Internal medicine severe illness clinical decision-making method and system based on multi-modal pre-training large model

Abstract

The invention relates to the technical field of medical severe clinical decision making, in particular to a medical severe clinical decision making method and system based on a multi-modal pre-training large model. The method comprises the following steps: performing focus description mapping on an internal medicine case focus image according to internal medicine case report data to obtain internal medicine case focus description data; performing focus diffusion behavior prediction according to the internal medicine case focus description data to obtain focus diffusion behavior prediction data; performing focus multi-dimensional influence degree evaluation on the focus diffusion behavior prediction data to obtain focus diffusion multi-dimensional influence degree data; performing depth feature extraction on the nidus diffusion multi-dimensional influence degree data to obtain nidus diffusion depth influence data; and performing clinical decision reference coupling and interactive learning according to the focus diffusion depth influence data to obtain focus diffusion decision reference interactive data. According to the method, the clinical decision-making technology is optimized, so that the clinical decision-making technology for serious illness in the internal medicine department is more perfect.

Description

Technical Field

[0001] The present invention relates to the technical field of clinical decision-making for critical internal medicine diseases, and in particular to a clinical decision-making method and system for critical internal medicine diseases based on a multimodal pre-trained large model. Background Art

[0002] With the advancement of medical technology, the field of critical care in internal medicine is facing increasingly complex challenges. Critically ill patients usually have multiple complex diseases, and their diseases progress rapidly and changeably. Traditional single-mode data analysis methods often find it difficult to provide comprehensive and accurate decision support. In this context, clinical decision-making methods based on multimodal pre-trained large models have emerged. Multimodal pre-trained large models can not only process a variety of different types of data, including medical images, electronic health records, genomic data, etc., but also provide more comprehensive disease diagnosis and treatment recommendations by integrating these diverse data sources. The core of this method lies in its powerful data processing and model training capabilities. It can learn disease patterns and clinical characteristics from massive historical data and predict disease development trends. The method based on multimodal pre-trained large models can better capture the multidimensional information of the disease through its deep learning and multi-level feature extraction capabilities, and reveal the potential laws hidden in complex data. Its training process utilizes large-scale medical data sets, so that it can quickly adapt and update the knowledge base when facing new diseases, improving the timeliness and accuracy of decision-making. However, the traditional clinical decision-making method for critical internal medicine diseases based on multimodal pre-trained large models has the problem of inaccurate analysis of the development trend of critical internal medicine diseases, resulting in low accuracy of clinical decision-making. Summary of the invention

[0003] Based on this, it is necessary to provide a method and system for clinical decision-making in critical internal medicine based on a multimodal pre-trained large model to solve at least one of the above technical problems.

[0004] To achieve the above purpose, a clinical decision-making method for critical medical diseases based on a multimodal pre-trained large model is provided, and the method comprises the following steps: Step S1: Obtaining medical case report data and medical case lesion images; performing lesion description mapping on the medical case lesion images according to the medical case report data to obtain medical case lesion description data; Step S2: performing lesion morphological structure analysis of different lesion locations according to the lesion description data of internal medicine cases to obtain lesion morphological structure data; performing lesion diffusion behavior prediction on the lesion morphological structure data to obtain lesion diffusion behavior prediction data; Step S3: simulating the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multimodal pre-trained large model to obtain the lesion diffusion behavior trend data; evaluating the multidimensional impact degree of the lesion according to the lesion diffusion behavior trend data to obtain the multidimensional impact degree data of the lesion diffusion; Step S4: perform deep feature extraction on the multi-dimensional impact degree data of lesion spread to obtain lesion spread depth impact data; perform clinical decision reference coupling based on the lesion spread depth impact data to obtain lesion spread clinical decision reference data; perform interactive learning on the lesion spread clinical decision reference data to obtain lesion spread decision reference interactive data.

[0005] The medical case report data of the present invention provides a detailed description of the patient's condition, symptoms and test results. These data include diagnosis, medical history, laboratory results, etc., which are essential for understanding the development of the disease. The medical case lesion images provide visual lesion locations and features, such as CT and MRI scan images. These images can help doctors accurately locate the lesion area. Combining the case report data with the lesion images and performing lesion description mapping can more accurately describe and annotate the lesions from the perspective of semantics to images, thereby improving the accuracy of lesion identification. The lesion description data finally obtained can clearly present the characteristics, location and cause of the lesion, providing an important basis for subsequent analysis and treatment decisions. Based on the lesion description data, morphological and structural analysis of different lesion locations can be performed to deeply understand the specific morphological characteristics of the lesion in tissues and organs, including size, shape, density, etc. These analyses help to assess the severity of the lesion and its impact on surrounding tissues. Using the lesion morphological and structural data, the lesion diffusion behavior is predicted. This includes predicting the expansion speed and direction of the lesion and its impact on surrounding tissues. The resulting lesion spread behavior prediction data provides a quantitative prediction of disease progression, which is very important for determining the priority and urgency of treatment options. Based on the multimodal pre-trained large model, the lesion spread behavior prediction data is simulated. This means using advanced deep learning techniques to simulate and predict the spread trend of lesions in the future. This work helps to understand the dynamic changes of lesions and provide predictive capabilities for future treatment and intervention strategies. The resulting lesion spread behavior trend data provides detailed information on the changes of lesions in time and space. These data can be used to predict the progression rate, spread direction and impact range of the disease, and to evaluate the multidimensional impact of lesion spread based on the lesion spread behavior trend data. This step aims to deeply analyze the multi-faceted impact of lesion spread on surrounding tissues, organs and overall physical health, including physiological and functional changes. The obtained lesion spread depth impact data provides more detailed and quantitative impact information. These data not only include the specific impact area and degree of the lesion in the body, but also can evaluate its potential impact on quality of life and health prognosis. According to the lesion spread depth impact data, clinical decision reference coupling is performed. This step combines scientifically analyzed data with clinical practice, and the final lesion spread decision reference interactive data is obtained. Therefore, the present invention is an optimization processing of a traditional clinical decision-making method for critical internal medicine diseases based on a multimodal pre-trained large model, which solves the problem of inaccurate analysis of development trends of critical internal medicine diseases and resulting in low accuracy of clinical decisions in the traditional clinical decision-making method for critical internal medicine diseases based on a multimodal pre-trained large model, improves the accuracy of development trend analysis of critical internal medicine diseases, and improves the accuracy of clinical decisions.

[0006] Preferably, step S1 comprises the following steps: Step S11: Acquire medical case report data and medical case lesion images; Step S12: desensitizing the medical case report data and the medical case lesion images to obtain the medical case report desensitized data and the medical case lesion desensitized images; Step S13: extracting the internal medicine case description from the desensitized data of the internal medicine case report to obtain the internal medicine case description data; Step S14: performing lesion description mapping on the desensitized images of lesions in internal medicine cases according to the internal medicine case description data to obtain the lesion description data in internal medicine cases.

[0007] The medical case report data of the present invention includes the patient's medical history, clinical symptoms, and laboratory test results. The medical case lesion images include medical imaging examination results such as CT scans and MRI images. These images provide detailed visual information about the location, morphology, size, and surrounding tissue conditions of the lesion, which helps doctors accurately locate the lesion and assess its severity. Under the premise of protecting the privacy of the patient, the case report data is desensitized to remove or replace sensitive information that identifies personal identity, such as name, address, etc. This ensures the security and privacy protection of the data in medical research and analysis. The lesion images are desensitized, including removing or blurring the facial features of the patient to prevent the disclosure of personal identity information. This processing ensures the safe use and sharing of medical images while complying with relevant laws and ethical standards. Based on the desensitized case report data, key case description information, such as condition description, symptom description, diagnosis results, etc., is extracted from it. After extraction and structuring, these data become important inputs for subsequent analysis and modeling. According to the extracted medical case description data, the desensitized lesion images are mapped to lesion descriptions. This process combines text descriptions with medical images.

[0008] Preferably, step S2 comprises the following steps: Step S21: performing position distribution analysis on the lesion images of internal medicine cases to obtain lesion position distribution data; Step S22: performing lesion morphological structure analysis of different lesion locations on the lesion location distribution data according to the lesion description data of the internal medicine case to obtain lesion morphological structure data; Step S23: performing isodensity characteristics recognition on the lesion morphological structure data to obtain isodensity data of the lesion structure; Step S24: predicting the lesion diffusion behavior of the lesion morphological structure data according to the lesion structure and other density data to obtain lesion diffusion behavior prediction data.

