A computer system-based medical data processing system and method

Abstract

The application discloses a kind of medical data processing system and method based on computer system, specifically relates to medical data processing technical field, including obtaining the vital sign data of target patient in target processing period, monitoring waveform data, equipment parameter data, inspection data and treatment action record, and by computer system execution time correlation arrangement, output observation medical data set;By constructing the observation medical data set under the intervention condition of treatment action, generate the candidate state explanation branch respectively representing the treatment action conduction relationship and the relationship of patient ontology state change, and by GPU, observation result reconstruction, index consistency back and remaining deviation solution are carried out to each candidate state explanation branch, determine target ontology state result.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and more specifically, to a computer-based medical data processing system and method. Background Technology

[0002] In existing medical data processing technologies such as intensive care, emergency resuscitation, and anesthesia recovery, the mainstream practice in the industry mainly addresses how to quickly complete risk identification, status assessment, or early warning output based on real-time patient monitoring data. Typically, monitoring waveforms, vital signs, bedside test results, ventilator parameters, infusion records, and medical order execution information are collected into a computer system, and then rule models or intelligent algorithms call GPUs to perform parallel processing of multi-source data at the current moment to improve response speed and processing capacity in high-frequency scenarios. For example, when making continuous risk assessments for shock patients in the ICU who are receiving vasopressors, mechanical ventilation, and rapid fluid resuscitation, the system is required to continuously receive monitoring parameters and equipment parameters at a rate of updates per second, and to maintain the accuracy of the assessment results in relation to the patient's original condition under conditions of frequent interventions and simultaneous fluctuations in multiple indicators. However, under this constraint, the mainstream approach will consistently reveal an observable and verifiable defect: the changes in indicators formed after the implementation of treatment actions will be directly sent to the GPU for processing as changes in the patient's natural state. This will result in the same patient experiencing a short-term rise in blood pressure immediately after the infusion of vasopressors, a rapid improvement in blood oxygen after the adjustment of ventilator parameters, and a temporary recovery of some circulatory indicators after fluid resuscitation. The risk output results of the system will show phenomena such as failure to report when they should, failure to rise when they should, or failure to fall when they should. The reason is that the existing plan usually only uses the latest data or the time window after fixed intervention to exclude the input, and cannot distinguish which data changes belong to the evolution of the patient's real state and which data changes belong to the transmission results of treatment actions on different indicator chains. Therefore, the technical problem to be solved by this application is: how to effectively distinguish between changes in indicators caused by treatment actions and changes in the patient's physical state during medical data processing based on computer systems and involving GPUs, so as to avoid misjudging the results of treatment intervention as changes in the patient's natural condition. Summary of the Invention

[0003] To overcome the aforementioned deficiencies in the prior art, embodiments of the present invention provide a medical data processing system and method based on a computer system. By constructing an observational medical dataset under the conditions of interventional treatment actions, candidate state interpretation branches are generated to represent the relationship between the transmission of treatment effects and the relationship between changes in the patient's ontological state. The GPU is then used to perform observation result reconstruction, index consistency back substitution, and residual bias solution on each candidate state interpretation branch to determine the target ontological state result, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a medical data processing method based on a computer system, comprising: S1. Acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and perform time correlation and processing through the computer system to output the observation medical dataset; S2. Based on the action indicators, starting order, duration, and change order of each indicator in the observed medical dataset corresponding to the treatment action record, the intervention transmission decomposition is performed on the observed medical dataset to generate multiple candidate state interpretation branches. The candidate state interpretation branches respectively represent the transmission relationship of different treatment effects and the relationship of different patient ontological state changes. S3. Load the candidate state interpretation branches to the GPU, perform observation result reconstruction, index consistency back substitution and residual bias solution for each candidate state interpretation branch, and output the reconstruction bias result and ontology state solution result for each candidate state interpretation branch. S4. Based on the reconstruction deviation results and ontology state solution results corresponding to each candidate state interpretation branch, perform branch competition judgment, determine the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch. S5. Generate the medical treatment result corresponding to the target patient based on the target ontology state result, and write the medical treatment result into the result reading path.

[0005] In a preferred embodiment, S1 includes: S1-1. Acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and extract the corresponding data source identifier, acquisition time identifier and record type identifier respectively, and output the original medical record set; S1-2. Based on the execution time, target, and execution order in the treatment action record, perform time-series correlation and positioning of each record in the original medical record set, determine the preceding, synchronous, or subsequent correlation position of each record relative to each treatment action record, and output a medical record set with treatment correlation position identifiers. S1-3. Organize and combine the medical record set with treatment-related location identifiers according to the target treatment period and the execution order of the treatment-related location identifiers to generate an observational medical dataset that characterizes the observational state of the target patient under the intervention of treatment actions.

[0006] In a preferred embodiment, S2 includes: S2-1. Based on the observational medical dataset and treatment action records corresponding to the target patient, and based on the effect indicators, starting order and duration of the treatment action records, combined with the change order of each indicator in the observational medical dataset, identify the reachable transmission interval of each treatment action on the change of each indicator, and output the treatment transmission constraint set. S2-2. Based on the treatment transmission constraint set, split and combine the changes of each indicator in the observed medical dataset, divide the changes of each indicator into the treatment effect change part and the ontology state change part, and generate multiple candidate state interpretation branches according to the combination relationship between different treatment effect change parts and different ontology state change parts, and output the candidate state interpretation branch set. S2-3. Establish therapeutic effect transmission relationship and ontological state change relationship for each candidate state interpretation branch in the candidate state interpretation branch set, and generate multiple candidate state interpretation branches that respectively characterize different therapeutic effect transmission relationships and different patient ontological state change relationships.

[0007] In a preferred embodiment, the process of outputting the candidate state interpretation branch set in S2-2 further includes: S2-21. Based on the correlation of the action indicators, the initial sequence constraint and the continuous process constraint in the treatment transmission constraint set, construct a treatment transmission candidate graph for the changes of each indicator in the observed medical dataset. The treatment transmission violation quantity and the ontology state jump violation quantity together constitute the attribution determination quantity. Under the constraints of the sequential relationship of treatment actions and the reachability relationship of indicator changes, perform initial attribution solution for each indicator change and output the initial split result of indicator change. S2-22. Based on the initial splitting results of indicator changes, perform time sequence consistency checks, indicator linkage consistency checks, and treatment response closure consistency checks on each indicator change under different treatment action paths and different ontology state evolution paths. Perform attribution reassignment on indicator changes with verification conflicts. Based on the number of verification-passing items and the number of verification-conflicting items after reassignment, form a combined confidence result, and output the candidate splitting combination set and the combined confidence result corresponding to each candidate splitting combination.

[0008] In a preferred embodiment, the process of outputting the candidate state interpretation branch set in S2-2 further includes: S2-23. Based on the candidate splitting and combination set and the combination confidence result, perform branch expansion and branch merging on each candidate splitting and combination, retain the splitting and combination whose combination confidence result is higher than the previous round and whose number of verification conflict items decreases, and stop processing when the combination confidence result no longer changes and the number of verification conflict items no longer decreases in two consecutive rounds of branch processing, and output the candidate state explanation branch set. S2-24. Write the treatment effect change part, ontology state change part, conduction path identifier and conflict resolution identifier into each candidate state interpretation branch in the candidate state interpretation branch set, respectively, to generate a candidate state interpretation branch set for subsequent ontology state solution processing.

[0009] In a preferred embodiment, S3 includes: S3-1. Based on the candidate state interpretation branch set and the observation medical dataset corresponding to the target patient, load the candidate state interpretation branch set onto the GPU according to the branch identifier, establish observation reconstruction units corresponding to the treatment effect change part and the ontological state change part for each candidate state interpretation branch, and output the observation reconstruction model set corresponding to each candidate state interpretation branch. S3-2. Perform observation result reconstruction and index consistency back substitution on each observation reconstruction model set respectively. Combine the therapeutic effect change part and ontology state change part in each candidate state interpretation branch into branch reconstruction observation results. Based on the corresponding difference between the branch reconstruction observation results and the observed medical dataset, solve the residual bias results and ontology state solution results corresponding to each candidate state interpretation branch. S3-3. Perform branch result organization on the remaining deviation results and ontology state solution results corresponding to each candidate state interpretation branch, and generate a set of reconstruction deviation results and ontology state solution results for subsequent branch competition determination.

[0010] In a preferred embodiment, S4 includes: S4-1. Obtain the candidate state interpretation branch set corresponding to the target patient, the reconstruction deviation result corresponding to each candidate state interpretation branch, and the ontology state solution result corresponding to each candidate state interpretation branch. For each candidate state interpretation branch, read the deviation distribution, deviation direction, and remaining deviation position in the reconstruction deviation result, as well as the state sequence, state transition relationship, and state continuity interval in the ontology state solution result, and generate the branch decision record set corresponding to each candidate state interpretation branch. S4-2. Based on the branch decision record set, perform reconstruction decision and state establishment decision for each candidate state interpretation branch. The reconstruction closure decision is used to determine whether the deviation distribution, deviation direction and deviation remaining position are consistent with the treatment effect change part and the ontological state change part corresponding to the candidate state interpretation branch. The state establishment decision is used to determine whether the state sequence, state transition relationship and state continuous interval are consistent with the index change order in the observed medical dataset and the starting order and continuous process in the treatment action record. Output the branch decision results corresponding to each candidate state interpretation branch.

