A vte intelligent assessment and prevention decision support system

Abstract

The application discloses a VTE intelligent evaluation and prevention decision support system, and belongs to the technical field of VTE risk evaluation, and comprises an adaptive acquisition module, an evaluation module, a decision module and a tracking optimization module; the adaptive acquisition module is used for receiving inventory medical data and dynamic physiological data of a target patient, setting a clinical trigger mechanism to obtain directional clinical data; the evaluation module is used for screening implicit features, assigning weights based on time decay correction, combining dominant features, generating a fusion feature set, and constructing a dynamic risk portrait; the decision module is used for constructing a multi-fusion model, performing collaborative evaluation, calculating the fusion risk probability of the target patient, generating a risk prediction sequence, and combining the bleeding risk probability to generate a net risk probability; the tracking optimization module is used for constructing a dynamic scheme library, performing intervention adaptation, generating a recommended scheme list, realizing closed-loop optimization from evaluation to intervention, and comprehensively improving the accuracy and effectiveness of VTE prevention.

Description

Technical Field

[0001] This invention belongs to the field of VTE risk assessment technology and relates to a VTE intelligent assessment and prevention decision support system. Background Technology

[0002] Venous thromboembolism (VTE) is a common complication, including deep vein thrombosis (DVT) and pulmonary embolism (PE), which affects patients' health. In order to effectively prevent the occurrence of VTE, it is necessary to conduct risk assessments on patients in clinical practice and take corresponding preventive measures based on the assessment results.

[0003] Chinese patent CN112447293B discloses a VTE risk warning system, comprising a VTE data acquisition terminal, a VTE database, a VTE data processing terminal, a VTE data analysis terminal, and a data display terminal. The system extracts VTE characteristic data from the health data of individuals to be assessed using a computer system, and analyzes the VTE risk level of each individual based on this characteristic data. Because VTE characteristic data is extracted from the health data of the individuals to be assessed, rather than solely relying on physical examination results, a more comprehensive data analysis of VTE risk can be achieved. Furthermore, using a computer system to assess VTE risk can save medical staff significant time in VTE risk assessment, allowing for faster acquisition of VTE risk assessment results.

[0004] Although existing technologies improve the efficiency of analysis and assessment and reduce the workload of medical staff by analyzing the VTE risk level of individuals through computer systems, they do not consider the collection and extraction mechanism of latent features. This results in blind spots in the data source, making it impossible to obtain information related to latent factors required for VTE assessment. Furthermore, information filtering relies solely on keyword matching, NLP word segmentation, and fixed indicator extraction, making it difficult to capture the unstructured and dynamic information corresponding to latent features. This leads to incomplete data dimensions for VTE risk assessment, ultimately affecting the comprehensiveness and accuracy of the assessment results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a VTE intelligent assessment and prevention decision support system. This system fills data gaps by using existing data and dynamic physiological data, extracts and processes latent features from unstructured text / dynamic data to supplement assessment dimensions, calculates net risk through multiple models to improve accuracy, and uses a dynamic solution library to match interventions, thus solving technical problems in a closed loop throughout the entire process.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A VTE intelligent assessment and prevention decision support system includes: an adaptive data acquisition module, an assessment module, a decision-making module, and a tracking and optimization module;

[0008] The adaptive acquisition module is used to receive the target patient's existing medical data and dynamic physiological data, set up a clinical triggering mechanism to obtain targeted clinical data, and configure differentiated permissions.

[0009] The evaluation module is used to screen latent features, and based on time decay correction, combined with the weights assigned to explicit features, generate a fused feature set and construct a dynamic risk profile.

[0010] The decision module is used to construct a multi-fusion model, perform collaborative assessment, calculate the fusion risk probability of the target patient, generate a risk prediction sequence, and generate a net risk probability by combining the bleeding risk probability.

[0011] The tracking and optimization module is used to build a dynamic solution library, intervene and adapt, initially screen suitable solutions, calculate the degree of adaptation, and generate a list of recommended solutions.

[0012] Specifically, the steps of the clinical triggering mechanism include:

[0013] Based on the VTE-related clinical event database, a unique identifier is defined for each type of event, including structured identifiers and unstructured identifiers, thereby generating an identifier lookup table;

[0014] The system monitors the structured data stream of the target patient in real time, extracts the standard coded fields from the data, filters out invalid codes, and generates a list of valid codes, which are then matched with the structured identifiers in the identifier lookup table.

[0015] If a match exists, mark the clinical event as having a structured association and record the event type and risk association degree; if no coded match exists, mark the clinical event as having no structured association.

[0016] The system monitors the unstructured text data stream of the target patient in real time, generates a list of text semantics, and calculates the semantic similarity with each unstructured identifier in the identifier lookup table.

[0017] If the highest similarity is greater than the matching verification threshold, an unstructured associated clinical event is determined to exist; if the highest similarity is less than or equal to the matching verification threshold, an unstructured associated clinical event is marked.

[0018] Specifically, the steps of the clinical triggering mechanism further include:

[0019] Set a time window and compare the structured judgment results with the unstructured judgment results;

[0020] If both events are marked as present and the event types are the same, then a related clinical event is detected; if both events are marked as present but the event types are different, then a review is required.

[0021] If both are marked as no event, it is determined that no associated clinical event was detected;

[0022] If an event is marked by only a single data source, a credibility check is performed. Based on the preset hospital rules, the preset credibility of the data source is retrieved to determine whether a related clinical event has been detected.

[0023] Generate a clinical event table; if no event is detected, do not trigger a targeted request.

[0024] If no event is detected, a targeted trigger signal is generated, and the request priority is calculated to sort the targeted trigger signals in order to obtain targeted clinical data.

[0025] By integrating targeted clinical data with existing medical data and dynamic physiological data, and assembling them into a data context according to the event timeline, an assembled medical dataset is generated.

[0026] Specifically, the steps for screening latent features include:

[0027] Based on the assembled medical dataset, texts that perfectly match the latent feature requirements in the directional trigger signal are selected to generate a set of text fragments;

[0028] Based on the constructed entity recognition model, the text fragment set is subjected to contextual semantic parsing, and a list of feature triples is generated based on contextual logic;

[0029] Based on the list of feature triples, repeating features of the target patient are filtered.