[0009] The present invention determines and records the position distribution of lesions in different parts by analyzing the images of lesions in internal medicine cases. These data provide detailed information about the position and distribution of lesions in the body, which helps to understand the extensiveness and scope of influence of lesions. Based on the lesion description data, morphological and structural analysis is performed on lesions in different positions. This step involves the extraction and analysis of detailed features such as the shape, size, and edge features of the lesions, which helps to depict the morphological characteristics of the lesions. Based on the lesion morphological and structural data, isodense characteristic recognition is performed. This includes identifying the characteristics of the internal density distribution of the lesions, such as uniformity, concentration, etc. These data help to further understand the internal structural characteristics and complexity of the diseased tissue. Using the lesion structure isodense data, the diffusion behavior of the lesion is predicted. This process includes analyzing how the internal structural characteristics of the lesion affect its propagation mode and speed in the body. The prediction data can provide quantitative predictions about the progression of the disease, helping to assess the potential danger of the lesion and the degree of impact on the surrounding tissues.

[0010] Preferably, step S24 comprises the following steps: Step S241: performing density evolution law analysis on the density data of the lesion structure to obtain density evolution law data of the lesion structure; Step S242: performing lesion spatial variability analysis on the lesion morphological structure data according to the lesion structure density evolution law data to obtain lesion spatial variability data; Step S243: constructing a lesion diffusion prediction model based on the lesion structure density evolution law data and the lesion spatial variability data based on the policy gradient algorithm to obtain a lesion diffusion prediction model; Step S244: predicting the lesion diffusion behavior according to the lesion diffusion prediction model to obtain lesion diffusion behavior prediction data.

[0011] The present invention determines the changing trend and law of the density inside the lesion by analyzing the density data of the lesion structure. These data provide detailed information about how the internal structure of the lesion evolves over time or other factors, which is helpful to understand the dynamic process of the development of the lesion. According to the data of the evolution law of the density of the lesion structure, the spatial variability analysis of the lesion is performed. This step involves evaluating the differences in the structural characteristics of the lesion at different parts, such as the changes in the morphology and density in different regions. These data help to reveal the complexity and diversity of the lesion in space. Based on the policy gradient algorithm, combined with the data of the evolution law of the density of the lesion structure and the data of the spatial variability of the lesion, a prediction model for the spread of the lesion is constructed. This model can systematically analyze the influence of the evolution of the lesion structure and the spatial variation on the spread behavior of the lesion, and provide predictive data support for the future development trend of the lesion. The constructed lesion spread prediction model is used to predict the spread behavior of the lesion. This includes analyzing the diffusion path, speed and influence range of the lesion in time and space according to the model. The prediction data helps to evaluate the progress of the lesion and the potential impact on the surrounding environment.

[0012] Preferably, constructing a lesion diffusion prediction model based on the lesion structure density evolution law data and the lesion spatial variability data comprises the following steps: Based on the data of evolution law of lesion structure density and spatial variability of lesions, spatial correlation analysis was performed to obtain spatial correlation data of lesion evolution; Calculate the local evolution correlation rate of the lesion evolution spatial correlation data to obtain the local evolution correlation rate data; Calculate the rate variance of the local evolution correlation rate data to obtain the local evolution rate variance data; Based on the policy gradient algorithm, the local evolution rate variance data is continuously evolved to learn the state, and the local continuous evolution state data is obtained; A lesion diffusion prediction model is constructed according to the local continuous evolution state data to obtain a lesion diffusion prediction model.

[0013] The present invention determines the spatial correlation of lesions by analyzing the evolution law data of lesion structure density and the spatial variability data of lesions. This step reveals the structural relationship between lesions in different regions, and provides the basis of spatial information for subsequent steps. According to the spatial correlation data of lesion evolution, the correlation rate of lesion evolution in different local regions is calculated. This calculation helps to identify the development speed and change trend of lesions in different parts, and provides basic data for subsequent rate variance calculation. The local evolution correlation rate data is statistically analyzed to calculate the variance of the lesion evolution rate in each local region. This step helps to identify and quantify the variability of the evolution speed of lesions in different regions, and provides input data for subsequent continuous evolution state learning. Based on the policy gradient algorithm, the local evolution rate variance data is learned to generate data describing the continuous evolution state of the lesion in time. This process can capture the dynamic changes of lesions at different time points, and lays the foundation for establishing a more accurate diffusion prediction model.

[0014] Preferably, step S3 comprises the following steps: Step S31: simulating the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multimodal pre-trained large model to obtain lesion diffusion behavior trend data; Step S32: performing spatial interpolation processing on the lesion diffusion behavior trend data to obtain lesion diffusion behavior spatial interpolation data; Step S33: performing hysteresis interpolation segmentation on the spatial interpolation data of the lesion diffusion behavior to obtain diffusion hysteresis interpolation segmentation data; Step S34: evaluating the multi-dimensional impact degree of lesions according to the diffusion hysteresis interpolation segmentation data and the lesion diffusion behavior trend data to obtain the multi-dimensional impact degree data of lesion diffusion.

[0015] The present invention utilizes a multimodal pre-trained large model to simulate the behavioral trend of lesion diffusion. This step can reveal the potential laws and future trends of lesion diffusion through the predictive ability of the model, and provide accurate diffusion behavior data for subsequent steps. The diffusion behavior trend data obtained in step S31 is subjected to spatial interpolation processing. This process can fill the data gaps in space and ensure the continuity and integrity of the diffusion trend distribution in the entire region. Based on the spatial interpolation data obtained in step S32, a hysteresis interpolation segmentation process is performed. This step allows the continuous diffusion trend data to be segmented into different time periods or spatial segments to more accurately describe the spatiotemporal variation characteristics of lesion diffusion. Combined with the diffusion hysteresis interpolation segmentation data and the lesion diffusion behavior trend data, a multidimensional impact degree assessment is performed. This assessment can quantify the contribution of various factors in different time periods or spatial segments to lesion diffusion, help identify key influencing factors and optimize the accuracy and applicability of the prediction model.

[0016] Preferably, step S32 includes the following steps: Step S321: performing trend scale effect analysis on the lesion diffusion behavior trend data to obtain diffusion trend scale effect data; Step S322: Calculating the diffusion rate increment discontinuity interval for the lesion diffusion behavior trend data according to the diffusion trend scale effect data to obtain the diffusion rate increment discontinuity interval; Step S323: performing spatial trend nonlinear difference quantification according to the diffusion rate increment discontinuity interval to obtain diffusion spatial trend difference quantification data; Step S324: performing spatial interpolation processing according to the diffusion rate increment discontinuity interval and the diffusion spatial trend difference quantified data to obtain the lesion diffusion behavior spatial interpolation data.

[0017] The present invention performs trend scale effect analysis on the lesion diffusion behavior trend data. This analysis can reveal the influence and change rules of the lesion diffusion trend at different spatial scales, and provide important reference data on scale for subsequent steps. Based on the diffusion trend scale effect data, the rate increment discontinuity interval in the lesion diffusion behavior trend data is calculated. This step helps to identify and describe the mutation and change trend of the lesion diffusion rate in different time or space segments. According to the diffusion rate increment discontinuity interval, a nonlinear difference quantitative analysis of the spatial trend is performed. This analysis helps capture and quantify the nonlinear change characteristics of the lesion diffusion in different regions, and provides a more accurate data basis for subsequent spatial interpolation processing. Based on the diffusion rate increment discontinuity interval and the diffusion space trend difference quantitative data, a spatial interpolation process is performed. This processing process ensures the continuity and integrity of the lesion diffusion trend data in the entire region, thereby more accurately reflecting the actual situation of the lesion diffusion.

[0018] Preferably, step S34 includes the following steps: Step S341: dividing the diffusion hysteresis interpolation segmented data into hysteresis-affected regions to obtain diffusion hysteresis-affected region data; Step S342: performing regional impact superposition analysis based on the diffusion hysteresis impact regional data to obtain hysteresis regional impact superposition data; Step S343: Identify the multi-dimensional influencing factors of the lesions according to the lagging regional impact superposition data and the lesion diffusion behavior trend data to obtain the multi-dimensional influencing factors of the lesions; Step S344: Evaluate the multidimensional impact degree of the lesion based on the multidimensional impact factor of the lesion on the superposition data of the lagging regional impact, and obtain the multidimensional impact degree data of the lesion diffusion.

