An AI-based clinical medical research evaluation system and method

Abstract

The present invention relates to the field of medical research evaluation technology, and specifically to an AI-based clinical medical research evaluation system and method, comprising: a research generation module, used to obtain basic information and test items during clinical medical research, and set a research path corresponding to the test items; a research verification module, used to set multiple research nodes on the research path, verify the information of each research node on the research path, and determine the target relationship vector corresponding to the research node; an indicator setting module, used to extract indicators from the target relationship vector, set research indicators matching the research nodes, and obtain research text data corresponding to the research indicators; a process verification module, used to identify the diagnostic behavior of the research text data, and process the process corresponding to the research indicator according to the diagnostic behavior, and evaluate the process difference degree of each research indicator under the corresponding process; the efficiency and accuracy of the evaluation are improved, and the setting of the research path and nodes is optimized.

Description

Technical Field

[0001] The present invention relates to the field of medical research evaluation technology, and specifically to an AI-based clinical medical research evaluation system and method. Background Art

[0002] In modern clinical medical research, it is crucial to efficiently and accurately evaluate the effects of different diagnosis and treatment pathways. With the development of medical technology and information technology, more and more data are collected to support clinical decision-making, including but not limited to patients' personal health information, laboratory test results, imaging examination data, etc. However, traditional data analysis methods often face many challenges, such as large amounts of data, complex and diverse data types, and time-consuming analysis processes. These problems limit the efficiency and depth of clinical research.

[0003] For example, Chinese patent publication number CN116959651A discloses a clinical trial efficacy evaluation method, system, device and medium, and the technical solution is: obtaining subject information; selecting a filling and evaluation terminal associated with a project according to the subject information, and allocating the subject information to the filling and evaluation terminal; obtaining the processing status of the imaging report of the subject by the filling and evaluation terminal; when it is judged that the imaging report is being processed, sending a reminder of the efficacy filling of the subject to the filling and evaluation terminal; obtaining the current efficacy report of the subject by the filling and evaluation terminal; and comparing the current efficacy report of the subject with the historical efficacy report according to the efficacy selection information to generate efficacy evaluation information.

[0004] For example, Chinese patent publication number CN116646041A discloses a method and system for improving the matching accuracy of clinical trial subjects based on a big model, which is applied to the technical field of big models, and includes obtaining medical text information; constructing a big language model in the medical field; extracting medical text information through the big language model in the medical field and obtaining a special disease data set; matching the special disease data set with corresponding clinical trial projects through preset project inclusion and exclusion criteria, and the project inclusion and exclusion criteria include different clinical trial projects and corresponding inclusion and exclusion criteria.

[0005] In the existing technology, after forming a data set for the patient according to the different conditions of the subjects, these data are directly compared with the historical data. However, this processing method tends to be targeted at clinical efficacy research and is easily affected by the huge amount of data, complex and diverse data types, and time-consuming analysis process. In this case, the current main research nodes and corresponding indicators cannot be dynamically adjusted according to the actual research, resulting in the final analysis report being unable to consider various influencing factors and lacking personalized support for patients. Summary of the invention

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: an AI-based clinical medicine research evaluation system, including: a research generation module, used to obtain basic information and test items in clinical medicine research, and set research paths corresponding to the test items.

[0007] The research verification module is used to set multiple research nodes on the research path, verify the information of each research node on the research path, and determine the target relationship vector corresponding to the research node.

[0008] The indicator setting module is used to extract indicators from the target relationship vector, set research indicators that match the research nodes, and obtain research text data corresponding to the research indicators.

[0009] The process verification module is used to identify the diagnostic behavior of the research text data, and process the process corresponding to the research indicators according to the diagnostic behavior, and evaluate the degree of process difference of each research indicator under the corresponding process.

[0010] The detection difference module is used to calculate the sample demand under the current process difference degree, and compare the standard detection project results according to the sample demand to generate the detection project difference results.

[0011] The result analysis module is used to analyze the difference results of the test items and combine the difference results of the test items into a research evaluation report according to the performance trend of the difference results of the test items at the research nodes.

[0012] An AI-based clinical medical research evaluation method includes: S1, obtaining basic information and test items in clinical medical research, and setting a research path corresponding to the test items.

[0013] S2, set up multiple research nodes on the research path, verify the information of each node, and determine the target relationship vector.

[0014] S3, extract indicators from the target relationship vector, set research indicators that match the research nodes, and obtain research text data.

[0015] S4, identify the diagnostic behavior of the research text data, process the processes corresponding to the research indicators according to the diagnostic behavior, and evaluate the degree of process difference.

[0016] S5, calculate the sample demand under the current process difference degree, and compare the standard test item results according to the sample demand to generate the test item difference results.

[0017] S6, analyze the difference results of the test items and combine them into a research evaluation report according to the performance trends.

[0018] The beneficial effects of the present invention are: 1. The present invention can automatically process and analyze clinical medical research data, thereby greatly improving the accuracy and efficiency of the evaluation; at the same time, it can also automatically verify information and extract indicators according to research paths and nodes, further reducing the interference of human factors.

[0019] 2. By constructing a research text library and performing similarity matching, the present invention can reasonably set research paths and nodes, and determine key nodes based on the clustering center of the vocabulary vector, which not only improves the scientific nature of the research, but also helps to ensure the reliability of the research results.