[0030] Based on the data source credibility in the hospital's rules, the repeated features are sorted in descending order, and the feature values ​​with the highest credibility are retained to generate a list of latent features;

[0031] Construct an effective physiological dataset, set an appropriate time period, and obtain the standard deviation of the target patient's hourly heart rate within the appropriate time period;

[0032] The stress level of the target patient is obtained by using a two-level stress threshold.

[0033] Set a minimum duration threshold for a single activity, count the total number of valid activities and the duration of a single valid activity within the appropriate time period, calculate the activity interval, and save the stress level and activity interval to the latent feature list.

[0034] Specifically, the steps for screening latent features include:

[0035] The rate of change in inflammation is calculated based on the D-dimer values ​​at the current time and before the adaptation period.

[0036] The propensity level of the current target patient is determined by using the secondary propensity threshold.

[0037] If only one D-dimer value exists, the prothrombin time is correlated, and the propensity level of the current target patient is determined by using a secondary time threshold.

[0038] Set drug risk intervals, calculate the cumulative duration of medication for VTE-risk drugs, determine the exposure level of the current target patient through the secondary exposure threshold, and save the propensity level and exposure level to the hidden feature list.

[0039] Based on the correlation strength between latent features and VTE risk, the features in the latent feature list are divided into categorical features and continuous features. Categorical features are labeled with levels using a label encoding method, while continuous features are processed using a normalization algorithm.

[0040] The classification features are corrected for time decay by calculating the exponential time coefficient, and the classification feature values ​​are corrected by combining the exponential function, and the latent feature list is updated.

[0041] Specifically, the steps for constructing a dynamic risk profile include:

[0042] Targeted selection of dominant features and quantification to generate a dominant feature set, and calculation of dominant feature weights;

[0043] For latent features, obtain the corresponding correlation strength and the risk correlation of the corresponding clinical events, set the latent correction coefficient, and calculate the latent feature weights;

[0044] By combining the weights of explicit and implicit features, normalization is performed to generate a weighted fusion feature set.

[0045] Key missing features were screened, and similar cases were selected for completion using the K-nearest neighbor imputation method.

[0046] Determine whether there is a logical contradiction in the combination of explicit and implicit features that are logically related. If a contradiction exists, mark the contradictory feature combination as pending review.

[0047] Based on the fusion feature set and combined with the timeline of clinical events, a dynamic risk profile is constructed;

[0048] The design incorporates a real-time update mechanism and sets a basic update frequency. Once a new clinical event or new patient data is detected, the screening of latent features is automatically triggered.

[0049] Specifically, the steps of collaborative assessment include:

[0050] Load the multi-fusion model for predicting risk probability, obtain the validation set accuracy of each model in the multi-fusion model, and calculate the fusion weight of the individual model.

[0051] The fused feature set is input into the multi-fusion model to obtain the risk probability output by each model;

[0052] Calculate the fusion risk probability of the target patient, and record the feature response vector of each model; generate a response list;

[0053] The characteristic change data of the target patient are extracted from the dynamic risk profile, and multimodal model fusion is performed to generate a historical risk sequence.

[0054] Using exponential smoothing, the historical risk sequence is used to predict trends and calculate confidence intervals, generating a risk prediction sequence.

[0055] Specifically, the steps of collaborative assessment also include:

[0056] Based on the risk prediction sequence, the predicted change slope is calculated, and the trend type of the target patient is determined by using a secondary trend threshold.

[0057] Based on the feature response vectors of all models, and combined with the fusion weights of each model, the contribution of a single feature is calculated.

[0058] Construct a single-feature contribution ranking table and retain the top performers. The core characteristic that has the greatest impact on risk;

[0059] Based on the single feature contribution ranking table, interactive feature pairs are constructed, the interaction contribution of the interactive feature pairs is calculated, and a feature interaction contribution table is generated.

[0060] Bleeding risk features are selected from the fused feature set, and the bleeding risk probability is obtained using the bleeding risk model.

[0061] Calculate the net risk probability and determine the net risk level using the secondary net judgment threshold;

[0062] The results of each stage of the collaborative assessment are integrated into a risk assessment result package, and the assessment results are updated synchronously to the dynamic risk profile.

[0063] Specifically, the steps for intervention and adaptation include:

[0064] Based on the prevention type, net risk level, and bleeding risk level, a basic intervention plan is generated.

[0065] Retrieve historical data from within the hospital, calculate the effectiveness and adherence rate of each basic intervention protocol in patients with the same combination, calculate the in-hospital fit weight of each protocol, and form a dynamic protocol library.

[0066] Based on the net risk level and bleeding risk level, suitable solutions are initially screened from the dynamic solution library;

[0067] The intervention single feature is extracted from the single feature contribution ranking, the coverage of each adaptation scheme is counted, the coverage coefficient is calculated, and thus the feature matching degree is obtained.

[0068] Intervention interaction pairs are extracted from the feature interaction contribution table to obtain coverage integrity and thus obtain the interaction matching degree. If there are multiple interaction feature pairs, the interaction matching degree is calculated by the arithmetic mean.

[0069] By combining the in-hospital adaptation weights, feature matching degrees, and interaction matching degrees of each adaptation scheme, and by weighted summation, the adaptation degree of each adaptation scheme is calculated, and a list of recommended schemes is generated.

[0070] After the plan is implemented, the dynamic risk profile is automatically updated, the net risk probability is recalculated, the risk reduction rate is calculated, and the effectiveness of the current intervention is determined by the secondary classification threshold, generating an intervention effect report.

[0071] The beneficial effects of this invention are:

[0072] By receiving existing medical data and incorporating dynamic physiological data, the system addresses data source blind spots. It dynamically assembles data context using clinical triggering mechanisms, avoiding static data lag. Semantic tags and access control anonymization further ensure data validity and security. Implicit features are filtered and corrected from unstructured text and dynamic data, and combined with explicit features to generate a fusion feature set and dynamic risk profile. This fills the gaps in implicit feature coverage, improving assessment accuracy and ease of use for healthcare professionals. Multi-fusion model collaborative assessment avoids single-model bias, predicts risk trends, quantifies uncertainty, clarifies core risk characteristics and synergistic effects, and calculates net risk based on bleeding risk, making assessments more precise. A dynamic treatment library is built based on risk and in-hospital data. Adaptability is calculated through feature and interaction matching to recommend suitable treatments. Post-implementation tracking updates the profile, and solutions address coarse-grained issues, achieving closed-loop optimization from assessment to intervention, comprehensively improving the accuracy and effectiveness of VTE prevention and treatment. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the structural principle of a VTE intelligent assessment and prevention decision support system.

