An interdisciplinary chronic disease diagnosis system based on an AI medical large model

Abstract

This invention discloses a multidisciplinary chronic disease diagnostic system based on an AI-powered medical big data model, belonging to the field of chronic disease auxiliary diagnosis technology. The system collects multidisciplinary medical data from patients, and a multidisciplinary knowledge graph construction module establishes a multidisciplinary knowledge graph for chronic diseases based on medical knowledge. The core engine of the AI-powered medical big data model combines this multidisciplinary knowledge graph to obtain a reference diagnosis for the patient. The multidisciplinary knowledge graph for chronic diseases established in this invention is a knowledge graph formed from multidisciplinary medical knowledge. It integrates multidisciplinary knowledge, obtains the connections between cross-disciplinary medical entities, and then uses this knowledge graph for reasoning and diagnosis, enabling the recognition of medical professional terminology and the integration of multidisciplinary information. In this invention, the core engine of the AI-powered medical big data model uses the established multidisciplinary knowledge graph for chronic diseases as a basis to comprehensively consider the patient's multidisciplinary medical data, thereby obtaining a complete diagnostic conclusion, which helps to find the cause of the patient's illness.

Description

Technical Field

[0001] This invention relates to the field of chronic disease auxiliary diagnosis technology, specifically to a multidisciplinary chronic disease diagnosis system based on an AI-powered medical big data model. Background Technology

[0002] Chronic diseases are a class of non-communicable diseases with long course, complex causes, and a tendency to relapse. Typical examples include cardiovascular diseases, type 2 diabetes, chronic obstructive pulmonary disease (COPD), chronic kidney disease, and depression. Their pathogenesis often involves the coordinated abnormalities of multiple systems in the human body, such as the circulatory, metabolic, nervous, and immune systems, and has become a major global public health challenge.

[0003] Traditional chronic disease diagnosis models are limited by the medical system architecture and technological means, and have many shortcomings that urgently need to be addressed:

[0004] 1. Disciplinary barriers lead to fragmented diagnosis and treatment: The current medical system is highly specialized, often requiring patients with chronic diseases to travel between multiple specialists such as cardiology, endocrinology, and nephrology. Doctors in each department make judgments based solely on data and knowledge from their own discipline, lacking effective integration of interdisciplinary information. For example, endocrinologists may overlook the impact of depression on blood sugar control in diabetic patients, leading to a "treating the symptoms but not the root cause" dilemma in diagnosis and treatment.

[0005] 2. A one-sided diagnostic perspective can lead to missed diagnoses and misdiagnoses: Doctors specializing in a single discipline have limited knowledge and may find it difficult to identify "cross-system related lesions." For example, in the case of hypertension patients, cardiologists may focus on the blood pressure index itself but may miss secondary hypertension caused by kidney damage (a nephrology specialty) or endocrine disorders (an endocrinology specialty), resulting in incomplete diagnosis and delayed treatment.

[0006] 3. Imbalance between supply and demand of primary healthcare resources: Primary healthcare institutions undertake a large number of chronic disease management tasks, but there is a shortage of general practitioners with interdisciplinary diagnosis and treatment capabilities, which cannot meet patients' needs for comprehensive diagnosis and treatment, resulting in an over-concentration of high-quality medical resources.

[0007] 4. Insufficient multi-source data processing capabilities: Chronic disease diagnosis and treatment data are characterized by "massive volume and polymorphism," including structured laboratory indicators (such as blood glucose and creatinine), unstructured electronic medical record texts, medical images (such as echocardiography), and real-time data from wearable devices (such as heart rate variability). Physicians, limited by their professional background and available time, find it difficult to perform in-depth correlation analysis of multiple data types.

[0008] 5. Medical knowledge lags behind clinical needs: Global medical research output is growing at a rate of millions of papers per year. In the field of diabetes alone, more than 100,000 related studies are published every year. Clinicians find it difficult to keep up with the latest interdisciplinary treatment guidelines and evidence-based data in real time.

[0009] While existing medical auxiliary diagnostic technologies have some applications, they suffer from significant limitations: Chinese invention patent application CN115830764A, which discloses a "diabetes diagnostic system based on machine learning," targets only a single disease and processes only structured data such as blood glucose and glycated hemoglobin; US patent US11455951B2, which proposes a "medical image-assisted diagnostic method," focuses on a single data type and lacks the ability to integrate text and numerical data. These technologies are generally based on traditional machine learning models or single expert systems, exhibiting poor generalization ability, insufficient natural language understanding, and an inability to simulate interdisciplinary consultation thinking, making it difficult to meet the needs of comprehensive chronic disease diagnosis.

[0010] Alibaba Group's large-scale language models (such as Qwen-7B / 14B) excel in Chinese semantic understanding, logical reasoning, and multimodal processing, providing underlying technical support for interdisciplinary diagnosis of chronic diseases. However, this general-purpose model suffers from "insufficient domain adaptation": it lacks precise knowledge of medical terminology, making it unable to understand the pathological differences between "renal hypertension" and "essential hypertension"; it fails to integrate interdisciplinary knowledge connections, making it difficult to establish a multi-systemic association chain of "gut flora imbalance - elevated inflammatory factors - insulin resistance," thus preventing its direct application to high-precision medical diagnostic scenarios.

[0011] Therefore, there is an urgent need for a system that can transform a general large model into an intelligent system specifically for the interdisciplinary diagnosis of chronic diseases through domain fine-tuning and system integration, in order to break through the existing technological bottlenecks. Summary of the Invention

[0012] In view of this, the purpose of this invention is to provide an interdisciplinary chronic disease diagnosis system based on an AI medical big data model, in order to solve the problems of existing systems lacking knowledge of medical professional terminology and failing to achieve interdisciplinary information integration.

[0013] To achieve the above objectives, the system of this invention is an intelligent system for the comprehensive interdisciplinary diagnosis of chronic diseases, based on a pre-trained large-scale medical model and integrating multidisciplinary medical data and knowledge systems. Specifically, the system includes: a multi-source data integration and processing module, an interdisciplinary knowledge graph construction module, an AI large-scale medical model core engine, and a diagnosis and report generation module;

[0014] The multi-source data integration and processing module is used to collect and standardize patient interdisciplinary medical data; the patient interdisciplinary medical data includes all of the patient's diagnosis and treatment data.

