Medical aid decision-making method and system fusing multi-mode and deep network

Abstract

The invention discloses a medical aid decision-making method and system fusing multiple modes and a deep network, and relates to the technical field of intelligent medical treatment.The method comprises the steps that first data of a user is collected and analyzed, and a physiological stage is judged; and calling a sub-library corresponding to the dynamic feature library, and matching and screening out a first physiological feature with high correlation. Second data are collected, the first physiological features are fused to construct individualized feature vectors, and a feature incidence matrix is generated. Third data are collected in real time, the incidence matrix is combined, the medicine in-vivo process is simulated through a physiological pharmacokinetic model, and an individualized treatment scheme is generated. According to the invention, through the full-process design of dynamic stage determination, multi-mode association mining, real-time dose optimization and continuous evolution iteration, core pain points in early-maturing diagnosis and treatment of children and adolescents are accurately broken, and transformation from an empirical general scheme to a data-driven individualized decision is realized; safer, more accurate and more timely medical support is provided for premature individuals.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical technology, and in particular to a medical auxiliary decision-making method and system integrating multimodality and deep network. Background Art

[0002] In the current era of social development, the growth and development of children and adolescents are affected by factors such as changes in dietary structure (such as increased intake of high-fat and high-calorie foods, increased exposure to foods containing hormones), exposure to environmental endocrine disruptors (such as chemicals in plastic products and cosmetics), and the phenomenon of precocious puberty is becoming increasingly prominent. Many teenagers aged 12 or 13 have reached the level of physical maturity of 16-year-olds, with characteristics such as early appearance of secondary sexual characteristics (such as breast development in girls before the age of 8 and enlarged testicles in boys before the age of 9), advanced bone age (actual age is 12 but bone age is over 14 years old), and early stagnation after a sudden increase in height. This disorder in developmental rhythm poses a severe challenge to the traditional medical diagnosis and treatment and medication system, and the existing background technology has obvious limitations:

[0003] From a diagnostic and treatment assessment perspective, traditional methods rely on fixed age-developmental comparison tables, failing to consider individual developmental differences. For example, a 13-year-old boy with a bone age of 16 is still assessed using the same growth and development standards as those used for adolescents. This ignores the accelerated rate of skeletal closure, leading to misjudgment of growth potential and overlooking psychosocial issues caused by precocious puberty.

[0004] In the formulation of medication plans, existing technologies have the drawback of a one-size-fits-all approach. On the one hand, drug dosage calculations are mostly based on actual age or standard weight, without considering the adjustment of advanced physiological development status. For example, recommending a pediatric dosage for a 13-year-old precocious teenager whose liver and kidney functions have reached adult levels may lead to insufficient drug efficacy due to rapid metabolism; on the other hand, without integrating multi-dimensional data such as genes and environment, it is difficult to accurately predict the absorption and distribution process of drugs in precocious individuals, increasing the risk of adverse reactions;

[0005] Furthermore, the existing system lacks a dynamic tracking mechanism and is unable to adapt to the rapid developmental changes of precocious individuals. For example, a 12-year-old boy's bone age progressed from 14 to 15 years within six months. Traditional treatment plans still use the treatment parameters of the initial assessment, without timely adjustments to drug dosages or intervention methods, which may miss the optimal intervention window. Furthermore, feature analysis is limited to a single dimension and fails to explore the relationship between diet structure, hormone levels, and bone age acceleration, resulting in a lack of targeted diagnosis and treatment plans. Summary of the Invention

[0006] The purpose of the present invention is to provide a medical decision-making assistance method and system that integrates multimodality and deep networks to address the shortcomings of the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a medical decision-making assistance method integrating multimodality and deep network, comprising:

[0008] Collecting first data, the first data including physiological data information of the user, including at least age, height, weight, and secondary sexual characteristics; analyzing the first data using a deep network learning model to determine the physiological stage of the user;

[0009] Based on the physiological stage determined for the user, a corresponding feature sub-library in the dynamic feature library is called, the first data is matched with the feature sub-library, and a first physiological feature that is highly correlated with the physiological stage is screened out;

[0010] collecting second data, where the second data includes multi-dimensional biomedical data, including at least genomic data, phenotypic group data, environmental exposure group data, and imaging group data; fusing the first physiological characteristics and the second data to construct an individualized feature vector; and capturing nonlinear correlations between the features through a graph neural network to generate a feature correlation matrix.

[0011] collecting third data, the third data including physiological indicator information of the user, including at least liver and kidney function, heart rate, and blood flow, collecting the third data in real time, combining it with the characteristic association matrix, simulating the process of the drug in the body through a physiological pharmacokinetic model, and generating an individualized treatment plan based on the simulation results;

[0012] In a preferred embodiment, real-time status data of the user is collected, wherein the real-time status data includes at least changes in the condition and drug reactions, wherein the changes in the condition include but are not limited to worsening or relief of symptoms, the occurrence of complications, and abnormal fluctuations in physiological indicators; and drug reactions include but are not limited to the time it takes for drug efficacy to appear, the type and extent of adverse reactions, and changes in serum drug concentrations; and the weights of corresponding features in the feature library are dynamically adjusted in combination with the first data, the second data, and the third data. The weights of corresponding features in the feature library are dynamically adjusted in the following manner:

[0013] Constructing a comprehensive feature set including the first data, the second data, the third data, and the real-time status data, and mapping the features corresponding to each data to the feature space of the feature library;

[0014] An attention mechanism is used to calculate the correlation between each feature and real-time status data. The higher the correlation, the greater the increase in the weight of the corresponding feature. The weight of features directly related to disease changes (such as specific physiological indicators corresponding to symptoms) is increased to 1.2-2.0 times the initial value, and the weight of features strongly related to drug response (such as drug metabolizing enzyme activity indicators) is increased to 1.1-1.8 times the initial value.

[0015] The weight adjustment results are verified with reference to the physiological stage information in the first data, the genome and environmental exposure characteristics in the second data, and the real-time physiological indicators in the third data to ensure that the adjusted weights are consistent with the user's individual physiological characteristics and current health status;

[0016] Record the weight adjustment process and results to form a weight adjustment log for iterative optimization of the feature library and decision tracing.