[0011] In a preferred embodiment, S4 further includes: S4-3. Based on the branch determination results corresponding to each candidate state interpretation branch, retain the candidate state interpretation branches that simultaneously satisfy the reconstruction determination and the state establishment determination, and eliminate the candidate state interpretation branches that do not simultaneously satisfy the reconstruction determination and the state establishment determination. In the retained candidate state interpretation branches, read the number of remaining deviation positions corresponding to the reconstruction deviation result and the number of continuous state intervals corresponding to the ontology state solution result. Determine the candidate state interpretation branch whose correspondence between the number of remaining deviation positions and the number of continuous state intervals is consistent as the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch.

[0012] In a preferred embodiment, S5 includes: S5-1. Based on the target ontology state results corresponding to the target patient, and according to the state category, state continuity interval and state transition relationship in the target ontology state results, perform medical processing mapping to generate the medical processing results corresponding to the target patient. S5-2. Read the corresponding processing type identifier and patient identifier for the medical processing result, and write the medical processing result into the corresponding result reading path according to the processing type identifier and patient identifier, and generate the result writing record corresponding to the target patient.

[0013] A computer-based medical data processing system includes an observation construction module, a transmission decomposition module, a state solution module, a competition determination module, and a result writing module. The observation construction module is used to acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and to perform time correlation and processing through the computer system to output the observation medical dataset; The transmission decomposition module is used to perform intervention transmission decomposition on the observed medical dataset based on the action indicators, starting order, duration, and change order of each indicator in the observed medical dataset corresponding to the treatment action record, and to generate multiple candidate state interpretation branches. The candidate state interpretation branches respectively represent the transmission relationship of different treatment effects and the relationship of different patient ontological state changes. The state solution module is used to load candidate state interpretation branches into the GPU, perform observation result reconstruction, index consistency back substitution and residual deviation solution on each candidate state interpretation branch, and output the reconstruction deviation result and ontology state solution result corresponding to each candidate state interpretation branch. The competition determination module is used to perform branch competition determination based on the reconstruction deviation results and ontology state solution results corresponding to each candidate state interpretation branch, determine the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch. The result writing module is used to generate the medical treatment result corresponding to the target patient based on the target ontology state result, and write the medical treatment result to the result reading path.

[0014] The technical effects and advantages of this invention are as follows: 1. This solution constructs an observational medical dataset, a treatment transmission constraint set, and candidate state interpretation branches, and distinguishes between the treatment effect change part and the ontology state change part in GPU processing. This can reduce the situation of misjudging the treatment intervention result as the patient's natural condition change and relatively improve the correspondence between state judgment and the actual condition in high-frequency intervention scenarios. 2. Based on the action indicators, starting sequence and continuous process in the treatment action record, the reachable conduction interval is identified, and the changes in indicators in the observed medical data are split and combined. This allows for the separation and modeling of treatment conduction changes and ontological state changes in mixed observations, thereby relatively suppressing the interpretation bias caused by directly calculating based solely on the latest data. 3. By loading multiple candidate state interpretation branches onto the GPU and performing observation result reconstruction, index consistency back substitution, and residual bias solution respectively, multiple state interpretation paths can be verified in parallel under the same observation constraints, thereby balancing the needs of multi-branch solution and the processing timeliness in high-frequency medical scenarios within a certain range. 4. Output the reconstruction deviation result and the ontological state solution result for each candidate state interpretation branch, and perform reconstruction closure judgment and state establishment judgment. This can filter out branches that cannot simultaneously explain the observation results and the treatment process, thereby relatively improving the judgment consistency of the target state interpretation branch determination process. 5. Map the state category, state continuity interval and state transition relationship in the target ontology state result to medical processing result, and write it into the result reading path according to the processing type identifier and patient identifier. This can directly convert the state solution result into a readable business result for subsequent early warning processing, auxiliary judgment processing or continuous monitoring processing, thereby improving the continuity of the result landing link. Attached Figure Description

[0015] Figure 1 This is a flowchart outlining the method steps of the present invention; Figure 2 This is a schematic diagram of the system module structure of the present invention. Detailed Implementation

[0016] 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.

[0017] Refer to the instruction manual appendix Figure 1-2 The present invention provides a medical data processing method based on a computer system, comprising: S1. Acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and perform time correlation and processing through the computer system to output the observation medical dataset; In this embodiment, the overall purpose of S1 is to: first, organize the original medical information of the target patient, which is scattered across different sources, sampling rhythms, and recording structures during the target treatment period, into a unified and readable observation basis; then, establish the relative temporal relationship between each record and the treatment action based on the execution time and target of the treatment action; finally, form an observational medical dataset that can characterize the patient's observed state under the intervention of the treatment action, for subsequent treatment transmission decomposition, candidate state interpretation branch generation, and ontology state resolution processing and reading; its principle flow is as follows: first, collect and identify the original records; then, locate the correlation position of the original records relative to the treatment action; then, organize and combine them according to the execution order of the target treatment period and the treatment correlation position, eliminating the confusion of multi-source data in terms of time reference, source identification, and record type; this implementation process includes the following steps: The purpose of S1-1 is to form a set of original medical records with unified fields, clear sources, and comparable time. Its mechanism is to first convert the original collected data into record units with a unified identification structure, and then use them as input objects for subsequent time-series correlation and positioning. The input includes the target patient's vital signs data, monitoring waveform data, equipment parameter data, test data, and treatment action records during the target treatment period. Among them, vital signs data include one or more of heart rate, blood pressure, respiratory rate, blood oxygen saturation, and body temperature; monitoring waveform data includes electrocardiogram waveform, arterial pressure waveform, or respiratory waveform; equipment parameter data includes ventilator parameters, infusion pump parameters, syringe pump parameters, or monitoring equipment operating parameters; test data includes blood gas, complete blood count, biochemistry, or bedside test results; and treatment action records include drug administration, fluid replacement, adjustment of ventilator parameters, starting and stopping equipment, or performing treatment operations. The processing steps include: reading the raw fields output by each source system; converting each record into a unified record structure; extracting the data source identifier, acquisition time identifier, and record type identifier and writing them into the corresponding fields; wherein, the data source identifier is determined according to the system to which the access interface belongs, the acquisition time identifier uses the original record timestamp and is uniformly converted to the same time base, and the record type identifier is classified and written according to vital sign records, waveform records, equipment records, test records, and treatment action records; for monitoring waveform data, segment records are generated according to sampling segments and the segment start time and segment end time are written; for test data, the test request time, result output time, and result confirmation status are written; for treatment action records, the execution time, target, and execution order are written. The output is a raw medical record set, which is grouped according to patient identifiers and written to the intermediate record buffer for S1-2 to read. The abnormal or missing data handling is as follows: when a certain type of record is missing, the field of that type of record is left empty and a missing identifier is written, without interrupting the entry of other records into the set; when there are duplicate records from the same source, they are deduplicated according to the collection time identifier and the source serial number; when the collection time is missing, the device upload time or system reception time is read first to fill in the collection time identifier, and the time filling identifier is written synchronously for identification during subsequent association and positioning. The purpose of S1-2 is to establish the relative temporal position between the original medical records and the treatment actions. Its mechanism is to use the treatment action records as reference points to map each original record to the corresponding interval before, during, or after the execution of the treatment action, thereby providing a positional basis for subsequent identification of the treatment action transmission chain and the ontological state change chain. The input is the set of original medical records output by S1-1, focusing on reading the execution time, target, and execution order of each treatment action record, as well as the collection time identifier, record type identifier, and data source identifier of the remaining records. The processing actions include: first, processing all treatments according to their execution time. The action records are sorted to form a treatment action sequence chain. Then, for each non-treatment action record in the original medical record set, its acquisition time marker is compared with the execution time of each treatment action record in the treatment action sequence chain. If the acquisition time of the record is earlier than the execution time of the target treatment action record, it is determined to be a preceding associated position relative to the treatment action record. If the acquisition time of the record falls within the synchronous action interval with the treatment action execution time as the starting point and the effective interval of the treatment action as the boundary, it is determined to be a synchronous associated position. If the acquisition time of the record is later than the synchronous action interval, it is determined to be a subsequent associated position. The scope of the synchronous action interval is determined according to the record type: for treatment actions involving infusion pumps, syringe pumps, and ventilator parameters, the effective interval is derived from the equipment rule constraint table; for drug administration treatment actions, the effective interval is derived from the drug action rule table; for treatment actions involving procedures, the effective interval is derived from the operation procedure configuration table. If the same record is associated with multiple treatment action records, the consistency of the action object is compared first, followed by the execution time distance. The treatment action record with the same action object and the closest execution time distance is written first as the primary association position, and the remaining associations are written to the extended association field. The output quantity includes treatment associations. The medical record set with location identifiers adds a primary associated treatment action identifier, a treatment associated location identifier, and an extended associated identifier to each record and writes them into the associated location result table for S1-3 to read. The handling of anomalies or omissions is as follows: when a treatment action record lacks an execution order, it is reordered by execution time and the execution order is added; when the target object is missing, the target object category is filled back according to the existing association rules between treatment type and equipment source; when a single record cannot uniquely locate the primary associated treatment action, the parallel association status is retained and a pending resolution identifier is written, which is further resolved during subsequent combination based on record type and time proximity. The purpose of S1-3 is to create an observational medical dataset that can directly characterize the observational state of the target patient under the intervention of treatment actions. Its mechanism involves rearranging, aggregating, and completing the fields of medical records that have completed treatment-related localization according to the execution order of the target processing time period and treatment-related location. This ensures that subsequent steps read not isolated records, but unified observational objects with treatment temporal semantics. The input is the set of medical records with treatment-related location identifiers output from S1-2. The focus is on reading the patient identifier, target processing time period, record type identifier, acquisition time identifier, main associated treatment action identifier, and treatment-related location identifier. The processing actions include: first, filtering records within the target processing time period; then, according to the previous... A location hierarchy is established in the order of associated location, synchronous associated location, and subsequent associated location. Within each location hierarchy, records are arranged in ascending order by acquisition time identifier. For multiple types of records at the same acquisition time, parallel field groups are established according to record type identifier. Vital sign data is written into the immediate indicator field, monitoring waveform data is written into the waveform segment reference field, equipment parameter data is written into the equipment status field, test data is written into the test result field, and treatment action records are written into the treatment action field. Then, using the main associated treatment action identifier as an index, records belonging to the same treatment action context are combined to form a treatment action context record block. Finally, the treatment action context record blocks are concatenated in the execution order to generate an observational medical dataset. The observed medical dataset includes at least patient identifier, treatment period identifier, record source field group, time field group, treatment-related location field group, real-time indicator field group, waveform segment field group, equipment status field group, test result field group, and treatment action field group. The output is the observed medical dataset, which is written to the observed data storage area and a dataset index is generated for S2 to read. Anomaly or missing data handling is as follows: when a certain type of record is missing in the same treatment action context record block, the field group is kept empty and a missing identifier is written; when there are multiple candidate values ​​for the same field, the primary value is selected according to the source priority table, and the remaining values ​​are written to the candidate value field; when parallel association states are not resolved in this step, the final resolution is performed according to the consistency relationship between the record type and the object of the treatment action. If the resolution still cannot be achieved, the multiple association identifier is retained and written to the conflict record table for identification and processing in subsequent intervention transmission decomposition steps. Through the above implementation method, multi-source raw medical information is first unified into a raw medical record set, then each record is located to the relative temporal position of the treatment action, and finally combined into an observational medical dataset with treatment action intervention semantics. This allows subsequent steps to directly read observational objects with source identifiers, time references, treatment association positions, and contextual structures instead of dealing with scattered, asynchronous, and context-lacking raw records. This provides a consistent input caliber for subsequent treatment transmission constraint construction, candidate state interpretation branch generation, and GPU ontology state solution, and reduces subsequent judgment biases caused by multi-source data temporal misalignment, missing treatment action associations, and inconsistent record structures. In practical applications, for example, for shock patients in the ICU receiving continuous norepinephrine infusion and simultaneous mechanical ventilation, the system first reads heart rate, arterial pressure, and blood oxygen data from monitoring equipment, tidal volume, positive end-expiratory pressure, and oxygen concentration parameters from the ventilator, blood gas results from the laboratory system, and vasopressor administration records and parameter adjustment records from the infusion pump and medical order system. Then, using the start time of vasopressor infusion and the time of ventilator parameter adjustment as references, each record is located as a pre-association position, a synchronous association position, or a post-association position. These records are then combined into an observational medical dataset according to the context of the treatment action. For example, the blood pressure drop record before vasopressor infusion, the blood pressure rise record during infusion, and the blood oxygen change record after ventilator parameter adjustment will be written into different position layers and form a unified context record block, thus providing complete input for subsequent differentiation between changes in treatment effects and changes in the patient's physical state.

