Warfarin dosage prediction and adjustment method and system based on multiple linear regression

Abstract

The invention relates to the technical field of warfarin dose regulation and control, and discloses a warfarin dose prediction and adjustment method and system based on multiple linear regression, and the system comprises a patient data collection module, a gene data acquisition module, a model construction module, a dose adjustment module, and a real-time prediction and adjustment execution module. The patient data acquisition module synchronously acquires multi-dimensional parameters in different treatment states and establishes a treatment data set; the gene data acquisition module takes gene polymorphism and age data, a treatment data set is combined to construct a refined patient feature space, and the model construction module constructs and trains a multiple linear regression prediction model according to the refined patient feature space and extracts dose indexes; the dose adjustment module establishes a mapping relation between dose adjustment levels and dose indexes, and the real-time prediction module obtains real-time treatment parameters and obtains real-time self-adaptive dose indexes in combination with the model; the adjustment execution module obtains a predicted adjustment level according to the mapping relation, and triggers a dose adjustment mechanism if the predicted adjustment level exceeds a preset threshold.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of warfarin dose regulation, in particular to a warfarin dose prediction adjustment method and system based on multiple linear regression. BACKGROUND

[0002] Warfarin is a commonly used oral anticoagulant in clinical practice, widely used in the prevention and treatment of thromboembolic diseases, such as atrial fibrillation, artificial heart valve replacement, etc. In the process of clinical application, the therapeutic window of warfarin is narrow, and the response of different patients to the drug shows significant individual differences. Insufficient dose can easily lead to thrombosis, and excessive dose may cause serious adverse reactions such as bleeding, and even endanger the life of the patient. Therefore, accurate dose adjustment is crucial to ensure treatment effect and drug safety.

[0003] Currently, warfarin dose adjustment in clinical practice mainly relies on the experience of doctors to judge according to the clinical indicators (such as international normalized ratio INR), age, weight, and concurrent medication of patients. However, this traditional method has obvious limitations: the experience of doctors is easily affected by subjective factors, and different doctors may have different dose adjustment plans for the same patient, making it difficult to form a unified and standardized adjustment standard; the traditional method does not fully consider the influence of genetic polymorphism on drug metabolism, and relevant studies have shown that the polymorphism of genes related to warfarin metabolism (such as CYP2C9, VKORC1, etc.) in patients can significantly affect the metabolism rate and efficacy of the drug, and ignoring this factor will reduce the accuracy of dose prediction and increase the risk of medication. With the development of medical information technology and artificial intelligence technology, some studies have attempted to build a warfarin dose prediction method based on data models, but existing models still have many shortcomings. Some models only rely on a single type of data (such as clinical indicators or genetic data) to build a prediction model, and fail to utilize multi-dimensional data, resulting in incomplete characterization of individual differences in patients and limited prediction accuracy; other models consider multi-dimensional data, but lack real-time data updating and dynamic adjustment mechanisms, and cannot optimize the dose plan in a timely manner according to the real-time changes in the patient's condition (such as INR value fluctuations, concurrent medication adjustments, etc.), making it difficult to meet the needs of dynamic treatment in clinical practice. In addition, most existing models only output predicted dose values, without establishing a clear dose adjustment level classification and corresponding trigger mechanism, making it difficult for doctors to quickly determine whether the dose needs to be adjusted and the adjustment range in actual application, reducing the clinical applicability and operability of the model. SUMMARY

[0004] The purpose of the present application is to provide a warfarin dose prediction adjustment method and system based on multiple linear regression to solve the problems raised in the background.