[0015] The interdisciplinary knowledge graph construction module is used to construct a chronic disease interdisciplinary knowledge graph by taking the acquired medical knowledge, using diseases, symptoms, examination indicators, genes, physiological pathways, drugs and nutritional elements in medical knowledge as entities, and causal relationships, correlation relationships, inhibition relationships, treatment relationships and aggravation relationships in medical knowledge as relationships, and adopting an entity-relationship-attribute triple structure.

[0016] The core engine of the AI ​​medical big data model includes a trained data understanding unit, a multimodal data fusion unit, and an interdisciplinary reasoning unit. The data understanding unit is used to extract data corresponding to the chronic disease interdisciplinary knowledge graph from the patient's interdisciplinary medical data. The multimodal data fusion unit is used to perform multi-type data alignment and fusion on the data extracted by the data understanding unit. The interdisciplinary reasoning unit is used to obtain a path matching the patient from the chronic disease interdisciplinary knowledge graph based on the data processed by the multimodal data fusion unit and using the chronic disease interdisciplinary knowledge graph as the basis for reasoning navigation and optimization.

[0017] The diagnosis and report generation module is used to generate interdisciplinary chronic disease diagnosis reports based on the path results obtained from the core engine of the AI ​​medical big data model.

[0018] Its beneficial effects are as follows: The interdisciplinary knowledge graph for chronic diseases established in this invention is a knowledge graph formed by medical knowledge from various medical fields. This knowledge graph includes various medical entities and the relationships between them, integrating interdisciplinary knowledge to obtain the connections between interdisciplinary medical entities. Furthermore, the process of reasoning and diagnosis using this knowledge graph enables the recognition of medical terminology and the integration of interdisciplinary information. In this invention, the core engine of the AI ​​medical big data model, based on all patient treatment data and using the established interdisciplinary knowledge graph for chronic diseases, can comprehensively consider the patient's multidisciplinary treatment data to obtain a complete diagnostic conclusion, which helps to find the cause of the patient's illness.

[0019] Furthermore, standardizing interdisciplinary medical data for patients includes: data cleaning, data normalization, and data structuring; the data cleaning involves removing outliers using a rule engine, the data normalization involves standardizing indicators from different units, and the data structuring involves extracting the required information from unstructured text using regular expressions and named entity recognition technology.

[0020] This invention takes into account the diverse forms of all patient medical data. Therefore, it employs data cleaning, data normalization, and data structuring to process this diverse medical data, facilitating subsequent reasoning processes based on the medical data.

[0021] Furthermore, the interdisciplinary knowledge graph construction module is also used to reacquire medical knowledge from the time of the previous establishment of the chronic disease interdisciplinary knowledge graph to the specified time after a specified time has been reached, and update the chronic disease interdisciplinary knowledge graph accordingly.

[0022] The interdisciplinary knowledge graph for chronic diseases established in this invention is also updated at specific points in time to acquire new medical knowledge, ensuring that the diagnostic criteria are consistent with the forefront of clinical practice, solving the problem of knowledge lag, and improving the compliance with the latest treatment guidelines.

[0023] Furthermore, the interdisciplinary knowledge graph construction module also stores the confidence scores calculated from the number of clinical evidences for each path in the interdisciplinary knowledge graph of chronic diseases.

[0024] This invention also stores confidence scores. The reasoning path and confidence score data of the knowledge graph can be used to calibrate the model reasoning results in real time, solving the problem of black box reasoning in traditional models and improving the interpretability of model reasoning.

[0025] Furthermore, based on the data processed by the multimodal data fusion unit, and using the interdisciplinary knowledge graph of chronic diseases as the basis for reasoning navigation and optimization, the path to match the patient obtained from the interdisciplinary knowledge graph of chronic diseases includes:

[0026] By traversing the interdisciplinary knowledge graph of chronic diseases through a graph neural network and combining the data processed by the multimodal data fusion unit, a set number of inference paths are selected before the matching degree is determined, wherein the matching degree is determined by the confidence degree.

[0027] The diagnostic probability is calculated for the selected inference paths, and the final path matching the patient is determined based on the result of the diagnostic probability calculation. The diagnostic probability is the product of the similarity between the inference path and the data processed by the multimodal data fusion unit and the confidence level.

[0028] Furthermore, by using the Qwen-14B model as a base and pre-training and fine-tuning it in the medical field, a core engine for a large AI medical model was constructed, which includes a trained data understanding unit, a multimodal data fusion unit, and an interdisciplinary reasoning unit.

[0029] Furthermore, the data understanding unit uses the Chinese NLP capabilities of the Qwen-14B model, employs a pre-trained and fine-tuned NER model to extract medical entities and contextual relationships from unstructured text, identifies medical terminology ambiguities through knowledge graph entity linking technology, and identifies negation intent through semantic role labeling technology, in order to extract data corresponding to the chronic disease interdisciplinary knowledge graph from the patient's interdisciplinary medical data.

[0030] Furthermore, the multimodal data fusion unit converts text data into text vectors via a BERT encoder, converts structured numerical data into numerical vectors via a fully connected layer, extracts medical images into image feature vectors via a CNN, calculates the weights of each vector through a cross-modal attention matrix, and fuses them to generate a patient global feature vector with unified dimensions, thereby aligning and fusing the data extracted by the data understanding unit with multiple types of data.

[0031] Furthermore, the medical field pre-training and fine-tuning adopts a three-stage training strategy: knowledge injection pre-training, path-guided fine-tuning, and real-time inference calibration.

