Neural critical care ai-assisted decision system based on knowledge graph and multi-modal temporal reasoning

Abstract

The application discloses a knowledge graph and multi-modal time sequence reasoning neural critical illness AI auxiliary decision system, and realizes the following processes through various functional units: constructing a self-adaptive weight-adjustable neural critical illness dynamic knowledge graph; collecting and processing multi-modal time sequence data and generating a feature matrix; identifying a diagnosis and treatment stage through double-mode reasoning, matching a knowledge graph to lock a standard path and a contraindication rule; adopting a graph neural network to fuse an attention mechanism and generating a whole-course diagnosis and treatment recommended path; carrying out medical order comparison, contraindication early warning and omission prompting; evaluating prognosis and layered intervention based on time sequence trends; completing diagnosis and treatment difference grading, quantitative scoring and tracing according to a clinical gold standard; building a full-closed-loop architecture, and realizing model and knowledge graph iterative optimization relying on man-machine feedback. The application realizes individualized adaptation of medical knowledge, improves the time sequence and systematization of diagnosis and treatment, has the ability of prospective prevention and control, compliance quality control and iterative optimization, and is suitable for neural critical illness whole-course intelligent auxiliary decision.

Description

Technical Field

[0001] This invention relates to the fields of smart medical technology and clinical decision support technology, and in particular to an AI-assisted decision-making system for neurocritical care based on knowledge graphs and multimodal temporal reasoning. Background Technology

[0002] Neurocritical care includes critical conditions such as severe traumatic brain injury, massive cerebral hemorrhage, and aneurysmal subarachnoid hemorrhage. These patients have complex and variable conditions, drastic fluctuations in physiological indicators, and strong constraints on the timing of diagnosis and treatment interventions. Clinical decisions need to comprehensively consider information from multiple dimensions, including vital signs, laboratory tests, imaging reports, diagnosis and treatment records, and medication records, and strictly follow clinical timelines and standardized diagnosis and treatment guidelines.

[0003] Currently used clinical decision support systems and treatment guideline reminder tools have significant technical deficiencies: First, existing knowledge bases are mostly static and fixed settings, which cannot dynamically adjust the association weights of knowledge nodes based on real-time clinical data of patients and lack the ability to adapt to individualized cases. Second, existing systems mostly provide isolated reminders at single points, which cannot generate a well-structured recommendation path for the entire disease course, resulting in fragmented and unsystematic clinical decision-making suggestions. Third, existing early warning systems are mostly passive, post-event alerts, which cannot proactively predict potential risks of complications and prolonged hospital stays based on multimodal temporal change trends, and lack the ability to intervene proactively in advance. Fourth, existing technologies are insufficient to integrate and process heterogeneous clinical data from multiple sources and form a standardized feature matrix, thus failing to provide a unified and reliable data foundation for intelligent reasoning; Fifth, existing technologies lack an integrated closed-loop operating architecture, and do not have the capabilities for comparing diagnostic and treatment differences, risk classification, quantitative scoring, and traceability quality control. They are also difficult to rely on clinical feedback to achieve iterative optimization of models and knowledge rules, resulting in limited overall applicability and clinical management value.

[0004] Therefore, there is an urgent need for an AI-assisted decision-making solution for neurocritical care that integrates dynamic knowledge graphs, multimodal temporal data processing, and temporal reasoning. This solution should rely on adaptive weighting of knowledge graphs, multimodal feature modeling, intelligent identification of diagnosis and treatment stages, temporal path generation, intelligent review of medical orders, prognostic risk assessment, traceability of diagnosis and treatment quality control, and closed-loop iterative optimization to achieve intelligent assisted decision-making throughout the entire neurocritical care process. Summary of the Invention

[0005] This invention aims to overcome the technical shortcomings of existing technologies, such as static and fixed knowledge, fragmented decision-making suggestions, passive early warning mechanisms, inability to cover the entire time-series diagnosis and treatment process, and lack of implementation architecture and visualization simulation support. It provides an AI-assisted decision-making system for neurocritical care based on knowledge graphs and multimodal time-series reasoning, applicable to the entire course of diagnosis and treatment of neurocritical care patients. Relying on dynamic knowledge graph construction, multimodal time-series data fusion, intelligent identification of diagnosis and treatment stages, time-series path reasoning and recommendation, real-time review and early warning of medical orders, and prognostic risk prediction and intervention, it provides dynamic, continuous, traceable, and quantifiable personalized full-path intelligent clinical auxiliary decision support for clinical practice.

[0006] This invention provides a knowledge graph and multimodal temporal reasoning-based AI-assisted decision-making system for neurocritical care, comprising: The graph construction unit is used to build and maintain a dynamic knowledge graph of neurocritical care, and adaptively adjusts the node association weights based on real-time clinical data of patients. The data processing unit is used to connect with the hospital's multi-business system to collect multimodal time-series data, and generate a multimodal time-series feature matrix after data processing; The matching and locking unit is used to intelligently identify the diagnosis and treatment stage based on the matrix through a reasoning model that supports dual-mode switching, match the dynamic knowledge graph and lock the standard diagnosis and treatment path, necessary intervention measures and diagnosis and treatment contraindication rules; The temporal reasoning unit is used to generate a structured recommendation path for the entire disease course by employing a temporal reasoning model that integrates graph neural networks and attention mechanisms. The multi-source prompting unit is used to perform real-time medical order comparison, proactive warning of contraindications, and prompting of missing items. The risk assessment unit is used to conduct prognostic risk assessment based on multimodal time-series change trends. The risk assessment prediction model predicts the risk of complications and prolonged hospital stay, and outputs stratified prospective intervention recommendations. The diagnosis and treatment assessment unit is used to conduct comparisons of differences in diagnosis and treatment, risk classification, total score assessment, and trace the source of deductions based on clinical gold standards; The closed-loop update unit is used to construct a fully closed-loop operation mode, relying on human-computer interaction and clinical feedback to optimize and update the model and dynamic knowledge graph rules.

[0007] Optionally, in the knowledge graph construction unit, the dynamic knowledge graph sets five types of entities: disease, symptom, examination, drug, and treatment stage, and configures eight core directional relationships: trigger, need, guidance, adaptation, contraindication, association, recommendation, and monitoring. The knowledge graph adopts a directed weighted graph structure, and the association weights range from 0 to 1. A data-driven cyclic iterative update mechanism is established to perform batch verification and adaptive correction of all association weights in the knowledge graph based on the latest patient clinical data.

[0008] Optionally, in the data processing unit, multimodal time-series data are divided into four categories: high-frequency time-series vital sign data, discrete time-point laboratory test results, unstructured text / image feature-based imaging reports, and event sequence-based diagnostic and treatment procedures and medication records. Data processing includes data cleaning and preprocessing, unified time axis construction, multimodal time-series alignment, and structured feature extraction. A unified reference time axis is constructed with the patient's admission time as the reference time T0 and a fixed time window of 1 hour. Time-series interpolation, mean imputation, and the 3σ criterion are used to complete anomaly removal and normalization. Feature extraction is completed through time-series feature calculation, clinical event coding, and image text feature vectorization.