[0017] In a preferred embodiment, the first data is physiological data information of the user, including at least age, height, weight, and secondary sexual characteristics; the determination of the user's physiological stage specifically comprises automatically analyzing the first data using a long short-term memory network (LSTM) model to determine the user's developmental stage (e.g., preschool, early adolescence), and introducing a developmental stage transition detection mechanism. When the user's physiological data crosses a preset stage threshold (e.g., a sudden height increase of more than 5 cm / year), a feature sub-library switch is automatically triggered;

[0018] The feature sub-library in the dynamic feature library is a subdivided library pre-constructed for different physiological stages. For example, the preschool feature sub-library focuses on basal metabolic rate and nutritional intake characteristics, and the adolescent feature sub-library contains growth hormone levels and sexual development indicator-specific characteristics. When matching the first data with the feature sub-library, the features ranked in the top 30% of the correlation are screened out through a feature similarity algorithm as the first physiological feature.

[0019] In a preferred embodiment, the second data is multi-dimensional biomedical data, including genomic (e.g., CYP2D6 genotype), phenotypic (e.g., BMI percentile), and environmental exposure (e.g., smoking history during pregnancy) data; a feature relationship graph is constructed using a graph neural network (GNN) to capture the nonlinear association between the first physiological feature and the second data, generating a feature association matrix, in which the matrix elements quantify the strength of the association between the features (with a value range of 0-1);

[0020] The third data is the user's real-time physiological indicator information, including liver and kidney function indicators, heart rate, blood flow, etc.; the physiological pharmacokinetic model combines the third data and metabolism-related features in the feature association matrix (such as CYP2D6 genotype-liver and kidney function association value) to simulate the drug's absorption, distribution, metabolism, and excretion processes in the body and predict the blood drug concentration-time curve.

[0021] In a preferred embodiment, the individualized treatment plan includes drug type, dosage, administration frequency and monitoring period; wherein the drug dosage is determined by the dosage adjustment coefficient calculation formula:

[0022] Dose adjustment factor = (predicted AUC value for current patients / AUC value for healthy children) × enzyme activity correction factor;

[0023] Among them, AUC is the area under the curve of the blood drug concentration-time curve, the enzyme activity correction factor is determined by the patient's CYP2D6 genotype (such as 1 / 1 type is 1.0, 17 / 17 type is 0.3), and the enzyme activity correction factor is set according to the CYP2D6 phenotypic classification, specifically including: ultra-rapid metabolizer (UM, correction factor 1.5), extensive metabolizer (EM, 1.0), intermediate metabolizer (IM, 0.5), and poor metabolizer (PM, 0.3), covering all common genotype combinations.

[0024] In a preferred embodiment, the user's real-time status data includes disease progression (such as bone age growth rate, secondary sexual characteristics development level), drug response (such as serum drug concentration, adverse reaction incidence rate) and lifestyle changes (such as nutritional intake, exercise frequency); the collection frequency is that basic data is collected once a day, and the real-time collection mechanism is activated when an abnormal state occurs.

[0025] In a preferred embodiment, the dynamic adjustment of the weights of corresponding features in the feature library is specifically achieved through an attention mechanism: the first data, the second data, the third data, and the real-time status data are input into the attention model, and the weight coefficient of each feature is output (the sum of the weights is 1); the weight of the feature with a correlation with the real-time status data exceeding 0.8 (such as the liver enzyme index corresponding to the adverse drug reaction) is increased to 1.5-2 times the initial value;

[0026] It also includes a feature weight decay strategy: for features that have not been activated in the feature association matrix for six consecutive months (association strength <0.1), the weight is automatically reduced to 50% of the initial value; features that have not been activated for 12 months are temporarily frozen and stored in the historical feature library for subsequent call. Feature activation is defined as the average association strength between the feature and other features in the feature association matrix being lower than 0.1 for six consecutive months (this threshold is determined based on the association strength distribution of invalid features in historical data).

[0027] In a preferred embodiment, a clinical feedback closed-loop mechanism is also included: the doctor's actual decision-making results (such as drug dosage adjustment records) are recorded as feedback data, which are used together with the first data, second data, third data and real-time status data to correct the feature library parameters. The feature association matrix and weight adjustment rules are updated once a quarter, and a dynamic update mechanism is introduced to adjust the update frequency according to the magnitude of data changes.

[0028] The present invention also provides a medical decision-making assistance system integrating multimodality and deep network, comprising:

[0029] A data collection module, used to collect first data, second data, third data and real-time status data of the user;

[0030] a stage determination module, configured to analyze the first data using a deep network learning model to determine the physiological stage of the user;

[0031] a feature screening module, configured to call a corresponding feature sub-library in the dynamic feature library based on the determined physiological stage, match the first data with the feature sub-library, and screen out the first physiological feature;

[0032] The feature fusion and association module is used to fuse the first physiological feature and the second data to construct an individualized feature vector, capture the nonlinear correlation between features through the graph neural network, and generate a feature correlation matrix;

[0033] The treatment plan generation module is used to collect third-party data in real time, combine it with the characteristic correlation matrix, simulate the drug's process in the body through the physiological pharmacokinetic model, and generate an individualized treatment plan based on the simulation results;

[0034] The weight adjustment module is used to dynamically adjust the weight of the corresponding feature in the feature library by combining the first data, the second data, the third data and the user's real-time status data.

[0035] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0036] 1. Accurately grasp the developmental stage of precocious individuals through dynamic judgment and targeted screening. By analyzing primary data with a long-short-term memory network and combining it with a developmental stage transition detection mechanism, we can break through traditional fixed age cognition, accurately identify the actual physiological stage of precocious individuals, and call upon the corresponding feature sub-library to focus on key features. At the same time, a feature similarity algorithm is used to screen the top 30% of relevant primary physiological features, prioritizing core features such as advanced bone age values, eliminating redundant information, and significantly improving the focus of diagnosis and treatment, avoiding misjudgment of growth potential or omission of psychological issues.

[0037] 2. By integrating the first physiological characteristics with the second data such as genome and environmental exposure group, and using graph neural network to capture implicit associations such as diet structure-hormone level-bone age acceleration, it provides comprehensive support for etiology analysis and makes up for the shortcomings of single-dimensional analysis. Based on the physiological pharmacokinetic model and the third data, combined with the dose adjustment coefficient formula containing CYP2D6 genotype correction, the drug dosage can be accurately customized for precocious individuals to avoid insufficient efficacy and reduce the risk of adverse reactions caused by genetic differences.

[0038] 3. Through real-time perception and dynamic optimization, we ensure that treatment plans closely track developmental changes in precocious individuals. We collect data on disease progression, drug response, and other issues daily, updating them in real time when anomalies occur. This addresses the lag inherent in traditional static assessment systems. We leverage attention mechanisms to increase the weight of key features, while combining them with a decay strategy to freeze irrelevant features. This allows treatment plans to adjust synchronously with development, avoiding missed intervention windows due to fixed parameters. For example, we might increase the weight of growth hormone inhibitory effects during rapid bone age closure. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0040] Figure 1 Flow chart of the method of the present invention.

