A model chain-based multi-knowledge base clinical decision method, electronic equipment and medium

Abstract

The application provides a model chain-based multi-knowledge base clinical decision method, an electronic device and a medium, relates to the technical field of medical data processing, and the method comprises the following steps: in response to a clinical decision instruction, converting an unstructured medical record description of a patient indicated by the clinical decision instruction into a structured medical record through a first model chain node, and generating a disease prediction result of the patient from the structured medical record by referring to a disease diagnosis knowledge block; referring to a medication knowledge block through a second model chain node, combining the structured medical record and the disease prediction result to give a medication suggestion; referring to an inspection knowledge block through a third model chain node, combining the structured medical record and the disease prediction result to give an inspection suggestion; and feeding back the disease prediction result, the medication suggestion and the inspection suggestion to a doctor terminal for display to assist the doctor in generating a clinical decision for the patient. The doctor can obtain more accurate clinical decisions.

Description

A multi-knowledge-base clinical decision-making method, electronic device, and medium based on model chains Technical Field

[0001] This application relates to the field of medical data processing technology, and more specifically, to a multi-knowledge base clinical decision-making method, electronic device, and medium based on model chains. Background Technology

[0002] As the medical field continues to evolve towards personalization and intelligence, Clinical Decision Support Systems (CDSS) are gradually becoming important auxiliary tools for improving the quality of diagnosis and treatment, reducing misdiagnosis rates, and enhancing treatment outcomes. Traditional CDSS systems, based on rule engines or expert systems, provide clinical decision suggestions to physicians through pre-set rules and knowledge bases. While they have a certain range of applications in areas such as assisted diagnosis, treatment recommendations, and drug interaction warnings, their decision-making mechanisms are singular and fixed, lacking flexibility and interpretability. Especially when faced with vast amounts of knowledge, complex medical environments, and significant individual differences among patients, the accuracy of decision support still needs improvement. Summary of the Invention

[0003] The purpose of this application is to provide a multi-knowledge base clinical decision-making method, electronic device, and medium based on model chains, in order to solve the technical problem of poor flexibility and accuracy of decision-making mechanisms in existing clinical decision support systems.

[0004] In a first aspect, the present invention provides a multi-knowledge-base clinical decision-making method based on model chains. The method includes, in response to a clinical decision instruction, converting the unstructured medical record description of the patient indicated by the clinical decision instruction into a structured medical record through a first model chain node, and generating a disease prediction result for the patient from the structured medical record with reference to a disease diagnosis knowledge block; providing medication recommendations by referring to a medication knowledge block through a second model chain node, combining the structured medical record and the disease prediction result; providing laboratory test recommendations by referring to a laboratory test knowledge block through a third model chain node, combining the structured medical record and the disease prediction result; and feeding back the disease prediction result, medication recommendations, and laboratory test recommendations to the doctor for display, so as to assist the doctor in generating clinical decisions for the patient.

[0005] In a second aspect, the present invention provides an electronic device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform any of the steps of a multi-knowledge base clinical decision-making based on a model chain as described in the foregoing embodiments.

[0006] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of any of the model chain-based multi-knowledge base clinical decision-making methods described in the foregoing embodiments.

[0007] This application provides a model chain-based multi-knowledge base clinical decision-making method, electronic device, and medium. The method includes, in response to a clinical decision instruction, converting the patient's unstructured medical record description as indicated by the clinical decision instruction into a structured medical record through a first model chain node, and extracting the patient's disease prediction results from the structured medical record with reference to a disease diagnosis knowledge block; providing medication recommendations by referencing a medication knowledge block through a second model chain node, combining the structured medical record and disease prediction results; and providing laboratory test recommendations by referencing a laboratory test knowledge block through a third model chain node, combining the structured medical record and disease prediction results; and feeding back the disease prediction results, medication recommendations, and laboratory test recommendations to the physician for display, thereby assisting the physician in generating clinical decisions for the patient. The model chain designed in this application breaks down decision support into three tasks: disease diagnosis, medication recommendation generation, and laboratory test recommendation. The model chain reasoning is staged, and medical knowledge is injected in batches from multiple knowledge bases. The reasoning process is transparent, allowing for better collaboration with physician experience, thereby assisting physicians in obtaining more accurate clinical decisions. It also makes diagnostic decisions more interpretable and better applicable to individuals with different differences, offering good flexibility. Attached Figure Description

[0008] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 is a structural diagram of an inference link provided in an embodiment of this application;

[0010] Figure 2 is a schematic diagram of an epilepsy classification framework in a disease guideline provided in an embodiment of this application;

[0011] Figure 3 is a flowchart of a multi-knowledge base clinical decision-making method based on model chains provided in this application;

[0012] Figure 4 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] First, the application scenario of this application will be described. This application can be applied to the design of clinical decision support systems.