[0009] Optionally, in the matching and locking unit, the diagnosis and treatment stages are divided into acute phase, stable phase, and recovery phase; the inference model adopts a temporal neural network classification model based on LSTM, which supports switching between temporal neural network classification and clinical professional rule inference modes; a confidence level of ≥0.8 is set as the effective judgment threshold, and when the confidence level does not meet the standard, a manual confirmation prompt is pushed, allowing doctors to manually verify the diagnosis and treatment stage and lock the standard diagnosis and treatment path and constraint rules.

[0010] Optionally, in the temporal reasoning unit, the temporal reasoning model is a hybrid model that integrates graph neural networks and attention mechanisms. The model input includes the patient's multimodal state vector at time T, the activated diagnosis and treatment stages, and the path constraints and rules of the dynamic knowledge graph. The multimodal state vector is generated by extracting the current time feature slice from the multimodal temporal feature matrix. The attention mechanism assigns high weight to the core clinical features of ICP and consciousness state, and outputs a temporalized diagnosis and treatment recommendation sequence containing examination, treatment, and medication according to different time windows, and annotates the corresponding medical basis.

[0011] Optionally, in the multi-source prompting unit, the time-series diagnosis and treatment recommendation sequence generated by the time-series reasoning unit is retrieved through the real-time comparison and analysis engine, and the executed clinical medical orders are obtained from the HIS in real time, and the comparison and verification are performed item by item and time window by time according to the three categories of examination, treatment and medication.

[0012] Optionally, in the risk assessment unit, the risk assessment prediction model uses a Transformer deep learning prediction network to calculate the risk probability through temporal feature extraction and fusion, and outputs high, medium and low risk levels for prolonged hospital stay; it automatically identifies persistent intracranial hypertension, postoperative infection and cerebral vasospasm as key influencing factors leading to prolonged hospital stay, and outputs prospective intervention recommendations in a differentiated manner according to risk level.

[0013] Optionally, in the diagnosis and treatment assessment unit, the actual diagnosis and treatment procedures are compared with the standard recommended path item by item based on the clinical gold standard, and the deviations in process logic are divided into high-risk and medium-risk levels; deviation scenarios such as delayed first dose of anti-infection, lack of intracranial pressure support, delayed etiological testing, and timeout of efficacy review are automatically marked and graded; a total score of 100 is used for evaluation, and the source of deductions is traced for each item that does not conform to process logic.

[0014] Optionally, the closed-loop update unit adopts a fully closed-loop operation mode of data collection—multimodal temporal fusion—knowledge graph reasoning—temporal path recommendation—intelligent review of medical orders—risk warning output—clinical feedback iteration; all treatment recommendation paths, high-risk warning information, prognostic risk prompts and prospective intervention suggestions are uniformly output through a standardized visual human-computer interaction interface.

[0015] Optionally, the closed-loop update unit supports doctors to manually correct the judgment results of the diagnosis and treatment stage, independently adopt or reject the system's recommended items, manually remove high-risk warning restrictions, and enter personalized clinical diagnosis and treatment feedback; clinical feedback data is automatically archived and serves as the core data source for model iteration training samples and dynamic knowledge graph rule optimization and updates, forming a continuous positive loop mechanism of clinical data—model reasoning—clinical feedback—model iteration optimization.

[0016] One or more technical solutions provided by this invention have at least the following technical effects or advantages: To address the shortcomings of existing knowledge bases that are static and fixed and cannot dynamically adapt to changes in patients' conditions, this invention constructs a dynamic knowledge graph for neurocritical care using graph construction units. This graph can adaptively adjust the node association weights and can dynamically optimize knowledge associations based on real-time clinical data of patients. This achieves individualized adaptation of medical knowledge to patients' conditions and overcomes the problems of poor adaptability and weak universality of traditional static knowledge bases.

[0017] To address the shortcomings of existing decision-making recommendations being fragmented and lacking structured guidance throughout the entire disease course, this invention processes multimodal time-series data through a data processing unit and constructs a feature matrix. A matching and locking unit identifies the diagnosis and treatment stages based on dual-mode reasoning, and a graph neural network fusion attention mechanism combined with a time-series reasoning unit generates a structured recommendation path throughout the entire disease course. This makes the diagnosis and treatment recommendations possess temporal logic and overall coherence, thereby enhancing their clinical guidance value.

[0018] To address the shortcomings of existing early warning methods, which are passive, lagging, and lack forward-looking prediction, this invention uses a multi-source prompting unit to perform real-time comparison of medical orders to achieve proactive early warning of contraindications and prompting of missing items. At the same time, a risk assessment unit conducts prognostic risk assessment based on multimodal time-series change trends, predicts the risk of complications and prolonged hospital stay, and provides stratified intervention suggestions, thus upgrading from passive post-event alerts to proactive and forward-looking prevention and control.

[0019] To address the shortcomings of existing multi-source clinical data being fragmented and unable to be used for unified modeling and reasoning, this invention connects a data processing unit to multiple business systems in a hospital to collect multimodal time-series data. After standardized data processing, a unified multimodal time-series feature matrix is ​​generated, providing standardized and reliable data support for the identification of diagnosis and treatment stages and the recommendation of time-series paths.

[0020] To address the shortcomings of existing technologies, such as the lack of an overall architecture and the difficulty in conducting full-process quality control and iterative optimization, this invention establishes a complete functional unit implementation system. Based on clinical gold standards, it realizes the comparison of diagnostic and treatment differences, risk classification, total score assessment, and traceability of deduction items. At the same time, it establishes a closed-loop operation mode, relying on human-computer interaction and clinical feedback to continuously optimize the model and knowledge graph rules, improve the ability of diagnosis and treatment compliance quality control, process traceability, and continuous iteration, and enhance the standardization level of clinical diagnosis and treatment of neurocritical care. Attached Figure Description

[0021] Figure 1 This is a flowchart of a neurocritical care AI-assisted decision-making system based on knowledge graphs and multimodal temporal reasoning, according to the present invention. Figure 2 This is a schematic diagram of the multimodal data temporal alignment and fusion processing and feature extraction process of the present invention; Figure 3 This is a flowchart of the diagnosis and treatment stage identification and path locking process of the present invention; Figure 4 This is a schematic diagram illustrating the reasoning and recommendation generation results of the diagnostic and treatment pathway in this invention; Figure 5 This is a schematic diagram illustrating the operation of the real-time medical order comparison and contraindication proactive early warning mechanism of the present invention; Figure 6 This is a schematic diagram illustrating the risk assessment and prediction of prolonged hospital stay and the generation of prospective intervention suggestions according to the present invention. Figure 7 This is a schematic diagram of the human-computer interaction interface for reviewing medical orders according to the present invention; Figure 8 This is a schematic diagram of the dynamic knowledge graph node relationships and node attribute details interface of this invention; Figure 9 This is a schematic diagram of the diagnosis and treatment recommendation results interface based on the gold standard of the present invention; Figure 10 This is a schematic diagram of the diagnostic and treatment difference comparison and scoring interface of the present invention. Detailed Implementation

[0022] This invention aims to overcome the technical shortcomings of existing technologies, such as static and fixed knowledge, fragmented decision-making suggestions, passive early warning mechanisms, inability to cover the entire time-series diagnosis and treatment process, and lack of implementation architecture and visualization simulation support. It provides an AI-assisted decision-making system for neurocritical care based on knowledge graphs and multimodal time-series reasoning, applicable to the entire course of diagnosis and treatment of neurocritical care patients. Relying on dynamic knowledge graph construction, multimodal time-series data fusion, intelligent identification of diagnosis and treatment stages, time-series path reasoning and recommendation, real-time review and early warning of medical orders, and prognostic risk prediction and intervention, it provides dynamic, continuous, traceable, and quantifiable personalized full-path intelligent clinical auxiliary decision support for clinical practice.

