Nervous system disease auxiliary evaluation method based on large model and related equipment

Abstract

The invention discloses a nervous system disease auxiliary evaluation method based on a large model and related equipment. The method comprises the following steps: constructing a structured knowledge graph based on multi-modal heterogeneous data of nervous system disease guide texts, symptom descriptions and examination results; a graph neural network module is embedded based on the structured knowledge graph in combination with a large language model, an inference network model of a dynamic graph structure is generated, and a dynamic diagnosis path tree is formed by calculating the relation weight between the medical nodes in the graph; and generating interpretable diagnosis suggestions through the reasoning network model according to the current symptom description, the examination result and the historical medical history of the patient. The problems that an existing model still has obvious bottlenecks and clinical challenges, depends on single modal data, is insufficient in quantification of dynamic signs such as a reasoning level, tremor frequency and gait abnormity, does not achieve disease development time sequence modeling and is difficult in clinical conversion can be solved.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the field of intelligent medical treatment, and more particularly, to a neural system disease auxiliary evaluation method based on a large model and related equipment. BACKGROUND

[0002] The current global artificial intelligence diagnosis and treatment system for neurology presents a differentiated development pattern. The NeuroDx system of MIT-MGH in the United States has an accuracy of 78% in early diagnosis of Parkinson's disease, the EpAlert epilepsy early warning system in Germany has a sensitivity of over 92%, and the stroke prediction model based on retinal images of Google DeepMind has an AUC of 0.91, showing the advantages of multi-modal fusion. In China, the Tencent brain blood vessel disease image analysis and the IFlytek epilepsy EEG detection technology have been certified by NMPA, and the Parkinson staging model of Fudan University Huashan Hospital has controlled the dynamic evaluation error within ±1.5 points.

[0003] However, the existing model still has obvious bottlenecks and clinical challenges. The existing system faces multi-dimensional limitations: at the data level, the model relies on single modal data, and the F1-score of Chinese medical entity recognition is lower than that of English system; at the inference level, the dynamic signs such as tremor frequency and gait abnormalities are not quantified, and the time series modeling of disease development is not realized; at the clinical transformation level, the participation rate of domestic prospective verification is insufficient, and some hospitals feedback that the existing artificial intelligence increases the operation burden. SUMMARY

[0004] A series of simplified concepts are introduced in the summary section, which will be further described in detail in the specific embodiment section. The summary section of the present application does not mean to attempt to limit the key features and necessary technical features of the claimed technical solutions, nor to determine the protection scope of the claimed technical solutions.

[0005] In order to solve the problems that the existing model still has obvious bottlenecks and clinical challenges, the model relies on single modal data, the dynamic signs such as tremor frequency and gait abnormalities are not quantified at the inference level, the time series modeling of disease development is not realized, and the clinical transformation is difficult, in a first aspect, the present application provides a neural system disease auxiliary evaluation method based on a large model, the method comprising:

[0006] Constructing a structured knowledge graph based on multi-modal heterogeneous data of neural system disease guideline texts, symptom descriptions and examination results;

[0007] Based on the structured knowledge graph, a graph neural network module is embedded based on a large language model to generate a dynamic graph structure inference network model, so as to form a dynamic diagnosis path tree by calculating the relationship weight between each medical node in the graph;

[0008] The explainable diagnosis suggestion is generated by the inference network model according to current symptom description, examination result and historical medical history of the patient.

[0009] Optionally, the multi-modal heterogeneous data based on the neurological disease guideline text, symptom description and examination result is used to construct a structured knowledge graph, and the method comprises the following steps:

[0010] The medical core entities of diseases, symptoms, examination results and laboratory indexes are extracted from the multi-modal heterogeneous data by entity recognition and term standardization technology respectively;

[0011] Based on the medical core entities and logical expressions and rule statements in the medical guideline text, a structured triple set of medical core entities is generated by using relationship extraction technology, so as to organize the structured triple set into the structured knowledge graph, and the structured knowledge graph comprises a disease ontology layer, a symptom-examination association layer and a diagnosis and treatment rule inference layer.

[0012] Optionally, the examination result comprises an image report and laboratory data, and the symptom comprises a medical record description, and the method further comprises:

[0013] By using a BERT-BiLSTM-CRF model, a radiomics feature extraction network and a Vision-Language Transformer model, consistent representation and associated coding of image features and text semantics are realized, so as to align entities in the guideline text, medical record description, image report and laboratory data.

[0014] Optionally, the method further comprises:

[0015] A graph conflict resolution rule engine is constructed based on a timestamp and an evidence level;

[0016] In the case of receiving new clinical data or guideline text changes, a graph incremental update protocol is triggered in real time through a medical event stream.

[0017] Optionally, the explainable diagnosis suggestion is generated by the inference network model according to current symptom description, examination result and historical medical history of the patient, and the method comprises the following steps:

[0018] According to the current symptom description, examination result and historical medical history of the patient, a graph neural network based on an embedded knowledge graph is used for disease identification and path, and in the inference process, a potential confounding factor and misleading association are identified by combining a causal inference mechanism, so as to correct the causal structure and optimize the diagnosis path;

[0019] After completing the causal correction, an interpretable diagnosis suggestion is generated according to the model output result, and a confidence score of the diagnosis suggestion is calculated based on the causal strength, attention consistency and model uncertainty indicators in the reasoning process, and if the confidence of the diagnosis suggestion does not reach a preset threshold, an artificial review mechanism or a multidisciplinary consultation collaborative process is automatically triggered.

[0020] Optionally, the disease recognition and path based on the embedded knowledge graph graph neural network according to the current symptom description, examination result and history of the patient, in the reasoning process, the potential confounding factors and misleading associations are identified in combination with the causal reasoning mechanism, so as to correct the causal structure and optimize the diagnosis path, including:

[0021] According to the current symptom description, examination result and history of the patient, the relevant diagnosis path is activated from the knowledge graph, the probability distribution of each candidate disease is output through the graph neural network, and the examination cost, diagnosis benefit and trauma risk factor are combined to optimize the DDx tree structure;

[0022] The text medical record features, MRI / CT image features and EEG time sequence physiological signals are input into a unified multi-modal attention fusion module, the Transformer and 1D-CNN network are used to extract high-order semantic representations of each mode, and the attention weight is dynamically allocated based on the mode credibility, and a joint diagnosis feature vector is output;

[0023] The joint diagnosis feature vector is input into the causal reasoning module, a medical causal graph is constructed by using a structure learning algorithm and an average causal effect analysis method, potential false associations and confounding factors are identified based on counterfactual data enhancement and adversarial training mechanism, and the causal explanation ability of the diagnosis result is optimized, so as to correct the causal structure and optimize the diagnosis path.

[0024] Optionally, the graph neural network module embedded in the large language model based on the structured knowledge graph is combined to generate a reasoning network model of a dynamic graph structure, including:

[0025] The graph neural network module is inserted into the Transformer backbone network of the large language model;

[0026] The large language model is semantically pre-trained and scene fine-tuned based on general medical corpus and neurology multi-center case data, so as to generate a reasoning network model of a dynamic graph structure.