[0018] S2. Based on the action indicators, starting order, duration, and change order of each indicator in the observed medical dataset corresponding to the treatment action record, the intervention transmission decomposition is performed on the observed medical dataset to generate multiple candidate state interpretation branches. The candidate state interpretation branches respectively represent the transmission relationship of different treatment effects and the relationship of different patient ontological state changes. In this implementation, the overall purpose of S2 is: under the condition that treatment action has been intervened, instead of directly regarding the changes in indicators in the observed medical dataset as changes in the patient's ontological state, it first identifies the reachable transmission intervals of the treatment effect on each indicator based on the treatment action records, then decomposes the indicator changes into treatment effect changes and ontological state changes, and constructs multiple candidate state interpretation branches around different decomposition results, ultimately forming an input object for subsequent GPU execution of observation reconstruction, indicator consistency back substitution, and residual bias solution; its principle process is as follows: first, treatment transmission constraints are formed from the observed medical dataset and treatment action records, then indicator change decomposition and combination are performed based on the treatment transmission constraints, and then the process gradually converges to the candidate state interpretation branch set through initial attribution solution, consistency verification, attribution redistribution, branch expansion, and branch merging, and writes the structure fields required for subsequent solution to each candidate state interpretation branch; this implementation process includes the following steps: The purpose of S2-1 is to form a set of treatment transmission constraints for subsequent decomposition and combination. Its mechanism is to first identify the range of influence of treatment actions on the changes of various indicators from the observed medical dataset, and then write the range of influence as a constraint object that can be read for subsequent solution. The input is the observed medical dataset and treatment action records output by S1. The observed medical dataset includes at least patient identifier, collection time identifier, record type identifier, treatment association location identifier, indicator value field and treatment action field. The treatment action records include at least execution time, target, starting order and duration. The processing actions include: first sorting the treatment action records according to the starting order to form a treatment action sequence chain. Next, the corresponding action indicators are read from the treatment action record according to the target. The values ​​of the action indicators are derived from the rule constraint table, which is formed by drug instructions, equipment operating procedures, and clinical treatment rules. For example, vasopressors correspond to blood pressure, heart rate, and perfusion-related indicators, while mechanical ventilation parameter adjustments correspond to blood oxygen, respiratory rate, and ventilator feedback parameters. Then, the order of change of each indicator in the observed medical dataset is read. The order of change is arranged in ascending order according to the acquisition time identifier, and the indicator change segments are identified by the direction, magnitude, and number of consecutive occurrences of the difference between two adjacent samples. Finally, based on the execution time and duration of the treatment action record, the indicator change segments corresponding to the action indicators are matched with the treatment action sequence chain. If the start time of a certain indicator change segment is after the execution time of the corresponding treatment action, within the duration of the process, and within the range of the affected indicator, then the indicator change segment is determined to be within the reachable transmission interval corresponding to the treatment action. If a certain indicator change segment spans multiple treatment action durations, then the primary reachable transmission interval is determined according to the distance between execution times and the consistency of the affected object, and the remaining reachability relationships are written into the extended constraint field. The output is a treatment transmission constraint set, which includes at least the treatment action identifier, the set of affected indicators, the start order constraint, the duration constraint, the start point of the reachable transmission interval, the end point of the reachable transmission interval, and the extended constraint field, and is written into the treatment transmission constraint storage area for S2-2 and S2-21 to read. The abnormal or missing handling is as follows: when the treatment action record is missing a duration, the duration is filled back according to the rule constraint table corresponding to the treatment type; when the affected object cannot be directly mapped to the affected indicator, the affected indicator is supplemented according to the fixed mapping relationship between the treatment type and the equipment source; when a certain affected indicator in the observed medical dataset is missing continuous sampling points, the missing identifier is retained and the indicator change segment is written into the incomplete constraint field, so that the priority of the segment is reduced during subsequent splitting. The purpose of S2-2 is to decompose the changes in indicators mixed in the same observed medical dataset into therapeutic effect changes and ontological state changes, and on this basis, to form multiple candidate state interpretation branches for subsequent solution. Its working mechanism is to put each indicator change segment into a therapeutic effect path or an ontological state path, and then combine the segments under different paths into different state interpretations. The input is the therapeutic transmission constraint set output by S2-1 and the indicator change segments in the observed medical dataset. The processing actions include: first, reading the set of effect indicators, the initial order constraint, and the reachable transmission interval in the therapeutic transmission constraint set; marking the indicator change segments located in the reachable transmission interval and belonging to the set of effect indicators as segments to be assigned; then, sending the segments to be assigned to the therapeutic effect candidate set and the ontological state candidate set respectively; then, performing decomposition and combination according to the temporal continuity relationship, the indicator linkage relationship, and the therapeutic action correspondence relationship. Among them, the temporal continuity relationship is formed based on the difference between adjacent sampling times, the indicator linkage relationship is formed based on the common change records of the same organ system indicators in the same processing time period, and the therapeutic action correspondence relationship is formed based on the mapping between the therapeutic action identifier and the set of effect indicators. For each splitting result, the fragment to be assigned to the treatment action path is written into the treatment action change part, and the fragment to be assigned to the ontology state path is written into the ontology state change part. Then, candidate state interpretation branches are formed according to the combination relationship between the treatment action change part and the ontology state change part. The output is a set of candidate state interpretation branches. This set of candidate state interpretation branches is first written into the intermediate branch buffer, and then read by S2-3, S2-21, and S2-22. The handling of anomalies or missing information is as follows: when the same indicator change fragment satisfies the conditions of both the treatment action path and the ontology state path, the dual attribution identifier is retained and written into the conflict field, and the attribution is subsequently reassigned by S2-22. When the indicator linkage relationship is missing, only the time continuity relationship and the treatment action correspondence relationship are retained for combination. When a branch after splitting lacks the treatment action change part or the ontology state change part, the branch is still retained, but a single-sided branch identifier is written into the branch field for subsequent branch expansion. The purpose of S2-3 is to solidify the splitting results of the candidate state interpretation branch set into computable structured branches. Its working mechanism is to establish a therapeutic effect transmission relationship and an ontological state change relationship for each candidate state interpretation branch, so that the subsequent GPU processing reads not fragment stacking results, but state interpretation objects with internal relationships. The input is the candidate state interpretation branch set output by S2-2. The processing actions include: first, reading the therapeutic effect change part and the ontological state change part of each candidate state interpretation branch; then connecting the indicator change fragments in the therapeutic effect change part according to the therapeutic action identifier, the order of the effect index, and the reachable transmission interval to form a therapeutic effect transmission relationship; at the same time, connecting the indicator change fragments in the ontological state change part according to the sampling time continuity relationship, the index linkage relationship, and the state continuation relationship to form an ontological state change relationship. Subsequently, the therapeutic effect transmission relationship and the ontological state change relationship are written into the relationship field of the same candidate state interpretation branch, and a branch identifier is generated. The output is a set of candidate state interpretation branches with therapeutic effect transmission relationship and ontological state change relationship. This set of candidate state interpretation branches is written into a structured branch table for S3 to read. The abnormal or missing handling is as follows: when the fragment in the therapeutic effect change part cannot form a continuous transmission relationship, the discrete transmission identifier is retained; when the fragment in the ontological state change part has a time break, it is split into multiple continuous state intervals according to the break point and written into the state relationship field respectively; when there are conflicting relationship connections in the same branch, a conflict resolution pending processing identifier is written for S2-24 to record. The purpose of S2-21 is to solve for the initial attribution of changes in each indicator. Its mechanism is to first write the transmission probability between indicator changes and treatment actions into a graph structure, and then use the cumulative result of the violation under the graph structure to determine the initial attribution of each indicator change segment. The input is the treatment transmission constraint set and the indicator change segments in the observed medical dataset. The processing actions include: first, taking each indicator change segment as a node, and taking the relationship between the treatment action and the indicator change segment that satisfies the effect indicator association, the starting order constraint, and the continuous process constraint as an edge, constructing a treatment transmission candidate graph; then, calculating the treatment transmission violation and the ontology state jump violation for each node. The treatment transmission violation is accumulated item by item according to three types of violations: the indicator change segment falling into the unreachable transmission interval, violating the order of treatment actions, and violating the effect indicator association. The ontology state jump violation is accumulated item by item according to three types of violations: the direction reversal within the continuous interval of the same state, the isolated change within the same linked indicator group, and the break between adjacent state segments. Subsequently, the treatment conduction violation and ontology state transition violation are written together into the attribution determination field. For each node, the attribution determination is compared with that when it is assigned to the treatment action path and when it is assigned to the ontology state path. The path with the smaller determination is taken as the initial attribution path of the node. If the attribution determination of the two paths is the same, the initial attribution of the two paths is retained. The output is the initial split result of the index change. This result includes at least the index change segment identifier, the initial attribution path, the treatment conduction violation, and the ontology state transition violation, and is written into the initial split result table for S2-22 to read. The abnormal or missing handling is as follows: when an index change segment has no connectable edges, the segment is directly written into the ontology state candidate path and an isolated segment identifier is written. When a closed loop connection appears in the treatment conduction candidate graph, the reverse edge is deleted according to the order of treatment actions, and the attribution determination is recalculated. The purpose of S2-22 is to perform consistency checks and conflict resolution on the initial splitting results. Its mechanism is to verify whether the initial assignment is valid from three perspectives: time sequence, indicator linkage, and treatment response closure, and to correct invalid parts through assignment reallocation. The inputs are the initial splitting results of indicator changes, treatment transmission constraint set, and observed medical dataset output by S2-21. The processing actions include: first, performing a time sequence consistency check for each candidate splitting combination to determine whether the indicator change segments within the same path satisfy the consistency between the sampling time order and the treatment action order; then performing an indicator linkage consistency check to determine whether the change direction and change time period in the same linkage indicator group remain corresponding; and finally performing a treatment response closure consistency check to determine whether the indicator change segments assigned to the treatment action path form a complete response chain of start, transmission, and end during the duration of the treatment action. If any check fails, the corresponding indicator change segment is written into the conflict item set, and the attribution is reassigned based on the temporal proximity, linkage, and treatment action correspondence between the conflict item and its adjacent segments. That is, it is moved out of the current attribution path and written into another attribution path, and then the three types of consistency checks are re-executed. When the number of items that pass the check increases and the number of conflict items decreases after reassignment, the reassignment result is retained. Then, a combination confidence result is generated based on the number of items that pass the check and the number of conflict items in each candidate split combination. The combination confidence result is composed of binary counts, that is, the number of items that pass the check and the number of conflict items are written into the field pair. The output is the candidate split combination set and the combination confidence result corresponding to each candidate split combination, and it is written into the combination result table for S2-23 to read. The abnormal or missing handling is as follows: when the linkage indicator group is missing some indicators, only the existing indicators are checked for linkage consistency, and the linkage missing item flag is written into the combination confidence result. When there are still conflict items after the attribution reassignment, the candidate split combination is retained, but an unclosed conflict flag is written for removal during subsequent branch merging. The purpose of S2-23 is to converge the candidate state interpretation branch set from the candidate splitting combination set. Its mechanism is to supplement the interpretation path that has not yet been covered by branch expansion, and then eliminate the duplicate or clearly contained branches by branch merging, and use the stability of the combination result as the stopping condition. The input is the candidate splitting combination set output by S2-22 and the combination confidence result corresponding to each candidate splitting combination. The processing actions include: first, reading the unassigned segments, double-assigned segments and unclosed conflict terms in each candidate splitting combination; if there are unassigned segments, then expand the segments into the treatment action path and the ontology state path respectively to form new extended branches; if there are double-assigned segments, then expand them into two branches according to the two types of assignment respectively. After expansion, the changes in therapeutic effects and ontological state between branches are compared. When the fragment sets, path connections, and conflict resolution results of two branches are consistent, branch merging is performed and one branch is retained. When the fragment set of one branch is completely contained in another branch and the number of passed items in the combined confidence result is small while the number of conflict items is large, the contained branch is deleted. The retention condition is: the number of passed items in the combined confidence result of this round is higher than that of the same source branch in the previous round, and the number of conflict items is lower than that of the same source branch in the previous round. The stopping condition is: in two consecutive rounds of branch processing, the combined confidence result field of the same source branch remains unchanged, and the value of the number of conflict items field remains unchanged. The output is a set of candidate state interpretation branches, which is written to the branch convergence result table for S2-24 to read. The handling of anomalies or missing items is as follows: when the number of branches after branch expansion exceeds the GPU's single-round reading capacity, they are written in batches in descending order of the number of passed items and ascending order of the number of conflict items. When there are no same source branches in two consecutive rounds of processing, the result of the current round is used as the stopping result and output directly. The purpose of S2-24 is to organize the candidate state interpretation branch set into the standard input for subsequent ontology state solution processing. Its mechanism is to supplement the fields required for subsequent GPU reading for each candidate state interpretation branch, and to write the splitting results, relation results, and conflict resolution results into the branch record. The input is the candidate state interpretation branch set output by S2-23. The processing actions include: reading the treatment effect change part, ontology state change part, treatment effect transmission relationship, ontology state change relationship, and conflict resolution result for each candidate state interpretation branch; writing the treatment effect change part into the treatment effect change field, writing the ontology state change part into the ontology state change field, encoding the treatment effect transmission relationship and ontology state change relationship together into a transmission path identifier, and encoding the conflict resolution result into a conflict resolution identifier. Simultaneously, branch identifier, patient identifier, and processing time period identifier are written to form a complete candidate state interpretation branch record; the output is a set of candidate state interpretation branches for subsequent ontology state solution processing to read. This set of candidate state interpretation branches is written to the GPU branch input area, and a branch index table is generated for S3-1 to load according to the branch identifier; the abnormal or missing handling is as follows: when a candidate state interpretation branch lacks conflict resolution results, a conflict-free identifier is written; when a candidate state interpretation branch lacks the ontology state change part or the treatment effect change part, the missing field is retained and written to the unilateral branch identifier for identification by the observation reconstruction unit in S3; Through the above implementation method, a set of treatment conduction constraints is first formed, then the index change is split and combined, and the candidate state interpretation branch set is gradually converged through the treatment conduction candidate graph, initial attribution solution, consistency check, attribution redistribution, branch expansion and branch merging. Finally, each candidate state interpretation branch is organized into a structured input object with treatment effect change part, ontological state change part, conduction path identifier and conflict resolution identifier. In this way, the GPU no longer faces the unexplained original index change, but directly reads multiple candidate state interpretation branches formed by separating treatment effect change and ontological state change, thereby providing a unified input caliber for observation result reconstruction, index consistency back substitution and residual bias solution, and reducing the risk of misjudging the treatment response as a change in the patient's ontological state when the treatment action has been intervened. In practical applications: For example, for shock patients in the ICU receiving continuous infusion of vasopressors and simultaneous adjustment of ventilator parameters, the system first forms a treatment conduction constraint set based on the blood pressure and heart rate effects of the vasopressors and the corresponding blood oxygen and respiratory rate effects of the ventilator parameters. Then, it performs splitting and combination on the segments of changes in indicators such as blood pressure recovery, heart rate fluctuation, blood oxygen improvement, and respiratory parameter changes after infusion. If a blood pressure recovery segment is within the continuous process of vasopressor administration and is consistent with the blood pressure effect indicator, it is preferentially written into the treatment effect change part. If lactate continues to rise during the same period and is not within the reachable conduction interval of the vasopressor, it is written into the ontology state change part. Subsequently, the system performs time sequence consistency verification, indicator linkage consistency verification, and treatment response closure consistency verification on the multiple candidate splitting combinations, and reassigns conflicting segments. Finally, multiple candidate state interpretation branches are obtained. For example, one branch interprets blood pressure recovery as the effect of vasopressors and lactate rise as the continuous deterioration of circulatory perfusion. Another branch interprets blood pressure recovery and lactate rise as short-term fluctuations after treatment. Subsequently, the GPU performs observation reconstruction and ontology state solution on these different interpretation branches respectively.