[0023] The following is combined with Figure 1 The following is an overall description of the processing flow of a neurocritical care AI-assisted decision-making system based on knowledge graphs and multimodal temporal reasoning (hereinafter referred to as the system) in this embodiment of the invention: The system takes multimodal patient data covering four categories of patient diagnosis and treatment data (vital signs, examinations, imaging, and drug treatment) and a dynamic neurocritical care knowledge graph providing clinical knowledge support as two core inputs. Both inputs enter the multimodal data integration and temporal alignment module to complete the standardized integration and temporal dimension alignment of the data. The processed data further enters the intelligent identification and treatment stage and the standard path activation module to complete the intelligent identification of the patient's diagnosis and treatment stage and the activation of the corresponding standard treatment path. The identification results and... The activation path involves a multimodal temporal reasoning module that shares the core input. This module outputs reasoning results in three directions: data prediction results to the data trend prediction module, treatment recommendation sequences to the structured recommendation sequence generation module, and intervention prompts to the prospective intervention prompt module. The output of the structured recommendation sequence generation module, along with the executed medical orders data, enters the real-time comparison and proactive warning module to complete the compliance comparison and risk warning of medical orders. Finally, the outputs of the prospective intervention prompt module, the multimodal temporal reasoning module, and the real-time comparison and proactive warning module are aggregated into the full-path intelligent auxiliary decision-making output module to complete the final auxiliary decision-making result output.

[0024] Specifically, the system includes a graph construction unit, a data processing unit, a matching and locking unit, a temporal reasoning unit, a multi-source prompting unit, a risk assessment unit, a diagnosis and treatment assessment unit, and a closed-loop update unit.

[0025] The knowledge graph construction unit is used to build and maintain a dynamic knowledge graph for neurocritical care: it integrates authoritative medical knowledge sources in the field of neurocritical care to build a knowledge graph containing five types of entities: disease, symptoms, examination, drugs, and treatment stages, and establishes directional relationships between these entities; the association weights between nodes in the dynamic knowledge graph can be dynamically and adaptively adjusted based on real-time clinical data of neurocritical care patients.

[0026] In the graph construction unit, the dynamic adaptive adjustment of node association weights is achieved by automatically correcting the weight values ​​of the association edges between entity nodes in the dynamic knowledge graph based on the real-time key physiological indicators and time-series changes of critically ill neurological patients, thereby realizing individualized dynamic adaptation of the association strength of medical knowledge.

[0027] The processing flow of the map construction unit specifically includes the following sub-steps: S1.1, Basic Construction of Knowledge Graph: Integrating the "Chinese Guidelines for the Diagnosis and Treatment of Neurocritical Care 2021", clinical research literature from core journals, standardized medical records of neurocritical care in tertiary hospitals and consensus of experts in the field as authoritative medical knowledge sources, the basic architecture of dynamic knowledge graph is built by adopting ontology modeling, mainly from three dimensions: entity definition, relation definition, and graph structure definition.

[0028] Entity Definitions: Five core entity types are defined: disease entity, symptom entity, examination entity, drug entity, and treatment stage entity. Among them, the disease entity covers typical symptoms of severe neurological diseases such as severe traumatic brain injury, massive cerebral hemorrhage, and aneurysmal subarachnoid hemorrhage; the symptom entity includes clinical manifestations such as increased intracranial pressure, altered consciousness, and hemiplegia; the examination entity includes diagnostic and treatment examinations such as head CT (Computed Tomography), head MRI (Magnetic Resonance Imaging), ICP (Intracranial Pressure) monitoring, and complete blood count; the drug entity includes commonly used clinical drugs such as mannitol, sodium valproate, and anticoagulants; and the treatment stage entity is divided into three standard phases: acute phase, stable phase, and rehabilitation phase.

[0029] Relationship Definition: Eight core directional relationships are set up, namely, triggering relationship, need relationship, guidance relationship, adaptation relationship, taboo relationship, association relationship, recommendation relationship, and monitoring relationship; each relationship constitutes a directional logical link between entities, providing a rule basis for intelligent reasoning of the knowledge graph.

[0030] Graph structure definition: The knowledge graph is set as a directed weighted graph structure, with various entities as graph nodes, and the relationships between entities as logical edges connecting the nodes. The weight values ​​configured on the logical edges are used to represent the strength of the relationship between entities. The entire system uses Neo4j (graph database) for underlying storage to build an initial static neurocritical care knowledge graph.

[0031] Depend on Figure 8As can be seen, the dynamic knowledge graph adopts a hierarchical architecture, with the overall structure organized sequentially through stage nodes, main process nodes, and planned item nodes. Each level of node achieves overall logical connection based on "planned projects," "step-by-step" processes, and "parallel execution." Simultaneously, each node has built-in multi-dimensional attribute parameter fields such as stage type, time sequence number, stage identifier, and classification channel, providing data and structural support for subsequent intelligent identification of diagnosis and treatment stages, automatic unfolding of standard treatment pathways, and precise matching and invocation of clinical rules.

[0032] Examples of entity association logic: Disease entities and symptom entities have a triggering relationship, such as "severe traumatic brain injury - triggering - increased intracranial pressure"; symptom entities and examination entities have a need relationship, such as "increased intracranial pressure - needing - head CT"; examination entities and drug entities have a guiding relationship, such as "intracranial pressure monitoring - guidance - mannitol use"; diagnosis and treatment stage entities and intervention measures have an adaptation relationship, such as "acute phase - adaptation - dehydration to reduce intracranial pressure"; diagnosis and treatment stage entities and drug entities have a contraindication relationship, such as "acute phase - contraindication - anticoagulants".

[0033] S1.2, Dynamic weight adjustment mechanism: The weight of the knowledge graph association is uniformly set to a value range of 0 to 1, and the initial default weight value of all associations is configured to 0.5; the larger the weight value, the stronger the association between entities; when the weight value is 0, it means that there is no effective association between the corresponding entities.

