Drug and disease correlation analysis method and system based on multi-modal feature fusion

Abstract

The invention provides a multi-modal feature fusion-based drug and disease correlation analysis method and system, and belongs to the field of drug correlation analysis. The method comprises the following steps: in response to a drug disease correlation analysis instruction sent by a doctor terminal, calling drug information associated with the drug disease correlation analysis instruction and patient characteristics of a target patient; acquiring a historical patient case set based on the disease characteristics and the medicine information of the patient; evaluating and acquiring a first correlation degree and a second correlation degree of each candidate drug option; and performing recommendation evaluation on each candidate drug option through the first association degree and the second association degree of each candidate drug option to obtain a comprehensive recommendation degree of each candidate drug option, and feeding back the comprehensive recommendation degree to the doctor terminal. The technical problem of insufficient drug selection accuracy in the prior art is solved, and the technical effect of improving the drug selection accuracy through multi-modal feature fusion and double-layer correlation analysis is achieved.

Description

Technical Field

[0001] This invention relates to the field of drug association analysis, and more particularly to a method and system for drug-disease association analysis based on multimodal feature fusion. Background Technology

[0002] In clinical treatment, drug selection is a crucial factor affecting treatment outcomes. When developing a treatment plan, physicians typically need to select the most suitable medication regimen from multiple candidate drugs, taking into account factors such as the patient's disease type, severity, and individual differences. However, due to the complexity and diversity of individual patient conditions, different patients respond significantly differently to the same drug, making drug selection a challenging task.

[0003] Currently, clinical medication decisions primarily rely on physicians' clinical experience and disease treatment guidelines. After establishing a treatment framework, physicians often make judgments based on past clinical experience when faced with multiple candidate drugs. This approach has the following shortcomings: First, physicians struggle to comprehensively consider the multidimensional individual characteristics of patients, such as age, gender, and liver and kidney function status; second, while the existing medical system has accumulated a large amount of historical medication data and treatment outcome information, this data lacks effective analysis and utilization methods. Physicians cannot quickly obtain historical medication experiences similar to those of their current patients for reference, leading to insufficient accuracy in drug selection. Summary of the Invention

[0004] This invention addresses the technical problem of insufficient accuracy in drug selection in existing technologies by providing a method and system for analyzing the correlation between drugs and diseases through multimodal feature fusion.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides a method for drug-disease correlation analysis based on multimodal feature fusion, comprising: responding to a drug-disease correlation analysis command issued by a doctor's terminal, retrieving drug information and patient characteristics of a target patient associated with the drug-disease correlation analysis command; wherein, the patient characteristics include patient basic characteristics and patient disease characteristics, the drug information includes multiple identified drugs and a drug item to be identified, the drug item to be identified having multiple candidate drug options; obtaining a historical patient case set based on the patient disease characteristics and the drug information; evaluating and obtaining a first correlation degree for each of the candidate drug options based on the historical patient case set, and evaluating and obtaining a second correlation degree for each of the candidate drug options based on the historical patient case set and the patient basic characteristics; performing a recommendation evaluation on each of the candidate drug options based on the first correlation degree and the second correlation degree to obtain a comprehensive recommendation degree for each of the candidate drug options, and feeding it back to the doctor's terminal.

[0007] Secondly, the present invention provides a drug-disease correlation analysis system based on multimodal feature fusion, comprising: a data retrieval module, used to retrieve drug information and patient characteristics of a target patient associated with the drug-disease correlation analysis command issued by a doctor's terminal; wherein the patient characteristics include patient basic characteristics and patient disease characteristics, the drug information includes multiple identified drugs and one unidentified drug item, the unidentified drug item having multiple candidate drug options; a case acquisition module, used to acquire a historical patient case set based on the patient disease characteristics and the drug information; a correlation evaluation module, used to evaluate and obtain a first correlation degree for each candidate drug option based on the historical patient case set, and evaluate and obtain a second correlation degree for each candidate drug option based on the historical patient case set and the patient basic characteristics; and a comprehensive recommendation module, used to evaluate and recommend each candidate drug option based on the first and second correlation degrees of each candidate drug option, obtain a comprehensive recommendation degree for each candidate drug option, and feed it back to the doctor's terminal.

[0008] The beneficial effects of this invention are:

[0009] In response to a drug-disease correlation analysis command issued by the doctor's terminal, the system retrieves drug information associated with the command and patient characteristics of the target patient. The patient characteristics include basic patient characteristics and disease characteristics. The drug information includes multiple identified drugs and one unidentified drug, with multiple candidate drug options. By acquiring the patient's multi-dimensional characteristics and candidate drug information, foundational data is provided for subsequent personalized correlation analysis. Based on the patient's disease characteristics and the drug information, a historical patient case set is obtained. By retrieving historical cases with similar diseases to the current patient who have used related drugs, real-world medication experience data is provided for correlation assessment. A first correlation degree is obtained for each candidate drug option based on the historical patient case set, and a second correlation degree is obtained for each candidate drug option based on the historical patient case set and the patient's basic characteristics. The first correlation degree reflects the overall effectiveness of the candidate drug for the disease, and the second correlation degree reflects the personalized efficacy of the candidate drug for a patient group similar to the current patient, thus establishing a two-tiered correlation assessment system. Each candidate drug option is evaluated using a first and second correlation coefficient to obtain a comprehensive recommendation score, which is then fed back to the physician's terminal. By integrating the correlation coefficients from both overall effectiveness and personalized efficacy dimensions, a comprehensive recommendation result is generated and fed back to the physician, providing a quantitative reference for clinical medication decisions.

[0010] Through the above technical solution, the present invention realizes drug recommendation based on multimodal feature fusion and two-layer correlation analysis, thereby improving the accuracy of drug selection. Attached Figure Description

[0011] Figure 1 A flowchart illustrating the drug-disease association analysis method based on multimodal feature fusion provided by this invention;

[0012] Figure 2 This is a schematic diagram of the structure of the drug-disease correlation analysis system based on multimodal feature fusion provided by the present invention.

[0013] In the attached diagram, the components represented by each number are as follows:

[0014] Data retrieval module 11, case acquisition module 12, correlation assessment module 13, and comprehensive recommendation module 14. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0016] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0017] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0018] Example 1, as Figure 1As shown, embodiments of the present invention provide a method for analyzing the association between drugs and diseases through multimodal feature fusion, including:

[0019] S1. In response to the drug-disease correlation analysis command issued by the doctor's terminal, retrieve the drug information and patient characteristics of the target patient associated with the drug-disease correlation analysis command; wherein, the patient characteristics include the patient's basic characteristics and the patient's disease characteristics, and the drug information includes multiple identified drugs and one unidentified drug item, the unidentified drug item having multiple candidate drug options.

[0020] Specifically, the system first receives a drug-disease correlation analysis command from the doctor's terminal. In actual clinical practice, when developing a treatment plan for a patient, doctors typically first determine the treatment framework based on the disease type and treatment guidelines, clarifying the required drug categories and general medication regimen. During this process, doctors may have already identified most of the drugs in the treatment plan, but for a specific drug position, they face the dilemma of choosing from multiple available options. For example, in treating hypertension, the doctor may have determined that a combination of antihypertensive drugs, diuretics, and lipid-lowering drugs is needed, and has already selected several specific drugs. However, for a certain class of drugs, there are still multiple candidate options, such as valsartan, irbesartan, and losartan, which belong to the angiotensin receptor antagonist class. In this case, the doctor needs to choose the most suitable drug for the patient from these candidate drugs. To address this situation, the doctor initiates a drug-disease correlation analysis command through the doctor's terminal, requesting auxiliary decision support.

[0021] In response to the drug-disease correlation analysis command, two types of core data relevant to the analysis task are retrieved: drug information and patient characteristics. Patient characteristics include two parts: basic patient characteristics and patient disease characteristics. Basic patient characteristics refer to fundamental information reflecting individual patient differences, including but not limited to structured data such as age, gender, weight, height, body mass index, liver function indicators, kidney function indicators, blood lipid levels, blood glucose levels, and genotype information. These basic characteristics characterize the patient's physiological state and individual differences from multiple dimensions, providing a basis for subsequent personalized analysis. Patient disease characteristics refer to characteristic information reflecting the patient's current disease state. The data sources are multimodal, including disease diagnosis text data and medical imaging data. Disease diagnosis text data includes diagnostic information in natural language forms such as disease name, disease description, symptom records, and medical history information; medical imaging data includes visualized disease feature data such as CT images, X-rays, ultrasound images, and MRI images. By fusing and encoding these multimodal disease data, a comprehensive representation of the patient's disease state can be obtained.

[0022] Drug information comprises two parts: multiple identified drugs and a list of drugs to be identified. The multiple identified drugs refer to those explicitly selected by the physician during the treatment plan development process; these drugs form the foundation of the treatment plan. The list of drugs to be identified refers to the position of drugs in the treatment plan that have not yet been finalized and require selection from multiple candidate drugs. This list of drugs to be identified has multiple candidate drug options, meaning there are several drugs available. These candidate drugs belong to the same pharmacological classification and are similar or substitutable in terms of mechanism of action and indications, but differ in specific efficacy, safety, metabolic characteristics, and applicable populations. For example, multiple candidate drugs could be products from different manufacturers of the same generic drug, or different compounds within the same drug class.