[0032] The knowledge injection pre-training includes: transforming the entity-relationship-attribute triples in the chronic disease interdisciplinary knowledge graph into structured text of entity-relationship-entity, forming a mixed corpus with a set number of multidisciplinary electronic medical records, incrementally pre-training the Qwen-14B model, using the AdamW optimizer, setting the learning rate and training epochs, so that the medical terminology vector space of the Qwen-14B model is aligned with the chronic disease interdisciplinary knowledge graph;

[0033] The path-guided fine-tuning includes: constructing a Few-Shot example set of patient features-knowledge graph path-diagnostic conclusion, and enabling the Qwen-14B model to master the mapping logic of features-path-conclusion through comparative learning;

[0034] The real-time calibration of inference includes: calling the inference path and confidence data of the chronic disease interdisciplinary knowledge graph in real time through the Neo4j Cypher query interface to verify the preliminary diagnosis results output by the Qwen-14B model. If the preliminary diagnosis results conflict with the high-confidence path in the chronic disease interdisciplinary knowledge graph, the Qwen-14B model is triggered to perform secondary inference, and the preliminary diagnosis results are corrected by combining the pathological mechanisms supplemented by the chronic disease interdisciplinary knowledge graph.

[0035] This invention utilizes a cross-disciplinary knowledge graph of chronic diseases in the core engine reasoning process of an AI medical big data model. This allows for the injection of incremental domain knowledge, solving the problem of fragmented knowledge in general models. By using the cross-disciplinary knowledge graph of chronic diseases as a navigation constraint for reasoning paths, the interpretability and accuracy of model diagnosis are improved. Data distribution deviation calibration of the cross-disciplinary knowledge graph of chronic diseases enhances the model's generalization ability. Medical terminology anchoring resolves semantic ambiguity, and dynamic knowledge synchronization updates ensure the timeliness of model knowledge.

[0036] Furthermore, the diagnosis and report generation module uses the Freemarker template engine to build report templates.

[0037] The present invention has the following advantages:

[0038] 1. Breaking down disciplinary barriers to achieve holistic diagnosis: Through interdisciplinary knowledge graphs and reasoning units, it simulates a multi-expert consultation scenario involving "endocrinology + nephrology + psychology", such as simultaneously diagnosing blood sugar problems and related kidney damage and depression in diabetic patients, thus solving the problem of fragmented diagnosis and treatment in traditional medicine.

[0039] 2. Improve diagnostic accuracy and reduce the rate of missed diagnoses and misdiagnoses: Multimodal data fusion technology improves diagnostic accuracy; through complication risk warning, the detection time of early diabetic nephropathy can be advanced.

[0040] 3. Empowering primary healthcare and balancing resource distribution: The system can be deployed in township health centers, enabling primary care physicians to receive interdisciplinary diagnostic and treatment support at the level of tertiary hospitals.

[0041] 4. Possesses continuous evolution and precise optimization capabilities: The knowledge graph is updated monthly with the latest medical knowledge, and is synchronized to the AI ​​large model through low-parameter fine-tuning to ensure that the diagnostic basis is consistent with the clinical forefront; at the same time, the reasoning path and confidence data of the knowledge graph can calibrate the model reasoning results in real time, solving the problems of "black box reasoning" and "knowledge lag" in traditional models, improving the interpretability of model reasoning, and having a high degree of compliance with the latest diagnosis and treatment guidelines.

[0042] 5. Optimize doctor-patient communication efficiency: The dual-mode report solves the problem of "professional terminology barriers", which improves patients' understanding of the diagnosis results and enhances treatment compliance.

[0043] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0044] Figure 1 This is a block diagram of the overall architecture of the system in this embodiment;

[0045] Figure 2 This is a flowchart illustrating the chronic disease diagnosis process implemented using the system described in this embodiment;

[0046] Figure 3 This is a schematic diagram of the interface for an interdisciplinary chronic disease diagnosis report generated using the system in this embodiment. Detailed Implementation

[0047] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art should understand that the embodiments described below are only for illustrating the present invention and should not be regarded as limiting the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] Example of an interdisciplinary chronic disease diagnosis system based on AI medical big data model

[0049] This embodiment of the system aims to overcome the shortcomings of existing technologies in chronic disease diagnosis, such as disciplinary barriers, insufficient data processing capabilities, and lagging knowledge updates. This embodiment's interdisciplinary chronic disease diagnosis system based on an AI medical big data model achieves comprehensive and accurate diagnosis of chronic diseases through multi-source data integration, interdisciplinary knowledge fusion, and intelligent reasoning, empowering primary healthcare. The system includes a multi-source data integration and processing module, an interdisciplinary knowledge graph construction module, an AI big data model core engine, and a diagnosis and report generation module. These modules work collaboratively to achieve interdisciplinary diagnostic functions. Specifically, the multi-source data integration and processing module collects and standardizes patients' interdisciplinary medical data; the interdisciplinary knowledge graph construction module stores entities and relationships of multidisciplinary medical knowledge; the AI ​​big data model core engine performs interdisciplinary reasoning on standardized data based on the knowledge graph; and the diagnosis and report generation module outputs a natural language diagnostic report. The specific structure of each module is as follows:

[0050] 1. Multi-source data integration and processing module.

[0051] This module is the core of the data input process, used to obtain the diagnostic and treatment information of patients to be diagnosed, such as... Figure 1 As shown, in this embodiment, the patient's medical information is obtained from the hospital information system (HIS) containing the patient's basic information / medical records, the patient's imaging examination reports in the picture archiving system (PACS), the patient's medical record information in the electronic medical record system (EMR), the patient's real-time heart rate / blood glucose data in wearable devices, the patient's laboratory examination data in the laboratory information system (LIS), and interdisciplinary medical data in the regional health information platform. All of the patient's medical information is obtained. The data to be integrated includes structured data, unstructured data, and semi-structured data. Structured data includes: laboratory test indicators (such as fasting blood glucose, serum creatinine, triglycerides), physiological monitoring data (such as heart rate, blood pressure), and demographic information (age, gender, family history). Unstructured data includes: electronic medical record (EMR) text, doctor's progress notes, medical imaging reports, and patient's chief complaint speech-to-text. Semi-structured data includes: discharge summaries, examination request forms, etc.

[0052] Specifically, this embodiment collects interdisciplinary medical data from hospital information systems (HIS), laboratory information systems (LIS), picture archiving systems (PACS), wearable devices (such as continuous glucose monitors and ECG bracelets), and regional health information platforms through standardized interfaces (such as HL7 FHIR and DICOM 3.0), and performs a data preprocessing process. The data preprocessing process in this embodiment includes data cleaning, data normalization, and data structuring.