[0019] The present invention can effectively identify the hysteresis effect area existing in the process of lesion diffusion by dividing the hysteresis influence area of ​​diffusion hysteresis interpolation segmented data. This division helps to accurately determine the time delay influence area of ​​lesion diffusion, so that the spatial and temporal characteristics of lesion diffusion can be more clearly understood. This is crucial for capturing the subtle changes of lesion diffusion and helps to improve the accuracy of subsequent analysis. By performing regional impact superposition analysis on diffusion hysteresis influence area data, the superposition effect of multiple hysteresis influence areas can be comprehensively evaluated. This step helps to identify the interaction and cumulative impact between different regions, and provides strong data support for the global understanding of lesion diffusion. According to the hysteresis regional impact superposition data and the lesion diffusion behavior trend data, the lesion multi-dimensional influence factor identification is carried out, and the various key factors affecting the lesion diffusion can be accurately found. This process provides a deep understanding of each dimension of lesion diffusion, including the impact path and mechanism, thereby providing a detailed basis for subsequent evaluation and decision-making. The multi-dimensional impact degree of lesions is evaluated based on the superposition data of the impact of multi-dimensional influencing factors on the lagging regions, which can comprehensively consider the combined effects of various influencing factors and provide a detailed evaluation of the multi-dimensional impact degree of lesion spread. This evaluation result can fully reflect the complexity and diversity of lesion spread and improve the accuracy and reliability of clinical decision-making methods for critical internal medicine.

[0020] Preferably, step S4 comprises the following steps: Step S41: extracting depth features from the multi-dimensional impact degree data of lesion diffusion to obtain lesion diffusion depth impact data; Step S42: performing commonality identification on the lesion diffusion depth influence data to obtain lesion diffusion influence commonality data; Step S43: Acquire historical clinical decision data; Step S44: performing clinical decision reference coupling on historical clinical decision data according to the common data on the impact of lesion spread, to obtain clinical decision reference data on lesion spread; Step S45: interactively learning the lesion spread clinical decision reference data according to the Q-Learning algorithm to obtain lesion spread decision reference interactive data.

[0021] The present invention performs deep feature extraction on the multidimensional impact degree data of lesion diffusion, and can extract key deep-level features from complex multidimensional data. This step helps to reveal the core influencing factors and potential patterns of lesion diffusion, thereby providing a more refined and accurate data basis for subsequent analysis. Commonality identification is performed on the lesion diffusion depth impact data, and common features in the lesion diffusion process can be extracted. These common features can help understand the general laws and common influencing factors of lesion diffusion, so as to better capture the overall trend and pattern of lesion diffusion. Acquiring historical clinical decision data provides valuable background information and reference basis. These data contain past decision-making processes and results, which help to compare and verify the current analysis results with historical experience, and improve the scientificity and practicality of the decision-making process. According to the common data of lesion diffusion, the historical clinical decision data is coupled with clinical decision reference, and the newly extracted common features can be combined with the decision-making mode in the historical data. This step enables the current decision analysis to integrate historical experience and improve the rationality and consistency of decision-making. The Q-Learning algorithm is used to interactively learn the clinical decision reference data of lesion diffusion, and the decision strategy can be continuously optimized and adjusted. Through interactive learning, the system is able to adjust its strategy through repeated trials, thereby improving the flexibility and adaptability of the decision-making process and ensuring that reasonable adjustments can be made quickly when faced with new situations.

[0022] Preferably, the present invention further provides an internal medicine critical care clinical decision system based on a multimodal pre-trained large model, which is used to execute the internal medicine critical care clinical decision method based on a multimodal pre-trained large model as described above. The internal medicine critical care clinical decision system based on a multimodal pre-trained large model comprises: The lesion description analysis module is used to obtain the medical case report data and the medical case lesion images; perform lesion description mapping on the medical case lesion images according to the medical case report data to obtain the medical case lesion description data; The lesion diffusion behavior prediction module is used to perform lesion morphological structure analysis of different lesion locations based on the lesion description data of internal medicine cases to obtain lesion morphological structure data; perform lesion diffusion behavior prediction on the lesion morphological structure data to obtain lesion diffusion behavior prediction data; The multi-dimensional impact assessment module of lesions is used to simulate the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multi-modal pre-trained large model to obtain the lesion diffusion behavior trend data; the multi-dimensional impact assessment of lesions is performed based on the lesion diffusion behavior trend data to obtain the multi-dimensional impact data of lesion diffusion; The decision reference learning module is used to perform deep feature extraction on the multi-dimensional impact degree data of lesion diffusion to obtain the lesion diffusion depth impact data; perform clinical decision reference coupling based on the lesion diffusion depth impact data to obtain the lesion diffusion clinical decision reference data; and perform interactive learning on the lesion diffusion clinical decision reference data to obtain the lesion diffusion decision reference interactive data.

[0023] The beneficial effects of the present invention are that the medical case report data provides a detailed description of the patient's condition, symptoms and examination results. These data include diagnosis, medical history, laboratory results, etc., which are essential for understanding the development of the disease. The medical case lesion images provide visual lesion locations and features, such as CT and MRI scan images. These images can help doctors accurately locate the lesion area. Combining the case report data with the lesion images and performing lesion description mapping can more accurately describe and annotate the lesions from the perspective of semantics to images, thereby improving the accuracy of lesion identification. The lesion description data finally obtained can clearly present the characteristics, location and cause of the lesion, providing an important basis for subsequent analysis and treatment decisions. Based on the lesion description data, morphological and structural analysis of different lesion locations can be performed to deeply understand the specific morphological characteristics of the lesion in tissues and organs, including size, shape, density, etc. These analyses help to assess the severity of the lesion and its impact on surrounding tissues. Using the lesion morphological and structural data, the lesion diffusion behavior is predicted. This includes predicting the expansion speed and direction of the lesion and its impact on surrounding tissues. The resulting lesion spread behavior prediction data provides a quantitative prediction of disease progression, which is very important for determining the priority and urgency of treatment options. Based on the multimodal pre-trained large model, the lesion spread behavior prediction data is simulated. This means using advanced deep learning techniques to simulate and predict the spread trend of lesions in the future. This work helps to understand the dynamic changes of lesions and provide predictive capabilities for future treatment and intervention strategies. The resulting lesion spread behavior trend data provides detailed information on the changes of lesions in time and space. These data can be used to predict the progression rate, spread direction and impact range of the disease, and to evaluate the multidimensional impact of lesion spread based on the lesion spread behavior trend data. This step aims to deeply analyze the multi-faceted impact of lesion spread on surrounding tissues, organs and overall physical health, including physiological and functional changes. The obtained lesion spread depth impact data provides more detailed and quantitative impact information. These data not only include the specific impact area and degree of the lesion in the body, but also can evaluate its potential impact on quality of life and health prognosis. According to the lesion spread depth impact data, clinical decision reference coupling is performed. This step combines scientifically analyzed data with clinical practice, and the final lesion spread decision reference interactive data is obtained. Therefore, the present invention is an optimization processing of a traditional clinical decision-making method for critical internal medicine diseases based on a multimodal pre-trained large model, which solves the problem of inaccurate analysis of development trends of critical internal medicine diseases and resulting in low accuracy of clinical decisions in the traditional clinical decision-making method for critical internal medicine diseases based on a multimodal pre-trained large model, improves the accuracy of development trend analysis of critical internal medicine diseases, and improves the accuracy of clinical decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1A schematic diagram of the steps of a clinical decision-making method for critical internal medicine based on a multimodal pre-trained large model; Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG.

[0025] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0026] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0027] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0028] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0029] To achieve this, please refer to Figure 1 to Figure 2 , a method for clinical decision-making in medical critical care based on a multimodal pre-trained large model, the method comprising the following steps: Step S1: Obtaining medical case report data and medical case lesion images; performing lesion description mapping on the medical case lesion images according to the medical case report data to obtain medical case lesion description data; Step S2: performing lesion morphological structure analysis of different lesion locations according to the lesion description data of internal medicine cases to obtain lesion morphological structure data; performing lesion diffusion behavior prediction on the lesion morphological structure data to obtain lesion diffusion behavior prediction data; Step S3: simulating the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multimodal pre-trained large model to obtain the lesion diffusion behavior trend data; evaluating the multidimensional impact degree of the lesion according to the lesion diffusion behavior trend data to obtain the multidimensional impact degree data of the lesion diffusion; Step S4: perform deep feature extraction on the multi-dimensional impact degree data of lesion spread to obtain lesion spread depth impact data; perform clinical decision reference coupling based on the lesion spread depth impact data to obtain lesion spread clinical decision reference data; perform interactive learning on the lesion spread clinical decision reference data to obtain lesion spread decision reference interactive data.