[0023] By retrieving the aforementioned drug information and patient characteristics, the necessary data support for correlation analysis was obtained, laying the foundation for intelligent recommendations based on historical data and individual patient characteristics in subsequent steps.

[0024] S2. Based on the patient's disease characteristics and the drug information, obtain a historical patient case set.

[0025] Specifically, based on patient disease characteristics and medication information, a historical patient case set is retrieved from a historical patient case database. This database stores a large amount of past patient medical records, including disease information, medication regimens, and treatment processes. This historical data records the actual treatment effects of different patients using different drug combinations in real clinical settings, containing rich information on medication experience. However, the data volume in historical case databases is typically enormous, making direct analysis and calculation of the entire dataset inefficient and lacking specificity. Therefore, it is necessary to select a subset of cases highly relevant to the current patient and medication scenario from the massive historical case database to form the historical patient case set. The historical patient case set needs to meet the following characteristics: First, the disease characteristics of historical patients should have a high degree of similarity with the disease characteristics of the current target patients, that is, the diseases suffered by historical patients are comparable to those of the current patients; second, the medication regimens of historical patients should be relevant to the current analysis scenario. Specifically, the medication regimens of historical patients should include multiple identified drugs from step S1 to ensure the consistency of the drug combination; third, the medication regimens of historical patients should include at least one candidate drug option for the drug to be identified, so as to provide data support for the efficacy evaluation of different candidate drugs.

[0026] To improve retrieval efficiency, Locality Sensitive Hashing (LSH) is employed to index patient disease features. Specifically, LSH is performed on the disease features of the target patient, mapping the patient's disease feature vectors to specific hash buckets. LSH is a hashing method that preserves data similarity; similar feature vectors are mapped to the same or adjacent hash buckets. By determining the target hash bucket to which the target patient belongs and its neighboring hash buckets, candidate historical cases with similar disease features to the target patient can be quickly located without traversing the entire database. Historical patient cases located within the target hash bucket and its neighboring hash buckets are retrieved from the historical patient case database, forming an initial candidate case set.

[0027] After obtaining the candidate case set, further refined screening is performed based on patient disease characteristics and drug information. Screening criteria include: the similarity between the disease characteristics of historical patients and those of the target patients must exceed a preset similarity threshold to ensure comparability; the medication regimens of historical patients must include multiple identified drugs from step S1 to ensure consistency in medication base; and the medication regimens of historical patients must include at least one candidate drug option for the drug to be identified, providing a data basis for assessing the correlation between candidate drugs. After the above screening, historical patient cases that simultaneously meet all screening criteria are retained, ultimately forming the historical patient case set.

[0028] Through the above retrieval and filtering process, a set of cases highly relevant to the current analysis task can be efficiently obtained from massive historical data, providing a reliable data foundation for correlation assessment in subsequent steps.

[0029] S3. Evaluate the first correlation degree of each candidate drug option based on the historical patient case set, and evaluate the second correlation degree of each candidate drug option based on the historical patient case set and the patient's basic characteristics.

[0030] Specifically, based on the historical patient case set obtained in step S2, a two-layer correlation assessment is performed on each candidate drug option, and the first correlation degree and the second correlation degree are calculated respectively.

[0031] The first correlation reflects the overall association between the candidate drug and the disease, that is, the overall therapeutic effectiveness of the candidate drug for this type of disease. Specifically, for each candidate drug option, all historical patient cases that used that candidate drug option are selected from the historical patient case set, forming a sub-case set corresponding to that candidate drug. In this sub-case set, the number of cases with effective treatment and the total number of cases are counted. The determination of treatment effectiveness can be based on the treatment outcome annotation information recorded in the case or on the efficacy evaluation criteria determined according to the disease type. The original efficacy rate of the candidate drug is obtained by calculating the ratio of the number of cases with effective treatment to the total number of cases.

[0032] Considering the potentially significant differences in sample sizes among different candidate drugs in historical data, the reliability of efficacy statistics for candidate drugs with smaller sample sizes is lower. To address this issue, a confidence coefficient is introduced to adjust the original efficacy rate. The total number of cases in the sub-case set is compared to a pre-set baseline sample size. When the sample size is less than the baseline, the confidence coefficient is reduced proportionally; when the sample size reaches or exceeds the baseline, the confidence coefficient is 1. Multiplying the original efficacy rate by the confidence coefficient yields the first correlation of the candidate drug option. In this way, the first correlation not only reflects the efficacy level of the candidate drug but also the reliability of the statistical results.

[0033] The second correlation reflects the matching degree between the candidate drug and the individual characteristics of the patient, that is, the personalized treatment effect of the candidate drug on a patient group similar to the current patient. Because different patients differ in basic characteristics such as age, gender, physical condition, liver and kidney function, and genotype, the efficacy of the same drug may vary significantly among different patients. Therefore, considering only the overall effectiveness of the drug is insufficient; it is also necessary to evaluate the drug's efficacy in patients with similar basic characteristics to the current patient. Specifically, for each candidate drug option, based on the corresponding sub-case set, historical patient cases with similar basic characteristics to the current target patient are further screened. By calculating the similarity between the basic characteristics of historical patients and the basic characteristics of the target patient, historical patient cases with similarity exceeding a preset threshold are selected to form a similar case set. The similarity calculation comprehensively considers various dimensions of patient basic characteristics, such as age differences, gender matching, and the degree of similarity in liver and kidney function status. In the similar case set, the number of cases with effective treatment and the total number of cases are counted, and the ratio between the two is calculated to obtain the second correlation of the candidate drug option. The second correlation reflects the actual efficacy of the candidate drug in a patient population highly similar to the current patients, and has stronger personalized reference value.

[0034] A two-tiered correlation assessment system was established by calculating the first and second correlation degrees. The first correlation degree reflects the overall therapeutic effect of the candidate drug on the disease at a macro level, while the second correlation degree reflects the personalized efficacy of the candidate drug for a specific patient group at a micro level. The two dimensions of correlation complement each other, jointly providing a quantitative basis for subsequent comprehensive recommendation and evaluation.

[0035] S4. The candidate drug options are evaluated and recommended based on the first and second correlation of each candidate drug option to obtain the comprehensive recommendation of each candidate drug option, and then fed back to the doctor's terminal.

[0036] Furthermore, based on the first and second correlation degrees of each candidate drug option obtained in step S3, a comprehensive recommendation evaluation is performed on each candidate drug option to generate a comprehensive recommendation score and feed it back to the doctor's terminal.

[0037] The first and second correlation scores reflect the applicability of candidate drugs from different perspectives. The first correlation score represents the overall therapeutic efficacy of the candidate drug for this type of disease, reflecting the drug's efficacy level in a general sense; the second correlation score represents the personalized efficacy of the candidate drug for a patient population similar to the current patients, reflecting the degree of suitability of the drug for specific patient characteristics. Each has its own emphasis and needs to be integrated to form a comprehensive recommendation.

[0038] However, the reliability and importance of the first and second correlation scores differ under different data conditions. When there is a large number of historical cases of similar patients, the second correlation score, based on sufficient personalized samples, has high statistical reliability, and should be given a higher weight. Conversely, when there is a small number of historical cases of similar patients, the statistical basis of the second correlation score is weak, and its reliability decreases. In this case, the weight of the second correlation score should be reduced, and more reliance should be placed on the overall effectiveness reflected by the first correlation score.

[0039] Therefore, a dynamic weighting calculation mechanism was designed. Specifically, for each candidate drug option, an initial second weighting coefficient is calculated based on the total number of cases in its corresponding similar case set, i.e., the second total number of cases. The second weighting coefficient is calculated by dividing the second total number of cases by the sum of the second total number of cases and the preset base total number of cases, i.e., initial second weighting coefficient = second total number of cases / (second total number of cases + base total number of cases). This calculation method makes the second weighting coefficient increase with the increase of the sample size of similar patients, reflecting the impact of sample size on the reliability of personalized assessment.

[0040] To avoid excessive reliance on personalized data due to an overly high second weighting coefficient, a preset upper limit weighting value was set. The initial second weighting coefficient was compared with the preset upper limit weighting value, and the smaller of the two was taken as the final second weighting coefficient. This design ensures that even with a large sample size of similar patients, the weight of the second association will not grow indefinitely, still retaining a certain proportion of overall validity for reference.

[0041] After determining the second weighting coefficient, the first weighting coefficient is calculated. Since both the first and second weighting coefficients need to satisfy a normalization condition (i.e., their sum is 1), the first weighting coefficient equals 1 minus the second weighting coefficient. In this way, the two weighting coefficients form a complementary relationship, jointly determining the fusion ratio of the correlation between the two dimensions.

[0042] Based on the calculated first and second weight coefficients, a weighted fusion of the first and second correlation scores is performed. Specifically, the first correlation score is multiplied by the first weight coefficient, and the second correlation score is multiplied by the second weight coefficient; the two are then added together to obtain the overall recommendation score for the candidate drug option. The overall recommendation score integrates the evaluation results of both the overall effectiveness and personalized suitability of the drug, and an adaptive adjustment under different data conditions is achieved through a dynamic weighting mechanism.

