Machine learning-based internal medicine disease treatment scheme optimization system and method

Abstract

The invention belongs to the technical field of medical informatics, and discloses an internal medicine disease treatment scheme optimization system and method based on machine learning, and the system comprises a heterogeneous data monitoring module which is integrated with a multi-source heterogeneous medical database through a standard medical API interface, and extracts and obtains internal medicine disease heterogeneous data; the immune response network module simulates an antigen recognition response mode, constructs a treatment scheme antigen representation matrix according to the internal medicine disease heterogeneous data, recognizes abnormal treatment modes through transfer learning, removes invalid treatment schemes or treatment schemes with side effects corresponding to all the abnormal treatment modes, and then obtains historical effective treatment schemes; a neural symbol memory mechanism is adopted, historical effective treatment scheme fragments are stored, and a treatment scheme variant is constructed based on an adversarial generation mechanism; and the accuracy, individuation and safety of a treatment scheme are improved, so that the treatment effect of internal medicine diseases is remarkably improved, side effects and wrong schemes are reduced, and the development of precise medical treatment is promoted.

Description

Technical Field

[0001] The present invention relates to the technical field of medical informatics, and more specifically, to a system and method for optimizing medical disease treatment plans based on machine learning. Background Art

[0002] Patent announcement number CN118430791A discloses a monitoring method and TCM rehabilitation treatment system based on cardiovascular diseases. The method uses a risk prediction model to predict the probability of cardiovascular diseases. If the probability of cardiovascular diseases is higher than a first threshold and lower than a second threshold, a TCM health regimen for cardiovascular diseases is pushed to the monitoring user. If the probability of cardiovascular diseases is higher than the second threshold, a medical consultation prompt is output. After the medical consultation, the cardiovascular disease is diagnosed and the user enters the rehabilitation system. A TCM rehabilitation guidance program for cardiovascular diseases is generated through a treatment and rehabilitation model, and the TCM rehabilitation guidance program is pushed to the monitoring user. This solution assists users in the rehabilitation treatment of cardiovascular diseases by pushing TCM health regimens or TCM rehabilitation guidance programs to users, thereby improving the rehabilitation treatment effect of cardiovascular diseases.

[0003] The existing internal medicine treatment plan optimization system and method have the following main problems:

[0004] Traditional methods typically use a preset fixed threshold for the probability of harmful antigens to determine whether a treatment plan is abnormal. However, in practice, fixed thresholds can lead to high false positives or missed detections. This static threshold setting can easily lead to high false positives or missed detections, compromising the accuracy and safety of the system. Existing technologies are unable to effectively adapt to low-quality data, significantly impacting model performance in the presence of poor data quality.

[0005] Most of the treatment plan optimization methods in the existing technology rely on standardized and unified rules. The evaluation criteria of the treatment plan are usually fixed, which makes it difficult to make personalized adjustments based on the characteristics of different patients. Therefore, the existing methods fail to fully consider individual differences, which may lead to inaccurate treatment effects or treatment plans that do not meet the specific needs of patients; traditional drug treatment evaluation methods often treat pharmacokinetics and pharmacodynamics separately, and lack a mechanism for comprehensively evaluating the metabolic efficiency of drug treatment. The existing schemes are unable to abstract the treatment process as a dynamic energy process, and in particular lack an efficacy evaluation system for biological energy circulation mechanisms; because the drug treatment metabolic process and pharmacodynamic process are usually ignored or handled improperly, the existing technology cannot accurately evaluate the therapeutic metabolic equivalent of the drug, making it difficult to predict potential side effect risks in the early stages of efficacy evaluation.

[0006] In view of this, the present invention proposes a system and method for optimizing medical disease treatment plans based on machine learning to solve the above problems. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention provides the following technical solutions: a system for optimizing medical disease treatment plans based on machine learning, comprising:

[0008] The heterogeneous data monitoring module integrates with multi-source heterogeneous medical databases through standard medical API interfaces to extract and obtain heterogeneous data on internal medicine diseases;

[0009] The immune response network module mimics the antigen recognition response pattern and constructs a treatment plan antigen representation matrix based on heterogeneous data of internal medicine diseases. It uses transfer learning to identify abnormal treatment patterns and eliminates ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment patterns, thereby obtaining historically effective treatment plans. It uses a neural symbol memory mechanism to store historically effective treatment plan fragments and constructs treatment plan variants based on an adversarial generation mechanism.

[0010] The pathological metabolic cycle module extracts key features of the treatment plan antigen representation matrix, draws on the energy conversion mechanism of the tricarboxylic acid cycle, and calculates the drug's metabolic rate and therapeutic effect intensity based on these key features to obtain the drug's therapeutic metabolic equivalent. This is then combined with treatment plan variants to optimize the treatment plan through a dynamic feedback control mechanism.

[0011] The negative feedback recombination module decomposes historical ineffective treatment plans or treatment plans with side effects into recombinable knowledge components, inputs them into the internal medicine disease treatment monitoring terminal, and continuously optimizes to obtain the optimal treatment plan.

[0012] Preferably, the method for acquiring heterogeneous data of internal medicine diseases includes:

[0013] The multi-source heterogeneous medical database includes an electronic medical record platform, a medical imaging archive management platform, and a wearable device data platform. By supporting the standard medical API interface, a communication connection is established with the multi-source heterogeneous medical database, data is retrieved through the standard medical API interface, heterogeneous data of internal medicine diseases are extracted, and the heterogeneous data of internal medicine diseases are normalized. The heterogeneous data of internal medicine diseases include electronic medical record data, patient physiological data, medical imaging data, and patient feedback data.

[0014] Preferably, the method for identifying abnormal treatment patterns through transfer learning includes:

[0015] The historical treatment plans corresponding to each internal medicine disease are obtained from a multi-source heterogeneous medical database. Each historical treatment plan is defined as an antigen. Historical treatment plans include effective treatment plans and ineffective treatment plans or treatment plans with side effects. Effective treatment plans are defined as friendly antigens, while ineffective treatment plans or treatment plans with side effects are defined as harmful antigens. A treatment plan antigen representation matrix is constructed through analogy.

[0016] Wavelet transform is used to extract features from electronic medical record data, patient physiological data, medical imaging data, and patient feedback data in heterogeneous data of internal medicine diseases, and then the features are spliced to obtain the structural feature vector of the unified representation space. The structural feature vector is used as the representation structure of the antigen to construct the treatment plan antigen representation matrix. The rows of the treatment plan antigen representation matrix represent each treatment plan antigen; the columns of the treatment plan antigen representation matrix represent the structural feature vector dimensions of the treatment plan in the unified representation space.

