Active Analysis Methods for Maternal Health Data Based on the DIKWP Large Model

Abstract

This invention provides a proactive analysis method for maternal and infant health data based on the DIKWP large-scale model. The method acquires and preprocesses multimodal data from inside and outside hospitals, inputs it into the DIKWP large-scale model to construct a network semantic model; constructs event sequences based on model nodes, determines risks, risk change rates, and sets of contributing factors, and generates evidence fragments; performs knowledge retrieval and verifies the consistency between risks, triggering criteria, and the DIKWP network association mapping results; under the constraint of the target profile, determines the target action package through model inference sub-services, generates a task dependency graph, and dispatches execution; calculates the latest intervention execution time and establishes a timeout escalation state machine, automatically escalating and alerting when the latest intervention execution time is exceeded or a risk transition occurs; generates an evidence chain object and writes it back, updating the version configuration of the threshold dictionary, rule base, and risk probability calibration parameters. This invention integrates multimodal data and semantic reasoning, improving the timeliness of risk response and the traceability of decisions.

Description

Technical Field

[0001] This invention relates to the field of medical and health information processing technology, and in particular to a proactive analysis method for maternal and infant health data based on the DIKWP large model. Background Technology

[0002] In the existing standardized management system for high-risk pregnancies (HRPs), traditional data analysis and risk warning models have significant engineering and technical deficiencies. Specific pain points include: First, data silos and temporal discontinuities are prominent issues. Currently, maternal health data is highly fragmented. Static medical data such as in-hospital electronic medical records (EMRs), laboratory indicators, and ultrasound diagnoses are isolated from continuous temporal physiological data collected by outpatient follow-up records, home self-testing data, and wearable devices. This makes it impossible to integrate and connect multi-source data, hindering the construction of a complete and traceable timeline of maternal health events, resulting in a single-dimensional and incomplete risk assessment data. Second, risk warning suffers from algorithmic black-box defects. Existing warning systems can only output single risk scores or simple risk classification results, failing to simultaneously output traceable and auditable risk evidence fragments and rule-hitting details. Medical staff cannot understand the basis for risk judgment, making it difficult to verify the effectiveness of warnings, significantly reducing the acceptance and implementation of warning results by the nursing staff. Third, risk management lacks a closed-loop control mechanism. Existing early warning functions only identify risks and cannot automatically transform abstract early warning signals into standardized, actionable task lists for nurses. They also lack a full-process mechanism for intervention, escalation of delays, and post-event review and tracing, leading to a disconnect between risk warnings and clinical treatment, and increasing the risk of delayed or missed management. Fourth, they suffer from poor scenario adaptability and high implementation difficulty. Medical resources, control objectives, and treatment conditions vary significantly across different treatment scenarios, such as outpatient, inpatient, and home care. Existing technologies lack calculable constraint logic based on "control objectives and scenario resource status," resulting in inconsistent treatment standards and execution paths for the same pregnant woman's risk in different scenarios. This leads to insufficient standardization and stability of control, making it difficult to meet the needs of refined management of high-risk pregnant women across the entire spectrum. Fifth, even when introducing general large-scale models, existing systems typically remain at the level of text summarization or question-and-answer, lacking a DIKWP large-scale model application mechanism that models data, information, knowledge, wisdom, and objectives in a network, integrating risk prediction, evidence verification, and action selection. This makes it difficult to form a purpose-driven, interpretable, and executable closed-loop analysis of maternal health. Summary of the Invention

[0003] The purpose of this invention is to provide a proactive analysis method for maternal health data based on the DIKWP large model, so as to solve or at least partially solve the problems of data fragmentation, uninterpretable risk warnings, disconnect between warnings and nursing care, and difficulty in calculating the constraints of different scenarios in the existing technology.

[0004] To achieve the above-mentioned objectives, this invention provides a method for proactive analysis of maternal health data based on the DIKWP large model, the method comprising: Acquire multimodal data from inside and outside the target hospital, and preprocess the multimodal data; The preprocessed multimodal data is input into the DIKWP large model to construct the DIKWP network semantic model. The DIKWP network semantic model includes D-type data nodes, I-type information nodes, K-type knowledge nodes, W-type wisdom nodes, P-type target nodes, and relation edges. Based on D-type data nodes and I-type information nodes, construct a timeline-based event sequence; Based on the event sequence, determine the risk, risk change rate and set of contributing factors, and generate evidence fragments. Each evidence fragment should include at least an indicator trajectory summary, triggering basis, node type and associated path identifier. Perform knowledge retrieval and verify the consistency between the risk, triggering basis, and DIKWP network association mapping results based on the knowledge retrieval results; Obtain or generate a target profile; under the constraints of the target profile, determine the target action package through the W-type smart node reasoning sub-service; generate a task dependency graph and dispatch the target action package; and track the execution status of the target action package. Calculate the latest execution time of the intervention of the target action package and establish a timeout escalation state machine; When the latest intervention time or preset risk transition conditions are exceeded, the target action package will be automatically upgraded to a higher-level action package and a reminder will be triggered. Generate an evidence chain object and record write-back information. Update the version configuration of the threshold dictionary, rule base, and risk probability calibration parameters based on the write-back information.

[0005] Furthermore, the multimodal data includes, but is not limited to, EMR structured fields, laboratory test data, ultrasound / imaging report text, ultrasound or imaging originals / keyframes, nursing records, follow-up scales, patient symptom check-in data, home monitoring data, and wearable time-series data.

[0006] Furthermore, the preprocessed multimodal data is extracted into event objects, each event object is encoded into an event vector, multiple event vectors are arranged into an event sequence by timestamp, and statistical analysis and trend characteristics are performed by time window.

[0007] Furthermore, the risk calculation is achieved through a first method or a second method, wherein the first method includes the following operations: The basic risk value is generated by the rule engine based on a preset guide / path rule library or preset risk trigger parameters; The risk prediction sub-service of the DIKWP large model is invoked to output the predicted risk. The final output risk is obtained by calibrating and fusing the basic risk value and the predicted risk. The second method includes the following operations: Using an event sequence as input, the event vector within a specified time window is encoded to obtain a window state vector; Output risk and risk change rate based on window state vector.