[0005] To achieve the above object, the application provides a warfarin dose prediction adjustment system based on multiple linear regression, which comprises: a patient data acquisition module for synchronously acquiring multi-dimensional parameter data of a patient in different treatment states and constructing a patient treatment data set; a gene data acquisition module for acquiring gene polymorphism data and patient age data of the patient, constructing a concise patient feature space based on the patient treatment data set, the gene polymorphism data and the patient age data; a model construction module for constructing and training a warfarin dose prediction model based on the concise patient feature space and extracting a warfarin dose index; a dose adjustment module for establishing a warfarin dose adjustment level and constructing a mapping relationship between the warfarin dose index and the warfarin dose adjustment level; a real-time prediction module for acquiring real-time treatment parameters of the patient and obtaining a real-time adaptive warfarin dose index according to the real-time treatment parameters of the patient and the warfarin dose prediction model; an adjustment execution module for obtaining a predicted dose adjustment level according to the real-time adaptive warfarin dose index and the mapping relationship between the warfarin dose index and the dose adjustment level, and judging whether the predicted dose adjustment level exceeds a preset threshold, and triggering a dose adjustment mechanism if the predicted dose adjustment level exceeds the preset threshold.

[0006] Preferably, the multi-dimensional parameter data comprises international normalized ratio data and patient weight feature data. The concise patient feature space is constructed based on the patient treatment data set, the gene polymorphism data and the patient age data. A treatment response feature matrix is obtained based on the patient treatment data set. A gene age feature matrix is obtained based on the gene polymorphism data and the patient age data. The treatment response feature matrix and the gene age feature matrix are fused to construct the concise patient feature space.

[0007] Preferably, the treatment response feature matrix is obtained based on the patient treatment data set, which comprises: The patient treatment data set is subjected to a change rate calculation to extract multi-parameter change rate data, wherein the multi-parameter change rate data comprises international normalized ratio change rate and patient weight feature change rate. The treatment response feature matrix is obtained based on the multi-parameter change rate data.

[0008] Preferably, the treatment response feature matrix is obtained based on the multi-parameter change rate data, which comprises: Based on the international normalized ratio change rate, calculate the statistical characteristics of the international normalized ratio data in different time periods to obtain an international normalized ratio feature vector; Based on the patient weight feature change rate, calculate the statistical characteristics of the patient weight feature data in different time periods to obtain a weight feature vector; Splice the international normalized ratio feature vector and the weight feature vector by column to form the treatment response feature matrix.

[0009] Preferably, the obtaining of the gene age feature matrix based on the gene polymorphism data and the patient age data comprises: Encoding the gene polymorphism data to obtain gene encoding data; Extracting features from the gene encoding data and the patient age data to obtain a gene age feature vector; Constructing the gene age feature matrix from the gene age feature vector.

[0010] Preferably, the constructing of the essence patient feature space by fusing the treatment response feature matrix and the gene age feature matrix comprises: Performing data standardization processing on the treatment response feature matrix and the gene age feature matrix; Splicing the standardized treatment response feature matrix and the gene age feature matrix by column to form a multi-dimensional patient feature matrix; Performing feature selection on the multi-dimensional patient feature matrix; Performing feature extraction on the multi-dimensional patient feature matrix after feature selection to obtain the essence patient feature space.

[0011] Preferably, the constructing and training of the warfarin dose prediction model based on the essence patient feature space comprises: Taking patient features in the essence patient feature space as sample features and taking artificially labeled warfarin dose indicators as sample labels to construct a training sample set; Using a multivariate linear regression model to construct an initial warfarin dose prediction model; Using the training sample set to train the initial warfarin dose prediction model to obtain weight coefficients and bias terms of each patient feature; Based on the weight coefficients and bias terms of each patient feature, obtaining a final warfarin dose prediction model.

[0012] Preferably, the dose adjustment module establishes a warfarin dose adjustment level and constructs a mapping relationship between the warfarin dose indicator and the warfarin dose adjustment level, comprising: Defining a plurality of warfarin dose adjustment levels; Based on historical data, a correspondence between the warfarin dose index and the warfarin dose adjustment level is constructed. The mapping relationship is stored.

