Individualized administration dosage optimization method based on deep learning

Abstract

The invention discloses an individualized administration dosage optimization method based on deep learning, and the method comprises the following steps: collecting multi-source medical data of a patient, and carrying out the preprocessing; modeling an association relationship between the individual features of the patient and the drug dose response based on the graph structure, and generating a preliminary dose prediction vector; performing time sequence feature extraction on the blood concentration feature vector and the physiological sign feature vector to form a self-organizing critical feature vector; calculating a comprehensive criticality score value based on the self-organizing criticality feature vector, dividing a drug dose reaction critical interval, and generating a critical state detection vector; and based on the critical state detection vector, correcting and updating the initial dose prediction vector, and generating a dynamic optimization dose recommendation result. According to the method, improved Grapher is adopted for modeling multi-source medical features, individualized administration dosage optimization is achieved, and the method has the advantages of being high in accuracy, high in safety and good in adaptability.

Description

Technical Field

[0001] This invention relates to the field of personalized medicine, and more particularly to a method for optimizing personalized drug delivery based on deep learning. Background Technology

[0002] Existing methods for optimizing individualized drug dosing largely rely on pharmacokinetic models based on population averages or traditional statistical methods. These calculations typically assume homogeneity in patient population characteristics, failing to adequately reflect individual differences among patients caused by variations in blood drug concentrations, physiological characteristics, and laboratory test results. Furthermore, traditional methods have significant limitations in handling the temporal fluctuations and nonlinear characteristics of drug dose-response, failing to effectively capture complex dose-response dynamics.

[0003] While some studies have attempted to introduce machine learning methods for predictive modeling of multi-source medical data, most methods remain at the level of static feature modeling. They lack the ability to model the power law distribution characteristics and fractal dimension characteristics of the drug dose-response process, and have failed to establish a mechanism to identify critical states in real time. This results in significant deficiencies in the prediction results regarding the identification of insufficient efficacy, overdose margins, and toxicity risks. Therefore, current technologies are still unable to achieve high-precision, dynamic, and adaptive individualized drug dosage optimization. Summary of the Invention

[0004] One objective of this invention is to propose a method for optimizing individualized drug administration based on deep learning. This invention uses an improved Graphormer model to model multi-source medical features to achieve individualized drug administration optimization, which has the advantages of high accuracy, high safety and good adaptability.

[0005] A method for optimizing individualized drug administration based on deep learning according to an embodiment of the present invention includes the following steps: Collect multi-source medical data from patients and preprocess it to form standardized medical feature vectors; Standardized medical feature vectors are input into an improved Graphormer network, which models the relationship between individual patient characteristics and drug dose response based on graph structure, generating a preliminary dose prediction vector. Temporal features were extracted from the blood drug concentration feature vector and the physiological sign feature vector to obtain the fluctuation features, power law distribution features and fractal dimension features in the drug dose response process, forming a self-organized critical feature vector; The comprehensive criticality score is calculated based on the self-organized criticality feature vector, the critical interval of drug dose response is divided, and a critical state detection vector is generated. Based on the critical state detection vector, the preliminary dose prediction vector is corrected and updated to generate dynamically optimized dose recommendation results.

[0006] Optionally, the multi-source medical data includes blood drug concentration data, physiological sign data, and laboratory test data. The preprocessing includes missing value imputation, outlier detection, noise filtering, time alignment, and normalization. The standardized medical feature vector includes blood drug concentration feature vector, physiological sign feature vector, and laboratory test feature vector.

[0007] Optionally, the generation of the preliminary dose prediction vector specifically includes: Standardized medical feature vectors are input into an improved Graphormer network; The improved Graphormer network includes an embedding layer, a graph structure encoding layer, a critical sensitivity attention layer, an individual dynamic pooling layer, and a prediction layer. In the embedding layer, standardized medical feature vectors are linearly mapped to initial node representations; In the graph structure coding layer, a graph structure adjacency matrix and a dose fluctuation prior modulation matrix are constructed based on the relationship between standardized medical feature vectors. Attention coefficients are calculated using the critically sensitive attention layer based on the graph structure adjacency matrix and the dose fluctuation prior modulation matrix. ; in, Represents a node With nodes No. Attention coefficient of layer This represents the natural exponential function. Indicates the total number of nodes. Represents a node The query matrix, Represents a node The key matrix, Represents a node The key matrix, This indicates the transpose operation. Representing feature dimension, The adjacency matrix of the graph structure is represented at the nodes. With nodes The element value, This indicates that the dose fluctuation prior modulation matrix is ​​at the node. With nodes The element value, The adjacency matrix of the graph structure is represented at the nodes. With nodes The element value, This indicates that the dose fluctuation prior modulation matrix is ​​at the node. With nodes The element value, Represents the modulation coefficients of the adjacency matrix. The modulation coefficients represent the prior modulation matrix of dose fluctuations; The initial node representation is iteratively updated to obtain the final node representation; The final node representation is input into the individual dynamic pooling layer to obtain the graph-level representation vector: ; in, Graph-level vector representation Represents a node In the Layer node representation, Represents a node Individual dynamic pooling weights; The graph-level representation vector is input into the prediction layer, and the initial dose prediction vector is obtained through multilayer perceptron mapping.