[0008] Furthermore, the determination of the set of contributing factors specifically includes: selecting a key time window, performing Top-K sampling on the contributing factors within the key time window to obtain the set of contributing factors; in the evidence fragment, the indicator trajectory summary is the trajectory summary of key features within the key time window, the triggering basis includes hit rules and similar case evidence, and the contributing factors are events, indicators, symptoms, text entities, or image signs whose contribution to risk or risk change rate exceeds a preset contribution threshold within the key time window.

[0009] Furthermore, consistency verification is used to handle conflicts between knowledge retrieval results and risk / trigger basis. Conflict types include hard constraint violation, insufficient evidence, and data anomaly. Different handling strategies are adopted for different types of conflicts: hard constraint violation triggers mandatory escalation or shortens the latest execution time of intervention, insufficient evidence triggers supplementary verification tasks, and data anomaly triggers measurement quality review or secondary data collection.

[0010] Furthermore, the target profile includes weight parameters corresponding to security, timeliness, resources, compliance, and preferences, as well as at least one hard constraint. Under the constraints of the target profile, the target action package is determined, specifically including the following operations: Candidate action packages are determined based on the knowledge retrieval results; Calculate the utility vector for each candidate action package, the utility vector including safety benefit, timeliness benefit, resource feasibility, compliance prediction value and preference matching value; The comprehensive score of the candidate action package is calculated based on the utility vector and the target profile weight parameters; Filter candidate action packets based on hard constraints; The candidate action package with the highest overall score is determined and output as the target action package. The reasons for selecting the target action package are written into the evidence chain.

[0011] Furthermore, the latest execution time of the intervention for the target action package is calculated, specifically including the following operations: The current case is represented by a timeline feature vector. The top N similar cases are retrieved to obtain the Top-N similar cases, forming a set of similar cases. The time from the similar time point to the adverse outcome / hospitalization / emergency visit is calculated to form a distribution. Select the quantile q of the distribution, and calculate the latest execution time of the basic intervention based on the quantile q; The latest execution time of basic intervention is revised based on the target profile, and the latest execution time of intervention is determined based on the revised latest execution time of basic intervention and rule boundary constraints.

[0012] Furthermore, the write-back information includes execution write-back, compliance write-back, and outcome write-back, which respectively trigger threshold dictionary updates, rule base iterations, and risk probability calibration parameter updates. During the update, the DIKWP mesh semantic model node and relationship versions, DIKWP large model versions, rule versions, and calibration parameter versions are recorded.

[0013] Compared with the prior art, the beneficial effects of the present invention are: 1) This invention constructs a network semantic model of multimodal data into five types of semantic nodes (D / I / K / W / P) and their relational edges through the DIKWP large model, so that risk identification, evidence interpretation and action selection have a unified and traceable semantic foundation; 2) This invention realizes a closed-loop engineering process of "early warning → task → time limit → upgrade → review", which significantly reduces the probability of implementation failure caused by the inability to execute early warnings; 3) This invention makes the objective constraints explicit, enabling the same risk to be controlled and differentiated in different scenarios / objectives without breaking the hard safety constraints; 4) This invention achieves standardized, programmable, and auditable nursing interventions through action packages; 5) This invention provides an interpretable latest execution time point based on the quantiles and risk growth buffer calculations of the time distribution of similar case outcomes; 6) This invention reduces misjudgments caused by noise / missing data through consistency verification and supplementary evidence task mechanisms, forming a traceable chain of evidence; 7) This invention avoids uncontrollable drift caused by black-box retraining by driving the hierarchical updates of thresholds, rules, and calibration through multi-dimensional write-back. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the overall process of a proactive analysis method for maternal health data based on the DIKWP large model provided in an embodiment of the present invention. Detailed Implementation

[0016] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0017] Reference Figure 1 This invention first provides a method for proactive analysis of maternal health data based on the DIKWP large model, the method comprising: S101. Obtain multimodal data of the target hospital inside and outside, and preprocess the multimodal data.

[0018] The multimodal data includes, but is not limited to, EMR structured fields, laboratory test data, ultrasound / imaging report text, original ultrasound or imaging images / keyframes, nursing records, follow-up scales, patient symptom check-in data, home monitoring data, and wearable time-series data.

[0019] For example, the preprocessing includes, but is not limited to, identity desensitization, field mapping, time alignment, unit conversion, outlier detection, missing value marking, duplicate record merging, data quality scoring, and original evidence location.

[0020] S102. Input the preprocessed multimodal data into the DIKWP large model to construct the DIKWP network semantic model. The DIKWP network semantic model includes D-type data nodes, I-type information nodes, K-type knowledge nodes, W-type wisdom nodes, P-type target nodes, and relation edges.

[0021] In this invention, the DIKWP large model is a multimodal large model service configured for maternal and child health management scenarios, employing a DIKWP network model rather than a nonlinear sequential model. The DIKWP network semantic model can be represented as follows: G DIKWP =(V,E) V includes D-type data nodes, I-type information nodes, K-type knowledge nodes, W-type wisdom nodes, and P-type purpose nodes; E represents relational edges, including at least one of source-related edges, time-related edges, evidence-related edges, rule-constrained edges, purpose-constrained edges, similar case-related edges, and action-dependent edges. Any node can establish directed or weighted relations with one or more other types of nodes, not limited to unidirectional sequential links from D to I, I to K, K to W, or W to P. The DIKWP large model is not only used as a natural language generation tool, but also participates in multimodal data standardization, network semantic modeling, risk prediction, evidence generation, knowledge retrieval, consistency verification, action package selection, and closed-loop learning.