[0043] Repeat the above calculation process for all candidate drug options to obtain the overall recommendation score for each candidate drug option. Sort the candidate drug options according to the overall recommendation score, and send the ranking along with the overall recommendation score for each candidate drug option to the doctor's terminal.

[0044] After receiving the recommendations, the doctor's terminal displays the ranking and correlation data of each candidate drug on the interface. The doctor can then make a final medication decision based on the recommendations, the patient's actual condition, and their own clinical experience.

[0045] Through the above steps, the process from patient feature collection, historical case retrieval, two-layer correlation assessment to comprehensive recommendation generation was completed, realizing intelligent drug recommendation based on multimodal feature fusion and two-layer correlation analysis, providing auxiliary support for clinical medication decision-making.

[0046] Furthermore, retrieving drug information and patient characteristics associated with the drug-disease correlation analysis command includes:

[0047] S11. Based on the preset basic information collection list, obtain the basic patient characteristics corresponding to the target patient from the electronic medical record system;

[0048] S12. Obtain disease-related multimodal data of the target patient, wherein the disease-related multimodal data includes disease diagnosis text data and medical image data;

[0049] S13. The disease diagnosis text data and medical image data are encoded separately and then fused through a multimodal fusion network to obtain the patient's disease characteristics;

[0050] S14. Summarize the patient's basic characteristics and the patient's disease characteristics as the patient characteristics;

[0051] S15. Retrieve the multiple confirmed drugs and multiple candidate drug options corresponding to the drugs to be determined from the prescription system to obtain drug information.

[0052] In a preferred embodiment, firstly, the basic patient characteristics corresponding to the target patient are obtained from the electronic medical record system according to a preset basic information collection list. The preset basic information collection list is a pre-configured list of standardized data fields that specifies the specific items of basic patient information to be collected. This list typically includes categories such as demographic characteristic fields, physiological indicator fields, and laboratory test fields. Specifically, demographic characteristic fields include basic information such as age, gender, height, weight, and ethnicity; physiological indicator fields include vital signs data such as blood pressure, heart rate, and body temperature; laboratory test fields include liver function indicators such as alanine aminotransferase (ALT), aspartate aminotransferase (AST), and total bilirubin (TBIL); kidney function indicators such as serum creatinine (Scr), blood urea nitrogen (BUN), and glomerular filtration rate (eGFR); blood lipid indicators such as total cholesterol (TC), low-density lipoprotein (LDL), high-density lipoprotein (HDL), and triglycerides (TG); and blood glucose indicators such as fasting plasma glucose (FPG) and glycated hemoglobin (HbA1c). By connecting to the electronic medical record system via a data interface, and according to the field names and data formats defined in the preset basic information collection list, the corresponding data values ​​are extracted from the target patient's electronic medical record. For example, for a 62-year-old male patient with hypertension, the extracted basic patient characteristics might include: age 62, sex male, weight 75 kg, BMI 26.5, blood pressure 150 / 95 mmHg, ALT 42 U / L, Scr 88 μmol / L, eGFR 72 mL / min, TC 5.8 mmol / L, LDL 3.6 mmol / L, etc.

[0053] Next, acquire disease-related multimodal data for the target patient. This data includes two different modalities: disease diagnosis text data and medical imaging data. Disease diagnosis text data is disease-related information recorded in natural language, including diagnosis results, symptom descriptions, medical history, and physical examination results. For example, for the aforementioned hypertensive patient, the disease diagnosis text data might include: "Grade 3 primary hypertension, very high risk. The patient has complained of dizziness and headaches for 3 years, with a highest blood pressure of 180 / 110 mmHg. There is a 5-year history of type 2 diabetes, currently controlled with oral hypoglycemic agents. Physical examination: no obvious abnormalities were found on cardiac auscultation, and no edema in the lower extremities." Medical imaging data is visualized disease feature data acquired through medical imaging equipment, including but not limited to CT images, X-rays, ultrasound images, MRI images, and electrocardiograms. For cardiovascular disease patients, medical imaging data might include echocardiograms showing left ventricular hypertrophy, carotid ultrasound images showing the degree of intimal thickening, and electrocardiograms showing heart rhythm. For cancer patients, medical imaging data may include CT images showing the location, size, shape, and other characteristics of the tumor.

[0054] Subsequently, the disease diagnosis text data and medical image data are encoded separately and then fused using a multimodal fusion network to obtain patient disease characteristics. The detailed process of this step will be further elaborated in subsequent descriptions. Next, the patient baseline characteristics obtained in step S11 and the patient disease characteristics obtained in step S13 are summarized as complete patient characteristics. Patient baseline characteristics reflect individual differences and physiological states, while patient disease characteristics reflect the patient's disease state and severity. The combination of these two characteristics forms a comprehensive feature description of the patient, providing multi-dimensional data support for subsequent correlation analysis.

[0055] Simultaneously, the system retrieves multiple candidate drug options corresponding to the doctor's prescribed medications and pending medications from the prescription system to obtain drug information. The prescription system is an important component of the hospital information system, recording prescription information issued by doctors to patients. Through a data interface with the prescription system, the system obtains the current patient's prescription draft or prescription intention. In this prescription, multiple prescribed medications are drugs that the doctor has clearly selected. For example, for a patient with hypertension, the doctor may have determined to use hydrochlorothiazide as a diuretic, amlodipine as a calcium channel blocker, and aspirin as an antiplatelet drug. Pending medications are drug positions that the doctor has not yet finalized, such as the primary antihypertensive drug. These pending medications correspond to multiple candidate drug options, which may include multiple angiotensin receptor blockers such as valsartan 80mg, irbesartan 150mg, and losartan 50mg. Detailed information on these candidate drugs is obtained, including the generic name, brand name, strength, dosage form, and manufacturer.

[0056] By acquiring drug information and patient characteristics, the foundation is laid for intelligent recommendations based on historical data and individual patient characteristics in subsequent steps.

[0057] Furthermore, the disease diagnosis text data and medical image data are encoded separately and then fused using a multimodal fusion network to obtain patient disease characteristics, including:

[0058] S131. The disease diagnosis text data is processed by a text data encoding analyzer to obtain a text feature vector, wherein the text data encoding analyzer is trained and generated based on the sample disease diagnosis text dataset and the sample text feature vector set.

[0059] S132. The medical image data is processed by an image data encoding analyzer to obtain image feature vectors, wherein the image data encoding analyzer is trained and generated based on a sample medical image dataset and a sample image feature vector set.

[0060] S133. Input the text feature vector and image feature vector into a multimodal fusion network. The multimodal fusion network adopts a gated fusion mechanism, dynamically generates fusion weights based on the correlation between the text feature vector and the image feature vector, and performs weighted fusion of the text feature vector and the image feature vector to obtain the patient's disease features.

[0061] In a preferred embodiment, a text data encoding analyzer is used to process disease diagnosis text data to obtain text feature vectors. The text data encoding analyzer is a deep learning-based natural language processing model that converts textual disease information into a computer-processable numerical vector form. The training process of this encoding analyzer is as follows: First, a large sample disease diagnosis text dataset is collected. This dataset contains text samples of real diagnosis records, symptom descriptions, and medical history information for various diseases. For example, the dataset may contain a large number of diagnostic text records covering multiple fields such as cardiovascular diseases, respiratory diseases, and endocrine diseases. Each record is a real diagnostic description written by doctors in clinical practice. Then, these sample disease diagnosis text data are labeled and processed to generate corresponding sample text feature vectors, resulting in a sample text feature vector set. The sample text feature vector is a numerical representation of the semantic information of the text and can be a fixed-dimensional real-number vector, such as 512-dimensional or 768-dimensional. The process of generating sample text feature vectors can be achieved through semantic understanding and feature extraction of the text by medical experts. After obtaining the sample disease diagnosis text dataset and the sample text feature vector set, a deep learning method is used to train the text data encoding analyzer. During training, the model learns how to extract semantic information from the input text and map it into feature vectors, ensuring that feature vectors corresponding to texts with similar meanings are close in the vector space. The text data encoding analyzer can be implemented using a pre-trained language model based on the Transformer architecture, such as BERT or BioBERT. The core structure of this type of model includes word embedding layers, multi-layer self-attention mechanisms, and feedforward neural networks. The specific workflow is as follows: When a disease diagnosis text is input, such as "primary hypertension grade 3, very high risk group, with left ventricular hypertrophy," the text data encoding analyzer first segments the text into tokens using a word segmenter, such as "primary," "hypertension," "grade 3," "very high risk group," "with," and "left ventricular hypertrophy." Then, each token is converted into an initial vector representation through a word embedding layer. Next, these token vectors enter a multi-layer Transformer encoder. In the Transformer encoder, the self-attention mechanism allows each token to pay attention to all other tokens in the text, thereby capturing the contextual relationships and semantic dependencies between tokens. For example, it can be understood that "Level 3" modifies the severity of "hypertension," "very high-risk group" describes risk stratification, and "left ventricular hypertrophy" is information about complications. Through multi-layered feature extraction and semantic understanding, a deep semantic representation of the entire text is gradually constructed. Finally, a fixed-dimensional text feature vector that integrates the semantic information of the entire text is output.Each dimension of the text feature vector corresponds to a semantic feature of a certain aspect of the text. For example, some dimensions may be related to the disease type, some to the severity, and some to complications. The overall vector can accurately represent the disease information of the input text.