[0019] S3. Load the candidate state interpretation branches to the GPU, perform observation result reconstruction, index consistency back substitution and residual bias solution for each candidate state interpretation branch, and output the reconstruction bias result and ontology state solution result for each candidate state interpretation branch. In this implementation, the overall purpose of S3 is to: feed the candidate state interpretation branch set generated in the preceding steps into the GPU for parallel state inverse solution, so that each candidate state interpretation branch can complete the reconstruction of observation results, back-substitution of index consistency, and solution of remaining bias under the same observation medical dataset constraints, thereby outputting the reconstruction bias results and ontology state solution results required for subsequent branch competition determination; the principle process is as follows: first, load the candidate state interpretation branches into the GPU according to the branch identifier and establish observation reconstruction units, then perform observation result reconstruction and index consistency back-substitution on each observation reconstruction unit, and then solve the remaining bias results and ontology state solution results based on the corresponding differences between the branch reconstruction observation results and the observation medical dataset, and finally organize the results of each branch into a standard result set that can be read subsequently; this implementation process includes the following steps: The purpose of S3-1 is to convert the candidate state interpretation branch set into an observation reconstruction model set that can be processed in parallel by the GPU. Its working mechanism is to map the treatment effect change part and the ontological state change part in each candidate state interpretation branch to the observation reconstruction unit within the same branch, so that the subsequent GPU processing faces a structured reconstruction object rather than discrete fragments. The input is the candidate state interpretation branch set corresponding to the target patient and the observation medical dataset. The candidate state interpretation branch set includes at least branch identifier, patient identifier, processing time identifier, treatment effect change part, ontological state change part, transmission path identifier, and conflict resolution identifier. The observation medical dataset includes at least collection time identifier, indicator value field, treatment association location identifier, and treatment action field. The processing steps include: first, sorting the candidate state interpretation branch set by branch identifier, and then dividing the candidate state interpretation branch set into several branch groups according to the number of branches that can be loaded in a single GPU round. The number of branches that can be loaded in a single GPU round is determined based on the GPU memory capacity, the length of a single branch field, and the single-thread block allocation rules, and the values ​​are derived from the device operation configuration table; then, loading each branch group into the GPU memory area and assigning an independent branch index to each candidate state interpretation branch; subsequently, reading the therapeutic effect change part and the ontological state change part in each candidate state interpretation branch, establishing a time alignment matrix according to the acquisition time identifier and the indicator identifier, and then writing the therapeutic effect change part into the treatment reconstruction channel and the ontological state change part into the ontological reconstruction channel according to the transmission path identifier, thereby forming an observation reconstruction unit; for multiple indicator change segments in the same candidate state interpretation branch, establishing an intra-branch index table according to the ascending order of sampling time and the order of indicator category; and then writing the original indicator sequence corresponding to the branch in the observed medical dataset into the control input area to generate the observation reconstruction model set corresponding to each candidate state interpretation branch. The output is an observation reconstruction model set, which is written to the GPU model cache and a model index table is generated synchronously for S3-2 to read. Anomaly or missing data handling is as follows: when a candidate state explanation branch lacks a therapeutic effect change portion, an empty channel identifier is written to the therapeutic reconstruction channel; when a ontology state change portion is missing, an empty channel identifier is written to the ontology reconstruction channel; when the number of branch groups in a single round exceeds the GPU memory capacity, branches are loaded in batches according to their identifiers, and unloaded branches are written to the pending processing queue; when a certain indicator in the observed medical dataset lacks a sample value during the processing period, linear point filling is performed based on two adjacent valid sample values, and a point filling identifier is written for identification during subsequent bias solution. The purpose of S3-2 is to perform branch observation reconstruction, ontology state back substitution, and bias solution on each observation reconstruction model set. Its mechanism involves first synthesizing the branch reconstruction observation results using the treatment reconstruction channel and the ontology reconstruction channel, then comparing these results item by item with the observed medical dataset. From the comparison differences, the remaining bias results not explained by the current candidate state's interpretation branch are extracted, and the corresponding ontology state solution results are deduced. The inputs are the observation reconstruction model set and the observed medical dataset output from S3-1. The processing actions include: first, synthesizing the treatment reconstruction channel and ontology reconstruction channel in each observation reconstruction model set item by item according to the acquisition time identifier and indicator identifier; if the same indicator at the same acquisition time simultaneously exhibits a treatment effect variable... The values ​​of change in the ontology state and the values ​​of change in the ontology state are calculated in the order of ontology change first and then the superposition of therapeutic effects to form the branch reconstruction observation results. Then, the index consistency back substitution is performed on the branch reconstruction observation results. Index consistency back substitution means: reading multiple reconstructed index values ​​in the same linkage index group, judging whether these reconstructed index values ​​meet the linkage relationship constraints. The linkage relationship constraints come from the rule constraint table and the clinical index association table, such as the linkage relationship between blood pressure, heart rate and changes in vasoactive drugs. If a certain reconstructed index value causes the linkage relationship to be invalid, the reconstructed index value is written back to the ontology reconstruction channel and the ontology change values ​​of adjacent linkage indicators are recalculated until no new linkage conflicts occur in the same linkage index group. Subsequently, the branch reconstruction observation results after back-substitution are compared item by item with the observed medical dataset. The difference direction, absolute value, and position are calculated for each acquisition time and each indicator. Then, the differences are summarized according to the indicator dimension and the time dimension to form the residual bias result, which includes at least the bias distribution, bias direction, and residual bias position. Next, the ontology change values ​​at each time point are read from the ontology reconstruction channel after back-substitution. An ontology state sequence is generated according to the temporal continuity relationship and the indicator linkage relationship. Then, the ontology state solution is obtained based on the change direction, duration, and consistency relationship of the linkage indicator changes between adjacent state segments. The ontology state solution includes at least the state category, state sequence, and state transition relationship. The output consists of the continuous interval of the state; the output is the remaining deviation result and the ontology state solution result corresponding to each candidate state interpretation branch. These two types of results are written to the GPU result cache area, and a result comparison table is established with the branch identifier for S3-3 to read; the abnormal or missing handling is as follows: when a certain linkage indicator group is missing some indicator values, only the existing indicators are back-substituted for indicator consistency, and the linkage missing item identifier is written; when the branch reconstruction observation result and the observed medical dataset cannot form a correspondence for a certain indicator at a certain moment, the position is directly written to the deviation remaining position field; when a circular back-substitution occurs during the back-substitution process, the value of the same indicator remains unchanged in two consecutive back-substitutions as the stopping condition, and the indicator is written to the back-substitution lock field; The purpose of S3-3 is to organize the remaining deviation results and the ontology state solution results output by each candidate state interpretation branch into a standard result set that can be directly read in subsequent branch competition decisions. Its working mechanism is to rearrange the branch-level results formed inside the GPU according to a unified field structure and write them into the result table, so that subsequent decision steps can read them one by one according to the branch identifier. The input is the remaining deviation results and ontology state solution results corresponding to each candidate state interpretation branch output by S3-2. The processing actions include: first, reading the remaining deviation results corresponding to each branch according to the branch identifier, extracting the deviation distribution, deviation direction and remaining deviation position, and generating a reconstructed deviation result record; then, reading the ontology state solution results corresponding to each branch according to the same branch identifier, extracting the state category, state sequence, state transition relationship and state continuous interval, and generating an ontology state solution result record. Subsequently, the two types of records under the same branch identifier are associated and written into the branch result table, and the patient identifier, processing time identifier, and result generation time are added to each record; then, all branch results are sorted in ascending order by branch identifier to generate a reconstruction deviation result set and an ontology state solution result set for subsequent branch competition judgment; the output is the reconstruction deviation result set and the ontology state solution result set, which are written into the branch judgment input area respectively, and a result index table is generated for S4 to read; the abnormal or missing handling is as follows: when a branch only outputs the remaining deviation result and does not output the ontology state solution result, the reconstruction deviation result record of the branch is retained and a missing identifier is written into the ontology state solution result field; when a branch only outputs the ontology state solution result and does not output the remaining deviation result, the ontology state solution result record of the branch is retained and a missing identifier is written into the reconstruction deviation result field; when there are duplicate result records in the same branch, the next record is retained according to the result generation time, and the previous record is written into the history result area; Through the above implementation method, the candidate state interpretation branch set is first loaded into the GPU according to the branch identifier and an observation reconstruction model set is established. Then, the observation results are reconstructed, the index consistency is back-substituted, and the remaining bias is solved. Finally, the reconstructed bias result set and the ontology state solution result set are organized. This allows the subsequent branch competition judgment to no longer directly face the unsolved candidate state interpretation branches, but directly read the bias results and ontology state results formed by each candidate state interpretation branch under the same observation medical dataset constraint. In this way, whether the change in treatment effect and the change in ontology state can jointly explain the current observation results can be transformed into comparable and judgmentable branch-level results, and the bias caused by directly determining the patient's ontology state based on a single rule filtering is reduced. In practical applications: For example, for shock patients in the ICU receiving continuous infusion of vasopressors and simultaneous adjustment of ventilator parameters, the system first loads candidate state interpretation branches—attributing blood pressure recovery to the effect of vasopressors and lactate elevation to deterioration of perfusion—as well as candidate state interpretation branches attributing both blood pressure recovery and lactate elevation to short-term fluctuations after treatment, onto the GPU. Then, an observation reconstruction unit is established for each candidate state interpretation branch, synthesizing the changes from vasopressor effects and changes in circulatory perfusion ontology into a branch reconstruction observation result. This result is then compared item by item with the blood pressure, heart rate, lactate, and respiratory parameters in the actual observed medical dataset. If a candidate state interpretation branch still has a continuous deviation remaining position in the blood pressure and lactate linkage relationship after back-substitution, then that candidate state interpretation branch will output the corresponding remaining deviation result. If another candidate state interpretation branch can form a continuous low-perfusion ontology state sequence after back-substitution, then it will output the corresponding ontology state solution result. Subsequent branch competition determination is based on these results to further determine the target state interpretation branch.