[0022] Specifically, Class D data nodes are used to carry raw or standardized factual data, such as blood pressure values, urine protein levels, gestational age, test values, symptom check-ins, and equipment sampling values; Class I information nodes are used to carry event semantics with time, unit, source, quality labels, and contextual constraints, such as "systolic blood pressure has been continuously elevated in the past 72 hours" and "report text indicates low amniotic fluid"; Class K knowledge nodes are used to carry guidelines, path rules, hospital SOPs, terminology ontology, similar cases, and their outcome distribution; Class W wisdom nodes are used to carry risk interpretation, intervention strategies, calculation of the latest intervention execution time, and escalation strategies. The latest intervention execution time is the latest time that the intervention task is allowed to be completed after the intervention task is generated for the target risk event, under the joint constraints of the current risk level, risk change trend, knowledge rules, and purpose profile; Class P purpose nodes are used to carry safety, timeliness, resources, compliance, and preference weights and hard constraints in outpatient, inpatient, and home scenarios. Various nodes form a traceable DIKWP mesh semantic model through relational edges. For example, the blood pressure value node is connected to the blood pressure rising trend node through time-related edges, the trend node is connected to the risk interpretation node through evidence-related edges, the risk interpretation node is connected to the guideline rule node through rule constraint edges, and is connected to the purpose profile node prioritizing maternal and fetal safety through purpose constraint edges.

[0023] S103. Construct a timeline-based event sequence based on D-type data nodes and I-type information nodes.

[0024] In this step, the preprocessed multimodal data is extracted into event objects, which include the following fields: event_id (event identifier), patient_id (patient identifier), event_time (event occurrence time), event_type (event type, including symptoms / signs / tests / medication / treatment / follow-up / medical treatment), source (data source), value / enum (event result value / enumeration value), unit (unit), severity (severity), confidence (confidence level), quality_tag (data quality tag, including missing / noise / anomalies / measurement quality score), and text_evidence (original text fragment location).

[0025] For each event object, the DIKWP large model generates a corresponding event vector x_e. The event vector x_e includes at least event type embedding, time location embedding, source embedding, normalized numerical value, unit embedding, severity, confidence, quality mask, text semantic vector, image semantic vector, and DIKWP node type label and relation edge label. For example, for report text, nursing records, and symptom check-in text, a text encoder is used to extract medical entities, negation words, degree words, time words, and symptom combinations; for ultrasound or original image / keyframe, an image encoder is used to extract visual vectors, which are then fused with the report text vector through cross-modal attention or gating to obtain the image semantic vector; when only the report text is obtained and the original image is not, the report text vector is used as the image semantic vector.

[0026] Multiple event vectors are arranged into an event sequence by timestamp, and statistical analysis and trend features are calculated within a time window Δt∈{24h,72h,7d,30d}. Trend features include, but are not limited to, maximum and minimum values, slope, abrupt change points, duration, duration of missing data, and changes in measurement quality. The purpose of encoding is to unify multimodal data, such as structured indicators, text, images, home monitoring, and wearable time-series data, into a single timeline representation for subsequent use in risk prediction, contribution attribution, similar case retrieval, knowledge retrieval, and action package selection.

[0027] S104. Based on the event sequence, determine the risk, risk change rate and set of contributing factors, and generate evidence fragment Ev(t). The evidence fragment shall include at least an indicator trajectory summary, trigger basis, node type and associated path identifier.

[0028] In this invention, risk calculation is achieved through a first method or a second method. The first method includes the following operations: S201. By calling the preset guidelines or clinical pathway rule base and the preset risk trigger parameters through the rule engine, the health data of the target object is matched with the corresponding rule conditions to obtain the rule hit result. Based on the rule hit result and the risk lower limit of the corresponding rule conditions, a normalized basic risk value between 0 and 1 is generated. The preset risk trigger parameters are a set of parameters that are extracted and structured from the hospital's standard operating procedures and can be directly read and judged by the rule engine.

[0029] In this invention, the preset guideline / path rule base stores national or industry-specific guidelines for maternal and child health management, rules for risk assessment and classification management of high-risk pregnant women, rules for the management of specific diseases such as gestational hypertension, gestational diabetes, abnormal fetal movement, postpartum hemorrhage, and infections, as well as hospital internal treatment pathways and nursing SOPs (Standard Operating Procedures). The rule engine converts threshold hits, red flag symptoms, critical values, deteriorating trends, and missing key examinations into normalized base risk values ​​between 0 and 1; for hard rule hit events, the base risk value is not lower than the minimum risk limit of the corresponding risk level.

[0030] In this invention, the preset risk trigger parameters are a set of parameters extracted and structured from hospital standard operating procedures, which can be directly read and judged by the rule engine. These parameters can include numerical limits, trends, time, and logical conditions. Specifically, numerical limits can be test indicator limits used to determine whether the current value matches the rule; trends can be the magnitude of change relative to the base value or a continuous slope used to determine whether a short-term or continuous trend is worsening; time can be retest / duration parameters, such as retesting every few minutes or lasting >15 minutes, used to determine whether the rule is continuously met or whether it is escalated; logical conditions can be red flag symptoms, critical values, missing key examinations, etc., used to determine whether event-based or combination-based rules are matched.

[0031] As a preferred implementation, the rule engine does not directly treat natural language normative text as numerical risk, but instead transforms the conditions within it that can be judged by a computer into machine-executable rule objects. Each rule object includes at least the fields shown in Table 1: Table 1. Rule Object Fields

[0032] After matching the target object's current health data with the rule objects item by item, a rule hit vector H (i.e., the rule hit result) is formed: H = {H_threshold, H_red, H_critical, H_trend, H_missing} Here, H_threshold represents the hit status of numerical threshold rules; H_red represents the hit status of red flag symptom rules; H_critical represents the hit status of critical values ​​or high-risk hard rules; H_trend represents the hit status of trend deterioration rules; and H_missing represents the hit status of critical check timeout missing rules. For a single rule, a hit is recorded as 1, and a miss is recorded as 0.

[0033] Taking severe hypertension during pregnancy as an example. According to the "Guidelines for the Diagnosis and Treatment of Hypertensive Disorders in Pregnancy (2020)," the diagnostic threshold for hypertension in pregnancy is a systolic blood pressure ≥140 mmHg and / or a diastolic blood pressure ≥90 mmHg. For pregnant women with severe hypertension, i.e., a systolic blood pressure ≥160 mmHg and / or a diastolic blood pressure ≥110 mmHg, repeated measurements can be taken at intervals of several minutes. An acute attack lasting more than 15 minutes can be labeled as persistent severe hypertension or a hypertensive emergency. This example only converts the above clinical triggering conditions into executable logic for the rule engine, and does not interpret them as the probability of disease occurrence.