[0062] Simultaneously, the medical image data is processed by an image data encoding analyzer to obtain image feature vectors. The image data encoding analyzer is a deep learning-based computer vision model whose function is to extract visual feature information from medical images and convert it into numerical vector form. The training process of this encoding analyzer is as follows: First, a large sample medical image dataset is collected, containing medical image samples of various diseases, such as cardiac ultrasound images, lung CT images, and carotid artery ultrasound images. The images in the sample medical image dataset were collected by professional radiologists, covering cases of different disease types and severity. Then, these sample medical image data are labeled and feature extracted to generate corresponding sample image feature vectors, resulting in a sample image feature vector set. The labeling process involves radiologists marking and describing key regions in the images. The sample image feature vectors are also fixed-dimensional real-number vectors. After obtaining the sample medical image dataset and the corresponding sample image feature vector set, a deep learning method is used to train the image data encoding analyzer. During training, it learns how to extract key visual features from input images and encode them into feature vectors. This image data encoding and analysis tool can be implemented using a convolutional neural network architecture, such as the ResNet or DenseNet model. The core structure of such models includes multiple convolutional layers, pooling layers, and fully connected layers. The specific workflow is as follows: When a medical image, such as a cardiac ultrasound image, is input, the basic visual features of the image are first extracted through initial convolutional layers. These convolutional layers extract features on the image by sliding convolutional kernels, recognizing basic visual elements such as edges, textures, and shapes. In cardiac ultrasound images, shallow convolutions can identify low-level features such as the boundaries of the heart chambers and the texture patterns of the myocardium. As the network depth increases, subsequent convolutional layers can extract higher-level semantic features. For example, mid-level convolutional layers can identify the overall contour of the heart chambers and the morphology of the ventricles and atria; deep convolutional layers can identify high-level semantic information such as the size of the left ventricle, the thickness of the ventricular wall, and the heart's contractile function. These high-level features directly correspond to key diagnostic indicators of the disease. During the feature extraction process of the convolutional layers, pooling layers play a role in dimensionality reduction and abstraction, gradually compressing the spatial size of the image while retaining the most important feature information. Through alternating layers of convolution and pooling, the model progressively transforms the original two-dimensional image into a highly abstract feature representation. Finally, a global average pooling layer compresses multiple feature maps into a fixed-dimensional image feature vector, such as a 512-dimensional vector. This image feature vector comprehensively encodes various disease-related visual information in medical images, with different dimensions reflecting different visual features of the image.

[0063] Next, the text feature vectors and image feature vectors are input into a multimodal fusion network and fused through a gated fusion mechanism to obtain the patient's disease features. The role of the multimodal fusion network is to effectively integrate features from two different information sources, text and image, to generate a unified and more comprehensive representation of the patient's disease features. The text modality excels at expressing explicit knowledge such as disease diagnosis and symptom description, while the image modality excels at displaying the visual features and spatial information of the disease. Simply concatenating or averaging the two feature vectors often fails to fully utilize their complementary information, thus requiring a specialized fusion mechanism. Specifically, a gated fusion mechanism is used for adaptive feature fusion. Its core idea is that the importance of features from different modalities in characterizing the patient's disease state is dynamically changing. For example, for some cases, textual diagnostic information may be more critical; while for others, image features may contain more valuable information. The gated fusion mechanism can automatically adjust the weight ratio of different modal features in the fusion process according to the characteristics of the specific case. The specific fusion process is as follows: First, the correlation between the text feature vector and the image feature vector is calculated. The correlation reflects the degree of semantic consistency between the two modal information. The calculation method involves multiplying the corresponding elements of two vectors and summing the results to obtain a relevance score. If the disease information described in the text is highly consistent with the lesion features shown in the image, the values ​​of the two feature vectors in the relevance dimension will be similar, resulting in a high relevance score. Conversely, if the two modalities are not very consistent, the relevance score will be low. For example, for a hypertensive patient, if the text diagnosis clearly describes "left ventricular hypertrophy," and the echocardiogram clearly shows left ventricular wall thickening, the dimensions representing "left ventricular hypertrophy" in the text feature vector and the dimensions representing "ventricular wall thickening" in the image feature vector will be relatively consistent, resulting in a high relevance score. Next, the text feature vector and the image feature vector are concatenated to form a joint feature vector. The concatenation operation links the two vectors end-to-end. If the text feature vector has a dimension of 512 and the image feature vector also has a dimension of 512, the concatenated joint feature vector will have a dimension of 1024. This joint feature vector simultaneously contains complete information from both the text and image modalities. Then, the joint feature vector and relevance score are fed into a gating network. The gating network is a small network consisting of fully connected neural network layers, whose function is to learn and generate appropriate fusion weights based on the current feature state and modal relevance. The input to the gating network is the combination of the joint feature vector and the relevance score, and the output is a fusion weight value between 0 and 1, whose numerical range is limited by an activation function. The fusion weight represents how much weight should be assigned to the text features. Correspondingly, the weight of the image features is 1 minus this fusion weight.When the fusion weight is close to 1, it indicates that more reliance should be placed on text features; when the fusion weight is close to 0, it indicates that more reliance should be placed on image features; and when the fusion weight is close to 0.5, it indicates that both modalities are equally important. The parameters of the gating network are learned through training data. During training, the model learns from a large number of samples when to trust text information more and when to trust image information more. For example, if training data shows that image features are more reliable than text descriptions for a certain type of disease, the gating network will learn to reduce the weight of text features for such cases. Then, based on the calculated fusion weights, the text feature vector and the image feature vector are weighted and fused. Specifically, the text feature vector is multiplied by the fusion weight, the image feature vector is multiplied by 1 and the value of the fusion weight is subtracted, and then the corresponding elements of the two weighted vectors are added together to obtain the final patient disease features.

[0064] Through a gated fusion mechanism, the fusion ratio of different modalities can be adaptively adjusted according to the characteristics of specific cases. When textual information is more critical or reliable, the weight of textual features is automatically increased; when image information is more important or clearer, the weight of image features is automatically increased; and when two modalities are complementary and equally important, they are given balanced weights. Compared to simple splicing or fixed-weighting, this dynamic fusion method can better integrate multimodal information and generate a more accurate and comprehensive representation of patient disease characteristics. The resulting patient disease characteristics integrate information from both disease diagnosis text and medical images, providing a more comprehensive characterization of the patient's disease state and offering a high-quality feature foundation for subsequent historical case retrieval and correlation assessment.

[0065] Furthermore, based on the patient's disease characteristics and the drug information, a historical patient case set is obtained, including:

[0066] S21. Perform local sensitive hash mapping on the patient disease characteristics of the target patient to determine the target hash bucket to which the target patient belongs and its neighboring hash buckets;

[0067] S22. Retrieve historical patient cases located in the target hash bucket and neighboring hash buckets from the historical patient case database to form a candidate case set;

[0068] S23. Based on the patient's disease characteristics and the drug information, the candidate case set is screened, and historical patient cases that simultaneously meet the preset screening conditions are retained to form a historical patient case set.

[0069] The preset screening conditions include: the similarity between the disease characteristics of historical patients and the disease characteristics of the patient exceeds a preset similarity threshold; the medication regimen of historical patients includes the multiple identified drugs; and the medication regimen of historical patients includes a candidate drug option of the drug to be identified.

[0070] In a preferred embodiment, firstly, Local Sensitive Hash Mapping (LSH) is performed on the disease features of the target patient to determine the target hash bucket and its neighboring hash buckets. LSH is an efficient similarity retrieval technique. Its basic principle is to map high-dimensional feature vectors to several hash buckets using a hash function, so that similar feature vectors are likely mapped to the same or adjacent hash buckets, while dissimilar feature vectors are likely mapped to different hash buckets. This method quickly narrows the search scope, avoiding one-by-one comparisons in massive amounts of data. In this embodiment, the historical patient case database stores a large number of past patient case records, each containing the patient's disease feature vector. To improve retrieval efficiency, an index is pre-built on the historical case database: LSH is performed on the disease feature vector of each historical patient in the database, assigning it to the corresponding hash bucket, and a hash bucket index structure is established. When it is necessary to retrieve similar cases for the target patient, the same hash mapping operation is first performed on the target patient's disease feature vector to determine the hash bucket to which the feature vector should belong, called the target hash bucket. Due to the characteristics of locality-sensitive hashing, historical patients with similar disease characteristics to the target patient are highly likely to have their disease feature vectors mapped to the target hash bucket or its neighboring hash buckets. Therefore, it is necessary to determine not only the target hash bucket but also its neighboring hash buckets. A neighboring hash bucket refers to a hash bucket in the hash space that is adjacent to or close to the target hash bucket. By simultaneously searching the target hash bucket and its neighboring hash buckets, similar cases in boundary cases can be avoided while ensuring search efficiency.