[0014] Existing knowledge-based and rule-based clinical decision support (CDSS) systems typically use fixed rules, such as IF-THEN rules, ontology, or logical conditions, to store clinical knowledge and form a knowledge base. A rule engine then matches patient data with these knowledge base rules to derive relevant diagnoses and recommendations. While the system structure is simple and easy to understand, rule and knowledge updates rely on manual maintenance, making it unable to automatically adapt to emerging new disease patterns, drug therapies, and laboratory tests. This results in poor scalability. Furthermore, significant individual differences among patients, varying symptoms, and complex diagnostic and treatment logic make it difficult to define with rigid rules, leading to insufficient flexibility and limiting the system's decision support capabilities.

[0015] With advancements in machine learning and natural language processing technologies, AI-driven intelligent CDSS systems, powered by artificial intelligence algorithms, perform large-scale data analysis and pattern recognition based on extensive historical patient data and real-time monitoring information, thereby making accurate predictions and decisions. However, these models are often considered "black boxes," meaning their decisions typically only provide the outcome, the decision-making process is opaque, difficult to integrate with doctors' experience, and not easily understood or verified by doctors. This hinders doctors' trust and the effectiveness of the system's application. Furthermore, the models lack explicit reasoning capabilities regarding medical knowledge logic, potentially conflicting with clinical common sense.

[0016] Based on this, this application provides a multi-knowledge base clinical decision-making method, electronic device, and medium based on model chains.

[0017] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0018] Example 1

[0019] The clinical decision support system provided in this application includes three knowledge bases: disease diagnosis, medication, and laboratory tests, as well as a LangChain-driven inference chain.

[0020] In one implementation, multiple knowledge bases are constructed, specifically including the construction of a disease diagnosis knowledge base, a medication knowledge base, and a laboratory test knowledge base.

[0021] Specifically, the target disease guidelines can be preprocessed to obtain standard-format disease guideline text. The medical knowledge recorded in the disease guidelines can include text, tables, images, etc. Here, the text and tables in the target disease guidelines can be hierarchically classified and converted to Markdown format. The images and the eighth prompt word template from the target disease guidelines are input into the fifth model chain node to obtain the image description text output by the fifth model chain node, which is then converted to Markdown format. The fifth model chain node can use a multimodal large model, such as the Qwen-vl series model.

[0022] For example, the original PDF or Word version of the disease guideline is preprocessed into Markdown format to supplement the knowledge in the large language model. The preprocessing stage requires restoring the guideline's hierarchical heading information as much as possible, dividing it into first-level, second-level, and third-level headings based on the table of contents and main text headings, ideally not exceeding six levels. Tables are organized in Markdown format, which can be done manually or generated by tools and then manually reviewed. Text formatting is restored to the guideline, including bolding, italics, and superscripts.

[0023] Images from disease guidelines can be input into a multimodal model to provide a detailed description in natural language. Taking Figure 2 as an example, uploading it to the multimodal model and writing the prompt "Describe the 2017 ILAE epilepsy classification framework in detail with reference to the image, outputting a paragraph" will result in the model providing a detailed natural language description of the image: "The 2017 International League Against Epilepsy (ILAE) epilepsy classification framework systematically integrates the multidimensional characteristics of epilepsy in a hierarchical structure, forming a complete classification system from clinical manifestations to potential causes. The leftmost side of the framework emphasizes 'comorbidities' (such as cognitive impairment or mental illness), suggesting that epilepsy management should focus on accompanying symptoms. The core part is divided into three layers: Seizure type (top layer) is divided into focal (abnormal discharge in a specific brain region), generalized (synchronous discharge throughout the brain), and unknown origin based on electrophysiological origin; Epilepsy type (middle layer) is further refined, including focal, generalized..." The epilepsy classification system categorizes epilepsy into mixed types (such as Lennox-Gastaut syndrome) and unknown types; the epilepsy syndrome (bottom level) defines the disease entity based on specific clinical characteristics (such as age of onset and EEG pattern) (such as childhood absence epilepsy). The right-hand etiology classification horizontally connects all levels, covering structural (such as brain tumors), genetic (such as SCN1A gene mutations), infectious (such as sequelae of encephalitis), metabolic (such as porphyria), immune (such as autoimmune encephalitis), and unknown causes, emphasizing a mechanism-based approach to guide precision diagnosis and treatment. This framework, through vertical stratification of "symptom-type-syndrome-etiology" and horizontal expansion of "comorbidity-etiology," constructs a multi-dimensional, traceable epilepsy classification system, providing a structured foundation for individualized treatment and research.

[0024] The standard format disease guideline text is decomposed into knowledge slices for disease diagnosis, medication, and laboratory tests.

[0025] Here, medical experts can deconstruct the pre-processed guideline text and reorganize the document into three categories: disease diagnosis knowledge, medication knowledge, and laboratory test knowledge. Taking the "Clinical Practice Guidelines - Epilepsy Volume" (2023 Revised Edition) as an example, the book has 13 chapters, systematically introducing the definition, classification, diagnosis, and prognosis of epilepsy; drug treatment and surgical treatment; diagnosis and management of status epilepticus; diagnosis and management of drug-resistant epilepsy; electroencephalography and neurophysiological monitoring; epilepsy-related imaging techniques; comorbidities of epilepsy; precautions for epilepsy patients' medical visits and treatment; comprehensive management of epilepsy patients; and traditional Chinese medicine diagnosis and treatment of epilepsy. After deconstruction, the original chapter content and specific details of the disease diagnosis knowledge, medication knowledge, and laboratory test knowledge are shown in Table 1 below.