[0074] Figure 2 This is a flowchart of the clinical triggering mechanism in this invention;

[0075] Figure 3 This is a flowchart of the process for selecting latent features in this invention;

[0076] Figure 4 This is a flowchart of the collaborative evaluation process in this invention. Detailed Implementation

[0077] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0078] Example 1

[0079] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a VTE intelligent assessment and prevention decision support system, including: an adaptive acquisition module, an assessment module, a decision module, and a tracking and optimization module;

[0080] The adaptive acquisition module receives existing medical data from target patients and incorporates dynamic physiological data generated by wearable devices within the hospital, such as vital signs and activity levels. It also sets up a clinical trigger mechanism. Once a clinical event is detected, such as the entry of medical orders or the start of surgery, a targeted request is triggered to dynamically assemble the data context required for evaluation. This allows for the acquisition of corresponding targeted clinical data from the target data source, avoiding static data lag and ensuring data timeliness and relevance to the scenario. Semantic tags are added to each data point, and differentiated data upload, viewing, modification, and decision-making permissions are set based on different roles. Sensitive information is anonymized or authorized for use, and an authorized data view is generated.

[0081] Among them, the existing medical data refers to the medical records that target patients have uploaded through the front end, such as medical records from other hospitals, examination reports from other places, long-term medication lists, historical medical records, and admission diagnoses. It covers both data from the hospital and other hospitals, solves the blind spots in data sources, and automatically verifies the data format, such as the clarity of PDFs and images. It triggers reminders to supplement and upload data for blurry or missing key information to ensure data validity.

[0082] Specifically, the steps involved in the clinical triggering mechanism include:

[0083] The system calls upon the existing VTE-related clinical event database within the hospital and defines a unique identifier for each type of event in the database, including structured identifiers and unstructured identifiers, thereby generating an identifier lookup table. The structured identifier is the medical standard code corresponding to the event, such as the surgical CPT code corresponding to general anesthesia, while the unstructured event identifier is the combination of text keywords or semantic rules corresponding to the event.

[0084] Through standardized interfaces at the hospital, the structured data stream of the target patient is monitored in real time. Based on the unique identifier of the target patient (such as the hospital number), related data is filtered, and standard coding fields such as medical order codes, examination codes, and diagnosis codes are extracted from the data. Invalid codes such as blank codes, incorrect codes that do not match the hospital coding standards, and non-VTE related codes are filtered out, thereby generating a list of valid codes.

[0085] The effective coding list is precisely matched with the structured identifiers in the identifier lookup table. A complete match is considered a match. If a match exists, a structured associated clinical event is marked, and the event type and risk correlation are recorded. If no coding matches, an unstructured associated clinical event is marked. Based on the historical VTE case dataset, the probability of data occurrence when the known VTE risk exists is calculated. The prior probability of VTE risk is set based on the hospital's VTE incidence rate over the past year. The probability of data occurrence in all patients is calculated, and the risk correlation between data and VTE risk is calculated using Bayesian confidence.

[0086] Through electronic medical records and nursing document interfaces, the unstructured text data stream of the target patient is monitored in real time. Using pre-trained NLP models in the medical field, such as the BioBERT fine-tuning model, the text data stream is segmented, word-segmented, and semantically parsed. Based on the three dimensions of time, behavior, and state, the core semantics in the text are extracted, such as the second day after surgery and bedridden, to generate a list of text semantics.

[0087] The cosine similarity algorithm is used to calculate the semantic similarity between the text semantic list and each unstructured identifier in the identifier lookup table, and the identifier with the highest similarity is selected as the candidate identifier. If the highest similarity is greater than the preset matching verification threshold, it is determined that there is an unstructured associated clinical event, and the event type, risk correlation, and implicit feature requirements are recorded, such as the implicit feature requirement of activity intention in postoperative activity assessment. If the highest similarity is not greater than the matching verification threshold, it is marked as an unstructured associated clinical event. Among them, implicit features cannot be directly obtained from hospital structured data (such as age, surgical code, single test value) and need to be mined and extracted from unstructured text and routine dynamic data, and are directly related to VTE risk. For non-obvious patient status or trend information, such as activity level, stress status, coagulation correlation, and drug metabolism, the following points need to be considered: For activity level, it is not possible to directly read whether the amount of activity is insufficient; it is necessary to extract patient activity status information from nursing records and manual dynamic records. For stress status, it is not possible to directly determine stress or inflammation status from a single test value; it is necessary to extract patient physiological / psychological status information from dynamic heart rate data and text emotional descriptions. For coagulation correlation, it is not possible to directly determine thrombosis tendency from a single coagulation test value; it is necessary to calculate trend information based on dynamic test data. For drug metabolism, it is not possible to directly determine thrombosis risk from the name of the medication; it is necessary to combine the duration of medication and medication scenario to associate risk information.

[0088] Because the same clinical event may contain both structured and unstructured text—for example, general anesthesia surgery may have both CPT codes and surgical record text—double verification is required to ensure accuracy. A time window is set based on the temporal continuity of the clinical event, and the structured and unstructured judgment results are compared. If both are marked as having an event and the event types are consistent, a preliminary confirmation of a related clinical event is made, integrating event type, risk correlation, implicit feature requirements, and target data source. If both are marked as having an event but the event types are inconsistent, manual review is performed. If both are marked as not having an event, a preliminary determination is made that no related clinical event was detected.

[0089] If an event is marked by only a single data source, there is uncertainty. In this case, a credibility check is performed. Based on the preset internal hospital rules, the preset credibility of the data source is retrieved. If the credibility exceeds the credibility check threshold, it is determined that a related clinical event has been detected, and the event information is integrated. If the credibility is less than the credibility check threshold, it is determined that no related clinical event has been detected. The internal hospital rules include the credibility of operating room / ICU data sources, the credibility of general ward nursing data sources, and the credibility of patient self-report data sources.

[0090] A clinical event table is generated based on the final clinical event determination results; if no event is detected, there is no need to trigger a targeted request.