[0034] In this example, the field settings for the rule object are shown in Table 2: Table 2. Rule object fields using severe hypertension during pregnancy as an example.

[0035] When the initial measurement reaches the threshold for severe hypertension, and a repeat measurement within minutes still meets the criteria of "SBP ≥ 160 mmHg and / or DBP ≥ 110 mmHg", the rule engine sets the hit status H_severeBP of the hard rule to 1; otherwise, it sets it to 0. H_severeBP = 1 (if it still hits after retesting); otherwise H_severeBP = 0.

[0036] This invention employs a deterministic transformation method of "hit identifier × risk lower limit, and taking the maximum value of the hit rules of the same type". For risk type k, its basic risk value can be expressed as: r_rule,k = max{ H_j × s_j | j ∈ J_k} In the formula, J_k represents the rule set corresponding to risk type k; H_j∈{0,1} indicates whether the j-th rule is matched; s_j∈[0,1] represents the pre-defined and version-managed risk_floor in the rule object. If no rule in J_k is matched, then r_rule,k=0. For the hard rule of severe hypertension in this example, s_severeBP=1.0, therefore: r_rule,severeBP = H_severeBP × 1.0 When the rule is not hit, the basic risk value r_rule,severeBP = 0; when the rule is hit, r_rule,severeBP = 1.0. The 1.0 only indicates that the normalized rule code is "fully activated in the rule engine", and does not mean that the probability of the target object developing a certain disease is 100%, nor does it replace clinical diagnosis.

[0037] Suppose the system obtains continuous blood pressure data: initial SBP = 166 mmHg, DBP = 112 mmHg; several minutes later, a repeat measurement shows SBP = 164 mmHg, DBP = 111 mmHg. Both measurements reach the threshold for severe hypertension, therefore H_severeBP = 1, and the rule engine outputs r_rule,severeBP = 1.0. If this condition persists for more than 15 minutes, the condition "persistent severe hypertension / hypertensive emergency" is simultaneously marked as a hit.

[0038] The system synchronously generates structured evidence objects, which at least record the rule ID, actual SBP / DBP and retest values, trigger limits, rule source, rule version, timestamp, persistence status, and original data location. This allows for a complete retrospective analysis of "which specification parameter → which rule → which set of data → which rule risk value".

[0039] S202, Call the DIKWP large model risk prediction sub-service to output predicted risk.

[0040] In this invention, the risk prediction sub-service takes event sequences, nodes of category P, and scene context as input, and obtains the window state vector h through a time encoder. w (t), and obtains the case state vector z(t) by fusing structured data vectors, wearable time-series vectors, text semantic vectors, and image semantic vectors through a cross-modal fusion layer. For each risk type k, the risk output header is configured according to r model ,k(t)=σ(w k ·z(t)+b k Output the predicted probability, where w k and b k These are the model parameters corresponding to risk type k.

[0041] S203. The risk obtained by calibrating and fusing the basic risk value and the predicted risk is expressed as follows: R k (t)=σ(α k ·logit(r model,k (t))+β k ·logit(r rule,k (t))+γ k T ·context(t)+b k ) In the formula, R k (t) represents the final risk probability of the k-th risk at time t; σ represents the Sigmoid function, used to map the fusion result to the interval between 0 and 1; α k β k γ represents the model risk weight and rule risk weight corresponding to the k-th risk, respectively;k The coefficient vector represents the scene context features; logit(x) = ln(x / (1-x)); r model,k (t) represents the k-th predicted risk output by the DIKWP large model risk prediction sub-service; r rule,k (t) represents the k-th basic risk output by the rule engine; context(t) represents the scenario context vector, including gestational age, age, medical history, comorbidities, medical scenario, data quality, target profile weight, and resource status; b k This represents the bias term. α k β k γ k and b k It can be pre-configured according to risk type, scenario and version or calibrated based on historical write-back data.

[0042] The resource status refers to the real-time availability of medical or nursing resources required to execute the action package, including but not limited to the number of available beds, the open status of emergency or green channels, outpatient appointment slots, laboratory test appointment times, doctor / nurse schedules, location and workload of home visit personnel, monitoring equipment inventory, patient-side communication accessibility, transportation distance, and hospital resource coverage. For example, for the same medium- to high-risk pregnant women, if patient-side communication is unavailable and home visit personnel are unavailable in a home setting, the resource status will lower the score of the online follow-up action package and increase the priority of the outpatient or emergency assessment action package.

[0043] The second method includes the following operations: S301. Using the event sequence as input, encode the event vector within the specified time window to obtain the window state vector.

[0044] For example, the time window in this step can be the last 7 days or the last 30 days. In the second method, the DIKWP large model adds relative event position encoding and a missing mask to the event vectors within the specified time window, and uses Transformer, GRU, TCN or a combination thereof to form the window state vector h. w (t), and fuse text vectors, image vectors and structured vectors into a unified vector z(t) through cross-modal attention or gating.

[0045] S302, Output risk and risk change rate.

[0046] In this step, for each risk type k, the risk output head outputs R based on the output z(t). k (t), and according to dR k / dt=(R k (t)-R k The risk change rate is calculated as (t-Δt) / Δt, where Rt k(t) represents the final risk probability of the k-th type of risk at time t, R k (t-Δt) represents the final risk probability of the k-th type of risk at time t-Δt, where Δt is the time difference, and dR k / dt is the rate of change of risk. When dR k / dt is greater than the preset growth threshold, or R k (t) When crossing the boundary between low-risk, medium-risk, high-risk, and extremely high-risk levels, it is marked as a risk transition.