[0013] Preferably, the real-time prediction module obtains real-time treatment parameters of the patient, and obtains a real-time adaptive warfarin dose index according to the real-time treatment parameters of the patient and the warfarin dose prediction model, which comprises: The real-time treatment parameters of the patient are obtained, and a mapping relationship between the real-time treatment parameters and a weight adjustment coefficient is constructed, wherein the construction of the weight adjustment coefficient mapping relationship relies on a clinical rule knowledge base, and contains weight correction rules in twenty-seven clinical scenarios, each rule is converted into a triple structure, including a trigger condition, an influence feature and an adjustment coefficient, and is stored in a hash mapping table of an in-memory database, supporting millisecond-level rule matching. According to the real-time treatment parameters of the patient and the mapping relationship, a real-time weight adjustment coefficient is calculated. According to the real-time weight adjustment coefficient, the weight coefficient of each patient feature in the warfarin dose prediction model is adjusted to obtain an adjusted weight coefficient. The adjusted weight coefficient is used to calculate the real-time adaptive warfarin dose index under the current working condition.

[0014] Preferably, the present application further comprises a warfarin dose prediction adjustment method based on multiple linear regression, which comprises all the modules and method processes of the warfarin dose prediction adjustment system based on multiple linear regression as described above.

[0015] Compared with the prior art, the present application has the following beneficial effects: In the data acquisition and processing link, the patient data acquisition module can synchronously acquire multi-dimensional parameter data of the patient under different treatment states, covering clinical routine indexes, medication conditions and other information, and the genetic data acquisition module further incorporates genetic polymorphism data and age data to construct a refined patient feature space. This multi-dimensional data fusion acquisition method breaks through the limitation of single data type in traditional dose adjustment, not only comprehensively captures the clinical basic information of the patient, but also fully integrates the key factors affecting drug metabolism at the genetic level, so that the constructed patient feature space is more in line with the actual situation of individual differences, lays a data foundation for the accurate prediction of the subsequent model, and makes the dose prediction no longer rely on single-dimensional information, but based on a more comprehensive patient portrait. In terms of model construction and dose prediction, the model construction module constructs and trains a warfarin dose prediction model based on the refined patient feature space using a multiple linear regression algorithm, which can effectively mine the linear correlation between multi-dimensional features and warfarin dose, accurately extract warfarin dose indicators, and significantly improve the objectivity and accuracy of dose prediction compared to traditional empirical judgment and single data model, reducing the interference of subjective factors on dose judgment, and making the dose prediction result more in line with the individual metabolic characteristics and disease needs of the patient. In terms of dose adjustment mechanism design, the dose adjustment module establishes warfarin dose adjustment levels and constructs a mapping relationship between dose indicators and adjustment levels. This hierarchical classification makes the dose adjustment standard clearer and more standardized, allowing doctors to intuitively understand the adjustment direction and amplitude corresponding to different dose indicators, avoiding the problem of inconsistent adjustment standards in traditional empirical judgment. At the same time, the real-time prediction module can obtain real-time treatment parameters of the patient, generate real-time adaptive dose indicators combined with the trained prediction model, realize dynamic tracking of the patient's disease changes, ensure that the dose prediction can follow the disease fluctuations in the patient's treatment process, and break the limitation of the lack of real-time updating mechanism in existing models, so that the dose plan always matches the current treatment status of the patient. In terms of clinical application safety and operability, the adjustment execution module triggers the dose adjustment mechanism by judging whether the predicted dose adjustment level exceeds the preset threshold. This threshold judgment mechanism sets a clear trigger condition for dose adjustment, effectively avoiding unnecessary dose adjustment, reducing the medication risks that may be caused by frequent dose adjustment, and also providing doctors with clear decision-making basis. When the adjustment level exceeds the threshold, the doctor is prompted to take adjustment measures in a timely manner, helping the doctor to make decisions quickly and improving clinical work efficiency. In addition, the entire system realizes full-process automation from data collection, model prediction to dose adjustment suggestion, reducing manual operation links and the probability of human error. At the same time, the functions of each module are clear and logical, and doctors do not need complex operation steps to obtain accurate dose adjustment suggestions, significantly improving the clinical practicality and generalizability of the system, and better adapting to the clinical needs of medical institutions of different levels, providing support for anticoagulant therapy for more patients. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A timing diagram of the warfarin dose prediction and adjustment system based on multiple linear regression described in the present application; Figure 2 A flowchart for constructing a refined patient feature space; Figure 3 A flowchart for calculating real-time adaptive warfarin dose indicators. DETAILED DESCRIPTION

[0017] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0018] With reference to Figure 1 The present application provides a warfarin dose prediction adjustment system based on multiple linear regression, which comprises: through integration of a patient data acquisition module, a genetic data acquisition module, a model construction module, a dose adjustment module, a real-time prediction module and an adjustment execution module, precise prediction and dynamic adjustment of warfarin dose are realized.