[0071] Then, historical patient cases located in the target hash bucket and neighboring hash buckets are retrieved from the historical patient case database to form a candidate case set. Based on the target hash bucket and neighboring hash buckets determined in step S21, all historical patient cases contained in these hash buckets are quickly located by querying the pre-established hash bucket index. Since the hash mapping has already completed the initial similarity screening, the disease characteristics of these historical patients located in the target hash bucket and neighboring hash buckets are somewhat similar to the disease characteristics of the target patient. All retrieved historical patient cases are summarized to form a candidate case set. The size of this candidate case set is much smaller than the entire historical case database, thus greatly improving the efficiency of subsequent fine-grained screening. For example, the historical case database may contain hundreds of thousands of case records, while the candidate case set obtained through hash retrieval may only contain thousands of cases, narrowing the search scope by hundreds of times.

[0072] Next, the candidate case set was screened based on patient disease characteristics and drug information, retaining historical patient cases that simultaneously met the preset screening criteria to form a historical patient case set. Although the historical patients in the candidate case set and the target patients have certain similarities in disease characteristics, not all candidate cases are suitable for correlation analysis. To ensure the quality and relevance of the historical patient case set, preset screening criteria were set to further refine the screening of candidate cases.

[0073] The preset filtering criteria include three aspects:

[0074] First, the similarity between the disease characteristics of historical patients and those of the target patient must exceed a preset similarity threshold. The similarity between the disease characteristics of each historical patient in the candidate case set and those of the target patient is calculated. Similarity calculation typically uses metrics such as cosine similarity or Euclidean distance. Only historical patient cases with a similarity exceeding the preset similarity threshold are retained. This condition ensures the comparability of historical patients and target patients in terms of disease status, avoiding the inclusion of cases with excessively large disease differences in the analysis. This preset similarity threshold is determined by an expert panel.

[0075] Second, the medication regimens of historical patients must include the multiple identified drugs from step S1. Since the remaining drugs are selected based on the already identified medications, the medication baseline of historical patients must be consistent with the current scenario. The medication regimen records of each historical patient in the candidate cases are examined to determine if they contain all identified drugs. Only historical patient cases with medication regimens containing all identified drugs are retained. This condition ensures the consistency between historical cases and the current medication scenario, making historical experience valuable as a reference.

[0076] Third, the patient history medication regimens must include at least one candidate drug option for the drug to be identified. Since the goal is to assess the correlation between candidate drugs, historical data showing cases of use of these candidate drugs is necessary. The patient history medication regimens are examined to determine if any candidate drug corresponding to the drug to be identified was used. This condition ensures that historical cases provide data support for evaluating the efficacy of candidate drugs.

[0077] Each historical patient case in the candidate case set was individually evaluated against the three conditions mentioned above, and only cases that met all three conditions were retained. The resulting historical patient case set, after screening, is highly similar to the target patient's disease status and highly relevant to the current medication scenario, providing a high-quality data foundation for subsequent correlation assessment.

[0078] Furthermore, the first association degree of each of the candidate drug options is obtained based on the historical patient case set assessment, including:

[0079] S311. Determine a first candidate drug option from the plurality of candidate drug options, and screen historical patient cases that used the corresponding first candidate drug option from the historical patient case set according to the first candidate drug option to form a first sub-case set;

[0080] S312. Calculate the number of cases that are effective in the first sub-case group and the total number of cases in the first sub-case group to obtain the first effective case number and the first total case number.

[0081] S313. Calculate the ratio of the first effective case count to the first total case count to obtain the first original effective rate;

[0082] S314. Calculate a first confidence coefficient based on the first total number of cases and the preset benchmark sample size. When the first total number of cases is less than the preset benchmark sample size, the first confidence coefficient is equal to the ratio of the first total number of cases to the preset benchmark sample size. When the first total number of cases is greater than or equal to the preset benchmark sample size, the first confidence coefficient is 1.

[0083] S315. Multiply the first original effectiveness rate by the first confidence coefficient to obtain the first correlation degree of the first candidate drug option;

[0084] S316. Obtain the first correlation of the remaining candidate drug options in the same way as obtaining the first correlation of the first candidate drug option, and obtain the first correlation of each candidate drug option.

[0085] In a preferred embodiment, firstly, a first candidate drug option is determined from multiple candidate drug options. Based on this first candidate drug option, historical patient cases that used the corresponding first candidate drug option are selected from the historical patient case set, forming a first sub-case set. The first candidate drug option is an arbitrary candidate drug selected from the multiple candidate drug options for the drug to be determined, used to illustrate the process of calculating the correlation. Each case record in the historical patient case set is traversed, checking whether the first candidate drug option was used in the patient's medication regimen. If the candidate drug was used, the case is included in the first sub-case set; otherwise, it is not included. After screening, the first sub-case set contains historical patient cases that are similar to the target patient's disease, have the same medication basis, and used the first candidate drug option. These cases record the use and treatment results of the candidate drug in a real clinical setting, providing a data foundation for evaluating the efficacy of the candidate drug.

[0086] Then, the number of cases with effective treatment in the first sub-case set and the total number of cases in the first sub-case set are counted to obtain the first number of effective cases and the first total number of cases. For each case in the first sub-case set, the treatment outcome information of that case is read to determine whether the treatment was effective. The criteria for determining treatment effectiveness are determined based on the disease type and clinical practice. If there is clear treatment outcome information in the case, such as "treatment effective," "symptom improvement," or "treatment goal achieved," the determination is made directly based on this information. If there is no clear labeling in the case, efficacy evaluation indicators and evaluation time windows are determined based on the disease type to determine whether the patient's relevant indicators after treatment have reached the preset standards. Specifically, efficacy evaluation rules corresponding to different disease types are pre-configured. These rules include two elements: efficacy evaluation indicators and evaluation time windows. Efficacy evaluation indicators are specific clinical indicators used to determine whether treatment is effective. The efficacy evaluation indicators differ for different diseases. For example, for patients with hypertension, efficacy assessment indicators could be systolic and diastolic blood pressure values; for patients with diabetes, efficacy assessment indicators could be fasting blood glucose and glycated hemoglobin levels; for patients with hyperlipidemia, efficacy assessment indicators could be total cholesterol and low-density lipoprotein levels; and for patients with heart failure, efficacy assessment indicators could be left ventricular ejection fraction and cardiac function classification. The assessment time window refers to the time interval from the start of medication to the assessment of efficacy. The onset time of treatment and the timing of assessment vary for different diseases. For example, for acute infectious diseases, the assessment time window could be three to seven days after medication; for chronic diseases such as hypertension, the assessment time window could be four to eight weeks after medication; and for cancer treatment, the assessment time window could be three months after treatment. Based on the patient's historical disease type, the corresponding efficacy assessment rules are queried to obtain the assessment indicators and assessment time window for that disease. Then, the relevant indicator values ​​for the patient within the assessment time window are extracted from the case records. For example, for patients with hypertension, the system extracts the blood pressure values ​​measured between four and eight weeks after medication. Next, the extracted indicator values ​​are compared with preset standards. Preset standards are treatment achievement standards determined according to clinical guidelines or medical consensus. For example, for patients with hypertension, the preset criteria could be a systolic blood pressure below 140 mmHg and a diastolic blood pressure below 90 mmHg; for patients with diabetes, the preset criteria could be a glycated hemoglobin level below 7%. If the patient's indicators meet the preset criteria within the assessment time window, the treatment is deemed effective; otherwise, it is deemed ineffective. This method allows for automated efficacy assessment of historical cases without clear efficacy markers, thus fully utilizing historical data for correlation evaluation. Subsequently, the number of cases deemed effective in the first sub-case set is counted, denoted as the first effective case count. Simultaneously, the total number of cases in the first sub-case set is counted, denoted as the first total case count.

[0087] Then, the ratio of the first effective cases to the first total cases is calculated to obtain the first original response rate. The first original response rate reflects the proportion of patients in the history who responded to treatment using the first candidate drug option. It is calculated by dividing the first effective cases by the first total cases. For example, if the first sub-case set has 80 cases, and 60 of them responded, the first original response rate is 75%. This response rate initially reflects the therapeutic effect of the candidate drug on this type of disease. Simultaneously, a first confidence coefficient is calculated based on the first total cases and the pre-set baseline sample size. The sample size in historical data may vary significantly for different candidate drugs. Some candidate drugs may be more widely used in clinical practice with a sufficient number of historical cases; while others may be less used with a limited number of historical cases. When the sample size is small, the statistically obtained response rate has lower reliability and may have greater randomness and chance. To reflect the impact of sample size on the reliability of statistical results, a confidence coefficient is introduced to adjust the original response rate. The pre-set baseline sample size is a pre-defined threshold, representing the minimum standard for considering the sample size sufficient. The pre-set baseline sample size can be set according to the disease type and clinical experience, for example, 100 or 200 cases. The initial total number of cases is compared with the pre-set baseline sample size. When the initial total number of cases is less than the pre-set baseline sample size, it indicates that the sample size for the candidate drug is insufficient, and the reliability of the statistical results needs to be discounted. In this case, the first confidence coefficient is equal to the ratio of the initial total number of cases to the pre-set baseline sample size. For example, if the pre-set baseline sample size is 100 cases and the initial total number of cases is 60 cases, then the first confidence coefficient is 0.6. This means that due to the insufficient sample size, the reliability of the efficacy statistical results for this candidate drug is only 60% of the baseline reliability. When the initial total number of cases is greater than or equal to the pre-set baseline sample size, it indicates that the sample size for the candidate drug is sufficient, and the statistical results have high reliability. In this case, the first confidence coefficient is 1, indicating that no discounting of the efficacy rate is required.