[0041] Figure 2 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0042] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0043] Example 1, please refer to Figure 1 As shown, the medical decision-making assistance method integrating multimodality and deep network described in this embodiment includes:

[0044] S1. Collecting first data, where the first data includes physiological data information of a user, including at least age, height, weight, and secondary sexual characteristics; analyzing the first data using a deep network learning model to determine the physiological stage of the user;

[0045] S2. Based on the physiological stage of the user, calling a corresponding feature sub-library in a dynamic feature library, matching the first data with the feature sub-library, and screening out a first physiological feature that is highly relevant to the physiological stage;

[0046] S3. Collect second data, where the second data includes multi-dimensional biomedical data, including at least genomic data, phenotypic group data, environmental exposure group data, and imaging group data; fuse the first physiological characteristics and the second data to construct an individualized feature vector; and capture the nonlinear correlation between the features through a graph neural network to generate a feature correlation matrix;

[0047] S4. Collect third data, where the third data includes physiological indicator information of the user, including at least liver and kidney function, heart rate, and blood flow. The third data is collected in real time, combined with the characteristic association matrix, and the process of the drug in the body is simulated through a physiological pharmacokinetic model, and an individualized treatment plan is generated based on the simulation results.

[0048] As described in steps S1-S4 above, in the current state of social development, the growth and development of children and adolescents are affected by factors such as changes in dietary structure and exposure to environmental endocrine disruptors, and the phenomenon of precocious puberty is becoming increasingly prominent. Many teenagers aged 12 or 13 have reached the level of physical maturity of 16 years old, with characteristics such as the early appearance of secondary sexual characteristics, advanced bone age, and early stagnation after a sudden increase in height. This disorder of developmental rhythm poses a severe challenge to traditional medical diagnosis and treatment and medication systems, and the existing background technology has obvious limitations:

[0049] From a diagnostic and treatment assessment perspective, traditional methods rely on fixed age-developmental comparison tables, failing to consider individual developmental differences. For example, a 13-year-old boy with a bone age of 16 is still assessed using the same growth and development standards as those used for adolescents. This ignores the accelerated rate of skeletal closure, leading to misjudgment of growth potential and overlooking psychosocial issues caused by precocious puberty.

[0050] In the formulation of medication plans, existing technologies have the drawback of a one-size-fits-all approach. On the one hand, drug dosage calculations are mostly based on actual age or standard weight, without considering the adjustment of advanced physiological development status. For example, recommending a pediatric dosage for a 13-year-old precocious teenager whose liver and kidney functions have reached adult levels may lead to insufficient drug efficacy due to rapid metabolism; on the other hand, without integrating multi-dimensional data such as genes and environment, it is difficult to accurately predict the absorption and distribution process of drugs in precocious individuals, increasing the risk of adverse reactions;

[0051] In addition, the existing system lacks a dynamic tracking mechanism and cannot adapt to the rapid developmental changes of precocious individuals. For example, the bone age of a 12-year-old boy progressed from 14 to 15 years old within six months. The traditional plan still uses the treatment parameters of the initial assessment and fails to adjust the drug dosage or intervention methods in time, which may miss the optimal intervention window. At the same time, feature analysis is limited to a single dimension and fails to explore the relationship between diet structure-hormone level-bone age growth rate, resulting in a lack of targeted diagnosis and treatment plans;

[0052] Through the first data collection in step S1 and deep network model analysis, combined with the LSTM model's capture of time-series physiological data, the actual physiological stage of precocious individuals can be accurately determined, rather than relying on fixed age comparison tables. This solves the problem of traditional methods misjudging individuals with advanced bone ages but younger chronological ages, and avoids misjudging the growth potential of precocious individuals by using universal standards for adolescents.

[0053] By integrating the first physiological characteristic with multi-dimensional biomedical data in step S2, and using graph neural networks to mine nonlinear associations such as diet structure, hormone levels, and bone age acceleration, the limitations of single-dimensional analysis using traditional methods are overcome. The combination of physiological pharmacokinetic models and characteristic association matrices in step S3 enables drug dosage calculations to break through the one-size-fits-all model of actual age / standard weight: for precocious individuals aged 13 years old whose liver and kidney functions have reached adult levels, the dosage adjustment coefficient formula is combined with the CYP2D6 genotype in the third data and the second data to accurately calculate the dosage, avoiding insufficient efficacy caused by children's doses; with the help of the genotype-liver and kidney function association value in the characteristic association matrix, drug metabolism differences are predicted in advance to reduce the risk of adverse reactions;

[0054] The real-time status data collection and dynamic feature weight adjustment in step S4 can adapt to the rapid development of precocious individuals. For example, when the bone age of a 12-year-old boy progresses from 14 to 15 years old within six months, the system automatically increases the bone age change feature weight based on real-time status data, triggering the re-simulation and adjustment of drug dosage in step S3, avoiding the intervention lag caused by the fixed parameters of traditional programs.

[0055] The feature association matrix generated by step S2 quantifies the correlation strength of key features such as CYP2D6 genotype, drug metabolism rate, and bone age acceleration, providing a precise solution for the linkage of etiology and efficacy for precocious individuals; the attention mechanism of step S4 enables the system to focus on the features most relevant to the current state in real time, solving the problem of lack of targeting caused by the lack of emphasis in feature analysis of traditional solutions.

[0056] In summary, through the closed-loop process of precise stage determination - multimodal data fusion - dynamic plan adjustment, we have effectively addressed the diagnostic and treatment challenges brought about by precocious puberty in children and adolescents, achieved the transition from empirical general plans to data-driven individualized decision-making, and significantly improved the accuracy, safety and timeliness of individual diagnosis and treatment of precocious puberty.

[0057] In one embodiment, the first data collection and the determination of the physiological stage of the user in step S1 include:

[0058] S101: Collect first data, where the first data is physiological data information of a user, including at least age, height, weight, and secondary sexual characteristics;

[0059] S102. Analyze the first data using a deep network learning model (specifically, a long short-term memory network (LSTM)) to determine the user's developmental stage (e.g., preschool, early adolescence).

[0060] S103. Introducing a developmental stage transition detection mechanism. When a user's physiological data crosses a preset stage threshold (e.g., a sudden increase in height exceeding 5 cm / year), the feature sub-library switch is automatically triggered.