[0034] The knowledge graph management module has a built-in weight calculation unit that pre-configures the weight mapping rules between key clinical indicators and graph entities. It extracts key indicators such as patient ICP, blood pressure, heart rate, inflammatory markers, and Glasgow Coma Scale scores in real time, and automatically and dynamically adjusts the weight of the edges between corresponding entity nodes in the knowledge graph based on the real-time values ​​and time-series fluctuation trends of the indicators.

[0035] The criteria for determining abnormal key indicators are uniformly set as exceeding the normal physiological reference range for the human body. The exemplary rules are as follows: The normal reference range for intracranial pressure is 5 mmHg to 15 mmHg. When a patient's ICP test value exceeds this range, it is determined to be elevated intracranial pressure. At the same time, the weight of the two correlations "elevated intracranial pressure - need to perform - head CT" and "elevated intracranial pressure - recommended - mannitol" is increased from 0.5 to 0.9. When the patient's condition changes from the acute phase to the stable phase and the intracranial pressure drops back to the normal physiological reference range, the correlation weight of "acute phase - appropriate - dehydration treatment" is decreased from 0.8 to 0.2.

[0036] At the same time, a data-driven cyclic iterative update mechanism is established. Every 30 minutes, the system performs batch verification and adaptive correction of all associated weights of the knowledge graph based on the latest patient clinical data, so as to realize the individualized real-time update of the dynamic knowledge graph as the disease progresses.

[0037] The data processing unit is used to acquire and integrate patients' multimodal time-series data: it connects to various business information systems of the hospital, collects patients' vital signs, laboratory tests, imaging reports, diagnosis and treatment operation records and medication records in real time, and sequentially completes data cleaning and preprocessing, unified time axis construction, multimodal time-series alignment, and structured time-series feature extraction, and finally generates a standardized multimodal time-series feature matrix to provide compliant input data for subsequent diagnosis and treatment identification and time-series reasoning models.

[0038] The data processing unit's processing flow specifically includes the following sub-steps: S2.1, Multi-source raw data acquisition: Please refer to... Figure 2 The multimodal data fusion module retrieves patients' raw multimodal time-series data in real time through standardized interfaces of HIS (Hospital Information System), LIS (Laboratory Information System), PACS (Picture Archiving and Communication System), and CCIS (Critical Care Information System).

[0039] The collected raw multimodal time series data are divided into four categories according to type: First, high-frequency time-series vital sign data: including blood pressure, heart rate, respiration, intracranial pressure (ICP), etc., with a collection frequency of 5 minutes / time. Second, discrete time-point laboratory test results: including data from tests such as complete blood count, coagulation function, and inflammatory markers; Thirdly, unstructured text / image feature-based imaging reports: including text descriptions and image feature data from head CT and MRI scans; Fourthly, event sequence-based medical procedures and medication records: including surgical and puncture procedure records, as well as medication behavior information such as medication time, dosage, and frequency.

[0040] S2.2, Temporal Alignment and Fusion Processing - Data Cleaning and Preprocessing: Standardized preprocessing is performed on the collected raw multimodal time-series data, including using temporal interpolation to complete missing vital sign data; using mean imputation to complete missing laboratory test data; identifying and removing abnormal data samples based on the 3σ criterion; normalizing various clinical indicators of different dimensions and magnitudes, uniformly mapping them to the standard numerical range of 0 to 1, eliminating interference caused by differences in dimensions, and achieving data standardization, noise reduction, and missing value imputation.

[0041] S2.3, Time Sequence Alignment and Fusion Processing - Unified Time Axis Construction: Taking the time when the patient is admitted to the neurocritical care unit as the baseline time T0, and using 1 hour as a fixed time window, the system is divided into multiple levels of standard time nodes T1, T2, T3...Tn to establish a globally unified baseline time axis, providing a unified time reference system for the full volume of multimodal data.

[0042] S2.4, Temporal Alignment and Fusion Processing - Multimodal Temporal Alignment: Through the multimodal temporal alignment engine, temporal alignment operations are performed on the full dataset, including downsampling / aggregation of high-frequency vital sign data; time window mapping of discrete laboratory test results and imaging reports; extraction and structuring of key diagnosis and treatment time points; and event time stamp alignment of diagnosis and treatment operations and medication event sequences. Finally, an aligned multimodal data view is generated, achieving accurate mapping of the full dataset on a unified time axis.

[0043] S2.5, Feature Extraction and Matrix Generation: After completing time-series alignment, feature mining and structured modeling are performed on all standardized data, including: Time series feature calculation: Extracting time series features such as the changing trends, volatility, and statistics of vital signs indicators; Clinical event coding: Standardize clinical event coding for key diagnostic and treatment procedures and medication status; Image text feature vectorization: NLP (Natural Language Processing) technology is used to extract entities and perform semantic parsing on the image report, and combined with image feature embedding to achieve full-modal feature vectorization; Finally, a standardized multimodal temporal feature matrix is ​​generated and integrated, serving as structured input data to provide standardized input support for subsequent diagnosis and treatment identification and temporal reasoning models.

[0044] Matching and locking units are used to intelligently identify the stage of diagnosis and treatment and lock in the standard treatment path: Please refer to Figure 3 Based on the multimodal temporal feature matrix output by the data processing unit, the inference model is invoked to identify the diagnosis and treatment stage. After determining the current diagnosis and treatment stage of the neurocritical patient, the dynamic knowledge graph is accessed and the corresponding stage knowledge is matched. The standard diagnosis and treatment path, necessary intervention measures and diagnosis and treatment contraindication rules corresponding to the diagnosis and treatment stage are activated in the dynamic knowledge graph. The integration and locking of the standard diagnosis and treatment path and constraint rules are completed, and the final path locking result is output for the temporal inference unit to generate diagnosis and treatment recommendations.

[0045] In the matching and locking unit, the diagnosis and treatment stages are divided into at least three categories: acute phase, stable phase, and recovery phase. The inference model supports dual-mode switching, specifically including temporal neural network classification mode and clinical professional rule inference mode, which can be flexibly adapted and called according to different clinical application scenarios to meet the intelligent recognition needs of different diagnosis and treatment scenarios.

[0046] This embodiment preferably uses a temporal neural network classification model based on LSTM (Long Short-Term Memory) as the inference model. During the model training phase, historical diagnosis and treatment data from the neurocritical care department are used as the base dataset. A multimodal temporal feature matrix labeled with acute, stable, and recovery phase classifications is used as the model input, and the classification results of the patient's corresponding diagnosis and treatment phase are used as the model output. Iterative training and optimal parameter tuning are completed to achieve the desired model performance. After training is successful, the model is deployed to the phase recognition and inference engine module for intelligent recognition and inference of the patient's real-time diagnosis and treatment phase.

[0047] The system reserves a clinical rule reasoning engine interface, supporting bidirectional switching between model reasoning and rule reasoning to adapt to different clinical application scenarios. The rule reasoning engine incorporates standardized thresholds for determining the diagnosis and treatment stage, along with authoritative clinical evidence. An example of the acute phase determination rule is: the patient's ICP > 20 mmHg, Glasgow Coma Scale score ≤ 8, and imaging evidence of intracranial hemorrhage or traumatic brain injury on head CT or MRI. Meeting all of these conditions determines the patient to be in the acute phase.