[0026] Table 1 Knowledge Decomposition Mapping Table

[0027]

[0028] Each knowledge slice is input into the fourth model chain node to obtain the corresponding disease diagnosis knowledge block vector output by the fourth model chain node. Here, the document can be further segmented according to chapters, paragraphs, or specified character lengths, and then vectorized using embedding models such as bge-large-zh-v1.5 and gte-Qwen2-7B-instruct.

[0029] Based on the association and storage of all knowledge block vectors and corresponding disease diagnosis knowledge slices, a disease diagnosis knowledge base, a medication knowledge base, and a laboratory test knowledge base are constructed respectively.

[0030] This application's knowledge base employs multi-knowledge base construction, knowledge block matching, and model chain reasoning to drive a large language model to generate complete disease diagnoses, medication recommendations, and laboratory test suggestions based on guideline knowledge and patient medical records, providing comprehensive clinical decision support for doctors and patients. The multi-knowledge base is vectorized and stored using an embedding model, simplifying construction and updates.

[0031] The inference link is used to obtain unstructured medical record descriptions input by users, and to generate disease prediction results, medication suggestions, and test and examination suggestions through the various models on the link.

[0032] Figure 1 is a structural diagram of an inference link provided in an embodiment of this application. As shown in Figure 1, the inference link may include a first model chain node, a second model chain node, a third model chain node, and a fourth model chain node. The first model chain node includes a first model chain sub-node, a second model chain node, and a third model chain node. The second model chain node includes a fourth model chain sub-node and a fifth model chain node. The third model chain node includes a sixth model chain node and a seventh model chain node.

[0033] The first, second, and third model chain nodes are connected sequentially. The input of the fourth model chain node is connected to the output of the first model chain node, and the output of the fourth model chain node is connected to the inputs of the second, fourth, and sixth model chain nodes, respectively.

[0034] The input of the fourth model chain node is also connected to the outputs of the first and third model chain nodes, and the output of the fourth model chain node is connected to the fifth model chain node. The input of the sixth model chain node is also connected to the outputs of the first and third model chain nodes, and the output of the sixth model chain node is connected to the seventh model chain node.

[0035] In one implementation, both the model chain nodes and model chain child nodes contain AI models, interfaces, and other components. These model chain nodes and child nodes can use the same or different models. Specifically, the fourth model chain node can include an embedding model, such as bge-large-zh-v1.5 or gte-Qwen2-7B-instruct, for knowledge vectorization. The first, second, third, fourth, fifth, sixth, and seventh model chain nodes can include language models, such as the Qwen series models.

[0036] Furthermore, the reasoning process of the inference chain may include the following steps:

[0037] Chain1-1: Based on the medical record collection template, unstructured natural language medical records are standardized into structured medical records.

[0038] First prompt template: You are a medical expert and need to generate structured medical records. Please generate a structured medical record for the patient based on the patient's medical record description {unstructured medical record description} and the {medical record collection template}.

[0039] Chain1-2: Based on the structured medical records generated by Chain1-1 and the matched TopN disease diagnosis-related knowledge blocks, it infers and generates a comprehensive diagnostic result and reasoning process.

[0040] Second prompt template: You are a medical expert and need to generate a comprehensive diagnosis. Based on the patient's medical record {structured medical record} and referring to the knowledge in {disease diagnosis knowledge blocks}, please generate a comprehensive diagnostic result, including diagnostic suggestions, suggestions for further examinations, differential diagnosis, etiological analysis, and treatment methods.

[0041] Chain1-3: Extract disease diagnosis results from the comprehensive diagnosis results generated by Chain1-2.

[0042] Third prompt template: You are a medical expert and need to extract core information from diagnostic results. Please extract the disease prediction results from {Comprehensive Diagnostic Results}.

[0043] Chain2-1: Based on the structured medical records generated by Chain1-1, the matched TopN medication-related knowledge, and the disease diagnosis results extracted by Chain1-3, it infers and generates medication suggestions and reasoning processes.

[0044] Fourth prompt template: You are a medical expert and need to generate a medication-specific module. Referencing the knowledge in the {Medication Knowledge Block}, and combining the medical records {Structured Medical Records} and preliminary diagnostic results {Diagnostic Results}, generate medication recommendations and reasoning processes.

[0045] Chain2-2: Extracts medication recommendations from the medication projects generated by Chain2-1.

[0046] Fifth prompt template: You are a medical expert and need to complete the extraction of core medical information. Please extract medication recommendations from {Medication-Specific Generation}.

[0047] Chain3-1: Based on the structured medical records generated by Chain1-1, the relevant knowledge of the matched TopN tests, and the disease diagnosis results extracted by Chain1-3, it infers and generates test recommendations.

[0048] Sixth prompt template: You are a medical expert and need to generate a specific test and examination report. Referencing the knowledge in the {Test and Examination Knowledge Block}, and combining the medical records {Structured Medical Records} and preliminary diagnostic results {Diagnostic Results}, generate test and examination recommendations and reasoning processes.