[0027] In a second aspect, the present application further provides a neural system disease auxiliary evaluation device based on a large model, including:

[0028] A graph construction unit is configured to construct a structured knowledge graph based on multi-modal heterogeneous data of neural system disease guideline texts, symptom descriptions and examination results;

[0029] A model construction unit is configured to generate an inference network model of a dynamic graph structure by combining a large language model embedding graph neural network module based on the structured knowledge graph, so as to form a dynamic diagnosis path tree by calculating the relationship weight between each medical node in the graph.

[0030] An evaluation analysis unit is configured to generate an interpretable diagnosis suggestion by the inference network model according to the current symptom description, examination result and historical medical history of the patient.

[0031] In a third aspect, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor is configured to implement the steps of the neural system disease auxiliary evaluation method based on a large model according to any one of the first aspect when executing the computer program stored in the memory.

[0032] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, and the computer program is executable by a processor to implement the neural system disease auxiliary evaluation method based on a large model according to any one of the first aspect.

[0033] In summary, the neural system disease auxiliary evaluation method based on a large model proposed in the present application constructs a structured knowledge graph based on the multi-modal heterogeneous data of the neural system disease guideline text, symptom description and examination result; generates an inference network model of a dynamic graph structure by combining a large language model embedding graph neural network module based on the structured knowledge graph, so as to form a dynamic diagnosis path tree by calculating the relationship weight between each medical node in the graph; and generates an interpretable diagnosis suggestion by the inference network model according to the current symptom description, examination result and historical medical history of the patient. Thus, a closed-loop process of mapping the diagnosis path of a complex neural system disease from symptom representation to structured path reasoning and to interpretable output is realized, and a plurality of frontier technologies such as language understanding, graph structure modeling, causal reasoning and multi-modal fusion are fused, thereby overcoming the bottlenecks of single diagnosis mode, uninterpretable logic and low conversion rate of traditional AI systems. The method has a wide application prospect in typical neural disease scenarios such as Parkinson's disease dynamic evaluation, epilepsy seizure mechanism modeling and acute cerebral apoplexy rapid screening, and can quickly adapt to different clinical environments through multi-center training samples, and has good universality and expansion ability.

[0034] The neural system disease auxiliary evaluation method based on a large model of the present application, other advantages, objects and features of the present application will be embodied in part through the following description, and will be understood by those skilled in the art through research and practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0035] Various other advantages and benefits will become apparent to those of ordinary skill in the art, upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of preferred embodiments, and are not meant to limit the present description. Moreover, the same reference numerals in different figures represent the same or similar components. In the drawings:

[0036] Figure 1 A large model-based neurological disease auxiliary evaluation method flowchart provided by an embodiment of the present application;

[0037] Figure 2 A large model-based neurological disease auxiliary evaluation device structure diagram provided by an embodiment of the present application;

[0038] Figure 3 A large model-based neurological disease auxiliary evaluation electronic device structure diagram provided by an embodiment of the present application. DETAILED DESCRIPTION

[0039] The terms "first", "second", "third", "fourth" and the like in the description and claims of the present application, and in the above drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order. It is to be understood that the use of the terms so construed can interchange, under appropriate circumstances, to refer to a similarly differentiated object. Moreover, the terms "comprise", "comprises", "comprising", "include", "includes", "including" and the like are used synonymously to encompass a process, a method, a system, a product, or a device that comprises a list of steps or units not necessarily limited to those explicitly listed, but can include other steps or units not expressly listed or inherent to such process, method, product, or device. The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments.

[0040] To solve the problems that the existing model still has obvious bottlenecks and clinical challenges, the model relies on single modal data, the dynamic sign quantification of tremor frequency, gait abnormalities and other dynamic signs is insufficient at the inference level, the disease development time series modeling has not been realized, and the clinical transformation is difficult. Please refer to Figure 1 A large model-based neurological disease auxiliary evaluation method flowchart provided by an embodiment of the present application, which can specifically include steps S110 to S130.

[0041] S110, constructing a structured knowledge graph based on multi-modal heterogeneous data of neurological disease guideline text, symptom description and examination results.

[0042] S120, Based on the structured knowledge graph combined with the large language model embedded graph neural network module, a dynamic graph structure inference network model is generated to form a dynamic diagnostic path tree by calculating the relational weights between each medical node in the graph.

[0043] S130, Based on the patient's current symptom description, examination results, and medical history, an interpretable diagnostic suggestion is generated through the inference network model.

[0044] For example, constructing a structured knowledge graph based on neurological disease guideline texts, symptom descriptions, and examination results involves the following steps: First, collect heterogeneous data from multiple sources, including national clinical pathway guidelines, neurology standard operating procedure documents, PubMed abstracts, chief complaints / present illness / physical examination / auxiliary examination records in electronic medical records (EMR), imaging reports (CT / MRI), and laboratory test results, forming a medical dataset with structural diversity and semantic richness. Second, use a BERT-based named entity recognition model (NER) to extract key medical entities, including disease names, symptom descriptions, imaging features, experimental indicators, drugs, and treatment methods, and uniformly map them to standard terminology systems such as SNOMED-CT and LOINC to ensure cross-institutional compatibility. Furthermore, by combining rule extraction and relation classification models, semantic relationships between different entities are identified, forming sets of triplets such as "disease-symptom-examination-indicator," for example, "Parkinson's disease - has_symptom - resting tremor," "epilepsy - triggered_by - sleep deprivation," and "MRI high signal - suggestions - cerebral infarction." Finally, a graph structure with an ontology layer (concept definition), a relation layer (causal logic), and a knowledge rule layer (diagnosis and treatment path) is constructed, realizing the systematic and structured expression of diagnostic knowledge for neurological diseases. This knowledge graph provides rich and verifiable graph structure priors for subsequent graph neural network reasoning, possesses cross-modal knowledge integration and dynamic update capabilities, and can be used to quickly locate multi-hop reasoning paths between medical entities, improving the completeness and traceability of diagnostic logic.

[0045] For example, based on the structured knowledge graph, a large language model inference network incorporating graph neural network modules is constructed to generate a diagnostic path tree with a dynamic graph structure. This step builds a semantic backbone network based on the Deepseek large language model and introduces graph neural network modules, such as graph convolutional networks (GCN), gated graph neural networks (GGNN), or graph attention mechanisms (GAT), into the middle structure of its Transformer, such as layer 12 or 24, to achieve explicit modeling of graph structure knowledge. The training phase employs a pre-training-fine-tuning strategy: first, the language model is pre-trained using large-scale medical guidelines and clinical summary corpora in the neuroscience field to improve the model's understanding of proper nouns and clinical context; then, multi-center EMR structured graph data, including patient symptom-examination-disease paths, is used as supervision signals to perform graph embedding joint optimization and node relationship learning. By introducing graph neural structures, the model can perceive multi-hop logical paths between nodes in the graph, calculate semantic and structural correlations between entities, and achieve causal path generation that "starts from symptoms and automatically infers possible examination items and corresponding diagnoses." During the inference phase, the model takes the patient's current medical data as input and dynamically activates associated nodes on the graph. It then calculates node importance and path strength using a graph neural network, outputting a diagnosis-oriented dynamic diagnostic path tree (DynamicDiagnosticTree). This path tree not only reflects the clinical logical chain from symptoms to examination to diagnosis but also automatically adjusts its branch structure based on different inputs. It possesses the ability to adapt to the phenotypic differences of various neurological diseases, significantly improving the model's interpretability, personalization, and clinical applicability.