[0008] Optionally, the generation of the self-organizing criticality feature vector includes: A time series is constructed based on the patient's blood drug concentration feature vector, the difference between adjacent time points is calculated to obtain the blood drug concentration fluctuation characteristics, and the blood drug concentration fluctuation characteristics are normalized. A time series is constructed based on the patient's physiological sign feature vector. The mean and variance are calculated within a preset time window to obtain the physiological sign fluctuation characteristics, and the physiological sign fluctuation characteristics are normalized. The distribution characteristics of blood drug concentration fluctuation sequence are modeled, the probability distribution of fluctuation amplitude is statistically analyzed, and the attenuation coefficient is calculated based on the power law characteristics to form the power law distribution characteristics. The power law distribution characteristics are then normalized. Based on the fluctuation characteristics of blood drug concentration and physiological signs, a multi-scale coverage method was adopted to count the minimum number of intervals required to cover the fluctuation trajectory at different scales. The fractal dimension feature was calculated based on the logarithmic relationship between scale and number of intervals, and the fractal dimension feature was normalized. The normalized blood drug concentration fluctuation characteristics, physiological sign fluctuation characteristics, power law distribution characteristics, and fractal dimension characteristics are concatenated into vectors to generate a self-organizing criticality feature vector.

[0009] Optionally, the generation of the critical state detection vector includes: The comprehensive criticality score is calculated based on the self-organized criticality feature vector. Specifically, this includes: extracting the values ​​of each dimension of the self-organized criticality feature vector, calculating the arithmetic mean of these values, and obtaining the average value; calculating the difference between the maximum and minimum values ​​of these values, and obtaining the range; normalizing the average value and the range, respectively, to obtain the normalized average value and the normalized range value; and taking the arithmetic mean of the normalized average value and the normalized range value to generate the comprehensive criticality score. The comprehensive criticality score is divided into four ranges: insufficient efficacy range, efficacy threshold range, excessive threshold range, and toxicity threshold range. Upper and lower limits are set for each range. For each interval, the difference between the comprehensive criticality score and the lower limit of the interval and the difference between the comprehensive criticality score and the upper limit of the interval are calculated to obtain the lower limit difference and the upper limit difference. The smaller value between the lower limit difference and the upper limit difference is taken as the final distance value of the interval, and the distance values ​​of insufficient efficacy, efficacy threshold, excessive edge, and toxicity threshold are obtained respectively. Standardize the four distance values ​​to generate probabilities of insufficient efficacy, efficacy threshold, excess margin, and toxicity threshold. The four probability values ​​are concatenated to form a critical state detection vector.

[0010] Optionally, the generation of the dynamically optimized dose recommendation result specifically includes: When the probability value of insufficient efficacy is the largest in the critical state detection vector, the mean value of the blood drug concentration feature vector is calculated to obtain the mean value of blood drug concentration. The difference between the mean value of blood drug concentration and the lower limit of the efficacy threshold is multiplied by the probability value of insufficient efficacy to determine the addition correction parameter. The addition correction parameter is applied to the preliminary dose prediction vector to generate the dose prediction vector with the addition correction. When the probability value of the therapeutic effect threshold in the critical state detection vector is the maximum, the hold correction parameter is set to zero, and the hold correction parameter is applied to the initial dose prediction vector to generate the hold-corrected dose prediction vector. When the excess margin probability value in the critical state detection vector is the largest, the fluctuation amplitude of the heart rate sequence and blood pressure sequence in the physiological sign feature vector is calculated to obtain the heart rate fluctuation amplitude value and blood pressure fluctuation amplitude value. The sum of the two is multiplied by the excess margin probability value to determine the reduction correction parameter. The reduction correction parameter is applied to the preliminary dose prediction vector to generate the dose prediction vector after reduction correction. When the probability value of the toxicity threshold in the critical state detection vector is the largest, the mean value of the toxicity-related index sequence in the laboratory detection feature vector is calculated. The difference between the mean value of the toxicity-related index and the upper limit of the toxicity threshold is multiplied by the probability value of the toxicity threshold to determine the significantly reduced correction parameter. The significantly reduced correction parameter is applied to the preliminary dose prediction vector to generate the significantly reduced and corrected dose prediction vector. The corrected dose prediction vector is used as the result of dynamic optimization dose recommendation.