[0049] Chain3-2: Extracts inspection and testing suggestions from the inspection and testing items generated by Chain3-1.

[0050] Seventh prompt template: You are a medical expert and need to complete the extraction of core medical information. Please extract test and examination suggestions from {Test and Examination Specialization Generation}.

[0051] This application provides a clinical decision support system that employs staged reasoning through a model chain, combined with explicit injection of medical knowledge from multiple knowledge bases, ensuring transparency in the reasoning process. Decision support is broken down into three tasks: disease diagnosis, medication recommendation generation, and laboratory test recommendation generation. Diagnosis-related knowledge is provided in the comprehensive diagnosis generation stage, medication-related knowledge in the medication-specific generation stage, and laboratory test recommendation generation stage, enabling batch injection of knowledge and improving the matching and relevance between tasks and knowledge. Furthermore, in the comprehensive diagnosis stage, although only disease diagnosis-related knowledge is provided, in addition to diagnostic suggestions, the model is required to generate further laboratory test recommendations, differential diagnoses, etiological analyses, and treatment methods, thereby improving the accuracy of disease diagnosis. The medication-specific generation and laboratory test recommendation generation stages reference the extracted preliminary diagnostic results, avoiding interference from the output of the comprehensive diagnosis stage, effectively improving the accuracy of medication and laboratory test recommendations.

[0052] Example 2

[0053] Figure 3 is a flowchart of a multi-knowledge base clinical decision-making method based on model chains provided in this application. As shown in Figure 3, the multi-knowledge base clinical decision-making method based on model chains provided in this application specifically includes:

[0054] S1. In response to clinical decision instructions, the unstructured medical record description of the patient indicated by the clinical decision instructions is converted into a structured medical record through the first model chain node, and the patient's disease prediction results are generated from the structured medical record with reference to the disease diagnosis knowledge block.

[0055] Clinical decision instructions here can be generated by doctors through the human-computer interaction interface provided by the doctor's end and sent to the inference link.

[0056] In step S1, for the unstructured medical record description input by the user, the unstructured medical record description of the patient indicated by the clinical decision instruction is first converted into a structured medical record through the first model chain node.

[0057] In one implementation, the unstructured medical record description of the patient indicated by the clinical decision instruction and the first prompt word template can be input into the first model chain sub-node to obtain the structured medical record output by the first model chain node, wherein the first prompt word template is used to indicate the preset case collection template.

[0058] Unstructured medical records typically include the patient's basic information, medical history, medical history, and auxiliary examination results. Chain1-1 can generate structured medical records for patients using a large model based on the following medical record collection template:

[0059] Outpatient medical records (all visit records):

[0060] (a) Basic patient information

[0061] Gender, age, and time of visit;

[0062] (ii) Chief Complaint

[0063] Symptoms, duration from the first onset to the present;

[0064] (iii) Present Illness

[0065] Clinical manifestations of the attack, duration of the attack, frequency of the attack, prodromal symptoms, triggers, time of day when the attack occurs (related to sleep and menstruation), name of the medication used in the past, dosage, frequency of use, efficacy of the medication, side effects of the medication, and time range of use of the medication;

[0066] (iv) Past medical history, personal history and growth and development history

[0067] Perinatal condition (whether it was full term, whether there was dystocia), medical history, history of brain surgery, history of trauma, growth and development, history of febrile seizures;

[0068] (v) Auxiliary examinations

[0069] Head MRI (date, description, conclusion); Head CT (date, description, conclusion); Head magnetoencephalography (date, description, conclusion); Electroencephalography (date, test duration, description, conclusion); Genetic testing (mutation site); Blood tests (date, test items, results); Cerebrospinal fluid examination (date, test items, results);

[0070] (vi) Diagnosis

[0071] Seizure type, syndrome;

[0072] (vii) Treatment received during this visit

[0073] Drugs (drug name, dosage, frequency of administration); neuromodulation (name of modulation technique); surgery (type of surgery);

[0074] (viii) For cases requiring long-term follow-up, add the diagnosis and treatment process.

[0075] Medications (drug name, dosage, frequency of use, efficacy [including seizure frequency and EEG changes], side effects, and duration of use); neuromodulation (treatment plan, efficacy [including seizure frequency and EEG changes], and side effects); surgery (surgery type, efficacy [including seizure frequency and EEG changes]); follow-up EEG (date, duration of examination, description, and conclusion).