[0046] For example, the system acquires the patient's current symptom description, examination results, and medical history information, and uses the aforementioned dynamic graph reasoning model to output interpretable diagnostic suggestions. The system receives user input such as a chief complaint (sudden dizziness with vomiting), auxiliary examination data (e.g., a negative head CT scan), physical signs (e.g., right nystagmus), and medical history (e.g., a history of hypertension). This information is uniformly converted into structured input that the model can process through a semantic encoding module. The model activates nodes related to the above input in a structured knowledge graph and performs multi-hop information propagation along the symptom-examination-disease chain through a graph neural module. It generates multiple possible diagnostic paths by combining the edge weights between nodes and the semantic path strength, and outputs a confidence score for each path based on pre-trained parameters. Building upon this foundation, the system further integrates the path interpretation results to generate structured decision support suggestions, including disease predictions such as a 60% probability of vestibular neuritis, a 25% probability of BPPV, and a 15% probability of brainstem stroke; suggested examinations such as supplementing with MRI-DWI scans and HINTSbedside examinations; clinical management suggestions such as immediately assessing the need for thrombolysis and simultaneously rechecking blood pressure; and precautions such as dynamically monitoring symptom changes after ruling out peripheral vertigo. If the differences between diagnostic pathways are high or the confidence level is below a set threshold (less than 60%), the system will automatically trigger a manual second review / MDT multidisciplinary collaboration mechanism to ensure diagnostic safety. Through this approach, the system achieves highly personalized, traceable, and interpretable auxiliary diagnostic output, not only improving doctors' diagnostic efficiency but also reducing the risk of misdiagnosis and enhancing patient trust.

[0047] In summary, the large-model-based auxiliary assessment method for neurological diseases provided in this application constructs a structured knowledge graph based on multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results. Based on this structured knowledge graph, a graph neural network module embedded in a large language model is used to generate a dynamic graph-structured inference network model. This model calculates the relational weights between medical nodes in the graph to form a dynamic diagnostic path tree. Based on the patient's current symptom description, examination results, and medical history, interpretable diagnostic suggestions are generated through the inference network model. This achieves a closed-loop process for diagnosing complex neurological diseases, mapping symptom representation to structured path inference and then to interpretable output. It integrates multiple cutting-edge technologies such as language understanding, graph structure modeling, causal reasoning, and multimodal fusion, overcoming the bottlenecks of traditional AI systems, such as single diagnostic modality, uninterpretable logic, and low conversion rate. This method has broad application prospects in typical neurological disease scenarios such as dynamic assessment of Parkinson's disease, modeling of epileptic seizure mechanisms, and rapid screening for acute stroke. Furthermore, it can quickly adapt to different clinical environments through multi-center training samples, demonstrating good versatility and scalability.

[0048] In some examples, the construction of a structured knowledge graph based on multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results includes:

[0049] Medical core entities representing diseases, symptoms, examination results, and laboratory indicators are extracted from the multimodal heterogeneous data using entity recognition and terminology standardization techniques, respectively.

[0050] Based on the logical expressions and rule statements in the medical core entities and medical guide texts, a set of structured triples of medical core entities is generated using relation extraction technology. The set of structured triples is then organized into the structured knowledge graph, which includes a disease ontology layer, a symptom-examination association layer, and a diagnosis and treatment rule reasoning layer.

[0051] For example, the process of constructing a structured knowledge graph based on multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results first includes the identification and standardization of core medical entities. Specifically, by applying Named Entity Recognition (NER) models based on BERT or RoBERTa architectures to multiple modalities in the input data source, such as PDF format clinical guidelines, structured EMR texts, unstructured chief complaint descriptions, imaging diagnostic reports, and laboratory test results, core medical entities such as disease names, typical symptoms, examination items, imaging findings, and laboratory test indicators are extracted. Simultaneously, terminology standardization is performed using word vector similarity matching, rule mapping tables, or medical ontology libraries such as SNOMED-CT, ICD-11, and LOINC, so that different expressions such as "dizziness," "vertigo," and "nausea" are uniformly encoded into standard entity nodes such as Symptom_00034_vertigo, to ensure semantic consistency across institutions. Secondly, for the standardized entity set mentioned above, we further analyze the large number of rule statements and logical expressions in the disease guideline text, such as "If A is accompanied by B, then it is recommended to check C" or "If X is present but Y is absent, Z should be considered". We use a rule extraction model based on syntax tree structure and relation classification network to mine the structured associations between entities, thereby generating a set of medical triples in the form of <entity A, relation R, entity B>, such as <Parkinson's disease, has_symptom, resting tremor>, <sudden headache, suggestions, subarachnoid hemorrhage>, <stroke, requires, cranial CTA>. Then, the set of triples is organized into a three-layer knowledge graph based on semantics and hierarchical logic: the first layer is the disease ontology layer, which defines the standard coding, definition, classification, and staging nodes of the disease; the second layer is the symptom-examination association layer, which reflects the causal or sequential relationship between symptoms, signs, and laboratory or imaging examination results, and is used to assist in diagnostic pathway planning; the third layer is the treatment rule reasoning layer, which explicitly models the conditional triggering logic and treatment suggestions according to guideline rules, and is used to support automatic consultation planning or treatment suggestion generation. Through this knowledge graph construction process, fragmented knowledge in heterogeneous data is transformed into a unified, structured, and computable entity network, which significantly enhances the subsequent graph neural network's ability to model clinical pathways, disease evolution, and multi-causal relationships, and provides a complete knowledge foundation and logical priors for large-scale model reasoning of diagnostic path trees.

[0052] For example, for cross-modal entity linking, a neuromedical entity recognizer can be built based on the BERT-BiLSTM-CRF model to uniformly extract disease entities such as "Parkinson's disease" from guideline texts, symptom descriptions such as "resting tremor" from EMR, and feature terms such as "thinning of the substantia nigra compacta" from imaging reports, and map them to the SNOMED-CT standard terminology set. This achieves unified entity extraction and standardized mapping of heterogeneous data. The specific steps are as follows:

[0053] Entity extraction pipeline: Text data such as guides and EMRs use the BERT-BiLSTM-CRF model, where the BERT layers are initialized with `bert-base-chinese` weights, the BiLSTM hidden layers have a dimension of 768, and the CRF layers learn the transition matrix. The loss function used is negative log-likelihood.

[0054]

[0055] Where, x i Given an input sequence such as a symptom description from an EMR, y i Entity labels such as diseases and symptoms are provided. Image data is combined with radiomics feature extractors, such as the PyRadiomics library, and a pre-trained ResNet-50 model, to output region feature maps. Laboratory metrics are parsed from structured tables using a regularization engine, such as the Python `re` module and the LOINC encoding mapping table.

