A method and system for cross-process scheduling of hematoma prediction tasks

Abstract

This invention provides a method and system for cross-process scheduling of hematoma prediction tasks. The method includes: reading the hematoma prediction task request corresponding to the object to be predicted, parsing it to form an original feature record, performing field integrity verification, numerical validity verification, numerical normalization, missing value completion, and feature alignment processing on the original feature record to generate a feature snapshot data object; writing the feature snapshot data object into a cross-process scheduling queue to generate a prediction task record to be executed, and distributing it to the target prediction process to call the hematoma prediction model to obtain an initial prediction result; associating and storing the initial prediction result with the corresponding feature snapshot data object to generate a historical prediction record, and calculating the prediction deviation value when receiving the result feedback record to generate a deviation record; and merging and sorting the deviation records according to their identity identifiers.

Description

Technical Field

[0001] This invention relates to the field of medical data processing and task scheduling technology, and in particular to a cross-process scheduling method and system for hematoma prediction task processing. Background Technology

[0002] Current hematoma risk prediction technologies typically involve inputting patients' basic indicators at the front end, followed by the server calling a prediction model to output the risk result. This approach can already collect data such as name, age, body mass index, fibrinogen, systolic blood pressure, albumin, and triglycerides, and provide a one-time hematoma probability value based on existing models, which has application value in basic screening and auxiliary judgment.

[0003] However, in practical applications, data collected from the front end often needs to be transmitted across processes before being sent to the prediction model. Different tasks are easily affected by the order of tasks, data encapsulation methods, and changes in process state during queuing, distribution, execution, and result return. If the input data lacks unified organization and fixed encapsulation before entering the model, problems such as inconsistent field order, inconsistent handling of missing fields, and unclear correspondence between task results can easily occur, thus affecting the stability and traceability of the prediction process. Therefore, this invention proposes a cross-process scheduling method and system for hematoma prediction task processing.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may contain prior art information that is not common knowledge to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a cross-process scheduling method and system for hematoma prediction tasks, thereby resolving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first part of this invention provides a method for processing hematoma prediction tasks via cross-process scheduling, comprising the following steps: S1. Read the hematoma prediction task request corresponding to the object to be predicted, parse it to form the original feature record, and perform field integrity verification, numerical legality verification, numerical normalization, missing value completion and feature alignment processing on the original feature record to generate a feature snapshot data object; S2. Write the feature snapshot data object into the cross-process scheduling queue, generate a record of prediction tasks to be executed, select the target prediction task record to be executed according to the preset scheduling rules and distribute it to the target prediction process, and call the hematoma prediction model to obtain the initial prediction result. S3. Associate and store the initial prediction results with the corresponding feature snapshot data objects to generate historical prediction records. Calculate the prediction deviation value and generate a deviation record when the corresponding result feedback record is received. S4. Merge and sort the deviation records according to their identity identifiers to construct deviation sequence data. Perform imitation learning processing based on the deviation sequence data to generate deviation correction parameters corresponding to the identity identifiers. S5. When a hematoma prediction task request for the corresponding object to be predicted is received again, the new feature snapshot data object and the corresponding deviation correction parameters are read to obtain a new initial prediction result. The prediction result is then corrected based on the deviation correction parameters, and the corrected prediction result is output as the final hematoma probability prediction result.

[0007] The second part of this invention provides a cross-process scheduling hematoma prediction task processing system, including: a raw feature processing module, a task scheduling and prediction execution module, a historical correlation and deviation generation module, a deviation learning and parameter generation module, and a prediction correction and result write-back module.

[0008] The beneficial effects of this invention are as follows: This invention processes the original feature records into feature snapshot data objects and combines them with a cross-process scheduling queue, a record of prediction tasks to be executed, and a distribution mechanism for the target prediction process. This achieves a stable connection between the prediction task and the model calling process, and can solve the problems of inconsistent field order, unclear task correspondence, and unstable processing links after cross-process transmission in the prior art. This improves the standardization and traceability of hematoma prediction task processing.

[0009] Before prediction, this invention performs field integrity verification, numerical validity verification, numerical normalization, missing value completion, and feature alignment on the original feature records. This ensures that the data structure entering the hematoma prediction model remains consistent, reducing prediction fluctuations caused by inconsistent input formats, inconsistent handling of missing fields, or differences in numerical scales. This improves the stability of the model calling process and the consistency of prediction results.

[0010] This invention associates and stores the initial prediction results with feature snapshot data objects to form historical prediction records. After receiving the result feedback record, it calculates the prediction deviation value and generates a deviation record. This enables the system to not only complete the current prediction, but also retain the complete data link required for subsequent corrections, thereby enhancing the backtracking capability of prediction results and the utilization efficiency of historical data.