[0048] The workflow of the stage identification and inference engine is as follows: It receives the multimodal temporal feature matrix output by the data processing unit, calls the trained LSTM temporal neural network classification model to identify the treatment stage, performs real-time inference calculations, and outputs the patient's current treatment stage determination result and corresponding confidence level. The system sets a confidence level ≥ 0.8 as the effective determination threshold. If the confidence level of the inference result meets the threshold, it automatically accesses the dynamic knowledge graph, matches the stage knowledge subset corresponding to the current treatment stage, and accurately activates the corresponding standard treatment path, necessary intervention measures, and treatment contraindication rules. If the confidence level of the inference result < 0.8, the determination result is not valid enough, and the system proactively pushes a manual confirmation prompt to the clinician. The doctor manually verifies and selects the patient's current treatment stage, completing the integration and locking of the standard treatment path and constraint rules. This provides reliable rule basis and path support for subsequent multimodal temporal inference to generate structured treatment recommendation paths.

[0049] The temporal reasoning unit is used to generate a structured recommendation path for the entire disease course based on multimodal temporal reasoning: combining dynamic knowledge graph path constraints and rules, individualized multimodal temporal features of patients, and patient diagnosis and treatment stages, a temporal reasoning model with graph neural network and attention mechanism is used to deduce and generate a structured recommendation path for the entire disease course that includes examination items, treatment operations, drug use, and suggested execution time windows.

[0050] In the temporal reasoning unit, the temporal reasoning model is a hybrid model that combines graph neural networks with attention mechanisms. It can deeply mine the entity association logic within the dynamic knowledge graph and the weight distribution of the patient's core clinical features, and generate a sequence of diagnosis and treatment operations and drug use that is highly adapted to the patient's condition in the time dimension.

[0051] Please refer to Figure 4 In this embodiment, the temporal reasoning process employs a temporal reasoning model that integrates graph neural networks with an attention mechanism. The model sets up three types of input architectures, namely: 1. Patient's multimodal state vector at time T (including individualized multimodal temporal features such as vital signs, test results, and imaging characteristics); 2. Activated treatment phase (output by the matching and locking unit, such as acute phase, stable phase, and recovery phase); 3. Path constraints and rules for dynamic knowledge graphs.

[0052] Specifically, the model learns and analyzes the inherent logical connections and constraints between medical entities within the knowledge graph through graph neural networks, and then uses an attention mechanism to deeply filter the multimodal temporal features of the input. It assigns high weights to core clinical features that directly affect diagnosis and treatment decisions, such as patient ICP and state of consciousness, so that the model focuses on key disease indicators, weakens invalid interference features, and ensures the clinical accuracy of temporal reasoning.

[0053] During the model training phase, standardized diagnostic and treatment pathway data for neurocritical care is used as the core training sample. The model fully learns the timing matching logic, combination rules, and adaptation conditions of various examination items, treatment plans, and drug use under different diagnostic and treatment stages and different disease severity. The model completes iterative training and optimal parameter tuning. After the training reaches the target, it is deployed to the pathway recommendation and medical order review modules.

[0054] The specific reasoning process is as follows: The path recommendation and medical order review module retrieves the treatment stage, standard treatment path, necessary intervention measures, and treatment contraindication rules locked by the matching and locking unit. Simultaneously, it extracts the feature slice corresponding to the current reasoning time T from the multimodal temporal feature matrix output by the data processing unit, generating the patient's multimodal state vector at time T, which is then input into the trained temporal reasoning model. The model combines knowledge graph path constraints and rules to adaptively optimize the weights of the patient's individualized multimodal temporal features, completing refined temporal deduction in the time dimension and outputting a temporalized treatment recommendation sequence. The recommended content is uniformly divided into three categories: examination, treatment, and medication, including standardized suggestion execution time windows, clinical execution priorities, and corresponding medical evidence. It can also be dynamically updated iteratively as the patient's condition evolves in real time, achieving individualized adaptation and adjustment.

[0055] The following are exemplary recommended generation results from this invention: 1. T+1h time window: Recommended examination: repeat head CT scan, based on worsening of consciousness impairment; Treatment recommendation: Adjust the dehydration and intracranial pressure reduction regimen based on intracranial pressure monitoring trends; Recommended medication: Initiate antiepileptic drugs (sodium valproate) based on abnormal EEG waves; 2. T+4h time window: Recommended tests: complete blood count + coagulation function, based on routine monitoring; Treatment recommendation: Maintain current treatment and closely monitor the patient's condition, based on the fact that the condition is relatively stable; Recommended medication: Maintain infusion of antihypertensive medication, based on blood pressure control targets; 3. T+24h time window: Recommended examination: Comprehensive review and assessment of surgical indications, based on periodic evaluations; Treatment recommendation: Consider early rehabilitation intervention, based on limb function status; Medication recommendation: Adjust the anti-infection regimen based on changes in inflammatory markers.

[0056] All recommended projects are clearly marked with their execution time window, priority level, and corresponding medical basis, forming a visual time-series diagnosis and treatment recommendation path that can be directly implemented.

[0057] Combination Figure 9 As shown in the clinical gold standard scenario, this invention takes severe intracranial neurological infection as a typical application scenario. Based on the preset clinical gold standard rule base, it automatically generates a standardized main process timeline of the entire disease course, covering the entire process of diagnosis classification, diagnosis confirmation, emergency treatment during the T0 window period, condition review during T0+24-72h, efficacy evaluation during T0+48-72h, long-term treatment management, rehabilitation prevention and complication management, and constructing a standardized time-series diagnosis and treatment path that conforms to clinical norms and can be implemented.

[0058] The complete implementation rules for the clinical gold standard for intracranial infection are as follows: I. Main process timeline; 1. Diagnosis and classification: The preferred examination is a plain MRI scan of the head with contrast enhancement. If no examination is available, a plain CT scan of the head with contrast enhancement is used. The disease assessment dimensions include physical signs such as fatigue and positive Kernig's sign / Brudzinski's sign, signs of systemic infection such as fever, vomiting, and chills, signs of brain parenchymal involvement such as focal neurological deficits, epilepsy, and mental and behavioral abnormalities, as well as core symptoms such as headache, neck stiffness, altered consciousness, increased intracranial pressure, and meningeal irritation signs.

[0059] 2. Diagnostic Confirmation: Pathogen confirmation is achieved through pathogen culture, multiplex PCR (Polymerase Chain Reaction), mNGS (metagenomic next-generation sequencing), routine cerebrospinal fluid analysis, cerebrospinal fluid biochemistry, acid-fast staining / tuberculosis culture, pathogen molecular PCR detection, Cryptococcus fungal antigen detection, and cerebrospinal fluid examination (clinical gold standard); drug susceptibility testing is carried out simultaneously to provide a basis for subsequent precise medication.