[0088] Next, the first original efficacy rate is multiplied by the first confidence coefficient to obtain the first correlation score of the first candidate drug option. The first correlation score is the efficacy rate after confidence adjustment, comprehensively reflecting the efficacy level of the candidate drug and the reliability of the statistical results. When the sample size is sufficient, the first correlation score equals the original efficacy rate; when the sample size is insufficient, the first correlation score is lower than the original efficacy rate, reflecting a cautious attitude towards statistical results from small samples. For example, if the first original efficacy rate is 75% and the first confidence coefficient is 0.6, then the first correlation score is 0.45. This means that although the historical efficacy rate of the candidate drug is 75%, due to insufficient sample size, the correlation score after confidence adjustment is 0.45.

[0089] For each candidate drug option in the drug category to be determined, repeat the calculation process from steps S311 to S315. That is, for each candidate drug, select cases that have used the candidate drug from the historical patient case set to form a subcase set, count the number of valid cases and the total number of cases, calculate the original effectiveness rate, calculate the confidence coefficient based on the sample size, and finally obtain the first association degree of the candidate drug.

[0090] Through the above process, the first correlation scores of all candidate drug options were obtained. These first correlation scores reflect the overall therapeutic efficacy of each candidate drug for this type of disease, providing a quantitative basis for subsequent comprehensive recommendation and evaluation.

[0091] Furthermore, a second correlation degree is obtained for each of the candidate drug options based on the historical patient case set and the patient's basic characteristics, including:

[0092] S321. Based on the first sub-case set, calculate the similarity between the patient baseline characteristics of each historical patient in the first sub-case set and the patient baseline characteristics of the target patient;

[0093] S322. Select historical patient cases with similarity exceeding a preset similarity threshold to form the first similar case set;

[0094] S323. Calculate the number of cases that are effective in the first set of similar cases and the total number of cases in the first set of similar cases to obtain the second number of effective cases and the second total number of cases.

[0095] S324. Calculate the ratio of the second effective case count to the second total case count to obtain the second correlation of the first candidate drug option;

[0096] S325. Obtain the second correlation of the remaining candidate drug options in the same way as obtaining the second correlation of the first candidate drug option, and obtain the second correlation of each candidate drug option.

[0097] In a preferred embodiment, for the first candidate drug option, firstly, based on a first sub-case set, the similarity between the patient baseline characteristics of each historical patient in the first sub-case set and the patient baseline characteristics of the target patient is calculated. The first sub-case set is constructed when calculating the first association degree and includes all historical patient cases using the first candidate drug option. Although these historical patients and the target patient are similar in disease characteristics, they may differ in patient baseline characteristics. Patient baseline characteristics include individualized information such as age, gender, weight, liver and kidney function, blood lipid and blood glucose levels, and genotype, which have a significant impact on drug efficacy. For each historical patient in the first sub-case set, the similarity between their patient baseline characteristics and the patient baseline characteristics of the target patient is calculated. The similarity calculation comprehensively considers various dimensions of the patient baseline characteristics. The specific calculation method is as follows: First, the similarity is calculated for each dimension of the patient baseline characteristics separately. For numerical characteristics, such as age, weight, and liver and kidney function indicators, the calculation method is based on normalization of numerical differences; the smaller the difference, the higher the similarity. For example, regarding age characteristics, if the target patient is 65 years old and a previous patient is 63 years old, the difference is 2 years. A similarity score can be calculated based on a pre-defined tolerance range for age differences. For categorical characteristics, such as gender and genotype, the calculation method is to determine if they are identical; if identical, the similarity score is 1; otherwise, it is 0. Then, the similarities of each dimension are weighted and aggregated to obtain a comprehensive similarity score for the patient's basic characteristics. The importance of different dimensions may vary, and the weights of each dimension are determined based on clinical experience or statistical analysis. For example, for some drugs, age and renal function may be key factors affecting efficacy, so these two dimensions will have relatively high weights; while other dimensions will have relatively lower weights. Through weighted aggregation, a comprehensive similarity score between 0 and 1 is obtained, with the score closer to 1 indicating greater similarity in basic characteristics between the previous and target patients.

[0098] Then, historical patient cases with similarity exceeding a preset similarity threshold are selected to form the first similar case set. The preset similarity threshold is a pre-defined value used to determine whether a historical patient is sufficiently similar to the target patient. This threshold is set according to clinical needs and data distribution characteristics; for example, it can be set to 0.7 or 0.8. Each historical patient case in the first sub-case set is iterated over, and its similarity value to the target patient is checked. If the similarity exceeds the preset similarity threshold, it indicates that the historical patient and the target patient are highly similar in basic characteristics, and the case is included in the first similar case set; if the similarity does not exceed the threshold, it indicates that the historical patient and the target patient differ significantly in basic characteristics, and the case is not included. After screening, the first similar case set contains historical patient cases that are highly similar to the target patient in both disease characteristics and basic characteristics. These cases represent historical medication experience most closely related to the current patient's condition, and have higher reference value for evaluating the personalized efficacy of candidate drugs.

[0099] Next, the number of cases in the first similar case set that responded to treatment and the total number of cases in the first similar case set are counted to obtain the second number of effective cases and the second total number of cases. For each case in the first similar case set, the system determines whether the case is effective in the same way as in step S312. That is, if there is clear treatment outcome information in the case, the determination is based on the label; if there is no clear label, the determination is based on the efficacy evaluation rules determined by the disease type. The number of all cases in the first similar case set that were determined to be effective is counted and recorded as the second number of effective cases. At the same time, the total number of cases in the first similar case set is counted and recorded as the second total number of cases. These two values ​​reflect the efficacy performance of the candidate drug in a patient population similar to the target patients.

[0100] Next, the ratio of the second effective cases to the second total cases is calculated to obtain the second correlation score of the first candidate drug option. The second correlation score reflects the proportion of patients with a history highly similar to the target patient who used the first candidate drug option and whose treatment was effective. It is calculated by dividing the second effective cases by the second total cases. For example, if the first similar case set has 30 cases, and 24 of them are effective, the second correlation score is 0.8. Compared to the first correlation score, the second correlation score is based on a patient group more similar to the target patient, and therefore better reflects the personalized suitability of the candidate drug for the current patient. If the second correlation score of a candidate drug is significantly higher than the first correlation score, it indicates that the drug is more effective for a patient group similar to the current patient and has stronger personalized recommendation value.

[0101] Subsequently, for each candidate drug option in the drug to be determined, the system repeats the calculation process of steps S321 to S324. That is, for each candidate drug's corresponding sub-case set, the basic feature similarity between each historical patient and the target patient is calculated, highly similar patients are selected to form a similar case set, the number of valid cases in this similar case set and the total number of cases are counted, and finally the second correlation of the candidate drug is obtained.

[0102] Through the above process, a second correlation coefficient was obtained for all candidate drug options. These second correlation coefficients reflect the therapeutic effects of each candidate drug on patient groups similar to the current patient at a personalized level, providing a second dimension of quantitative basis for subsequent comprehensive recommendation evaluation.

[0103] Furthermore, each candidate drug option is evaluated and recommended using a first correlation and a second correlation to obtain a comprehensive recommendation for each candidate drug option, including:

[0104] S41. For the first candidate drug option, calculate an initial second weighting coefficient based on the second total number of cases and the preset basic total number of cases. The initial second weighting coefficient is equal to the second total number of cases divided by the sum of the second total number of cases and the basic total number of cases.

[0105] S42. Compare the initial second weight coefficient with the preset upper limit weight value, and take the smaller value between the initial second weight coefficient and the preset upper limit weight value as the second weight coefficient;

[0106] S43. Calculate the first weight coefficient, which is equal to 1 minus the second weight coefficient;

[0107] S44. The first correlation degree and the second correlation degree are weighted and fused according to the first weight coefficient and the second weight coefficient to obtain the comprehensive recommendation degree of the first candidate drug option;

[0108] S45. Obtain the comprehensive recommendation scores of the remaining candidate drug options in the same way as obtaining the comprehensive recommendation score of the first candidate drug option, and thus obtain the comprehensive recommendation scores of each candidate drug option.

[0109] In a preferred embodiment, firstly, for the first candidate drug option, an initial second weighting coefficient is calculated based on the second total number of cases and a preset baseline total number of cases. The initial second weighting coefficient is equal to the second total number of cases divided by the sum of the second total number of cases and the baseline total number of cases. The first correlation and the second correlation reflect the applicability of the candidate drug from different perspectives. The first correlation represents the overall therapeutic efficacy of the candidate drug for this type of disease, while the second correlation represents the personalized efficacy of the candidate drug for a patient population similar to the current patient. To generate comprehensive recommendation results, the two correlations need to be fused. However, the importance of the two correlations is not fixed but depends on the sufficiency of personalized data. When there are many historical cases similar to the target patient, the second correlation, based on sufficient personalized samples, has high statistical reliability, and should be given a higher weight. Conversely, when there are few historical cases of similar patients, the statistical basis of the second correlation is weak, and its reliability is reduced. In this case, the weight of the second correlation should be reduced, and more reliance should be placed on the overall effectiveness reflected by the first correlation.