[0011] This invention constructs deviation sequence data by merging and sorting deviation records according to identity identifiers and by time sequence, and generates deviation correction parameters based on imitation learning. This enables the system to form targeted correction criteria based on the historical prediction deviation change patterns of the same object to be predicted, thereby improving the relevance and individual adaptability of subsequent hematoma prediction results.

[0012] This invention improves upon receiving a hematoma prediction task request for the corresponding target object again, by reading a new feature snapshot data object and the corresponding deviation correction parameters, performing correction calculations on the new initial prediction result, and outputting the corrected prediction result. This enables the system to continuously optimize subsequent prediction results based on historical deviations, thereby overcoming the shortcomings of existing technologies that can only output one-time prediction results and are difficult to gradually optimize as they are used. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of a cross-process scheduling method for hematoma prediction task processing according to the present invention; Figure 2 This is a schematic diagram of the framework of a cross-process scheduling hematoma prediction task processing system according to the present invention. Detailed Implementation

[0014] 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.

[0015] Example 1: As Figure 1 As shown, this embodiment provides a cross-process scheduling method for hematoma prediction tasks, including the following steps: S1. Read the hematoma prediction task request corresponding to the object to be predicted, parse it to form the original feature record, and perform field integrity verification, numerical legality verification, numerical normalization, missing value completion and feature alignment processing on the original feature record to generate a feature snapshot data object; S2. Write the feature snapshot data object into the cross-process scheduling queue, generate a record of prediction tasks to be executed, select the target prediction task record to be executed according to the preset scheduling rules and distribute it to the target prediction process, and call the hematoma prediction model to obtain the initial prediction result. S3. Associate and store the initial prediction results with the corresponding feature snapshot data objects to generate historical prediction records. Calculate the prediction deviation value and generate a deviation record when the corresponding result feedback record is received. S4. Merge and sort the deviation records according to their identity identifiers to construct deviation sequence data. Perform imitation learning processing based on the deviation sequence data to generate deviation correction parameters corresponding to the identity identifiers. S5. When a hematoma prediction task request for the corresponding object to be predicted is received again, the new feature snapshot data object and the corresponding deviation correction parameters are read to obtain a new initial prediction result. The prediction result is then corrected based on the deviation correction parameters, and the corrected prediction result is output as the final hematoma probability prediction result.

[0016] S1 specifically includes the following sub-steps: S110. Read the hematoma prediction task request corresponding to the object to be predicted, parse it to form the original feature record, perform field integrity verification on the original feature record first, then perform numerical validity verification, delete invalid original feature records that fail either verification, and output the set of verified original feature records.

[0017] The original feature record includes at least the following fields: name, age, body mass index (BMI), fibrinogen (FIB), systolic blood pressure (SBP), albumin (ALB), and triglycerides (TAG). Among these, the name and age fields are used together to generate the identity identifier. The age field is an integer field, while the BMI, FIB, SBP, ALB, and TAG fields are numeric fields.

[0018] Field integrity checks are used to determine whether the name, age, BMI, FIB, SBP, ALB, and TAG fields are empty; numerical validity checks are used to determine whether the age field is greater than 0, and whether the BMI, FIB, SBP, ALB, and TAG fields are greater than 0.

[0019] For original feature records where the name field or age field is empty, delete the entire record directly, because the record cannot generate a unique identifier; for original feature records where the age field is less than or equal to 0, or where any of the BMI, FIB, SBP, ALB, and TAG fields is less than or equal to 0, mark them as invalid original feature records and delete them, because such data cannot be used as valid input for the hematoma prediction model.

[0020] For example, a record with the name field "Zhang San", age field 58, BMI field 24.6, and FIB field 3.2 can pass this step's verification, while a record with an age field value of 0 is deleted in this step. Each record in the verified set of original feature records retains its name, age, BMI, FIB, SBP, ALB, and TAG fields, which are used as input for S120.

[0021] S120. Read the verified set of original feature records, and perform field filtering, numerical normalization, missing value completion and feature alignment processing on each original feature record in sequence. Delete records that cannot meet the input requirements of the hematoma prediction model, and output the standardized set of feature records.

[0022] Field filtering removes redundant fields that do not belong to the preset feature name set of the hematoma prediction model, ensuring that only the fields actually needed by the hematoma prediction model are retained in subsequent processing records. Numerical normalization maps fields with different dimensions to a unified numerical range, ensuring that the numerical distribution used by the hematoma prediction model during the inference phase is consistent with the numerical distribution used during the training phase. In this embodiment, fields requiring normalization are processed using the following formula: in, The normalized field value; The field value before normalization; This represents the minimum value of the corresponding field in the preset training samples. This represents the maximum value of the corresponding field in the preset training samples. The calculation result of the above formula is used to replace the corresponding original value in the current original feature record and form a normalized field value.