[0060] 3. Emergency Management (T0, within 1-3 hours): Basic management includes maintaining the patient's electrolyte balance, elevating the head of the bed to 30°, managing fever (using acetaminophen or physical cooling), symptomatic control of epilepsy (using levetiracetam, etc.), and preventing hyponatremia complications; Targeted conditional treatment plans are as follows: Adult patients with community-acquired bacterial meningitis are treated with ceftriaxone or cefotaxime combined with vancomycin, and the glucocorticoid dexamethasone may be added or discontinued as appropriate; Postoperative, post-traumatic, and nosocomial infection patients are treated with vancomycin combined with meropenem or cefepime; Age > 5 For patients aged 0 years or those with immunosuppression and suspected Listeria infection, add ampicillin; for patients suspected of herpes simplex virus encephalitis, immediately initiate intravenous acyclovir treatment; for patients suspected of cryptococcal meningitis, use amphotericin B liposome combined with flucytosine; for patients suspected of tuberculous meningitis, use a quadruple anti-tuberculosis regimen of isoniazid, rifampin, pyrazinamide, and ethambutol. Adults can use dexamethasone at a dose of 0.15 mg / kg every 6 hours for 2-4 days; for patients with intracranial hypertension, mannitol osmotic dehydration therapy or EVD (External Ventricular Drainage) intracranial pressure intervention can be carried out according to the condition.

[0061] 4. Mid-term adjustment at T0+24—72h: Based on the cerebrospinal fluid culture results and pathogen molecular detection results, broad-spectrum antibiotics are adjusted to narrow-spectrum targeted drugs, drugs with good blood-brain barrier penetration are given priority, and the medication regimen is optimized in combination with drug sensitivity results.

[0062] 5. T0+48—72h efficacy evaluation: Conduct efficacy review based on the patient's clinical signs and symptom improvement. If the condition does not improve or continues to worsen, repeat cerebrospinal fluid indicators and cranial imaging, and adjust the treatment plan in a timely manner.

[0063] 6. Long-term diagnosis and treatment management: Herpes simplex virus encephalitis is treated with acyclovir for 14-21 days with a total follow-up period of more than 1 year; bacterial meningitis is routinely treated for 10-14 days, which is dynamically adjusted according to the type of pathogen and clinical recovery; tuberculous meningitis is treated for a total of 9-12 months; cryptococcal meningitis is managed in a step-by-step manner with induction therapy, consolidation therapy and maintenance therapy.

[0064] 7. Rehabilitation and Prevention Management: Specialized neurorehabilitation treatments are provided for sequelae such as hearing loss, cognitive impairment, and limb movement disorders; vaccinations against Streptococcus pneumoniae, Neisseria meningitidis, and Haemophilus influenzae are guided according to the type of pathogen infecting the patient; and preventive drug interventions are carried out for close contacts of patients with Neisseria meningitidis infection.

[0065] II. Rules for the Management of Concurrent Complications; Throughout the process, monitor the patient's head circumference and level of consciousness, promptly screen for hydrocephalus complications, and perform external ventriculoperitoneal shunt (EVD) or VPS (Ventriculoperitoneal shunt) surgical intervention when necessary. For patients with brain abscess or subdural empyema, when the lesion diameter is >2.5cm, has a significant mass effect, and drug treatment is ineffective, evaluate and perform surgical drainage surgery. For patients with venous sinus thrombosis, use low molecular weight heparin anticoagulation therapy. For patients with cortical involvement or recurrent epilepsy, implement long-term standardized antiepileptic treatment.

[0066] The multi-source prompting unit is used to perform real-time medical order comparison, proactive contraindication warnings, and omission item prompts: Please refer to [link / reference]. Figure 5 , Figure 7 The path recommendation and medical order review module incorporates a real-time comparison and analysis engine. It pre-sets standardized core comparison algorithms and clinical rule verification logic, comparing the time-series treatment recommendation sequences generated by the time-series reasoning unit with clinically executed medical orders item by item and time window by time window to complete intelligent compliance review of medical orders. During the review process, it relies on dynamic knowledge graph association rules for reasoning support, proactively issuing warnings and blocking items with contraindications and conflicts, and intelligently prompting for missing necessary items. Simultaneously, for scenarios where necessary treatment items are omitted, it uses dynamic knowledge graph reasoning to match clinically equivalent solutions and outputs supplementary solutions such as alternative examinations, alternative treatments, or alternative medications, achieving fully automated compliance quality control and intelligent optimization of medical orders.

[0067] The specific comparison and review process is as follows: The real-time comparison and analysis engine obtains core data sources bidirectionally. On the one hand, it retrieves the time-series diagnosis and treatment recommendation sequence, which includes examinations, treatments, medications, and corresponding time windows and priorities, as ultimately output by the time-series reasoning unit. On the other hand, it retrieves the executed clinical orders from clinicians in real time by connecting to the HIS. According to the three categories of examination items, treatment operations, and medication use, the time-series diagnosis and treatment recommendation sequence and the executed clinical orders are compared and verified precisely item by item and time window by time window. At the same time, the entity association rules and clinical contraindication rules of the dynamic knowledge graph provide intelligent reasoning support for the entire comparison process, ensuring the professionalism, accuracy, and clinical rationality of the order comparison results.

[0068] The multi-source prompting unit implements two core intelligent auditing functions: proactive warning and blocking of taboo conflicts, and intelligent prompting and supplementation of equivalent solutions for missing necessary items.

[0069] Firstly, proactive warning and blocking of contraindications: If the real-time comparison and analysis engine detects that executed clinical orders conflict with the dynamic knowledge graph's contraindication rules, such as the unauthorized use of anticoagulants in the acute phase of neurological critical care or the irrational combination of mannitol and specific diuretics, the system immediately triggers a proactive high-risk warning mechanism. Through the warning and interactive output module, the complete warning information is displayed on the doctor's terminal in a visual format (pop-up / list / highlight), clearly indicating the contraindicated item, the corresponding conflicting clinical rule, the risk level, and standardized clinical blocking recommendations, enabling early intervention and risk blocking of high-risk medical behaviors.

[0070] Secondly, the system provides intelligent prompts and equivalent solutions for missing essential items: If the comparison results show that essential acute-phase treatment items specified in the time-sequential treatment recommendation sequence are missing from the clinically executed medical orders, such as failure to promptly perform intracranial pressure monitoring or timely follow-up of head CT / MRI images, the system proactively pops up a reminder interface to provide intelligent prompts, accurately marking the missing essential items, the standard recommended execution time window, and the corresponding medical basis. For various scenarios involving missing essential treatment items, the system relies on a dynamic knowledge graph to perform deep reasoning, accurately matching the clinically equivalent treatment logic suitable for the patient's condition, and generating alternative or supplementary treatment suggestions. Especially for missing treatment items that cannot be implemented immediately due to factors such as the patient's physical condition, equipment conditions, and clinical scenario limitations, the system can intelligently output personalized supplementary treatment suggestions for alternative examinations, alternative treatments, or alternative medications, filling in the missing links in clinical treatment, maximizing the integrity and feasibility of the treatment plan, and making up for shortcomings in clinical treatment.