[0091] If no event is detected, a targeted trigger signal is immediately generated, including event type, risk correlation, implicit feature requirements, target data source list, and judgment timestamp. Various events in the VTE-related clinical event database are divided into three correlation levels based on risk correlation: high correlation events, medium correlation events, and low correlation events. Based on the correlation level of the clinical event in the targeted trigger signal, the corresponding implicit correlation level is configured, and a balance coefficient is set as the weight of the implicit correlation level. The timeliness is configured through the time elapsed since the data generation. The request priority is calculated through the weighted operation of the implicit correlation level and the timeliness to ensure that high-value data is obtained first. The targeted trigger signals are sorted according to the request priority to obtain the targeted clinical data responded by the target data source.

[0092] Integrate targeted clinical data (if available) with existing medical data and dynamic physiological data, assemble them into a data context according to the event timeline, such as the end of surgery - heart rate 1 hour after surgery - activity assessment 2 hours after surgery, add association markers between events and data, and generate an assembled medical dataset.

[0093] A three-level semantic tagging system is constructed. The first-level tags are data types, such as existing medical data, dynamic physiological data, and unstructured text data. The second-level tags are risk-related dimensions, such as coagulation function, activity ability, stress state, and surgical trauma. The third-level tags are implicit feature indicators, such as the number of steps, frequency of getting out of bed, and description of activity intention in activity ability.

[0094] For the assembled medical dataset, differential mapping is used according to data type. For structured data, the field name is directly matched with the secondary label. For unstructured text, an attention-based text-label matching model is used to calculate the semantic similarity between the text and the tertiary label. For example, the patient's reluctance to get out of bed corresponds to the description of activity willingness in the activity ability. Based on the tertiary label with the highest similarity, the most matching label is automatically labeled.

[0095] The assessment module is used to filter latent features that match VTE risk scenarios from unstructured text and dynamic physiological data, exclude irrelevant information, correct latent feature values ​​through time decay, filter and quantify explicit features, assign weights to the two types of features, complete missing features, verify logical consistency, generate a fused feature set, and construct a dynamic risk profile in combination with the clinical timeline to clearly display information and warn of risks. It also has a real-time update mechanism to ensure synchronization with the patient's status, improve the accuracy of VTE risk assessment and the convenience of medical staff to use.

[0096] Specifically, the steps for screening latent features include:

[0097] To avoid extracting features solely based on semantic labels without considering clinical events, and to ensure a strong correlation between features and the current VTE risk scenario, we screen unstructured text from the assembled medical dataset that perfectly matches the implicit feature requirements in the clinical event trigger signals, excluding irrelevant text. For texts that match the requirements, we further screen text fragments whose semantic labels are classified into three levels as pointing to implicit features, ensuring a strong correlation between text and implicit feature types. We also bind the text to the time window of the clinical event, such as within 72 hours after surgery, retaining only text within this time range and avoiding the use of historical texts unrelated to the current event. This generates a set of text fragments that contain a triple match of event requirements, semantic labels, and time windows.

[0098] Based on the constructed entity recognition model, contextual semantic parsing is performed on the text fragment set to analyze the main logic of sentences. For example, if a patient is unwilling to get out of bed due to postoperative wound pain for nearly 24 hours, the main logic is postoperative pain. This avoids extracting the unwillingness to get out of bed in isolation and ignoring the causal context of postoperative pain. Based on the contextual logic, triples containing feature type, feature, and causal / temporal basis are extracted to generate a list of feature triples with contextual basis. In this process, the BioBERT+CRF model is loaded and fine-tuned through in-hospital VTE medical records to optimize the recognition accuracy of professional terms in order to generate the entity recognition model.

[0099] Based on the feature triplet list, the same type of features repeatedly extracted from the same patient within the same time period are marked as duplicate features. Combining the data source credibility in the hospital rules, the data source credibility of the duplicate features is sorted in descending order, and the feature value with the highest credibility is retained. The contextual basis of different data sources is merged to generate a list of latent features.

[0100] The system filters device data that is associated with latent feature types from dynamic physiological data, such as heart rate monitoring data corresponding to stress levels, and removes invalid data due to device failure or signal interruption. The valid data is then correlated with the timeline of clinical events, and historical data before the events occur is excluded, thereby generating a valid physiological dataset.

[0101] Because different patients have different physiological bases and equipment conditions, a fixed threshold will lead to calculation bias. The type of clinical event will affect the stress level benchmark, so the threshold needs to be corrected. Set an appropriate time period, obtain the standard deviation of the hourly heart rate of the target patient within the appropriate time period based on the effective physiological dataset, and obtain the stress level of the target patient, including high stress, moderate stress and low stress, through the set secondary stress classification threshold.

[0102] A minimum duration threshold for a single activity is set to avoid misjudging a 1-minute time spent getting out of bed to retrieve an item as a valid activity. The total number of valid activities and the duration of a single valid activity are counted within the adaptation period to calculate the cumulative activity duration within the adaptation period. The activity interval is obtained by dividing the difference between the adaptation period and the cumulative activity by the number of activities. The stress level and the activity interval are saved to a latent feature list, and the source of the calculated data is labeled as a dynamic physiological dataset. The duration of a single valid activity is determined by obtaining the start and end times.

[0103] From the assembled medical dataset, dynamic structured data within the past 72 hours is prioritized for screening, while static data is excluded. Based on the D-dimer value at the current time and before the matching period, the rate of change of inflammation is calculated using the deviation rate calculation formula. The propensity level of the current target patient is determined by a preset secondary propensity threshold, including high propensity to thrombosis, moderate propensity to thrombosis, and low propensity to thrombosis. If only one D-dimer value exists, it is correlated with the prothrombin time. The propensity level of the current target patient is determined by a preset secondary time threshold, thus avoiding the failure of static values ​​to reflect risk changes.

[0104] Set drug risk intervals, retrieve medication records for drug risk intervals, calculate the cumulative duration of medication for VTE risk drugs, determine the exposure level of the current target patient through preset secondary exposure thresholds, including high exposure, moderate exposure and low exposure, and save the propensity level and exposure level to a hidden feature list.

[0105] Based on the correlation strength between latent features and VTE risk, the features in the latent feature list are divided into categorical features and continuous features. Categorical features are strongly correlated features, such as stress level and exposure level, and are labeled using a tag encoding method. Continuous features are weakly correlated features, such as bed rest duration and inflammation change rate, and are processed using a Min-Max normalization algorithm. The extreme values ​​are taken from the hospital's VTE case set of the past year. Among them, the correlation strength is based on the clinical mechanism and guidelines of VTE, relies on medical common sense and risk pathology logic, and is a rule-driven classification result, which can be directly called in the hospital terminal or set by those skilled in the art.