[0023] Taking the SBP field as an example, when the SBP field of a certain object to be predicted is 142, and the minimum value of the SBP field in the preset training samples is 90 and the maximum value is 180, the corresponding normalization result can be obtained according to the above formula, and the normalization result can be written into the current record for subsequent encapsulation.

[0024] After numerical normalization, missing value completion is performed on the fields that are allowed to be completed. The name and age fields are mandatory; if they are missing, the record has already been deleted from S110. The BMI, FIB, SBP, ALB, and TAG fields are fields that are allowed to be completed. When any of these fields is missing, completion is performed using the mean of valid samples from the preset training samples. The completion method is calculated using the following formula: in, The complete value for the target missing field; This is the number of valid samples for the corresponding field in the preset training samples. For the first The values ​​of the field corresponding to each valid sample, where i is a positive integer, and The completed values ​​are written to the corresponding missing field positions in the current record after calculation, ensuring that the record meets the input requirements of the hematoma prediction model in terms of the number of fields. For example, when the ALB field is missing, the mean value can be calculated using all valid ALB field values ​​from the training samples, and this mean value can be written to the ALB field of the current record.

[0025] After completing numerical normalization and missing value completion, feature alignment is performed. Feature alignment is used to rearrange the field order in the current record according to the preset feature name set and preset field order of the hematoma prediction model, and to fill in the preset default values ​​for fields required by the hematoma prediction model but not yet present in the current record.

[0026] If a record, after completion, still fails to meet the requirements of the hematoma prediction model regarding the number and order of fields, then that record is deleted. Each record in the standardized feature record set has a uniform field structure, uniform field order, and uniform number of fields, and serves as input to S130.

[0027] S130. Read the standardized feature record set, encapsulate each standardized feature record into a feature snapshot data object, perform encapsulation consistency verification, delete feature snapshot data objects that fail to encapsulate or have abnormal field mapping, and output the feature snapshot data object set.

[0028] The feature snapshot data object includes at least an identity identifier, a set of field key-value pairs, a set of field order indexes, an object generation timestamp, and an object version number. The identity identifier is generated by combining the name and age fields. The set of field key-value pairs is used to record the field name and normalized field value of each field. The set of field order indexes is used to clarify the order of each field in the hematoma prediction model input vector. The object generation timestamp is used to record the current encapsulation time, and the object version number is used to distinguish encapsulation results of different batches.

[0029] The encapsulation consistency check is used to determine whether the number of fields in the key-value pair set is consistent with the number of indices in the field order index set, and whether the field order index set is consistent with the preset field order of the hematoma prediction model. Feature snapshot data objects that fail the encapsulation consistency check are directly deleted and do not enter the cross-process scheduling process. Each feature snapshot data object in the feature snapshot data object set is treated as an independent object to be scheduled and serves as input to S210.

[0030] S2 specifically includes the following sub-steps: S210. Read the feature snapshot data object set, write each feature snapshot data object into the cross-process scheduling queue, generate a record of prediction tasks to be executed, perform enqueue verification, delete the record of prediction tasks to be executed that failed to write or failed to generate the task identifier, and output the set of prediction tasks to be executed.

[0031] The record of the prediction task to be executed includes at least a unique task identifier, an identity identifier, a feature snapshot data object, a task status field, a task generation timestamp, and a scheduling count field. The unique task identifier consists of an identity identifier, a task generation timestamp, and a task sequence number, and is used to establish a unique association in subsequent steps. The task status field includes at least four states: pending scheduling, in execution, completed, and execution failed. The scheduling count field is used to record the number of times the current task has been scheduled.

[0032] The enqueue verification checks whether the unique task identifier is empty, whether the feature snapshot data object is empty, and whether the task status field is initialized to pending scheduling. For any pending prediction task record that does not meet any of the conditions, the record is deleted. Each record in the set of pending prediction tasks serves as input to S220.

[0033] S220. Read the set of prediction tasks to be executed, select the target prediction task record to be executed according to the preset scheduling rules, distribute the target prediction task record to the target prediction process according to the target prediction process status, and output the target prediction task record in the execution state.

[0034] To ensure the executability of cross-process scheduling rules, this embodiment calculates a task priority score for each pending predicted task record and prioritizes the execution of the pending predicted task record with the highest task priority score. The task priority score is calculated using the following formula: in, Rate task priority; The task waiting time is calculated by subtracting the task generation timestamp from the current time. This refers to the number of retries corresponding to the scheduling count field. and This is the priority weight coefficient, and , These are used to adjust the impact of task waiting time and retries on the scheduling order, respectively.

[0035] The results of the task priority scoring above are used to establish a comparable ranking basis among the records to be scheduled. For example, when the waiting times of two predicted task records to be executed are 12 seconds and 5 seconds respectively, and the number of retries are 1 and 3 respectively, the ranking can be determined based on the selected priority score. and Calculate the corresponding task priority score and select the record with the higher score first.