[0020] 3. The present invention can reasonably set research indicators according to the target relationship vector, and evaluate the degree of difference of the indicators under the corresponding process through the process verification module, which not only helps researchers to better understand the research results, but also provides strong support for subsequent improvements and optimizations.

[0021] Fourth, the present invention can accurately calculate the required sample size based on the current degree of process differences and historical data, which not only helps to reduce research costs but also improves research efficiency and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0023] Figure 1 It is a system framework diagram of an AI-based clinical medical research evaluation system.

[0024] Figure 2 It is a flowchart of the research and inspection module of an AI-based clinical medicine research evaluation system.

[0025] Figure 3 It is a flowchart of the process inspection module of the AI-based clinical medical research evaluation system.

[0026] Figure 4 It is a flowchart of an AI-based clinical medical research evaluation method. DETAILED DESCRIPTION

[0027] The embodiments of the present invention are described in detail below. The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. If no specific techniques or conditions are specified in the embodiments, the techniques or conditions described in the literature in the art or the product specifications are used.

[0028] See also Figure 1A clinical medicine research evaluation system based on AI includes: a research generation module, a research inspection module, an indicator setting module, a process inspection module, a detection difference module and a result analysis module; the output end of the research generation module is connected to the research inspection module, the output end of the research inspection module is connected to the indicator setting module, the output end of the indicator setting module is connected to the process inspection module, the output end of the process inspection module is connected to the detection difference module, and the output end of the detection difference module is connected to the result analysis module.

[0029] The research generation module is used to obtain basic information and test items in clinical medical research and set up research paths corresponding to the test items.

[0030] The research verification module is used to set multiple research nodes on the research path, verify the information of each research node on the research path, and determine the target relationship vector corresponding to the research node.

[0031] The indicator setting module is used to extract indicators from the target relationship vector, set research indicators that match the research nodes, and obtain research text data corresponding to the research indicators.

[0032] The process verification module is used to identify the diagnostic behavior of the research text data, and process the process corresponding to the research indicators according to the diagnostic behavior, and evaluate the degree of process difference of each research indicator under the corresponding process.

[0033] The detection difference module is used to calculate the sample demand under the current process difference degree, and compare the standard detection project results according to the sample demand to generate the detection project difference results.

[0034] The result analysis module is used to analyze the difference results of the test items and combine the difference results of the test items into a research evaluation report according to the performance trend of the difference results of the test items at the research nodes.

[0035] In clinical medical research, basic information generally refers to the basic personal data of the research subjects and their health status, including but not limited to age, gender, weight, height, past medical history, family medical history, current medication status, etc.; for specific types of clinical research, more detailed genetic information may be required, such as gene sequencing results, lifestyle factors, smoking, drinking habits, and psychological status assessments. These contents will represent the specific circumstances of the research subjects, and the research subjects will be classified according to these circumstances, and multiple groups of data will be obtained.

[0036] Testing items refer to the various examinations and tests required to achieve the purpose of the study. These may include laboratory tests such as blood biochemistry analysis, tumor marker testing, imaging tests such as CT scans, MRIs, physiological function tests such as cardiopulmonary function tests, and other special tests such as tissue biopsies.

[0037] Depending on the specific needs of the study, there may be specific time points or frequencies at which these tests are required to be performed to monitor treatment efficacy or disease progression.

[0038] The research pathway defines a series of steps to be followed from patient recruitment to completion of the entire research process. It includes the task arrangement of each stage, such as initial diagnosis, treatment intervention, follow-up, the goals of each stage, the tests to be performed, the data analysis plan, and the expected results.

[0039] When the research is generated, a comprehensive evaluation will be performed based on the information described above to set the processing method of the current research and complete the overall report of the research.

[0040] In one embodiment of the present invention, the research verification module mainly selects research nodes that need to be processed from the research path, and after completing the verification of each research node, generates a target relationship vector corresponding to the research node.

[0041] When setting the research nodes, they will be set according to the key checkpoints or nodes on the research path, so that the research nodes can represent important moments such as data collection, analysis, and decision making; the research nodes will include four types of nodes: initial evaluation nodes, treatment intervention nodes, mid-term evaluation nodes, and final evaluation nodes. These nodes will be prominently displayed when targeting clinical research; by verifying these nodes, we can find the target relationship vector currently needed, as well as the logical associations between different research nodes and the degree of influence of the research nodes on the overall research process; for example, the target relationship vector will be divided into direct association vectors and indirect association vectors, so as to divide the associations between research nodes when they are set.

[0042] The initial assessment node includes the establishment of patient recruitment criteria and the collection of baseline data, such as medical history and laboratory results.

[0043] The treatment intervention node involves the implementation of specific treatment measures, such as drug administration, surgical operations, and the setting of follow-up time points.

[0044] Interim assessment nodes are key moments used to monitor treatment effectiveness or disease progression and may include changes in specific biomarker levels, imaging test results, etc.

[0045] The final evaluation node covers data aggregation, statistical analysis and conclusion drawing at the end of the entire research cycle.

[0046] The direct correlation vector indicates the strength of the relationship between two directly related nodes. For example, the status of certain gene variants in the initial evaluation node can directly affect the subsequent treatment selection.