[0106] Time decay correction is applied to categorical features by calculating an exponential time coefficient using the time elapsed since the clinical event and a time decay constant. This coefficient is then combined with an exponential function to correct the categorical feature values ​​and update the latent feature list. The expression for time decay correction is shown below:

[0107]

[0108] In the formula, , These are the classification feature values ​​before and after time decay correction. The time elapsed since the occurrence of the clinical event. is the time decay constant.

[0109] Specifically, the steps for constructing a dynamic risk profile include:

[0110] From the assembled medical dataset, structured explicit features directly related to VTE risk assessment were selectively screened, while explicit data unrelated to VTE risk, such as patient occupation and history of non-anticoagulant allergies, were excluded. Explicit features included basic characteristics, disease and medical history characteristics, treatment and procedural characteristics, and physiological indicators. Basic characteristics included patient age, body mass index, and gender; disease and medical history characteristics included whether the patient had a history of VTE, whether they had malignant tumors, and whether they had experienced a stroke within the past month; treatment and procedural characteristics included the type of surgery the patient underwent and whether they had undergone central venous catheterization; and physiological indicators included the patient's absolute D-dimer value, platelet count, and prothrombin time.

[0111] Based on the hospital's clinical guidelines for VTE risk assessment, such as the Padua score and Caprini score, the dominant features were standardized and quantified. For example, after assigning values ​​based on the scoring criteria, normalization was performed to form a set of dominant features with a uniform format that can be directly integrated with the latent features.

[0112] After standardizing the dominant features, weights need to be assigned to the dominant and latent features respectively to distinguish the importance of different features to VTE risk assessment and avoid weakly associated features from interfering with the assessment results. For dominant features, the feature weights specified in the hospital scoring criteria are directly used. The dominant feature weights are obtained by calculating the ratio between the feature score obtained without normalization and the full score of the scoring criteria, ensuring that the weight allocation conforms to clinical consensus.

[0113] For latent features, obtain the corresponding correlation strength and the risk correlation of the corresponding clinical events, set the latent correction coefficient, calculate the correction weight by multiplying the latent correction coefficient, the risk correlation, and the correlation strength, and calculate the latent feature weight by summing the correction weight and the correlation strength.

[0114] By combining the weights of all explicit and implicit features and normalizing them, the total weight of all explicit and implicit features is ensured to be 1, thus avoiding deviations in subsequent risk assessment results due to imbalances in the total weight. This generates a weighted fusion feature set.

[0115] Due to the certain missing rate of hospital clinical data, such as missing nursing records and delayed test results, it is necessary to handle the missing key features during the fusion process. At the same time, the logical consistency between explicit and implicit features must be verified to ensure that the fused feature set is complete and reasonable. Missing features are identified by traversing the fused feature set, screening features with a correlation strength greater than the preset correlation strength threshold for VTE risk, and removing features that are irrelevant to the current patient scenario. The retained features are marked as key missing features, such as the activity interval in the implicit feature being missing due to the lack of nursing records, and the platelet count in the explicit feature being missing due to test delays. At the same time, irrelevant missing features that do not need to be filled in are excluded. For example, if the patient has not undergone surgery, the surgery type feature does not need to be filled in.

[0116] For key missing features, a similar case imputation method was used to complete them. From the hospital's historical case data, the K-nearest neighbor imputation method was used to screen for similar cases that matched the target patient in age range, primary diagnosis, and current clinical event type. The mean of the missing features corresponding to the similar cases was taken as the feature value of the target patient, and the imputation source was marked, such as the similar case number and data collection time, to facilitate subsequent verification by medical staff. Specifically, the K-nearest neighbor imputation method was used to calculate the consistency between the target patient and the hospital's historical case data, and the data was sorted in descending order, with the most consistent cases ranked first. This medical record is considered a similar medical record;

[0117] After the feature set is filled in, a logical consistency check is performed on the dominant and latent features. For combinations of dominant and latent features that are logically related, such as a high D-dimer index in the dominant feature and a thrombosis tendency level in the latent feature, it is determined whether there is a logical contradiction between the two. For example, the dominant feature indicates high risk but the latent feature indicates low risk. If a contradiction exists, the contradictory feature combination is marked as pending review and sent to a clinical specialist for manual confirmation, such as verifying whether there is a test error or data extraction deviation. If there is no contradiction or the contradiction is resolved after review, and the overall feature consistency meets the preset standard, the fused feature set is deemed valid.

[0118] Based on a fusion feature set and combined with a clinical event timeline, a dynamic risk profile is constructed, including a basic information block, a full-feature real-time block, a feature trend block, a risk warning block, and a data traceability block. The basic information block displays the target patient's basic identity information, current clinical event, and the latest profile update timestamp, facilitating quick identification of the patient and current clinical stage by medical staff. The full-feature real-time block displays the quantified values ​​and weights of all explicit and implicit features, with features reaching high-risk thresholds highlighted to help medical staff quickly locate core risk features. The feature trend block uses a line graph to show the dynamic changes of key features over the past 72 hours, marking important clinical event nodes on the curves to visually present the changing trend of the patient's risk status. The risk warning block automatically identifies high-weight, high-quantified feature combinations and generates targeted warning prompts. The data traceability block labels each feature with detailed data source information, allowing medical staff to trace the data source and verify data authenticity when there are doubts about the features.

[0119] At the same time, a real-time update mechanism is designed and a basic update frequency is set. Once a new clinical event or new patient data is detected, the screening of latent features is automatically triggered to ensure that the profile is synchronized with the patient's clinical status.

[0120] The decision-making module is used to construct multi-fusion models for collaborative evaluation, avoid single-model bias to improve generalization, calculate the fusion risk probability of the target patient, record feature responses simultaneously, generate risk prediction sequences based on the target patient's recent characteristics, predict risk trends and quantify uncertainty, calculate the contribution of single features to identify core features and clarify risk sources, analyze interactive feature pairs to obtain synergistic risk effects; then, a bleeding risk model is constructed and the bleeding risk probability is calculated, and combined with the fusion risk probability, the net risk probability is generated.

