Clinical diagnosis and treatment medical information retrieval system based on big data

Abstract

This invention relates to the field of medical information processing and intelligent retrieval technology, specifically a clinical diagnosis and treatment medical information retrieval system based on big data. The system includes: a data acquisition module, an encoding module, a database reconstruction module, a core retrieval module, and a decoding output module. The data acquisition module acquires real-time time-series medical record streams from the electronic medical record system and encodes them to generate a current physiological state tensor. The database reconstruction module parses massive amounts of medical text to generate a set of diagnosis and treatment intervention matrices. The core retrieval module uses the graphics processing unit's tensor core for parallel computation to generate a tensor of expected future states and calculates the physiological residual entropy between it and the standard healthy state tensor. The decoding output module outputs the original medical text retrieval results in ascending order of physiological residual entropy. This invention realizes clinically targeted medical information retrieval based on the real-time physiological evolution of patients.

Description

Technical Field

[0001] This invention relates to the field of medical information processing and intelligent retrieval technology, specifically a clinical diagnosis and treatment medical information retrieval system based on big data. Background Technology

[0002] Clinical diagnosis and treatment information retrieval is an important component of medical decision support. With the continuous development of electronic medical record systems, laboratory systems, monitoring systems, and historical medical record databases, the scale of diagnosis and treatment data accumulated by medical institutions is constantly expanding, and the types of data are becoming increasingly rich. In order to improve the efficiency of diagnosis and treatment reference in complex patient scenarios, it is usually necessary to conduct unified retrieval of massive clinical guideline texts, historical valid medical record texts, and related medical materials, so as to provide doctors with diagnosis and treatment information that matches the current condition.

[0003] However, when retrieving medical information for patients with complex disease courses, existing methods mostly employ keyword matching, topic retrieval, or semantic similarity retrieval. While these methods can filter out superficially relevant content from massive amounts of medical text, they often struggle to simultaneously consider the patient's real-time vital signs, changes in laboratory indicators, and the timing of medication use. In particular, they fail to express the differentiated impacts that different diagnostic and treatment texts may have on the patient's future physiological state. In other words, existing medical information retrieval results are easily interfered with by static text similarity, resulting in insufficient adaptability of retrieval results to the patient's current dynamic condition. This not only affects the accuracy of the retrieval results but may also adversely affect the clinical reference value for patients with complex complications. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a clinical diagnosis and treatment medical information retrieval system based on big data. Specifically, the technical solution of this invention includes:

[0005] The system comprises a data acquisition module, an encoding module, a database reconstruction module, a core retrieval module, and a decoding output module.

[0006] The data acquisition module is connected to the electronic medical record system to acquire the real-time time-series medical record stream from the electronic medical record system;

[0007] The encoding module is connected to the data acquisition module and is used to encode the real-time time-series medical record stream using a multi-dimensional time series encoder to generate the current physiological state tensor; the database reconstruction module is used to parse the original medical text in the medical big data containing massive amounts of medical text using a pre-trained causal extraction model to generate a set of diagnosis and treatment intervention matrices corresponding to each original medical text.

[0008] The core retrieval module is equipped with a graphics processor and is connected to the encoding module and the database reconstruction module respectively. It is used to perform parallel multiplication operations on the current physiological state tensor and each treatment intervention matrix in the treatment intervention matrix set using the tensor core in the graphics processor, so as to generate the expected future state tensor corresponding to each treatment intervention matrix.

[0009] The core retrieval module is also used to obtain the preset standard health state tensor representing the health baseline, and to calculate the physiological residual entropy between each expected future state tensor and the preset standard health state tensor representing the health baseline. The physiological residual entropy represents the degree to which the expected future state tensor deviates from the standard health state tensor.

[0010] The decoding output module communicates with the core retrieval module and is used to sort the original medical texts corresponding to each diagnosis and treatment intervention matrix in ascending order of physiological residual entropy in order to output the retrieval results.

[0011] Preferably, the real-time time-series medical record stream includes vital sign data, laboratory indicator data, and medication timestamps; the encoding module is used to divide the vital sign data, laboratory indicator data, and medication timestamps into continuous preset time slices according to the time axis, and perform alignment processing in a unified sequence to obtain structured time-series data; the encoding module is also used to input the structured time-series data into a multidimensional time series encoder for feature extraction to obtain the current physiological state tensor.

[0012] Preferably, the original medical text includes clinical guideline text and historical valid medical record text, and the medical big data consists of the original medical text; the pre-trained causal extraction model includes a temporal causal extraction network; the database reconstruction module is used to extract temporal causal relationship features from the clinical guideline text and historical valid medical record text using the temporal causal extraction network; the database reconstruction module is also used to construct a state transition matrix based on the temporal causal relationship features, and use the state transition matrix as each diagnosis and treatment intervention matrix in the set of diagnosis and treatment intervention matrices.

[0013] Preferably, the core retrieval module is pre-configured with a forward simulation operator; the core retrieval module is used to load the set of diagnostic and treatment intervention matrices into the video memory of the graphics processor; the core retrieval module is also used to call the tensor core in the graphics processor using the forward simulation operator to perform phase space trajectory mapping between the current physiological state tensor and each diagnostic and treatment intervention matrix in the video memory to obtain the mapping result; the core retrieval module is also used to use the mapping result as the expected future state tensor.

[0014] Preferably, the expected future state tensor includes various state indicators, and the standard health state tensor includes corresponding baseline indicators; the core retrieval module is used to subtract each state indicator from the corresponding baseline indicator and take the absolute value to obtain the indicator residual value of each state indicator.

[0015] The core retrieval module is also used to calculate the ratio of the residual value of each indicator to the sum of the residual values ​​of all indicators with a preset non-zero positive real number smoothing term as the probability distribution value, and to sum the products of each probability distribution value and its logarithm in the negative direction to obtain the physiological residual entropy.

[0016] Preferably, the decoding output module is used to sort the various diagnostic and treatment intervention matrices in the diagnostic and treatment intervention matrix set according to the physiological residual entropy from small to large, to obtain a matrix sorting sequence; the decoding output module is also used to map the matrix sorting sequence back to the corresponding original medical text to obtain a text sorting list; the decoding output module is also used to output the text sorting list as the retrieval result.

[0017] Preferably, the big data-based clinical diagnosis and treatment medical information retrieval system also includes a security monitoring module; the security monitoring module is connected to the core retrieval module, and the security monitoring module is used to extract the feature vectors of the number of underlying diseases, age factors and adverse drug reaction records of the current patient, calculate the basic residual tolerance through a preset function, and establish the basic residual tolerance as a safety threshold.