[0017] The historical treatment plan is set as the source domain, and the current treatment plan is set as the target domain. A transfer learning model consisting of a shared feature extractor and a classifier sub-network is used. The model's training samples include historical treatment plans and the probability that the historical treatment plan is a friendly antigen or a harmful antigen. The corresponding shared feature extractor is used to extract latent semantic features from the antigen representation vector, and the classifier sub-network is used to determine whether the current treatment plan is an abnormal pattern.

[0018] After training is completed, the current treatment plan in the target domain is input into the transfer learning model, and the probability of the historical treatment plan being a friendly antigen or a harmful antigen is output; the harmful antigen probability threshold is preset Among them, θ0 represents the preset basic harmful antigen probability threshold; α represents the adjustment factor; represents the integrity score of the current training sample; b represents the number of antigens in each treatment regimen; m represents the number of dimensions of the structural feature vector of the treatment regimen in the unified representation space;

[0019] When the probability of harmful antigens is greater than or equal to the preset harmful antigen probability threshold, it is judged that the current treatment plan is abnormal. All abnormal treatment modes are identified through the transfer learning model, and the ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment modes are eliminated to obtain effective treatment plans.

[0020] Preferably, the method for storing historical effective treatment plan fragments includes:

[0021] Obtain historical effective treatment plans from multi-source heterogeneous medical databases and store them in segments using a neural symbolic memory mechanism; historical effective treatment plans contain u-dimensional treatment process data; discretize the treatment process into N time points, and then construct a multidimensional time series Among them, t' k is the kth discrete time point; v k is the treatment state vector at the kth discrete time point; for each discrete time point, all treatment process data are collected to construct the treatment state vector v k ; k is the index of the scattered time point, k = 1, 2, ..., N; u is the dimension of the treatment process data; R is the real number domain;

[0022] Perform standard deviation normalization on the multidimensional time series, perform change point detection on the multidimensional time series, use cosine similarity to calculate the difference in the treatment state vectors of two adjacent discrete time points in the multidimensional time series, and obtain the treatment state jump value The preset treatment state transition threshold δ is satisfied. The k+1th discrete time point is regarded as a change point, and all change points are combined to obtain a change point sequence; the multidimensional time series is segmented according to the change point sequence, and a neural encoder is constructed to perform neural encoding on each segment;

[0023] Construct a neural encoder, take the multidimensional time series data contained in each segment as input to the neural encoder, output the corresponding high-dimensional embedding vector, preset a set of symbol annotation rules, use vocabulary encoding to semantically classify each segment according to the preset symbol annotation rule set, obtain the corresponding symbol label, encode the symbol label through the word embedding model, convert it into a vector representation, and then obtain the symbol embedding vector;

[0024] The multidimensional time series data contained in any segment is used as the value of the memory unit, the high-dimensional embedding vector corresponding to the segment is used as the key of the memory unit, and the symbol embedding vector corresponding to the segment is used as the annotation of the memory unit to construct a set of key-value-annotation memory pairs; the key-value-annotation memory pairs of all segments are collected and stored in the historical effective treatment plan memory library to store historical effective treatment plan fragments.

[0025] Preferably, the method for constructing the treatment regimen variant comprises:

[0026] Construct and train a generative adversarial network, which includes a generator and a discriminator. The generator accepts random noise and generates a false effective treatment plan, which is input into the discriminator. The discriminator compares the false effective treatment plan with the memory bank of historical effective treatment plans and outputs a true or false discrimination result. If the discrimination result is false, a false effective treatment plan is regenerated and compared again through the discriminator. The iteration is repeated until the discrimination result is true, thereby obtaining a trained generative adversarial network, and outputting a treatment plan variant through the trained generative adversarial network.

[0027] Preferably, the key characteristics include the pharmacokinetic characteristics and pharmacodynamic characteristics of the treatment regimen; the pharmacokinetic characteristics of the treatment regimen include the efficiency of the drug entering the blood circulation, the time required for the drug concentration in the patient's body to drop to half, the distribution range of the drug in the patient's body, and the rate of elimination of the drug from the patient's body; the pharmacodynamic characteristics include the maximum therapeutic effect of the drug, the blood concentration required for the drug to produce half-maximal effect, and the intensity of the drug's effect on different diseases.

[0028] Preferably, the method for obtaining the drug therapeutic metabolic equivalent comprises:

[0029] The therapeutic metabolic equivalent of a drug is analogized to the energy conversion mechanism in the tricarboxylic acid cycle, and the drug metabolism process is regarded as a dynamic process similar to the tricarboxylic acid cycle. The dynamic process of the tricarboxylic acid cycle represents the various stages from the drug entering the body to the drug's efficacy, including absorption, distribution, metabolism, and excretion. It is assumed that each step of drug metabolism contributes different amounts of energy. The total energy released during the treatment process is estimated by the drug's metabolic rate and the intensity of the therapeutic effect, thereby obtaining the drug's therapeutic metabolic equivalent.

[0030] The metabolic rate of the drug is described by the kinetic rate equation of drug concentration changing with time. The kinetic rate equation is C(t) = C0·e -λt ; Where C(t) represents the blood drug concentration at time point t; C0 represents the initial drug concentration; λ represents the elimination rate constant of the drug; t represents the time the drug has passed in the patient's body;

[0031] The metabolic rate of the drug is obtained by derivatizing C(t). Combining the metabolic rate of the drug and the intensity of the therapeutic effect, the therapeutic metabolic equivalent of the drug is calculated. Among them, E max Indicates the maximum therapeutic effect of the drug; V max Indicates the maximum metabolic rate of the drug; d met Represents the metabolic rate constant of the drug.

[0032] Preferably, the method for optimizing the treatment plan through a dynamic feedback control mechanism includes:

[0033] Using the therapeutic metabolic equivalent of the drug as a reward signal, the treatment variants are continuously adjusted through reinforcement learning. Based on patient feedback data, the treatment plan is dynamically optimized based on reinforcement learning, and ultimately the treatment variant that best suits the patient is identified.