[0056] Entity standardization: All extracted entities are mapped using the SNOMED-CT terminology set. The mapping algorithm employs vector retrieval based on cosine similarity.

[0057]

[0058] Among them, e raw For the original entity, e std 'v' represents the SNOMED-CT standard entity, and 'v' represents the FastText word vector. Additionally, a conflict resolution rule is added: when multiple candidate entities have a similarity > 0.85, the entity with the highest evidence level is selected, with the evidence level quantified by PubMed citation frequency.

[0059] Image-Text Semantic Bridging: The Vision-LanguageTransformer model is used to associate MRI / CT image features, such as calcifications in the basal ganglia, with text descriptions to generate encodeable structured feature vectors. The specific process is as follows:

[0060] Model architecture:

[0061] The image branch takes an MRI / CTDICOM file as input and extracts visual features using a ViT-Base model, such as a 16x16 patch size. The text branch takes the image report text as input and extracts text features using a BERT model. Finally, image-text attention weights are calculated through a cross-modal attention layer.

[0062]

[0063] Where Q = TW q $,$K=VW k $,$V=VWv (W represents the learnable weights).

[0064] Generative structured features:

[0065] Output encoding vector z = MLP([V att ;T att The MLP is a two-layer fully connected layer with an output dimension of 256. A dynamic update mechanism is also incorporated; when new image data is input, the ViLT model is fine-tuned through online learning with a learning rate set to 10. -5 Use a cosine annealing scheduler.

[0066] For dynamic knowledge fusion, a conflict resolution rule engine can be used. When new guidelines conflict with old data, such as updates to epilepsy medication dosages, knowledge weight adjustments can be automatically triggered based on timestamps and evidence levels (formula: W_new=α·PubDate+β·Evidence_Level). An incremental update protocol for the knowledge graph can be designed to integrate new case data in real time through MedicalEventStreamProcessing, maintaining the timeliness of the knowledge graph.

[0067] For example, for disease guideline texts, PDF parsing combined with NLP entity relation extraction can be used to output RDF triples; for MRI / CT images, 3DResNet can be used to segment lesions and combined with radiomics feature extraction to output feature vectors in JSON format; for laboratory indicators, LIS system can be used for interface and unit standardization to output structured tables.

[0068] For example, for knowledge fusion and graph construction, multi-graph neural networks (Multi-GNNs) can be used to fuse heterogeneous features:

[0069] m represents the modal type. It is the modal intra-adjacency matrix.

[0070] For example, clinical pathway derivation can be implemented based on the OWL inference engine, such as: headache combined with papilledema presuming suspected increased intracranial pressure and recommending cranial MRI. Accuracy was validated using 100,000 real medical records from a tertiary hospital (F1-score reached 92.7%). Taking the generation of acute stroke treatment pathways as an example, the input is: sudden onset of hemiplegia (symptom text) and negative CT scan (image data). The atlas inference trigger rule is: negative CT combined with focal neurological deficit → suspected ischemic stroke. The atlas inference output suggestion is: urgently complete DWI-MRI, exclude contraindications for thrombolysis within the time window. For dynamic updates, if new guidelines recommend extending the thrombectomy time window to 24 hours, the system automatically updates the treatment rule base.

[0071] In some examples, the examination results include imaging reports and laboratory data, the symptoms include medical record descriptions, and the method further includes:

[0072] By employing the BERT-BiLSTM-CRF model, radiomics feature extraction network, and Vision-LanguageTransformer model, consistent representation and associative encoding of image features and text semantics are achieved to align entities in guideline texts, medical record descriptions, image reports, and laboratory data.

[0073] For example, the examination results include imaging reports and laboratory data, and the symptoms include natural language descriptions such as chief complaints, present medical history, and physical examination records in medical record texts. To achieve unified expression and alignment of medical entities across different modalities in a structured knowledge graph, the method further includes a semantic alignment strategy based on joint text and image modeling. Specifically, firstly, for text modal data such as medical record descriptions, guideline texts, and imaging reports, a joint sequence labeling model based on BERT-BiLSTM-CRF is used for named entity recognition (NER) to extract medical terms such as diseases, symptoms, anatomical structures, imaging features, and quantitative indicators, and these terms are standardized and encoded using dictionary mapping and semantic vectors. Secondly, for raw imaging data such as CT and MRI, a radiomics feature extraction network based on CNN or Transformers is used to extract image features of semantically sensitive regions such as shape, density, texture, and boundary discontinuities. Simultaneously, under the condition of annotated samples, regional attention maps or heatmaps (e.g., high-signal plaques, low-density lesions, etc.) corresponding to the imaging features are generated. Furthermore, a cross-modal semantic encoding module for images and text is constructed using Vision-LanguageTransformers (such as ViLT and Med-VLP). This module aligns image region features with entity representations in the text through a shared multimodal attention mechanism. In the multimodal vector space, high signal intensity in the right frontal cortex, high signal intensity lesions on frontal lobe MRI, and high signal intensity in the right frontal lobe region are uniformly mapped to the same knowledge graph node entity, such as "Imaging_Feature_012_frontal lobe high signal intensity". In this way, different representations of the same clinical fact in images, text, and experimental data are integrated, forming a more consistent and context-rich set of graph entities. This significantly improves the ability of subsequent inference models to integrate cross-modal medical evidence and allows the knowledge graph to serve as a unified input interface for graph neural networks, enabling higher-precision causal path reasoning and diagnostic assistance functions.

[0074] For example, the multimodal medical named entity recognition system covers five core medical entities and designs different structured extraction strategies based on data modalities to support knowledge graph construction. For disease guideline texts, an entity recognition model based on BERT-BiLSTM-CRF is used, combined with the SNOMED-CT medical terminology database to achieve terminology standardization, which can extract structured disease entities such as "vestibular migraine" and "TIA". For symptom descriptions in structured or free text electronic medical records (EMR), a RoBERTa model with enhanced medical context is used to focus on solving semantic disambiguation problems such as "dizziness" vs. "vertigo", which can identify symptom entities such as "rotational vertigo" and "pulsating headache". In terms of medical imaging data, a joint learning model is constructed by combining a visual Transformer network and a radiology report parsing module to jointly model image regions and text labels, which can extract feature entities such as "DWI high signal" and "pontine 'cross sign'". For laboratory test indicators, the system uses the LONIC standard engine and regular rule-guided approach to perform structured mapping of test results, identifying indicator entities such as "CSF oligoclonal band positive" and "NMDA receptor antibody". In terms of treatment rule extraction, a structured rule parser transforms IF-THEN logical rules from medical guidelines into triple structures, supporting the automatic extraction of treatment conditions and triggering mechanisms. For example, "atrial fibrillation patients must first check anticoagulation status" is transformed into a condition-action rule entity. Through this multimodal NER mechanism, the system can achieve high-quality extraction of core entities from raw medical data, providing a highly consistent and comprehensive knowledge graph node foundation for subsequent knowledge graph construction, path reasoning, and causal assessment.