[0053] Data cleaning: Remove outliers (such as abnormal data with blood glucose values ​​>33.3 mmol / L and no clinical notes) using a rule engine, and fill in missing values ​​(using the mean of the same disease or multiple imputation methods).

[0054] Data normalization: standardizing indicators with different units (e.g., converting blood pressure "130 / 85 mmHg" into a unified numerical vector);

[0055] Data structuring: Key information in unstructured text is extracted using regular expressions and named entity recognition (NER) technology (e.g., extracting "symptoms: dry mouth and excessive thirst, duration: 3 months, fasting blood glucose: 8.2 mmol / L" from "patient has had dry mouth and excessive thirst for the past 3 months, fasting blood glucose 8.2 mmol / L").

[0056] 2. Interdisciplinary knowledge graph construction module.

[0057] The knowledge graph constructed by this module adopts a "entity-relationship-attribute" triple structure, as follows:

[0058] Core entity types include: disease entities (such as hypertension, type 2 diabetes), symptom entities (such as dizziness, edema), test indicator entities (such as glomerular filtration rate eGFR), gene entities (such as APOE gene), physiological pathway entities (such as the renin-angiotensin-aldosterone system RAAS), drug entities (such as metformin), and nutrient element entities (such as dietary fiber).

[0059] Core relationship types include causal relationships (e.g., "hypertension → leading to → left ventricular hypertrophy"), association relationships (e.g., "gut flora imbalance → association → insulin resistance"), inhibitory relationships (e.g., "dietary fiber → inhibition → release of inflammatory factors"), therapeutic relationships (e.g., "metformin → treatment → type 2 diabetes"), and aggravating relationships (e.g., "depressive mood → aggravation → blood pressure fluctuations"), etc.

[0060] Knowledge Sources and Update Mechanism: The initial construction of the knowledge graph is based on UMLS (Unified Medical Language System), SNOMEDCT (Systematic Medical Terminology), the Chinese Guidelines for the Prevention and Control of Chronic Diseases (2024 Edition), and highly cited literature (impact factor ≥ 5) from the PubMed database in the past 5 years. Papers from top journals such as The New England Journal of Medicine and The Lancet are regularly crawled using web crawling technology, and new knowledge triplets are automatically extracted using natural language processing technology. After being reviewed by multidisciplinary experts from tertiary hospitals, the knowledge graph is updated.

[0061] This module not only provides structured medical knowledge support for the system, but also achieves precise optimization of the AI ​​medical model through a triple mechanism of "knowledge injection, inference navigation, and data calibration." ① Knowledge injection mechanism: Medical entities and relationships are transformed into structured corpora for incremental pre-training of the model, improving the accuracy of medical terminology comprehension to 94%. ② Inference navigation mechanism: Matching diagnostic inference paths and confidence levels are provided to guide the model in generating interpretable diagnostic evidence, achieving 90% interpretability of inference. ③ Data calibration mechanism: Evidence-based data from the knowledge graph is used to correct diagnostic biases caused by data bias, improving the generalization accuracy in primary care scenarios to 89%. It fundamentally solves the domain adaptation defects of general-purpose large models, and its optimization and improvement effects are reflected in the following five core dimensions:

[0062] 1) Incremental injection of domain knowledge to solve the problem of knowledge fragmentation in general models:

[0063] General large-scale models (such as Qwen-14B) rely heavily on general corpora for medical knowledge, resulting in fragmented and unrelated characteristics that fail to form interdisciplinary diagnostic and treatment logic. This module transforms multidisciplinary knowledge into structured vectors that can be parsed by large-scale models through a process of "entity alignment - relation mapping - knowledge triple extraction."

[0064] First, align the 1.2 million+ medical entities (diseases, indicators, pathways, etc.) in the knowledge graph with the large model vocabulary to establish unique vector identifiers for professional terms such as "renal hypertension";

[0065] The 800,000+ relationships such as "disease → cause → complications" were then converted into semantic association weights and injected into the model through incremental pre-training. This enabled the model to master cross-system association chains such as "type 2 diabetes → insulin resistance → hyperuricemia → kidney damage". Compared with the model without injected knowledge, the accuracy of medical terminology understanding was improved by 62%, and the recall rate of cross-disciplinary knowledge association recognition reached 89%.

[0066] 2) Inference path navigation constraints improve the interpretability and accuracy of model diagnosis:

[0067] Existing medical AI models often suffer from ambiguous diagnostic criteria due to "black box reasoning." This module provides explicit constraints for large-scale model reasoning by leveraging the path search capabilities of knowledge graphs.

[0068] During the model fine-tuning phase, the confidence path of "symptom-indicator-disease" in the knowledge graph (such as "dry mouth and excessive thirst + fasting blood glucose ≥7.0mmol / L + family history → type 2 diabetes") is used as a Few-Shot example to guide the model to learn the logic of "evidence-based reasoning".

[0069] During the inference phase, the model needs to call the knowledge graph query interface to obtain the matching path and confidence level (e.g., the confidence level of the path "hypertension → left ventricular hypertrophy" is 0.92), and generate diagnostic criteria based on the path, which increases the interpretability of the model's inference from 35% of the traditional model to 90%, while avoiding the error of "isolated interpretation of indicators" caused by noisy data, and reducing the false positive rate of diagnosis by 41%.

[0070] 3) Data distribution bias calibration to enhance model generalization ability:

[0071] Primary healthcare data often suffers from bias, with an overabundance of data on certain diseases and a lack of data on rare complications, which can easily lead to model generalization failure in primary healthcare settings. This module constructs a "standard diagnostic space" using evidence-based data from a knowledge graph (such as diagnostic criteria recommended in the "Guidelines for the Prevention and Treatment of Chronic Diseases" and pathological mechanisms from highly cited PubMed literature) to calibrate the model training data: when the proportion of samples for a certain type of high-incidence disease in primary healthcare (such as hypertension) in the model input data exceeds 60%, it automatically calls the relevant triples for "secondary causes of hypertension" in the knowledge graph (such as "renal artery stenosis → leading to → secondary hypertension") to supplement virtual samples and balance the data distribution; at the same time, it uses authoritative thresholds from the knowledge graph (such as eGFR < 60 ml / min / 1.73 m) to supplement virtual samples. 2 The model's judgment of fuzzy indicators was corrected (based on the standard for kidney injury), which improved the diagnostic generalization accuracy of the model in primary healthcare institutions from 71% to 89%, an improvement of 18 percentage points compared to traditional data augmentation methods.