[0030] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of a method for clinical decision-making in medical critical illness based on a multimodal pre-trained large model of the present invention. In this example, the method for clinical decision-making in medical critical illness based on a multimodal pre-trained large model includes the following steps: Step S1: Obtaining medical case report data and medical case lesion images; performing lesion description mapping on the medical case lesion images according to the medical case report data to obtain medical case lesion description data; In an embodiment of the present invention, first, it is necessary to extract the medical case report data from the hospital electronic health record system (EHR). This process usually includes accessing the database interface, using SQL query statements or API requests, and extracting relevant case information from the structured data table. When extracting data, attention should be paid to the patient's basic information, medical history, examination results, laboratory test data, and the doctor's clinical records. These data are then desensitized to hide the patient's key personal privacy information. The data format includes CSV, JSON or XML, etc. The data needs to be cleaned and formatted to ensure its accuracy and completeness. For lesion image data, DICOM format image files need to be extracted from the hospital image storage system (PACS). Use medical image processing software (such as OsiriX, 3D Slicer or tools provided by the PACS system) to download the images and perform preprocessing, including noise removal, contrast enhancement, etc., to improve image quality. After obtaining the medical case report data and lesion images, it is necessary to map the text information in the case report to the corresponding position in the image. This process first involves text parsing, extracting and structuring the descriptive information (such as lesion location, size, type, etc.) in the case report. This can be done using natural language processing (NLP) tools, such as SpaCy, NLTK, or NLP models in specific medical fields. Next, the image data needs to be annotated to map the description information in the case report to the image. This process can be done using medical image annotation software (such as ITK-SNAP or Labelbox) to mark the lesion area on the image through manual annotation or semi-automatic annotation techniques. Specifically, the lesion area can be delineated by setting the region of interest (ROI) tool, and the description information in the case report can be attached to these annotations. Through the aforementioned lesion description mapping steps, the final generated internal medicine case lesion description data includes the combination of the text information in the case report and the lesion annotation information in the image. These data should be stored in a structured format, such as creating a database table containing case ID, lesion description, image annotation coordinates, lesion size, and other features. The data format can be JSON or CSV, where each record contains the description text in the case report and its corresponding image annotation location. When storing, the integrity and consistency of the data must be ensured for subsequent analysis and model training. The data validation step is also very important, and cross-validation with clinical experts is performed to ensure the consistency and accuracy of the description and image annotation.

[0031] Step S2: performing lesion morphological structure analysis of different lesion locations according to the lesion description data of internal medicine cases to obtain lesion morphological structure data; performing lesion diffusion behavior prediction on the lesion morphological structure data to obtain lesion diffusion behavior prediction data; In an embodiment of the present invention, when performing lesion morphological structure analysis, it is first necessary to parse the lesion description data of the medical case, and extract the lesion location and morphological features contained therein. Each lesion area is segmented using a medical image analysis algorithm (e.g., an algorithm based on image segmentation technology). In the image segmentation step, a threshold-based segmentation method is used to distinguish the lesion area from the background area. For each segmented lesion area, its morphological features are further calculated, including the shape, size, boundary contour, and volume of the lesion. This process can be achieved by calculating geometric features, such as using a connected domain analysis algorithm to calculate the boundary and area of ​​the lesion. Finally, the lesion morphological data obtained by sorting out includes information such as the location, shape characteristics, and volume of each lesion to form a structured lesion morphological structure data set. When predicting the spread behavior of the lesion, the lesion morphological structure data is first input into a preset diffusion behavior prediction model. The model is based on the diffusion theory in physics and is calculated in combination with the geometric characteristics of the lesion and the surrounding tissue structure. A mathematical model of lesion diffusion is established using a simulation algorithm, such as a finite element analysis method, and the diffusion coefficient and environmental factors of the lesion are set for simulation. In the specific operation, the initial diffusion conditions are set for each lesion area, and its diffusion range and morphological changes at different time points are calculated. The lesion diffusion behavior data generated by the simulation results include the predicted diffusion image and diffusion measurement data, such as diffusion radius, diffusion speed, etc. These data are used to evaluate the future diffusion trend of the lesion and provide support for clinical decision-making.

[0032] Step S3: simulating the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multimodal pre-trained large model to obtain the lesion diffusion behavior trend data; evaluating the multidimensional impact degree of the lesion according to the lesion diffusion behavior trend data to obtain the multidimensional impact degree data of the lesion diffusion; In an embodiment of the present invention, when simulating the diffusion behavior trend of the lesion diffusion behavior prediction data, the prediction data is first input into a multimodal pre-trained large model. The model integrates information from different data sources (such as imaging data and clinical data) and analyzes it through deep learning technology. First, the lesion diffusion prediction data is standardized so that it is consistent with the data format during model pre-training. Then, the standardized data is input into the model, and the model processes the data through its learned diffusion mode and lesion behavior mode. The convolutional neural network (CNN) part of the model extracts the characteristics of the diffusion behavior, and the subsequent recurrent neural network (RNN) part captures the dynamic changes in the time series. Finally, the lesion diffusion behavior trend data is generated through the output layer of the model, including the diffusion trend map and the predicted time series data. These data show the future trend of lesion diffusion and provide a basis for subsequent analysis. When evaluating the multidimensional impact of lesions, relevant indicators are first extracted from the lesion diffusion behavior trend data obtained in step 1. These indicators include diffusion rate, diffusion range, and impact area. Then, multidimensional analysis methods, such as principal component analysis (PCA) or factor analysis, are applied to reduce the dimensionality of multiple influencing factors to identify the main influencing dimensions. For each dimension, the degree of influence of lesion spread is calculated, including the degree of compression on surrounding tissues, the impact of lesions on system function, and the impact of lesion spread on overall health status. In specific operations, mathematical models (such as linear regression models) are established to quantify the weight of each influencing factor, and combined with diffusion trend data, multidimensional influence degree data of lesion spread are generated. Ultimately, these data show the degree of influence of lesion spread in different dimensions, providing a detailed evaluation basis for clinical decision-making.

[0033] Step S4: perform deep feature extraction on the multi-dimensional impact degree data of lesion spread to obtain lesion spread depth impact data; perform clinical decision reference coupling based on the lesion spread depth impact data to obtain lesion spread clinical decision reference data; perform interactive learning on the lesion spread clinical decision reference data to obtain lesion spread decision reference interactive data.

[0034] In an embodiment of the present invention, when performing deep feature extraction, the multi-dimensional impact degree data of lesion spread is first preprocessed to ensure that the data format is consistent and suitable for the input of the deep learning model. Feature extraction is performed by a convolutional neural network (CNN), in which the input data is first subjected to hierarchical convolution processing to extract high-dimensional features. The convolution layer identifies local features in the data, and then reduces the dimension through the pooling layer to retain the main features. The features extracted by the convolution layer are further integrated using a fully connected layer to form a representation with deep features. The representation is the depth impact data of lesion spread, which includes the depth and breadth of the impact of the lesion on different tissues and systems. These data provide detailed impacts of lesion spread in multiple dimensions, providing a basis for clinical analysis. In the process of clinical decision reference coupling, the depth impact data of lesion spread needs to be matched with the clinical decision criteria first. The depth impact data of lesion spread is input into the clinical decision reference model, which integrates different clinical decision rules and parameters. The model compares the depth impact data with the standardized clinical decision framework to identify the impact features that meet the decision criteria. The specific operation includes calculating the fitness of the lesion impact data under different decision parameters, and screening out the features that meet the decision criteria by setting the threshold. The generated clinical decision reference data for lesion spread includes specific decision suggestions and reference indicators, such as whether further examination is needed, treatment plan suggestions, etc. These data provide a specific reference basis for clinical decision-making. When interactive learning is carried out, the clinical decision reference data for lesion spread is first input into the interactive learning model. The model analyzes the applicability and effect of decision reference data in different clinical scenarios by simulating the feedback loop in the actual clinical decision-making process. In the specific operation, the model generates prediction results based on the preliminary decision suggestions and compares these results with the actual clinical feedback. The model gradually optimizes the accuracy of the decision reference by adjusting the decision parameters and strategies. The process of interactive learning involves real-time updating of the parameters of the decision model to better adapt to the actual clinical situation. The final generated interactive data for lesion spread decision reference includes optimized decision suggestions and strategy adjustment data. These data provide optimized reference and guidance for further clinical decision-making.

[0035] Preferably, step S1 comprises the following steps: Step S11: Acquire medical case report data and medical case lesion images; Step S12: desensitizing the medical case report data and the medical case lesion images to obtain the medical case report desensitized data and the medical case lesion desensitized images; Step S13: extracting the internal medicine case description from the desensitized data of the internal medicine case report to obtain the internal medicine case description data; Step S14: performing lesion description mapping on the desensitized images of lesions in internal medicine cases according to the internal medicine case description data to obtain the lesion description data in internal medicine cases.