[0076] For example, consider the locally deployed large model Qwen-72B and the unstructured medical record description: "Zhang, male, 10 years old, October 15, 2024. Chief complaint: paroxysmal convulsions for 6 months. Present illness: Paroxysmal convulsions began 6 months ago, with one episode in April, one in September, and one in October of this year. The convulsions occur during sleep, manifesting as left-sided facial twitching accompanied by limb jerking. Ancillary examinations: Head MRI showed no abnormalities. Video EEG: a few sharp waves in the T4 region." The corresponding structured medical record for patient Zhang would be as follows:

[0077] "**Outpatient Medical Records (All Visit Records)**"

[0078] (a) Basic patient information

[0079] Gender: Male; Age: 10 years old; Date of visit: 2024-10-15;

[0080] (ii) Chief Complaint

[0081] Symptoms: Paroxysmal seizures; Duration from first episode to present: 6 months;

[0082] (iii) Present Illness

[0083] Clinical manifestations of the attack: The attack occurs during sleep, manifesting as left-sided facial twitching accompanied by limb twitching; Duration of the attack: Unknown; Frequency of the attack: 3 attacks in the past 6 months (1 each in April, September and October of this year); Prodromal symptoms: No obvious prodromal symptoms; Trigger: Unknown; Time of onset: During sleep; Previous medication: None; Drug efficacy: None; Drug side effects: None; Duration of medication use: None;

[0084] (iv) Past medical history, personal history and growth and development history

[0085] Perinatal condition: Unknown; Medical history: None; History of brain surgery: None; History of trauma: None; Growth and development: Normal; History of febrile seizures: None;

[0086] (v) Auxiliary examinations

[0087] Head MRI (Date: Unknown; Description: No abnormalities; Conclusion: Normal); Electroencephalogram (Date: Unknown; Test duration: Unknown; Description: A few sharp waves in the T4 region; Conclusion: Abnormal);

[0088] (vi) For cases requiring long-term follow-up, add the diagnosis and treatment process.

[0089] This is the first visit of this patient, and there is currently no long-term follow-up record.

[0090] The above is the content of the outpatient medical record information collection template generated based on the provided medical record description. Please note that certain specific details, such as the specific dates of auxiliary examinations, are not listed in detail in the above medical record because they were not provided in the original description. If necessary, it is recommended to supplement the complete information to ensure the accuracy and completeness of the medical record.

[0091] After obtaining the structured medical records, you can first search for knowledge blocks. Specifically, this can be achieved through the following steps:

[0092] The structured medical records are input into the fourth model chain node to obtain the structured medical record vectors corresponding to the structured medical records output by the fourth model chain node. The similarity between the structured medical record vectors and the relevant knowledge block vectors in the preset knowledge base is calculated. The preset knowledge base includes at least a disease diagnosis knowledge base, a medication knowledge base, and a laboratory test knowledge base. Based on the similarity, the top N target knowledge block vectors are determined to identify the corresponding disease diagnosis knowledge block, medication knowledge block, and laboratory test knowledge block, respectively.

[0093] Specifically, cosine similarity can be used to calculate the similarity between vectors.

[0094] Next, the patient's disease prediction results can be extracted using the following methods:

[0095] Input the structured medical records, disease diagnosis knowledge blocks, and second prompt word templates into the second model chain sub-nodes to obtain the comprehensive diagnostic results and reasoning process output by the second model chain node.

[0096] For example, the comprehensive diagnostic structure and reasoning process corresponding to Zhang can be represented as follows:

[0097] "### Diagnostic Recommendations"

[0098] #### Seizure Types and Syndromes

[0099] Preliminary diagnosis: Focal seizures, possibly benign childhood epilepsy with centrotemporal spikes (BECTS).

[0100] Diagnostic basis:

[0101] 1. The patient is a 10-year-old male who has had three episodes in the past six months since his first episode.

[0102] 2. The attacks mainly occur during sleep, manifesting as left-sided facial twitching, accompanied by limb twitching.

[0103] 3. No obvious warning signs, and the frequency of attacks is low.

[0104] 4. The cranial MRI showed no abnormalities, but the electroencephalogram (EEG) showed a few sharp waves in the T4 region, suggesting focal abnormalities.

[0105] Confidence level of inference diagnosis: 70%

[0106] Further testing and examination

[0107] (1) Follow-up MRI of the head: to confirm whether there are any new structural abnormalities. (2) Long-term video EEG monitoring: to capture EEG during the seizure period and to clarify the seizure type and origin. (3) Genetic testing: to rule out the possibility of hereditary epilepsy.

[0108] ###Differential Diagnosis

[0109] (1) Psychogenic non-epileptic seizures: ruled out through detailed medical history and long-term video EEG monitoring. (2) Sleep disorders: especially NREM parasomnias, ruled out through long-term video EEG monitoring and sleep monitoring. (3) Metabolic causes: ruled out through blood and urine metabolic screening.

[0110] ###Treatment Plan

[0111] Drug treatment:

[0112] Oxcarbazepine: The initial dose is 10 mg / kg / day, taken orally in two divided doses. The dose should be gradually adjusted based on efficacy and tolerability.

[0113] Sodium valproate: If oxcarbazepine is ineffective, it can be considered for use. The initial dose is 10-15 mg / kg / day, divided into two oral doses.

[0114] Lifestyle adjustments:

[0115] Ensure sufficient sleep and avoid excessive fatigue. Avoid triggering factors such as bright light and noise.

[0116] Regular follow-up:

[0117] An electroencephalogram (EEG) should be performed every 3 months to assess changes in the condition. A comprehensive evaluation should be conducted every 6 months, including the effectiveness and side effects of medications.