[0011] The beneficial effects of this invention are: This invention introduces an improved Graphormer network into the personalized drug dosage optimization process. It fully utilizes multi-source medical data, including blood drug concentration feature vectors, physiological sign feature vectors, and laboratory test feature vectors, to establish a graph structure model that reflects individual patient differences. In this model, by constructing a graph structure adjacency matrix and a dose fluctuation prior modulation matrix, it simultaneously models the potential correlations between patient features and the dynamic fluctuations in dose, thus avoiding the shortcomings of traditional population models that cannot capture the complex interactions between individual features. Furthermore, by employing a critically sensitive attention mechanism and an individual dynamic pooling mechanism, it ensures that the weights of different features in the graph structure can be dynamically adjusted according to the patient's real-time state, making the prediction results more closely aligned with individualized needs.

[0012] In terms of drug dose-response time-series modeling, this invention proposes to extract fluctuation features, power law distribution features, and fractal dimension features based on blood drug concentration characteristics and physiological signs to form a self-organized criticality feature vector. Based on a comprehensive criticality score, criticality state detection is achieved. This design can accurately identify whether the patient's dose-response process is at a stage of insufficient efficacy, efficacy threshold, overdose borderline, or toxicity threshold, thus compensating for the lack of criticality state identification capabilities in existing technologies. Furthermore, this invention uses the criticality state detection vector to correct and update the initial dose prediction vector, achieving dynamic optimization of dose recommendations. Compared with traditional static prediction methods, this invention can increase the dose when efficacy is insufficient, reduce the dose in a timely manner when there is an overdose borderline or toxicity risk, and maintain dose stability within the efficacy range, thereby achieving individualized dose recommendations. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a deep learning-based individualized drug administration dosage optimization method proposed in this invention; Figure 2 This is a schematic diagram of the improved Graphormer network structure for a deep learning-based individualized drug delivery optimization method proposed in this invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0015] refer to Figures 1-2 A deep learning-based method for optimizing individualized drug administration dosage includes the following steps: Collect multi-source medical data from patients and preprocess it to form standardized medical feature vectors; Standardized medical feature vectors are input into an improved Graphormer network, which models the relationship between individual patient characteristics and drug dose response based on graph structure, generating a preliminary dose prediction vector. Temporal features were extracted from the blood drug concentration feature vector and the physiological sign feature vector to obtain the fluctuation features, power law distribution features and fractal dimension features in the drug dose response process, forming a self-organized critical feature vector; The comprehensive criticality score is calculated based on the self-organized criticality feature vector, the critical interval of drug dose response is divided, and a critical state detection vector is generated. Based on the critical state detection vector, the preliminary dose prediction vector is corrected and updated to generate dynamically optimized dose recommendation results.

[0016] In this embodiment, the multi-source medical data includes blood drug concentration data, physiological sign data, and laboratory test data. The preprocessing includes missing value imputation, outlier detection, noise filtering, time alignment, and normalization. The blood drug concentration data is a time-varying sequence of target drug concentration measurements in plasma. The physiological sign data includes heart rate, blood pressure, body temperature, and respiratory rate. The laboratory test data includes complete blood count indicators, liver function indicators, and kidney function indicators. The standardized medical feature vector includes a blood drug concentration feature vector, a physiological sign feature vector, and a laboratory test feature vector.