[0036] In an embodiment of the present invention, the process of obtaining medical case report data and medical case lesion images involves extracting data from a medical information system. First, case report data containing patient medical history, laboratory results, and examination records are downloaded from an electronic health record (EHR) system using a database interface through an SQL query or an API request. These data are usually in a structured format (such as CSV or JSON). Next, lesion images are downloaded from a hospital image storage system (PACS) in a DICOM format. Use a medical image processing tool (such as OsiriX or 3D Slicer) to access the PACS and extract relevant image data. The downloaded images need to be saved in a high-quality format to ensure the accuracy of subsequent processing. The process of desensitizing medical case report data and lesion images includes removing or blurring all personal information that identifies the patient. Desensitization in medical case report data involves using data cleaning techniques to replace sensitive information such as patient name, ID number, and contact information with anonymous identifiers. For lesion images, image processing tools (such as ImageJ or Matlab) are used to blur or mask the identifiable areas in the image. The desensitized data should be stored as a new file or database table to ensure the privacy protection of the original data while maintaining the usability and integrity of the data. Internal medicine case description extraction involves natural language processing (NLP) on the desensitized case report data. First, text processing tools are applied to segment and annotate the case reports to identify key terms and descriptive information in the cases. Next, the specific description of the lesion, such as lesion location, type, size, and other relevant features, is extracted from the text using a vocabulary or entity recognition technology in the medical field. This process includes applying a named entity recognition (NER) algorithm to extract and structure the key information in the case description. The final generated internal medicine case description data should include specific description information of the case to support subsequent analysis and mapping. The lesion description mapping process involves matching the internal medicine case description data with the lesion area in the desensitized image. First, the desensitized image is opened using a medical image annotation tool (such as ITK-SNAP or Labelbox) to map the lesion location and feature information in the case description data to the image. During the annotation process, select appropriate marking tools (such as ROI tools) to draw lesion areas in the image and attach description data to these marks. After the annotation is completed, the generated medical case lesion description data should contain the precise location, morphological characteristics and detailed information corresponding to the case description of each lesion. These data provide a visual and structured reference for subsequent analysis and decision-making.

[0037] Preferably, step S2 comprises the following steps: Step S21: performing position distribution analysis on the lesion images of internal medicine cases to obtain lesion position distribution data; Step S22: performing lesion morphological structure analysis of different lesion locations on the lesion location distribution data according to the lesion description data of the internal medicine case to obtain lesion morphological structure data; Step S23: performing isodensity characteristics recognition on the lesion morphological structure data to obtain isodensity data of the lesion structure; Step S24: predicting the lesion diffusion behavior of the lesion morphological structure data according to the lesion structure and other density data to obtain lesion diffusion behavior prediction data.

[0038] As an example of the present invention, refer to Figure 2 As shown, in this example, step S2 includes: Step S21: performing position distribution analysis on the lesion images of internal medicine cases to obtain lesion position distribution data; In an embodiment of the present invention, when performing position distribution analysis on lesion images of medical cases, the images need to be preprocessed first, including denoising and contrast enhancement. Medical image processing tools (such as 3D Slicer) are used to preprocess the image data to improve the accuracy of the analysis. Subsequently, image segmentation techniques (such as threshold-based segmentation or deep learning-based segmentation networks) are applied to identify lesion areas. The position coordinates and size of each lesion area are obtained through the segmentation results. Spatial statistical analysis is performed on the identified lesions to determine the distribution characteristics of the lesions in the images, including the density, distribution range, and degree of aggregation of the lesion locations. This analysis generates lesion location distribution data, describes the spatial distribution pattern of the lesions in the images, and provides a basis for further analysis.

[0039] Step S22: performing lesion morphological structure analysis of different lesion locations on the lesion location distribution data according to the lesion description data of the internal medicine case to obtain lesion morphological structure data; In an embodiment of the present invention, when performing lesion morphological structure analysis on lesion location distribution data, it is first necessary to integrate the lesion description data with the location distribution data. Through medical image annotation tools, the specific location and morphological features in the lesion description are matched with the lesion area in the image. Computer vision algorithms (such as edge detection and morphological transformation) are used to extract the shape features of the lesion. The geometric features of the lesion area, including area, perimeter, shape irregularity, etc., are analyzed, and the geometric properties of the lesion are calculated using morphological measurement methods. This process generates lesion morphological structure data, provides detailed morphological features of each lesion, including shape, size and structural features, and provides necessary information for predicting the spread behavior of the lesion.

[0040] Step S23: performing isodensity characteristics recognition on the lesion morphological structure data to obtain isodensity data of the lesion structure; In an embodiment of the present invention, isodensity feature recognition involves analyzing the lesion morphological structure data to identify density changes in the lesion area. A density estimation algorithm (such as kernel density estimation) is used to calculate the density distribution in the lesion area. First, the window size for density calculation is defined, and its local density is calculated in each lesion area. Then, by setting a density threshold, regions with similar density characteristics are identified and grouped. These isodensity regions are marked and their characteristics, such as density uniformity and regional area, are recorded. Finally, lesion structure isodensity data are generated to display the density distribution characteristics in the lesion and its changes in the lesion area.

[0041] Step S24: predicting the lesion diffusion behavior of the lesion morphological structure data according to the lesion structure and other density data to obtain lesion diffusion behavior prediction data.

[0042] In an embodiment of the present invention, the prediction of lesion diffusion behavior is based on density data such as lesion structure. First, a lesion diffusion model is established, and a diffusion simulation algorithm (such as finite element method or diffusion equation) is used to predict the future diffusion behavior of the lesion by combining the density data and the lesion morphological structure data. The model input includes the initial morphology, density distribution and surrounding environment characteristics of the lesion. By simulating the diffusion process of the lesion under different conditions, a prediction of the diffusion path and diffusion range is generated. Finally, the model output is integrated to generate lesion diffusion behavior prediction data, including diffusion prediction maps and diffusion metrics (such as diffusion rate, diffusion range). These data provide a detailed prediction of the future diffusion trend of the lesion, providing a basis for clinical decision-making.

[0043] Preferably, step S24 comprises the following steps: Step S241: performing density evolution law analysis on the density data of the lesion structure to obtain density evolution law data of the lesion structure; Step S242: performing lesion spatial variability analysis on the lesion morphological structure data according to the lesion structure density evolution law data to obtain lesion spatial variability data; Step S243: constructing a lesion diffusion prediction model based on the lesion structure density evolution law data and the lesion spatial variability data based on the policy gradient algorithm to obtain a lesion diffusion prediction model; Step S244: predicting the lesion diffusion behavior according to the lesion diffusion prediction model to obtain lesion diffusion behavior prediction data.

[0044] In an embodiment of the present invention, when density data such as lesion structure are analyzed for density evolution law, the density data are first grouped according to time series or different lesion stages. Statistical analysis tools (such as R language or Matlab) are used to calculate the density change trend of each stage. Regression analysis or time series analysis methods (such as linear regression or smooth trend analysis) are used to identify the law of density change. The spatial distribution and temporal dynamics of the data need to be considered during the analysis process to ensure that the analysis results can reflect the true evolution law of lesion density. The lesion structure density evolution law data finally generated contains a statistical description and evolution trend of density change, providing basic data for subsequent variability analysis and diffusion prediction. When performing lesion spatial variability analysis, spatial statistical analysis is first performed on the lesion morphological structure data, and the spatial distribution characteristics of the lesion area are analyzed using tools (such as ArcGIS or QGIS). Spatial variability analysis methods (such as spatial autocorrelation analysis or variogram analysis) are used to calculate the spatial variability indicators of the lesion position, including spatial autocorrelation coefficients and variograms. These indicators reflect the distribution characteristics of lesions in space and their laws of change. Combined with the data of the evolution law of lesion structure density, the spatial variability is further analyzed to determine the spatial heterogeneity and evolution trend of the lesion. Finally, the spatial variability data of the lesion is generated, including the variability analysis results and the description of the spatial distribution characteristics, which provides input data for the construction of the diffusion prediction model. When constructing the lesion diffusion prediction model, the lesion structure density evolution law data and the lesion spatial variability data are first input into the policy gradient algorithm. The policy gradient algorithm trains the model by optimizing the objective function, which usually includes the prediction error and the model complexity. The model is trained using the data set, and the input features of the model include lesion density, spatial variability and other related factors (such as tissue type, environmental factors). Through multiple iterations and policy updates, the parameters of the model are optimized so that the model can accurately predict the diffusion behavior of the lesion. The lesion diffusion prediction model finally obtained can be used to simulate the diffusion of lesions under different conditions and provide a prediction basis for actual clinical applications. When using the constructed lesion diffusion prediction model to predict the lesion diffusion behavior, the lesion data to be predicted (including structural density and spatial variability characteristics) are first input into the model. The model generates predictions of diffusion paths and diffusion ranges through calculations. The output of the application model is visualized (such as generating diffusion prediction maps) and diffusion metrics (such as diffusion rate and diffusion range) are calculated. These prediction data should be output in the form of charts and data reports to show the future diffusion trend and range of the lesion. The final prediction data of lesion diffusion behavior provides detailed diffusion prediction information for clinical decision-making, which helps to formulate effective intervention measures.