[0061] As described in steps S101-S103 above, S101 specifies that the first data must include core physiological information such as age, height, weight, and secondary sexual characteristics to ensure that the data covers key dimensions of individual growth and development. Compared to the single age or weight data in traditional methods, this multi-dimensional collection method can more comprehensively reflect the individual's physiological state and lay a data foundation for subsequent stage determinations. S102 uses a long short-term memory network to analyze the first data, leveraging its advantage in capturing time series characteristics to achieve dynamic determination of physiological stages. For example, for a precocious individual who is 12 years old but whose secondary sexual characteristics have reached mid-puberty, the LSTM model can accurately determine that they are in mid-puberty by combining their developmental data from the past two years, rather than early adolescence based solely on age as in traditional methods. This solves the problem of misjudgment caused by fixed age thresholds. The stage transition detection mechanism introduced in S103 can capture key turning points in individual development in real time. For example, when school-age children show signs of puberty such as a sudden increase in height and the emergence of secondary sexual characteristics, the system can immediately switch from the school-age feature sub-library to the early puberty sub-library, ensuring that feature analysis keeps pace with development and avoiding the lag in traditional static assessments where development has progressed but the assessment standards have not been updated.

[0062] In one embodiment, the feature sub-library calling and first physiological feature screening in step S2 include:

[0063] S201. Based on the determined physiological stage, the corresponding feature sub-library in the dynamic feature library is called (e.g., the preschool sub-library focuses on basal metabolic rate and nutritional intake characteristics, while the adolescent sub-library includes exclusive features such as growth hormone levels and sexual development indicators);

[0064] S202, matching the first data with the feature sub-library, and selecting the top 30% of the features ranked by relevance as the first physiological features through a feature similarity algorithm;

[0065] As described in steps S201-S202 above, S201 calls the corresponding feature sub-library based on the determined physiological stage, focusing feature analysis on the core needs of that stage. For example, the school-age sub-library focuses on basic growth characteristics such as basal metabolism and nutritional intake, while the adolescent sub-library focuses on specific features such as growth hormone and sexual development indicators. This precise stage-sub-library matching model avoids the interference of irrelevant features in traditional general feature libraries and improves the targeted nature of feature analysis. S202 uses a feature similarity algorithm to select the top 30% of relevant features as the first physiological feature, extracting the most relevant core information for the current physiological stage from the first data. This streamlined mechanism reduces the interference of redundant data on subsequent analysis and improves the efficiency and accuracy of feature vector construction.

[0066] In one embodiment, the second data collection, individualized feature vector construction, and feature correlation matrix generation in step S3 include:

[0067] S301, collecting second data, where the second data includes multi-dimensional biomedical data, including at least genomic data (such as CYP2D6 genotype), phenotypic group data (such as BMI percentile), environmental exposure group data (such as smoking history during pregnancy), and imaging group data;

[0068] S302, fusing the first physiological feature and the second data to construct an individualized feature vector; constructing a feature relationship graph through a graph neural network (GNN), capturing nonlinear correlations between features, and generating a feature correlation matrix (matrix elements quantify the correlation strength between features, with values ​​ranging from 0 to 1);

[0069] As described in steps S301-S302 above, at the data coverage level, the second data collected in step S301 includes multi-dimensional biomedical data, including genomic data (e.g., CYP2D6 genotype), phenotypic data (e.g., BMI percentiles), environmental exposure data (e.g., smoking history during pregnancy), and imaging data. This transcends the limitations of a single data type and captures comprehensive information from genes to the environment, from molecules to phenotypes. This multi-dimensional data collection approach more comprehensively reflects the physiological status and influencing factors of the research subject, providing a rich and solid data foundation for subsequent personalized analysis and avoiding analytical bias caused by one-sided data. At the feature construction and association mining level, step S302 fuses the first physiological characteristics and the second data to construct a personalized feature vector, effectively integrating data from different sources and types. This allows each research subject's characteristics to be presented in a unified vector format, facilitating standardized analysis and comparison. Constructing a feature relationship graph through a graph neural network (GNN) can accurately capture the nonlinear correlations between features. The generated feature correlation matrix (matrix elements quantify the correlation strength between features, with values ​​ranging from 0 to 1) presents this correlation in an intuitive and quantifiable form. This process breaks through the limitations of traditional linear analysis methods and can discover complex potential connections between features. It provides a powerful tool for in-depth understanding of physiological mechanisms and the discovery of key influencing factors, helping to improve the accuracy and interpretability of subsequent analytical models. These two steps provide a high-quality, high-value data processing foundation for related research or applications, ensuring the comprehensiveness and richness of the data while achieving deep integration and correlation mining of features through advanced algorithms. This lays a solid foundation for subsequent individualized assessment, prediction, or decision-making, effectively improving the scientific nature and reliability of the entire process.

[0070] In one embodiment, the generation of the personalized treatment plan in step S4 includes third data collection and drug metabolism simulation and treatment plan generation:

[0071] S401: Collect third data, where the third data includes physiological indicator information of the user, including at least real-time physiological indicators such as liver and kidney function, heart rate, and blood flow;

[0072] S402: In combination with the feature correlation matrix (such as CYP2D6 genotype-liver and kidney function correlation value), the absorption, distribution, metabolism and excretion of the drug in the body are simulated by the physiological pharmacokinetic model to predict the blood concentration-time curve; based on the simulation results, individualized treatment plan is generated, including drug type, dose, dosing frequency and monitoring period;

[0073] Wherein the drug dose is determined by the dose adjustment coefficient calculation formula:

[0074] Dose adjustment coefficient = (current patient AUC predicted value / healthy child AUC) x enzyme activity correction factor;

[0075] Wherein, AUC is the area under the curve of the blood concentration-time curve, and the enzyme activity correction factor is determined by the CYP2D6 genotype of the patient (such as 1 / 1 type is 1.0, 17 / 17 type is 0.3);