[0017] In this embodiment, the generation of the preliminary dose prediction vector specifically includes: Standardized medical feature vectors are input into an improved Graphormer network; The improved Graphormer network includes an embedding layer, a graph structure encoding layer, a critical sensitivity attention layer, an individual dynamic pooling layer, and a prediction layer. In the embedding layer, standardized medical feature vectors are linearly mapped to initial node representations; In the graph structure coding layer, based on the relationship between standardized medical feature vectors, a graph structure adjacency matrix and a dose fluctuation prior modulation matrix are constructed. The graph structure adjacency matrix indicates whether there is a topological connection between nodes, and the dose fluctuation prior modulation matrix indicates the modulation prior between nodes in the modeling of drug dose time-series fluctuations. The generation of the graph structure adjacency matrix specifically includes: using the blood drug concentration feature vector, physiological sign feature vector, and laboratory test feature vector from the standardized medical feature vector as nodes; comparing each node pairwise; calculating the Euclidean distance to reflect the degree of numerical difference; calculating the cosine similarity to reflect directional consistency; taking the reciprocal of the Euclidean distance to obtain the numerical similarity index; averaging it with the cosine similarity to obtain a comprehensive similarity score; judging the comprehensive similarity score against a preset connection threshold; if the comprehensive similarity score is greater than the preset connection threshold, establishing an edge connection between the corresponding nodes and assigning a value of 1 to the adjacency matrix; otherwise, assigning a value of 0 to obtain the graph structure adjacency matrix. The generation of the dose fluctuation prior modulation matrix specifically includes: calculating the numerical difference between adjacent time points based on historical blood drug concentration feature vectors to obtain a dose fluctuation amplitude sequence, which reflects the instantaneous fluctuation intensity of blood drug concentration in the time dimension; calculating the mean and standard deviation of each indicator based on physiological sign feature vectors within a preset time window, and using the ratio of the standard deviation to the mean as the single indicator coefficient of variation value; normalizing and weighting the single indicator coefficient of variation values ​​of all indicators to obtain a comprehensive coefficient of variation value, which characterizes the patient's overall sensitivity to dose perturbations during this time period; calculating the rate of change of values ​​within a continuous detection period based on laboratory test feature vectors to obtain a dose fluctuation trend value, which describes the cumulative fluctuation direction of drug effect in the time dimension; weighting and fusing the dose fluctuation amplitude sequence, comprehensive coefficient of variation value, and dose fluctuation trend value according to preset weight coefficients to obtain a comprehensive fluctuation factor, which characterizes the overall fluctuation characteristics of the patient in the current dose response process; mapping the comprehensive fluctuation factor to a matrix of the same dimension as the graph structure adjacency matrix according to the correspondence between node pairs to form a dose fluctuation prior modulation matrix, which is used to correct the correlation between nodes during the attention weight calculation process; Attention coefficients are calculated using the critically sensitive attention layer based on the graph structure adjacency matrix and the dose fluctuation prior modulation matrix. ; ; in, Represents a node With nodes No. Attention coefficient of layer This represents the natural exponential function. Indicates the total number of nodes. Represents a node The query matrix, Represents a node The key matrix, Represents a node The key matrix, This indicates the transpose operation. Representing feature dimension, The weight matrix represents the query matrix. The weight matrix represents the key matrix. Represents a node In the Layer node representation, Represents a node In the Layer node representation, Represents a node In the Layer node representation, The adjacency matrix of the graph structure is represented at the nodes. With nodes The element value, This indicates that the dose fluctuation prior modulation matrix is ​​at the node. With nodes The element value, The adjacency matrix of the graph structure is represented at the nodes. With nodes The element value, This indicates that the dose fluctuation prior modulation matrix is ​​at the node. With nodes The element value, Represents the modulation coefficients of the adjacency matrix. The modulation coefficients represent the prior modulation matrix of dose fluctuations; The generation of the modulation coefficients of the adjacency matrix and the dose fluctuation prior modulation matrix specifically includes: based on the comprehensive similarity score of each node pair in the graph structure adjacency matrix, calculating the structural normalization value of each node pair according to the normalization processing method to obtain the modulation coefficient of the adjacency matrix, which is used to characterize the relative importance of the feature nodes at the structural level; based on the comprehensive fluctuation factor corresponding to each node pair in the dose fluctuation prior modulation matrix, calculating the fluctuation sensitivity value according to the standardization processing method to obtain the modulation coefficient of the dose fluctuation prior modulation matrix, which is used to characterize the relative effect strength of the feature nodes under dose fluctuation conditions; Iteratively update the initial node representation to obtain the final node representation: ; in, Represents a node In the Layer node representation, Represents a non-linear activation function. Indicates the total number of nodes. Indicates the first The weight matrix of the layer, Indicates the first Attention coefficient of layer Represents a node In the Layer node representation, The weight matrix represents the value matrix; The final node representation is input into the individual dynamic pooling layer to obtain the graph-level representation vector: ; in, Graph-level vector representation Represents a node In the Layer node representation, Represents a node Individual dynamic pooling weights; The generation of the individual dynamic pooling weights specifically includes: calculating the mean and variance of the blood drug concentration feature vector, physiological sign feature vector, and laboratory test feature vector within a preset time window to obtain mean and variance indices, used to characterize the numerical level and fluctuation range of each feature during that time period; normalizing the mean and variance indices to obtain normalized mean and normalized variance values, and taking the average to obtain a single-feature stability index, used to characterize the stability of the feature within the time window; uniformly normalizing all single-feature stability indices to obtain feature stability weight values, used to characterize the relative stability differences of different features among individual patients; and combining the patient's... The latest blood drug concentration feature vector is used to calculate the concentration change rate. Combined with the physiological sign feature vector, the deviation value is calculated. Combined with the laboratory test feature vector, the change rate is calculated to obtain a set of immediate response sub-values, which are used to reflect the patient's multidimensional immediate response level at the current dose. Then, the set of immediate response sub-values ​​is weighted and fused to obtain the individual sensitivity value, which is used to comprehensively characterize the patient's immediate response intensity at the current dose. The feature stability weight value and the individual sensitivity value are averaged to obtain the individual weighting factor. The individual weighting factor is then mapped to the weight allocation of each node in the pooling operation to form the individual dynamic pooling weight, which is used to realize personalized control for individual patients during the feature convergence process. The graph-level representation vector is input into the prediction layer, and a preliminary dose prediction vector is obtained by mapping through a multilayer perceptron. The preliminary dose prediction vector includes four dimensions: the first dimension is the recommended baseline dose value, which represents the initial dosing level before the dose adaptive adjustment mechanism is triggered; the second dimension is the dose response strength value, which characterizes the fit between the current individual characteristics and the dose; the third dimension is the prediction confidence value, which represents the reliability of the model at the recommended baseline dose; and the fourth dimension is the individual difference correction value, which reflects the adjustment results of the patient characteristic relationship under the combined effect of the graph structure adjacency matrix and the dose fluctuation prior modulation matrix during the modeling process.