[0047] Indirect association vectors reflect relationships that indirectly affect the final research target through a series of intermediate nodes. For example, early lifestyle adjustments may not show immediate efficacy, but may affect the rate of disease progression in the long term.

[0048] like Figure 2 As shown, the implementation method of the research verification module includes: constructing a research text library corresponding to the research path, the research text library contains multiple vocabulary vectors corresponding to the current basic information and the detection items; the vocabulary vector can be represented by any one of the vocabulary vectors in Word2Vec, GloVe, FastText, ELMo and BERT. At this time, the vocabulary vector will represent the text in the research text library about the current research topic and the research subjects in the corresponding situation, and these texts will be represented in the form of vectors to determine the position dynamics of different words and the relative relationship between texts.

[0049] Word2Vec includes two models: Continuous Bag of Words (CBOW) and Skip-gram. They both learn vector representations of words by predicting context; these vectors capture the semantic similarity between words, for example, the vectors of "king" and "queen" are close in space.

[0050] GloVe learns word vectors from word co-occurrence statistics through matrix decomposition method, emphasizing global statistical information. GloVe vectors can also well represent the semantic relationship between words.

[0051] FastText extends the concept of Word2Vec by not only considering the entire word as the basic unit, but also decomposing it into character n-grams, thereby having better representation for unknown words.

[0052] ELMo uses a bidirectional long short-term memory network to dynamically generate embeddings based on the position of each word in its context. It provides deeply contextualized word representations and can better understand the different meanings of polysemous words.

[0053] BERT uses the Transformer architecture and learns deep contextualized embeddings through pre-training tasks such as masked language model MLM and next sentence prediction NSP. It can generate word representations for specific contexts, greatly improving the results of various NLP tasks.

[0054] The vocabulary vectors corresponding to these technologies can reflect the different needs of vocabulary vectors in medical clinical research, and realize the evaluation of medical research based on these different needs.

[0055] The vocabulary vectors are matched for similarity, and a disease-specific data set of the vocabulary vectors is set according to the value of the similarity matching; the disease-specific data set of the vocabulary vectors is also set by: extracting the medication type and medication frequency of the vocabulary vector according to the detection items and basic information corresponding to the vocabulary vectors, and setting the medical keywords of the vocabulary vectors according to the medication type and medication frequency, mapping the medical keywords with the vocabulary vectors, and matching the medical keywords for similarity according to the distribution of the vocabulary vectors to obtain the cosine similarity of the medical keywords. When the cosine similarity of the medical keywords is greater than the preset similarity threshold, the corresponding vocabulary vectors are formed into a disease-specific data set. At this time, the medical keywords corresponding to the vocabulary vectors are used for similarity matching to realize the step of similarity matching of the vocabulary vectors. At the same time, the medical keywords are matched for similarity in the following manner: according to the distribution position of the vocabulary vectors, the cosine similarities of the closest group of medical keywords are compared one by one, and the direction in which the cosine similarity of the medical keywords takes the maximum value is set as the direction in which the medical keywords are matched for similarity; the cosine similarity calculated for the medical keywords in the direction of similarity matching is output as the value of the vocabulary vector for similarity matching.

[0056] The preset similarity threshold is set using the average value of the cosine similarity set for the current basic information and the detection item using historical data.

[0057] Extract the cluster center of the vocabulary vector from the disease-specific dataset, and set the cluster center as the research node. The cluster center extracted here represents the vocabulary after the vocabulary vector in the disease-specific dataset is clustered after the disease-specific dataset is constructed. These words represent the key words corresponding to the multiple types of disease-specific datasets divided for disease research and analysis. The analysis of these medical data can be completed based on these key words. The cluster center of the vocabulary vector can be implemented using K-means, hierarchical clustering, DBSCAN, etc.; since the cluster center of the vocabulary vector will be used as a research node for subsequent use, when selecting the cluster center, the vocabulary vector is first clustered to obtain multiple clusters. Each cluster is allocated according to the value of the vocabulary vector, and the vocabulary vector corresponding to the average value of the cluster is selected as the cluster center.

[0058] After obtaining the research node, it is necessary to verify the type of the current research node, for example, to determine whether the research node belongs to the initial evaluation node, treatment intervention node, mid-term evaluation node, or final evaluation node; so that the research node can include these four types, and each type corresponds to at least one research node.

[0059] The subsequent processing of the research node and the implementation method of determining the target relationship vector corresponding to the research node include: for each research node, according to the position of the research node on the research path, obtaining the alternative setting label of the research node within a preset time period, and setting the target alternative value for the research node; when the target alternative value on the research node is not less than the first threshold, the research node is judged to be a key node, and the corresponding research node is composed of a target relationship vector according to the value of the target alternative value; when the target alternative value on the research node is less than the first threshold, the research node is judged to be a node to be improved, and the corresponding research node is traversed according to the alternative setting label of the research node, and the research node with the largest similarity value after traversal is set as the target relationship vector.