[0036] After identifying the target prediction task record to be executed, a busy / idle status check is performed on the target prediction process. If the target prediction process is idle, the task status field of the target prediction task record to be executed is updated to "in execution," and the feature snapshot data object is passed to the target prediction process as the sole input parameter. If the target prediction process is busy, the target prediction task record to be executed is retained in the cross-process scheduling queue, its task status field is not changed, and it continues to participate in the task priority scoring calculation in the next scheduling cycle. For target prediction task records to be executed whose scheduling count field exceeds a preset threshold, the task status field is directly updated to "execution failed," and the record is deleted from the current main processing chain, thus preventing the formation of a separate set of failed tasks. Target prediction task records to be executed that are in the "in execution" state are used as input to S230.

[0037] S230. Read the target prediction task record that is in the execution state, extract the feature snapshot data object in the target prediction process and call the hematoma prediction model to perform hematoma probability prediction processing, verify the validity of the prediction execution result, delete the task record corresponding to the invalid prediction result, and output the initial prediction result set.

[0038] After receiving the feature snapshot data object, the target prediction process first restores the input order required by the hematoma prediction model according to the field order index set, then assembles the field values ​​into an input vector and sends it into the hematoma prediction model to obtain the initial prediction result.

[0039] The result validity check is used to determine whether the initial prediction result is empty, whether it is numerical, and whether it falls within the probability range of 0 to 1. For prediction task records whose initial prediction results are empty, non-numerical, or less than 0 or greater than 1, the task status field is updated to "execution failed," and the record is deleted from the current main processing chain. For prediction task records that pass the result validity check, the task status field is updated to "completed." Each initial prediction result in the initial prediction result set includes at least a unique task identifier, an identity identifier, an initial prediction result value, and a result generation time, which are used as input to S310.

[0040] In this embodiment, the hematoma prediction model can be any one of a logistic regression model, a random forest model, or a multilayer perceptron network model, pre-trained based on historical real case data. The hematoma prediction model uses the input vector as input features, performs forward propagation calculations through its internal algorithm nodes or network weights, and finally outputs a floating-point value between 0 and 1 from the classification layer at the end of the model as the initial prediction result. The basic architecture of this model belongs to a mature machine learning classification model, and its specific number of network layers or tree depth can be conventionally set according to actual computing resources.

[0041] S3 specifically includes the following sub-steps: S310. Read the initial prediction result set, associate each initial prediction result with the corresponding feature snapshot data object, generate historical prediction records, delete historical prediction records that failed to associate or had missing fields, and output the historical prediction record set. Historical prediction records must include at least a unique task identifier, identity identifier, feature snapshot data object, initial prediction result, result generation time, task status field, and record write time.

[0042] When storing data in association, the initial prediction result is first matched with the feature snapshot data object in the prediction task record based on the unique task identifier. If the match is successful, a historical prediction record is generated and written to the prediction result database. If the match fails, the corresponding initial prediction result is deleted and no historical prediction record is generated.

[0043] After historical prediction records are written, a field integrity check is performed on them. This check determines whether the unique task identifier, identity identifier, feature snapshot data object, and initial prediction result are complete. Historical prediction records that fail the field integrity check are deleted. Each record in the historical prediction record set serves as input to S320.

[0044] S320: Read the historical prediction record set and the subsequent received result feedback records. First, perform matching and retrieval based on the identity identifier and unique task identifier. Then, calculate the prediction deviation value, delete records that failed to match or whose actual result value is invalid, and output the deviation record set.

[0045] The result feedback record includes at least the identity identifier, unique task identifier, actual result value, and result generation time. The actual result value is expressed in a numerical form that is comparable to the initial prediction result. In this embodiment, the actual result value is expressed in a binary form of 0 or 1, where 0 indicates that no hematoma has occurred and 1 indicates that a hematoma has occurred.

[0046] During the matching retrieval, a unique corresponding historical prediction record is retrieved from the historical prediction record set based on the identity identifier and unique task identifier. If no corresponding record is found, or multiple records are found, the result feedback record is deleted, and deviation calculation is not performed. For records with successful matching and valid actual result values, the prediction deviation value is calculated using the following formula: in, This represents the predicted deviation value; This represents the actual result value; This represents the initial prediction result corresponding to this historical task. The calculation result of this formula is used to represent the direction and magnitude of the difference between the hematoma prediction model output and the actual result. A value greater than 0 indicates that the initial prediction result for that historical task was too low; when... A value less than 0 indicates that the initial prediction result for that historical task was too high; when... When the value is 0, it means that the initial prediction result for the historical task is consistent with the actual result.