[0118] ###Precautions

[0119] Monitor for drug side effects: Regularly check liver function, kidney function, and blood routine to ensure drug safety.

[0120] Educate parents: Teach parents how to record seizures in order to better adjust the treatment plan.

[0121] Psychological support: Provide necessary psychological support to help patients and their families cope with the stress brought about by the disease.

[0122] The above diagnosis and treatment recommendations are based on the current medical history and examination results. The subsequent treatment plan will be adjusted according to changes in the patient's condition.

[0123] Input the comprehensive diagnostic results, reasoning process, and third prompt word template into the third model chain sub-node to obtain the disease prediction results output by the third model chain node.

[0124] The third model chain node here can extract disease prediction results from the comprehensive diagnosis, which can be "focal seizures - benign childhood epilepsy with centrotemporal spikes (BECTS)".

[0125] S2. By referencing the medication knowledge block through the second model chain node, medication recommendations are given in combination with structured medical records and disease prediction results.

[0126] In step S2, medication recommendations can be given in the following ways:

[0127] The structured medical record, N medication knowledge blocks, disease prediction results, and the fourth prompt word template are input into the fourth model chain sub-node to obtain the medication reasoning process output by the fourth model chain node. The medication reasoning process includes at least a medication recommendation item.

[0128] For example, Zhang's reasoning process regarding medication can be as follows:

[0129] ### Medication Recommendations

[0130] Preliminary Diagnosis

[0131] Focal seizures - Benign childhood epilepsy with centrotemporal spikes (BECTS).

[0132] #### Medication Regimen

[0133] Medication: Carbamazepine (CBZ); Dosage: 5 mg / (kg·d), divided into 2 doses; Frequency of administration: twice daily.

[0134] #### Reasoning Process

[0135] 1. Seizure type: According to the medical history, the patient presented with seizures during sleep, with left-sided facial twitching accompanied by limb twitching, which is consistent with the characteristics of focal seizures.

[0136] 2. Electroencephalogram (EEG): The EEG showed a small number of sharp waves in the T4 region, supporting the diagnosis of focal seizures. The sharp waves were located in the centrotemporal region, which is consistent with the characteristics of benign childhood epilepsy with centrotemporal spikes (BECTS).

[0137] 3. Age: The patient is a 10-year-old male in childhood. Benign childhood epilepsy with centrotemporal spikes (BECTS) is a common focal epilepsy syndrome in childhood.

[0138] 4. (1) Drug selection: Carbamazepine (CBZ) is one of the first-line drugs for the treatment of benign epilepsy in children with centrotemporal spikes (BECTS). Carbamazepine has a good effect on focal seizures and has a high safety profile in children.

[0139] (2) Dosage: The initial dose is 5 mg / (kg·d), divided into 2 doses, and gradually adjusted to the effective dose.

[0140] 5. Monitoring: During the initial stage of treatment, close monitoring of the patient's drug response and adverse reactions is essential. If symptoms such as rash, dizziness, or nausea occur, the treatment plan should be adjusted promptly. Regular follow-up electroencephalograms (EEGs) should be performed to assess the treatment effectiveness.

[0141] 6. Precautions: (1) Carbamazepine has the potential to induce liver enzymes, and liver function should be monitored during long-term use. (2) Pay attention to drug interactions and avoid using it at the same time with other drugs that may affect liver enzymes.

[0142] ####Follow-up Recommendations

[0143] Short-term follow-up: Return for a follow-up visit after 1 month to assess drug efficacy and adverse reactions. Adjust the dosage or change the medication if necessary.

[0144] Long-term follow-up: An electroencephalogram (EEG) should be performed every 3-6 months to assess changes in the condition. If there are no seizures for more than 2 years, gradual tapering and discontinuation of medication can be considered, but this must be done under the guidance of a doctor.

[0145] Through the above medication recommendations and follow-up plan, patients' epileptic seizures can be effectively managed and controlled, improving their quality of life.

[0146] Input the medication reasoning process and the fifth prompt word template into the fifth model chain sub-node to obtain the medication advice extracted and output by the fifth model chain node from the medication advice project.

[0147] The medication recommendation for Zhang, output by the fifth model chain node, can be "Carbamazepine, 5mg / (kg·d), divided into 2 doses, twice daily".

[0148] S3. By referencing the knowledge blocks of the third model chain nodes, and combining structured medical records and disease prediction results, suggestions for examination and testing are given.

[0149] In step S3, inspection recommendations can be given in the following ways:

[0150] The structured medical record, N test and examination knowledge blocks, disease prediction results, and the sixth prompt word template are input into the sixth model chain sub-node to obtain the test and examination reasoning process output by the fourth model chain node. The test and examination reasoning process includes at least the test and examination recommendation item.