[0076] As described in steps S401-S402 above, S401 explicitly collects tertiary data to ensure the system can capture dynamic changes in an individual's physiological state. Compared to traditional methods that rely on fixed physiological parameters, this real-time data collection mode accurately reflects the individual's current physical function, providing real-time updated input parameters for drug metabolism simulation and avoiding simulation bias caused by changes in physiological state. For example, if a precocious adolescent's liver and kidney function suddenly increases to adult levels, the real-time collection of tertiary data allows the system to promptly detect this change, providing a basis for subsequent dosage adjustments. In S402, drug metabolism simulation is combined with a feature association matrix, breaking through the limitations of traditional single-dimensional analysis. The feature association matrix quantifies the implicit associations between multidimensional features, enabling physiological pharmacokinetic models to comprehensively consider the combined effects of genotype and real-time physiological indicators. For example, for precocious individuals with a CYP2D6 mutant gene and strong liver and kidney function, the model can use the association matrix to identify the counteracting effects of the gene mutation-induced slowed metabolism and the strong liver and kidney function-induced accelerated metabolism, enabling more accurate predictions of drug absorption and distribution in the body and reducing the risk of adverse reactions. Individualized drug dosage calibration is achieved through the dose adjustment coefficient calculation formula (dose adjustment coefficient = (current patient AUC predicted value / healthy child AUC) × enzyme activity correction factor). For precocious adolescents aged 13 years whose liver and kidney function has reached adult levels, the system can calculate a dose higher than the pediatric standard based on their predicted AUC value (higher than that of healthy children) and the enzyme activity correction factor (e.g., 1.0 for CYP2D61 / 1 type), thereby avoiding insufficient drug efficacy due to rapid metabolism. For individuals with CYP2D617 / 17 type (enzyme activity correction factor 0.3), even if real-time physiological indicators are normal, the formula will automatically reduce the dose to prevent cumulative toxicity caused by slow drug metabolism. This dosage calculation method considers both the metabolic differences between individuals and the healthy population and the genetically determined enzyme activity characteristics, making dosage recommendations more tailored to individual physiological characteristics.

[0077] The personalized treatment plan generated based on the simulation results includes the type of drug, dosage, dosing frequency, and monitoring cycle, forming a closed-loop recommendation of drug selection-dose-frequency-monitoring. For example, for precocious individuals with rapidly progressing bone age, the plan will not only recommend appropriate inhibitory drugs and dosages, but also set the dosing frequency based on the drug's metabolic rate (such as half-life), and formulate a monitoring plan based on the frequency of third-party data collection (such as weekly monitoring of liver and kidney function). This ensures that while improving efficacy, potential risks are discovered and avoided in a timely manner, solving the problem of traditional plans focusing only on dosage and ignoring full-process management. Steps S401-S402 achieve a transformation from empirical to data-driven precision drug treatment plans through a coherent process of real-time physiological data collection-multi-feature correlation simulation-individualized dose calculation-full-cycle plan generation. This effectively improves the safety and efficacy of medication and reduces the risk of adverse reactions, especially for special populations such as precocious puberty.

[0078] In one embodiment, real-time status data of the user is collected, and the weights of corresponding features in the feature library are dynamically adjusted by combining the first data, the second data, and the third data;

[0079] The user's real-time status data includes disease progression (such as bone age acceleration, secondary sexual characteristics development level), drug response (such as serum drug concentration, adverse reaction incidence) and lifestyle changes (such as nutritional intake, exercise frequency). The collection frequency is once a day for basic data + real-time collection when abnormalities occur;

[0080] The dynamic adjustment of feature weights is achieved through an attention mechanism: the first data, the second data, the third data, and the real-time status data are input into the attention model, and the weight coefficient of each feature is output (the sum of the weights is 1); the weight of features with a correlation of more than 0.8 with the real-time status data (such as liver enzyme indicators corresponding to adverse drug reactions) is increased to 1.5-2 times the initial value; the original output of the attention model is normalized using a softmax function so that the weight coefficients of each feature meet the condition of summing to 1, thereby avoiding decision bias caused by unbalanced weight distribution;

[0081] It also includes a feature weight decay strategy: for features that have not been activated in the feature association matrix for six consecutive months (association strength <0.1), the weight is automatically reduced to 50% of the initial value; features that have not been activated for 12 months are temporarily frozen and stored in the historical feature library for subsequent use.

[0082] It also includes a clinical feedback closed-loop mechanism: the doctor's actual decision-making results (such as drug dosage adjustment records) are recorded as feedback data, which is used together with the first data, second data, third data, and real-time status data to modify the feature library parameters. The feature association matrix and weight adjustment rules are updated every quarter.

[0083] As mentioned above, it covers multi-dimensional information such as disease progression (bone age growth rate, secondary sexual characteristics development level), drug response (serum drug concentration, adverse reaction incidence) and lifestyle changes (nutritional intake, exercise frequency), and adopts a model of daily basic data + real-time collection when abnormalities occur. It can not only continuously track the normal status, but also respond to sudden changes (such as the sudden onset of adverse drug reactions) in a timely manner. Compared with the traditional method that relies on static data from regular follow-up visits, this collection method can accurately reflect the real-time status of the individual during the treatment process (such as accelerated bone age growth in precocious adolescents, sudden increase in serum concentration of antidepressants, etc.), providing a fresh basis for subsequent weight adjustments and avoiding decision-making bias caused by information lags;

[0084] The attention model fuses and analyzes the four types of data, automatically outputting weight coefficients for each feature. This increases the weight of features with a correlation of more than 0.8 with real-time status data (such as liver enzyme indicators for adverse drug reactions) to 1.5-2 times their initial value. This on-demand weighting mechanism addresses the limitations of traditional fixed-weight models: when premature individuals experience adverse drug reactions, the system prioritizes the weighting of relevant features such as liver and kidney function indicators and drug-metabolizing enzyme genotypes, ensuring that subsequent treatment adjustments focus on the core triggers of the adverse reaction. When lifestyle changes (such as a sudden decrease in nutrient intake) affect growth and development, the weight of nutrition-related features is automatically increased, making the intervention plan more tailored to individual behavioral habits and improving compliance.

[0085] Features that have been inactive for six consecutive months (correlation strength <0.1) are weighted down to 50%, and features that have been inactive for 12 months are temporarily frozen to prevent redundant features (such as baseline indicators unrelated to the current condition) from occupying system resources. For example, for precocious individuals who have entered late puberty, nutritional characteristics from school age have been attenuated due to long-term lack of correlation. This allows the system to more efficiently focus on core features such as bone age closure rate and sex hormone levels, reducing inefficient computation and improving decision response speed. Physicians' actual decision results (such as medication dosage adjustment records) are used as feedback data, combined with four types of data to refine feature library parameters. The feature correlation matrix and weight adjustment rules are updated quarterly. This closed-loop design of model prediction, clinical validation, and parameter optimization can promptly correct model deviations (e.g., reverse optimization of feature correlation strength when actual dosage adjustments do not match predictions) and continuously adapt the feature library to changes in clinical practice (e.g., incorporating metabolic characteristics of new drugs into the correlation matrix), ensuring the system's long-term accuracy and practicality. This is particularly applicable to areas such as precocious puberty in children and adolescents, where diagnostic and treatment plans must be continuously updated with medical advances.