[0047] For example, if an initial prediction result is 0.35, and the subsequent actual result is 1, then the prediction deviation value is 0.65. After calculation, a deviation record is generated. The deviation record includes at least the identity identifier, unique task identifier, initial prediction result, actual result value, prediction deviation value, and deviation generation time. Each record in the deviation record set serves as input to S330.

[0048] S330. Read the set of deviation records, perform effective sample filtering and object merging on the deviation records, delete deviation records with missing identity identifiers, empty predicted deviation values, or abnormal deviation generation times, and output the set of effective deviation records merged by identity identifier.

[0049] Valid sample screening ensures that the deviation records entering subsequent imitation learning processing are calculable and traceable; object merging processing groups multiple deviation records belonging to the same identity into the same subset of deviation records, allowing the historical error evolution trajectory of the same object to be predicted to be continuously used in subsequent steps. The output set of valid deviation records is grouped by identity and used as input to S410.

[0050] S4 specifically includes the following sub-steps: S410. Read the set of valid deviation records. For each identity identifier, sort the deviation record subset in ascending order by deviation generation time and construct deviation sequence data. Then calculate the historical prediction deviation mean, delete the deviation record subset with insufficient record quantity or abnormal sorting, and output the deviation sequence set.

[0051] The deviation sequence data includes at least the sequence position index, initial prediction result, prediction deviation value, and sample timestamp. To ensure that the state vectors in subsequent imitation learning samples possess stable error aggregation information, this embodiment calculates the average historical prediction deviation for the valid deviation records corresponding to the current identity identifier. The calculation method is as follows: in, This represents the average historical prediction deviation. This represents the number of valid deviation records corresponding to the current identity identifier. For the first The prediction deviation value corresponding to each deviation record, where j is a positive integer, and The aforementioned historical prediction average is used to characterize the overall deviation level of the object to be predicted over a historical period, and is continuously invoked during the construction of the state vector in S420 and the execution of correction calculations in S520.

[0052] For example, when the predicted deviation values ​​of the three most recent deviation records are 0.20, 0.35, and 0.10 respectively, the average historical predicted deviation can be calculated using the above formula. For a subset of deviation records whose number of sorted records is less than a preset sample threshold, they are not included in the imitation learning modeling process but are instead marked as default parameter generation objects; in this embodiment, the preset sample threshold can be set to 3. Each deviation sequence data in the deviation sequence set and each default parameter generation object serve as the input source for S420.

[0053] S420: Read the set of deviation sequences, construct a learning sample sequence for each deviation sequence data, and generate deviation correction parameters based on the learning sample sequence; at the same time, generate default deviation correction parameters for the default parameter generation object; delete deviation sequence data with missing sample fields, abnormal sample timestamps, or training failures, and output the set of deviation correction parameters.

[0054] The imitation learning sample sequence consists of multiple imitation learning samples. Each imitation learning sample includes at least a state vector, a demonstration action value, and a sample timestamp. The state vector represents the error state at the current correction decision moment, and the demonstration action value represents the target correction amount to be applied to the current initial prediction result. To ensure a fixed input structure, the state vector includes at least the current initial prediction result, the most recent prediction deviation value, and the average of the most recent three prediction deviations. When there are fewer than three most recent three prediction deviation values, the average of the existing entries is used instead of the average of the most recent three prediction deviations.

[0055] Since the impact of historical samples on current prediction corrections varies, newer deviation records should have higher weights. Therefore, a sample-level time decay weight is introduced for each deviation sample, calculated using the following formula: in, For the first Time decay weight of each deviation record; The parameters for calculating the time decay coefficient are as follows: ; The timestamp of the current prediction task; For the first The timestamp of each bias record is used to generate the bias, where e is a natural constant. The sample-level time decay weights described above are used to adjust the contribution of each historical bias sample to the generation of the current correction parameters, so that bias records that are closer in time to the current prediction task have a higher proportion in parameter generation. For example, if two bias records were formed 1 hour and 24 hours before the current time, respectively, then according to the above formula, the time decay weight of the former is higher than that of the latter.

[0056] After obtaining the sample-level time decay weights corresponding to each sample Then, the time decay weights of all samples corresponding to the same identity are aggregated to generate parameter-level time decay coefficients. Parameter-level time decay coefficient Used for subsequent calculation of revised prediction results. Specifically, it can be used for all those corresponding to the same identity identifier. Calculate the weighted average or arithmetic mean to obtain a unique parametric time decay coefficient. and the parameter-level time decay coefficient Write the bias correction parameter. The purpose of this setting is to compress sample-level time information into a single parameter that can be directly called during the inference stage, ensuring that the correction calculation in S520 has a clear input source.

[0057] Imitation learning is used to learn the mapping relationship from error states to corrective behaviors based on state vectors, demonstration action values, and sample-level time decay weights, and outputs deviation correction parameters.