[0018] In this embodiment, the generation of the self-organizing criticality feature vector includes: A time series is constructed based on the patient's blood drug concentration feature vector. The difference between adjacent time points is calculated to obtain the blood drug concentration fluctuation characteristics. The blood drug concentration fluctuation characteristics are then normalized to reflect the dynamic change intensity of the drug during the continuous dosing cycle. A time series is constructed based on the patient's physiological sign feature vector. The mean and variance are calculated within a preset time window to obtain the physiological sign fluctuation characteristics. The physiological sign fluctuation characteristics are then normalized to characterize the dynamic response of the patient's body under the action of drugs. The distribution characteristics of the blood drug concentration fluctuation feature sequence are modeled, the probability distribution of the fluctuation amplitude is statistically analyzed, and the attenuation coefficient is calculated based on the power law feature to form a power law distribution feature. The power law distribution feature is then normalized to characterize the sudden response mode that may be triggered by changes in drug dosage. The power law feature is based on the amplitude statistical distribution of the blood drug concentration fluctuation feature sequence. By fitting the correspondence between the fluctuation amplitude and the probability of occurrence with a power law distribution, the attenuation trend description parameter is obtained and used to characterize the power law law of the fluctuation distribution during the drug dosage response process. Based on the fluctuation characteristics of blood drug concentration and physiological signs, a multi-scale coverage approach is adopted. The minimum number of intervals required to cover the fluctuation trajectory at different scales is statistically analyzed. The fractal dimension is calculated based on the logarithmic relationship between scale and the number of intervals, and then normalized to characterize the multi-scale complexity of drug metabolism. Specifically, this involves: dividing the fluctuation trajectory into intervals at a preset maximum scale and calculating the number of intervals required to cover the complete fluctuation trajectory at that scale, obtaining the maximum scale coverage value; progressively reducing the scale and calculating the number of intervals required to cover the fluctuation trajectory at each scale, obtaining a coverage sequence at different scales; performing a logarithmic transformation between the coverage sequence and the corresponding scale size to obtain a scale logarithmic sequence and a coverage logarithmic sequence; linearly fitting the scale logarithmic sequence and the coverage logarithmic sequence to obtain the fitting slope value, which characterizes the complexity of the fluctuation trajectory at multiple scales; and using the fitting slope value as the fractal dimension feature to describe the multi-scale complexity of the blood drug concentration fluctuation feature sequence and the physiological sign fluctuation feature sequence. ; in, Indicates the fractal dimension characteristic. Indicates taking the limit, Indicated in scale The minimum number of intervals required to cover the lower cover fluctuation trajectory; The normalized blood drug concentration fluctuation characteristics, physiological sign fluctuation characteristics, power law distribution characteristics, and fractal dimension characteristics are concatenated into vectors to generate a self-organizing criticality feature vector.

[0019] In this embodiment, the generation of the critical state detection vector includes: The comprehensive criticality score is calculated based on the self-organized criticality feature vector. Specifically, this includes: extracting the values ​​of each dimension of the self-organized criticality feature vector, calculating the arithmetic mean of these values, and obtaining the average value; calculating the difference between the maximum and minimum values ​​of these values, and obtaining the range; normalizing the average value and the range, respectively, to obtain the normalized average value and the normalized range value; and taking the arithmetic mean of the normalized average value and the normalized range value to generate the comprehensive criticality score, which is used to characterize the overall stability and risk level of the drug dose-response process. The comprehensive criticality score is divided into four ranges: insufficient efficacy interval, efficacy threshold interval, excessive dosage border interval, and toxicity threshold interval. Upper and lower limits are set for each interval. Specifically, based on the patient's comprehensive criticality score sequence, the minimum and maximum scores are extracted to form a score range. Within this score range, the range is divided into segments according to a preset ratio to generate the insufficient efficacy interval, efficacy threshold interval, excessive dosage border interval, and toxicity threshold interval in sequence, and upper and lower limits are set for each interval. For each interval, the difference between the comprehensive criticality score and the lower limit of the interval and the difference between the comprehensive criticality score and the upper limit of the interval are calculated to obtain the lower limit difference and the upper limit difference. The smaller value between the lower limit difference and the upper limit difference is taken as the final distance value of the interval, and the distance values ​​of insufficient efficacy, efficacy threshold, excessive edge, and toxicity threshold are obtained respectively. Standardize the four distance values ​​to generate a probability value for insufficient efficacy, a probability value for efficacy threshold, a probability value for excessive margin, and a probability value for toxicity threshold, such that the sum of the four probability values ​​equals one. The four probability values ​​are concatenated to form a critical state detection vector.