[0151] For example, the verification and inspection reasoning process for Zhang, derived from the child node of the fourth model chain, can be as follows:

[0152] "Based on the provided medical record information and epilepsy diagnostic methods, for this 10-year-old boy suspected of having focal seizures-benign childhood epilepsy with centrotemporal spikes (BECTS), the following are recommendations for further examinations and the reasoning process:"

[0153] Recommendations for further testing and examination

[0154] 1. Follow-up electroencephalogram (EEG)

[0155] Reason: The initial EEG showed a small number of sharp waves in the T4 region, suggesting abnormal discharge. However, considering that epileptiform discharges are relatively common in children and may manifest differently at different times, it is recommended to perform ambulatory electroencephalography (AEEG) or video electroencephalography (VEEG) to capture more interictal brain activity, especially during sleep, as the patient's seizures mostly occur during sleep.

[0156] Specific indicators: Pay attention to sharp wave activity in the central temporal region, especially changes during sleep, and whether there are abnormal discharges in other areas.

[0157] 2. Follow-up MRI of the head

[0158] Reason: Although the initial head MRI showed no abnormalities, considering that BECTS may be accompanied by minor structural abnormalities, it is recommended to use more detailed scanning parameters (such as 3.0T MRI, 3D). Weighted imaging and 3D FLAIR sequences, with particular attention to the structures of the temporal lobe and central region, are used to rule out subtle cortical dysplasia or other structural lesions.

[0159] Specific indicators: Focus on structural and signal changes in the hippocampus, central sulcus, and temporal lobe.

[0160] 3. Blood test

[0161] Reason: Although there is no clear history of drug use, blood tests can help rule out metabolic abnormalities, infections, or other systemic diseases that may trigger or worsen seizures.

[0162] Specific indicators include complete blood count, blood glucose, electrolytes, liver and kidney function, pyruvate, lactate, and, when necessary, screening for inherited metabolic diseases.

[0163] ### Reasoning Process

[0164] 1. Follow-up EEG:

[0165] Preliminary examination results: A small number of sharp waves in the T4 region, indicating the presence of focal abnormal discharge.

[0166] Necessity of further examination: Ambulatory EEG or video EEG can capture brain activity more comprehensively, especially changes during attacks and during sleep, which is crucial for the diagnosis of BECTS.

[0167] Expected goal: To confirm whether the abnormal discharge in the central temporal region is persistent, and to identify other possible areas of abnormal discharge.

[0168] 2. Follow-up MRI of the head:

[0169] Preliminary examination results: No abnormalities were found, but BECTS may be accompanied by minor structural abnormalities.

[0170] Necessity of further examination: Using more detailed scanning parameters can improve the sensitivity and specificity of the detection, and help to discover potential subtle lesions.

[0171] Expected outcome: To rule out structural abnormalities in the temporal lobe and central region, such as cortical dysplasia and hippocampal sclerosis.

[0172] 3. Blood test:

[0173] Preliminary examination results: No obvious abnormalities.

[0174] Further investigation is necessary to rule out metabolic abnormalities, infections, or other systemic diseases that may trigger or exacerbate seizures.

[0175] Expected outcome: To ensure that there are no other underlying causes affecting the patient's seizures.

[0176] These further examinations can provide a more comprehensive assessment of the patient's condition, offering a scientific basis for diagnosis and treatment.

[0177] Input the inspection and verification reasoning process and the seventh prompt word template into the seventh model chain sub-node to obtain the inspection and verification suggestions extracted and output by the seventh model chain node from the inspection and verification suggestion project.

[0178] For example, the examination and testing suggestions for Zhang output by the seventh model chain sub-node could be "electroencephalogram examination, head MRI re-examination, blood test".

[0179] S4. Feed back the disease prediction results, medication recommendations, and test recommendations to the doctor for display, in order to assist the doctor in making clinical decisions for the patient.

[0180] Doctors can combine the disease prediction results, medication recommendations, and test results they obtain with their own experience to make clinical diagnoses and treat patients.

[0181] This application provides a multi-knowledge base clinical decision-making method based on model chains. The designed model chain breaks down decision support into three tasks: disease diagnosis, medication recommendation generation, and test and examination recommendations. The model chain performs staged reasoning and combines multiple knowledge bases to inject medical knowledge in batches. The reasoning process is transparent and can better collaborate with doctors' experience, thereby assisting doctors in making more accurate clinical decisions. At the same time, it can also make diagnostic decisions more interpretable and better applicable to individuals with different differences, with good flexibility.

[0182] In one implementation, after obtaining the structured medical record in step S1, the structured medical record can be fed back to the doctor and displayed through a human-computer interaction interface. The structured medical record display interface allows for the segmentation of the structured medical record through editable text boxes.

[0183] In this way, doctors can add to and verify the structured medical records, and also modify them based on their own experience. Therefore, in steps S2 and S3, the medication and laboratory test recommendations generated based on the modified structured medical records are more tailored to the patient's actual situation, further ensuring the decision support capability of this application even when there are significant individual differences among patients.

[0184] Example 3

[0185] Please refer to Figure 4, which is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. As shown in Figure 4, based on the same inventive concept, an electronic device 400 is also provided in this embodiment of the application. The electronic device 400 includes a processor 410, a memory 420, and a bus 430.