[0045] Preferably, constructing a lesion diffusion prediction model based on the lesion structure density evolution law data and the lesion spatial variability data comprises the following steps: Based on the data of evolution law of lesion structure density and spatial variability of lesions, spatial correlation analysis was performed to obtain spatial correlation data of lesion evolution; Calculate the local evolution correlation rate of the lesion evolution spatial correlation data to obtain the local evolution correlation rate data; Calculate the rate variance of the local evolution correlation rate data to obtain the local evolution rate variance data; Based on the policy gradient algorithm, the local evolution rate variance data is continuously evolved to learn the state, and the local continuous evolution state data is obtained; A lesion diffusion prediction model is constructed according to the local continuous evolution state data to obtain a lesion diffusion prediction model.

[0046] In an embodiment of the present invention, when performing a spatial correlation analysis of lesion evolution, first, the lesion structure density evolution law data and the lesion spatial variability data are input into a spatial analysis tool (such as an ArcGIS Spatial Analyst module). Spatial statistical methods (such as spatial autocorrelation analysis and Moran index calculation) are used to analyze the relationship between the density evolution and spatial variability of the lesion area. By calculating the spatial correlation coefficient and the spatial autocorrelation diagram, the evolution pattern and its spatial correlation in the lesion area are identified. The analysis provides correlation information between different regions during the lesion evolution process, thereby generating lesion evolution spatial correlation data to describe the evolution correlation of the lesion in space. When the local evolution correlation rate is calculated for the lesion evolution spatial correlation data, a spatial variogram model (such as a weighted regression model) is used to calculate the local evolution correlation rate in the lesion area. First, the lesion evolution spatial correlation data is divided into local areas according to a predetermined spatial window. Regression analysis is performed on the data in each local area to estimate the evolution rate. A distance weighted method is used in the regression analysis to calculate the evolution rate in each local area. Finally, local evolution associated rate data are generated, which represent the rate and intensity of lesion evolution in each region. When calculating the variance of local evolution rate, statistical analysis tools (such as Matlab or Python's Numpy library) are used to calculate the variance of local evolution associated rate data. First, the mean of the evolution rate in each local region is calculated. Then, the variance of the evolution rate in each local region is calculated using the mean. The calculation formula of the variance is the average of the square of the difference between the evolution rate and the mean of each local region. Finally, local evolution rate variance data are generated to describe the degree of variation of the evolution rate in the lesion region, providing detailed rate variation information for subsequent model construction. When using the policy gradient algorithm for local continuous evolution state learning, the local evolution rate variance data is first used as the input feature, and the objective function is designed to minimize the prediction error. The policy gradient algorithm trains the model by optimizing the objective function and uses the gradient ascent method to update the model parameters to improve the prediction accuracy of the evolution state. During the training process, the rate variance data of each local region is used to update the policy and calculate the gradient of each local state. After multiple training iterations, local continuous evolution state data are obtained to describe the continuous changes in the evolution state in the lesion area and its influencing factors. In the construction of the lesion spread prediction model, the local continuous evolution state data is first used as the input feature to build a diffusion prediction model. Select a model building tool (such as TensorFlow or PyTorch) to design the model architecture, including the input layer, hidden layer, and output layer. Use the continuous evolution state data to train the model and adjust the network parameters to minimize the prediction error. During the training process, the model learns the law of lesion spread through gradual optimization. Finally, the model outputs the lesion spread prediction result and obtains the lesion spread prediction model, which can predict the future spread behavior of the lesion and provide a basis for clinical decision-making.

[0047] Preferably, step S3 comprises the following steps: Step S31: simulating the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multimodal pre-trained large model to obtain lesion diffusion behavior trend data; Step S32: performing spatial interpolation processing on the lesion diffusion behavior trend data to obtain lesion diffusion behavior spatial interpolation data; Step S33: performing hysteresis interpolation segmentation on the spatial interpolation data of the lesion diffusion behavior to obtain diffusion hysteresis interpolation segmentation data; Step S34: evaluating the multi-dimensional impact degree of lesions according to the diffusion hysteresis interpolation segmentation data and the lesion diffusion behavior trend data to obtain the multi-dimensional impact degree data of lesion diffusion.

[0048] In an embodiment of the present invention, first, the lesion diffusion behavior prediction data is input into a multimodal pre-trained large model. The model combines visual data (such as lesion images) and text data (such as case reports), and uses deep learning technology to extract complex features in the data. When training the large model, an autoregressive model is used to model the lesion diffusion trend, and historical data and spatial information are considered to simulate the diffusion behavior trend through the feedforward neural network of the model. The simulation process generates lesion diffusion behavior trend data by post-processing the output of the model to characterize the trend and future evolution path of lesion diffusion. The lesion diffusion behavior trend data is spatially interpolated using a spatial interpolation method (such as Kriging interpolation). First, the spatial distribution of the interpolation grid is defined, and then the interpolation algorithm is applied on the grid to convert the discrete lesion diffusion behavior trend data into continuous spatial data. The interpolation process involves calculating the spatial correlation between sample points and estimating unknown points based on these correlations. The interpolated data fills the blank areas in the data and generates lesion diffusion behavior spatial interpolation data, which is used to show the changes in the lesion diffusion trend at different spatial locations. When processing the spatial interpolation data of lesion diffusion behavior, the hysteresis interpolation method is used to segment the data. Hysteresis interpolation adjusts the data by identifying the hysteresis phenomenon in the data, that is, the delayed effect of diffusion behavior in time. The specific steps include: first, determine the hysteresis period (such as the time delay of diffusion), then segment the spatial interpolation data and divide the data into several segments according to the hysteresis period. The interpolation result of each segment reflects the hysteresis of lesion diffusion in different time periods. After the processing is completed, the diffusion hysteresis interpolation segmented data is generated to describe the hysteresis effect of lesion diffusion in different time periods. When evaluating the multidimensional impact degree of lesions, the diffusion hysteresis interpolation segmented data is first integrated with the lesion diffusion behavior trend data. Multidimensional evaluation tools (such as multivariate statistical analysis) are used to comprehensively analyze these data to evaluate the multidimensional influencing factors of lesion diffusion. The evaluation process includes determining the evaluation indicators (such as diffusion rate, spatial coverage, hysteresis effect, etc.), and applying the weighted average method to summarize the various influencing factors into a comprehensive impact. During the analysis process, statistical models (such as principal component analysis) are used to extract the main influencing factors. Finally, multidimensional impact data of lesion spread are generated, which integrates the influence of diffusion trend and hysteresis effect, providing detailed multidimensional analysis results for clinical decision-making.

[0049] Preferably, step S32 includes the following steps: Step S321: performing trend scale effect analysis on the lesion diffusion behavior trend data to obtain diffusion trend scale effect data; Step S322: Calculating the diffusion rate increment discontinuity interval for the lesion diffusion behavior trend data according to the diffusion trend scale effect data to obtain the diffusion rate increment discontinuity interval; Step S323: performing spatial trend nonlinear difference quantification according to the diffusion rate increment discontinuity interval to obtain diffusion spatial trend difference quantification data; Step S324: performing spatial interpolation processing according to the diffusion rate increment discontinuity interval and the diffusion spatial trend difference quantified data to obtain the lesion diffusion behavior spatial interpolation data.