[0020] In this embodiment, the generation of the dynamically optimized dose recommendation result specifically includes: When the probability value of insufficient efficacy is the largest in the critical state detection vector, the mean value of the blood drug concentration feature vector is calculated to obtain the mean value of blood drug concentration. The difference between the mean value of blood drug concentration and the lower limit of the efficacy threshold is multiplied by the probability value of insufficient efficacy to determine the addition correction parameter. The addition correction parameter is applied to the preliminary dose prediction vector to generate the dose prediction vector with the addition correction. When the probability value of the therapeutic effect threshold in the critical state detection vector is the maximum, the hold correction parameter is set to zero, and the hold correction parameter is applied to the initial dose prediction vector to generate the hold-corrected dose prediction vector. When the excess margin probability value in the critical state detection vector is the largest, the fluctuation amplitude of the heart rate sequence and blood pressure sequence in the physiological sign feature vector is calculated to obtain the heart rate fluctuation amplitude value and blood pressure fluctuation amplitude value. The sum of the two is multiplied by the excess margin probability value to determine the reduction correction parameter. The reduction correction parameter is applied to the preliminary dose prediction vector to generate the dose prediction vector after reduction correction. When the probability value of the toxicity threshold in the critical state detection vector is the largest, the mean value of the toxicity-related index sequence in the laboratory detection feature vector is calculated. The toxicity-related index sequence is the detection items closely related to drug toxicity in the laboratory detection feature vector, including liver function indicators and kidney function indicators. The difference between the mean value of the toxicity-related index and the upper limit of the toxicity threshold is multiplied by the probability value of the toxicity threshold to determine the significantly reduced correction parameter. The significantly reduced correction parameter is applied to the preliminary dose prediction vector to generate the significantly reduced and corrected dose prediction vector. The corrected dose prediction vector is used as the result of dynamic optimization dose recommendation, and the output is consistent with the judgment state of the critical state detection vector.

[0021] Example 1: To verify the feasibility of this invention in practice, it was applied to the clinical medication management of a tertiary hospital. This hospital's oncology department has long faced the challenge of individualized chemotherapy dosage determination. Previously, clinicians mainly relied on traditional pharmacokinetic formulas and empirical reference values ​​for dosage recommendations. However, patients' blood drug concentration fluctuations, physiological responses, and laboratory test indicators often vary significantly. This method can easily lead to insufficient dosage resulting in poor efficacy in some patients, or excessive dosage causing toxic side effects. Especially in complex cases, when a patient's drug metabolism is affected by multiple factors, a single model struggles to capture potential critical changes, and dosage recommendations often lag behind the patient's actual condition, posing potential risks.

[0022] In clinical trials at the hospital's oncology department, this invention collected multi-source medical data from patients throughout the entire treatment cycle, including continuous monitoring of blood drug concentration over time, dynamic changes in physiological signs such as heart rate and blood pressure, and multidimensional data from laboratory tests. In practice, these data first underwent preprocessing to remove obvious outliers and missing values, and normalization was used to ensure that features of different dimensions could be compared on the same scale. Subsequently, the improved Graphormer network used the processed features as node inputs to automatically construct graph structure relationships between patient features. Based on graph structure encoding, the system further introduced a dose fluctuation prior modulation matrix to characterize the patient's drug efficacy fluctuations over time, ensuring that it can dynamically reflect the individual's sensitivity to dose changes.

[0023] In the actual clinical environment of this hospital, the system no longer simply outputs a static value for each patient's dosage prediction. Instead, it updates in real time through a criticality-sensitive attention mechanism. Especially when there are sudden fluctuations in the patient's blood drug concentration and physiological signs, the system can promptly identify the critical state and generate a critical state detection vector to determine whether the patient is in a state of insufficient efficacy, efficacy threshold, borderline overdose, or toxicity risk. When the determination indicates that the patient is close to the toxicity risk, the system automatically triggers an adaptive adjustment mechanism to reduce the dosage, generating a dynamically optimized dosage recommendation plan to avoid toxic reactions caused by overdose. When the patient is in the state of insufficient efficacy, the system will increase the dosage within the safety boundary to ensure that the drug can exert its due therapeutic effect.

[0024] In follow-up of clinical medication use in oncology, the recommendations provided by the physician feedback system have significantly improved the accuracy of individualized regulation. In the past, under traditional dosage recommendations, physicians needed to frequently adjust the dosage based on experience, often only making passive adjustments after side effects or insufficient efficacy occurred. The method of this invention can predict risks in advance and provide a basis for adjustment, enabling physicians to manage the treatment process more proactively.