[0058] The bias correction parameters include at least a correction coefficient, a parameter-level time decay coefficient, and an update version number. The correction coefficient characterizes the correction magnitude, the parameter-level time decay coefficient reflects the timeliness of historical samples in subsequent correction prediction calculations, and the update version number distinguishes parameter records generated in different batches. For default parameter generation objects, default bias correction parameters are directly generated and written to the bias correction parameter set; the default bias correction parameters can be set to a correction coefficient of 0 and a parameter-level time decay coefficient of 1. Each set of parameters in the bias correction parameter set corresponds one-to-one with an identifier and serves as input to S430.

[0059] As a specific implementation method, the imitation learning process employs a behavior cloning network architecture based on a multilayer perceptron. The specific processing steps include: using the state vector as input layer data for the behavior cloning network, passing it through at least one hidden layer of nonlinear activation mapping, and then outputting the predicted corrected action value from the output layer; constructing a weighted mean squared error loss function with the demonstrated action value as the true label; during the calculation of the weighted mean squared error loss function, using the sample-level time decay weight corresponding to each sample as the dynamic penalty coefficient of the sample's loss term, so that samples with more recent time contribute more to the overall loss; Based on the weighted mean square error loss function, the gradient descent algorithm is used to iteratively update the network weight parameters of the behavioral clone network in multiple rounds until the weighted mean square error loss function converges. After the model converges, the output layer weight parameters or scaling coefficients of specific neurons of the behavioral clone network are extracted as the correction coefficients to ensure that the correction coefficients reflect the correction rules learned based on the historical error sequence.

[0060] S430. Read the deviation correction parameter set, establish a mapping relationship between each set of deviation correction parameters and the corresponding identity identifier and persist the mapping relationship, delete records with non-unique mapping relationships or incomplete parameter fields, and output the set of callable deviation correction parameters.

[0061] When establishing the mapping relationship, the identity identifier is used as the primary search key, and the update version number is used as the version differentiation field. The correction coefficient, parameter-level time decay coefficient, and update time are written into the deviation correction parameter storage table. The deviation correction parameter storage table is a data table used to store the mapping relationship between identity identifiers and deviation correction parameters. Parameter field integrity verification is used to determine whether the identity identifier, correction coefficient, parameter-level time decay coefficient, and update version number are complete. Mapping uniqueness verification is used to determine whether the same identity identifier corresponds to only one set of deviation correction parameters under the same update version number. Parameter records that fail any verification are directly deleted and do not proceed to the subsequent correction process.

[0062] Each set of parameters in the callable set of deviation correction parameters can be retrieved by identity and used as input to S510 and S520.

[0063] S5 specifically includes the following sub-steps: S510. When a hematoma prediction task request for the corresponding object to be predicted is received again, S110, S120 and S130 are executed repeatedly to generate a new feature snapshot data object, and the set of callable deviation correction parameters corresponding to the object to be predicted is read synchronously. Parameter records with mismatched identity or invalid parameter versions are deleted, and the set of prediction tasks to be corrected is output.

[0064] The generation method of the new feature snapshot data object is consistent with S110-S130 to ensure the consistency of the input structure across multiple prediction tasks. When reading the bias correction parameters, the identifier in the new feature snapshot data object is used as the search condition to find the corresponding parameter in the set of callable bias correction parameters. When multiple sets of parameters exist, the deviation correction parameter with the largest updated version number is selected as the current valid parameter; when no parameter is found, the default deviation correction parameter defined in S420 is directly called. Each record in the set of prediction tasks to be corrected includes a new feature snapshot data object and its corresponding valid deviation correction parameter, which are used as input to S520.

[0065] S520: Read the set of prediction tasks to be corrected, write the new feature snapshot data object and the effective bias correction parameters into the cross-process scheduling queue, call the target prediction process to obtain the new initial prediction result, then perform correction calculation based on the effective bias correction parameters, delete records where the correction conditions are not met, and output the set of corrected prediction results. The method of obtaining the new initial prediction result is consistent with S220 and S230. To ensure that the bias correction parameters can directly affect the new initial prediction result during the inference phase, the corrected prediction result is calculated according to the following formula in this embodiment: in, To correct the prediction results; This is the new initial prediction result for the current task; For correction coefficients, and ; This represents the average historical prediction deviation. This is the parameter-level time decay coefficient.

[0066] The calculation logic of the formula is as follows: First, the new initial prediction result corresponding to the current task is used as the basic prediction value; second, the average historical prediction deviation is used to reflect the overall deviation direction and deviation magnitude formed by the current object to be predicted in the historical prediction process; then, the correction coefficient is used to adjust the influence of the average historical prediction deviation on the current basic prediction value; finally, the parameter-level time decay coefficient is used to further reflect the timeliness of the historical deviation sample.