[0110] To achieve the aforementioned adaptive weight adjustment, a dynamic weight calculation mechanism was designed. First, a preset baseline total number of cases is obtained. This baseline total number of cases serves as a reference value, representing the minimum standard for considering a sufficient personalized sample. This baseline value is set based on clinical experience and data distribution characteristics; for example, it can be set to 50 or 100 cases. The calculation method of the initial second weight coefficient reflects the impact of the personalized sample size on the weights. When the second total number of cases equals the baseline total number of cases, the initial second weight coefficient is 0.5, indicating that each of the two associations accounts for 50% of the weight. When the second total number of cases is less than the baseline total number of cases, the initial second weight coefficient is less than 0.5, indicating insufficient personalized samples and a greater reliance on overall efficacy. When the second total number of cases is greater than the baseline total number of cases, the initial second weight coefficient is greater than 0.5, indicating sufficient personalized samples and a greater reliance on personalized efficacy. For example, if the baseline total number of cases is set to 100 and the second total number of cases is 150, then the initial second weight coefficient is 150 divided by (100 + 150), which equals 0.6. This indicates that there are sufficient personalized samples, and the second correlation should be given a weight of 60%.

[0111] Then, the initial second weight coefficient is compared with the preset upper limit weight value, and the smaller value between the initial second weight coefficient and the preset upper limit weight value is taken as the second weight coefficient. To avoid the second weight coefficient being too high and leading to over-reliance on personalized data while ignoring the overall pattern, a preset upper limit weight value is set. This preset upper limit weight value limits the maximum value of the second weight coefficient, ensuring that even if the sample size of similar patients is large, the weight of the second association degree will not increase indefinitely, and a certain proportion of overall validity reference is still retained. The preset upper limit weight value is set according to clinical needs; for example, it can be set to 0.8, indicating that the weight of the second association degree will not exceed 80%, and the weight of the first association degree will be retained at least 20%. Afterward, the initial second weight coefficient calculated in step S41 is compared with the preset upper limit weight value. If the initial second weight coefficient is less than or equal to the preset upper limit weight value, it means that the weight is within a reasonable range, and the initial second weight coefficient is directly used as the final second weight coefficient; if the initial second weight coefficient is greater than the preset upper limit weight value, it means that the weight is too high and needs to be limited, and the preset upper limit weight value is used as the final second weight coefficient. For example, if the initial second weight coefficient is 0.6 and the preset upper limit weight value is 0.8, then 0.6 is less than 0.8, and the final second weight coefficient is 0.6. If the initial second weight coefficient is 0.9 and the preset upper limit weight value is 0.8, then 0.9 is greater than 0.8, and the final second weight coefficient is limited to 0.8.

[0112] Next, the first weighting coefficient is calculated, which equals 1 minus the second weighting coefficient. Since the first and second correlations need to be merged according to their weight ratios, and the sum of the two weighting coefficients should be 1 to ensure the reasonableness of the fusion result, the first and second weighting coefficients are complementary. Once the second weighting coefficient is determined, the first weighting coefficient is automatically set to 1 minus the second weighting coefficient. For example, if the second weighting coefficient is 0.6, then the first weighting coefficient is 0.4. This means that in the fusion process, the first correlation accounts for 40% of the weight, and the second correlation accounts for 60% of the weight.

[0113] Next, the first correlation and the second correlation are weighted and fused according to the first and second weight coefficients to obtain the comprehensive recommendation of the first candidate drug option. The weighted fusion is calculated by multiplying the first correlation by the first weight coefficient, multiplying the second correlation by the second weight coefficient, and then adding the two results to obtain the comprehensive recommendation. The comprehensive recommendation integrates the evaluation results of the overall effectiveness and personalized suitability of the candidate drug, and achieves adaptive adjustment under different data conditions through a dynamic weight mechanism. At the same time, the comprehensive recommendation of the remaining candidate drug options is obtained in the same way as the comprehensive recommendation of the first candidate drug option, to obtain the comprehensive recommendation of each candidate drug option. For each candidate drug option to be determined, the calculation process of steps S41 to S44 is repeated. That is, for each candidate drug, a dynamic weight coefficient is calculated based on its second total number of cases, and its first correlation and the second correlation are weighted and fused to obtain the comprehensive recommendation of the candidate drug.

[0114] Through the above process, a comprehensive recommendation score for all candidate drug options was obtained. These comprehensive recommendation scores reflect the overall effectiveness and individual suitability of each candidate drug, providing a quantitative reference for physicians' medication decisions. Candidate drugs can be ranked according to their comprehensive recommendation scores, and the results can be fed back to the physician's terminal to assist in making more scientific and precise medication choices.

[0115] Example 2, as Figure 2 As shown, based on the same inventive concept as the drug-disease association analysis method using multimodal feature fusion provided in Embodiment 1, this embodiment of the invention also provides a drug-disease association analysis system using multimodal feature fusion, comprising:

[0116] The data retrieval module 11 is used to retrieve drug information and patient characteristics of the target patient associated with the drug-disease correlation analysis command issued by the doctor's terminal in response to the command.

[0117] The patient characteristics include basic patient characteristics and patient disease characteristics, and the drug information includes multiple identified drugs and one unidentified drug item, wherein the unidentified drug item has multiple candidate drug options.

[0118] The case acquisition module 12 is used to acquire a historical patient case set based on the patient's disease characteristics and the drug information;

[0119] The correlation assessment module 13 is used to assess and obtain a first correlation of each of the candidate drug options based on the historical patient case set, and to assess and obtain a second correlation of each of the candidate drug options based on the historical patient case set and the patient's basic characteristics.

[0120] The comprehensive recommendation module 14 is used to evaluate each candidate drug option based on the first and second correlation of each candidate drug option, obtain the comprehensive recommendation degree of each candidate drug option, and feed it back to the doctor terminal.

[0121] Furthermore, the data retrieval module 11 is also used for:

[0122] Based on the preset basic information collection list, obtain the basic patient characteristics corresponding to the target patient from the electronic medical record system;

[0123] Acquire disease-related multimodal data of the target patient, including disease diagnosis text data and medical image data;

[0124] The disease diagnosis text data and medical image data are encoded separately and then fused through a multimodal fusion network to obtain the patient's disease characteristics;

[0125] The patient's basic characteristics and the patient's disease characteristics are summarized as the patient's characteristics;

[0126] The prescription system retrieves the multiple confirmed drugs already prescribed by the doctor and the multiple candidate drug options corresponding to the drugs to be determined, thus obtaining drug information.

[0127] Furthermore, the data retrieval module 11 is also used for:

[0128] The disease diagnosis text data is processed by a text data encoding analyzer to obtain text feature vectors, wherein the text data encoding analyzer is trained and generated based on the sample disease diagnosis text dataset and the sample text feature vectors;

[0129] The medical image data is processed by an image data encoding and analysis device to obtain image feature vectors, wherein the image data encoding and analysis device is trained and generated based on a sample medical image dataset and sample image feature vectors.

[0130] The text feature vector and image feature vector are input into a multimodal fusion network. The multimodal fusion network adopts a gated fusion mechanism, dynamically generates fusion weights based on the correlation between the text feature vector and the image feature vector, and performs weighted fusion of the text feature vector and the image feature vector to obtain the patient's disease features.

[0131] Furthermore, the case acquisition module 12 is also used for:

[0132] Perform locality-sensitive hash mapping on the patient disease characteristics of the target patient to determine the target hash bucket to which the target patient belongs and its neighboring hash buckets;

[0133] Historical patient cases located in the target hash bucket and neighboring hash buckets are retrieved from the historical patient case database to form a candidate case set;

[0134] The candidate case set is screened based on the patient's disease characteristics and the drug information, and historical patient cases that simultaneously meet the preset screening conditions are retained to form a historical patient case set.

[0135] The preset screening conditions include: the similarity between the disease characteristics of historical patients and the disease characteristics of the patient exceeds a preset similarity threshold; the medication regimen of historical patients includes the multiple identified drugs; and the medication regimen of historical patients includes a candidate drug option of the drug to be identified.

[0136] Furthermore, the correlation assessment module 13 is also used for:

[0137] A first candidate drug option is determined from the plurality of candidate drug options, and historical patient cases that used the corresponding first candidate drug option are screened from the historical patient case set according to the first candidate drug option to form a first sub-case set;

[0138] The number of cases that responded to treatment in the first sub-case group and the total number of cases in the first sub-case group were counted to obtain the first effective case number and the first total case number.

[0139] Calculate the ratio of the first effective case count to the first total case count to obtain the first original effectiveness rate;

[0140] A first confidence coefficient is calculated based on the first total number of cases and the preset baseline sample size. When the first total number of cases is less than the preset baseline sample size, the first confidence coefficient is equal to the ratio of the first total number of cases to the preset baseline sample size. When the first total number of cases is greater than or equal to the preset baseline sample size, the first confidence coefficient is 1.