[0072] 4) Medical terminology anchoring and resolution to address semantic ambiguity:

[0073] Medical terminology suffers from multiple meanings and multiple terms for the same meaning (e.g., "kidney failure" can refer to acute kidney injury or stage 5 chronic kidney disease), and the accuracy of general large-scale model for ambiguity identification is less than 50%.

[0074] This module provides terminology anchoring for the model through entity linking technology in knowledge graphs: During the data understanding phase, the model matches extracted terms (such as "renal failure") with entities in the knowledge graph, combined with context (such as "disease duration 10 years + eGFR 15ml / min / 1.73m"). 2The system precisely links to the entity “Stage 5 Chronic Kidney Disease” and obtains the entity’s attributes (such as “Irreversible Kidney Damage”) and relationships (such as “Requires Dialysis Treatment”), thereby improving the accuracy of disambiguation of model terminology to 94% and avoiding missed diagnoses due to misunderstandings of terminology.

[0075] 5) Dynamic knowledge is updated synchronously to ensure the timeliness of model knowledge:

[0076] Medical knowledge iterates rapidly (e.g., the 2024 American Diabetes Association Standards of Care added "SGLT2 inhibitors for primary prevention of diabetic nephropathy"), and general large-scale models lag behind in knowledge updates by 6-12 months. This module uses a closed-loop mechanism of "literature crawling - triple extraction - expert review - model incremental fine-tuning" to synchronously inject the latest medical knowledge into the model: 5000+ new knowledge triples (such as "dietary fiber → regulation → gut microbiota → improvement → insulin resistance") are extracted monthly from top journal articles. After review by multidisciplinary experts, knowledge synchronization is achieved through low-parameter fine-tuning (updating only 10% of the model's attention layer parameters). This increases the model's adherence to the latest treatment guidelines from 48% in traditional models to 92%, ensuring that diagnostic criteria are consistent with the clinical forefront.

[0077] 3. AI Large Model Core Engine.

[0078] This module is the core of the system's reasoning. It is based on the "Qwen-14B" large model and has been pre-trained and fine-tuned in the medical field. It includes three sub-modules: a data understanding unit, a multimodal data fusion unit, and an interdisciplinary reasoning unit.

[0079] Data Understanding Unit:

[0080] This study analyzes unstructured text using named entity recognition and entity linking techniques. Specifically, leveraging the Chinese NLP capabilities of a large-scale model, a pre-trained and fine-tuned NER model is employed to extract medical entities and contextual relationships from unstructured text. To address ambiguity in medical terminology, knowledge graph entity linking technology is introduced; for example, "kidney failure" is precisely linked to "stage 5 chronic kidney disease" instead of "acute kidney injury." For negative sentences (such as "no chest tightness or chest pain"), semantic role labeling technology is used to identify the negation intent and avoid erroneous extraction.

[0081] Multimodal data fusion unit:

[0082] An attention mechanism is employed to fuse text, numerical, and image feature vectors, achieving aligned fusion of multiple data types. Text data is transformed into text vectors using a BERT encoder, structured numerical data is transformed into numerical vectors using a fully connected layer, and medical images are extracted into image feature vectors using a CNN (such as ResNet50). The weights of each vector are calculated using a cross-modal attention matrix (e.g., the weight of blood glucose data for diabetic patients is higher than that of routine physical examination indicators), and then fused to generate a unified global feature vector for the patient (768 dimensions).

[0083] Interdisciplinary Reasoning Unit:

[0084] Based on knowledge graph path search and cosine similarity calculation to generate diagnostic probabilities, the system uses knowledge graph as the basis for reasoning navigation and optimization, and adopts a phased mechanism of "path search-probability calculation" to give full play to the role of knowledge graph in constraining and improving model reasoning.

[0085] The first stage (path search) involves traversing the knowledge graph using a graph neural network (GNN) and combining patient feature vectors (such as "type 2 diabetes + elevated urinary microalbumin + family history") to select the top 5 matching reasoning paths, such as "type 2 diabetes → insulin resistance → glomerular hyperfiltration → kidney damage" and "family history of diabetes → genetic susceptibility → abnormal glucose metabolism → kidney damage". The priority of path selection is determined by the confidence of the paths in the knowledge graph (calculated based on the number of evidence-based literature).

[0086] The second stage (probability calculation): The diagnostic probability is calculated based on the following formula, generating a mechanistic explanation. The formula embeds the path confidence (P(KG Paths)) provided by the knowledge graph to calibrate the model inference results:

[0087] ;

[0088] Wherein, P(KG Paths) is the confidence level of the knowledge graph path (calculated based on the amount of clinical evidence, such as the confidence level of "hypertension → left ventricular hypertrophy" being 0.92, because it is supported by 120 evidence-based articles); "Positively correlated" means that the probability of the diagnosis is proportional to the subsequent similarity calculation result; Sim is the cosine similarity function; Embed(Patient Data) means that the patient's multimodal data is mapped to a high-dimensional vector space through a large model to obtain its representation vector; Embed(KGPaths) means that the reasoning paths in the knowledge graph related to the current diagnostic task (e.g., hypertension → leading to → cardiac hypertrophy → association → heart failure) are also mapped to the same vector space.

[0089] The above formula calculates the matching degree between the patient's feature vector and the path vector; the final diagnosis probability is the product of the similarity and the path confidence.

[0090] 4. Diagnosis and Report Generation Module.

[0091] This module transforms AI reasoning results into standardized diagnostic reports, offering a dual-mode output function combining professional expertise and popular science information. Figure 3 As shown.