[0020] S4. Based on the reconstruction deviation results and ontology state solution results corresponding to each candidate state interpretation branch, perform branch competition judgment, determine the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch. In this implementation, the overall purpose of S4 is to perform a joint judgment on the reconstruction deviation results and ontology state solution results corresponding to each candidate state interpretation branch output by S3, filter out candidate state interpretation branches that cannot simultaneously explain the current observed medical dataset and treatment action records, retain candidate state interpretation branches that can form a closed interpretation, determine the target state interpretation branch from among them, and output the target ontology state result; its principle flow is as follows: first, generate a branch judgment record for each candidate state interpretation branch, then perform reconstruction closure judgment and state establishment judgment based on the branch judgment record, and finally perform retention, elimination, and target branch confirmation based on the judgment results; this implementation process includes the following steps: The purpose of S4-1 is to organize the deviation information and state information required for subsequent judgment into a unified record. Its mechanism is to merge the reconstruction deviation results and ontology state solution results corresponding to the same candidate state interpretation branch according to the branch identifier, forming a branch judgment record that can be directly used for judgment. The input quantities are the candidate state interpretation branch set corresponding to the target patient, the reconstruction deviation results corresponding to each candidate state interpretation branch, and the ontology state solution results corresponding to each candidate state interpretation branch. Among them, the reconstruction deviation results include at least the deviation distribution, deviation direction, and deviation remaining position, and the ontology state solution results include at least the state sequence, state transition relationship, and state continuity interval. The processing actions include: first, reading the treatment effect change part, ontology state change part, and conduction path identifier in the candidate state interpretation branch set according to the branch identifier; then reading the corresponding reconstruction deviation results and ontology state solution results according to the same branch identifier; and then writing the deviation distribution, deviation direction, deviation remaining position, state sequence, state transition relationship, and state continuity interval into the same branch judgment record. Then, patient identifiers, processing time identifiers, and treatment action identifiers are added to form a branch decision record set corresponding to each candidate state interpretation branch; the output is the branch decision record set, which is written to the branch decision buffer for S4-2 to read; the abnormal or missing handling is as follows: when a candidate state interpretation branch lacks reconstruction deviation results, a missing identifier is written to the deviation field and the candidate state interpretation branch is marked as a branch to be removed; when a candidate state interpretation branch lacks ontology state solution results, a missing identifier is written to the state field and the candidate state interpretation branch is marked as a branch to be removed; when there are multiple result records for the same branch identifier, the last record is retained according to the result generation time, and the remaining records are written to the history record area; The purpose of S4-2 is to determine whether each candidate state interpretation branch can simultaneously explain the deviation result and the state result. Its mechanism is to perform judgments from two directions: the closure of the reconstruction result and the validity of the state evolution, and output the branch judgment result. The inputs are the branch judgment record set, the observed medical dataset, and the treatment action record output by S4-1. The processing actions include: first, performing a reconstruction closure judgment on each candidate state interpretation branch, specifically reading the deviation distribution, deviation direction, and remaining deviation position of the candidate state interpretation branch, and comparing them item by item with the corresponding treatment effect change part and the ontological state change part of the candidate state interpretation branch; if the remaining deviation position falls outside the treatment effect change part and the ontological state change part, or the deviation direction is opposite to the corresponding change direction, it is recorded as reconstruction closure not valid; if the remaining deviation positions can all be mapped to the unexplained segments of the candidate state interpretation branch, and the deviation direction is consistent with the corresponding change direction, it is recorded as reconstruction closure valid. Then, a state validity determination is performed on the same candidate state interpretation branch. Specifically, the state sequence, state transition relationship, and state continuous interval are read and compared with the order of indicator changes in the observed medical dataset and the starting order and duration of treatment action records item by item. If the order of the state sequence is inconsistent with the order of indicator changes, or the state transition relationship crosses the boundary of the treatment action duration without a corresponding state connection, or the state continuous interval cannot be aligned with the continuous change segment in the observation period, it is recorded as a state validity failure. If all three comparisons are valid, it is recorded as a state validity success. The output is the branch determination result corresponding to each candidate state interpretation branch. The branch determination result includes at least the reconstruction closure determination result and the state validity determination result, and is written into the branch determination result table for S4-3 to read. The abnormal or missing items are handled as follows: when there are missing items in the deviation distribution, only the existing deviation items are reconstructed for closure determination, and the deviation missing item identifier is written into the determination result. When the state continuous interval is broken, it is split according to the break point and the state validity determination is performed separately. When there are multiple unaligned segments in the same candidate state interpretation branch, the number of unaligned segments is accumulated and written into the determination result field. The purpose of S4-3 is to determine the target state interpretation branch from all candidate state interpretation branches. Its mechanism involves first retaining and eliminating branches based on the branch decision results, and then confirming the unique target state interpretation branch based on the number of remaining deviation positions and the number of continuous state intervals in the retained branches. The inputs are the branch decision results, reconstruction deviation results, and ontology state solution results corresponding to each candidate state interpretation branch output by S4-2. The processing actions include: first, retaining candidate state interpretation branches that simultaneously satisfy the reconstruction closure criterion and the state establishment criterion, and eliminating candidate state interpretation branches that do not simultaneously satisfy both criterion; then, reading the number of remaining deviation positions in the reconstruction deviation results and the number of continuous state intervals in the ontology state solution results for each retained candidate state interpretation branch; subsequently, determining the correspondence between the number of remaining deviation positions and the number of continuous state intervals. A consistent correspondence means that each continuous state interval corresponds to a remaining deviation position segment, and the number of remaining deviation position segments is the same as the number of continuous state intervals; the candidate state interpretation branch that satisfies this correspondence is determined as the target state interpretation branch. If multiple branches among the retained candidate state interpretation branches simultaneously satisfy the corresponding relationship, then the total number of remaining deviation positions of each branch is compared, and the candidate state interpretation branch with the smaller total number of remaining deviation positions is retained; if the total number of remaining deviation positions is still the same, then the number of alignments between the continuous interval of the state and the continuous change segment of the observed medical dataset is compared, and the candidate state interpretation branch with the larger number of alignments is retained; the output is the target state interpretation branch and its corresponding target ontology state result, which are written to the target state result table for S5 to read; the abnormal or missing handling is as follows: when the retained candidate state interpretation branches are empty, return to S2 to regenerate the candidate state interpretation branch set; when only one candidate state interpretation branch is retained, directly determine the candidate state interpretation branch as the target state interpretation branch; when the retained candidate state interpretation branches have missing numbers of remaining deviation positions or missing numbers of continuous state intervals, directly perform elimination; Through the above implementation method, the deviation results and state results of each candidate state interpretation branch are first organized into a branch decision record, then the reconstruction closure decision and state validity decision are performed, and finally the target state interpretation branch is determined according to the consistency rule of the decision result and the correspondence relationship, thereby converging the multi-branch results output by the preceding GPU into a unique target ontology state result; this can avoid directly determining the patient ontology state based on a single deviation result or a single state result, and reduce the situation where invalid state interpretation branches are mistakenly retained after treatment intervention. In practical applications: For example, for a shock patient receiving continuous vasopressor infusion and whose ventilator parameters are adjusted, the system first reads the reconstruction deviation results and ontology state solutions corresponding to two candidate state interpretation branches. The remaining deviation positions corresponding to one candidate state interpretation branch are mainly concentrated in the segment of continuously rising lactate, and the state sequence shows a continuous decrease in circulatory perfusion. The remaining deviation positions corresponding to the other candidate state interpretation branch are scattered across multiple segments of blood pressure, blood oxygen, and heart rate, and the continuous interval of the state cannot be aligned with the continuous change segments in the observed medical dataset. Based on this, the system determines that the former candidate state interpretation branch satisfies both the reconstruction closure criterion and the state establishment criterion, while the latter candidate state interpretation branch does not satisfy the state establishment criterion. Therefore, the system retains the former candidate state interpretation branch and outputs its corresponding target ontology state result for subsequent medical processing result generation.