[0075] For example, the cross-modal relation extraction (RE) module is used to identify semantic and inference relationships between extracted entities from multimodal medical data, and to construct structured triples to support the construction of knowledge graph edges. To this end, this invention first defines a standard medical relation type matrix, including but not limited to the following five core relations: ① Disease-symptom relation (has_symptom), used to describe the symptom characteristics of a typical manifestation of a disease, such as <epilepsy, has_symptom, tonic-clonic seizure>; ② Symptom-examination recommendation relation (recommend_exam), representing examination suggestions guided by a certain symptom, such as <sudden headache, recommend_exam, cranial CTA>; ③ Examination-diagnosis confirmation relation (confirm_diagnosis), used to express supporting evidence for diagnosis from imaging / laboratory examinations, such as <D WI high signal, confirm diagnosis, acute cerebral infarction; ④ Disease-differential diagnosis relationship (differential diagnosis), used to indicate the clinical semantics that need to be differentiated between different diseases, such as <Parkinson's disease, differential diagnosis, MSA>; ⑤ Indicator-clinical interpretation (clinical interpretation), used to establish an interpretive mapping between laboratory / imaging indicators and potential pathological states, such as <abnormal decrease in serum ceruloplasmin, clinical interpretation, Wilson's disease>.

[0076] In its implementation, the RE module employs a joint extraction architecture based on PubMedBERT. The model input is a medical text fragment, and it simultaneously outputs entity boundaries and relation labels, thus forming structured triples. For example, the text "Patients with acute vertigo and nystagmus require HINTS examination" can generate the triple <acute vertigo, require_exam, HINTS examination>. To achieve collaborative recognition of the relationship between image data and text, an "image-text alignment" mechanism is also constructed, using the ViTFeatureReport module to achieve consistent alignment of image region semantics and text terms. For example, the report mentioning "high T1 signal in the right thalamus" can be transformed into the triple <Wilson's disease, imaging_feature, high T1 signal in the right thalamus> through region localization and embedding encoding. Furthermore, with the support of a structured rule base, the RE module supports the parsing and transformation of complex rule chains. For example, "new-onset headache combined with papilledema THEN recommends cranial MRI" can be automatically decomposed into multiple rule triples such as <new-onset headache, urgent_exam, cranial MRI>, thereby achieving the goal of dynamically extracting from medical knowledge texts and case corpora. Through the above-mentioned cross-modal RE mechanism, the system can efficiently construct "entity-relationship-entity" knowledge units based on the NER output, forming a complete set of structured triples. This supports the expression of multi-level relational edges in the knowledge graph, providing accurate graph structure input for downstream graph neural network reasoning, and effectively improving the logical rigor and causal explanation ability of the diagnostic path.

[0077] In some examples, it also includes:

[0078] A graph conflict resolution rule engine is built based on timestamps and evidence levels;

[0079] Upon receiving new clinical data or changes to guideline text, the atlas incremental update protocol is triggered in real time via medical event stream.

[0080] It is understandable that the process of constructing a structured knowledge graph based on multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results further includes dynamic updating and conflict resolution mechanisms for the knowledge graph, in order to adapt to the situational needs of clinical guideline evolution and the continuous introduction of new case data.

[0081] For example, addressing the temporal heterogeneity and multi-source contradictions inherent in actual clinical practice, the system constructs a graph conflict resolution rule engine based on timestamp information between entities, such as examination time, literature publication time, guideline version number, and evidence level stratification, such as A / B / C level recommendation strength. This engine dynamically determines the current valid representation of a path in the graph when there are differences in treatment paths between the same entities, such as different versions of guidelines recommending different examination methods. This is achieved through a dual weighting mechanism of time priority and evidence level. For instance, if there are two conflicting paths, <TIA, requires, head CT> and <TIA, requires, MRI-DWI>, the latter will receive a higher path weight in the graph after the release of a new guideline. Furthermore, to ensure the graph has real-time update capabilities in clinical deployment scenarios, the method introduces a HealthcareEventStream mechanism into the knowledge graph platform. When new structured case data, newly released guideline documents, or research paper abstracts are received, an incremental graph update protocol is automatically triggered to identify newly added entities, edges, or relationship correction items. These are then structurally merged through graph insertion, updates, or soft replacements, avoiding overall reconstruction. Through this dynamic maintenance mechanism, the atlas system can continuously absorb the differences in real-world diagnostic and treatment pathways from multi-center clinical practice, while maintaining logical consistency with authoritative guidelines. This allows for the construction of a clinical-grade neurological disease knowledge graph with "version evolution capability, conflict robustness, and real-time updability," further enhancing its adaptability and accuracy in downstream large-scale model diagnostic reasoning.

[0082] In some examples, the generation of interpretable diagnostic suggestions through the inference network model based on the patient's current symptom description, examination results, and medical history includes:

[0083] Based on the patient's current symptom description, examination results, and medical history, disease identification and path determination are performed using a graph neural network embedded with a knowledge graph. During the reasoning process, a causal reasoning mechanism is combined to identify potential confounding factors and misleading associations in order to correct the causal structure and optimize the diagnostic path.

[0084] After completing the causal correction, interpretable diagnostic suggestions are generated based on the model output results. The confidence score of the diagnostic suggestions is calculated based on the causal strength, attention consistency and model uncertainty indicators in the reasoning process. If the confidence score of the diagnostic suggestions does not reach the preset threshold, the manual review mechanism or multidisciplinary consultation and collaboration process is automatically triggered.

[0085] Understandably, the process of generating interpretable diagnostic suggestions based on the patient's current symptom description, examination results, and medical history through the inference network model first includes structural mapping and disease path identification of the input patient information based on knowledge graph embedding, such as GGNN (Gated Graph Neural Network). The system maps the user's main complaint (e.g., limb numbness with slurred speech), auxiliary examination items (e.g., head MRI showing acute infarction in the right frontal lobe), and medical history (e.g., a 5-year history of hypertension) to corresponding nodes in the structured knowledge graph, and activates relevant symptom, examination, indicator, and disease nodes in the graph through the graph neural network propagation mechanism. To enhance the model's clinical causal explanation capability, a causal reasoning mechanism is further introduced into the graph structure reasoning process. A do-calculus rule framework and a structural causal model (SCM) embedding are used to identify potential confounding variables (e.g., comorbidities, missing examinations, misaligned time sequences, etc.) and misleadingly strong associations (e.g., the relationship between "hyponatremia-epilepsy" and "tumor-induced hyponatremia"), dynamically correcting the causal structure in the graph and optimizing path weight allocation. After causal correction, the model generates structured diagnostic suggestions, including disease prediction (multi-class output probability vectors), recommended examination items (based on the attention focus area of ​​path nodes), preliminary intervention suggestions, and precautions. Furthermore, the model introduces a multi-index confidence assessment module at the output layer to comprehensively evaluate the average causal contribution of each causal path during the inference process, the consistency of the attention mechanism distribution (entropy / dispersal), and output layer uncertainty (such as the distribution variance after Monte Carlo dropout), calculating the confidence score of the current diagnostic suggestion (e.g., a continuous value between 0 and 1). When the confidence score of the diagnostic suggestion is lower than the system's preset confidence threshold (e.g., 0.65), or when there are multiple alternative paths with insignificant differences in confidence scores (e.g., the difference between the top-2 paths < 0.05), the system automatically triggers a manual review mechanism or an MDT (Multidisciplinary Team) collaborative prompting process, pushing the intelligent preliminary screening results and causal chain interpretation graph to the attending neurologist or related specialists to assist in the final decision-making process. Through the above mechanism, the system not only improves the interpretability, path transparency and clinical safety of the reasoning results, but also realizes the human-machine collaborative closed-loop control capability of artificial intelligence in the process of assisted diagnosis of neurological diseases.