[0034] Preferably, the method for obtaining the optimal treatment plan comprises:

[0035] Decompose each historical ineffective treatment regimen or treatment regimen with side effects into L components, including drug type, drug dosage, treatment method, administration route, and treatment duration; each component represents a knowledge component, and each knowledge component contains its impact on treatment efficacy and side effect label information;

[0036] The internal medicine disease treatment monitoring terminal issues a comparison instruction to compare the patient feedback data with the decomposed knowledge components, automatically adjusts and reorganizes the knowledge components in each historical ineffective treatment plan or treatment plan with side effects, thereby forming a new treatment plan. By calculating the Jaccard similarity between the patient feedback data and the new treatment plan, the most matching treatment plan is recommended for each patient, and then the optimal treatment plan for each patient is generated.

[0037] A method for optimizing treatment plans for internal medicine diseases based on machine learning, comprising:

[0038] S1, heterogeneous data monitoring module, integrates with multi-source heterogeneous medical databases through standard medical API interfaces to extract and obtain heterogeneous data on internal medicine diseases;

[0039] S2: Mimicking the antigen recognition response pattern, constructing a treatment plan antigen representation matrix based on heterogeneous data of internal medicine diseases, identifying abnormal treatment patterns through transfer learning, eliminating ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment patterns, and then obtaining historically effective treatment plans; using a neural symbolic memory mechanism to store historically effective treatment plan fragments, and constructing treatment plan variants based on an adversarial generation mechanism;

[0040] S3. Extract key features of the therapeutic antigen representation matrix. Drawing on the energy conversion mechanism of the tricarboxylic acid cycle, calculate the drug's metabolic rate and therapeutic effect intensity based on these key features to obtain the drug's therapeutic metabolic equivalent. Combine the drug's therapeutic metabolic equivalent with the therapeutic variants and optimize the treatment plan through a dynamic feedback control mechanism.

[0041] S4. Decompose historical ineffective treatment plans or treatment plans with side effects into recombinable knowledge components, input them into the internal medicine disease treatment monitoring terminal, and continuously optimize to obtain the optimal treatment plan.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] This invention dynamically adjusts the threshold by introducing a completeness score, which allows for flexible adjustment of the judgment criteria based on the quality of different data, thereby improving detection accuracy, especially when faced with uneven data quality. By assessing the completeness of the sample, the model can make targeted adjustments to low-quality and high-quality data, improving the reliability and stability of the model on data of varying quality.

[0044] The transfer learning model based on dynamic threshold adjustment can flexibly adjust the judgment criteria for abnormal patterns according to the characteristics of different patients and the actual effects of the treatment plan, and optimize the evaluation process of the treatment plan in real time; the model can automatically adjust the evaluation criteria based on the patient's individualized data (such as historical treatment effects, feedback data, etc.), making the treatment plan more personalized and targeted, thereby improving the accuracy of the treatment effect; through the adjustable threshold mechanism and integrity score, doctors can understand how the model makes treatment plan evaluations based on sample quality in specific situations, thereby enhancing the transparency and trust of the model.

[0045] The introduction of the therapeutic metabolic equivalent (TME) comprehensively considers a drug's metabolic rate, maximum metabolic capacity, and maximum therapeutic effect, enabling a more comprehensive assessment of therapeutic efficacy. By analogy with the dynamic energy release mechanism of the tricarboxylic acid cycle, the drug's absorption, distribution, metabolism, and excretion processes are modeled as a continuous energy conversion chain. A first-order pharmacokinetic model is used to describe the temporal evolution of drug concentration, and its derivative is used to model the metabolic rate, effectively depicting the drug's true dynamic performance during the treatment cycle. The dynamic assessment capability of the TME can predict potential toxic and side effects caused by excessive metabolic load or overly potent therapeutic effects early in the efficacy evaluation process, reducing the probability of side effects caused by blind dosing and incorrect drug combinations. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of the structure of a system for optimizing medical disease treatment plans based on machine learning according to the present invention;

[0047] Figure 2 This is a flow chart of a method for optimizing medical disease treatment plans based on machine learning according to the present invention;

[0048] Figure 3 This is a flow chart of the method for constructing variants of the treatment regimen provided by the present invention. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] Example 1

[0051] See also Figure 1 and Figure 3 As shown, this embodiment 1 further illustrates a system for optimizing medical disease treatment plans based on machine learning proposed by the present invention, including:

[0052] In the treatment of internal medicine diseases, due to individual differences among patients and the complexity of the disease, traditional treatment methods often rely on experience and rules, lacking dynamic optimization and scientific analysis of individualized data, resulting in unstable treatment effects, large side effects, and insufficient efficacy evaluation. Especially in the optimization of treatment plans, existing technologies mostly rely on simple rule models or empirical formulas, making it difficult to achieve accurate and efficient treatment plan adjustments in practical applications. With the rapid development of artificial intelligence and machine learning technologies, treatment plan optimization in the medical field has begun to develop in the direction of data-driven, personalized, and precise. However, existing treatment plan optimization methods based on machine learning still face several significant problems:

[0053] Medical data comes from a wide range of sources, including electronic medical records, medical images, patient physiological data, and feedback data. These data not only come from different sources but also exhibit heterogeneity in structure, format, and protocol. Effectively integrating this heterogeneous data and extracting valuable information has always been a technical challenge.

[0054] Traditional treatment plans are often based on past experience and standardized models. However, differences in patient constitutions, individualized disease manifestations, and treatment responses require that treatment plans be dynamically adjusted in real time to address varying therapeutic effects and side effects. Within this process, how to precisely adjust treatment plans and improve their effectiveness remains a pressing challenge.

[0055] Most existing treatment optimization methods lack effective dynamic feedback mechanisms, making it impossible to adjust treatment plans in a timely manner based on real-time patient responses (such as side effects and changes in treatment efficacy). This makes risk management and individualized adjustments during treatment difficult, and can easily lead to unstable efficacy or exacerbated side effects.

[0056] Most current treatment optimization methods fail to fully consider drug metabolism and its relationship to efficacy. In particular, during the dynamic adjustment of treatment plans, there is a lack of quantitative assessment of drug metabolic efficiency and efficacy. A key issue in treatment optimization is how to integrate pharmacokinetic and pharmacodynamic characteristics for a comprehensive assessment and improve the accuracy of treatment plans.

[0057] In order to effectively solve the above problems, the present invention proposes a system for optimizing medical disease treatment plans based on machine learning, comprising:

[0058] The heterogeneous data monitoring module integrates with multi-source heterogeneous medical databases through standard medical API interfaces to extract and obtain heterogeneous data on internal medicine diseases;

[0059] The immune response network module mimics the antigen recognition response pattern and constructs a treatment plan antigen representation matrix based on heterogeneous data of internal medicine diseases. It uses transfer learning to identify abnormal treatment patterns and eliminates ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment patterns, thereby obtaining historically effective treatment plans. It uses a neural symbol memory mechanism to store historically effective treatment plan fragments and constructs treatment plan variants based on an adversarial generation mechanism.