[0092] The professional model is geared towards doctors and includes:

[0093] ① Core diagnostic conclusions (e.g., "1. Type 2 diabetes (HbA1c 8.5%); 2. Stage 3 diabetic nephropathy (eGFR 58 ml / min / 1.73 mcg)"). 2 ); 3. Depressive state (PHQ-9 score of 12)

[0094] ② Reasoning basis (e.g., "Based on the patient's fasting blood glucose of 8.2 mmol / L and HbA1c of 8.5%, combined with a family history of diabetes, it is consistent with the diagnosis of type 2 diabetes; the decrease in eGFR is accompanied by an increase in urinary microalbumin, combined with the 'hyperglycemia → kidney damage' pathway in the knowledge graph, the diagnosis is diabetic nephropathy").

[0095] ③ Confidence score (e.g., 95% confidence for core diagnosis and 88% confidence for associated complications);

[0096] ④ Personalized recommendations (e.g., "It is recommended to complete a fundus examination to rule out diabetic retinopathy, and to have joint diagnosis and treatment by the endocrinology and nephrology departments").

[0097] The science popularization model is geared towards patients, explaining the diagnosis results in layman's terms (such as "Your blood sugar is poorly controlled, and long-term high blood sugar has affected your kidney function. It is recommended that you eat less sweets and take your hypoglycemic drugs on time").

[0098] To enable the modules described above in this embodiment to perform their corresponding functions, the system hardware and software environment are configured as follows:

[0099] Hardware configuration: The server uses an Intel Xeon Gold 6330 processor (28 cores), 128GB of memory, an NVIDIA A100 GPU (80GB of video memory), and a 10TB SSD for storage; the wearable device is a dynamic blood glucose monitor with Bluetooth 5.0 (sampling frequency of 5 minutes / time).

[0100] Software environment: The operating system is Ubuntu 20.04 LTS, the AI ​​large model is deployed based on the PyTorch 2.0 framework, the database uses Neo4j (knowledge graph storage) and MySQL (structured data storage), and the interface development uses the Spring Boot framework, supporting the HL7 FHIR v4.0 standard.

[0101] The specific implementation of each module is as follows:

[0102] 1. Implementation of multi-source data integration and processing module.

[0103] By connecting to the hospital's HIS system via the FHIR API, basic patient information and medical records can be obtained; by parsing ultrasound and CT image reports from the PACS system via the DICOMViewer SDK; and by receiving real-time data (such as heart rate and blood glucose) from wearable devices via the MQTT protocol.

[0104] Data cleaning was performed using Python's Pandas library, and outlier detection was based on... In principle, missing value completion uses the KNN algorithm (K=5); unstructured text processing uses HanLP's medical NER model, achieving an entity recognition accuracy of 91.5%.

[0105] 2. Implementation of interdisciplinary knowledge graph construction.

[0106] A graph database was built based on Neo4j, initially importing 1.2 million medical entities and 800,000 relationships from UMLS. The latest literature from PubMed and the Chinese Journal of Internal Medicine was crawled using the Python Scrapy framework. New knowledge triples were extracted using the BERT model. The database was updated after being reviewed by three experts at the associate chief physician level or above (approximately 5,000 data entries were updated monthly).

[0107] The path search of the knowledge graph uses Neo4j's Cyphe query statement, such as "MATCH (d:disease{name:'type 2 diabetes'})-[r:cause]->(c:complications) RETURN d,r,c" to quickly locate diabetes-related complications.

[0108] 3. Implementation of the core engine for large-scale AI models.

[0109] Using Qwen-14B as the base model and combining it with the optimization mechanism of interdisciplinary knowledge graphs, a three-stage training strategy of "knowledge injection pre-training - path-guided fine-tuning - real-time inference calibration" is adopted to realize the transformation of general large-scale models into professional medical models:

[0110] 1) Knowledge-injected pre-training:

[0111] Medical entities and relation triples in the knowledge graph were transformed into structured text of "entity 1-relation-entity 2" (e.g., "type 2 diabetes-leads to-kidney damage"). This text was then mixed with 100,000 multidisciplinary electronic medical records to form a hybrid corpus. The base model was incrementally pre-trained using the AdamW optimizer with a learning rate of 2e-5 and 3 training epochs. This aligned the model's medical terminology vector space with the knowledge graph, improving the terminology recognition accuracy to 94%.

[0112] 2) Path-guided fine-tuning:

[0113] A Few-Shot example set (e.g., "Fasting blood glucose 9.1 mmol / L + HbA1c 8.7% - Path: Type 2 diabetes → Insulin resistance → Hyperglycemia - Diagnosis: Type 2 diabetes") was constructed, containing 5000 examples. Through comparative learning, the model mastered the mapping logic of "feature-path-conclusion", and the inference accuracy was improved by 23% compared with unguided fine-tuning.

[0114] 3) Real-time inference calibration:

[0115] During the inference process, the inference path and confidence data of the knowledge graph are called in real time through Neo4j's Cypher query interface to verify the preliminary diagnostic results output by the model. If the model's diagnostic conclusion conflicts with the high-confidence path in the knowledge graph (e.g., the model diagnoses "primary hypertension" but the knowledge graph shows that "renal hypertension" is more consistent with the patient's renal function indicators), the model is triggered to conduct secondary inference. The conclusion is corrected by combining the pathological mechanism supplemented by the knowledge graph (e.g., "renal artery stenosis → increased renin secretion → elevated blood pressure") to ensure diagnostic consistency.

[0116] Multimodal fusion employs the cross-modal attention mechanism of Transformer, implemented in PyTorch through a custom Attention layer; path matching of the inference unit uses cosine similarity calculation, implemented through Python's Scikit-learn library, and the product of the similarity calculation result and the path confidence of the knowledge graph serves as the basis for the final diagnostic probability.

[0117] 4. Implementation of the diagnosis and report generation module.

[0118] This system enables the creation of report templates based on the Freemarker template engine. The professional template includes ICD-11 disease coding fields for easy integration with hospital systems; the popular science template uses Natural Language Generation (NLG) technology to transform technical terms into more accessible expressions through a large model (e.g., "decreased glomerular filtration rate" becomes "renal detoxification function is somewhat reduced"). Reports support three output methods: online preview, PDF download, and HL7 push notifications.