[0086] For example, to support unified modeling and causal reasoning of multimodal clinical information of patients with neurological diseases, the system constructs a joint diagnostic network comprising a dynamic graph structure reasoning layer and a multimodal attention fusion module. First, after the structured knowledge graph is constructed, its medical entities (such as diseases, symptoms, examinations, etc.) and their corresponding edges are organized into a heterogeneous medical graph structure, which serves as the input to the dynamic graph neural reasoning module. Each node represents a medical entity, and each edge carries weights and semantic labels, such as <partial>.

[0087] The edge weight for <headache, has_symptom, shaking headache> can be 0.92, and the edge for <shaking headache, recommend_exam, cranial MRI> carries the evidence label "Guideline Level A". To perform inference propagation between nodes, a Gated Graph Neural Network (GGNN) is used as the core model, and the state vector of each node is iteratively updated using the following formula:

[0088]

[0089] Among them, h v Represents the current hidden state of a node (e.g., a disease), where N(v) is its adjacent entities, and W... edgeThe propagation weight matrix represents the edge type. This mechanism enables propagation from symptom nodes along the graph structure to relevant examination and diagnostic nodes, outputting the probability distribution P(Di|S1,S2,…) of each candidate disease as the basis for subsequent diagnosis. For example, given the input "middle-aged woman + shaking headache + photophobia and phonophobia", the system can automatically infer the reasoning path <shaking headache → migraine → cranial MRI>, and activate relevant examination nodes on the graph. If the MRI results support the diagnosis, it further matches the treatment path <triptan treatment>, achieving an automatic reasoning closed loop from symptoms to treatment. Furthermore, to enhance the fusion capability of multimodal diagnostic data, a multimodal attention mechanism is introduced into the diagnostic path reasoning. This mechanism calculates the attention scores between different modalities using a Vision-LanguageTransformer, specifically for joint modeling of medical record text, MRI / CT images, and EEG sequence signals. The system first extracts the dynamic spectral features of EEG epileptiform waves using 1D-CNN. Then, it inputs these features into a unified multimodal attention module to calculate the cross-attention matrix Attention(Q,K,V) to identify high-confidence modalities. During inference, the system dynamically adjusts the attention weights based on data type. For example, the weight is 0.6 when data_type = "text medical record" and is "emergency room record," and can reach 0.8 when data_type = "EEG" and is "seizure phase." This achieves contextual adaptation in modality selection and enhances the collaborative diagnostic capabilities between different modalities. This mechanism significantly improves the model's ability to locate epileptic foci using combined EEG and MRI findings. In real-world testing, the model's AUC improved by approximately 0.04, and the diagnostic accuracy increased by over 7%.

[0090] In some examples, the disease identification and path determination are based on a graph neural network embedded with a knowledge graph, taking into account the patient's current symptom description, examination results, and medical history. During the reasoning process, a causal reasoning mechanism is incorporated to identify potential confounding factors and misleading associations, thereby correcting the causal structure and optimizing the diagnostic path. This includes:

[0091] Based on the patient's current symptom description, examination results, and medical history, relevant diagnostic paths are activated from the knowledge graph. The probability distribution of each candidate disease is output through a graph neural network, and the DDx tree structure is optimized by combining examination costs, diagnostic benefits, and trauma risk factors.

[0092] The multimodal attention fusion module integrates text medical record features, MRI / CT image features, and EEG time-series physiological signals. It uses Transformer and 1D-CNN networks to extract high-order semantic representations of each modality and dynamically allocates attention weights based on modality credibility to output a joint diagnostic feature vector.

[0093] The joint diagnostic feature vector is input into the causal reasoning module, and a medical causal graph is constructed using a structure learning algorithm and an average causal effect analysis method. Based on counterfactual data augmentation and adversarial training mechanisms, potential spurious associations and confounding factors are identified to optimize the causal explanatory power of the diagnostic results, thereby correcting the causal structure and optimizing the diagnostic path.

[0094] For example, the process of disease identification and diagnostic path optimization based on a graph neural network embedded in a knowledge graph, based on the patient's current symptom description, examination results, and medical history, first includes a candidate disease reasoning mechanism based on the graph neural network. After receiving structured or unstructured information such as the patient's chief complaint text, imaging report, and past medical history, the system first performs entity matching in the knowledge graph, activates relevant symptom, examination, indicator, and disease nodes, and propagates contextual features in the graph neural network embedded in the graph structure. Through node aggregation and path awareness mechanisms, it outputs the predicted probability distribution of each candidate disease. At the same time, combining the examination costs (such as cost levels), diagnostic benefits (such as early disease detection benefit scores), and trauma risk factors (such as whether it is an invasive operation) already modeled in the knowledge graph, a cost function evaluation matrix is ​​constructed. The DDxTree (Differential Diagnosis Tree) structure is pruned and optimized, thereby outputting an intelligent diagnostic path map that is more in line with clinical practice. Secondly, to enable multi-source data to collaboratively empower diagnostic decisions, the system inputs multimodal data from textual medical records such as chief complaint / present illness history, structured imaging MRI / CT, ​​and electrophysiological time-series signals (EEG) into a unified multimodal attention fusion module. This module uses a Transformer-based attention mechanism to extract contextual dependencies in the medical record text, while simultaneously using 1D-CNN to capture local patterns in the time-series signals. Furthermore, it dynamically adjusts the weight contributions of each modality in the decision-making process through a trainable modality confidence gating unit, ultimately fusing them into a joint diagnostic feature vector. This joint feature vector is then input into the causal inference module, where structural learning algorithms such as NOTEARS and CAM automatically learn the causal relationship graph structure from the features. Finally, it employs ATE (Average Treatment Effect) and counterfactual inference mechanisms to construct an interpretable medical causal graph. To further enhance the robustness of the causal structure to spurious strong associations and data bias, this module combines counterfactual sample generation and adversarial training mechanisms to identify, correct, or reduce the influence weights of suspected confounding factors and spurious causal paths, thereby outputting an optimized causal diagnostic path. The final system outputs diagnostic results that are both accurate and causally interpretable, improving the interpretability of multimodal data collaborative diagnosis and enhancing the model's generalization ability and clinical reliability in complex cases.