[0060] The pathological metabolic cycle module extracts key features of the treatment plan antigen representation matrix, draws on the energy conversion mechanism of the tricarboxylic acid cycle, and calculates the drug's metabolic rate and therapeutic effect intensity based on these key features to obtain the drug's therapeutic metabolic equivalent. This is then combined with treatment plan variants to optimize the treatment plan through a dynamic feedback control mechanism.

[0061] The negative feedback recombination module decomposes historical ineffective treatment plans or treatment plans with side effects into recombinable knowledge components, inputs them into the internal medicine disease treatment monitoring terminal, and continuously optimizes to obtain the optimal treatment plan.

[0062] Methods for obtaining heterogeneous data on internal medicine diseases include:

[0063] Multi-source heterogeneous medical databases include electronic medical record platforms (EMR / EHR platforms), medical imaging archive management platforms (PACS platforms), and wearable device data platforms. These databases originate from different medical institutions, health platforms, or IoT devices, and are heterogeneous in terms of data structure, storage format, and communication protocols. Standard medical API interfaces, such as HL7, FHIR, and DICOM, are used to establish communication connections with these multi-source heterogeneous medical databases. These interfaces are used to retrieve data, extract heterogeneous data on internal medicine diseases, and normalize this data. Heterogeneous data on internal medicine diseases includes electronic medical record data, patient physiological data, medical imaging data, and patient feedback data. Electronic medical record data includes disease diagnosis data, medication records, medical history data, and allergy history data. Patient physiological data includes vital sign monitoring data, metabolic index data, lifestyle behavior data, and chronic disease monitoring data. Medical imaging data includes image type data, structured image information, and image report text. Patient feedback data includes treatment effect feedback, side effect feedback, and health status change data.

[0064] Methods for identifying anomalous treatment patterns through transfer learning include:

[0065] The researchers used a multi-source heterogeneous medical database to obtain historical treatment plans for each internal medicine disease. Each historical treatment plan was defined as an antigen, similar to how the human immune system identifies and responds to foreign substances. These historical treatment plans included effective treatment plans and ineffective treatment plans or treatment plans that produced side effects. Effective treatment plans were defined as friendly antigens, while ineffective treatment plans or treatment plans that produced side effects were defined as harmful antigens. By analogy, a treatment plan antigen representation matrix was constructed.

[0066] Wavelet transform is used to extract features from electronic medical record data, patient physiological data, medical imaging data, and patient feedback data in heterogeneous data of internal medicine diseases, and then the features are spliced to obtain the structural feature vector of the unified representation space. The structural feature vector is used as the representation structure of the antigen to construct the treatment plan antigen representation matrix. The rows of the treatment plan antigen representation matrix represent each treatment plan antigen; the columns of the treatment plan antigen representation matrix represent the structural feature vector dimensions of the treatment plan in the unified representation space.

[0067] Historical treatment plans are set as the source domain, and current treatment plans are set as the target domain. A transfer learning model consisting of a shared feature extractor and a classifier subnetwork is used. The model's training samples include historical treatment plans and the probability of them being friendly or harmful antigens. The shared feature extractor is used to extract latent semantic features from the antigen representation vector, and the classifier subnetwork is used to determine whether the current treatment plan is an anomaly (i.e., whether it is a harmful antigen). The source domain refers to a dataset with complete labels, a large number of samples, and clear therapeutic effects. In this case, it refers to historically proven effective and safe treatment plans with clear therapeutic results (e.g., high cure rates, low side effects, and stable efficacy). The target domain refers to treatment plan data with incomplete or uncertain labels. The model aims to use the knowledge learned from the source domain to determine whether these target domain treatment plans are abnormal or risky (i.e., whether they are "harmful antigens"). To use an analogy, it is like a doctor who first studies a large number of "successful classic treatment cases" (the source domain) and then faces some unknown or difficult cases (the target domain). Based on experience and comparison, they can identify which new treatment plans may have potential problems.

[0068] After training is completed, the current treatment plan in the target domain is input into the transfer learning model, and the probability of the historical treatment plan being a friendly antigen or a harmful antigen is output; in the immune response network module, after the transfer learning model training is completed, the treatment pattern to be identified in the target domain is input into the model, and the probability of the historical treatment plan being a friendly antigen or a harmful antigen is output. Traditional methods usually use a preset fixed harmful antigen probability threshold to determine whether the treatment plan is abnormal. However, in actual applications, fixed thresholds may lead to a high misjudgment rate or an increased missed detection rate, affecting the accuracy and safety of the system. In order to solve the above problems, the following dynamic optimization mechanism is introduced to adaptively adjust the harmful antigen probability threshold to improve the accuracy of abnormal treatment pattern recognition;

[0069] Preset harmful antigen probability threshold Among them, θ0 represents the preset basic harmful antigen probability threshold; α represents the adjustment factor, which determines the impact of feedback information (completeness score) on the threshold adjustment, similar to the hyperparameters of the model in machine learning (such as learning rate, regularization factor, etc.). According to the expert experience method, α∈(0,1]; represents the integrity score of the current training sample; b represents the number of antigens in each treatment regimen; m represents the number of dimensions of the structural feature vector of the treatment regimen in the unified representation space;

[0070] It should be noted that when the data quality is high, the model can more easily identify abnormal and effective treatment plans. The ratio of the number of antigens to the dimensions of the representation space is used to measure the adequacy or representativeness of the data. The quality of the sample is determined by calculating the ratio of the number of antigens to the dimensions of the representation space; this method is similar to measuring the effectiveness of the sample through information entropy or information gain in information theory. A higher ratio means that the sample contains more information or more feature dimensions, so the recognition accuracy can be improved by lowering the threshold.

[0071] If the sample contains more information (such as more antigens or higher feature dimensions), the model will be more sensitive to the treatment plan and can allow a lower threshold to increase the probability of detecting abnormalities. This is similar to the learning rate adjustment in deep learning. As the model gradually converges, the learning rate will gradually decrease. Similarly, the threshold can also be dynamically adjusted according to the quality of the training samples. Many intelligent systems use feedback mechanisms to dynamically adjust the judgment criteria. For example, in a control system, as the feedback increases, the system will adjust the parameters in real time to optimize the control effect. In this formula, feedback is reflected in the impact of the integrity score. The more feedback information, the lower the threshold, and the model's ability to identify abnormal treatment plans is enhanced.