[0050] In an embodiment of the present invention, first, a trend scale effect analysis method is selected to process the lesion diffusion behavior trend data. A scale effect model (such as wavelet transform) is used to decompose the trend data into components of different scales, and the influence of each scale on the lesion diffusion behavior is analyzed. The specific operation includes: defining the scale range, and applying wavelet transform to decompose the data into low-frequency and high-frequency components. Then, the energy and volatility of each component are calculated to identify the degree of influence of the scale effect on the diffusion behavior. By analyzing the results, diffusion trend scale effect data are generated, which provide detailed characteristics of the diffusion trend at different scales. Based on the diffusion trend scale effect data, the diffusion rate increment discontinuous interval calculation is performed. First, the lesion diffusion behavior trend data is incremented to determine the change in the diffusion rate. The difference method is applied to calculate the increment between data points, and the change interval of the increment is identified. Then, the threshold analysis method (such as mean plus or minus standard deviation) is used to determine the discontinuity points of the increment. The data is divided into multiple intervals according to these discontinuity points, and the diffusion rate change characteristics of each interval are marked. Finally, the diffusion rate increment discontinuous interval data is obtained, which describes the change of the diffusion rate in different intervals. Based on the diffusion rate increment discontinuity interval, the nonlinear difference of spatial trend is quantified. The nonlinear regression model (such as polynomial regression) is applied to fit the data of each increment discontinuity interval to analyze the nonlinear difference. First, the parameters of the nonlinear model are defined, and the data are regressed to calculate the fitting error and model parameters of each interval. Then, the spatial trend difference in each interval is quantified, including the goodness of fit and residual of the model. Finally, the quantitative data of diffusion spatial trend difference is generated, which provides the nonlinear change information of diffusion behavior in space. The spatial interpolation processing is performed using the diffusion rate increment discontinuity interval and the diffusion spatial trend difference quantitative data. First, the incremental discontinuity interval data and the spatial trend difference quantitative data are integrated into the input of the interpolation model. The data is interpolated by selecting a suitable spatial interpolation method (such as the inverse distance weighted method). The interpolation process includes: defining the interpolation grid and applying the interpolation algorithm on the grid to convert the discrete diffusion data points into continuous spatial data. Then, the spatial interpolation data of lesion diffusion behavior are generated according to the interpolation results. These data show the continuous distribution of lesion diffusion behavior in space and reflect the diffusion trend of different spatial positions.

[0051] Preferably, step S34 includes the following steps: Step S341: dividing the diffusion hysteresis interpolation segmented data into hysteresis-affected regions to obtain diffusion hysteresis-affected region data; Step S342: performing regional impact superposition analysis based on the diffusion hysteresis impact regional data to obtain hysteresis regional impact superposition data; Step S343: Identify the multi-dimensional influencing factors of the lesions according to the lagging regional impact superposition data and the lesion diffusion behavior trend data to obtain the multi-dimensional influencing factors of the lesions; Step S344: Evaluate the multidimensional impact degree of the lesion based on the multidimensional impact factor of the lesion on the superposition data of the lagging regional impact, and obtain the multidimensional impact degree data of the lesion diffusion.

[0052] In an embodiment of the present invention, the diffusion hysteresis interpolation segmented data is divided into hysteresis-affected regions. First, according to the spatial distribution of the diffusion hysteresis interpolation segmented data, the data is grouped using a threshold-based regional division method (such as a K-means clustering algorithm). Select a suitable hysteresis threshold to define the boundary of the affected region. By analyzing the hysteresis data values ​​in different segments, the hysteresis-affected regions are identified, and the boundaries and features of different regions are marked. Finally, the diffusion hysteresis-affected region data is generated to provide the spatial position and hysteresis-affected degree of each affected region. The diffusion hysteresis-affected region data is subjected to regional impact superposition analysis. First, the data of each hysteresis-affected region are superimposed into a unified analysis model. The impact values ​​of different regions are synthesized using a superposition method (such as weighted average) to calculate the comprehensive impact degree of each position. The impact superposition results are analyzed to extract the overall impact characteristics of each region. Finally, the hysteresis-affected region superposition data is generated to describe the comprehensive impact degree and spatial distribution of each region in the diffusion behavior. The multi-dimensional impact factors of the lesion are identified based on the hysteresis-affected region superposition data and the lesion diffusion behavior trend data. First, the lagging regional impact superposition data and the lesion diffusion behavior trend data are integrated to construct a multidimensional analysis model. The main influencing factors are identified by applying factor analysis methods (such as principal component analysis). By analyzing the contribution of each factor to the diffusion behavior, the key multidimensional influencing factors are extracted. Finally, the multidimensional influencing factors of the lesion are generated to identify the factors and their weights that have a major impact on the behavior during the lesion diffusion process. The multidimensional influencing factors of the lesion are used to evaluate the multidimensional impact of the lesion on the lagging regional impact superposition data. First, the identified multidimensional influencing factors are combined with the lagging regional impact superposition data, and the multidimensional impact of each region is calculated using a weighted evaluation model (such as a weighted sum model). For each region, the weights of each influencing factor and its contribution to the diffusion behavior are comprehensively considered to obtain a comprehensive impact score. Finally, the multidimensional impact data of the lesion diffusion are generated to provide a specific impact evaluation of the lesion diffusion behavior in multiple dimensions, describing the impact of each region and its impact on the overall diffusion behavior.

[0053] Preferably, step S4 comprises the following steps: Step S41: extracting depth features from the multi-dimensional impact degree data of lesion diffusion to obtain lesion diffusion depth impact data; Step S42: performing commonality identification on the lesion diffusion depth influence data to obtain lesion diffusion influence commonality data; Step S43: Acquire historical clinical decision data; Step S44: performing clinical decision reference coupling on historical clinical decision data according to the common data on the impact of lesion spread, to obtain clinical decision reference data on lesion spread; Step S45: interactively learning the lesion spread clinical decision reference data according to the Q-Learning algorithm to obtain lesion spread decision reference interactive data.

[0054] In an embodiment of the present invention, deep feature extraction is performed on the multidimensional impact degree data of lesion diffusion. First, a feature extraction algorithm (such as a convolutional neural network) is applied to process the multidimensional impact degree data to extract deep features in the data. A feature extraction model is constructed, and the multidimensional data is input into the model. After multiple layers of convolution and pooling operations, representative deep features are extracted. The extracted features are standardized and scaled to reduce the differences between the data. Finally, lesion diffusion depth impact data is generated, which describes the deep features of lesion diffusion behavior and its performance under different influencing factors. Commonality identification is performed on lesion diffusion depth impact data. First, a commonality analysis method (such as cluster analysis) is applied to process the depth impact data. Select a suitable clustering algorithm (such as K-means clustering or hierarchical clustering), group the depth impact data, and identify the common features of each group of data. The identified common features are extracted and summarized to remove noise data. Finally, lesion diffusion impact commonality data is generated to describe the commonality patterns of lesion diffusion behavior and its common influencing factors under different circumstances. Obtain historical clinical decision data. First, extract the historical clinical decision records related to internal medicine from the hospital database. These records include past case analysis, doctors' decision basis and their results. Integrate the data to ensure that different cases, treatment plans and their effects are covered. Perform preliminary cleaning on the extracted data to remove duplicate or invalid records. Finally, obtain complete historical clinical decision data to provide a basis for subsequent analysis and model building. Couple the historical clinical decision data with clinical decision references based on the common data of lesion spread. First, build a reference model to combine the common data of lesion spread with the historical clinical decision data. Use data matching algorithms (such as association rule mining) to match the features in the common data with the decision records in the historical data. Analyze the matching results to extract the key reference factors in the decision-making process. Finally, generate the clinical decision reference data of lesion spread to describe how the influencing factors in lesion spread behavior are associated with historical decisions. Interactively learn the clinical decision reference data of lesion spread according to the Q-Learning algorithm. First, build a Q-Learning algorithm model and input the clinical decision reference data of lesion spread into the model. Through interactive learning, the algorithm continuously updates the Q-value function to optimize the choice of lesion spread decision. The learning results after each interaction are stored and analyzed to evaluate the learning effect of the model. Finally, the lesion spread decision reference interaction data is generated to describe how the model adjusts the decision-making strategy during the learning process, and combines historical data to optimize the future decision-making process.

[0055] Preferably, the present invention further provides an internal medicine critical care clinical decision system based on a multimodal pre-trained large model, which is used to execute the internal medicine critical care clinical decision method based on a multimodal pre-trained large model as described above. The internal medicine critical care clinical decision system based on a multimodal pre-trained large model comprises: The lesion description analysis module is used to obtain the medical case report data and the medical case lesion images; perform lesion description mapping on the medical case lesion images according to the medical case report data to obtain the medical case lesion description data; The lesion diffusion behavior prediction module is used to perform lesion morphological structure analysis of different lesion locations based on the lesion description data of internal medicine cases to obtain lesion morphological structure data; perform lesion diffusion behavior prediction on the lesion morphological structure data to obtain lesion diffusion behavior prediction data; The multi-dimensional impact assessment module of lesions is used to simulate the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multi-modal pre-trained large model to obtain the lesion diffusion behavior trend data; the multi-dimensional impact assessment of lesions is performed based on the lesion diffusion behavior trend data to obtain the multi-dimensional impact data of lesion diffusion; The decision reference learning module is used to perform deep feature extraction on the multi-dimensional impact degree data of lesion diffusion to obtain the lesion diffusion depth impact data; perform clinical decision reference coupling based on the lesion diffusion depth impact data to obtain the lesion diffusion clinical decision reference data; and perform interactive learning on the lesion diffusion clinical decision reference data to obtain the lesion diffusion decision reference interactive data.