[0095] According to some embodiments, the step of generating a dynamic graph structured inference network model based on the structured knowledge graph combined with a large language model embedded in a graph neural network module includes:

[0096] Insert a graph neural network module into the Transformer backbone network of the large language model;

[0097] Semantic pre-training and scenario fine-tuning of a large language model are performed based on general medical corpus and multi-center neurological case data to generate a dynamic graph structured inference network model.

[0098] For example, the process of generating a dynamic graph structure inference network model based on a structured knowledge graph combined with a large language model and embedded graph neural network modules first includes modifying the structure of the large language model, that is, inserting graph neural network modules into its Transformer backbone network. Specifically, a large language model with open parameters and embeddable structures, such as Deepseek, is selected, and graph neural network modules (such as GatedGraphNeuralNetwork, GraphConvolutionalNetwork, or RelationalGNN) are inserted between multiple layers in the middle of the Transformer backbone, usually between the Mth and Nth layers, to allow the model to perceive the graph structure semantics in the knowledge graph while maintaining natural language inference capabilities. This insertion method can adopt a parallel (residual) or injection strategy, synchronously processing the adjacency relationships and linguistic context of nodes in the graph in each forward propagation cycle, so that the output latent variable representation has both semantic consistency and causal path constraints in the graph structure.

[0099] For example, to adapt the inference network model to tasks in the neurological disease domain, the large language model requires knowledge enhancement in two stages after structural modification: semantic pre-training and scenario fine-tuning. First, the semantic pre-training stage uses general medical corpora such as PubMed summaries, open guidelines, and NMPA review materials for masked language modeling and relation prediction tasks to supplement the model's understanding of basic medical terminology, standard diagnostic procedures, and cross-modal expressions. Second, the scenario fine-tuning stage introduces multi-center neurological case data, including structured data tables, medical records, image labels, and diagnostic conclusions. Task-oriented supervised fine-tuning strategies, such as etiology classification, path recognition, and examination suggestion generation, are employed to further improve the model's structural awareness and professional generalization performance in neurological disease inference tasks. During this training process, the parameters of the graph neural network are jointly optimized with the language model to ensure that their representation capabilities of knowledge graph entities remain synergistic, and the final inference model outputs a semantic-entity joint embedding representation with a dynamic graph structure. The resulting dynamic graph reasoning network model can not only activate personalized graph subgraphs and perform efficient path reasoning based on the input of different patients, but also output structured decision suggestions through a language interface. This balances the language understanding ability of large models with the medical graph reasoning ability, significantly improving the professionalism, transparency and clinical adaptability of the diagnostic system.

[0100] According to some embodiments, during the data processing phase, the system collects structured and unstructured data resources from multiple hospitals, including medical record texts, image reports, laboratory indicators, EEG signals, and treatment guidelines. To ensure data security and privacy compliance, a differential privacy mechanism is used to perform layered desensitization of sensitive patient information, protecting identity attributes while preserving semantic structure. Subsequently, based on the constructed medical ontology alignment framework, terms from various data sources are mapped to unified standards, such as SNOMED-CT, LOINC, and RadLex, resolving the issue of inconsistent data encoding across multiple hospitals. The standardized multimodal data is used as input into the multimodal NER and RE modules, where core triples such as "disease-symptom-examination-indicator" are extracted and organized into a structured knowledge graph using a rule engine and graph embedding algorithm. This graph includes a disease ontology layer, a symptom-examination association layer, and a treatment rule layer. This effectively solves the problem of clinical data heterogeneity and improves data accuracy and semantic consistency during the graph reasoning phase.

[0101] Next, in the system deployment phase, this invention constructs a dynamic graph structure inference network based on large language models such as Deepseek-Med embedded graph neural network modules, and deploys it in a physician collaboration platform to support real-time clinical interaction. The platform integrates modules for visualized backtracking of treatment pathways, identification and analysis of misdiagnosis pathways, and intelligent DICOM image parsing and 3D spatial positioning. It is compatible with mainstream medical system standards such as HL7 / FHIR and DICOM WebAPI, and can seamlessly connect to hospital HIS / PACS systems to achieve real-time synchronization and processing of diagnostic data. During inference, the system uses the patient's current symptom description, examination results, and medical history as input, activates relevant paths in the knowledge graph, and uses the GatedGraphNeuralNetwork model to achieve causal graph propagation and diagnostic node probability output. Simultaneously, combined with a multimodal attention mechanism, it dynamically weights and fuses MRI images, EEG signals, and text medical record features, adaptively adjusting the diagnostic contribution weights based on modal confidence to generate interpretable diagnostic suggestions (including candidate diseases, recommended examinations, preliminary treatment suggestions, etc.). If the system identifies potentially misleading associations or inference confidence levels below a threshold in the diagnostic pathway, it will automatically trigger the MDT collaboration mechanism to ensure clinical safety. Through modular deployment and intelligent interface integration, closed-loop control is achieved from data access and inference modeling to diagnostic feedback, ensuring high integrability and practicality of the system in real hospital environments.

[0102] Finally, in the validation phase, this invention employs a multi-center, three-blind comparative experimental design to construct a module for analyzing the discrepancies between the AI ​​system's output and the diagnostic conclusions of neurological experts. In typical diseases (such as stroke, epilepsy, and Parkinson's disease), instances of discrepancies between AI diagnostic suggestions and actual outcomes are collected, and the root causes of errors are traced. The robustness of the model under confounding interference is evaluated by combining graph node contribution, modal attention heatmaps, and causal path structure visualization analysis. The system automatically records the critical paths and uncertainties in each inference process and triggers a knowledge graph reconstruction mechanism, modal weight adjustment, and causal structure relearning when deviations occur, achieving closed-loop self-updating of the inference network. Statistical results from multi-center experiments show that this system significantly improves diagnostic accuracy and clinical adoption rates in early stroke classification, epileptic focus localization, and differentiation of chronic headache etiologies. The misdiagnosis rate for some diseases decreases by more than 15%, validating the system's high performance and sustainable optimization capabilities in real-world clinical environments.

[0103] In summary, by organically integrating privacy-preserving data processing procedures, graph neural network inference engines, and cross-modal causal reasoning mechanisms, and combining them with end-to-end deployment and closed-loop verification strategies, this invention constructs an end-to-end, dynamically interpretable, and clinically practical auxiliary diagnosis and treatment system for neurological diseases, breaking through the technical bottlenecks of data silos, single-modal dependence, and weak clinical translatability in existing artificial intelligence systems.

[0104] Please see Figure 2 One embodiment of the large-model-based auxiliary assessment device for neurological diseases in this application may include:

[0105] Graph construction unit 21 is used to construct a structured knowledge graph based on multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results;

[0106] The model building unit 22 is used to generate a dynamic graph structured inference network model based on the structured knowledge graph combined with the large language model embedded graph neural network module, so as to form a dynamic diagnostic path tree by calculating the relation weights between each medical node in the graph.

[0107] The assessment and analysis unit 23 is used to generate interpretable diagnostic suggestions based on the patient's current symptom description, examination results, and medical history through the inference network model.