[0072] The beneficial effects compared to the prior art are:

[0073] By introducing a completeness score to dynamically adjust the threshold, the judgment criteria can be flexibly adjusted according to the quality of different data, thereby improving the accuracy of detection, especially when facing uneven data quality. By evaluating the completeness of the sample, the model can make targeted adjustments to low-quality data and high-quality data, improving the reliability and stability of the model on data of different quality levels.

[0074] The transfer learning model based on dynamic threshold adjustment can flexibly adjust the judgment criteria for abnormal patterns according to the characteristics of different patients and the actual effects of the treatment plan, and optimize the evaluation process of the treatment plan in real time; the model can automatically adjust the evaluation criteria based on the patient's individualized data (such as historical treatment effects, feedback data, etc.), making the treatment plan more personalized and targeted, thereby improving the accuracy of the treatment effect; through the adjustable threshold mechanism and integrity score, doctors can understand how the model makes treatment plan evaluations based on sample quality in specific situations, thereby enhancing the transparency and trust of the model.

[0075] When the probability of harmful antigens is greater than or equal to the preset harmful antigen probability threshold, it is judged that the current treatment plan is abnormal. All abnormal treatment modes are identified through the transfer learning model, and the ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment modes are eliminated to obtain effective treatment plans.

[0076] Methods for storing historical effective treatment plan fragments include:

[0077] Historically effective treatment plans are obtained from multi-source heterogeneous medical databases and stored in segments using a neural symbolic memory mechanism. Historically effective treatment plans contain u-dimensional treatment process data. Treatment process data includes electronic medical record data, drug combinations, drug dosages, routes of administration, treatment cycles, and clinical monitoring indicators.

[0078] Discretize the treatment process into N time points and then construct a multidimensional time series Among them, t' k is the kth discrete time point; v k is the treatment state vector at the kth discrete time point; for each discrete time point, all treatment process data are collected to construct the treatment state vector v k ; k is the index of the scattered time point, k = 1, 2, ..., N; u is the dimension of the treatment process data; R is the real number domain;

[0079] Perform standard deviation normalization on the multidimensional time series, perform change point detection on the multidimensional time series, use cosine similarity to calculate the difference in the treatment state vectors of two adjacent discrete time points in the multidimensional time series, and obtain the treatment state jump value The preset treatment state transition threshold δ is satisfied. The k+1th discrete time point is regarded as a change point, and all change points are combined to obtain a change point sequence; the multidimensional time series is segmented according to the change point sequence, and a neural encoder is constructed to perform neural encoding on each segment;

[0080] Construct a neural encoder, take the multidimensional time series data contained in each segment as input, and output the corresponding high-dimensional embedding vector, which contains the temporal features and key information within the corresponding segment; preset a set of symbol annotation rules, use vocabulary encoding according to the preset symbol annotation rule set to semantically classify each segment, obtain the corresponding symbol label, encode the symbol label through a word embedding model, convert it into a vector representation, and then obtain the symbol embedding vector;

[0081] The multidimensional time series data contained in any segment is used as the value of the memory unit, the high-dimensional embedding vector corresponding to the segment is used as the key of the memory unit, and the symbol embedding vector corresponding to the segment is used as the annotation of the memory unit to construct a set of key-value-annotation memory pairs; the key-value-annotation memory pairs of all segments are collected and stored in the historical effective treatment plan memory library to store historical effective treatment plan fragments.

[0082] Methods for constructing treatment regimen variants include;

[0083] Construct and train a generative adversarial network, which includes a generator and a discriminator. The generator accepts random noise and generates a false effective treatment plan, which is input into the discriminator. The discriminator compares the false effective treatment plan with the memory bank of historical effective treatment plans and outputs a true or false discrimination result. If the discrimination result is false, a false effective treatment plan is regenerated and compared again through the discriminator. The iteration is repeated until the discrimination result is true, thereby obtaining a trained generative adversarial network, and outputting a treatment plan variant through the trained generative adversarial network.

[0084] Key characteristics include the pharmacokinetic and pharmacodynamic characteristics of the treatment regimen; the pharmacokinetic characteristics of the treatment regimen include the efficiency of the drug entering the blood circulation, the time required for the drug concentration in the patient's body to drop to half, the distribution range of the drug in the patient's body, and the rate of drug elimination from the patient's body; the pharmacodynamic characteristics include the maximum therapeutic effect of the drug (maximum efficacy), the blood concentration required for the drug to produce half-maximal effect, and the intensity of the drug's effect on different diseases.

[0085] Methods for obtaining therapeutic metabolic equivalents of drugs include:

[0086] The tricarboxylic acid cycle is a key metabolic pathway in cellular energy metabolism, primarily converting organic molecules (such as glucose) into high-energy compounds (such as ATP) through a series of chemical reactions, providing energy for cells. In drug metabolism, this analogy can be applied to how drugs are converted by enzyme systems in the body through different metabolic pathways, releasing "energy" or producing pharmacological effects, thereby affecting therapeutic outcomes.

[0087] The therapeutic metabolic equivalent of a drug is analogized to the energy conversion mechanism in the tricarboxylic acid cycle, and the drug metabolism process is regarded as a dynamic process similar to the tricarboxylic acid cycle. The dynamic process of the tricarboxylic acid cycle represents the various stages from the drug entering the body to the drug's efficacy, including absorption, distribution, metabolism, and excretion. It is assumed that each step of drug metabolism contributes different amounts of energy. The total energy released during the treatment process is estimated by the drug's metabolic rate and the intensity of the therapeutic effect, thereby obtaining the drug's therapeutic metabolic equivalent.

[0088] The metabolic rate of the drug is described by the kinetic rate equation of drug concentration changing with time. The kinetic rate equation is C(t) = C0·e -λt ; Where C(t) represents the blood drug concentration at time point t; C0 represents the initial drug concentration; λ represents the elimination rate constant of the drug, which is completed by metabolism and excretion; t represents the time the drug spends in the patient's body;

[0089] The metabolic rate of the drug is obtained by derivatizing C(t). Combining the metabolic rate of the drug and the intensity of the therapeutic effect, the therapeutic metabolic equivalent of the drug is calculated. Among them, E max Indicates the maximum therapeutic effect of the drug; V max Indicates the maximum metabolic rate of the drug; d met Represents the metabolic rate constant of the drug.