[0047] In this invention, the contributing factors are events, indicators, symptoms, text entities, or image signs whose contribution to risk or risk change rate exceeds a preset contribution threshold within a key time window. The key feature is that they are either Top-K contributing factors ranked high in the contributing factor set, or strongly constrained factors that are hit by hard rules even if their contribution ranking is not high.

[0048] The determination of the set of contributing factors specifically includes: selecting key time windows (e.g., the last 24 hours / 72 hours / 7 days), and evaluating the contributing factors E within the key time windows. k (t) Perform Top-K sampling to obtain a set of contribution factors, represented as {feature_name, feature_type, trend, max / min / slope, trigger point, contribution, confidence, dikwp_tag}. Sampling can be achieved using one or more of attention weighting, gradient attribution, SHAP, rule hit strength, or difference contribution from similar cases. In the evidence fragment, the indicator trajectory summary is a summary of the trajectory of key features within a key time window, such as a 7-day upward trend in blood pressure, changes in urine protein grading, duration of headache symptoms, or imaging reports indicating placental abnormalities. The triggering basis includes hit rules and similar case evidence. Hit rules include the hit guideline / SOP entry number or rule ID and triggering conditions. Similar case evidence includes similar case search results and a summary of their outcome time distribution.

[0049] S105. Conduct knowledge retrieval through the knowledge retrieval sub-service of the DIKWP large model, and verify the consistency between the risk, triggering basis and the DIKWP network association mapping results based on the knowledge retrieval results.

[0050] The knowledge retrieval subservice is based on risk type, evidence fragment Ev(t), and contribution factor set E. k (t) Generate structured search terms or search vectors from the nodes of category P and the context of the search scenario, and retrieve guide entries, disease-specific pathways, hospital SOPs, contraindications, supplementary evidence templates, and similar cases from the knowledge nodes of category K. The search results return the knowledge item ID, applicable conditions, recommended actions, contraindications, evidence level, version number, and applicable scenarios.

[0051] The DIKWP network association mapping result consists of risk nodes, evidence nodes, knowledge nodes, target nodes and their relationship paths; the consistency verification uses the applicable conditions, taboo conditions and recommended actions in the search results as verification constraints, and compares the risk type, triggering basis, association path and candidate action item by item.

[0052] In this invention, consistency verification is used to handle conflicts between knowledge retrieval results and risk / trigger criteria. Conflict types include: C1: Hard constraint violation (critical value / red flag symptoms appear, action package must be upgraded, but model has not been upgraded); C2: Insufficient evidence (lack of key test / retest data leads to uncertainty); C3: Abnormal data (low measurement quality or inconsistent with history).

[0053] For different types of conflicts, appropriate handling strategies should be adopted: For C1, force an upgrade of the action package level or shorten the latest execution time of the intervention, and generate evidence of rule hit; For C2, a supplementary evidence task is generated, and the shortest intervention and latest execution time point are given. The supplementary evidence task is a retest, supplementary examination, or supplementary consultation. For C3, trigger measurement quality verification and secondary data acquisition, and perform manual verification if necessary.

[0054] The results of the consistency verification are ultimately written into the evidence chain to support review and retrospective analysis.

[0055] S106. Obtain or generate a target profile. Under the constraints of the target profile, determine the target action package through the W-type intelligent node inference sub-service of the DIKWP large model, generate a task dependency graph and dispatch the target action package, and track the execution status of the target action package.

[0056] In this invention, the purpose profile corresponds to the P-class purpose node of the DIKWP mesh semantic model, including weight parameters corresponding to safety, timeliness, resources, compliance, and preference, as well as at least one hard constraint. For example, the data structure of the purpose profile includes the following fields: purpose_id (task identifier), scenario (outpatient / home / hospitalization), weights (including w_safety (safety weight), w_timeliness (timeliness weight), w_resource (resource consumption weight), w_adherence (compliance weight), w_preference (preference weight)), constraints (maximum allowed intervention latest execution time, allowed escalation limit, resource availability list, red flag symptom mandatory escalation rule), and update_rule (dynamic update trigger condition).

[0057] Under the constraints of the target profile, the target action package is determined, which specifically includes the following operations: S401. Determine candidate action packages based on knowledge retrieval results.

[0058] In this invention, the data structure of the action package includes the following fields: package_id (action package identifier), risk_level (risk level), risk_type (risk type), purpose_id (task identifier), actions (action set), frequency (frequency / duration), channel (execution channel, including online / outpatient / inpatient / home visits), role_requirement (responsible person qualification), record_fields (record field dictionary), recheck_node (re-evaluation node / condition), education_content (educational points), and escalation_rule_id (escalation rule identifier).

[0059] The actions are a structured list, and each action includes action_code, action_desc, deadline (derived from the latest execution time of the intervention), priority, owner, dependency, trigger_condition, and evidence_required.

[0060] S402. Calculate the utility vector for each candidate action package. Let the set of candidate action packages be A = {A1, ..., A2}. i A m Candidate action package A i The utility vector U(A) i The formula for calculating ) is as follows: U(A i )=[u safety ,u timeliness ,u resource ,u adherence ,u preference ] In the formula, u safety Indicates the expected security benefits or risk reduction of the candidate action package; u timeliness This indicates the time-efficiency benefit of the candidate action package being completed within the latest possible execution time of the intervention; u resource This indicates the inverse score of the feasibility or resource cost of the candidate action package under the current resource state; u adherenceThis represents the predicted adherence value for the target group upon completion of retesting, uploading, visiting a clinic, or reading educational materials; u preference This represents the degree of match between the candidate action package and the target object's preferences, distance, communication style, and management scenario. All utility values ​​are normalized to the range of 0 to 1.

[0061] S403. Calculate the comprehensive score of the candidate action package based on the utility vector and the target profile weight parameters. The formula for calculating the comprehensive score is as follows: Score(A i )=Σ j∈J w j ·u j (A i )-λ pen ·penalty(A i ), J={safety,timeliness,resource,adherence,preference} In the formula, Score(A) i ) represents candidate action package A i Overall score; w j Let Σw represent the weight of the j-th item in the target profile, and Σw j =1; u j (A i ) represents candidate action package A i Normalized score on the j-th utility; penalty(A i ) indicates penalties resulting from data uncertainty, resource conflicts, or patient unavailability; λ pen λ represents the weight of the penalty term. When the penalty term is not enabled, λ pen Let J be 0, where J represents the set of multi-objective decision optimization objectives, safety represents the safety objective, timeliness represents the timeliness objective, resource represents the resource rationality objective, adherence represents the compliance objective, and preference represents the personalized preference objective.