[0018] The safety monitoring module is used to determine the relationship between the physiological residual entropy and the safety threshold. When the physiological residual entropy is greater than or equal to the safety threshold, a red warning mark is generated and attached to the corresponding original medical text. When the physiological residual entropy is less than the safety threshold, the safety monitoring module generates a normal release instruction and transmits the corresponding original medical text to the decoding output module based on the normal release instruction.

[0019] Preferably, the security monitoring module is also used to generate a weight reduction instruction when the physiological residual entropy is greater than or equal to the security threshold; the security monitoring module sends the weight reduction instruction to the decoding output module; the decoding output module is used to reduce the ranking weight of the corresponding original medical text in the search results after receiving the weight reduction instruction.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. This invention continuously acquires real-time time-series medical record streams from the electronic medical record system through a data acquisition module, and uses a multi-dimensional time series encoder to perform time axis alignment and feature extraction on vital sign data, test index data, and medication timestamps to generate a current physiological state tensor. This enables the system to simultaneously combine the patient's real-time vital signs, test changes, and medication sequence to uniformly express the current condition, effectively avoiding the problem that existing static keyword matching or semantic retrieval cannot reflect the dynamic disease progression, and significantly improving the adaptability of retrieval results for patients with complex complications to the current condition.

[0022] 2. This invention utilizes a pre-trained causal extraction model, especially a temporal causal extraction network, through a database reconstruction module to extract temporal causal relationship features from clinical guideline texts and historical valid medical record texts, and constructs a state transition matrix as a diagnosis and treatment intervention matrix. This transforms massive amounts of original medical texts from simply readable content into calculable intervention representations that can be applied to the patient's state, thereby overcoming the shortcomings of existing retrieval methods that can only filter superficially related texts and are difficult to express the differentiated impact of different diagnosis and treatment texts on the patient's future physiological state.

[0023] 3. This invention loads the set of diagnostic and treatment intervention matrices into the graphics processor's memory through a core retrieval module, and uses a forward simulation operator to call the tensor core to perform parallel multiplication operations and phase space trajectory mapping on the current physiological state tensor and each diagnostic and treatment intervention matrix, generating the expected future state tensor corresponding to each candidate text. This transforms clinical retrieval from text-by-text comparison to a real-time forward extrapolation process with multiple options, which not only improves the efficiency of large-scale medical text retrieval, but also directly assesses the impact of different diagnostic and treatment paths on the patient's subsequent state, enhancing the clinical effectiveness and timeliness of the retrieval results. Attached Figure Description

[0024] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0025] Figure 1 This is a schematic diagram of a module of a big data-based clinical diagnosis and treatment medical information retrieval system provided in an embodiment of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0027] A big data-based clinical diagnosis and treatment medical information retrieval system, which includes:

[0028] The data acquisition module communicates with the electronic medical record system and is used to acquire the real-time time-series medical record stream from the electronic medical record system.

[0029] The encoding module is connected to the data acquisition module. The encoding module is used to encode the real-time time series medical record stream using a multi-dimensional time series encoder to generate the current physiological state tensor.

[0030] The database reconstruction module is used to parse the original medical text in medical big data containing massive amounts of medical text using a pre-trained causal extraction model, so as to generate a set of diagnosis and treatment intervention matrices corresponding to each original medical text.

[0031] The core retrieval module is equipped with a graphics processor and is connected to the encoding module and the database reconstruction module. The core retrieval module is used to perform parallel multiplication operations on the current physiological state tensor and each treatment intervention matrix in the treatment intervention matrix set using the tensor core in the graphics processor, so as to generate the expected future state tensor corresponding to each treatment intervention matrix.

[0032] The core retrieval module is also used to obtain the preset standard health state tensor representing the health baseline, and to calculate the physiological residual entropy between each expected future state tensor and the preset standard health state tensor representing the health baseline. The physiological residual entropy represents the degree to which the expected future state tensor deviates from the standard health state tensor.

[0033] The decoding output module communicates with the core retrieval module. The decoding output module is used to sort the original medical texts corresponding to each diagnosis and treatment intervention matrix in ascending order of physiological residual entropy in order to output the retrieval results.

[0034] This embodiment provides a clinical diagnosis and treatment medical information retrieval mechanism based on big data, such as... Figure 1 As shown; specifically, the system is deployed in the clinical decision support platform of the hospital's chest pain center, serving a patient suspected of having non-ST-segment elevation acute coronary syndrome with a history of type 2 diabetes, chronic renal insufficiency, and hypertension; during the observation period, the patient's condition is not a static qualitative text attribute, but a continuously evolving physiological process. Therefore, the system does not simply set the search target to find the most similar article, but regards the patient's current state as the physiological state to be observed, and regards the massive amount of medical text as the source of intervention rules that may change the course of this state, thereby completing the search for clinical effectiveness;

[0035] Specifically, the data acquisition module communicates with the electronic medical record system, laboratory system, monitoring system and medication execution record system through the hospital intranet to continuously acquire the patient's real-time time-series medical record stream; this real-time time-series medical record stream includes continuous vital signs generated by bedside monitoring, laboratory test results uploaded in time periods, and medication records after the doctor's orders are executed;

[0036] The encoding module receives the above data, organizes it into a multidimensional input that reflects the patient's current pathophysiological state, and uses a multidimensional time-series encoder to generate a tensor of the current physiological state. The multidimensional time-series encoder specifically includes an attention mechanism layer and a long short-term memory network layer, which are used to identify the cross-dimensional correlation weights of various physiological indicators under different time slices and the temporal dependencies along the time axis. In engineering implementation, the tensor can be understood as a set of state data matrices organized by time and indicators, which simultaneously retains the comprehensive characterization of circulation, metabolism, inflammation, coagulation, renal function and recent drug exposure. Therefore, it corresponds to the patient's current state, rather than a simple text summary.

[0037] Furthermore, the database reconstruction module performs offline parsing of the original medical text in the medical big data; the original medical text can come from clinical guidelines, pathway consensus, historical valid medical records, specialist case summaries, etc.; this module does not directly save the full text vector, but extracts information from the text under what circumstances what intervention will cause which indicators to change in what direction within what time range through a pre-trained causal extraction model, and converts it into a set of diagnosis and treatment intervention matrices.