[0086] In summary, these steps, through the full-process design of real-time perception-dynamic weighting-streamlined optimization-clinical iteration, ensure that the feature library is always synchronized with individual status and clinical needs. This not only improves the accuracy of current decisions, but also ensures the long-term evolutionary capabilities of the system, providing sustainable technical support for personalized medical decision-making.

[0087] In one embodiment, when applied to medication for adolescent mental illness, dynamic adjustment of antidepressant dosage requires synchronous reference to the growth curve-serum drug concentration correlation value in the characteristic association matrix. When the PBPK model-predicted concentration is greater than the upper limit of the therapeutic window, a dosage adjustment reminder is automatically triggered and an alternative plan is pushed.

[0088] As mentioned above, adolescents are in a critical period of growth and development, with rapidly changing growth curves. This significantly impacts the metabolism and efficacy of antidepressants. Simultaneously referencing the growth curve-serum drug concentration correlation value in the characteristic correlation matrix can fully account for the impact of growth and development on drug concentrations. For example, adolescents undergoing a growth spurt may experience significant changes in their metabolic capacity. This correlation value reflects the relationship between this change and serum drug concentration, providing a more tailored basis for dose adjustments tailored to individual growth status and avoiding inappropriate dosing due to neglecting growth and development factors. When the PBPK model-predicted concentration exceeds the upper limit of the therapeutic window, an automatic dose adjustment reminder and alternative plan notification are automatically triggered, effectively mitigating the risks of excessive drug concentrations. Adolescents' organs are not yet fully developed, and their tolerance to drugs is relatively low. Excessive drug concentrations can trigger serious adverse reactions, such as excessive central nervous system suppression. This automatic reminder and alternative plan notification mechanism can promptly alert physicians to potential risks, enabling them to quickly make adjustments and ensure safe medication use in adolescents.

[0089] By combining the relationship between growth curves and serum drug concentrations with model predictions, antidepressant dosages can be adjusted more precisely to maintain effective therapeutic concentrations in the body. This helps improve treatment efficacy, shorten treatment cycles, and reduce instances of poor treatment efficacy or repeated dose adjustments due to inappropriate dosages, allowing adolescents to recover more quickly from mental illness.

[0090] In one embodiment, when applied to neonatal jaundice management, the feature association matrix needs to include the three-dimensional feature association of age, bilirubin level, and feeding method. The formula for calculating the duration of phototherapy is:

[0091] Phototherapy time = age in days × peak bilirubin × breastfeeding frequency × correction factor

[0092] The correction factor is determined by the patient's gestational age (e.g., 1.2 for <37 weeks) and blood type (e.g., 1.5 for Rh negative);

[0093] As mentioned above, individualized phototherapy plans can be formulated: the severity and recovery of neonatal jaundice are closely related to multiple factors such as age, bilirubin peak, and feeding method. The feature association matrix incorporates the three-dimensional feature association of age-bilirubin level-feeding method, which can comprehensively consider the impact of these factors on jaundice. For example, newborns of different ages have different abilities to metabolize bilirubin, and there are also differences in bilirubin excretion between breastfed and formula-fed newborns. The three-dimensional association can accurately capture these differences and provide a comprehensive basis for the formulation of phototherapy plans;

[0094] The phototherapy duration calculation formula combines gestational age and blood type to determine a correction factor, accurately calculating the appropriate phototherapy duration for each newborn. For premature infants with a gestational age of less than 37 weeks, whose physical functions are relatively weak and bilirubin metabolism is slow, a correction factor of 1.2 can appropriately extend the phototherapy duration and ensure effective treatment of jaundice. For newborns with Rh-negative blood, due to special circumstances such as possible hemolysis, a correction factor of 1.5 is set to ensure sufficient phototherapy duration. This precise calculation can avoid persistent jaundice due to insufficient phototherapy time or unnecessary harm to the newborn caused by excessive phototherapy time, thereby improving treatment effectiveness. A reasonable phototherapy duration can effectively reduce the risk of hyperbilirubinemia damaging the newborn's nervous system. Through accurate phototherapy time calculation and personalized plan development, bilirubin levels can be reduced to a safe range in the shortest possible time, reducing the risk of serious complications such as kernicterus and ensuring the healthy growth of the newborn.

[0095] In one embodiment, a federated learning framework is used to optimize the cross-institutional feature library: each medical institution locally retains the first data, second data, third data, and real-time status data, and only uploads the feature weight adjustment gradient to the central server. Laplace noise (ε=1.0) is added during the gradient upload process to ensure data privacy. The central server aggregates the gradients and sends the updated parameters.

[0096] As mentioned above, each medical institution retains the original information such as the first data, second data, third data and real-time status data locally, and only uploads the feature weight adjustment gradient to the central server, which solves the problem of data islands in traditional cross-institutional collaboration. Different institutions (such as children's hospitals, mental health centers, and maternal and child health hospitals) can contribute model optimization directions based on their own clinical data, so that the feature library can integrate multi-center and multi-scenario medical experience (such as comprehensive diagnosis and treatment data of precocious puberty in children and adolescents in different regions, and differentiated cases of neonatal jaundice). For example, a children's hospital has accumulated rich data on the use of drugs for adolescent mental illness, and another institution excels in the management of neonatal jaundice. Through federated learning, the feature weight adjustment gradients of the two can be aggregated to form more comprehensive optimization parameters, so that the scope of application of the feature library is expanded from a single institution to multiple scenarios, thereby improving the universality of the system;

[0097] During the gradient upload process, Laplace noise (ε=1.0) is added to anonymize sensitive information using differential privacy techniques, ensuring that personal patient data (such as genomic information and medical records) cannot be deduced from the uploaded gradients. A Laplace noise parameter of ε=1.0, evaluated against a differential privacy budget, maintains model convergence efficiency while protecting individual privacy (ensuring that patient data cannot be deduced from the gradients). In actual deployments, the ε value can be adjusted based on the sensitivity of the institution's data (e.g., ε=0.5 for scenarios with higher privacy requirements). This "data remains static, model remains dynamic" model not only meets regulatory requirements for medical data privacy protection (such as HIPAA and GDPR), but also alleviates institutional concerns about data leakage, encouraging more medical institutions to participate in collaboration. For example, noise processing can prevent the identification of private data (such as medical history and medication response) of adolescent patients with mental illness during cross-institutional transmission, allowing institutions to confidently contribute data value without worrying about compliance risks.