[0021] S5. Generate the medical treatment result corresponding to the target patient based on the target ontology state result, and write the medical treatment result into the result reading path; In this embodiment, the overall purpose of S5 is to convert the target ontology state result output by S4 into a medical processing result that can be directly read by subsequent business processes, and to write the medical processing result into the corresponding result reading path, so that the target ontology state result completes the conversion from state determination result to business processing result. The principle process is as follows: first, medical processing mapping is performed based on the state category, state continuity interval, and state transition relationship in the target ontology state result to form the medical processing result corresponding to the target patient; then, the medical processing result is written into the corresponding result reading path based on the processing type identifier and patient identifier, and a result writing record is generated. This implementation process includes the following steps: The purpose of S5-1 is to transform the target ontology state result into an executable medical treatment result. Its working mechanism is to map the state determination information to the corresponding treatment type, treatment content, and treatment level. The input is the target ontology state result corresponding to the target patient. The target ontology state result includes at least the patient identifier, state category, state continuity interval, and state transition relationship. The processing actions include: first, reading the state category and looking up the corresponding treatment type in the medical treatment mapping table based on the state category. The values ​​of the medical treatment mapping table are derived from the rule constraint table, which is formed by clinical treatment specifications, departmental treatment rules, and system configuration rules; then, reading the state continuity interval, determining the duration of the same state within the target treatment period, and writing the state continuity interval into the treatment period field. Subsequently, the state transition relationship is read to determine whether the state has transitioned from low risk to high risk, from a single anomaly to a compound anomaly, or from stable to fluctuating, and the state transition relationship is written into the processing intensity field. Then, the medical processing result is generated by combining the state category, the continuous state interval, and the state transition relationship. The medical processing result includes at least the processing type identifier, the processing content field, the processing time period field, and the processing intensity field. The output is the medical processing result corresponding to the target patient, which is written into the medical processing result cache for S5-2 to read. The anomaly or missing information handling is as follows: when the state category is missing, the target ontology state result is written into the pending review queue; when the continuous state interval is missing, the target processing time period is written into the medical processing result as the processing time period field; when the state transition relationship is missing, the processing intensity field is written according to the basic processing intensity corresponding to the state category, and a transition missing identifier is written. The purpose of S5-2 is to write medical processing results into a specified result reading path. Its mechanism is to locate and write the medical processing results based on the processing type identifier and patient identifier, so that subsequent early warning processing, auxiliary judgment processing, or continuous monitoring processing can be directly read according to the path. The input is the medical processing results output by S5-1, which includes at least the processing type identifier, patient identifier, processing content field, processing time field, and processing intensity field. The processing actions include: first, reading the processing type identifier and looking up the corresponding result reading path in the result path mapping table. The values ​​of the result path mapping table are derived from the system configuration rules and include at least the early warning processing path, auxiliary judgment path, and continuous monitoring path; then, reading the patient identifier and combining the patient identifier with the result reading path to generate the target writing address. The medical processing result is then written to the target write address, and the result generation time, write time, and result version identifier are written synchronously. A result write record is then generated, which includes at least the patient identifier, processing type identifier, result read path, write time, and write status. The output is the result write record corresponding to the target patient, which is written to the log table for subsequent traceability. Anomaly or missing information is handled as follows: when the processing type identifier cannot be mapped to the result read path, the medical processing result is written to the default review path; when the patient identifier is missing, writing is aborted and a write failure record is generated; when the target write address returns a conflict status, the old version is overwritten or a new version is appended according to the result version identifier, with the overwrite or append rule derived from the system configuration rules. Through the above implementation method, the target ontology state result is first mapped to the medical processing result, and then written into the corresponding result reading path according to the processing type identifier and the patient identifier, thereby completing the closed loop landing from the state solution result to the business processing result; this allows the target ontology state result output by the previous steps to directly enter the subsequent business link, and ensures that different types of medical processing results enter the corresponding reading channel, avoiding the disconnect between the state result and the processing path. In practical applications: For example, for a shock patient in the ICU receiving continuous vasopressor infusion, the system determines the target ontology state result as continuously deteriorating circulatory perfusion in S4, and reads the continuous state interval and state transition relationship in the target ontology state result; then, in the medical processing mapping table, the state category is mapped to the early warning processing type, the continuous deterioration is mapped to the continuous monitoring period, and the transition from a single abnormality to a compound abnormality is mapped to the high-intensity processing field, generating a medical processing result; then, according to the early warning processing type and the patient identifier, the medical processing result is written into the early warning processing path, and the result is simultaneously generated and written to the record for direct reading by subsequent early warning services.