[0038] Through the above processing, each original medical text is reconstructed from a static expression into an intervention representation that can act on the current physiological state; among them, the pre-trained causal extraction model is specifically a joint information extraction model based on graph neural networks, which is generated in advance by supervised feature alignment training using a set of specialist guideline texts containing known medical intervention logic and the corresponding expert-annotated causal knowledge graph.

[0039] The core retrieval module is equipped with a graphics processor, preferably one with a tensor core. During retrieval, the core retrieval module calls the tensor core in the graphics processor to perform parallel multiplication operations on the current physiological state tensor and the set of diagnostic and treatment intervention matrices, generating multiple expected future state tensors respectively. This process is used to simulate in actual clinical practice, if a patient is treated according to a certain guideline or a certain type of previously effective treatment, in what direction the patient's vital signs, organ burden and laboratory indicators might evolve in the next stage.

[0040] Then, the core retrieval module compares each expected future state tensor with the preset standard health state tensor to obtain the physiological residual entropy; the standard health state tensor here is not an ideal set of values ​​detached from the scenario, but a baseline reference used to characterize when the patient's physiological system tends to be stable, the organ function load is acceptable, and the risk is low.

[0041] The standard health status tensor is generated by the data acquisition module based on the stable period vital signs data in the current patient's previous electronic medical records, or calculated based on the statistical mean physiological baseline of a subgroup of patients with similar underlying disease profiles to the current patient in the local clinical database; the smaller the physiological residual entropy, the more likely the corresponding diagnosis and treatment text is to push the condition in a relatively safe, stable and controllable direction in the current state of the patient.

[0042] After receiving the sorting results, the decoding output module remaps each treatment intervention matrix back to the corresponding original medical text and outputs the search results in ascending order of physiological residual entropy. The output results can be clinical guideline entries, summaries of similar historical valid medical records, recommended path descriptions, and corresponding warning information. In this way, the terminal outputs not a list of high-frequency cited literature based on word frequency or shallow semantic matching, but a more likely safe and effective treatment reference sequence filtered based on the patient's current dynamic status.

[0043] As an anomaly handling mechanism, when the electronic medical record system is temporarily unable to provide continuous time-series data, key test results are not returned, medication administration time is missing, or the patient data source does not have structured time stamps, the data acquisition module will mark the missing segments, and the encoding module will only use data within the confidence interval to generate a downgraded version of the current physiological state tensor, and add a prompt for incomplete time-series information to the output results; if the missing key data exceeds the preset level, such as myocardial injury markers, renal function indicators, or recent anticoagulation medication time cannot be obtained, the core retrieval module will pause the high-risk regimen sorting and only return conservative reference text to avoid giving misleading high-intensity recommendations when information is insufficient;

[0044] In this chest pain center scenario, the patient experienced recurrent chest tightness within 6 hours of admission. The monitoring system recorded increased heart rate fluctuations, the laboratory system reported a dynamic increase in troponin, and the nursing execution system showed that the patient had received antiplatelet drugs and low molecular weight heparin. After acquiring the above real-time time-series medical record stream, the system encodes the current physiological state tensor.

[0045] The database contains texts from various sources, including conventional intensive treatment without renal impairment and conservative adjustment treatment for patients with renal impairment. After parallel analysis by the core retrieval module, it was found that the expected future state of the latter type of text was closer to the stable baseline, so it was ranked higher in the output list. Although the former type was semantically highly related to acute coronary syndrome, it was automatically ranked lower because it may increase the renal burden and bleeding risk.

[0046] The purpose of this step is to transform clinical retrieval from static text matching into a state assessment process oriented towards the real-time physiological evolution of patients, thereby enabling more clinically targeted medical information retrieval for patients with complex complications and providing a unified data foundation for subsequent security control and ranking optimization.

[0047] Furthermore, the real-time time-series medical record stream includes vital sign data, laboratory test data, and medication timestamps; the encoding module is used to divide the vital sign data, laboratory test data, and medication timestamps into continuous preset time slices according to the time axis, and perform alignment processing in a unified sequence to obtain structured time-series data; the encoding module is also used to input the structured time-series data into a multidimensional time series encoder for feature extraction to obtain the current physiological state tensor.

[0048] This embodiment provides an alignment coding mechanism for real-time time-series medical record streams. Specifically, in the previous implementation scenario, simply stacking vital signs, test results, and medication records together can form a dataset, but it is difficult to accurately express key clinical relationships such as whether a certain drug was used before or after a certain test result, or whether a deterioration in vital signs occurred after the administration of additional anticoagulants. For chest pain patients with multiple underlying diseases, the temporal relationship itself is part of the risk assessment, so it is necessary to align the timeline first and then encode it.

[0049] Specifically, the encoding module receives three types of data streams: one is vital sign data, such as heart rate, blood pressure, blood oxygen saturation, and respiratory rate; another is laboratory indicator data, such as troponin, creatinine, blood glucose, D-dimer, and blood potassium; and the third is medication timestamps, including the execution time of aspirin, P2Y12 receptor inhibitors, heparin, nitrates, insulin, and antihypertensive drugs. Because the three types of data have different original sampling frequencies, vital signs are often generated at the minute or even second level, laboratory results are returned at the hour level, and medication execution is an event-type record. The encoding module will first map the three to the same sequence using a unified timeline to form structured time-series data.

[0050] In implementation, a certain observation window can be divided into continuous time slices. Each time slice records the summary of vital signs that best represent the patient's condition at that moment or within that time period, the latest available test results, and whether a drug administration event occurred within that time period. The structured time series data is logically represented as a multidimensional feature matrix sequence. The step size of the sequence is fixed to the preset time slice length. Each channel of the matrix corresponds to the downsampled mean of vital signs, the interpolated and completed test values, and the drug administration event vector represented by one-hot encoding, respectively.

[0051] The significance of this approach is that the system no longer processes isolated data fields, but rather the comprehensive correlation features of the patient's circulatory status, laboratory evidence, and recently applied interventions within that time slice; the multidimensional time series encoder then extracts the changing trends, lag effects, and linkage features across time slices, ultimately generating the current physiological state tensor.

[0052] To facilitate understanding, the following principle can be illustrated: Suppose there are three consecutive time slices within the observation window. The first time slice mainly records persistent chest pain and blood pressure exceeding the upper limit of the standard reference range. The second time slice records antiplatelet administration. The third time slice reports elevated troponin and elevated creatinine. After alignment, the multidimensional time series encoder can identify the complex state of administration occurring before further increases in myocardial injury indicators and the patient simultaneously experiencing renal overload, rather than processing each piece of information in isolation.