[0186] The memory 420 stores machine-readable instructions that can be executed by the processor 410. When the electronic device 400 is running, the processor 410 and the memory 420 communicate through the bus 430. When the machine-readable instructions are executed by the processor 410, the steps of a multi-knowledge base clinical decision-making method based on a model chain, as described in the above method embodiment, can be executed. For specific implementation details, please refer to the method embodiment, which will not be repeated here.

[0187] This application also provides a computer-readable storage medium storing a computer program. When the computer program is run by a processor, it can execute the steps of a multi-knowledge base clinical decision-making method based on a model chain as described in the above method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0188] Those skilled in the art will 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.

[0189] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0190] Furthermore, 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.

[0191] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0192] It should be noted that if the function is implemented as a software module 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 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, a server, or a network device, etc.) to execute all or part of the steps of the methods described in 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.

[0193] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0194] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A multi-knowledge-base clinical decision-making method based on model chains, characterized in that, The method includes: responding to a clinical decision instruction, converting the unstructured medical record description of the patient indicated by the clinical decision instruction into a structured medical record through a first model chain node, and generating a disease prediction result for the patient from the structured medical record with reference to a disease diagnosis knowledge block; providing medication recommendations by referring to a medication knowledge block through a second model chain node, combining the structured medical record and the disease prediction result; providing laboratory test recommendations by referring to a laboratory test knowledge block through a third model chain node, combining the structured medical record and the disease prediction result; and feeding back the disease prediction result, the medication recommendations, and the laboratory test recommendations to the doctor for display, in order to assist the doctor in generating clinical decisions for the patient; the first model chain node includes a first model chain sub-node, and the step of converting the unstructured medical record description of the patient indicated by the clinical decision instruction into a structured medical record through the first model chain node specifically includes: inputting the unstructured medical record description of the patient indicated by the clinical decision instruction and a first prompt word template into the first model chain sub-node to obtain the structured medical record output by the first model chain sub-node. The medical record, wherein the first prompt word template is used to indicate a preset case collection template; the structured medical record is input into a fourth model chain node to obtain a structured medical record vector corresponding to the structured medical record output by the fourth model chain node; the similarity between the structured medical record vector and the relevant knowledge block vectors in a preset knowledge base is calculated, wherein the preset knowledge base includes at least a disease diagnosis knowledge base, a medication knowledge base, and a laboratory test knowledge base; based on the similarity, the top N target knowledge block vectors are determined to determine the corresponding disease diagnosis knowledge block, medication knowledge block, and laboratory test knowledge block respectively; the first model chain node also includes a second model chain sub-node and a third model chain node, and the patient's disease prediction results are extracted in the following way: the structured medical record, the disease diagnosis knowledge block, and the second prompt word template are input into the second model chain sub-node to obtain the comprehensive diagnosis result and reasoning process output by the second model chain node; the comprehensive diagnosis result and reasoning process and the third prompt word template are input into the third model chain node to obtain the disease prediction result output by the third model chain node.

2. The method according to claim 1, characterized in that, The second model chain node includes a fourth model chain sub-node and a fifth model chain node, and provides medication recommendations in the following way: The structured medical record, N medication knowledge blocks, disease prediction results, and the fourth prompt word template are input into the fourth model chain sub-node to obtain the medication reasoning process output by the fourth model chain node. The medication reasoning process includes at least a medication recommendation item. The medication reasoning process and the fifth prompt word template are input into the fifth model chain node to obtain the medication recommendations extracted and output by the fifth model chain node from the medication recommendation item.

3. The method according to claim 1, characterized in that, The third model chain node includes a sixth model chain sub-node and a seventh model chain node, and provides examination and testing suggestions in the following way: The structured medical record, N examination and testing knowledge blocks, disease prediction results, and the sixth prompt word template are input into the sixth model chain sub-node to obtain the examination and testing reasoning process output by the fourth model chain node. The examination and testing reasoning process includes at least the examination and testing suggestion item; The examination and testing reasoning process and the seventh prompt word template are input into the seventh model chain node to obtain the examination and testing suggestions extracted and output by the seventh model chain node from the examination and testing suggestion item.

4. The method according to claim 1, characterized in that, The disease diagnosis knowledge base is constructed as follows: Preprocessing the target disease guidelines to obtain standard-format disease guideline text; decomposing the standard-format disease guideline text into knowledge slices to obtain disease diagnosis knowledge slices; inputting each disease diagnosis knowledge slice into a fourth model chain node to obtain the corresponding disease diagnosis knowledge block vector output by the fourth model chain node; and associating and saving all disease diagnosis knowledge block vectors and their corresponding disease diagnosis knowledge slices to construct the disease diagnosis knowledge base.

5. The method according to claim 4, characterized in that, The steps of preprocessing the target disease guideline to obtain the standard format disease guideline text specifically include: classifying the text and tables in the target disease guideline and converting them into Markdown format; inputting the images and the eighth prompt word template in the target disease guideline into the fifth model chain node to obtain the image description text output by the fifth model chain node, and converting the image description text into Markdown format.

6. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the model-chain-based multi-knowledge-base clinical decision-making method as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the model chain-based multi-knowledge base clinical decision-making method as described in any one of claims 1 to 5.

Images