[0062] S404. Filter candidate action packets based on hard constraints.

[0063] In this invention, candidate action packages must meet the hard constraints of the target profile, such as the need to upgrade the action package when red flag symptoms appear, the maximum allowed intervention and the latest execution time, and resource availability.

[0064] S405. Determine the candidate action package with the highest comprehensive score, output it as the target action package, and write the reasons for selecting the target action package into the evidence chain. The reasons include hit rules, evidence fragments, resource states, etc.

[0065] In this invention, a task dependency graph G=(V,E) is generated based on the target action package, where V represents the task nodes in the action package and E represents the dependencies between tasks. Task nodes include task identifier, action code, responsible role, execution deadline, priority, required evidence, and write-back field. Dependencies are jointly determined by the dependency (preceding dependent action) field, the trigger_condition field, review nodes, temporal sequence, and evidence collection requirements. For example, "doctor review" is triggered only after "retesting blood pressure" is completed and written back; "education outreach" and "retest reminder" can be dispatched in parallel; "appearance of red flag symptoms" can bypass the preceding evidence-gathering node and directly trigger emergency assessment.

[0066] After receiving the task dependency graph, the dispatch system first sorts the task nodes topologically to identify dispatchable nodes with no prerequisites or whose prerequisites have been completed. Then, based on the owner (executor), role_requirement (executor role requirements), resource status, latest execution time of intervention, and priority (task priority), the task is assigned to the nurse, doctor, patient, follow-up staff, or emergency / outpatient work queue. After dispatching, the system continuously monitors the execution writeback. When the prerequisite task is completed, the supplementary certificate result is returned, or a risk transition occurs, the system automatically releases subsequent nodes or triggers upgrade nodes.

[0067] S107. Calculate the latest execution time of the intervention of the target action package and establish a timeout upgrade state machine.

[0068] This step specifically includes the following operations: S501. Represent the current case using the timeline feature vector v(t). Retrieve the top N similar cases to obtain the Top-N similar cases, forming a set of similar cases. Calculate the time To from the similar case set to the adverse outcome / admission / emergency visit, forming a distribution D(To). The timeline feature vector v(t) is composed of the case state vector, key contributing factors, gestational age, scene context, and target profile output by the DIKWP large model.

[0069] S502. Select a quantile q (e.g., 0.2~0.3), and calculate the latest execution time of the basic intervention based on the quantile q. The calculation formula is as follows: TTA base =now+Quantile(D(To),q)-buffer In the formula, TTA base q represents the latest time point at which basic intervention was implemented; now represents the current time; Quantile(D(To),q) represents the qth quantile of the time distribution D(To) of similar cases from similar time points to adverse outcomes / hospitalization / emergency room visits, where a smaller q indicates greater conservatism.

[0070] buffer=b0+b1·max(dR / dt,0)+b2·duration key In the formula, b0 represents the fixed safety buffer time; b1 represents the risk growth rate buffer coefficient; dR / dt represents the risk growth rate; b2 represents the duration buffer coefficient of the key contribution factor; and duration key This indicates the duration of the key contribution factor or hard rule hit status. b0, b1, and b2 can be pre-configured according to risk type, scenario, and version, or calibrated based on the write-back outcome.

[0071] S503. Based on the target profile, the latest execution time of the basic intervention is revised, and the latest execution time of the intervention is determined based on the revised latest execution time of the basic intervention and the rule boundary constraints.

[0072] In this step, the correction coefficient ρ is calculated. p =clip(1-η s ·w safety -η t ·w timeliness +η r ·w resource ,ρ min ,ρ max ), and obtain TTA p =now+ρ p ·(TTA base -now), where η s η t η r ρ is the objective weighting influence coefficient. min and ρ max To adjust the upper and lower limits of the coefficient. The latest execution time of the final intervention = min(TTA) p TTA rule,max TTA redflag ), of which TTA rule,max This indicates the maximum time limit allowed by the guidelines / SOPs, TTA. redflag w represents the latest time limit corresponding to the red flag hard rule. safety For security weights, w timeliness As a timeliness weight, w resource This represents the weight of resource consumption.

[0073] In this step, the rule boundary constraints are the allowable intervention time ranges pre-set for different risk levels, intervention types, and purpose profiles, based on hospital SOPs, clinical diagnosis and treatment guidelines, risk management requirements, or expert-preset rules.

[0074] S108. When the latest execution time of intervention or the preset risk transition condition is exceeded, the target action package will be automatically upgraded to a higher-level action package and a reminder will be triggered.

[0075] In this step, if now > the latest time point for intervention, or ΔR = R k (t)-R k (t-Δt)>δ k If the risk level exceeds the preset risk threshold, it will automatically escalate to a higher-level action package and trigger a multi-channel alert and order dispatch. k (t) represents the final risk probability of the k-th type of risk at time t, R k (t-Δt) represents the final risk probability of the k-th type of risk at time t-Δt, δ k This represents the risk transition threshold corresponding to the k-th risk category, which can be configured according to risk type, scenario, or historical risk volatility quantiles. The upgrade reason, trigger threshold, evidence fragments, and dispatch records are written into the evidence chain.

[0076] S109. Generate an evidence chain object and record write-back information. Update the version configuration of the threshold dictionary, rule base, and risk probability calibration parameters based on the write-back information.