[0090] The present invention solves the following technical problems existing in the prior art:

[0091] Traditional treatment optimization methods are often based on rules or empirical models and are unable to quantitatively evaluate the therapeutic metabolic efficiency of different treatment options based on the drug metabolism-effect pathway.

[0092] Existing solutions are unable to abstract the treatment process into a dynamic energy process, and in particular lack an efficacy evaluation system based on a biological energy circulation mechanism.

[0093] Existing technologies usually process pharmacokinetic and pharmacodynamic characteristics separately, which cannot be integrated into treatment plan evaluation indicators, making it difficult for intelligent systems to effectively judge "metabolic efficiency";

[0094] It should be noted that this program draws on the energy conversion mechanism in the tricarboxylic acid cycle and proposes a method for calculating the metabolic equivalent of drug therapy. By integrating the pharmacokinetic parameters and pharmacodynamic parameters of the drug, it comprehensively characterizes the therapeutic energy efficiency ratio of the drug in the metabolic process, providing a quantitative basis for the optimization of the treatment plan.

[0095] In the tricarboxylic acid cycle, metabolic substrates (such as acetyl-CoA) gradually release energy (such as NADH and ATP) through multiple enzymatic reaction stages. The energy release efficiency is reflected as the "total energy produced per unit substrate metabolism." By analogy, the present invention abstracts the metabolic pathway of drugs in the body as a staged dynamic process similar to the TCA cycle, and considers the therapeutic effect as energy output to obtain the therapeutic metabolic equivalent of the drug.

[0096] The beneficial effects compared with the prior art are:

[0097] The introduction of therapeutic metabolic equivalents comprehensively considers the metabolic rate, maximum metabolic capacity, and maximum therapeutic effect of the drug, enabling a more comprehensive assessment of therapeutic energy efficiency. By analogy with the dynamic energy release mechanism of the tricarboxylic acid cycle, the absorption, distribution, metabolism, and excretion processes of the drug in the body are modeled as a continuous energy conversion chain. A first-order pharmacokinetic model is used to describe the temporal evolution of drug concentration, and its derivative is used to model the metabolic rate, thereby effectively depicting the true dynamic performance of the drug during the treatment cycle. By analogy with the tricarboxylic acid cycle, the drug metabolic pathway is introduced with a multi-stage energy release and dynamic efficiency modeling approach, making the evaluation of the treatment process closer to physiological logic and facilitating the subsequent introduction of feedback regulation and self-healing mechanisms. The dynamic assessment capability of the therapeutic metabolic equivalent index can predict potential toxic and side effects caused by excessive metabolic load or excessive efficacy in the early stages of efficacy evaluation, reducing the probability of side effects caused by blind dosing and incorrect combinations.

[0098] Methods for optimizing treatment plans through dynamic feedback control mechanisms include:

[0099] Using the therapeutic metabolic equivalent of the drug as a reward signal, the treatment variants are continuously adjusted through reinforcement learning. Based on patient feedback data, the treatment plan is dynamically optimized based on reinforcement learning, and ultimately the treatment variant that best suits the patient is identified.

[0100] Methods for achieving the best treatment plan include:

[0101] Decompose each historical ineffective treatment regimen or treatment regimen with side effects into L components, including drug type, drug dosage, treatment method, administration route, and treatment duration; each component represents a knowledge component, and each knowledge component contains its impact on treatment efficacy and side effect label information;

[0102] The internal medicine disease treatment monitoring terminal issues a comparison instruction to compare the patient feedback data with the decomposed knowledge components, automatically adjusts and reorganizes the knowledge components in each historical ineffective treatment plan or treatment plan with side effects, thereby forming a new treatment plan. By calculating the Jaccard similarity between the patient feedback data and the new treatment plan, the most matching treatment plan is recommended for each patient, and then the optimal treatment plan for each patient is generated.

[0103] The preset basic harmful antigen probability threshold is set by the staff. By collecting different harmful antigen probabilities, the average value of multiple harmful antigen probabilities is taken as the preset basic harmful antigen probability threshold; similarly, the preset treatment state jump threshold is set.

[0104] This embodiment introduces a completeness score to dynamically adjust the threshold, allowing for flexible adjustments to the judgment criteria based on the quality of different data, thereby improving detection accuracy, especially when faced with uneven data quality. By assessing sample completeness, the model can make targeted adjustments to low-quality and high-quality data, improving the reliability and stability of the model on data of varying quality.

[0105] The transfer learning model based on dynamic threshold adjustment can flexibly adjust the judgment criteria for abnormal patterns according to the characteristics of different patients and the actual effects of the treatment plan, and optimize the evaluation process of the treatment plan in real time; the model can automatically adjust the evaluation criteria based on the patient's individualized data (such as historical treatment effects, feedback data, etc.), making the treatment plan more personalized and targeted, thereby improving the accuracy of the treatment effect; through the adjustable threshold mechanism and integrity score, doctors can understand how the model makes treatment plan evaluations based on sample quality in specific situations, thereby enhancing the transparency and trust of the model.

[0106] The introduction of the therapeutic metabolic equivalent (TME) comprehensively considers a drug's metabolic rate, maximum metabolic capacity, and maximum therapeutic effect, enabling a more comprehensive assessment of therapeutic efficacy. By analogy with the dynamic energy release mechanism of the tricarboxylic acid cycle, the drug's absorption, distribution, metabolism, and excretion processes are modeled as a continuous energy conversion chain. A first-order pharmacokinetic model is used to describe the temporal evolution of drug concentration, and its derivative is used to model the metabolic rate, effectively depicting the drug's true dynamic performance during the treatment cycle. The dynamic assessment capability of the TME can predict potential toxic and side effects caused by excessive metabolic load or overly potent therapeutic effects early in the efficacy evaluation process, reducing the probability of side effects caused by blind dosing and incorrect drug combinations.

[0107] Example 2

[0108] See also Figure 2As shown, for the parts not described in detail in this embodiment, please refer to the description of Example 1. A method for optimizing the treatment plan of internal medicine diseases based on machine learning is provided, comprising:

[0109] S1, heterogeneous data monitoring module, integrates with multi-source heterogeneous medical databases through standard medical API interfaces to extract and obtain heterogeneous data on internal medicine diseases;

[0110] S2: Mimicking the antigen recognition response pattern, constructing a treatment plan antigen representation matrix based on heterogeneous data of internal medicine diseases, identifying abnormal treatment patterns through transfer learning, eliminating ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment patterns, and then obtaining historically effective treatment plans; using a neural symbolic memory mechanism to store historically effective treatment plan fragments, and constructing treatment plan variants based on an adversarial generation mechanism;

[0111] S3. Extract key features of the therapeutic antigen representation matrix. Drawing on the energy conversion mechanism of the tricarboxylic acid cycle, calculate the drug's metabolic rate and therapeutic effect intensity based on these key features to obtain the drug's therapeutic metabolic equivalent. Combine the drug's therapeutic metabolic equivalent with the therapeutic variants and optimize the treatment plan through a dynamic feedback control mechanism.