[0060] The above-mentioned alternative setting label represents the label set for the specific content of the current research node. At the same time, this label will also indicate that the research node is in the four types of situations: initial evaluation node, treatment intervention node, mid-term evaluation node and final evaluation node. The target alternative value is used to quantify the importance or urgency of a node in the entire study. For example, in a cancer treatment study, the initial evaluation node may be assigned a higher label value because it directly affects all subsequent decisions. When the target alternative value of the research node reaches or exceeds the preset threshold, the system will automatically trigger the corresponding processing mechanism; this mechanism can be to notify researchers to perform further operations or directly enter the next research stage; the first threshold used here is to use the average value of the target alternative value set by the research node in the historical data to represent the implementation of the research at this time.

[0061] The preset time period under the current treatment means the time required for the research node to conduct the research normally. For example, the initial evaluation node is selected within the last month before the patient is enrolled, with the purpose of collecting baseline data, including medical history, laboratory test results, etc. The treatment intervention node is selected from the 0th day to the 30th day after the start of treatment, with the purpose of recording the specific implementation of the first treatment and early response; the mid-term evaluation node is selected around the 90th day after the start of treatment, with the purpose of evaluating the treatment effect, such as changes in tumor size, biomarker levels, etc., and deciding whether to continue the current treatment plan; the final evaluation node is selected at the end of the entire research cycle (such as two years), with the purpose of summarizing all data for statistical analysis, drawing conclusions and writing research reports. Different preset time periods will be required for different types of current research nodes to complete the evaluation of the entire medical research.

[0062] When a research node is a node to be improved, traversing the research nodes according to the alternative setting labels of the research nodes is essentially to calculate the similarity between the research nodes under different alternative setting labels. The calculated similarity will adopt the value of cosine similarity, and the research node with the maximum cosine similarity calculated under the corresponding alternative setting label will be set as the target relationship vector. This set target relationship vector can reflect the outstanding problems and relevance of the research node when the importance of the research node is relatively low, which is convenient for tracing back the medical research conducted.

[0063] In one embodiment of the present invention, the indicator setting module mainly extracts the research indicator through the target relationship vector and obtains the research text data corresponding to the research indicator.

[0064] When extracting indicators for the target relationship vector, the main indicators extracted are efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators; and according to the form of these indicators, research indicators corresponding to the research nodes are set. Finally, the research indicators can evaluate the medical research to judge the efficacy of the corresponding patients in the corresponding situations, and whether the implementation of the medical plan is effective.

[0065] Efficacy indicators: such as changes in tumor size, biomarker levels, etc.

[0066] Safety indicators: incidence and severity of side effects, etc.

[0067] Patient status indicators: quality of life score, symptom improvement, etc.

[0068] Treatment compliance indicators: frequency of medication, follow-up completion, etc.

[0069] Time-related indicators: the time interval between each research node, the time point when the treatment response occurs, etc.

[0070] The implementation method of the indicator setting module includes: extracting indicators from the target relationship vector, obtaining the efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators in the target relationship vector; setting the research indicators corresponding to the research nodes according to the efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators; if the type and number of research indicators corresponding to the current target relationship vector are less than the preset number of indicators, then obtaining the target relationship vector adjacent to the current target relationship vector, and selecting the data in the adjacent target relationship vector except the research indicators corresponding to the current target relationship vector and adding them to the research indicators corresponding to the current target relationship vector, thereby determining the number of research indicators.

[0071] The type of research indicator is expressed as any one or more of efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators.

[0072] In clinical research, some nodes may lack sufficient information due to various reasons, such as missing data, incomplete records, etc. By obtaining relevant data from adjacent target relationship vectors, these missing data can be supplemented, thereby improving the completeness of the overall data set. Each target relationship vector should cover all predetermined research indicator types. If the data of a certain node is insufficient to cover all types of indicators, obtaining relevant data from adjacent nodes can help ensure that all types of indicators are considered.

[0073] Moreover, when setting the preset number of indicators for research indicators, each study should at least meet the average value of the overall treatment, which will put the research on a certain drug in a more stable state; by obtaining data from adjacent target relationship vectors, the research path can be dynamically adjusted to ensure that each research node can meet the preset indicator requirements. This helps to optimize the design of the entire study and make it more scientific and reasonable. Data from different research nodes may provide different perspectives and information. Integrating data from multiple research nodes can enable a more comprehensive multi-dimensional analysis, thereby better understanding the interactions between various variables.

[0074] Suppose a clinical trial of a new anticancer drug is underway: Initial assessment node: This node originally only contains the patient's baseline health status and medical history, but lacks the important efficacy indicator of gene mutation status. By obtaining relevant gene mutation information from adjacent treatment intervention nodes, this missing data can be supplemented to obtain a more complete description of patient characteristics.

[0075] Treatment intervention node: This node records the detailed treatment plan, but does not include the patient's quality of life score. By obtaining quality of life score data from the mid-term assessment node, this missing indicator can be supplemented, enabling a more comprehensive assessment of the treatment effect.

[0076] Final evaluation node: This node summarizes the data of the entire study cycle, but due to incomplete follow-up data of some patients, the treatment compliance index is insufficient. This deficiency can be compensated by obtaining follow-up records of some patients from the adjacent mid-term evaluation nodes to ensure a comprehensive evaluation of treatment compliance.

[0077] In this way, not only can the data quality and integrity of individual nodes be improved, but a more coherent and comprehensive data network can also be formed throughout the entire research path, thus providing a solid foundation for subsequent analysis and decision-making.