[0077] In this invention, the data structure of the evidence chain object includes the following fields: alert_id (alert identifier), patient_id (patient identifier), time_generated (evidence chain generation time), Rk(t) (final risk probability of the k-th risk at time t), dRk / dt (risk change rate), Ek(t) (set of contributing factors), Ev(t) (evidence fragment), DIKWP_object_ids (DIKWP associated object identifier set), matched_rules (matching rule set), similar_case_ids (similar case identifier set), selected_package_id (selected action package identifier), TTA (latest intervention execution time), task_ids (task identifier set), execution_log (execution log), adherence_log (compliance log), outcome (treatment outcome), model_version (model version number), rule_version (rule version number), calibration_version (calibration version number), and review_note (review notes). The evidence chain object supports playback by alert identifier and outputs audit reports or quality control indicators.

[0078] The write-back information includes execution write-back, compliance write-back, and outcome write-back, which respectively trigger threshold dictionary updates, rule base iterations, and risk probability calibration parameter updates. During the update, the DIKWP mesh semantic model node and relationship versions, DIKWP large model versions, rule versions, and calibration parameter versions are recorded.

[0079] The write-back process includes whether tasks were completed on time, the quality of completion, and the completeness of recorded fields. Compliance write-back includes the completion rate of outreach / reading / uploading / retesting and the timeout reason tag. Outcome write-back includes hospital admission, emergency room visits, complications, NICU admission, and gestational age outcomes.

[0080] The aforementioned write-backs trigger layered versioned releases of threshold dictionary updates, rule base iterations, and risk probability calibration parameter updates, respectively: Threshold dictionary update is performed by grouping and updating the trigger threshold and the quantile q of the latest execution time of intervention according to scenario / purpose profile; The rule base iteration transforms the review conclusions into optimizations for rule ID addition, deletion, modification, and supplementary verification task templates; Model calibration calibrates the risk probability output (e.g., Platt / Isotonic) and releases it in a versioned manner to avoid black-box drift. DIKWP node and relation version update records the source, time, and scope of changes to D / I / K / W / P semantic nodes and relation edges.

[0081] This invention also provides a proactive analysis system for maternal health data based on the DIKWP large model, the system comprising: The data access and governance module is used to acquire multimodal data from inside and outside the target hospital and to preprocess the multimodal data. The DIKWP large model service module is used to build DIKWP mesh semantic models based on preprocessed multimodal data. The DIKWP mesh semantic model includes D-type data nodes, I-type information nodes, K-type knowledge nodes, W-type wisdom nodes, P-type target nodes, and relation edges. The event sequence construction module is used to construct timeline-based event sequences based on D-type data nodes and I-type information nodes. The Risk and Evidence module is used to determine the risk, risk change rate, and set of contributing factors based on the event sequence, and generate evidence fragments. The evidence fragments include at least an indicator trajectory summary, triggering basis, node type, and associated path identifier. The knowledge retrieval and consistency verification module is used to perform knowledge retrieval and, based on the knowledge retrieval results, verify the consistency between the risk, triggering basis, and the DIKWP network association mapping results. The target profile management module is used to acquire or generate target profiles. Under the constraints of the target profiles, the target action packages are determined through the W-type smart node reasoning sub-service, a task dependency graph is generated and the target action packages are dispatched, and the execution status of the target action packages is tracked. The intervention latest execution time calculation and upgrade module is used to automatically upgrade the target action package to a higher-level action package and trigger a reminder when the intervention latest execution time or preset risk transition conditions are exceeded. The evidence chain and closed-loop learning module is used to generate evidence chain objects and record write-back information, and update the version configuration of threshold dictionary, rule base and risk probability calibration parameters based on the write-back information.

[0082] The system is used to execute the aforementioned method, and its working principle and technical effects can be referred to the aforementioned method, so they will not be repeated here.

[0083] The present invention will be described below through specific embodiments: Example 1: Analysis of full-cycle health data of high-risk pregnant women In this embodiment, throughout the entire cycle of maternal and infant outpatient visits, home care, and hospitalization, multimodal data, including EMR structured fields, laboratory test and ultrasound report texts, follow-up scales / symptom self-assessments, home monitoring, and wearable time-series data, are input into the system. The DIKWP big data model first maps the above data into factual data in D-class data nodes and event semantics in I-class information nodes, then retrieves guidelines, rules, and similar cases in K-class knowledge nodes, and finally, W-class intelligent nodes output executable action packages under the constraints of P-class goal nodes.

[0084] The system constructs an event sequence based on the input multimodal data and outputs risk, risk change rate, set of contributing factors, and evidence fragments. If a red flag symptom is detected or consistency checks reveal insufficient evidence, a supplementary evidence task is triggered. An action package is selected based on the target profile, and the latest intervention execution time is calculated. After selecting the action package, a task dependency graph is generated and dispatched, a reminder is sent to the dispatched nurse, and a record is maintained. After the nurse completes the intervention based on the action package, the system writes back the execution / compliance / outcome information and updates the threshold dictionary, rule base, and calibration parameters.

[0085] Example 2: Closed-loop analysis of gestational hypertension / preeclampsia In this embodiment, for pregnant women with gestational hypertension or preeclampsia, the system inputs their blood pressure data collected at home, as well as multimodal data such as symptom logs (e.g., headache / blurred vision / upper abdominal pain / edema / fetal movement changes), weight / edema, gestational age, urine protein, and liver and kidney function tests.

[0086] The DIKWP big data model records blood pressure, urine protein, liver and kidney function, and symptom facts at the D-class data nodes; forms event semantics such as "blood pressure rises in the past 72 hours" and "persistent headache symptoms" at the I-class information nodes; retrieves disease-specific rules and similar cases at the K-class knowledge nodes; outputs risk interpretations, the latest intervention time point, and escalation strategies at the W-class wisdom nodes; and retrieves purpose profiles such as prioritizing maternal and fetal safety or reducing unnecessary hospitalizations at the P-class purpose nodes. The system forms evidence fragments based on blood pressure change trends, symptom combinations, and changes in laboratory data. When the purpose profile prioritizes maternal and fetal safety, the latest intervention time point is shortened and the action package is upgraded; when the purpose profile aims to reduce unnecessary hospitalizations, supplementary evidence and expedited outpatient assessments are prioritized, but without violating the red flag hard rules. If no intervention is performed within the time limit, the system automatically upgrades to a higher-level action package, such as emergency or green channel admission assessment.