[0053] As an anomaly handling mechanism, if certain types of data are missing in some time slices, such as nighttime test results not yet being returned, the encoding module will use the most recently verified valid result and add a freshness mark; if the missing data is a key dosing timestamp, then this segment of data will not participate in the extraction of drug exposure strongly correlated features, and only basic vital signs and test changes will be retained to avoid misjudging unknown dosing situations as no dosing; if the vital sign monitoring data shows obvious abnormal jumps and is judged as equipment noise, then the time slice will be masked or smoothed to prevent the encoding results from being dominated by transient artifacts.

[0054] During the 8th hour of the patient's observation, the monitoring system uploaded heart rate and blood pressure every 5 minutes. The laboratory returned troponin and creatinine results at the 6th and 8th hours, respectively. The nursing execution system generated a low molecular weight heparin administration time at the 7th hour. After the system was aligned with a unified time axis, it could form structured time-series data containing fluctuations in vital signs, drug administration events, and test responses. The coding module then generated a tensor of the current physiological state. The state obtained at this time not only reflects the severity of the illness but also the reality of drug exposure but insufficient organ reserves.

[0055] The purpose of this step is to ensure that the coding results faithfully reflect the sequence of clinical events and the patient's physiological response chain, thereby achieving a unified expression of long-term, multi-source, asynchronously sampled medical record streams and avoiding subsequent retrievals from the actual condition due to temporal misalignment.

[0056] Furthermore, the original medical texts include clinical guideline texts and historical valid medical record texts; medical big data consists of the original medical texts; the pre-trained causal extraction model includes a temporal causal extraction network; the database reconstruction module is used to extract temporal causal relationship features from the clinical guideline texts and historical valid medical record texts using the temporal causal extraction network; the database reconstruction module is also used to construct a state transition matrix based on the temporal causal relationship features, and use the state transition matrix as each diagnosis and treatment intervention matrix in the set of diagnosis and treatment intervention matrices.

[0057] This embodiment provides a medical big data reconstruction mechanism. Specifically, in the aforementioned scenario, if the backend database still only stores the original guideline text and the full text of historical medical records, the system will still get stuck in superficial similar word matching during retrieval, making it difficult to express the causal relationship of which patients are suitable for a certain treatment under what prerequisites and in what direction the treatment will evolve. Especially for patients with chest pain and renal insufficiency, the applicable boundaries and subsequent risks of texts that mention anticoagulation and contrast enhancement therapy may be completely different. Therefore, it is necessary to reconstruct the causal relationship of medical texts.

[0058] Specifically, the database reconstruction module collects clinical guideline texts and historical valid medical record texts to form the source of medical big data; clinical guideline texts usually have standardized conditions, recommendation levels and indications, while historical valid medical record texts contain real-world evolution paths and processing results;

[0059] When the temporal causal extraction network parses these texts, it focuses on identifying four types of information: first, the initial state conditions, such as persistent chest pain, elevated troponin, decreased renal function, and abnormal glucose metabolism; second, the intervention actions, such as antiplatelet therapy, anticoagulation, fluid replacement, insulin adjustment, delayed angiography, or dose reduction strategies; third, the direction of evolution, such as hemodynamic stabilization, slowed progression of myocardial injury, increased bleeding risk, and deterioration of renal function; and fourth, the sequence of occurrence, such as assessing renal function first, then deciding on the timing of angiography, and adjusting coagulation indicators if coagulation indicators are abnormal after medication.

[0060] Based on this, the database reconstruction module constructs a state transition matrix; in an engineering sense, this matrix represents the direction and strength of the effects of the intervention path contained in a certain text on multiple physiological dimensions of the patient; for guideline texts, the matrix emphasizes the standardized path and applicable conditions; for historical valid medical record texts, the matrix emphasizes the actual disease progression and organ response; after such transformation, texts from different sources can participate in retrieval in a unified state space.

[0061] As a specific explanatory example: if a guideline states that antithrombotic therapy should be initiated as early as possible for high-risk patients with chest pain, but drug exposure should be controlled and bleeding closely monitored when there is significant renal dysfunction, then the reconstructed treatment intervention matrix will not only retain the term "antithrombotic therapy," but will also include several state transition relationships, such as reducing thrombosis-related risks, increasing bleeding-related vigilance, and being constrained by renal function conditions; if a valid historical medical record shows that in the context of diabetes and mild to moderate renal impairment, myocardial markers stabilized after conservative antithrombotic therapy and delayed invasive procedures, then the corresponding matrix will better reflect the causal chain between mild intervention and organ protection;

[0062] As an anomaly handling mechanism, if the original medical text contains confusing abbreviations, vague time descriptions, or missing efficacy outcomes, the database reconstruction module marks it as low-confidence text, retaining only the confirmed state constraints and intervention directions, and not using it as a high-priority matrix source; if there are obviously contradictory descriptions in the text, such as the same paragraph containing both contraindications and recommendations without explanation of applicable conditions, it will enter the manual review queue to avoid directly writing ambiguous causal structures into the matrix database.

[0063] Before the patient search task was initiated, the system backend had already collected guidelines for acute coronary syndrome, consensus on the management of diabetes mellitus with coronary heart disease, recommendations for adjusting antithrombotic therapy in patients with renal insufficiency, and multiple historical valid medical records. The reconstruction module uniformly parses these texts into a set of diagnosis and treatment intervention matrices. Among them, for medical record texts on reducing angiography-related damage in the context of chronic renal insufficiency, a state transition matrix that emphasizes the renal protection pathway is generated; for guideline entries on standard high-intensity antithrombotic procedures, a matrix that emphasizes ischemic control but also includes bleeding and renal load constraints is generated. During subsequent searches, the system can directly compare the effects of these matrices on the current patient's status.

[0064] The purpose of this mechanism is to transform temporal causal knowledge from massive amounts of raw medical texts into computationally computable and parallel-accessible intervention representations, thereby achieving a shift from retrieving texts to retrieving medical knowledge units that can be applied to the patient's condition.

[0065] Furthermore, the core retrieval module is pre-installed with a forward simulation operator; the core retrieval module is used to load the set of diagnostic and treatment intervention matrices into the video memory of the graphics processor; the core retrieval module is also used to call the tensor core in the graphics processor using the forward simulation operator to perform phase space trajectory mapping between the current physiological state tensor and each diagnostic and treatment intervention matrix in the video memory to obtain the mapping result; the core retrieval module is also used to use the mapping result as the expected future state tensor.