[0056] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0057] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for clinical decision-making in critical medical care based on a multimodal pre-trained large model, characterized in that: The following steps are involved: Step S1: Obtaining medical case report data and medical case lesion images; performing lesion description mapping on the medical case lesion images according to the medical case report data to obtain medical case lesion description data; Step S2: performing lesion morphological structure analysis of different lesion locations according to the lesion description data of internal medicine cases to obtain lesion morphological structure data; performing lesion diffusion behavior prediction on the lesion morphological structure data to obtain lesion diffusion behavior prediction data; Step S3: simulating the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multimodal pre-trained large model to obtain the lesion diffusion behavior trend data; The multi-dimensional impact degree of the lesion is evaluated according to the lesion diffusion behavior trend data to obtain the multi-dimensional impact degree data of the lesion diffusion; Step S4: perform deep feature extraction on the multi-dimensional impact degree data of lesion spread to obtain lesion spread depth impact data; perform clinical decision reference coupling based on the lesion spread depth impact data to obtain lesion spread clinical decision reference data; perform interactive learning on the lesion spread clinical decision reference data to obtain lesion spread decision reference interactive data.

2. According to claim 1, the method for clinical decision-making of critical medical diseases based on multimodal pre-trained large models is characterized in that: Step S1 includes the following steps: Step S11: Acquire medical case report data and medical case lesion images; Step S12: desensitizing the medical case report data and the medical case lesion images to obtain the medical case report desensitized data and the medical case lesion desensitized images; Step S13: extracting the internal medicine case description from the desensitized data of the internal medicine case report to obtain the internal medicine case description data; Step S14: performing lesion description mapping on the desensitized images of lesions in internal medicine cases according to the internal medicine case description data to obtain the lesion description data in internal medicine cases.

3. The method for clinical decision-making of critical medical diseases based on a multimodal pre-trained large model according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: performing position distribution analysis on the lesion images of internal medicine cases to obtain lesion position distribution data; Step S22: performing lesion morphological structure analysis of different lesion locations on the lesion location distribution data according to the lesion description data of the internal medicine case to obtain lesion morphological structure data; Step S23: performing isodensity characteristics recognition on the lesion morphological structure data to obtain isodensity data of the lesion structure; Step S24: predicting the lesion diffusion behavior of the lesion morphological structure data according to the lesion structure and other density data to obtain lesion diffusion behavior prediction data.

4. The method for clinical decision-making in critical medical diseases based on a multimodal pre-trained large model according to claim 3, characterized in that: Step S24 includes the following steps: Step S241: performing density evolution law analysis on the density data of the lesion structure to obtain density evolution law data of the lesion structure; Step S242: performing lesion spatial variability analysis on the lesion morphological structure data according to the lesion structure density evolution law data to obtain lesion spatial variability data; Step S243: constructing a lesion diffusion prediction model based on the lesion structure density evolution law data and the lesion spatial variability data based on the policy gradient algorithm to obtain a lesion diffusion prediction model; Step S244: predicting the lesion diffusion behavior according to the lesion diffusion prediction model to obtain lesion diffusion behavior prediction data.

5. The method for clinical decision-making of critical medical diseases based on a multimodal pre-trained large model according to claim 4, characterized in that: The construction of a lesion diffusion prediction model based on the lesion structure density evolution law data and the lesion spatial variability data includes the following steps: Based on the data of evolution law of lesion structure density and the data of spatial variability of lesions, spatial correlation analysis was performed to obtain the data of spatial correlation of lesion evolution; Calculate the local evolution correlation rate of the lesion evolution spatial correlation data to obtain the local evolution correlation rate data; Calculate the rate variance of the local evolution correlation rate data to obtain the local evolution rate variance data; Based on the policy gradient algorithm, the local evolution rate variance data is continuously evolved to learn the state, and the local continuous evolution state data is obtained; A lesion diffusion prediction model is constructed according to the local continuous evolution state data to obtain a lesion diffusion prediction model.

6. The method for clinical decision-making of critical medical diseases based on a multimodal pre-trained large model according to claim 1, characterized in that: Step S3 includes the following steps: Step S31: simulating the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multimodal pre-trained large model to obtain lesion diffusion behavior trend data; Step S32: performing spatial interpolation processing on the lesion diffusion behavior trend data to obtain lesion diffusion behavior spatial interpolation data; Step S33: performing hysteresis interpolation segmentation on the spatial interpolation data of the lesion diffusion behavior to obtain diffusion hysteresis interpolation segmentation data; Step S34: evaluating the multi-dimensional impact degree of lesions according to the diffusion hysteresis interpolation segmentation data and the lesion diffusion behavior trend data to obtain the multi-dimensional impact degree data of lesion diffusion.

7. The method for clinical decision-making of critical medical diseases based on a multimodal pre-trained large model according to claim 6, characterized in that: Step S32 includes the following steps: Step S321: performing trend scale effect analysis on the lesion diffusion behavior trend data to obtain diffusion trend scale effect data; Step S322: Calculating the diffusion rate increment discontinuity interval for the lesion diffusion behavior trend data according to the diffusion trend scale effect data to obtain the diffusion rate increment discontinuity interval; Step S323: performing spatial trend nonlinear difference quantification according to the diffusion rate increment discontinuity interval to obtain diffusion spatial trend difference quantification data; Step S324: performing spatial interpolation processing according to the diffusion rate increment discontinuity interval and the diffusion spatial trend difference quantified data to obtain the lesion diffusion behavior spatial interpolation data.

8. The method for clinical decision-making of critical medical diseases based on a multimodal pre-trained large model according to claim 6, characterized in that: Step S34 includes the following steps: Step S341: dividing the diffusion hysteresis interpolation segmented data into hysteresis-affected regions to obtain diffusion hysteresis-affected region data; Step S342: performing regional impact superposition analysis based on the diffusion hysteresis impact regional data to obtain hysteresis regional impact superposition data; Step S343: Identify the multi-dimensional influencing factors of the lesions according to the lagging regional impact superposition data and the lesion diffusion behavior trend data to obtain the multi-dimensional influencing factors of the lesions; Step S344: Evaluate the multidimensional impact degree of the lesion based on the multidimensional impact factor of the lesion on the superposition data of the lagging regional impact, and obtain the multidimensional impact degree data of the lesion diffusion.

9. The method for critical medical clinical decision-making based on a multimodal pre-trained large model according to claim 1, characterized in that: Step S4 includes the following steps: Step S41: extracting depth features from the multi-dimensional impact degree data of lesion diffusion to obtain lesion diffusion depth impact data; Step S42: performing commonality identification on the lesion diffusion depth influence data to obtain lesion diffusion influence commonality data; Step S43: Acquire historical clinical decision data; Step S44: performing clinical decision reference coupling on historical clinical decision data according to the common data on the impact of lesion spread, to obtain clinical decision reference data on lesion spread; Step S45: interactively learning the lesion spread clinical decision reference data according to the Q-Learning algorithm to obtain lesion spread decision reference interactive data.

10. A clinical decision system for critical medical diseases based on a multimodal pre-trained large model, characterized in that: For executing the method for clinical decision-making of critical medical diseases based on a multimodal pre-trained large model as claimed in claim 1, the clinical decision-making system for critical medical diseases based on a multimodal pre-trained large model comprises: The lesion description analysis module is used to obtain the medical case report data and the medical case lesion images; perform lesion description mapping on the medical case lesion images according to the medical case report data to obtain the medical case lesion description data; The lesion diffusion behavior prediction module is used to perform lesion morphological structure analysis of different lesion locations based on the lesion description data of internal medicine cases to obtain lesion morphological structure data; perform lesion diffusion behavior prediction on the lesion morphological structure data to obtain lesion diffusion behavior prediction data; The multi-dimensional impact assessment module of lesions is used to simulate the diffusion behavior trend of the lesion diffusion behavior prediction data based on the multi-modal pre-trained large model to obtain the lesion diffusion behavior trend data; the multi-dimensional impact assessment of lesions is performed based on the lesion diffusion behavior trend data to obtain the multi-dimensional impact data of lesion diffusion; The decision reference learning module is used to perform deep feature extraction on the multi-dimensional impact degree data of lesion diffusion to obtain the lesion diffusion depth impact data; perform clinical decision reference coupling based on the lesion diffusion depth impact data to obtain the lesion diffusion clinical decision reference data; and perform interactive learning on the lesion diffusion clinical decision reference data to obtain the lesion diffusion decision reference interactive data.

Images