[0087] Example 3: Generation of Risk Evidence through Text-Image Fusion In this embodiment, the system acquires the ultrasound report text and corresponding image keyframes. The text encoder extracts medical entities and degree terms such as "low amniotic fluid index" and "abnormal placental position," while the image encoder extracts visual features from the keyframes. The DIKWP large model fuses the text semantic vector and image semantic vector into a unified image event vector through cross-modal attention. If the text report indicates an abnormality but the image quality score is low, a consistency check generates an image review task. If the text and image evidence are consistent and the risk exceeds a threshold, the W-type intelligent node inference subservice writes the image feature as a contributing factor into the evidence fragment and selects a follow-up examination, expedited outpatient visit, or inpatient assessment action package based on the target profile.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A proactive analysis method for maternal health data based on the DIKWP large model, characterized in that, The method includes: Acquire multimodal data from inside and outside the target hospital, and preprocess the multimodal data; The preprocessed multimodal data is input into the DIKWP large model to construct the DIKWP network semantic model. The DIKWP network semantic model includes D-type data nodes, I-type information nodes, K-type knowledge nodes, W-type wisdom nodes, P-type target nodes, and relation edges. Based on D-type data nodes and I-type information nodes, construct a timeline-based event sequence; Based on the event sequence, determine the risk, risk change rate and set of contributing factors, and generate evidence fragments. Each evidence fragment should include at least an indicator trajectory summary, triggering basis, node type and associated path identifier. Perform knowledge retrieval and verify the consistency between the risk, triggering basis, and DIKWP network association mapping results based on the knowledge retrieval results; Obtain or generate a target profile; under the constraints of the target profile, determine the target action package through the W-type smart node reasoning sub-service; generate a task dependency graph and dispatch the target action package; and track the execution status of the target action package. Calculate the latest execution time of the intervention of the target action package and establish a timeout escalation state machine; When the latest intervention time or preset risk transition conditions are exceeded, the target action package will be automatically upgraded to a higher-level action package and a reminder will be triggered. Generate an evidence chain object and record write-back information. Update the version configuration of the threshold dictionary, rule base, and risk probability calibration parameters based on the write-back information.

2. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 1, characterized in that, The multimodal data includes EMR structured fields, laboratory test data, ultrasound or imaging report text, original ultrasound or imaging images or keyframes, nursing records, follow-up scales, patient symptom check-in data, home monitoring data, and wearable time-series data.

3. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 1, characterized in that, The preprocessed multimodal data is extracted into event objects, each event object is encoded into an event vector, multiple event vectors are arranged into an event sequence by timestamp, and statistical analysis and trend characteristics are performed by time window.

4. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 3, characterized in that, The risk is calculated through either a first method or a second method, wherein the first method includes the following operations: By calling preset guidelines or clinical pathway rule bases and preset risk trigger parameters through the rule engine, the health data of the target object is matched with the corresponding rule conditions to obtain the rule hit result. Based on the rule hit result and the risk lower limit of the corresponding rule conditions, a normalized basic risk value between 0 and 1 is generated. The preset risk trigger parameters are a set of parameters that are extracted and structured from the hospital's standard operating procedures and can be directly read and judged by the rule engine. The risk prediction sub-service of the DIKWP large model is invoked to output the predicted risk. The final output risk is obtained by calibrating and fusing the normalized base risk value and the predicted risk. The second method includes the following operations: Using an event sequence as input, the event vector within a specified time window is encoded to obtain a window state vector; Output risk and risk change rate based on window state vector.

5. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 1, characterized in that, The determination of the set of contributing factors specifically includes: selecting a key time window, performing Top-K sampling on the contributing factors within the key time window to obtain the set of contributing factors; in the evidence fragment, the indicator trajectory summary is the trajectory summary of the key features within the key time window, the triggering basis includes hit rules and similar case evidence, and the contributing factors are events, indicators, symptoms, text entities or image signs whose contribution to risk or risk change rate exceeds a preset contribution threshold within the key time window.

6. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 1, characterized in that, Consistency verification is used to handle conflicts between knowledge retrieval results and risks or triggering criteria. Conflict types include hard constraint violations, insufficient evidence, and data anomalies. Different handling strategies are adopted for different types of conflicts: hard constraint violations trigger mandatory escalation or shorten the latest execution time of intervention; insufficient evidence triggers supplementary verification tasks; and data anomalies trigger measurement quality review or secondary data collection.

7. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 1, characterized in that, The target profile includes weighted parameters corresponding to security, timeliness, resources, compliance, and preferences, as well as at least one hard constraint. Under the constraints of the target profile, the target action package is determined, specifically including the following operations: Candidate action packages are determined based on the knowledge retrieval results; Calculate the utility vector for each candidate action package, the utility vector including safety benefit, timeliness benefit, resource feasibility, compliance prediction value and preference matching value; The comprehensive score of the candidate action package is calculated based on the utility vector and the target profile weight parameters; Filter candidate action packets based on hard constraints; The candidate action package with the highest overall score is determined and output as the target action package. The reasons for selecting the target action package are written into the evidence chain.

8. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 1, characterized in that, Calculate the latest execution time of the intervention for the target action package, specifically including the following operations: The current case is represented by a timeline feature vector. The top N similar cases are retrieved to obtain the Top-N similar cases, forming a set of similar cases. The time from the similar time point to the adverse outcome, hospitalization or emergency room visit is calculated to form a distribution. Select the quantile q of the distribution, and calculate the latest execution time of the basic intervention based on the quantile q; The latest execution time of basic intervention is revised based on the target profile, and the latest execution time of intervention is determined based on the revised latest execution time of basic intervention and rule boundary constraints.

9. The method for proactive analysis of maternal health data based on the DIKWP large model according to claim 1, characterized in that, The write-back information includes execution write-back, compliance write-back, and outcome write-back, which respectively trigger threshold dictionary updates, rule base iterations, and risk probability calibration parameter updates. During the update, the DIKWP mesh semantic model node and relationship versions, DIKWP large model versions, rule versions, and calibration parameter versions are recorded.

Images