[0066] This embodiment provides a core retrieval mechanism based on forward simulation. Specifically, the previous implementation has completed the formation of the patient state tensor and the construction of the diagnosis and treatment intervention matrix. However, if the retrieval stage still uses serial comparison of text entries one by one, the response speed will be difficult to meet the timeliness requirements of real-time auxiliary decision-making in chest pain centers when faced with a large number of guideline entries and historical medical record databases. Especially in the acute exacerbation stage, doctors need to see candidate solutions with confidence greater than a preset threshold within a preset response period. Therefore, this embodiment introduces a forward simulation arithmetic unit and a graphics processor memory loading mechanism.

[0067] Specifically, after receiving a retrieval task, the core retrieval module first loads the set of diagnostic and treatment intervention matrices into the graphics processor's video memory in batches. The significance of doing this is to reduce the frequent data transfer between main memory and graphics processor, so that a large number of subsequent matrix operations can be completed continuously on the video memory side.

[0068] The forward simulation processor is a software inference service process based on Markov decision process deployed inside the core retrieval module. It calls the application programming interface of the graphics processor. The forward simulation processor is responsible for scheduling the current physiological state tensor and various diagnostic and treatment intervention matrices in the video memory to perform parallel operations, and completes high-throughput state mapping through the tensor core.

[0069] The phase space trajectory mapping mentioned here can be understood as follows: the patient is not just a binary state parameter, but a point in a high-dimensional state space composed of multiple organs and multiple indicators; different diagnostic and treatment intervention matrices are equivalent to forces in different directions, which will push the patient's state to different subsequent trajectories.

[0070] Specifically, the mapping operation process is as follows: the current physiological state tensor is multiplied in parallel with each diagnostic and treatment intervention matrix pre-loaded in the video memory; since the physiological state tensor reflects the quantitative status of multi-dimensional indicators, and the diagnostic and treatment intervention matrix stores the transition weights between state variables, that is, the positive and negative change coefficients caused by a specific intervention, the result of the multiplication of the two directly represents the new state distribution after the intervention is applied, that is, the mapping result; the operating mechanism of the forward simulation operator is to quickly evaluate how each candidate intervention strategy affects the patient's future state trajectory, rather than relying solely on the similarity measure of the text surface;

[0071] For ease of understanding, the following simplified mapping explanation can be provided: Assume that there are three candidate matrices in the video memory at the same time, which correspond to the standard enhanced antithrombotic pathway, the reduced antithrombotic pathway under renal function limitation, and the conservative adjustment pathway that prioritizes the stability of the internal environment.

[0072] After the current physiological state tensor is input, the forward simulation calculator will simultaneously obtain three mapping results; the first result improves the ischemia-related dimensions more significantly, but at the same time increases the renal load and hemorrhage sensitivity dimensions; the second result takes into account both ischemia control and organ burden; the third result improves at a slower rate than the first preset rate of change, but the overall fluctuation variance is smaller than the first preset variance.

[0073] If the improvement rate is greater than or equal to the first preset rate of change, or the overall fluctuation variance is greater than or equal to the first preset variance, then other corresponding mapping results are generated; the first preset rate of change and the first preset variance are preset by the system based on the statistical characteristics of the physiological index fluctuations of patients with the same type of underlying disease in the local historical valid medical record database during the stable period; the three mapping results are respectively used as expected future state tensors to enter the next stage of evaluation.

[0074] As an exception handling mechanism, if the video memory capacity is insufficient to load the entire set of diagnostic and intervention matrix at once, the core retrieval module will load the matrix library in batches while keeping the same patient state tensor unchanged, and finally summarize the expected future state tensors of each batch; if the graphics processor is temporarily unavailable or in a faulty state, the system can switch to the degraded mode on the central processing unit, only search the high-priority matrix subset, and mark it as having entered the degraded inference on the results page; if the current physiological state tensor contains too many missing labels, the forward simulation calculator will reduce the interpretation weight of the long-term trajectory to avoid over-extrapolation when the input is incomplete;

[0075] On the 10th hour of the patient's observation, chest pain recurred and blood sugar control was unstable. The doctor needed to quickly review the treatment references more suitable for the current stage. The system preloaded thousands of reconstructed treatment intervention matrices into the graphics processor's memory. The forward simulation calculator called the tensor core to complete the parallel mapping, forming a set of expected future state tensors within seconds. Compared with traditional text-by-text retrieval, the doctor did not need to wait for the system to analyze the text line by line. He only needed to view the candidate solutions that were closest to the stable baseline after simulation.

[0076] The purpose of this mechanism is to leverage the parallel capabilities of hardware to transform clinical knowledge retrieval into a real-time forward extrapolation process, thereby enabling rapid, multi-option, and comparable assessments for patients with complex and long-term illnesses.

[0077] Furthermore, the expected future state tensor includes various state indicators, and the standard health state tensor includes corresponding baseline indicators. The core retrieval module is used to subtract each state indicator from its corresponding baseline indicator and take the absolute value to obtain the indicator residual value of each state indicator. The core retrieval module is also used to calculate the ratio of each indicator residual value to the sum of all indicator residual values ​​with a preset non-zero positive real number smoothing term as the probability distribution value, and to negatively accumulate and sum the products of each probability distribution value and its logarithm to obtain the physiological residual entropy.

[0078] This embodiment provides a physiological residual entropy assessment mechanism. Specifically, after forward simulation is completed, if only a single indicator is used to judge the merits of candidate solutions, such as extracting only a single test indicator or focusing only on blood pressure changes, it is easy to make one-sided decisions. For patients with chest pain complicated by diabetes and renal insufficiency, the clinical high-risk state is usually not due to deviation of a single indicator, but rather to deviation of multiple systems at the same time and an unbalanced distribution of deviations. Therefore, this embodiment uses physiological residual entropy to describe the overall degree of deviation of the expected future state from the standard healthy state.

[0079] Specifically, the core retrieval module extracts various state indicators from each expected future state tensor, such as indicators related to circulatory stability, myocardial injury, renal function, coagulation risk, and metabolic control, and extracts corresponding baseline indicators from the standard health state tensor.

