candidate fitting node combination, and the obtained nonlinear fitting results are incorporated into the survival risk statistical model to obtain the risk change relationship corresponding to each candidate fitting node combination; Determine whether the direction of change of each risk change relationship before and after the nonlinear turning point is consistent with the risk change sequence of the interval, and determine whether the turning point obtained by fitting is located within the corresponding nonlinear sensitive interval. Evaluate the model discrimination ability of candidate fitting node combinations that meet the above conditions. Based on the evaluation results of the model's ability to distinguish, the target fitting node combination is determined, and the nonlinear turning points, nonlinear sensitive intervals, fitting node positions, and risk change relationships corresponding to the target fitting node combination are organized into nonlinear relationship characterization data.
7. The method for processing nonlinear relationship data of stroke prognosis according to claim 6, characterized in that: The execution stability check specifically includes: Multiple resampling with replacement was performed on the stroke prognosis statistics set to form multiple resampled data sets. Candidate sensitive interval refinement and nonlinear fitting were then performed again on each resampled data set to obtain the corresponding resampled nonlinear inflection points and resampled nonlinear sensitive intervals. The position offset of each resampled nonlinear transition position relative to the original nonlinear transition position is determined, and the proportion of resampling times where the position offset does not exceed the preset position tolerance is counted to obtain the stability ratio of the transition positions. The overlap length of the resampled nonlinear sensitive interval and the original nonlinear sensitive interval and the combined coverage length of the two are determined respectively. The corresponding interval coverage consistency is determined according to the ratio of the overlap length to the combined coverage length. When the stability ratio at the turning point reaches the preset stability threshold, and the number of resampled nonlinear sensitive intervals that meet the preset coverage consistency threshold reaches the preset stability ratio, the stability check is determined to be passed; otherwise, the stability check is determined to be failed.
8. The method for processing nonlinear relationship data of stroke prognosis according to claim 7, characterized in that: When the stability check fails, the candidate sensitivity interval is adjusted, specifically including: Obtain the offset direction of each resampled nonlinear transition position relative to the original nonlinear transition position, and collect the resampled nonlinear transition positions that have experienced positional offset according to the offset direction; When the resampling nonlinear transition positions that reach the preset ratio are concentrated on the same side of the candidate sensitive interval, the corresponding side boundary of the candidate sensitive interval is expanded along the corresponding offset direction. When the resampling nonlinear transition positions are distributed on both sides of the candidate sensitive interval and both reach the corresponding preset ratio, the side boundaries of the candidate sensitive interval are expanded at the same time. Based on the candidate sensitive interval after boundary adjustment, the progressive refinement, nonlinear fitting and stability check are re-executed until the stability check is passed or the preset maximum number of corrections is reached. When the stability check passes, the corresponding nonlinear relationship representation data is determined as the target nonlinear relationship representation data. If the check fails after reaching the preset maximum number of corrections, an unstable nonlinear relationship marker is generated for the corresponding nonlinear candidate variable.
9. The method for processing nonlinear relationship data of stroke prognosis according to claim 8, characterized in that: The process of determining the target nonlinear relationship characterization data and generating stroke prognostic statistical processing results specifically includes: The set of candidate prognostic variables includes age, sex, systemic inflammatory response index, body mass index, and albumin, among which albumin is identified as a nonlinear candidate variable. The data interval containing albumin data of the stroke subjects to be treated is determined based on the target nonlinear relationship characterization data, and the nonlinear risk contribution characterization of albumin is determined based on the risk change relationship corresponding to the data interval. The variable data corresponding to age, gender, systemic inflammatory response index and body mass index are combined with the nonlinear risk contribution characterization and incorporated into the survival risk statistical model. The survival risk statistical model is used to generate stroke prognostic statistics for stroke subjects to be treated. The prognostic statistics of stroke are correlated with the corresponding target nonlinear relationship characterization data to form a record of prognostic statistical data for the stroke subjects to be treated.
10. The method for processing nonlinear relationship data of stroke prognosis according to claim 9, characterized in that: The generation of stroke prognostic statistical processing results also includes: Based on the survival risk statistical model and stroke prognostic statistics, the all-cause mortality risk statistics for stroke patients to be treated at the one-year, three-year and five-year follow-up time points were determined respectively. Obtain the stroke prognostic statistics of each stroke object in the stroke prognostic statistics set, determine the risk stratification boundary with the median of the stroke prognostic statistics, compare the stroke prognostic statistics of the stroke object to be treated with the risk stratification boundary, and determine the corresponding high-risk level or low-risk level. The stroke patient identifier, stroke prognostic statistics, all-cause mortality risk statistics at each follow-up time point, risk level, nonlinear candidate variable identifier, nonlinear inflection point, nonlinear sensitive interval, and nonlinear risk contribution characterization are correlated to form and output the stroke prognostic statistical processing results.