[0112] S4. Decompose historical ineffective treatment plans or treatment plans with side effects into recombinable knowledge components, input them into the internal medicine disease treatment monitoring terminal, and continuously optimize to obtain the optimal treatment plan.

[0113] Since the electronic device introduced in this embodiment is an electronic device used to implement a system and method for optimizing a treatment plan for internal medicine based on machine learning in the embodiment of this application, based on a system and method for optimizing a treatment plan for internal medicine based on machine learning introduced in the embodiment of this application, a person skilled in the art can understand the specific implementation of the electronic device of this embodiment and its various variations, so how the electronic device implements the method in the embodiment of this application will not be described in detail here. As long as a person skilled in the art implements the electronic device used by a system and method for optimizing a treatment plan for internal medicine based on machine learning in the embodiment of this application, it falls within the scope of protection of this application.

[0114] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0115] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for users of ordinary skill in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A system for optimizing medical disease treatment plans based on machine learning, characterized in that: include: The heterogeneous data monitoring module integrates with multi-source heterogeneous medical databases through standard medical API interfaces to extract and obtain heterogeneous data on internal medicine diseases; The immune response network module mimics the antigen recognition response pattern and constructs a treatment plan antigen representation matrix based on heterogeneous data of internal medicine diseases. It uses transfer learning to identify abnormal treatment patterns and eliminates ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment patterns, thereby obtaining historically effective treatment plans. It uses a neural symbol memory mechanism to store historically effective treatment plan fragments and constructs treatment plan variants based on an adversarial generation mechanism. The pathological metabolic cycle module extracts key features of the treatment plan antigen representation matrix, draws on the energy conversion mechanism of the tricarboxylic acid cycle, and calculates the drug's metabolic rate and therapeutic effect intensity based on these key features to obtain the drug's therapeutic metabolic equivalent. This is then combined with treatment plan variants to optimize the treatment plan through a dynamic feedback control mechanism. The negative feedback recombination module decomposes historical ineffective treatment plans or treatment plans with side effects into recombinable knowledge components, inputs them into the internal medicine disease treatment monitoring terminal, and continuously optimizes to obtain the optimal treatment plan.

2. The internal medicine disease treatment plan optimization system based on machine learning according to claim 1 is characterized in that: The method for acquiring heterogeneous data of internal medicine diseases includes: The multi-source heterogeneous medical database includes an electronic medical record platform, a medical imaging archive management platform, and a wearable device data platform. By supporting the standard medical API interface, a communication connection is established with the multi-source heterogeneous medical database, data is retrieved through the standard medical API interface, heterogeneous data of internal medicine diseases are extracted, and the heterogeneous data of internal medicine diseases are normalized. The heterogeneous data of internal medicine diseases include electronic medical record data, patient physiological data, medical imaging data, and patient feedback data.

3. The internal medicine disease treatment plan optimization system based on machine learning according to claim 2, characterized in that: The method for identifying abnormal treatment patterns through transfer learning includes: The historical treatment plans corresponding to each internal medicine disease are obtained from a multi-source heterogeneous medical database. Each historical treatment plan is defined as an antigen. Historical treatment plans include effective treatment plans and ineffective treatment plans or treatment plans with side effects. Effective treatment plans are defined as friendly antigens, while ineffective treatment plans or treatment plans with side effects are defined as harmful antigens. A treatment plan antigen representation matrix is constructed through analogy. Wavelet transform is used to extract features from electronic medical record data, patient physiological data, medical imaging data, and patient feedback data in heterogeneous data of internal medicine diseases, and then the features are spliced to obtain the structural feature vector of the unified representation space. The structural feature vector is used as the representation structure of the antigen to construct the treatment plan antigen representation matrix. The rows of the treatment plan antigen representation matrix represent each treatment plan antigen; the columns of the treatment plan antigen representation matrix represent the structural feature vector dimensions of the treatment plan in the unified representation space. The historical treatment plan is set as the source domain, and the current treatment plan is set as the target domain. A transfer learning model consisting of a shared feature extractor and a classifier sub-network is used. The model's training samples include historical treatment plans and the probability that the historical treatment plan is a friendly antigen or a harmful antigen. The corresponding shared feature extractor is used to extract latent semantic features from the antigen representation vector, and the classifier sub-network is used to determine whether the current treatment plan is an abnormal pattern. After training is completed, the current treatment plan in the target domain is input into the transfer learning model, and the probability of the historical treatment plan being a friendly antigen or a harmful antigen is output; the harmful antigen probability threshold is preset Among them, θ0 represents the preset basic harmful antigen probability threshold; α represents the adjustment factor; represents the integrity score of the current training sample; b represents the number of antigens in each treatment regimen; m represents the number of dimensions of the structural feature vector of the treatment regimen in the unified representation space; When the probability of harmful antigens is greater than or equal to the preset harmful antigen probability threshold, it is judged that the current treatment plan is abnormal. All abnormal treatment modes are identified through the transfer learning model, and the ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment modes are eliminated to obtain effective treatment plans.