[0080] In the specific simulation, the core retrieval module extracts a preset number of... status indicators ,in and obtain the corresponding baseline indicators. Before calculating the deviation, the core retrieval module uses a normalization algorithm to normalize various state indicators. and corresponding baseline indicators Dimensionless processing is performed to eliminate the differences in physical dimensions between different physiological indicators;

[0081] The module first calculates each status indicator. Relative to the corresponding baseline index The deviation range is used to obtain the residual values ​​of each indicator:

[0082]

[0083] Calculate the sum of the residuals for all indicators:

[0084]

[0085] Based on each residual value In overall residual The proportion in the probability distribution value is:

[0086]

[0087] The probability distribution values ​​of each item Instead of summing the products of numerical values ​​in the negative order, we can calculate the physiological residual entropy, which characterizes the overall degree of deviation:

[0088]

[0089] in, For summation, Indicates The natural logarithm function with base 0;

[0090] In engineering implementation, to avoid calculation anomalies where the denominator is zero when all residual values ​​of various indicators are zero, the core retrieval module will sum the residual values ​​of all indicators. On top of this, a preset non-zero positive real constant, such as 1e-6, is added as a smoothing term; at the same time, when calculating the logarithmic product, if the probability distribution value of a certain term is zero, then according to the mathematical limit principle, the product result of that term is directly taken as zero, thereby ensuring that the physiological residual entropy calculation process has rigorous mathematical completeness and engineering robustness under any extreme physiological state input.

[0091] From a physiological perspective, this value reflects not only the magnitude of the deviation, but also whether the deviation is concentrated in a few key systems or dispersed across multiple systems. If a candidate treatment causes a patient to deviate slightly in only one dimension while other dimensions remain close to baseline, the overall risk structure may be relatively manageable. If a treatment causes multiple organ load indicators to rise simultaneously, even if a single indicator does not reach an extreme level, it may indicate a more complex clinical risk. Therefore, physiological residual entropy is suitable as a unified ranking criterion across multiple indicators and systems.

[0092] As a concrete explanatory example: Suppose that a certain expected future state deviates in three core dimensions, corresponding to ischemia control, renal load, and bleeding tendency, respectively; if the deviation is mainly concentrated in the renal load dimension, it indicates that the main risk points of the plan are relatively clear; if all three dimensions deviate significantly, it suggests that the plan has a wider impact on the patient's overall homeostasis; the difference between the former and the latter in the residual composition will be reflected through physiological residual entropy.

[0093] As an anomaly handling mechanism, if certain state indicators do not have comparable health baselines in the current scenario, such as patients with long-term chronic abnormalities who should not be brought back to the normal range of the general population in the short term, the standard health status tensor can adopt the clinical stable baseline of this patient subgroup instead of the ideal textbook value; if an indicator cannot participate in the calculation due to missing data, it is masked in the total residual and re-normalized to the set of available indicators; if the number of available indicators is lower than the preset lower limit, the physiological residual entropy of the candidate scheme is only used as a low confidence reference and does not directly participate in the highest priority ranking.

[0094] Among the multiple candidate options for this patient, one option could reduce the chest pain-related risk assessment value within the first preset time period, but it also caused significant deviations in the two dimensions of renal function load and coagulation risk. Another option improved myocardial ischemia at a slightly slower rate, but it kept circulation, renal function and coagulation within a more controllable range. By comparing the residual structure of the two options with respect to the tensor of standard healthy state, the system usually judges the latter to have a smaller physiological residual entropy, and thus presents it first in the results.

[0095] The purpose of this mechanism is to compress complex multi-organ risk differences into a unified, rankable indicator, thereby achieving an overall steady-state friendliness evaluation of candidate treatment options and reducing the interference of single-indicator bias on clinical search results.

[0096] Furthermore, the decoding output module is specifically used to sort the various diagnostic and treatment intervention matrices in the set of diagnostic and treatment intervention matrices according to the physiological residual entropy from small to large, to obtain a matrix sorting sequence; the decoding output module is also used to map the matrix sorting sequence back to the corresponding original medical text to obtain a text sorting list; the decoding output module is also used to output the text sorting list as the retrieval result.

[0097] This embodiment provides a decoding output mechanism; specifically, the aforementioned steps have obtained candidate results after evaluation according to physiological residual entropy, but clinical users ultimately need to read guideline entries, medical record summaries, and treatment recommendations, rather than the state matrix itself; therefore, it is necessary to restore the computation-oriented diagnosis and treatment intervention matrix back to readable, traceable, and auditable original medical text, and form a stable output sequence.

[0098] Specifically, the decoding output module sorts the set of diagnostic and treatment intervention matrices according to the physiological residual entropy to obtain a matrix sorting sequence. This sequence is essentially an order of which intervention path is more likely to lead the current patient to a stable state. The decoding output module reads the source index of each matrix in the database and maps it back to the corresponding original medical text to form a text sorting list. The mapping relationship can be one-to-one or one-to-many. For example, the same guideline text may form multiple matrix branches with different applicable conditions after reconstruction. In this case, the output can be restored and displayed according to the condition description corresponding to the optimal branch.

[0099] To facilitate physician comprehension, the output can display not only the text title but also the source category, summary of applicable conditions, key risk warnings, and matching reasons related to the current patient's condition. For example, it can display a section from the Acute Coronary Syndrome Guidelines, applicable to patients with limited renal function, emphasizing dose reduction of antithrombotic drugs and close monitoring, enabling clinicians to clearly understand the source of the ranking criteria and avoid outputting uninterpretable and untraceable calculation results.

[0100] The following mapping example can be used to illustrate this: If the matrix sorting sequence corresponds to M2, M5, and M1 in sequence, the decoding output module will map M2 back to the antithrombotic adjustment suggestion for patients with impaired renal function, map M5 back to the medical record summary of the conservative control path for diabetes mellitus with coronary heart disease, and map M1 back to the standard intensive treatment guideline entry, thus forming the final text sorting list.

[0101] As an exception handling mechanism, if a diagnosis and treatment intervention matrix loses its original text mapping relationship during database updates, the matrix will not be directly output to the clinical end, but will be marked as requiring maintenance of the source mapping; if multiple matrices correspond to different branches of the same original medical text, the decoding output module can merge and display them, and add the applicable conditions of the current hit below the text; if the text ranked at the top comes from a low-quality or low-confidence data source, the system can retain its position, but at the same time add a source level label to facilitate doctors' careful judgment.