[0078] The supplemented and processed data are output as research text data, and then the comparison of the process is improved according to the research text data and the corresponding diagnostic behavior. The research text data will contain multiple research nodes, research indicators, and data on the corresponding target relationship vector, which will represent the data needed for the evaluation of the corresponding patients and drugs.

[0079] In one embodiment of the present invention, the process verification module mainly verifies the diagnostic behavior existing in the research text data, determines whether there is a difference in the diagnostic behavior under the process corresponding to the research indicator, and generates the degree of process difference.

[0080] When processing research text data, the waveform factor and kurtosis factor of the diagnostic behavior in the corresponding process will be obtained according to the implementation of the diagnostic behavior. The waveform factor represents the frequency of implementation of the diagnostic behavior, and the kurtosis factor represents the peak value of the research indicator of the corresponding patient after the implementation of the diagnostic behavior. According to these two factors, the patient's response to the diagnostic treatment is used to describe the degree of process difference of the research indicator.

[0081] Diagnostic behavior refers to the process by which medical professionals identify, classify, and evaluate diseases based on symptoms, signs, laboratory test results, and other relevant information provided by patients. These behaviors cover multiple stages from initial diagnosis to follow-up and are the basis for formulating treatment plans. Diagnostic behavior has obvious effects in initial diagnosis, multiple diagnoses, follow-up diagnoses, disease tracking, and treatment effect evaluation.

[0082] For example, information collected when first contacting the patient, such as medical history, physical examination, preliminary laboratory or imaging test results, is collected, and a preliminary judgment is made based on the collected information, and then the data in these judgments are identified as diagnostic behavior to identify the content that needs to be evaluated and processed at this time. For example, the patient's chief complaint, past medical history, family history, etc., information is obtained through inspection, palpation, percussion, auscultation, etc., and may also include basic examinations such as blood tests, urine analysis, and X-rays.

[0083] After that, other information based on the preliminary diagnosis is obtained. This information is generally tests performed, such as CT scans, MRI, PET-CT, etc., as well as pathological examinations for special pathologies; these data will also be considered as processes for corresponding research indicators, and the collected data will be analyzed.

[0084] At the same time, the condition will be tracked and multiple examinations will be conducted to find the corresponding data for a single research indicator, as well as whether there is relief and effective treatment of the condition during the treatment process. These data will be used as the process corresponding to the research indicator and compared with the normal treatment method to obtain the degree of difference in the processes of different research indicators.

[0085] like Figure 3 As shown, the implementation method of the process verification module includes: obtaining the waveform factor of the frequency of implementation of the diagnostic behavior and the kurtosis factor of the peak value of the corresponding research indicator after the implementation of the diagnostic behavior.

[0086] Construct a decision tree corresponding to the waveform factor and the kurtosis factor, and determine the degree of process difference of each research indicator under the corresponding process according to the output results of each leaf node on the decision tree.

[0087] Constructing a decision tree corresponding to the waveform factor and the kurtosis factor also includes: recording the scene information of the waveform factor, and when the scene information of the current waveform factor belongs to the top-level scene information, setting the current waveform factor as the root node of the decision tree.

[0088] At this time, the scene information of the waveform factor is included in the scene for obtaining the waveform factor, such as the electrocardiogram data currently obtained and the data recorded for each diagnosis. When the scene information is at the top level, it means that the current diagnosis is the initial diagnosis, or it is the data containing the most scene information. This data can cover subsequent more specific scenes to facilitate the representation of the current implementation process.

[0089] For example, in a decision tree for heart disease diagnosis, the root node may be "data source and type = electrocardiogram signal", and then branch out to secondary nodes such as different diagnostic behavior implementation frequencies, environmental parameters, and preprocessing steps. Each leaf node will eventually output a specific combination of waveform factor and kurtosis factor, and evaluate the degree of difference of the research indicators under the corresponding process.

[0090] Calculate the Gini impurity of the waveform factor and the kurtosis factor, and divide the waveform factor and the kurtosis factor into intermediate nodes and leaf nodes in the decision tree according to the value of Gini impurity. At this time, the intermediate node will be divided into multiple branches according to the value of Gini impurity, and each branch contains at least one set of waveform factors and kurtosis factors; then sort the intermediate nodes on each branch from left to right according to the value of the waveform factor, and distribute the kurtosis factor to multiple intermediate nodes according to the value of the kurtosis factor.

[0091] Iterate the leaf nodes. When the Gini impurity values ​​of the waveform factor and the kurtosis factor on the leaf nodes are both maximum values, the corresponding leaf nodes are output as the output results of each leaf node on the decision tree.

[0092] The leaf nodes output at this time represent the best effect of a specific diagnosis and treatment path; for example, in the diagnosis of heart disease, a leaf node may represent a specific electrocardiogram waveform pattern (waveform factor) and its corresponding peak characteristics (peakedness factor), which indicate the best treatment response; based on this information, patient groups in specific conditions can be identified, and doctors can adjust treatment plans for patient groups based on this information to improve the treatment effect on patients.

[0093] For each waveform factor and kurtosis factor, calculate their Gini impurity. Gini impurity measures the probability of misclassifying an element when it is randomly selected from the data set. Based on the value of Gini impurity, the waveform factor and kurtosis factor are divided into intermediate nodes and leaf nodes. The attributes with the largest reduction in Gini impurity are preferentially selected for segmentation.