4. The internal medicine disease treatment plan optimization system based on machine learning according to claim 3 is characterized in that: The method for storing historical effective treatment plan fragments includes: Obtain historical effective treatment plans from multi-source heterogeneous medical databases and store them in segments using a neural symbolic memory mechanism; historical effective treatment plans contain u-dimensional treatment process data; discretize the treatment process into N time points, and then construct a multidimensional time series Among them, t' k is the kth discrete time point; v k is the treatment state vector at the kth discrete time point; for each discrete time point, all treatment process data are collected to construct the treatment state vector v k ; k is the index of the scattered time point, k = 1, 2, ..., N; u is the dimension of the treatment process data; R is the real number domain; Perform standard deviation normalization on the multidimensional time series, perform change point detection on the multidimensional time series, use cosine similarity to calculate the difference in the treatment state vectors of two adjacent discrete time points in the multidimensional time series, and obtain the treatment state jump value The preset treatment state transition threshold δ is satisfied. The k+1th discrete time point is regarded as a change point, and all change points are combined to obtain a change point sequence; the multidimensional time series is segmented according to the change point sequence, and a neural encoder is constructed to perform neural encoding on each segment; Construct a neural encoder, take the multidimensional time series data contained in each segment as input to the neural encoder, output the corresponding high-dimensional embedding vector, preset a set of symbol annotation rules, use vocabulary encoding to semantically classify each segment according to the preset symbol annotation rule set, obtain the corresponding symbol label, encode the symbol label through the word embedding model, convert it into a vector representation, and then obtain the symbol embedding vector; The multidimensional time series data contained in any segment is used as the value of the memory unit, the high-dimensional embedding vector corresponding to the segment is used as the key of the memory unit, and the symbol embedding vector corresponding to the segment is used as the annotation of the memory unit to construct a set of key-value-annotation memory pairs; the key-value-annotation memory pairs of all segments are collected and stored in the historical effective treatment plan memory library to store historical effective treatment plan fragments.

5. The system for optimizing medical disease treatment plans based on machine learning according to claim 4, characterized in that: The method for constructing the treatment regimen variant includes: Construct and train a generative adversarial network, which includes a generator and a discriminator. The generator accepts random noise and generates a false effective treatment plan, which is input into the discriminator. The discriminator compares the false effective treatment plan with the memory bank of historical effective treatment plans and outputs a true or false discrimination result. If the discrimination result is false, a false effective treatment plan is regenerated and compared again through the discriminator. The iteration is repeated until the discrimination result is true, thereby obtaining a trained generative adversarial network, and outputting a treatment plan variant through the trained generative adversarial network.

6. The internal medicine disease treatment plan optimization system based on machine learning according to claim 5, characterized in that: The key characteristics include the pharmacokinetic and pharmacodynamic characteristics of the therapeutic regimen; The pharmacokinetic characteristics of a treatment regimen include the efficiency of the drug entering the blood circulation, the time required for the drug concentration in the patient's body to drop to half, the distribution range of the drug in the patient's body, and the rate of drug elimination from the patient's body; the pharmacodynamic characteristics include the maximum therapeutic effect of the drug, the blood concentration required for the drug to produce half-maximal effect, and the intensity of the drug's effect on different diseases.

7. The system for optimizing medical disease treatment plans based on machine learning according to claim 6, characterized in that: The method for obtaining the drug therapeutic metabolic equivalent comprises: The therapeutic metabolic equivalent of a drug is analogized to the energy conversion mechanism in the tricarboxylic acid cycle, and the drug metabolism process is regarded as a dynamic process similar to the tricarboxylic acid cycle. The dynamic process of the tricarboxylic acid cycle represents the various stages from the drug entering the body to the drug's efficacy, including absorption, distribution, metabolism, and excretion. It is assumed that each step of drug metabolism contributes different amounts of energy. The total energy released during the treatment process is estimated by the drug's metabolic rate and the intensity of the therapeutic effect, thereby obtaining the drug's therapeutic metabolic equivalent. The metabolic rate of the drug is described by the kinetic rate equation of drug concentration changing with time. The kinetic rate equation is C(t) = C0·e -λt ; Where C(t) represents the blood drug concentration at time point t; C0 represents the initial drug concentration; λ represents the elimination rate constant of the drug; t represents the time the drug has passed in the patient's body; The metabolic rate of the drug is obtained by derivatizing C(t). Combining the metabolic rate of the drug and the intensity of the therapeutic effect, the therapeutic metabolic equivalent of the drug is calculated. Among them, E max Indicates the maximum therapeutic effect of the drug; V max Indicates the maximum metabolic rate of the drug; d met Represents the metabolic rate constant of the drug.

8. The system for optimizing medical disease treatment plans based on machine learning according to claim 7, characterized in that: The method for optimizing the treatment plan through a dynamic feedback control mechanism includes: Using the therapeutic metabolic equivalent of the drug as a reward signal, the treatment variants are continuously adjusted through reinforcement learning. Based on patient feedback data, the treatment plan is dynamically optimized based on reinforcement learning, and ultimately the treatment variant that best suits the patient is identified.

9. The system for optimizing medical disease treatment plans based on machine learning according to claim 8, characterized in that: The method for obtaining the optimal treatment plan includes: Decompose each historical ineffective treatment regimen or treatment regimen with side effects into L components, including drug type, drug dosage, treatment method, administration route, and treatment duration; each component represents a knowledge component, and each knowledge component contains its impact on treatment efficacy and side effect label information; The internal medicine disease treatment monitoring terminal issues a comparison instruction to compare the patient feedback data with the decomposed knowledge components, automatically adjusts and reorganizes the knowledge components in each historical ineffective treatment plan or treatment plan with side effects, thereby forming a new treatment plan. By calculating the Jaccard similarity between the patient feedback data and the new treatment plan, the most matching treatment plan is recommended for each patient, and then the optimal treatment plan for each patient is generated.

10. A method for optimizing treatment plans for internal medicine diseases based on machine learning, used to implement a system for optimizing treatment plans for internal medicine diseases based on machine learning according to any one of claims 1 to 9, characterized in that: include: S1, heterogeneous data monitoring module, integrates with multi-source heterogeneous medical databases through standard medical API interfaces to extract and obtain heterogeneous data on internal medicine diseases; S2: Mimicking the antigen recognition response pattern, constructing a treatment plan antigen representation matrix based on heterogeneous data of internal medicine diseases, identifying abnormal treatment patterns through transfer learning, eliminating ineffective treatment plans or treatment plans with side effects corresponding to all abnormal treatment patterns, and then obtaining historically effective treatment plans; using a neural symbolic memory mechanism to store historically effective treatment plan fragments, and constructing treatment plan variants based on an adversarial generation mechanism; S3. Extract key features of the therapeutic antigen representation matrix. Drawing on the energy conversion mechanism of the tricarboxylic acid cycle, calculate the drug's metabolic rate and therapeutic effect intensity based on these key features to obtain the drug's therapeutic metabolic equivalent. Combine the drug's therapeutic metabolic equivalent with the therapeutic variants and optimize the treatment plan through a dynamic feedback control mechanism. S4. Decompose historical ineffective treatment plans or treatment plans with side effects into recombinable knowledge components, input them into the internal medicine disease treatment monitoring terminal, and continuously optimize to obtain the optimal treatment plan.

Images