[0071] Combination Figure 7As can be seen from the human-computer interaction interface, the system supports manual input of patient medical record summary text. After the operator triggers the "start review" command, the background automatically links and calls the dynamic knowledge graph analysis module and the time sequence reasoning module to carry out fully automatic and multi-dimensional compliance review and diagnosis and treatment logic verification on the execution of all current medical orders of the patient. Finally, it visualizes and outputs the complete medical order review conclusion, existing diagnosis and treatment deviations and standardization optimization suggestions.

[0072] An exemplary case of medical order review in this invention is as follows: "The patient was diagnosed with intracranial infection. Upon admission, only antipyretics and fluid replacement were administered initially, without immediate cranial MRI / CT evaluation. Empirical anti-infective therapy was initiated 6 hours after T0, and a lumbar puncture for cerebrospinal fluid examination was not performed until the day after administration. Dexamethasone was administered again 8 hours after the first dose of antibiotics. The anti-infective regimen was not adjusted until 96 hours after T0, even after the etiological results were available. Cerebrospinal fluid and cranial imaging were re-examined 120 hours after T0 to assess efficacy. During this period, only intermittent fever management was provided, with no clear measures for intracranial pressure management or epilepsy treatment." The system can automatically identify the above-mentioned issues such as deviations in the timing of diagnosis and treatment, missing core items, and non-standard medication timing, and accurately output corresponding quality control rectification prompts and standardized diagnosis and treatment optimization suggestions.

[0073] The risk assessment unit is used for prognostic risk assessment and prospective intervention based on multimodal time-series change trends: the risk assessment prediction model conducts prognostic risk assessment based on multimodal time-series change trends, predicts the risk of complications and prolonged hospital stay, and outputs prospective intervention recommendations.

[0074] Combination Figure 6 The model architecture shown includes the following detailed steps in the risk assessment unit's processing flow: S6.1, Patient Multimodal Time Series Data Input: The input to the risk assessment and prediction model is patient multimodal time series data, which includes three core features: vital sign trends, laboratory test dynamics, and imaging change sequences. It is also compatible with time series sequences of diagnosis and treatment operations and drug use records, providing the model with comprehensive patient status data support.

[0075] S6.2, Risk Assessment Prediction Model Inference: In this embodiment, the risk assessment of prolonged hospital stay uses a Transformer deep learning prediction network. The model inference process is divided into two levels: Temporal feature extraction and fusion: mining long-term variation patterns of multimodal time series data, and completing the alignment and deep fusion of cross-modal features; Deep learning predictive network inference: It adopts a long short-term memory network / transformer architecture, uses the actual length of hospital stay in historical cases as training labels, and outputs the risk probability calculation results of the patient's extended hospital stay.

[0076] S6.3, Prediction Results Output: Based on the risk probability output by the model, automatically output the risk level (high, medium, low) of the patient's prolonged hospital stay, and automatically identify key influencing factors that lead to prolonged hospital stay, such as persistent intracranial hypertension, postoperative infection, cerebral vasospasm, etc.

[0077] S6.4, Prospective Intervention Recommendation Generation: The early warning and interactive output module generates personalized prospective intervention recommendations based on the prediction model's output results, targeting different risk levels. Core intervention directions include recommending early specific treatment intervention, strengthening the frequency monitoring of key indicators, and initiating rehabilitation plan assessments. Among these: For high-risk patients: it is recommended to increase the frequency of monitoring key indicators, intervene with specialized treatment in advance, and activate early rehabilitation plans; For patients at medium risk: it is recommended to optimize existing treatment plans and increase the frequency of routine assessments; For low-risk patients: It is recommended to maintain the current treatment pathway and prepare for rehabilitation assessment in advance.

[0078] For high-risk patients, the system will automatically recommend the activation of an early rehabilitation plan to reduce the risk of prolonged hospital stay through multidisciplinary collaborative intervention.

[0079] The diagnosis and treatment assessment unit is used for quantitative scoring and quality control traceability of diagnosis and treatment differences: The system compares the actual diagnosis and treatment procedures with the standard recommended pathway item by item according to the clinical gold standard, and classifies the deviations in process logic into high-risk and medium-risk levels; it automatically marks and grades deviation scenarios such as delayed first-dose anti-infection treatment, lack of concurrent intracranial pressure support treatment, delayed etiological testing, and efficacy re-evaluation exceeding the standard time window; at the same time, it adopts a total score evaluation method and traces the source of each process logic inconsistency, so as to realize the quality control of diagnosis and treatment process, compliance assessment and clinical standardization training support.

[0080] Combination Figure 10 The diagnostic and treatment difference comparison and scoring module shown includes the following detailed steps in its diagnostic and treatment assessment unit processing flow: S7.1, Gold Standard-Oriented Comparison of Treatment Differences: The system uses the clinical gold standard for neurocritical care as a benchmark to conduct precise comparisons of actual patient treatment and standardized recommended pathways item by item and time window by time window. Identified procedural logic deviations are categorized into high and medium levels based on clinical risk. Typical difference markers are shown below: High-risk procedural logic deviation: The first dose of anti-infection is administered later than T0+3h (currently approximately T0+6.00h). Medium-risk procedural logic deviation: No record of intracranial pressure / supportive treatment concurrent with anti-infection therapy during the acute phase; Medium-risk process logic deviation: Etiological outcome-oriented treatment is not within the T0+24-72h window; Medium-risk process logic deviation: efficacy reassessment is not within the T0+48-72h window.

[0081] S7.2, Quantitative Scoring and Deduction Source Tracing: The system has a built-in standardized scoring system to quantitatively assess the compliance of the diagnosis and treatment process on a 100-point scale, while also enabling precise source tracing of deductions. Overall Score Assessment: Automatically outputs the overall score for compliance with medical treatment guidelines, such as... Figure 10 The example in the middle has a total score of "61 / 100"; Source of deductions: Deductions are calculated and marked for each deviation that does not conform to the process logic. For example, the deduction item in the figure is "Process logic inconsistency: 4 items (-39)". It can accurately locate the deviation type, risk level, time window deviation and gold standard basis corresponding to each deduction, and provide a traceable and complete evidence chain for clinical quality control.

[0082] The closed-loop update unit is used to update the model and dynamic knowledge graph rules: The neurocritical care AI clinical auxiliary decision-making system of this invention adopts a fully closed-loop operation mode of "data acquisition - multimodal temporal fusion - knowledge graph reasoning - temporal path recommendation - intelligent review of medical orders - risk warning output - clinical feedback iteration". After the system starts, each functional module automatically works in coordination to complete the fully automated reasoning process of real-time acquisition of patient multimodal data, standardized preprocessing and temporal alignment, intelligent identification of diagnosis and treatment stages, recommendation of temporal paths throughout the entire disease course, comparison and review of medical order compliance, and prediction of prognostic risk stratification.