[0094] In the scenario of processing electrocardiograms, waveform factors may include QRS complex width, PR interval length, etc.; kurtosis factors may include the kurtosis value of the QRS complex, the kurtosis value of the T wave, etc.; the QRS complex width is used as the first split point to divide it into two branches: the QRS complex width is wider and the narrower, and the Gini impurity of the remaining factors is continued to be calculated to select the next optimal split point, such as the PR interval length. On each branch, the intermediate nodes are sorted according to the value of the waveform factor; for example, in the branch with a wider QRS complex width, it is further subdivided according to the PR interval length. The kurtosis factor is assigned to the corresponding intermediate node according to the value of the kurtosis factor; for example, a QRS complex with a high kurtosis is assigned to one node, and a QRS complex with a low kurtosis is assigned to another node. The leaf nodes are iterated until a leaf node with the maximum Gini impurity is found; for example, a leaf node represents a patient group with a wide QRS complex width and high kurtosis and a long PR interval length, which indicates a higher risk of heart disease.

[0095] The leaf nodes of the final output will represent the values ​​of the QRS complex width of the waveform factor, the PR interval length, the kurtosis value of the QRS complex, and the kurtosis value of the T wave; these are only for the scenarios of electrocardiogram processing. When processing different scenarios, the output content will be output according to the content of the corresponding research indicators to obtain multiple values.

[0096] After obtaining the output results of each leaf node, determining the degree of process difference of each research indicator under the corresponding process includes: processing the values ​​on the leaf nodes, calculating the average values ​​of the waveform factor and kurtosis factor on each leaf node respectively, and calculating the combined standard deviation of the waveform factors and kurtosis factors on all leaf nodes; the combined standard deviation represents the standard deviation values ​​of all waveform factors and kurtosis factors.

[0097] After subtracting the average value of the waveform factor and kurtosis factor on each leaf node from the average value of the waveform factor and kurtosis factor on the adjacent leaf node, divide it by the combined standard deviation of the waveform factor and kurtosis factor; finally, the ratio of the waveform factor and the kurtosis factor is obtained, and the sum of the ratio of the waveform factor and the ratio of the kurtosis factor is taken as the process difference degree of a single leaf node; the process difference degree of a single leaf node is averaged and output as the process difference degree of the research indicator under the corresponding process.

[0098] In one embodiment of the present invention, the detection difference module mainly verifies whether the sample requirement used by the current research indicator under the process difference degree can achieve a significant effect, and generates the detection project difference result according to the result obtained by this sample requirement.

[0099] At this time, the sample requirement will be set according to the degree of difference of the research indicators under the corresponding process. The main purpose is to improve the overall calculation, the significance level and test efficiency during the calculation, so as to verify the implementation of the project in the medical evaluation scenario.

[0100] At this time, the sample demand will be calculated according to the value of the process difference degree and the corresponding research indicators. The correlation value will be calculated between the process difference degree available under the current research indicator and the sample demand used for the corresponding research indicator in the historical data. The currently selected sample demand will be adjusted according to the correlation value, and finally the sample demand and test item difference results will be obtained. The test item difference results will include the sample demand, research indicators, process difference degree and other data at this time.

[0101] The implementation method of estimating the sample demand under the current process difference degree includes: calculating the Euclidean distance between the current process difference degree and the historical difference degree in the historical data, finding the historical difference degree closest to the current process difference degree as the first correlation value of the current difference degree; this first correlation value is to find the data with the smallest difference between the current process difference degree and the historical difference degree. The historical difference degree is the process difference degree that has been calculated for the current research indicator. The Euclidean distance is often used as a measure of the similarity between two data points. The smaller the distance, the more similar the two points are; the larger the distance, the less similar the two points are. At this time, the Euclidean distance between the process difference degree and the historical difference degree is calculated using the values ​​of these two data to determine the current process difference degree.

[0102] Compare the number of inconsistencies in the data between the current process difference degree and the historical difference degree in the historical data, and use the number of inconsistencies as the second correlation value; the second correlation value measures the monotonic relationship between the current process difference degree and the historical difference degree, that is, whether the process difference degree will change due to the setting size of the sample requirement, and mark the number of inconsistencies identified at this time to verify whether the current research indicator is affected by the sample requirement during the research.

[0103] The Pearson correlation coefficient is calculated by combining the first correlation value and the second correlation value with the preset correlation threshold. When the Pearson correlation coefficient takes the maximum value, the sample demand corresponding to the preset correlation threshold is used as the sample demand under the current process difference degree. At this time, the Pearson correlation coefficient is calculated by calculating the first correlation value and the second correlation value with the preset correlation threshold. The preset correlation threshold will set the corresponding threshold for the first correlation value and the second correlation value. At the same time, the preset correlation threshold will correspond to the sample demand. After completing the calculation of the Pearson correlation coefficient, the corresponding sample demand can be directly obtained.

[0104] After obtaining the sample demand, the standard test item results are compared according to the sample demand. According to the current sample demand, the standard test item results with the same value as the sample demand are selected, and the current research indicator is compared to see if they are consistent with the standard test item results. The values ​​under the research indicator, the calculated standard deviation, the average value and the standard test item results are output together to obtain the test item difference results. At this time, the compared items will be set in adjacent positions in the table to facilitate the comparison of the specific difference positions between the standard item test results and the current research indicator.