[0121] Specifically, the steps of collaborative assessment include:

[0122] Single models are susceptible to the influence of data distribution. For example, XGBoost is sensitive to extreme values, and logistic regression has poor adaptability to nonlinear features. To avoid the bias of a single model affecting the contribution assessment, multiple models are fused to combine their advantages, resulting in better generalization on small sample data and avoiding black-box reliance on a single model. A multi-fusion model predicting risk probabilities is loaded, including an XGBoost model, a logistic regression model, and a lightweight neural network. The multi-fusion model is trained using a historical VTE case set from within the hospital, covering different risk assessment scenarios to improve generalization. Among these, the XGBoost model is used... For performing small-sample classification tasks, the training data consists of VTE and non-VTE cases within the hospital over the past 3 years. Grid search optimization is used to ensure adaptability to the hospital data. The logistic regression model is used to assist in judging linear correlation features, such as the linear relationship between D-dimer and VTE risk. The training data is consistent with the XGBoost model. The lightweight neural network (MLP) contains only 2 hidden layers, which is suitable for the low computing power environment of the hospital and avoids the computational delay caused by complex models. All models have been retrospectively validated within the hospital, and the prediction accuracy has been verified using clinical cases from the past six months to ensure that they meet the clinical needs of the hospital.

[0123] The validation set accuracy of each model in the multi-fusion model is obtained. The fusion weight of a single model is calculated by the ratio of the validation set accuracy of a single model to the sum of the validation set accuracies of the three models. This ensures that the model with higher accuracy has a higher proportion in the fusion, thereby improving the overall prediction reliability. The fusion feature set is then input into the multi-fusion model to obtain the risk probability output by each model. The fusion risk probability of the target patient is calculated by weighted averaging. At the same time, the feature response vector of each model is recorded, which is the contribution value of each feature in the fusion feature set to the risk probability of a single model. For example, if an increase in a feature value causes the model's prediction probability to rise or fall, this provides multi-dimensional evidence for subsequent interpretability analysis, thereby generating a response list.

[0124] Extract characteristic change data of the target patient in the past 72 hours from the dynamic risk profile, such as bed rest time, D-dimer value and stress level recorded every 12 hours, perform multimodal model fusion to generate historical risk sequence, which includes the fused risk probability of multiple time points to fully reflect the patient's recent risk changes.

[0125] Exponential smoothing methods, such as the Holt-Winters seasonalless model, are used to predict trends in historical risk sequences in order to generate forecasts for future periods. The risk prediction probability at each time point is calculated. At the same time, based on the historical prediction error dataset, the 95% confidence interval of each risk prediction probability is calculated using the normal distribution confidence interval method to quantify the prediction uncertainty and thus generate a risk prediction sequence.

[0126] Based on the risk prediction sequence, the predicted change slope is calculated, and the trend type of the target patient is determined by a preset secondary trend threshold, including an upward trend, a stable trend, and a downward trend. The secondary trend threshold includes a first trend threshold and a second trend threshold, which have opposite signs and the same absolute value.

[0127] Since the contribution values ​​in the response list are the independent contributions of features in a single model to the model's own risk probability, rather than the contribution of features to the fused risk probability, and since the fused risk probability is the result of the combined effect of all features in the multimodal model, it is impossible to directly reverse-engineer the contribution of a single feature. Based on the feature response vectors of all models and combined with the fusion weights of each model, the contribution of a single feature is calculated. A positive contribution indicates that an increase in the corresponding feature value will lead to an increase in risk probability, such as an increase in bed rest time. A negative contribution indicates that an increase in the feature value will lead to a decrease in risk probability, such as an increase in postoperative activity time. The single feature contributions are sorted in descending order by absolute value to generate a ranking table, retaining the top... Identify the core characteristics that have the greatest impact on risk and clarify the main sources of current risk.

[0128] The top three core features are selected from the single feature contribution ranking table. Based on the interaction of two features, interactive feature pairs are constructed. Gradient boosting tree and interaction effect algorithm are used to analyze the joint impact of the simultaneous change of the two features on the probability of fusion risk, so as to calculate the interaction contribution of the interactive feature pairs and generate a feature interaction contribution table. Among them, once the interaction contribution is positive, it means that when the two exist at the same time, the risk contribution exceeds the sum of the individual contributions of the two, reflecting the synergistic risk effect.

[0129] Bleeding risk features, including direct and indirect features, were screened from the fusion feature set. A bleeding risk model was constructed using a logistic regression model and trained using historical VTE prophylaxis cases and cases without bleeding. The bleeding risk features and their assigned weights were then input into the bleeding risk model to obtain the probability of bleeding risk. Direct features include explicit features (platelet count, prothrombin time, presence of active bleeding) and implicit features (exposure duration of combined anticoagulant and antiplatelet therapy). Indirect features are those VTE risk features that have cross-correlation with bleeding risk, such as a significantly elevated D-dimer level which may indicate coagulation dysfunction and indirectly link to bleeding risk. The assigned weights are the weights corresponding to the direct and indirect features.

[0130] The net risk probability is calculated based on the difference between the product of the fusion risk probability, the balance coefficient, and the bleeding risk probability. The net risk level is then determined by a preset secondary net judgment threshold, including high net risk, medium net risk, and low net risk.

[0131] The results of each stage of the collaborative assessment are integrated into a risk assessment result package, and the assessment results are updated synchronously to the dynamic risk profile. The net risk level and risk trend prompts are added to the risk warning block, and the risk prediction curve is added to the characteristic trend block. The results are pushed in a tiered manner according to the net risk level. For example, high net risk results are pushed to the attending physician's workstation within 1 hour, medium net risk results are pushed to the ward nurse's mobile terminal within 4 hours, and low net risk results are only updated in the dynamic risk profile without being actively pushed, so as to avoid information overload for medical staff and ensure that high priority risks are addressed in a timely manner.

[0132] The tracking and optimization module is used to construct basic intervention plans based on prevention type and risk level, label core intervention feature tags, match implicit needs to ensure the plan's relevance, calculate in-hospital fit weights based on historical in-hospital data, supplement contraindications and monitoring requirements, build a dynamic plan library, and perform intervention adaptation. It initially screens out suitable plans, calculates feature matching degree and interaction matching degree, and calculates fit degree by combining in-hospital fit weights, thereby generating a list of recommended plans. After the plan is implemented, it updates the dynamic risk profile and evaluates the intervention effect.

[0133] Specifically, the steps for intervention and adaptation include:

[0134] Based on the hospital's clinical guidelines for VTE prevention and treatment, we break down the basic intervention plans recommended by the guidelines according to three dimensions: prevention type, net risk level, and bleeding risk level. We then associate these plans with the implicit feature requirements in the dynamic risk profile and label each plan with core intervention feature tags to ensure that the plan can specifically address the identified core risk features.