[0141] Multiply the first original effectiveness rate by the first confidence coefficient to obtain the first correlation degree of the first candidate drug option;

[0142] The first correlation of the remaining candidate drug options is obtained in the same way as the first correlation of the first candidate drug option, thus obtaining the first correlation of each candidate drug option.

[0143] Furthermore, the correlation assessment module 13 is also used for:

[0144] Based on the first sub-case set, calculate the similarity between the basic patient characteristics of each historical patient in the first sub-case set and the basic patient characteristics of the target patient;

[0145] Historical patient cases with similarity exceeding a preset similarity threshold are selected to form the first set of similar cases;

[0146] The number of effective cases in the first set of similar cases and the total number of cases in the first set of similar cases are counted to obtain the second number of effective cases and the second total number of cases.

[0147] The ratio of the second effective case count to the second total case count is used to obtain the second correlation of the first candidate drug option;

[0148] The second correlation of the remaining candidate drug options is obtained in the same way as the second correlation of the first candidate drug option, thus obtaining the second correlation of each candidate drug option.

[0149] Furthermore, the correlation assessment module 13 is also used for:

[0150] For the first candidate drug option, an initial second weighting coefficient is calculated based on the second total number of cases and the preset base total number of cases. The initial second weighting coefficient is equal to the second total number of cases divided by the sum of the second total number of cases and the base total number of cases.

[0151] The initial second weight coefficient is compared with the preset upper limit weight value, and the smaller value between the initial second weight coefficient and the preset upper limit weight value is taken as the second weight coefficient.

[0152] Calculate the first weighting coefficient, which is equal to 1 minus the second weighting coefficient;

[0153] The first correlation and the second correlation are weighted and fused according to the first weight coefficient and the second weight coefficient to obtain the comprehensive recommendation of the first candidate drug option;

[0154] The overall recommendation scores for the remaining candidate drug options are obtained by using the same method as for obtaining the overall recommendation score for the first candidate drug option, thus obtaining the overall recommendation score for each candidate drug option.

[0155] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for analyzing the association between drugs and diseases through multimodal feature fusion, characterized in that, The method includes: In response to a drug-disease correlation analysis command issued by the doctor's terminal, the drug information associated with the drug-disease correlation analysis command and the patient characteristics of the target patient are retrieved. The patient characteristics include basic patient characteristics and patient disease characteristics, and the drug information includes multiple identified drugs and one unidentified drug item, wherein the unidentified drug item has multiple candidate drug options. Based on the patient's disease characteristics and the drug information, obtain a historical patient case set; The first correlation degree of each candidate drug option is obtained by evaluating the historical patient case set, and the second correlation degree of each candidate drug option is obtained by evaluating the historical patient case set and the patient's basic characteristics. The candidate drug options are evaluated and recommended based on the first and second correlation scores to obtain a comprehensive recommendation score, which is then fed back to the doctor's terminal.

2. The method according to claim 1, characterized in that, Retrieve drug information and patient characteristics associated with the drug-disease correlation analysis command, including: Based on the preset basic information collection list, obtain the basic patient characteristics corresponding to the target patient from the electronic medical record system; Acquire disease-related multimodal data of the target patient, including disease diagnosis text data and medical image data; The disease diagnosis text data and medical image data are encoded separately and then fused through a multimodal fusion network to obtain the patient's disease characteristics; The patient's basic characteristics and the patient's disease characteristics are summarized as the patient's characteristics; The prescription system retrieves the multiple confirmed drugs already prescribed by the doctor and the multiple candidate drug options corresponding to the drugs to be determined, thus obtaining drug information.

3. The method according to claim 2, characterized in that, The disease diagnosis text data and medical image data are encoded separately and then fused using a multimodal fusion network to obtain patient disease characteristics, including: The disease diagnosis text data is processed by a text data encoding analyzer to obtain text feature vectors, wherein the text data encoding analyzer is trained and generated based on the sample disease diagnosis text dataset and the sample text feature vectors; The medical image data is processed by an image data encoding and analysis device to obtain image feature vectors, wherein the image data encoding and analysis device is trained and generated based on a sample medical image dataset and sample image feature vectors. The text feature vector and image feature vector are input into a multimodal fusion network. The multimodal fusion network adopts a gated fusion mechanism, dynamically generates fusion weights based on the correlation between the text feature vector and the image feature vector, and performs weighted fusion of the text feature vector and the image feature vector to obtain the patient's disease features.

4. The method according to claim 1, characterized in that, Based on the patient's disease characteristics and the drug information, a historical patient case set is obtained, including: Perform locality-sensitive hash mapping on the patient disease characteristics of the target patient to determine the target hash bucket to which the target patient belongs and its neighboring hash buckets; Historical patient cases located in the target hash bucket and neighboring hash buckets are retrieved from the historical patient case database to form a candidate case set; The candidate case set is screened based on the patient's disease characteristics and the drug information, and historical patient cases that simultaneously meet the preset screening conditions are retained to form a historical patient case set. The preset screening conditions include: the similarity between the disease characteristics of historical patients and the disease characteristics of the patient exceeds a preset similarity threshold; the medication regimen of historical patients includes the multiple identified drugs; and the medication regimen of historical patients includes a candidate drug option for the drug to be identified.

5. The method according to claim 1, characterized in that, The first association degree of each of the candidate drug options was obtained by evaluating historical patient case sets, including: A first candidate drug option is determined from the plurality of candidate drug options, and historical patient cases that used the corresponding first candidate drug option are screened from the historical patient case set according to the first candidate drug option to form a first sub-case set; The number of cases that responded to treatment in the first sub-case group and the total number of cases in the first sub-case group were counted to obtain the first effective case number and the first total case number. Calculate the ratio of the first effective case count to the first total case count to obtain the first original effectiveness rate; A first confidence coefficient is calculated based on the first total number of cases and the preset baseline sample size. When the first total number of cases is less than the preset baseline sample size, the first confidence coefficient is equal to the ratio of the first total number of cases to the preset baseline sample size. When the first total number of cases is greater than or equal to the preset baseline sample size, the first confidence coefficient is 1. Multiply the first original effectiveness rate by the first confidence coefficient to obtain the first correlation degree of the first candidate drug option; The first correlation of the remaining candidate drug options is obtained in the same way as the first correlation of the first candidate drug option, thus obtaining the first correlation of each candidate drug option.

6. The method according to claim 5, characterized in that, A second correlation degree is obtained for each of the candidate drug options based on the historical patient case set and the patient's basic characteristics, including: Based on the first sub-case set, calculate the similarity between the basic patient characteristics of each historical patient in the first sub-case set and the basic patient characteristics of the target patient; Historical patient cases with similarity exceeding a preset similarity threshold are selected to form the first set of similar cases; The number of effective cases in the first set of similar cases and the total number of cases in the first set of similar cases are counted to obtain the second number of effective cases and the second total number of cases. The ratio of the second effective case count to the second total case count is used to obtain the second correlation of the first candidate drug option; The second correlation of the remaining candidate drug options is obtained in the same way as the second correlation of the first candidate drug option, thus obtaining the second correlation of each candidate drug option.

7. The method according to claim 6, characterized in that, The candidate drug options are evaluated using a first correlation and a second correlation to obtain a comprehensive recommendation for each candidate drug option, including: For the first candidate drug option, an initial second weighting coefficient is calculated based on the second total number of cases and the preset base total number of cases. The initial second weighting coefficient is equal to the second total number of cases divided by the sum of the second total number of cases and the base total number of cases. The initial second weight coefficient is compared with the preset upper limit weight value, and the smaller value between the initial second weight coefficient and the preset upper limit weight value is taken as the second weight coefficient. Calculate the first weighting coefficient, which is equal to 1 minus the second weighting coefficient; The first correlation and the second correlation are weighted and fused according to the first weight coefficient and the second weight coefficient to obtain the comprehensive recommendation of the first candidate drug option; The overall recommendation scores for the remaining candidate drug options are obtained by using the same method as for obtaining the overall recommendation score for the first candidate drug option, thus obtaining the overall recommendation score for each candidate drug option.

8. A drug-disease association analysis system based on multimodal feature fusion, characterized in that, The system for implementing the method as described in any one of claims 1 to 7, the system comprising: The data retrieval module is used to respond to the drug-disease correlation analysis command issued by the doctor's terminal and retrieve drug information and patient characteristics of the target patient associated with the drug-disease correlation analysis command. The patient characteristics include basic patient characteristics and patient disease characteristics, and the drug information includes multiple identified drugs and one unidentified drug item, wherein the unidentified drug item has multiple candidate drug options. The case acquisition module is used to acquire a historical patient case set based on the patient's disease characteristics and the drug information; The correlation assessment module is used to assess and obtain a first correlation of each of the candidate drug options based on the historical patient case set, and to assess and obtain a second correlation of each of the candidate drug options based on the historical patient case set and the patient's basic characteristics. The comprehensive recommendation module is used to evaluate each candidate drug option based on its first and second correlation, obtain a comprehensive recommendation score for each candidate drug option, and feed it back to the doctor's terminal.

Images