[0019] The patient data acquisition module synchronously acquires multi-dimensional parameter data of the patient under different treatment states, including international normalized ratio data and patient weight characteristic data, and constructs a patient treatment data set. The genetic data acquisition module acquires genetic polymorphism data and patient age data of the patient, and combines the patient treatment data set to construct a refined patient characteristic space. The model construction module constructs and trains a warfarin dose prediction model based on the refined patient characteristic space by using a multiple linear regression method, and extracts a warfarin dose index. The dose adjustment module establishes a warfarin dose adjustment level, and constructs a mapping relationship between the warfarin dose index and the dose adjustment level. The real-time prediction module acquires real-time treatment parameters of the patient, combines the warfarin dose prediction model, and calculates a real-time adaptive warfarin dose index. The adjustment execution module obtains a predicted dose adjustment level according to the real-time adaptive warfarin dose index and the mapping relationship, judges whether the predicted dose adjustment level exceeds a preset threshold, and triggers a dose adjustment mechanism. The system realizes personalized and adaptive management of warfarin dose through cooperation of multiple modules.

[0020] The patient data acquisition module is connected with the electronic medical record system through the medical monitoring device, and the dynamic parameters of the patient in the whole cycle of anticoagulant therapy are captured in real time. The international normalized ratio data is collected by a coagulation function detector every 24 hours, and the patient weight characteristic data is automatically recorded by an intelligent bed sensor every morning. All data is transmitted to the central database through encryption to form a time-stamped treatment data set. The gene data acquisition module calls the original report of the gene sequencing laboratory, extracts the polymorphism data of CYP2C92, CYP2C93 and VKORC1-1639G>A loci, and directly calls the birth date from the identity information system to calculate the actual age of the patient. The original gene data is subjected to a quality control process to exclude samples with insufficient sequencing depth. When constructing the treatment response feature matrix, the system performs time series slicing processing on the treatment data set, calculates the sliding window change rate of the international normalized ratio with a time window of 72 hours, and specifically uses the relative change percentage formula of the detection values at adjacent time points. The change rate of the patient's weight characteristics is calculated according to the daily morning measurement value, and the calculation process automatically excludes abnormal values of instantaneous weight fluctuations caused by treatment such as infusion. The generation of the international normalized ratio feature vector includes four statistical dimensions: the mean value is calculated in eight consecutive time windows to reflect the overall trend, the variance represents the fluctuation degree, the maximum value marks the risk peak, and the minimum value identifies the low coagulation risk point. The weight characteristic vector synchronously calculates the same four types of statistics in the same time window, and after data alignment, the two vectors are horizontally spliced into a 24-column treatment response feature matrix. The construction of the gene age feature matrix first performs function coding on the gene polymorphism data. The CYP2C9 2 locus is encoded as a binary variable using the dominant model, and the VKORC1 genotype is converted into a continuous variable according to the haplotype dose. The missing data of all gene loci is filled with the high-frequency genotype of this population. The patient age data is processed into three biological age interval features, and the coded gene variables and age interval variables are input into the feature engineering layer. Through principal component analysis, 8-dimensional latent variables representing genetic metabolic capacity are extracted, and finally a gene age feature vector containing 11 features is formed. After sample alignment, the gene age feature matrix is expanded. The feature fusion stage adopts a two-level processing architecture. The first level standardizes the treatment response feature matrix and the gene age feature matrix respectively. The treatment response feature uses the RobustScaler method to eliminate the influence of outliers, and the gene age feature uses the MinMaxScaler method to compress to the [0, 1] interval. The standardized double matrix is spliced by column through feature ID matching to generate a 35-dimensional original feature space. In the second level processing, the feature selection module uses random forest importance evaluation to retain the top 20 features in the importance score, and then compresses the features to a 12-dimensional concise feature space through linear discriminant analysis. The conversion parameters of each feature are automatically recorded when the feature space is established, providing the basis for inverse transformation in real-time prediction.