[0119] Based on the system of this embodiment, it can achieve the following in practical applications: Figure 2 The process is shown below.

[0120] S1: Interdisciplinary data collection: Obtain comprehensive medical data of the target patient through multi-source data interfaces, covering data from at least two relevant departments (e.g., for diabetic patients, blood glucose data from the endocrinology department and renal function data from the nephrology department should be included), and supplement information such as family medical history and lifestyle habits (e.g., smoking, drinking).

[0121] S2: Data Standardization Processing: Calls the multi-source data integration and processing module to perform cleaning, normalization, and structuring processing, and outputs a standardized data set to ensure that the data format meets the input requirements of the AI ​​engine.

[0122] S3: Multimodal Data Fusion: Standardized data is input into the core engine of the AI ​​large model. After being parsed by the data understanding unit, the multimodal fusion unit generates a global feature vector of the patient.

[0123] S4: Interdisciplinary Intelligent Reasoning: The reasoning unit calls upon the knowledge graph, searches for matching paths, calculates diagnostic probabilities, and identifies interdisciplinary related lesions (such as "depression → sleep disorders → blood sugar fluctuations") and potential complications (such as "type 2 diabetes patients have a 32% risk of kidney damage in the next 5 years"). The reasoning results include a confidence score for the diagnostic conclusion; a confidence score ≥90% is considered a core diagnosis, while a confidence score between 75% and 90% is considered a diagnosis requiring confirmation and suggests further investigation.

[0124] S5: Diagnostic Report Generation: The report module generates diagnostic reports from both the doctor's and patient's perspectives based on the reasoning results, and supports PDF export for integration with hospital systems.

[0125] Taking Mr. Zhang, a 58-year-old male patient, as an example, his medical data is as follows:

[0126] ①Basic information: Family history of diabetes, 30-year smoking history;

[0127] ② Endocrinology data: fasting blood glucose 9.1 mmol / L, HbA1c 8.7%, metformin treatment was ineffective;

[0128] ③ Nephrology data: Urinary microalbumin / creatinine ratio 320 mg / g, eGFR 56 ml / min / 1.73 m 2 ;

[0129] ④ Psychological data: Low mood for the past two months, PHQ-9 score of 13;

[0130] ⑤ Unstructured data: EMR records "nocturnal polyuria, fatigue, and decreased appetite".

[0131] System diagnostic process in this embodiment:

[0132] S1-S2: Collect the above multidisciplinary data, clean it, and extract the key entities "type 2 diabetes, elevated urinary microalbumin, depressive state, family history";

[0133] S3: The fusion module generates patient feature vectors, with blood glucose and renal function indicators having the highest weights;

[0134] S4: The reasoning unit searches the knowledge graph path "Type 2 diabetes → Insulin resistance → Kidney damage" and "Depressive state → Sleep disorder → Blood glucose fluctuation", and calculates the diagnosis probability: Type 2 diabetes (98%), Diabetic nephropathy stage 3 (92%), Moderate depressive state (89%).

[0135] S5: Generate report:

[0136] The professional model recommends "adjusting the blood glucose lowering regimen by the endocrinology department (adding SGLT2 inhibitors), assessing the progression of kidney disease by the nephrology department, and providing intervention by the psychology department";

[0137] The popular science explanation states, "Your blood sugar is not well controlled, which has already affected your kidneys. Emotional problems can also make blood sugar control more difficult. You need both medication and psychological adjustment for management."

[0138] The diagnosis was confirmed by a joint consultation of the endocrinology, nephrology and psychology departments of a top-tier hospital. It was completely consistent with the system diagnosis, and the complication warning was 3 months earlier than the traditional single-discipline diagnosis.

[0139] The system in this embodiment has the following significant advantages:

[0140] 1. Breaking down disciplinary barriers to achieve holistic diagnosis: Through interdisciplinary knowledge graphs and reasoning units, it simulates a multi-expert consultation scenario involving "endocrinology + nephrology + psychology", such as simultaneously diagnosing blood sugar problems and related kidney damage and depression in diabetic patients, thus solving the problem of fragmented diagnosis and treatment in traditional medicine.

[0141] 2. Improve diagnostic accuracy and reduce the rate of missed diagnoses and misdiagnoses: Multimodal data fusion technology improves the diagnostic accuracy to 92.3% (based on a test set of 3,000 patients with chronic diseases, which is 28% higher than the accuracy of single-discipline diagnosis); through complication risk warning, the detection time of early diabetic nephropathy can be advanced by 6-12 months.

[0142] 3. Empowering primary healthcare and balancing resource distribution: The system can be deployed in township health centers, enabling primary care physicians to receive interdisciplinary diagnostic and treatment support at the level of tertiary hospitals. Tests show that after primary care physicians used this system, the accuracy rate of chronic disease diagnosis increased from 65% to 89%.

[0143] 4. Possesses continuous evolution and precise optimization capabilities: The knowledge graph is updated with 5,000+ pieces of the latest medical knowledge every month, and is synchronized to the AI ​​large model through low-parameter fine-tuning to ensure that the diagnostic basis is consistent with the clinical forefront; at the same time, the reasoning path and confidence data of the knowledge graph can calibrate the model reasoning results in real time, solving the problems of "black box reasoning" and "knowledge lag" in traditional models. The interpretability of model reasoning has been improved from 35% to 90%, and the compliance with the latest diagnosis and treatment guidelines has reached 92%.

[0144] 5. Optimize doctor-patient communication efficiency: The dual-mode reporting solves the problem of "professional terminology barriers", increasing patients' understanding of the diagnosis results from 42% to 78%, thereby improving treatment compliance.