[0108] In summary, the large-model-based auxiliary assessment device for neurological diseases provided in this application constructs a structured knowledge graph based on multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results. Based on this structured knowledge graph, a graph neural network module embedded with a large language model is used to generate a dynamic graph-structured inference network model. This model calculates the relational weights between medical nodes in the graph to form a dynamic diagnostic path tree. Based on the patient's current symptom description, examination results, and medical history, interpretable diagnostic suggestions are generated through the inference network model. This achieves a closed-loop process for diagnosing complex neurological diseases, mapping symptom representation to structured path reasoning and then to interpretable output. It integrates multiple cutting-edge technologies such as language understanding, graph structure modeling, causal reasoning, and multimodal fusion, overcoming the bottlenecks of traditional AI systems, such as single diagnostic modality, uninterpretable logic, and low conversion rate. This method has broad application prospects in typical neurological disease scenarios such as dynamic assessment of Parkinson's disease, modeling of epileptic seizure mechanisms, and rapid screening for acute stroke. Furthermore, it can quickly adapt to different clinical environments through multi-center training samples, demonstrating good versatility and scalability.

[0109] like Figure 3 As shown, this application embodiment also provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 320 and executable on the processor. When the processor 320 executes the computer program 311, it implements the steps of any of the above-described methods for auxiliary assessment of neurological diseases based on a large model.

[0110] Since the electronic device described in this embodiment is the device used to implement the large-model-based auxiliary assessment device for neurological diseases in this application embodiment, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in this application embodiment. Therefore, how the electronic device implements the method in this application embodiment will not be described in detail here. Any device used by those skilled in the art to implement the method in this application embodiment is within the scope of protection of this application.

[0111] In practical implementation, when the computer program 311 is executed by the processor, it can achieve the following: Figure 1 Any of the corresponding implementation methods in the embodiments.

[0112] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] This application also provides a computer program product, which includes computer software instructions that, when executed on a processing device, cause the processing device to perform actions such as... Figure 1 The corresponding embodiment describes the process for auxiliary assessment of neurological diseases based on a large model.

[0118] A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0119] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0120] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0121] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0122] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0123] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for auxiliary assessment of neurological diseases based on a large model, characterized in that, include: A structured knowledge graph was constructed based on multimodal heterogeneous data of guideline texts, symptom descriptions, and examination results for neurological diseases. Based on the structured knowledge graph and the large language model embedded in the graph neural network module, a dynamic graph structure inference network model is generated to form a dynamic diagnostic path tree by calculating the relational weights between medical nodes in the graph. Based on the patient's current symptom description, examination results, and medical history, the inference network model generates interpretable diagnostic suggestions.

2. The method as described in claim 1, characterized in that, The structured knowledge graph constructed from multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results includes: Medical core entities representing diseases, symptoms, examination results, and laboratory indicators are extracted from the multimodal heterogeneous data using entity recognition and terminology standardization techniques, respectively. Based on the logical expressions and rule statements in the medical core entities and medical guide texts, a set of structured triples of medical core entities is generated using relation extraction technology. The set of structured triples is then organized into the structured knowledge graph, which includes a disease ontology layer, a symptom-examination association layer, and a diagnosis and treatment rule reasoning layer.

3. The method as described in claim 2, characterized in that, The examination results include imaging reports and laboratory data, the symptoms include medical record descriptions, and the method further includes: By employing the BERT-BiLSTM-CRF model, radiomics feature extraction network, and Vision-LanguageTransformer model, consistent representation and associative encoding of image features and text semantics are achieved to align entities in guideline texts, medical record descriptions, image reports, and laboratory data.

4. The method as described in claim 1, characterized in that, Also includes: A graph conflict resolution rule engine is built based on timestamps and evidence levels; Upon receiving new clinical data or changes to guideline text, the atlas incremental update protocol is triggered in real time via medical event stream.

5. The method as described in claim 1, characterized in that, The process of generating interpretable diagnostic suggestions based on the patient's current symptom description, examination results, and medical history through the inference network model includes: Based on the patient's current symptom description, examination results, and medical history, disease identification and path determination are performed using a graph neural network embedded with a knowledge graph. During the reasoning process, a causal reasoning mechanism is combined to identify potential confounding factors and misleading associations in order to correct the causal structure and optimize the diagnostic path. After completing the causal correction, interpretable diagnostic suggestions are generated based on the model output results. The confidence score of the diagnostic suggestions is calculated based on the causal strength, attention consistency and model uncertainty indicators in the reasoning process. If the confidence score of the diagnostic suggestions does not reach the preset threshold, the manual review mechanism or multidisciplinary consultation and collaboration process is automatically triggered.

6. The method as described in claim 1, characterized in that, The process involves disease identification and path determination based on a graph neural network embedded with an embedded knowledge graph, taking into account the patient's current symptom description, examination results, and medical history. During the reasoning process, a causal reasoning mechanism is incorporated to identify potential confounding factors and misleading associations, thereby correcting the causal structure and optimizing the diagnostic path. This includes: Based on the patient's current symptom description, examination results, and medical history, relevant diagnostic paths are activated from the knowledge graph. The probability distribution of each candidate disease is output through a graph neural network, and the DDx tree structure is optimized by combining examination costs, diagnostic benefits, and trauma risk factors. The multimodal attention fusion module integrates text medical record features, MRI / CT image features, and EEG time-series physiological signals. It uses Transformer and 1D-CNN networks to extract high-order semantic representations of each modality and dynamically allocates attention weights based on modality credibility to output a joint diagnostic feature vector. The joint diagnostic feature vector is input into the causal reasoning module, and a medical causal graph is constructed using a structural learning algorithm and an average causal effect analysis method. Based on counterfactual data augmentation and adversarial training mechanisms, potential spurious associations and confounding factors are identified to optimize the causal explanatory power of the diagnostic results, thereby correcting the causal structure and optimizing the diagnostic path.

7. The method according to any one of claims 1-6, characterized in that, The method for generating a dynamic graph-structured inference network model based on the structured knowledge graph combined with a large language model and embedded in a graph neural network module includes: Insert a graph neural network module into the Transformer backbone network of the large language model; Semantic pre-training and scenario fine-tuning of a large language model are performed based on general medical corpus and multi-center neurological case data to generate a dynamic graph structured inference network model.

8. A large-scale model-based auxiliary assessment device for neurological diseases, characterized in that, include: The graph construction unit is used to construct a structured knowledge graph based on multimodal heterogeneous data of neurological disease guideline texts, symptom descriptions, and examination results. The model building unit is used to generate a dynamic graph structured inference network model based on the structured knowledge graph combined with a large language model embedded in a graph neural network module, so as to form a dynamic diagnostic path tree by calculating the relational weights between each medical node in the graph. The assessment and analysis unit is used to generate interpretable diagnostic suggestions based on the patient's current symptom description, examination results, and medical history through the inference network model.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program stored in the memory, implements the steps of the large-model-based auxiliary assessment method for neurological diseases as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the large-model-based auxiliary assessment method for neurological diseases as described in any one of claims 1-7.

Images