[0105] Comparing the results of standard test items according to the sample demand to generate the test item difference results can be expressed according to the following example, using two processes to express the relative difference at this time; for example, the diagnostic accuracy of process A is 85%, the standard deviation is 5%; the patient satisfaction score is 4.5 (full score 5 points), the standard deviation is 0.5; the average treatment cost is 100 yuan, and the standard deviation is 10 yuan; the diagnostic accuracy of process B is 75%, the standard deviation is 6%; the patient satisfaction score is 4.0, the standard deviation is 0.6; the average treatment cost is 90 yuan, and the standard deviation is 12 yuan. Through these two processes, we can know the relative situation of the current clinical medical evaluation. Only part of the evaluation values ​​are pointed out here, and the efficacy and implementation effects of different research indicators under the current research node will also be included, thereby reflecting the overall research content.

[0106] In one embodiment of the present invention, the result analysis module mainly combines all the detection project difference results according to the positions of the research indicators on the research nodes to obtain a comprehensive trend display report to obtain the research evaluation report at this time, so as to facilitate subsequent staff to process this report.

[0107] In the result analysis module, the data in the detection difference module will be converted into a trend chart, and the converted data will be output as a research evaluation report; at this time, the research indicators will be aggregated according to the position of the research nodes, and the final research evaluation report will be displayed in the form of research indicators and research node distribution. The final report will fully display the situation under different research indicators.

[0108] For example, the initial assessment node mainly displays baseline data (medical history, laboratory results), and then draws a bar chart to show the comparison of baseline characteristics of the two groups of patients.

[0109] The treatment intervention node mainly displays the implementation of specific treatment measures and the setting of follow-up time points, and uses a line graph to show the trend of condition changes in the two groups of patients after treatment.

[0110] The mid-term assessment node mainly displays the changes in the levels of specific biomarkers and the results of imaging examinations, and uses a box plot to show the distribution of various indicators of the two groups of patients at the mid-term assessment.

[0111] The final evaluation node mainly displays the data summary, statistical analysis and conclusions at the end of the entire study cycle, and uses a radar chart to show the comprehensive performance of the two groups of patients during the entire study period.

[0112] These nodes represent the specific calculation of the research indicators, and extract the statistical mean and standard deviation of each research node for aggregate display.

[0113] like Figure 4 As shown, the present invention also provides an AI-based clinical medical research evaluation method, including: S1, obtaining basic information and test items during clinical medical research, and setting a research path corresponding to the test items.

[0114] S2, set up multiple research nodes on the research path, verify the information of each node, and determine the target relationship vector.

[0115] S3, extract indicators from the target relationship vector, set research indicators that match the research nodes, and obtain research text data.

[0116] S4, identify the diagnostic behavior of the research text data, process the processes corresponding to the research indicators according to the diagnostic behavior, and evaluate the degree of process difference.

[0117] S5, calculate the sample demand under the current process difference degree, and compare the standard test item results according to the sample demand to generate the test item difference results.

[0118] S6, analyze the difference results of the test items and combine them into a research evaluation report according to the performance trends.

[0119] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention and they are still covered by the protection scope of the present invention.

Claims

1. A clinical medical research evaluation system based on AI, characterized in that: include: Research generation module, used to obtain basic information and test items in clinical medical research, and set research paths corresponding to the test items; A research verification module is used to set multiple research nodes on the research path, verify the information of each research node on the research path, and determine the target relationship vector corresponding to the research node; The indicator setting module is used to extract indicators from the target relationship vector, set research indicators that match the research nodes, and obtain research text data corresponding to the research indicators; The implementation methods of the indicator setting module include: Extract indicators from the target relationship vector to obtain efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators in the target relationship vector; set research indicators corresponding to the research nodes according to the efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators; if the type and number of research indicators corresponding to the current target relationship vector are less than the preset number of indicators, obtain the target relationship vector adjacent to the current target relationship vector, and select the data in the adjacent target relationship vector except the research indicator corresponding to the current target relationship vector and add it to the research indicator corresponding to the current target relationship vector; The process verification module is used to identify the diagnostic behavior of the research text data, and process the process corresponding to the research indicator according to the diagnostic behavior, and evaluate the degree of process difference of each research indicator under the corresponding process; The detection difference module is used to calculate the sample demand under the current process difference degree, and compare the standard detection project results according to the sample demand to generate the detection project difference results; The result analysis module is used to analyze the difference results of the test items and combine the difference results of the test items into a research evaluation report according to the performance trend of the difference results of the test items at the research nodes.

2. According to claim 1, a clinical medical research evaluation system based on AI is characterized in that: The implementation method of the research inspection module includes: constructing a research text library corresponding to the research path, wherein the research text library contains multiple vocabulary vectors corresponding to the current basic information and the inspection items; Perform similarity matching on the vocabulary vectors, and set the disease-specific dataset of the vocabulary vectors according to the similarity matching values; Extract the cluster centers of vocabulary vectors from the disease-specific dataset and set the cluster centers as research nodes.