[0025] To verify the performance of the present invention in practice, it was compared with traditional methods, and the results are shown in Table 1.

[0026] Table 1. Comparison of Individualized Dosing Optimization Methods and Traditional Methods

[0027] As can be seen from the data in Table 1, the individualized dose optimization method based on improved Graphormer network and self-organized criticality detection proposed in this invention is superior to traditional methods in terms of efficacy, safety and dynamic adjustment capability.

[0028] Regarding efficacy-related indicators, the average effective rate of the method of this invention reached 85.4%, which is 12.9 percentage points higher than that of the traditional method. The incidence of insufficient efficacy was significantly reduced, indicating that the invention can make timely dose corrections when patients approach the efficacy threshold, effectively avoiding the situation of insufficient efficacy. At the same time, the time for maintaining stable efficacy was extended by 16 days, further demonstrating the advantages of the dynamic adjustment mechanism in improving the durability of treatment.

[0029] In terms of safety, the method of the present invention reduces the incidence of drug toxicity from 15.7% to 6.3%, the rate of serious adverse reactions is reduced by more than half, and the recovery time of adverse reactions is shortened by 3.8 days. This shows that by using the critical state detection vector for toxicity risk identification and dose correction, it is possible to reduce the side effects caused by overdose and improve the overall safety of medication.

[0030] Regarding dynamic adjustment capabilities, the average number of dose corrections per treatment course in this invention has been reduced from 4.3 to 2.1, and the correction response time has been shortened from 48 hours to 12 hours. This result shows that the system can identify problems and make adjustments more quickly, reducing the risk window for patients waiting for manual intervention. Especially in the case of acute toxicity or sudden decline in efficacy, the rapid response capability reflects the clinical value of this invention.

[0031] In terms of prediction and control capabilities, the accuracy of critical state early warning increased from 61.4% to 84.2%, and the mean confidence level of dose recommendation increased from 0.71 to 0.86. This means that the method of the present invention can not only improve the reliability of prediction results, but also enhance doctors' reliance on the system's recommendation results and reduce the workload of repeated manual adjustments.

[0032] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing individualized drug administration based on deep learning, characterized in that, Includes the following steps: Collect multi-source medical data from patients and preprocess it to form standardized medical feature vectors; Standardized medical feature vectors are input into an improved Graphormer network, which models the relationship between individual patient characteristics and drug dose response based on graph structure, generating a preliminary dose prediction vector. Temporal features were extracted from the blood drug concentration feature vector and the physiological sign feature vector to obtain the fluctuation features, power law distribution features and fractal dimension features in the drug dose response process, forming a self-organized critical feature vector; The comprehensive criticality score is calculated based on the self-organized criticality feature vector, the critical interval of drug dose response is divided, and a critical state detection vector is generated. Based on the critical state detection vector, the preliminary dose prediction vector is corrected and updated to generate dynamically optimized dose recommendation results.

2. The method for optimizing individualized drug administration based on deep learning according to claim 1, characterized in that, The multi-source medical data includes blood drug concentration data, physiological sign data, and laboratory test data. The preprocessing includes missing value imputation, outlier detection, noise filtering, time alignment, and normalization. The standardized medical feature vectors include blood drug concentration feature vectors, physiological sign feature vectors, and laboratory test feature vectors.

3. The method for optimizing individualized drug administration based on deep learning according to claim 1, characterized in that, The generation of the preliminary dose prediction vector specifically includes: Standardized medical feature vectors are input into an improved Graphormer network; The improved Graphormer network includes an embedding layer, a graph structure encoding layer, a critical sensitivity attention layer, an individual dynamic pooling layer, and a prediction layer. In the embedding layer, standardized medical feature vectors are linearly mapped to initial node representations; In the graph structure coding layer, a graph structure adjacency matrix and a dose fluctuation prior modulation matrix are constructed based on the relationship between standardized medical feature vectors. Attention coefficients are calculated using the critically sensitive attention layer based on the graph structure adjacency matrix and the dose fluctuation prior modulation matrix. ； in, Represents a node With nodes No. Attention coefficient of layer This represents the natural exponential function. Indicates the total number of nodes. Represents a node The query matrix, Represents a node The key matrix, Represents a node The key matrix, This indicates the transpose operation. Representing feature dimension, The adjacency matrix of the graph structure is represented at the nodes. With nodes The element value, This indicates that the prior modulation matrix of dose fluctuation is at the node. With nodes The element value, The adjacency matrix of the graph structure is represented at the nodes. With nodes The element value, This indicates that the prior modulation matrix of dose fluctuation is at the node. With nodes The element value, Represents the modulation coefficients of the adjacency matrix. The modulation coefficients represent the prior modulation matrix of dose fluctuations; The initial node representation is iteratively updated to obtain the final node representation; The final node representation is input into the individual dynamic pooling layer to obtain the graph-level representation vector: ； in, Graph-level vector representation Represents a node In the Layer node representation, Represents a node Individual dynamic pooling weights; The graph-level representation vector is input into the prediction layer, and the initial dose prediction vector is obtained through multilayer perceptron mapping.