[0083] All recommended treatment pathways, high-risk warning information, prognostic risk alerts, and prospective intervention suggestions are uniformly output through a standardized, visual, and user-friendly interface, ensuring that clinicians can intuitively and efficiently obtain comprehensive diagnostic and treatment support information. The system supports full-process manual interaction by doctors, including but not limited to manually correcting the results of treatment stage assessments, independently adopting or rejecting system-recommended items, manually lifting high-risk warning restrictions, and entering personalized clinical treatment feedback.

[0084] All doctor interactions and clinical feedback data are automatically recorded and archived by the system, serving as core data sources for model iteration training samples and dynamic knowledge graph rule optimization and updates. This forms a continuous positive cycle mechanism of "clinical data - model reasoning - clinical feedback - model iteration optimization," enabling the clinical adaptability and reasoning accuracy of the system's diagnosis and treatment recommendations to be continuously improved through iteration, achieving deep integration and dynamic adaptation with the entire time-series process of neurocritical care clinical diagnosis and treatment.

[0085] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention also includes such modifications and variations.

Claims

1. A neurocritical care AI-assisted decision-making system based on knowledge graphs and multimodal temporal reasoning, characterized in that, include: The graph construction unit is used to build and maintain a dynamic knowledge graph of neurocritical care, and adaptively adjusts the node association weights based on real-time clinical data of patients. The data processing unit is used to connect with the hospital's multi-business system to collect multimodal time-series data, and generate a multimodal time-series feature matrix after data processing; The matching and locking unit is used to intelligently identify the diagnosis and treatment stage based on the matrix through a reasoning model that supports dual-mode switching, match the dynamic knowledge graph and lock the standard diagnosis and treatment path, necessary intervention measures and diagnosis and treatment contraindication rules; The temporal reasoning unit is used to generate a structured recommendation path for the entire disease course by employing a temporal reasoning model that integrates graph neural networks and attention mechanisms. The multi-source prompting unit is used to perform real-time medical order comparison, proactive warning of contraindications, and prompting of missing items. The risk assessment unit is used to conduct prognostic risk assessment based on multimodal time-series change trends. The risk assessment prediction model predicts the risk of complications and prolonged hospital stay, and outputs stratified prospective intervention recommendations. The diagnosis and treatment assessment unit is used to conduct comparisons of differences in diagnosis and treatment, risk classification, total score assessment, and trace the source of deductions based on clinical gold standards; The closed-loop update unit is used to construct a fully closed-loop operation mode, relying on human-computer interaction and clinical feedback to optimize and update the model and dynamic knowledge graph rules.

2. The system according to claim 1, characterized in that, In the knowledge graph construction unit, the dynamic knowledge graph sets five types of entities: disease, symptom, examination, drug, and treatment stage, and configures eight core directional relationships: trigger, need, guidance, adaptation, contraindication, association, recommendation, and monitoring. The knowledge graph adopts a directed weighted graph structure, with the association weights ranging from 0 to 1. A data-driven iterative update mechanism is established to perform batch verification and adaptive correction of all association weights in the knowledge graph based on the latest patient clinical data.

3. The system according to claim 1, characterized in that, In the data processing unit, multimodal time-series data are divided into four categories: high-frequency time-series vital sign data, discrete time-point laboratory test results, unstructured text / image feature-based imaging reports, and event sequence-based diagnostic and treatment procedures and medication records. Data processing includes data cleaning and preprocessing, unified time axis construction, multimodal time-series alignment, and structured feature extraction. A unified reference time axis is constructed with the patient's admission time as the baseline time T0 and a fixed time window of 1 hour. Time-series interpolation, mean imputation, and the 3σ criterion are used to complete anomaly removal and normalization. Feature extraction is completed through time-series feature calculation, clinical event coding, and image text feature vectorization.

4. The system according to claim 1, characterized in that, In the matching and locking unit, the diagnosis and treatment stages are divided into acute phase, stable phase, and recovery phase; the inference model adopts a temporal neural network classification model based on LSTM, which supports switching between temporal neural network classification and clinical professional rule inference modes; a confidence level of ≥0.8 is set as the effective judgment threshold, and when the confidence level does not meet the standard, a manual confirmation prompt is pushed, allowing doctors to manually verify the diagnosis and treatment stage and lock the standard diagnosis and treatment path and constraint rules.

5. The system according to claim 1, characterized in that, In the temporal reasoning unit, the temporal reasoning model is a hybrid model that combines graph neural networks with an attention mechanism. The model input includes the patient's multimodal state vector at time T, the activated treatment stage, and the path constraints and rules of the dynamic knowledge graph. The multimodal state vector is generated by extracting the current time-series feature slice from the multimodal temporal feature matrix. The attention mechanism assigns high weight to ICP and core clinical features of consciousness state, outputs a time-sequential diagnosis and treatment recommendation sequence including examination, treatment and medication according to different time windows, and marks the corresponding medical basis.

6. The system according to claim 1, characterized in that, In the multi-source prompting unit, the time-series diagnosis and treatment recommendation sequence generated by the time-series reasoning unit is retrieved through the real-time comparison and analysis engine. At the same time, executed clinical medical orders are obtained from the HIS in real time and compared and verified item by item and time window by time according to the three categories of examination, treatment and medication.

7. The system according to claim 1, characterized in that, In the risk assessment unit, the risk assessment prediction model uses the Transformer deep learning prediction network to calculate the risk probability through temporal feature extraction and fusion, and outputs high, medium and low risk levels for prolonged hospital stay; it automatically identifies persistent intracranial hypertension, postoperative infection and cerebral vasospasm as key influencing factors leading to prolonged hospital stay, and outputs prospective intervention recommendations in a differentiated manner according to risk level.

8. The system according to claim 1, characterized in that, In the diagnosis and treatment assessment unit, the clinical gold standard is used as the benchmark to compare the actual diagnosis and treatment with the standard recommended path item by item, and the process logic deviations are divided into high-risk level and medium-risk level. The system automatically labels and grades deviation scenarios such as delayed first-dose anti-infection, lack of intracranial pressure support, delayed etiological testing, and time-out of efficacy review; it adopts a 100-point total score evaluation method and deducts points for each item with inconsistent process logic.

9. The system according to claim 1, characterized in that, The closed-loop update unit adopts a fully closed-loop operation mode of data collection, multimodal temporal fusion, knowledge graph reasoning, temporal path recommendation, intelligent review of medical orders, risk warning output, and clinical feedback iteration. All treatment recommendation paths, high-risk warning information, prognostic risk prompts, and prospective intervention suggestions are uniformly output through a standardized and visual human-computer interaction interface.

10. The system according to claim 9, characterized in that, Within the closed-loop update unit, doctors can manually correct the judgment results of the diagnosis and treatment stage, independently adopt or reject system-recommended items, manually remove high-risk warning restrictions, and enter personalized clinical diagnosis and treatment feedback. Clinical feedback data is automatically archived and serves as the core data source for model iteration training samples and dynamic knowledge graph rule optimization updates, forming a continuous positive loop mechanism of clinical data—model reasoning—clinical feedback—model iteration optimization.

Images