[0135] Historical data on VTE prevention and treatment protocols and clinical outcomes over the past three years were retrieved from within the hospital. The actual effectiveness (e.g., whether the incidence of VTE decreased after implementation) and adherence rate (e.g., whether nurses followed instructions and whether patients cooperated) of each basic intervention protocol were statistically analyzed in patients with the same net risk level and bleeding risk combination. Based on the protocol effectiveness and adherence rate, the in-hospital fit weight of each protocol was calculated; protocols with higher effectiveness and adherence rates received higher weights. Contraindications and monitoring requirements were also added to each protocol to form a dynamic protocol library. The expression is shown below:

[0136]

[0137] In the formula, To adapt weights within the hospital, The effectiveness of the solution in the corresponding risk portfolio. To achieve compliance rate in program implementation;

[0138] Based on net risk level and bleeding risk level, suitable protocols were initially screened from the dynamic protocol library, while protocols with contraindications were excluded. The suitable protocols were those that met the net risk level and bleeding risk level of the target patients.

[0139] For the adaptation schemes after initial screening, the top two features with the highest absolute value of contribution are selected from the single feature contribution ranking and defined as intervention single features. Through the preset intrusive single feature labels for each scheme, the number of intervention single features covered by the adaptation scheme is counted. Based on the ratio of the number of covered features to the total number of intervention single features, the coverage coefficient is calculated. The feature matching degree is calculated by multiplying the coverage coefficient and the intervention intensity coefficient.

[0140] Interaction feature pairs with interaction contribution greater than a preset threshold are selected from the feature interaction contribution table and defined as intervention interaction pairs. Based on whether the adaptation scheme covers the two features in the key interaction pair, the coverage integrity is configured, including simultaneous coverage, coverage of only one feature, and no coverage, in order to obtain the interaction matching degree. If there are multiple interaction feature pairs, the interaction matching degree is calculated by the arithmetic mean.

[0141] By combining the in-hospital adaptation weights, feature matching degrees, and interaction matching degrees of each adaptation scheme, the adaptation degree of each adaptation scheme is calculated by weighted summation, and the adaptation schemes are sorted in descending order based on the adaptation degree to generate a list of recommended schemes.

[0142] After the plan is implemented, the dynamic risk profile is automatically updated, the net risk probability is recalculated, the risk is compared with the risk before the intervention, the risk reduction rate is calculated, and the effectiveness of the current intervention is determined by the preset secondary classification threshold, including effective, partially effective, and ineffective. An intervention effect report is generated and pushed to the doctor's workstation to help determine whether the plan should be adjusted.

[0143] Working principle and effects:

[0144] It receives existing medical data from both internal and external hospitals of target patients, incorporates dynamic physiological data, and automatically verifies data validity to address data source blind spots. Simultaneously, it sets up a clinical trigger mechanism to trigger targeted data requests when events such as medical order entry and surgery commencement are detected, dynamically assembling data context to avoid static data lag and ensure data timeliness and scenario relevance, providing a complete and up-to-date data foundation for subsequent assessments. It filters VTE risk-related latent features from unstructured text and dynamic physiological data, corrects for time decay, and combines this with explicit feature quantification, weighting, missing data completion, and logical consistency verification to generate a fused feature set and construct a dynamic risk profile, effectively capturing the unstructured and dynamic data corresponding to latent features. This system provides dynamic information to fill gaps in assessment data dimensions, updates profiles in real time, enhances ease of use for medical staff, and provides comprehensive data support for risk assessment. Through collaborative assessment using multiple fusion models, it avoids the bias of single models to improve generalization, calculates fusion risk probabilities, analyzes the contribution of single features to clarify risk sources, captures synergistic risks through interaction effects, and obtains net risk by combining with a bleeding risk model, significantly improving the accuracy of VTE risk assessment. A dynamic protocol library is built based on prevention types and risk levels, labeling core intervention features to match implicit needs, matching protocols according to risk and features and calculating suitability recommendations. After execution, the profile is updated to assess the effect, ensuring the precision and effectiveness of intervention protocols and contributing to efficient VTE risk prevention and control.

[0145] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A VTE intelligent assessment and prophylaxis decision support system, characterized in that, include: Adaptive acquisition module, evaluation module, decision-making module, and tracking optimization module; The adaptive acquisition module is used to receive the target patient's existing medical data and dynamic physiological data, set up a clinical triggering mechanism to obtain targeted clinical data, and configure differentiated permissions. The evaluation module is used to screen latent features, and based on time decay correction, combined with the weights assigned to explicit features, generate a fused feature set and construct a dynamic risk profile. The decision module is used to construct a multi-fusion model, perform collaborative assessment, calculate the fusion risk probability of the target patient, generate a risk prediction sequence, and generate a net risk probability by combining the bleeding risk probability. The tracking and optimization module is used to build a dynamic solution library, perform intervention and adaptation, initially screen suitable solutions, calculate the degree of adaptation, and generate a list of recommended solutions; where VTE refers to venous thromboembolism. The steps of the clinical triggering mechanism include: Based on the VTE-related clinical event database, a unique identifier is defined for each type of event, including structured identifiers and unstructured identifiers, thereby generating an identifier lookup table; The system monitors the structured data stream of the target patient in real time, extracts the standard coded fields from the data, filters out invalid codes, and generates a list of valid codes, which are then matched with the structured identifiers in the identifier lookup table. If a match exists, mark the clinical event as having a structured association and record the event type and risk association degree; if no coded match exists, mark the clinical event as having no structured association. The system monitors the unstructured text data stream of the target patient in real time, generates a list of text semantics, and calculates the semantic similarity with each unstructured identifier in the identifier lookup table. If the highest similarity is greater than the matching verification threshold, an unstructured associated clinical event is determined to exist; if the highest similarity is less than or equal to the matching verification threshold, an unstructured associated clinical event is marked.