[0067] In other words, if a certain object to be predicted has a long history of underestimating its value, then The value is often greater than 0, and the revised prediction result will be adjusted upwards based on the new initial prediction result; if a certain object to be predicted has a long history of being overestimated, then... The value is often less than 0, and the corrected prediction result will be adjusted downwards based on the new initial prediction result. For example, when the new initial prediction result is 0.42, the correction coefficient is 0.6, the average historical prediction deviation is 0.20, and the parameter-level time decay coefficient is 0.8, the corresponding corrected prediction result can be obtained according to the above formula.

[0068] To ensure that the corrected prediction results always retain their probabilistic meaning, boundary truncation is performed after the corrected calculation is completed: when the corrected prediction result is less than 0, the corrected prediction result is truncated to 0; when the corrected prediction result is greater than 1, the corrected prediction result is truncated to 1.

[0069] For records where the initial prediction result is empty, the correction coefficient is empty, the parameter-level time decay coefficient is empty, or the historical prediction deviation mean cannot be calculated, the corresponding record is deleted, and no corrected prediction result is output. Each result in the corrected prediction result set includes at least a unique task identifier, an identity identifier, a corrected prediction result value, and a corrected generation time, and is used as input to S530.

[0070] S530: Read the set of corrected prediction results, output each corrected prediction result as the final hematoma probability prediction result, and write the corrected prediction results back to the historical prediction records to generate an updated set of historical prediction records. Delete records that failed to output or write back, so that they can continue to be used in subsequent imitation learning processing. During output, return the corrected prediction result value to the calling end that initiated the current hematoma prediction task request and mark it as the final hematoma probability prediction result.

[0071] During the write-back process, the corrected prediction results are written to the corrected result field and corrected time field of the corresponding historical prediction record based on the unique task identifier. This ensures that the same historical prediction record simultaneously saves both the initial prediction result and the corrected prediction result. This is because when a result feedback record is received again later, it is possible to trace back both the feature snapshot data object used in this prediction and the initial and corrected prediction results, thus providing a continuous data foundation for the next round of deviation record generation and imitation learning updates.

[0072] For corrected prediction results for which no unique task identifier corresponds to a historical prediction record, the corresponding write-back operation is deleted and not included in the updated historical prediction record set. The updated historical prediction record set continues to be used as input to S320 when result feedback records are received again, thus forming a closed-loop processing chain of "cross-process scheduling prediction - deviation feedback - imitation learning correction - cross-process scheduling prediction again".

[0073] Example 2: Figure 2 As shown, this embodiment provides a cross-process scheduling hematoma prediction task processing system, including: The original feature processing module is used to read the hematoma prediction task request corresponding to the object to be predicted, parse it to form an original feature record, perform field integrity verification, numerical legality verification, numerical normalization, missing value completion and feature alignment processing on the original feature record, generate a standardized feature record, and encapsulate the standardized feature record into a feature snapshot data object; The task scheduling and prediction execution module is used to write feature snapshot data objects into a cross-process scheduling queue, generate prediction task records to be executed, select target prediction task records to be executed according to preset scheduling rules and distribute them to the target prediction process, call the hematoma prediction model to perform hematoma probability prediction processing, and obtain the initial prediction results. The historical association and deviation generation module is used to associate and store the initial prediction results with the corresponding feature snapshot data objects to generate historical prediction records. When the corresponding result feedback record is received, the prediction deviation value is calculated based on the historical prediction record and the result feedback record to generate a deviation record. The deviation learning and parameter generation module is used to merge and sort deviation records by identity identifier, construct deviation sequence data, perform imitation learning processing based on deviation sequence data, and generate deviation correction parameters corresponding to identity identifier. The prediction correction and result write-back module is used to read the new feature snapshot data object and the corresponding deviation correction parameters when a hematoma prediction task request for the corresponding object to be predicted is received again, call the target prediction process to obtain a new initial prediction result, correct the new initial prediction result based on the deviation correction parameters, output the corrected prediction result as the final hematoma probability prediction result, and write the corrected prediction result back to the corresponding historical prediction record.

[0074] All the above formulas are performed using dimensionless numerical calculations; the relevant formulas are based on empirical models that approximate the real situation, obtained through extensive data collection and software simulation fitting. The preset parameters and thresholds involved in the formulas can be conventionally set and adjusted by those skilled in the art according to the physical constraints of the actual application scenario.

[0075] The above are merely specific embodiments 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.