4. The method for optimizing individualized drug administration based on deep learning according to claim 1, characterized in that, The generation of the self-organizing criticality feature vector includes: A time series is constructed based on the patient's blood drug concentration feature vector, the difference between adjacent time points is calculated to obtain the blood drug concentration fluctuation characteristics, and the blood drug concentration fluctuation characteristics are normalized. A time series is constructed based on the patient's physiological sign feature vector. The mean and variance are calculated within a preset time window to obtain the physiological sign fluctuation characteristics, and the physiological sign fluctuation characteristics are normalized. The distribution characteristics of blood drug concentration fluctuation sequence are modeled, the probability distribution of fluctuation amplitude is statistically analyzed, and the attenuation coefficient is calculated based on the power law characteristics to form the power law distribution characteristics. The power law distribution characteristics are then normalized. Based on the fluctuation characteristics of blood drug concentration and physiological signs, a multi-scale coverage method was adopted to count the minimum number of intervals required to cover the fluctuation trajectory at different scales. The fractal dimension feature was calculated based on the logarithmic relationship between scale and number of intervals, and the fractal dimension feature was normalized. The normalized blood drug concentration fluctuation characteristics, physiological sign fluctuation characteristics, power law distribution characteristics, and fractal dimension characteristics are concatenated into vectors to generate a self-organizing criticality feature vector.

5. The method for optimizing individualized drug administration based on deep learning according to claim 1, characterized in that, The generation of the critical state detection vector includes: The comprehensive criticality score is calculated based on the self-organized criticality feature vector. Specifically, this includes: extracting the values ​​of each dimension of the self-organized criticality feature vector, calculating the arithmetic mean of these values, and obtaining the average value; calculating the difference between the maximum and minimum values ​​of these values, and obtaining the range; normalizing the average value and the range, respectively, to obtain the normalized average value and the normalized range value; and taking the arithmetic mean of the normalized average value and the normalized range value to generate the comprehensive criticality score. The comprehensive criticality score is divided into four ranges: insufficient efficacy range, efficacy threshold range, excessive threshold range, and toxicity threshold range. Upper and lower limits are set for each range. For each interval, the difference between the comprehensive criticality score and the lower limit of the interval and the difference between the comprehensive criticality score and the upper limit of the interval are calculated to obtain the lower limit difference and the upper limit difference. The smaller value between the lower limit difference and the upper limit difference is taken as the final distance value of the interval, and the distance values ​​of insufficient efficacy, efficacy threshold, excessive edge, and toxicity threshold are obtained respectively. Standardize the four distance values ​​to generate probabilities of insufficient efficacy, efficacy threshold, excess margin, and toxicity threshold. The four probability values ​​are concatenated to form a critical state detection vector.

6. The method for optimizing individualized drug administration based on deep learning according to claim 1, characterized in that, The generation of the dynamically optimized dose recommendation results specifically includes: When the probability value of insufficient efficacy is the largest in the critical state detection vector, the mean value of the blood drug concentration feature vector is calculated to obtain the mean value of blood drug concentration. The difference between the mean value of blood drug concentration and the lower limit of the efficacy threshold is multiplied by the probability value of insufficient efficacy to determine the addition correction parameter. The addition correction parameter is applied to the preliminary dose prediction vector to generate the dose prediction vector with the addition correction. When the probability value of the therapeutic effect threshold in the critical state detection vector is the maximum, the hold correction parameter is set to zero, and the hold correction parameter is applied to the initial dose prediction vector to generate the hold-corrected dose prediction vector. When the excess margin probability value in the critical state detection vector is the largest, the fluctuation amplitude of the heart rate sequence and blood pressure sequence in the physiological sign feature vector is calculated to obtain the heart rate fluctuation amplitude value and blood pressure fluctuation amplitude value. The sum of the two is multiplied by the excess margin probability value to determine the reduction correction parameter. The reduction correction parameter is applied to the preliminary dose prediction vector to generate the dose prediction vector after reduction correction. When the probability value of the toxicity threshold in the critical state detection vector is the largest, the mean value of the toxicity-related index sequence in the laboratory detection feature vector is calculated. The difference between the mean value of the toxicity-related index and the upper limit of the toxicity threshold is multiplied by the probability value of the toxicity threshold to determine the significantly reduced correction parameter. The significantly reduced correction parameter is applied to the preliminary dose prediction vector to generate the significantly reduced and corrected dose prediction vector. The corrected dose prediction vector is used as the result of dynamic optimization dose recommendation.

Images