[0102] In the patient's search results, the system ultimately ranked the guidelines for adjusting antithrombotic therapy for acute coronary syndrome in the context of renal impairment first, the historical valid medical record summary for patients with diabetes mellitus and chronic kidney damage second, and the standard intensive treatment items without obvious organ load limitations third. What the doctor sees on the interface is a list of texts sorted by clinical applicability, and behind each text is a corresponding treatment intervention path that has been completed through forward extrapolation.

[0103] The purpose of this mechanism is to convert the state sorting results of the computation layer into clinically readable, reviewable, and directly referential text results, thereby achieving a closed-loop implementation from algorithm deduction to medical information output.

[0104] Furthermore, the system also includes a security monitoring module;

[0105] The security monitoring module communicates with the core retrieval module.

[0106] The safety monitoring module is used to extract the feature vectors of the number of underlying diseases, age factors, and adverse drug reaction records of the current patient, calculate the baseline residual tolerance through a preset function, and establish the baseline residual tolerance as a safety threshold.

[0107] The safety monitoring module is used to determine the relationship between the physiological residual entropy and the safety threshold; when the physiological residual entropy is greater than or equal to the safety threshold, a red warning mark is generated and appended to the corresponding original medical text;

[0108] When the physiological residual entropy is less than the safety threshold, the safety monitoring module generates a normal release instruction and transmits the corresponding original medical text to the decoding output module based on the normal release instruction.

[0109] This embodiment provides a security monitoring mechanism for search results. Specifically, while ranking the output can identify relatively better candidate texts, in high-risk clinical scenarios, being merely better is insufficient; it is also necessary to clearly identify obviously unsafe results. This is because some texts, although highly relevant to the current condition and even routinely recommended for general patients, may significantly deviate from the safety baseline for patients with impaired renal function, bleeding tendency, or metabolic disorders. Therefore, this embodiment adds a security monitoring module to the search process.

[0110] Specifically, the security monitoring module pre-stores security thresholds; these thresholds are used to distinguish between candidate results that can enter the sorting list and high-risk results that require significant warning; after the core retrieval module calculates the physiological residual entropy corresponding to each candidate text, the security monitoring module judges them one by one; if the physiological residual entropy corresponding to a certain text reaches or exceeds the security threshold, it indicates that the expected future state tensor corresponding to the candidate text exceeds the preset security tolerance range. At this time, the security monitoring module generates a red warning mark and attaches the mark to the text; if the physiological residual entropy is lower than the security threshold, a normal release instruction is generated, and the corresponding text is transmitted to the decoding output module for regular display;

[0111] The safety threshold can be dynamically configured according to the department type, disease type, and patient subgroup. In specific configuration, the safety monitoring module extracts the feature vectors of the current patient's underlying disease quantity, age factor, and historical adverse drug reaction records, calculates the personalized basic residual tolerance through a preset multiple linear regression function, and establishes this tolerance as the safety threshold.

[0112] The multiple linear regression function is pre-trained using a historical medical safety dataset. This dataset takes the number of underlying diseases, age factors, and adverse drug reaction records of patients as inputs, and the corresponding safety residual boundary values ​​annotated by medical experts as output labels. It is generated by training and converging by minimizing the prediction error. For example, for high-risk comorbidities in chest pain centers, stricter thresholds can be set to prioritize the identification of potential bleeding and organ failure risks. For follow-up scenarios with relatively stable conditions, relatively lenient thresholds can be used to avoid excessive interception of routine reference information.

[0113] As an anomaly handling mechanism, if the security monitoring module detects that the physiological residual entropy is in the confidence critical range near the threshold, the result can be marked as requiring manual review. It will neither be completely blocked nor directly released as a normal result. If the security threshold parameter is not successfully loaded, the system will default to a conservative security strategy and uniformly add warnings to high-risk text categories. If some texts have low confidence in physiological residual entropy due to insufficient data, the principle of prudence will be given priority, and they will be included in the observation list with a warning that the risk assessment evidence is insufficient.

[0114] During the patient search process, a text from the conventional high-intensity antithrombotic pathway, although highly relevant to coronary syndrome semantics, showed forward extrapolation results indicating a significant increase in future bleeding tendency and renal burden, causing the physiological residual entropy to exceed the preset safety threshold; therefore, the safety monitoring module added a red warning mark to this text; at the same time, a text emphasizing renal function protection and phased antithrombotic adjustment had a lower residual entropy, was generated as a normal release instruction and entered into the regular output list;

[0115] The purpose of this mechanism is to perform a safety-oriented screening process in advance during the retrieval phase, thereby providing a significant indication of high-risk recommendations and reducing the probability of clinicians being misled by superficial relevance.

[0116] Furthermore, the security monitoring module is also used to generate a weight reduction instruction when the physiological residual entropy is greater than or equal to the security threshold; the security monitoring module sends the weight reduction instruction to the decoding output module; the decoding output module is used to reduce the ranking weight of the corresponding original medical text in the search results after receiving the weight reduction instruction.

[0117] This embodiment provides an output optimization mechanism with risk reduction. Specifically, in the previous embodiment, the safety monitoring module was able to add a red warning mark to high-risk text. However, in some actual clinical scenarios, adding a warning is still insufficient to prevent misuse. The reason is that doctors may read the items at the top of the list under high pressure. If high-risk text is marked in red but still remains at the top, it may still interfere with clinical judgment. Therefore, this embodiment further introduces a reduction instruction.

[0118] Specifically, when the security monitoring module determines that the physiological residual entropy corresponding to a certain original medical text reaches or exceeds the security threshold, in addition to generating a red warning mark, it also generates a weight reduction instruction and sends it to the decoding output module. After receiving the instruction, the decoding output module reduces the ranking weight of the text in the search results. This reduction does not mean complete deletion, but rather moving it from the priority recommendation layer to the risk reference layer or the warning and reference layer. In this way, the system still retains the traceability and integrity of medical information, but avoids it competing with safer candidate results for the first place in the ranking.

[0119] This mechanism is particularly suitable for texts with obvious limitations on applicability; for example, some guideline texts are effective for ordinary high-risk coronary patients, but are not the optimal choice for patients with impaired renal function and potential bleeding risk; after the weighting is reduced, doctors can still view these texts, but the order in which they are presented on the interface reflects the system's recalibration for the specific risks of the current patient.

[0120] As an exception handling mechanism, if the total number of candidate texts is small and all texts trigger the deweighting condition, the system will not simply reduce the ranking weight of all results to the lowest priority range. Instead, it will retain the texts with relatively lower risk at the top and uniformly display that the overall risk of the current search results is high, suggesting a comprehensive manual evaluation. If a text is judged to need deweighting by multiple rules at the same time, such as having both high physiological residual entropy and coming from a low-confidence data source, the deweighting magnitude can be stacked. If a text is deweighted but belongs to a normative guideline that is forcibly retained, the system will retain its display, but will not enter the preferred recommendation area.