[0022] Furthermore, the present invention also includes a computer-based medical data processing system, the system comprising an observation construction module, a transmission decomposition module, a state solution module, a competition determination module, and a result writing module: The observation construction module is used to acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and to perform time correlation and processing through the computer system to output the observation medical dataset; The transmission decomposition module performs intervention transmission decomposition on the observed medical dataset based on the action indicators, starting sequence, duration, and change sequence of each indicator in the observed medical dataset corresponding to the treatment action record, generating multiple candidate state interpretation branches. The candidate state interpretation branches respectively represent the transmission relationship of different treatment effects and the relationship of different patient ontological state changes. The state solution module is used to load candidate state interpretation branches into the GPU, perform observation result reconstruction, index consistency back substitution and residual deviation solution on each candidate state interpretation branch, and output the reconstruction deviation result and ontology state solution result corresponding to each candidate state interpretation branch. The competition determination module is used to perform branch competition determination based on the reconstruction deviation results and ontology state solution results corresponding to each candidate state interpretation branch, determine the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch. The result writing module is used to generate the medical treatment result corresponding to the target patient based on the target ontology state result, and write the medical treatment result to the result reading path.

[0023] Working principle: This solution first places the patient's monitoring data, test data, equipment parameters, and treatment records on the same timeline, then determines which changes are caused by treatment and which changes are more likely to be changes in the patient's condition. Based on this, multiple state interpretations are generated, and with the help of GPU, these interpretations are simultaneously verified to see which one best matches all current observations. Finally, the one that best reflects the patient's true state is selected, and the corresponding medical processing result is generated and written into the subsequent reading path. The core is not to directly look at what the current data is, but to first understand why the data has changed in this way, and then deduce the patient's true condition. For example, in the intensive care unit, a patient's blood pressure rises after taking vasopressors, but lactate continues to rise. Traditional approaches tend to directly interpret the rise in blood pressure as an improvement in the patient's condition, while this approach takes into account the effect of vasopressors: on the one hand, the rise in blood pressure is interpreted as a change brought about by treatment, while on the other hand, indicators such as lactate are checked to see if they still indicate a deterioration in the patient's condition; then, the interpretation that is most consistent with all the data is selected from multiple interpretations, and the result that is closer to the actual situation is given, such as the patient's circulatory perfusion is still deteriorating, and an early warning or continuous monitoring is triggered accordingly.