3. According to claim 2, an AI-based clinical medical research evaluation system is characterized in that: The disease-specific datasets for setting vocabulary vectors also include: According to the detection items and basic information corresponding to the vocabulary vector, the medication type and frequency of the vocabulary vector are extracted, and according to the medication type and frequency of medication, the medical keywords of the vocabulary vector are set, and the medical keywords are mapped with the vocabulary vector. According to the distribution of the vocabulary vector, the medical keywords are matched by similarity to obtain the cosine similarity of the medical keywords. When the cosine similarity of the medical keywords is greater than the preset similarity threshold, the corresponding vocabulary vectors are combined into a special disease data set.

4. The AI-based clinical medical research evaluation system according to claim 1, characterized in that: The implementation methods of determining the target relationship vector corresponding to the research node include: For each research node, according to the position of the research node on the research path, the alternative setting label of the research node within the preset time period is obtained, and a target alternative value is set for the research node; when the target alternative value on the research node is not less than a first threshold, the research node is judged to be a key node, and the corresponding research node is composed of a target relationship vector according to the value of the target alternative value; when the target alternative value on the research node is less than the first threshold, the research node is judged to be a node to be improved, and the corresponding research node is traversed according to the alternative setting label of the research node, and the research node with the largest similarity value after traversal is set as the target relationship vector.

5. The AI-based clinical medical research evaluation system according to claim 1, characterized in that: The implementation method of the process inspection module includes: obtaining a waveform factor regarding the frequency of implementation of the diagnostic behavior and a kurtosis factor regarding the peak value of the corresponding research indicator after the implementation of the diagnostic behavior; Construct a decision tree corresponding to the waveform factor and the kurtosis factor, and determine the degree of process difference of each research indicator under the corresponding process according to the output results of each leaf node on the decision tree.

6. The AI-based clinical medical research evaluation system according to claim 5, characterized in that: Constructing a decision tree corresponding to the waveform factor and the kurtosis factor also includes: Record the scene information of the waveform factor, and when the scene information of the current waveform factor belongs to the top-level scene information, set the current waveform factor as the root node of the decision tree; Calculate the Gini impurity of the waveform factor and the kurtosis factor, and divide the waveform factor and the kurtosis factor into intermediate nodes and leaf nodes in the decision tree according to the values ​​of the Gini impurity; Iterate the leaf nodes. When the Gini impurity values ​​of the waveform factor and the kurtosis factor on the leaf nodes are both maximum values, the corresponding leaf nodes are output as the output results of each leaf node on the decision tree.

7. The AI-based clinical medical research evaluation system according to claim 6, characterized in that: The degree of process difference in evaluating each research indicator under the corresponding process includes: Process the values ​​on the leaf nodes, calculate the average values ​​of the waveform factor and the kurtosis factor on each leaf node, and calculate the combined standard deviation of the waveform factor and the kurtosis factor on all leaf nodes; After subtracting the average value of the waveform factor and the kurtosis factor on each leaf node from the average value of the waveform factor and the kurtosis factor on the adjacent leaf node, divide it by the combined standard deviation of the waveform factor and the kurtosis factor to obtain the corresponding ratio of the waveform factor and the kurtosis factor. The sum of the ratio of the waveform factor and the ratio of the kurtosis factor is taken as the process difference degree of a single leaf node; the process difference degree of a single leaf node is averaged and output as the process difference degree of the research indicator under the corresponding process.

8. The AI-based clinical medical research evaluation system according to claim 1, characterized in that: The implementation methods for estimating the sample demand under the current process difference degree include: Calculate the Euclidean distance between the current process difference degree and the historical difference degree in the historical data, and find the historical difference degree closest to the current process difference degree as the first correlation value of the current difference degree; comparing the current process difference degree with the number of data inconsistencies of the historical difference degree in the historical data, and using the number of inconsistencies as a second correlation value; The Pearson correlation coefficient is calculated by using the first correlation value, the second correlation value and the preset correlation threshold. When the Pearson correlation coefficient is maximized, the sample demand corresponding to the preset correlation threshold is used as the sample demand under the current process difference degree.

9. A clinical medical research evaluation method based on AI, characterized in that: include: S1, obtain basic information and test items for clinical medical research, and set up research paths corresponding to the test items; S2, setting multiple research nodes on the research path, verifying the information of each node, and determining the target relationship vector; S3, extracting indicators from the target relationship vector, obtaining efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators in the target relationship vector; setting research indicators corresponding to the research nodes according to the efficacy indicators, safety indicators, patient status indicators, treatment compliance indicators and time-related indicators; if the types and quantities of research indicators corresponding to the current target relationship vector are less than the preset number of indicators, obtaining target relationship vectors adjacent to the current target relationship vector, and selecting data from adjacent target relationship vectors except for the research indicators corresponding to the current target relationship vector and adding them to the research indicators corresponding to the current target relationship vector, thereby obtaining research text data; S4, identify the diagnostic behavior of the research text data, process the processes corresponding to the research indicators according to the diagnostic behavior, and evaluate the degree of process difference; S5, calculating the sample demand under the current process difference degree, and comparing the standard test item results according to the sample demand to generate the test item difference results; S6, analyze the difference results of the test items and combine them into a research evaluation report according to the performance trends.

Images