[0121] In the patient's search results, the standard enhanced antithrombotic pathway text was given a red warning and a downweight instruction by the safety monitoring module because it was expected to amplify the risk of bleeding and kidney damage. Ultimately, it was moved from a position that might have been at the top of the results list to the back. On the other hand, the staged antithrombotic and monitoring pathway for patients with limited renal function maintained a higher ranking weight and appeared in the preferred reference area. As a result, doctors saw text that was more in line with the current pathological context, while high-risk text was retained as information that should be referred to with caution.

[0122] The purpose of this mechanism is to directly feed the safety assessment results back to the ranking layer, thereby enabling risk reorganization of search results and ensuring that the clinical interface prioritizes the presentation of safer and more suitable medical information for the current patient's condition.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A clinical diagnosis and treatment medical information retrieval system based on big data, characterized in that, The system includes: A data acquisition module is communicatively connected to the electronic medical record system; the data acquisition module is used to acquire the real-time time-series medical record stream of the electronic medical record system. An encoding module is communicatively connected to the data acquisition module; the encoding module is used to encode the real-time time-series medical record stream using a multidimensional time series encoder to generate a current physiological state tensor. The database reconstruction module is used to parse the original medical text in the medical big data containing massive amounts of medical text using a pre-trained causal extraction model, so as to generate a set of diagnosis and treatment intervention matrices corresponding to each original medical text. The core retrieval module is equipped with a graphics processor and is communicatively connected to the encoding module and the database reconstruction module. The core retrieval module is used to perform parallel multiplication operations on the current physiological state tensor and each treatment intervention matrix in the treatment intervention matrix set using the tensor core in the graphics processor, so as to generate the expected future state tensor corresponding to each treatment intervention matrix. The core retrieval module is also used to obtain a preset standard health state tensor representing a health baseline, and to calculate the physiological residual entropy between each expected future state tensor and the preset standard health state tensor representing a health baseline, wherein the physiological residual entropy represents the degree to which the expected future state tensor deviates from the standard health state tensor. The decoding output module is communicatively connected to the core retrieval module. The decoding output module is used to sort the original medical texts corresponding to each of the diagnostic and treatment intervention matrices in ascending order of physiological residual entropy in order to output retrieval results.

2. The clinical diagnosis and treatment medical information retrieval system based on big data according to claim 1, characterized in that, The real-time time-series medical record stream includes vital signs data, laboratory test data, and medication timestamps; The encoding module is used to divide the vital signs data, the test index data, and the medication timestamp into continuous preset time slices according to the time axis, and perform alignment processing in a unified sequence to obtain structured time series data. The encoding module is also used to input the structured time series data into the multidimensional time series encoder for feature extraction to obtain the current physiological state tensor.

3. The clinical diagnosis and treatment medical information retrieval system based on big data according to claim 1, characterized in that, The original medical texts include clinical guideline texts and historical valid medical record texts; the medical big data consists of the original medical texts. The pre-trained causal extraction model includes a temporal causal extraction network; The database reconstruction module is used to extract temporal causal relationship features from the clinical guideline text and the historical valid medical record text using the temporal causal extraction network; The database reconstruction module is also used to construct a state transition matrix based on the temporal causal relationship features, and to use the state transition matrix as each diagnosis and treatment intervention matrix in the diagnosis and treatment intervention matrix set.

4. The clinical diagnosis and treatment medical information retrieval system based on big data according to claim 1, characterized in that, The core retrieval module is pre-installed with a forward simulation calculator; The core retrieval module is used to load the set of diagnostic and intervention matrices into the video memory of the graphics processor; The core retrieval module is also used to call the tensor core in the graphics processor using the forward simulation arithmetic unit, and perform phase space trajectory mapping between the current physiological state tensor and each of the diagnostic and treatment intervention matrices in the video memory to obtain the mapping result; The core retrieval module is also used to use the mapping result as the expected future state tensor.

5. The clinical diagnosis and treatment medical information retrieval system based on big data according to claim 1, characterized in that, The expected future state tensor includes various state indicators, and the standard health state tensor includes corresponding baseline indicators. The core retrieval module is used to subtract each of the status indicators from the corresponding baseline indicators and take the absolute value to obtain the indicator residual value of each status indicator. The core retrieval module is also used to calculate the ratio of the residual value of each indicator to the sum of all indicator residual values ​​with a preset non-zero positive real number smoothing term as the probability distribution value, and to negatively accumulate and sum the products of each probability distribution value and its logarithm to obtain the physiological residual entropy.

6. The clinical diagnosis and treatment medical information retrieval system based on big data according to claim 1, characterized in that, The decoding output module is specifically used to sort the various diagnostic and treatment intervention matrices in the diagnostic and treatment intervention matrix set according to the physiological residual entropy from small to large, so as to obtain a matrix sorting sequence; The decoding output module is also used to map the matrix sorting sequence back to the corresponding original medical text to obtain a text sorting list; The decoding output module is also used to output the sorted text list as the search result.

7. The clinical diagnosis and treatment medical information retrieval system based on big data according to claim 1, characterized in that, The system also includes a security monitoring module; the security monitoring module is communicatively connected to the core retrieval module. The safety monitoring module is used to extract the feature vectors of the number of underlying diseases, age factors, and adverse drug reaction records of the current patient, calculate the basic residual tolerance through a preset function, and establish the basic residual tolerance as a safety threshold. The safety monitoring module is used to determine the relationship between the physiological residual entropy and the safety threshold; when the physiological residual entropy is greater than or equal to the safety threshold, a red warning mark is generated and the red warning mark is attached to the corresponding original medical text. When the physiological residual entropy is less than the safety threshold, the safety monitoring module generates a normal release instruction and transmits the corresponding original medical text to the decoding output module based on the normal release instruction.

8. The clinical diagnosis and treatment medical information retrieval system based on big data according to claim 7, characterized in that, The safety monitoring module is also used to generate a weight reduction instruction when the physiological residual entropy is greater than or equal to the safety threshold; The security monitoring module sends the demotion instruction to the decoding output module; The decoding output module is used to reduce the ranking weight of the corresponding original medical text in the search results after receiving the weight reduction instruction.

Images