[0145] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A cross-disciplinary chronic disease diagnosis system based on an AI medical large model, characterized by, It includes a multi-source data integration and processing module, an interdisciplinary knowledge graph construction module, an AI medical big data model core engine, and a diagnosis and report generation module; The multi-source data integration and processing module is used to collect and standardize patient interdisciplinary medical data; the patient interdisciplinary medical data includes all of the patient's diagnosis and treatment data. The interdisciplinary knowledge graph construction module is used to construct a chronic disease interdisciplinary knowledge graph by taking the acquired medical knowledge, using diseases, symptoms, examination indicators, genes, physiological pathways, drugs and nutritional elements in medical knowledge as entities, and causal relationships, correlation relationships, inhibition relationships, treatment relationships and aggravation relationships in medical knowledge as relationships, and adopting an entity-relationship-attribute triple structure. The core engine of the AI ​​medical big data model includes a trained data understanding unit, a multimodal data fusion unit, and an interdisciplinary reasoning unit. The data understanding unit is used to extract data from the patient's interdisciplinary medical data that corresponds to the data in the interdisciplinary knowledge graph of chronic diseases. The multimodal data fusion unit is used to perform multi-type data alignment and fusion on the data extracted by the data understanding unit. The interdisciplinary reasoning unit is used to obtain a path matching the patient from the interdisciplinary knowledge graph of chronic diseases, based on the data processed by the multimodal data fusion unit and using the interdisciplinary knowledge graph of chronic diseases as the basis for reasoning navigation and optimization. The diagnosis and report generation module is used to generate interdisciplinary chronic disease diagnosis reports based on the path results obtained from the core engine of the AI ​​medical big data model. 2.The cross-disciplinary chronic disease diagnosis system based on an AI medical large model according to claim 1, wherein, Standardized patient interdisciplinary medical data includes: data cleaning, data normalization, and data structuring; data cleaning involves removing outliers using a rule engine, data normalization involves standardizing indicators from different units, and data structuring involves extracting the required information from unstructured text using regular expressions and named entity recognition technology. 3.The AI medical large model-based cross-disciplinary chronic disease diagnosis system of claim 1, wherein, The interdisciplinary knowledge graph construction module is also used to retrieve medical knowledge from the time the interdisciplinary knowledge graph for chronic diseases was previously established to the specified time after a specified time has been reached, and update the interdisciplinary knowledge graph for chronic diseases accordingly. 4.The AI medical large model-based cross-disciplinary chronic disease diagnosis system of claim 1, wherein, The interdisciplinary knowledge graph construction module also stores the confidence scores calculated from the number of clinical evidences for each pathway in the interdisciplinary knowledge graph of chronic diseases.

5. The interdisciplinary chronic disease diagnosis system based on an AI-powered medical big data model according to claim 4, characterized in that, Based on the data processed by the multimodal data fusion unit, and using the interdisciplinary knowledge graph of chronic diseases as the basis for reasoning navigation and optimization, the path to match the patient obtained from the interdisciplinary knowledge graph of chronic diseases includes: By traversing the interdisciplinary knowledge graph of chronic diseases through a graph neural network and combining the data processed by the multimodal data fusion unit, a set number of inference paths are selected before the matching degree is determined, wherein the matching degree is determined by the confidence degree. The diagnostic probability is calculated for the selected inference paths, and the final path matching the patient is determined based on the result of the diagnostic probability calculation. The diagnostic probability is the product of the similarity between the inference path and the data processed by the multimodal data fusion unit and the confidence level.

6. The interdisciplinary chronic disease diagnosis system based on an AI-powered medical big data model according to claim 4, characterized in that, By using the Qwen-14B model as a base and pre-training and fine-tuning it in the medical field, a core engine for a large AI medical model was constructed, which includes a trained data understanding unit, a multimodal data fusion unit, and an interdisciplinary reasoning unit.

7. The interdisciplinary chronic disease diagnosis system based on an AI-powered medical big data model according to claim 6, characterized in that, The data understanding unit uses the Chinese NLP capabilities of the Qwen-14B model, employs a pre-trained and fine-tuned NER model to extract medical entities and contextual relationships from unstructured text, identifies medical terminology ambiguities through knowledge graph entity linking technology, and identifies negation intent through semantic role labeling technology, in order to extract data corresponding to the chronic disease interdisciplinary knowledge graph from the patient's interdisciplinary medical data.

8. The interdisciplinary chronic disease diagnosis system based on an AI-powered medical big data model according to claim 6, characterized in that, The multimodal data fusion unit converts text data into text vectors using a BERT encoder, converts structured numerical data into numerical vectors using a fully connected layer, and extracts image feature vectors from medical images using a CNN. It then calculates the weights of each vector using a cross-modal attention matrix and fuses them to generate a unified global feature vector for the patient, thereby aligning and fusing the data extracted by the data understanding unit across multiple data types.

9. The interdisciplinary chronic disease diagnosis system based on an AI-powered medical big data model according to claim 6, characterized in that, The aforementioned pre-training and fine-tuning in the medical field employs a three-stage training strategy: knowledge-injected pre-training, path-guided fine-tuning, and real-time inference calibration. The knowledge injection pre-training includes: transforming the entity-relationship-attribute triples in the chronic disease interdisciplinary knowledge graph into structured text of entity-relationship-entity, forming a mixed corpus with a set number of multidisciplinary electronic medical records, incrementally pre-training the Qwen-14B model, using the AdamW optimizer, setting the learning rate and training epochs, so that the medical terminology vector space of the Qwen-14B model is aligned with the chronic disease interdisciplinary knowledge graph; The path-guided fine-tuning includes: constructing a Few-Shot example set of patient features-knowledge graph path-diagnostic conclusion, and enabling the Qwen-14B model to master the mapping logic of features-path-conclusion through comparative learning; The real-time calibration of inference includes: calling the inference path and confidence data of the chronic disease interdisciplinary knowledge graph in real time through the Neo4j Cypher query interface to verify the preliminary diagnosis results output by the Qwen-14B model. If the preliminary diagnosis results conflict with the high-confidence path in the chronic disease interdisciplinary knowledge graph, the Qwen-14B model is triggered to perform secondary inference, and the preliminary diagnosis results are corrected by combining the pathological mechanisms supplemented by the chronic disease interdisciplinary knowledge graph.

10. The interdisciplinary chronic disease diagnosis system based on an AI-powered medical big data model according to claim 1, characterized in that, The diagnostics and report generation module uses the Freemarker template engine to build report templates.

Images