[0098] The central server aggregates the gradients of each institution and then issues updated parameters, allowing the feature library to quickly absorb multi-source feedback and complete iterations. Compared with the isolated optimization of a single institution, this distributed collaborative model greatly increases the sample size and diversity of data, making the optimization of parameters such as the feature association matrix and weight adjustment rules more statistically significant. For example, for the topic of precocious puberty in children and adolescents, cases from a single institution may be limited to a certain region or age group, while cross-institutional collaboration can include samples from different races and living environments, making the mining of feature associations such as dietary structure, hormone levels, and bone age growth rate more accurate. The update speed of the feature library has also increased from the quarterly level of a single institution to the monthly or even weekly level, adapting more quickly to changes in clinical needs.

[0099] Institutions no longer need to standardize data formats or perform large-scale data migrations; they only need to complete model training locally and upload gradients, significantly reducing the technical barriers and costs (such as the manpower required for data cleaning and format conversion) of cross-institutional collaboration. Furthermore, the federated learning framework supports incremental learning. When new institutions join or existing ones expand their business scenarios, they can seamlessly integrate into the collaborative network and continuously contribute to the optimization of the feature library. For example, a newly established maternal and child health care hospital can quickly access the system, and its unique data on neonatal jaundice management (such as phototherapy responses of specific blood types) can be incorporated into the feature library through gradient upload, eliminating the need to restructure the entire collaborative system and ensuring the system's scalability and compatibility.

[0100] In summary, the federated learning framework makes cross-institutional feature library optimization sustainable by balancing collaborative efficiency and privacy protection. It not only integrates scattered medical resources but also maintains the bottom line of data security, providing a strong driving force for the long-term evolution of medical decision-making support systems. It is especially suitable for complex medical scenarios such as precocious puberty in children and adolescents, neonatal jaundice, etc. that require multi-center data support.

[0101] In the diagnosis and treatment scenario of precocious puberty in children and adolescents, a comparative experiment between this method and the traditional fixed-age plan showed that the accuracy of medication was improved by 32% (p < 0.05), the incidence of adverse reactions was reduced by 27%, and the misjudgment rate of bone age was reduced from 18% to 5%, verifying the clinical effectiveness of the plan.

[0102] Example 2, please refer to Figure 2 As shown, the medical decision-making support system integrating multimodality and deep network described in this embodiment includes:

[0103] A data collection module, used to collect first data, second data, third data and real-time status data of the user;

[0104] The data acquisition module includes:

[0105] Basic physiological data collection unit, used to collect the first data, supporting manual input and automatic synchronization of smart devices (such as data docking of smart height and weight scales);

[0106] The biomedical data collection unit is used to collect secondary data and supports the import of genomic data files (such as FastQ format), standardized entry of phenotypic group data, and the generation and collection of environmental exposure group data questionnaires. When collecting genomic data, quality control processes (such as FastQC to check sequencing quality and remove low-quality reads) are used to ensure data reliability. Environmental exposure group data are verified through questionnaire logic (such as mandatory item checking and range verification) to reduce subjective bias.

[0107] Real-time physiological index acquisition unit, used to collect third-party data, supports docking with the hospital HIS system and monitor interface, and realizes real-time transmission of indicators such as liver and kidney function and heart rate;

[0108] The status data collection unit is used to collect the user's real-time status data, support doctors to manually record changes in condition and drug reactions, and support users to upload lifestyle data.

[0109] a stage determination module, configured to analyze the first data using a deep network learning model to determine the physiological stage of the user;

[0110] The module for determining the user's physiological stage has a built-in long short-term memory network (LSTM) model. This model is trained with multi-center physiological data. The input is the time-series first data, and the output is the probability distribution of the user's physiological stage (such as a 70% probability for preschool and a 30% probability for early adolescence). It also has a developmental stage transition detection submodule. When it detects that the physiological data crosses the stage threshold, it automatically triggers the feature sub-library switching instruction.

[0111] a feature screening module, configured to call a corresponding feature sub-library in the dynamic feature library based on the determined physiological stage, match the first data with the feature sub-library, and screen out the first physiological feature;

[0112] The feature screening module includes a feature sub-library management unit and a similarity matching unit; the feature sub-library management unit is used to store and call feature sub-libraries of different physiological stages, and the similarity matching unit calculates the matching degree between the first data and each feature in the feature sub-library using a cosine similarity algorithm, and screens out the features with the top 30% correlation as the first physiological feature.

[0113] The feature fusion and association module is used to fuse the first physiological feature and the second data to construct an individualized feature vector, capture the nonlinear correlation between features through the graph neural network, and generate a feature correlation matrix;

[0114] The feature fusion and association module includes:

[0115] a vector construction unit, configured to integrate the first physiological feature and the second data into an individualized feature vector after normalization processing;

[0116] The graph neural network unit is used to construct a feature relationship graph, capture the nonlinear correlation between features through multi-layer graph convolution operations, generate a feature correlation matrix and store it in a matrix database.

[0117] The treatment plan generation module is used to collect third-party data in real time, combine it with the characteristic correlation matrix, simulate the drug's process in the body through the physiological pharmacokinetic model, and generate an individualized treatment plan based on the simulation results;

[0118] The treatment plan generation module includes a PBPK model unit and a plan optimization unit; the PBPK model unit combines third-party data and a characteristic association matrix to simulate the drug metabolism process and outputs a blood drug concentration prediction result; the plan optimization unit generates an individualized treatment plan including drug type, dosage, and dosing frequency based on the prediction result and the dose adjustment coefficient formula, and supports visual display of the plan (such as a blood drug concentration-time curve).

[0119] The weight adjustment module is used to dynamically adjust the weight of the corresponding feature in the feature library by combining the first data, the second data, the third data and the user's real-time status data.

[0120] After collecting the user's real-time status data (including at least changes in condition and drug reactions), the weight adjustment module dynamically adjusts the weights of corresponding features in the feature library based on the first data, the second data, and the third data, specifically including:

[0121] a data association unit, configured to associate and match the real-time status data with the first data, the second data, and the third data, and determine features related to the real-time status in each data set;

[0122] The weight calculation unit calculates the weight adjustment coefficient of each feature based on the association matching results through the attention model. The coefficient size is positively correlated with the degree of influence of the feature on the real-time state;

[0123] The weight update unit updates the weight of the corresponding feature in the feature library according to the weight adjustment coefficient, and synchronously stores the weight value before and after the update and the update time;

[0124] The verification unit verifies whether the updated weights are compatible with the user's physiological stage (determined based on the first data), biomedical characteristics (based on the second data), and real-time physiological indicators (based on the third data). If there is any adaptation anomaly, a secondary adjustment is triggered.