2. The VTE intelligent assessment and prophylaxis decision support system according to claim 1, wherein, The steps of the clinical triggering mechanism also include: Set a time window and compare the structured judgment results with the unstructured judgment results; If both events are marked as present and the event types are the same, then a related clinical event is detected; if both events are marked as present but the event types are different, then a review is required. If both are marked as no event, it is determined that no associated clinical event was detected; If an event is marked by only a single data source, a credibility check is performed. Based on the preset hospital rules, the preset credibility of the data source is retrieved to determine whether a related clinical event has been detected. Generate a clinical event table; if no event is detected, do not trigger a targeted request. If no event is detected, a targeted trigger signal is generated, and the request priority is calculated to sort the targeted trigger signals in order to obtain targeted clinical data. By integrating targeted clinical data with existing medical data and dynamic physiological data, and assembling them into a data context according to the event timeline, an assembled medical dataset is generated.

3. The VTE intelligent assessment and prophylaxis decision support system according to claim 2, wherein, The steps for screening latent features include: Based on the assembled medical dataset, texts that perfectly match the latent feature requirements in the directional trigger signal are selected to generate a set of text fragments; Based on the constructed entity recognition model, the text fragment set is subjected to contextual semantic parsing, and a list of feature triples is generated based on contextual logic; Based on the list of feature triples, repeating features of the target patient are filtered. Based on the data source credibility in the hospital's rules, the repeated features are sorted in descending order, and the feature values ​​with the highest credibility are retained to generate a list of latent features; Construct an effective physiological dataset, set an appropriate time period, and obtain the standard deviation of the target patient's hourly heart rate within the appropriate time period; The stress level of the target patient is obtained by using a two-level stress threshold. Set a minimum duration threshold for a single activity, count the total number of valid activities and the duration of a single valid activity within the appropriate time period, calculate the activity interval, and save the stress level and activity interval to the latent feature list.

4. The VTE intelligent assessment and prophylaxis decision support system according to claim 3, wherein, The steps for screening latent features include: The rate of change in inflammation is calculated based on the D-dimer values ​​at the current time and before the adaptation period. The propensity level of the current target patient is determined by using the secondary propensity threshold. If only one D-dimer value exists, the prothrombin time is correlated, and the propensity level of the current target patient is determined by using a secondary time threshold. Set drug risk intervals, calculate the cumulative duration of medication for VTE-risk drugs, determine the exposure level of the current target patient through the secondary exposure threshold, and save the propensity level and exposure level to the hidden feature list. Based on the correlation strength between latent features and VTE risk, the features in the latent feature list are divided into categorical features and continuous features. Categorical features are labeled with levels using a label encoding method, while continuous features are processed using a normalization algorithm. The classification features are corrected for time decay by calculating the exponential time coefficient, and the classification feature values ​​are corrected by combining the exponential function, and the latent feature list is updated.

5. The VTE intelligent assessment and prevention decision support system according to claim 4, characterized in that, The steps to build a dynamic risk profile include: Targeted selection of dominant features and quantification to generate a dominant feature set, and calculation of dominant feature weights; For latent features, obtain the corresponding correlation strength and the risk correlation of the corresponding clinical events, set the latent correction coefficient, and calculate the latent feature weights; By combining the weights of explicit and implicit features, normalization is performed to generate a weighted fusion feature set. Key missing features were screened, and similar cases were selected for completion using the K-nearest neighbor imputation method. Determine whether there is a logical contradiction in the combination of explicit and implicit features that are logically related. If a contradiction exists, mark the contradictory feature combination as pending review. Based on the fusion feature set and combined with the timeline of clinical events, a dynamic risk profile is constructed; The design incorporates a real-time update mechanism and sets a basic update frequency. Once a new clinical event or new patient data is detected, the screening of latent features is automatically triggered.

6. The VTE intelligent assessment and prevention decision support system according to claim 5, characterized in that, The steps of collaborative assessment include: Load the multi-fusion model for predicting risk probability, obtain the validation set accuracy of each model in the multi-fusion model, and calculate the fusion weight of the individual model. The fused feature set is input into the multi-fusion model to obtain the risk probability output by each model; Calculate the fusion risk probability of the target patient, and record the feature response vector of each model; generate a response list; The characteristic change data of the target patient are extracted from the dynamic risk profile, and multimodal model fusion is performed to generate a historical risk sequence. Using exponential smoothing, the historical risk sequence is used to predict trends and calculate confidence intervals, generating a risk prediction sequence.

7. The VTE intelligent assessment and prevention decision support system according to claim 6, characterized in that, The steps of collaborative assessment also include: Based on the risk prediction sequence, the predicted change slope is calculated, and the trend type of the target patient is determined by using a secondary trend threshold. Based on the feature response vectors of all models, and combined with the fusion weights of each model, the contribution of a single feature is calculated. Construct a single-feature contribution ranking table and retain the top performers. The core characteristic that has the greatest impact on risk; Based on the single feature contribution ranking table, interactive feature pairs are constructed, the interaction contribution of the interactive feature pairs is calculated, and a feature interaction contribution table is generated. Bleeding risk features are selected from the fused feature set, and the bleeding risk probability is obtained using the bleeding risk model. Calculate the net risk probability and determine the net risk level using the secondary net judgment threshold; The results of each stage of the collaborative assessment are integrated into a risk assessment result package, and the assessment results are updated synchronously to the dynamic risk profile.

8. The VTE intelligent assessment and prevention decision support system according to claim 7, characterized in that, The steps involved in intervention and adaptation include: Based on the prevention type, net risk level, and bleeding risk level, a basic intervention plan is generated. Retrieve historical data from within the hospital, calculate the effectiveness and adherence rate of each basic intervention protocol in patients with the same combination, calculate the in-hospital fit weight of each protocol, and form a dynamic protocol library. Based on the net risk level and bleeding risk level, suitable solutions are initially screened from the dynamic solution library; The intervention single feature is extracted from the single feature contribution ranking, the coverage of each adaptation scheme is counted, the coverage coefficient is calculated, and thus the feature matching degree is obtained. Intervention interaction pairs are extracted from the feature interaction contribution table to obtain coverage integrity and thus obtain the interaction matching degree. If there are multiple interaction feature pairs, the interaction matching degree is calculated by the arithmetic mean. By combining the in-hospital adaptation weights, feature matching degrees, and interaction matching degrees of each adaptation scheme, and by weighted summation, the adaptation degree of each adaptation scheme is calculated, and a list of recommended schemes is generated. After the plan is implemented, the dynamic risk profile is automatically updated, the net risk probability is recalculated, the risk reduction rate is calculated, and the effectiveness of the current intervention is determined by the secondary classification threshold, generating an intervention effect report.

Images