[0076] In conclusion, the above are merely preferred embodiments of the present invention and are 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 hematoma prediction tasks through cross-process scheduling, characterized in that, Includes the following steps: S1. Read the hematoma prediction task request corresponding to the object to be predicted, parse it to form the original feature record, and perform field integrity verification, numerical legality verification, numerical normalization, missing value completion and feature alignment processing on the original feature record to generate a feature snapshot data object; S2. Write the feature snapshot data object into the cross-process scheduling queue, generate a record of prediction tasks to be executed, select the target prediction task record to be executed according to the preset scheduling rules and distribute it to the target prediction process, and call the hematoma prediction model to obtain the initial prediction result. S3. Associate and store the initial prediction results with the corresponding feature snapshot data objects to generate historical prediction records. Calculate the prediction deviation value and generate a deviation record when the corresponding result feedback record is received. S4. Merge and sort the deviation records according to their identity identifiers to construct deviation sequence data. Perform imitation learning processing based on the deviation sequence data to generate deviation correction parameters corresponding to the identity identifiers.

2. The method for cross-process scheduling of hematoma prediction task processing according to claim 1, characterized in that, Also includes: S5. When a hematoma prediction task request for the corresponding object to be predicted is received again, the new feature snapshot data object and the corresponding deviation correction parameters are read to obtain a new initial prediction result. The prediction result is then corrected based on the deviation correction parameters, and the corrected prediction result is output as the final hematoma probability prediction result.

3. The method for cross-process scheduling of hematoma prediction task processing according to claim 1, characterized in that, S1 specifically includes: Read the hematoma prediction task request corresponding to the object to be predicted, and parse it to form an original feature record that includes at least the name field, age field, body mass index (BMI) field, fibrinogen (FIB) field, systolic blood pressure (SBP) field, albumin (ALB) field, and triglyceride (TAG) field; Perform field integrity verification and numerical validity verification on the original feature records, and delete invalid original feature records; Perform numerical normalization, missing value completion, and feature alignment on the verified original feature records to generate standardized feature records; Standardized feature records are encapsulated into feature snapshot data objects that include identity identifiers, a set of field key-value pairs, and a set of field sequential indexes.

4. The method for cross-process scheduling of hematoma prediction task processing according to claim 1, characterized in that, S2 specifically includes: Write the feature snapshot data object into the cross-process scheduling queue to generate a record of the prediction task to be executed, which includes a unique task identifier, identity identifier, feature snapshot data object, task status field, task generation timestamp and scheduling count field. Calculate the task priority score of each predicted task record to be executed according to the preset scheduling rules, and select the target predicted task record to be executed with the highest task priority score; The target prediction task record is distributed to the target prediction process, the hematoma prediction model is called to perform hematoma probability prediction processing, and the initial prediction result is obtained.

5. The method for cross-process scheduling of hematoma prediction task processing according to claim 1, characterized in that, S3 specifically includes: The initial prediction results are associated with the corresponding feature snapshot data objects and stored to generate historical prediction records that include at least a unique task identifier, an identity identifier, a feature snapshot data object, the initial prediction results, and the result generation time. Upon receiving the corresponding result feedback record, retrieve the corresponding historical prediction record based on the identity identifier and unique task identifier.

6. The method for cross-process scheduling of hematoma prediction task processing according to claim 5, characterized in that, Also includes: The prediction deviation value is calculated based on the initial prediction results in the historical prediction records and the actual result values ​​in the result feedback records, and a deviation record is generated.

7. The method for cross-process scheduling of hematoma prediction task processing according to claim 1, characterized in that, S4 specifically includes: Deviation records are merged according to their identity identifiers and sorted chronologically according to the time of deviation generation to construct deviation sequence data; The imitation learning sample sequence is constructed based on the deviation sequence data. The imitation learning sample sequence includes at least a state vector, a demonstration action value, and a sample timestamp.

8. The method for cross-process scheduling of hematoma prediction task processing according to claim 7, characterized in that, Also includes: Imitation learning processing is performed based on the imitation learning sample sequence to generate bias correction parameters including correction coefficients, parameter-level time decay coefficients, and update version numbers, and a mapping relationship between bias correction parameters and identity identifiers is established.

9. The method for cross-process scheduling of hematoma prediction task processing according to claim 1, characterized in that, S5 specifically includes: When a hematoma prediction task request for the corresponding object to be predicted is received again, a new feature snapshot data object is generated and the deviation correction parameter corresponding to its identity is read. Invoke the target prediction process to obtain new initial prediction results; Based on the deviation correction parameters, a correction calculation is performed on the new initial prediction results to obtain the corrected prediction results; The corrected prediction result is output as the final hematoma probability prediction result, and the corrected prediction result is written back to the corresponding historical prediction record.

10. A cross-process scheduling hematoma prediction task processing system, employing the cross-process scheduling hematoma prediction task processing method according to any one of claims 1 to 9, characterized in that, include: The module includes: original feature processing module, task scheduling and prediction execution module, historical association and deviation generation module, deviation learning and parameter generation module, and prediction correction and result write-back module.

Images