[0125] Introducing regularization methods in the weight adjustment module:

[0126] For highly correlated features (e.g., Pearson correlation coefficient > 0.8), L1 regularization is used to constrain the weight adjustment amplitude to avoid model overfitting caused by multicollinearity and ensure the independence and interpretability of feature weights;

[0127] The system also includes a clinical feedback module for receiving the actual decision results input by the doctor, associating the results with the first data, the second data, the third data and the real-time status data, and storing them as a basis for revising the feature library parameters, and triggering the update of the feature association matrix and weight adjustment rules once a quarter;

[0128] It also includes a federated learning module for achieving cross-institutional feature library optimization. The module consists of a local training unit, a gradient encryption unit, and a parameter aggregation unit. The local training unit performs model training locally in each medical institution, the gradient encryption unit adds Laplace noise (ε=1.0) to the feature weight adjustment gradient, and the parameter aggregation unit receives the encrypted gradients of each institution and aggregates them before issuing updated parameters.

[0129] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A medical decision-making assistance method integrating multimodality and deep networks, characterized in that: include: Collecting and analyzing the user's first data to determine the user's physiological stage; Based on the physiological stage of the user, calling a corresponding feature sub-library in a dynamic feature library, matching the first data with the feature sub-library, and screening out a first physiological feature that is highly correlated with the physiological stage of the user; collecting second data, fusing the first physiological characteristic and the second data to construct an individualized feature vector, and generating a feature correlation matrix; The third data is collected in real time, combined with the characteristic association matrix, and the process of the drug in the user's body is simulated through a physiological pharmacokinetic model, and an individualized treatment plan is generated based on the simulation results.

2. The medical decision-making assistance method integrating multimodality and deep network according to claim 1, characterized in that: Collecting real-time status data of the user, combining the first data, the second data, and the third data, and dynamically adjusting the weights of corresponding features in the feature library; Collect real-time status data of the user, the real-time status data including at least changes in the condition and drug reactions, wherein changes in the condition include but are not limited to worsening or relief of symptoms, occurrence of complications, and abnormal fluctuations in physiological indicators, and drug reactions include but are not limited to the time it takes for drug efficacy to appear, the type and degree of adverse reactions, and changes in serum drug concentrations; dynamically adjust the weights of corresponding features in the feature library based on the first data, the second data, and the third data.

3. The medical decision-making assistance method integrating multimodality and deep network according to claim 1, characterized in that: The first data is physiological data information of the user, including at least age, height, weight, and secondary sexual characteristics; the determination of the user's physiological stage specifically comprises automatically analyzing the first data using a long short-term memory network model to determine the user's developmental stage, and introducing a developmental stage transition detection mechanism to automatically trigger a feature sub-library switch when the user's physiological data crosses a preset stage threshold; The feature sub-library in the dynamic feature library is a subdivided library pre-constructed for different physiological stages. When matching the first data with the feature sub-library, the features ranked in the top 30% of the correlation are screened out as the first physiological features through a feature similarity algorithm.

4. The medical decision-making assistance method integrating multimodality and deep network according to claim 1, characterized in that: The second data is multi-dimensional biomedical data, including genome, phenotype, and environmental exposure group data; a feature relationship graph is constructed through a graph neural network to capture the nonlinear association between the first physiological feature and the second data, and a feature association matrix is ​​generated, in which the matrix elements quantify the association strength between the features; The third data is the user's real-time physiological index information, including at least liver and kidney function indicators, heart rate, and blood flow; The physiological pharmacokinetic model combines the third data and the metabolism-related features in the feature association matrix to simulate the absorption, distribution, metabolism, and excretion processes of the drug in the body and predict the blood drug concentration-time curve.

5. The medical decision-making assistance method integrating multimodality and deep network according to claim 4, characterized in that: The individualized treatment plan includes drug type, dosage, administration frequency and monitoring period; the drug dosage is determined by the dosage adjustment coefficient calculation formula: Dose adjustment factor = (predicted AUC value for current patients / AUC value for healthy children) × enzyme activity correction factor; Wherein, AUC is the area under the curve of the blood drug concentration-time curve, and the enzyme activity correction factor is determined by the patient's CYP2D6 genotype.

6. The medical decision-making assistance method integrating multimodality and deep network according to claim 1, characterized in that: The user's real-time status data includes disease progression, drug response and lifestyle changes; the collection frequency is once a day for basic data, and the real-time collection mechanism is activated when an abnormal state occurs.

7. The medical decision-making assistance method integrating multimodality and deep network according to claim 2, characterized in that: The dynamic adjustment of the weights of corresponding features in the feature library is specifically achieved through the attention mechanism: the first data, the second data, the third data and the real-time status data are input into the attention model, and the weight coefficient of each feature is output; the weight of the feature whose correlation with the real-time status data exceeds 0.8 is increased to 1.5-2 times the initial value.

8. The medical decision-making assistance method integrating multimodality and deep network according to claim 2, characterized in that: It includes a feature weight decay strategy: for features that have not been activated in the feature association matrix for 6 consecutive months, the weight is automatically reduced to 50% of the initial value; features that have not been activated for 12 months are temporarily frozen and stored in the historical feature library for subsequent use.

9. The medical decision-making assistance method integrating multimodality and deep network according to claim 2, characterized in that: It includes a clinical feedback closed-loop mechanism: the doctor's actual decision-making results are recorded as feedback data, which are used together with the first data, second data, third data and real-time status data to correct the feature library parameters. The feature association matrix and weight adjustment rules are updated once a quarter, and a dynamic update mechanism is introduced to adjust the update frequency according to the magnitude of data changes.

10. A medical decision-making assistance system integrating multimodality and deep network, for implementing the medical decision-making assistance method integrating multimodality and deep network according to any one of claims 1 to 9, characterized in that: include: A data collection module, used to collect first data, second data, third data and real-time status data of the user; a stage determination module, configured to analyze the first data using a deep network learning model to determine the physiological stage of the user; a feature screening module, configured to call a corresponding feature sub-library in the dynamic feature library based on the determined physiological stage, match the first data with the feature sub-library, and screen out the first physiological feature; The feature fusion and association module is used to fuse the first physiological feature and the second data to construct an individualized feature vector, capture the nonlinear correlation between features through the graph neural network, and generate a feature correlation matrix; The treatment plan generation module is used to collect third-party data in real time, combine it with the characteristic correlation matrix, simulate the drug's process in the body through the physiological pharmacokinetic model, and generate an individualized treatment plan based on the simulation results; The weight adjustment module is used to dynamically adjust the weight of the corresponding feature in the feature library by combining the first data, the second data, the third data and the user's real-time status data.

Images