[0024] The above description is merely 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 medical data processing method based on a computer system, characterized in that, include: S1. Acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and perform time correlation and processing through the computer system to output the observation medical dataset; S2. Based on the action indicators, starting order, duration, and change order of each indicator in the observed medical dataset corresponding to the treatment action record, the intervention transmission decomposition is performed on the observed medical dataset to generate multiple candidate state interpretation branches. The candidate state interpretation branches respectively represent the transmission relationship of different treatment effects and the relationship of different patient ontological state changes. S3. Load the candidate state interpretation branches to the GPU, perform observation result reconstruction, index consistency back substitution and residual bias solution for each candidate state interpretation branch, and output the reconstruction bias result and ontology state solution result for each candidate state interpretation branch. S4. Based on the reconstruction deviation results and ontology state solution results corresponding to each candidate state interpretation branch, perform branch competition judgment, determine the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch. S5. Generate the medical treatment result corresponding to the target patient based on the target ontology state result, and write the medical treatment result into the result reading path.

2. The medical data processing method based on a computer system according to claim 1, characterized in that: S1 includes: S1-1. Acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and extract the corresponding data source identifier, acquisition time identifier and record type identifier respectively, and output the original medical record set; S1-2. Based on the execution time, target, and execution order in the treatment action record, perform time-series correlation and positioning of each record in the original medical record set, determine the preceding, synchronous, or subsequent correlation position of each record relative to each treatment action record, and output a medical record set with treatment correlation position identifiers. S1-3. Organize and combine the medical record set with treatment-related location identifiers according to the target treatment period and the execution order of the treatment-related location identifiers to generate an observational medical dataset that characterizes the observational state of the target patient under the intervention of treatment actions.

3. The medical data processing method based on a computer system according to claim 2, characterized in that: S2 includes: S2-1. Based on the observational medical dataset and treatment action records corresponding to the target patient, and based on the effect indicators, starting order and duration of the treatment action records, combined with the change order of each indicator in the observational medical dataset, identify the reachable transmission interval of each treatment action on the change of each indicator, and output the treatment transmission constraint set. S2-2. Based on the treatment transmission constraint set, split and combine the changes of each indicator in the observed medical dataset, divide the changes of each indicator into the treatment effect change part and the ontology state change part, and generate multiple candidate state interpretation branches according to the combination relationship between different treatment effect change parts and different ontology state change parts, and output the candidate state interpretation branch set. S2-3. Establish therapeutic effect transmission relationship and ontological state change relationship for each candidate state interpretation branch in the candidate state interpretation branch set, and generate multiple candidate state interpretation branches that respectively characterize different therapeutic effect transmission relationships and different patient ontological state change relationships.

4. A medical data processing method based on a computer system according to claim 3, characterized in that: The process of outputting the candidate state interpretation branch set in S2-2 also includes: S2-21. Based on the correlation of the action indicators, the initial sequence constraint and the continuous process constraint in the treatment transmission constraint set, construct a treatment transmission candidate graph for the changes of each indicator in the observed medical dataset. The treatment transmission violation quantity and the ontology state jump violation quantity together constitute the attribution determination quantity. Under the constraints of the sequential relationship of treatment actions and the reachability relationship of indicator changes, perform initial attribution solution for each indicator change and output the initial split result of indicator change. S2-22. Based on the initial splitting results of indicator changes, perform time sequence consistency checks, indicator linkage consistency checks, and treatment response closure consistency checks on each indicator change under different treatment action paths and different ontology state evolution paths. Perform attribution reassignment on indicator changes with verification conflicts. Based on the number of verification-passing items and the number of verification-conflicting items after reassignment, form a combined confidence result, and output the candidate splitting combination set and the combined confidence result corresponding to each candidate splitting combination.

5. A medical data processing method based on a computer system according to claim 4, characterized in that: The process of outputting the candidate state interpretation branch set in S2-2 also includes: S2-23. Based on the candidate splitting and combination set and the combination confidence result, perform branch expansion and branch merging on each candidate splitting and combination, retain the splitting and combination whose combination confidence result is higher than the previous round and whose number of verification conflict items decreases, and stop processing when the combination confidence result no longer changes and the number of verification conflict items no longer decreases in two consecutive rounds of branch processing, and output the candidate state explanation branch set. S2-24. Write the treatment effect change part, ontology state change part, conduction path identifier and conflict resolution identifier into each candidate state interpretation branch in the candidate state interpretation branch set, respectively, to generate a candidate state interpretation branch set for subsequent ontology state solution processing.

6. A medical data processing method based on a computer system according to claim 5, characterized in that: S3 includes: S3-1. Based on the candidate state interpretation branch set and the observation medical dataset corresponding to the target patient, load the candidate state interpretation branch set onto the GPU according to the branch identifier, establish observation reconstruction units corresponding to the treatment effect change part and the ontological state change part for each candidate state interpretation branch, and output the observation reconstruction model set corresponding to each candidate state interpretation branch. S3-2. Perform observation result reconstruction and index consistency back substitution on each observation reconstruction model set respectively. Combine the therapeutic effect change part and ontology state change part in each candidate state interpretation branch into branch reconstruction observation results. Based on the corresponding difference between the branch reconstruction observation results and the observed medical dataset, solve the residual bias results and ontology state solution results corresponding to each candidate state interpretation branch. S3-3. Perform branch result organization on the remaining deviation results and ontology state solution results corresponding to each candidate state interpretation branch, and generate a set of reconstruction deviation results and ontology state solution results for subsequent branch competition determination.

7. A medical data processing method based on a computer system according to claim 6, characterized in that: S4 includes: S4-1. Obtain the candidate state interpretation branch set corresponding to the target patient, the reconstruction deviation result corresponding to each candidate state interpretation branch, and the ontology state solution result corresponding to each candidate state interpretation branch. For each candidate state interpretation branch, read the deviation distribution, deviation direction, and remaining deviation position in the reconstruction deviation result, as well as the state sequence, state transition relationship, and state continuity interval in the ontology state solution result, and generate the branch decision record set corresponding to each candidate state interpretation branch. S4-2. Based on the branch decision record set, perform reconstruction decision and state establishment decision for each candidate state interpretation branch. The reconstruction closure decision is used to determine whether the deviation distribution, deviation direction and deviation remaining position are consistent with the treatment effect change part and the ontological state change part corresponding to the candidate state interpretation branch. The state establishment decision is used to determine whether the state sequence, state transition relationship and state continuous interval are consistent with the index change order in the observed medical dataset and the starting order and continuous process in the treatment action record. Output the branch decision results corresponding to each candidate state interpretation branch.

8. A medical data processing method based on a computer system according to claim 7, characterized in that: S4 also includes: S4-3. Based on the branch determination results corresponding to each candidate state interpretation branch, retain the candidate state interpretation branches that simultaneously satisfy the reconstruction determination and the state establishment determination, and eliminate the candidate state interpretation branches that do not simultaneously satisfy the reconstruction determination and the state establishment determination. In the retained candidate state interpretation branches, read the number of remaining deviation positions corresponding to the reconstruction deviation result and the number of continuous state intervals corresponding to the ontology state solution result. Determine the candidate state interpretation branch whose correspondence between the number of remaining deviation positions and the number of continuous state intervals is consistent as the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch.

9. A medical data processing method based on a computer system according to claim 8, characterized in that: S5 includes: S5-1. Based on the target ontology state results corresponding to the target patient, and according to the state category, state continuity interval and state transition relationship in the target ontology state results, perform medical processing mapping to generate the medical processing results corresponding to the target patient. S5-2. Read the corresponding processing type identifier and patient identifier for the medical processing result, and write the medical processing result into the corresponding result reading path according to the processing type identifier and patient identifier, and generate the result writing record corresponding to the target patient.

10. A computer-based medical data processing system for implementing the computer-based medical data processing method according to any one of claims 1-9, the system comprising an observation construction module, a transmission decomposition module, a state solution module, a competition determination module, and a result writing module, characterized in that: The observation construction module is used to acquire vital sign data, monitoring waveform data, equipment parameter data, test data and treatment action records of the target patient during the target treatment period, and to perform time correlation and processing through the computer system to output the observation medical dataset; The transmission decomposition module is used to perform intervention transmission decomposition on the observed medical dataset based on the action indicators, starting order, duration, and change order of each indicator in the observed medical dataset corresponding to the treatment action record, and to generate multiple candidate state interpretation branches. The candidate state interpretation branches respectively represent the transmission relationship of different treatment effects and the relationship of different patient ontological state changes. The state solution module is used to load candidate state interpretation branches into the GPU, perform observation result reconstruction, index consistency back substitution and residual deviation solution on each candidate state interpretation branch, and output the reconstruction deviation result and ontology state solution result corresponding to each candidate state interpretation branch. The competition determination module is used to perform branch competition determination based on the reconstruction deviation results and ontology state solution results corresponding to each candidate state interpretation branch, determine the target state interpretation branch, and output the target ontology state result corresponding to the target state interpretation branch. The result writing module is used to generate the medical treatment result corresponding to the target patient based on the target ontology state result, and write the medical treatment result to the result reading path.

Images