[0021] In the clinical implementation scenario, the dynamic updating mechanism of the treatment response feature matrix is set to trigger calculation automatically every 6 hours, the gene age feature matrix is only generated as static data when the patient first enrolls, the feature fusion process adopts incremental learning mode, and only the relevant feature blocks are updated when new treatment data is input. The system establishes an independent version management of the feature space for each patient, retains the historical version for tracing function every time the feature is updated, and stores the dimension information and feature explanation of the feature space in the metadata database for clinical personnel to query and verify the feature generation logic. The calculation of the international normalized ratio change rate adopts a double verification mechanism, the original detection value is first reviewed by the laboratory information system, and the change rate calculation result is compared with the clinical nursing record again. When the difference exceeds 5%, an artificial review process is triggered. The weight feature change rate calculation integrates bed sensor data and daily manual record data, the system automatically identifies the deviation of the two data sources and adopts a weighted average value, and the statistical validity check is set in the feature vector generation process. When the effective data points in the time window are less than 70%, the time window range is automatically expanded. The clinical explainability conversion rule is set in the gene coding link. When the CYP2C9 genotype is converted into the metabolic capacity score, the international warfarin pharmacogenomics consortium standard is adopted, the dose influence coefficient of the VKORC1 genotype is quoted from the PharmGKB database, and the age segmentation processing adopts a nonlinear binning algorithm. When divided by 10 years, the segmentation boundary is automatically adjusted to balance the sample distribution. The feature extraction stage retains the gene loading matrix of each principal component, so that the reduced features can still be traced back to the original gene site contribution. The feature selection process adopts a dynamic threshold mechanism, and the feature importance threshold is adjusted adaptively according to the size of the training set. When the sample size is greater than 500, the top 25% features are retained, and in the small sample scenario, the features with importance greater than the average value are retained. The linear discriminant analysis model used in feature extraction updates the projection matrix regularly, re-trains the projection model every 200 new samples, and automatically outputs the feature correlation heat map after generating the essence feature space, helping clinical personnel to understand the interaction between features.

[0022] The entire implementation process is embedded with a data quality monitoring module, which detects the rationality of the value in real time when the treatment data set is collected, and immediately alarms when the international normalized ratio is out of the range of [0.5, 10]. When the weight changes by more than 5 kg in a single day, data review is triggered. A sequencing quality filter is set in the gene data acquisition stage, and sites with allele detection confidence less than 99% are marked as low-quality data. An abnormal value interception mechanism is provided in each link of feature construction, and the process is suspended and manual intervention is prompted when the feature value deviates from the group distribution by 3 standard deviations. The system establishes a feature reconstruction log, which records 287 conversion parameters from raw data to essence features, meeting the traceability requirements of medical AI systems.

[0023] The construction of the treatment response feature matrix starts from the patient treatment data set loading link. The system automatically identifies the time series structure in the data set. The international standardization ratio data is sorted according to the detection time stamp to form a continuous observation sequence. The patient weight feature data is established according to the measurement time and the ratio data to establish the time alignment relationship. The change rate calculation module uses a sliding window mechanism to process the time series. The international standardization ratio change rate is calculated based on the absolute difference percentage of the adjacent two detection values. The weight feature change rate takes the deviation degree of the daily morning measurement value and the average value of the previous seven days. The calculation process embeds an anomaly detection algorithm to automatically filter non-physiological weight fluctuations caused by postoperative edema or dehydration treatment. The generation of the international standardization ratio feature vector defines four statistical dimensions. The system performs parallel calculation in the preset eight consecutive time windows: the first dimension calculates the arithmetic mean of the detection values in each window, reflecting the overall level of anticoagulation effect; the second dimension calculates the variance value, capturing the fluctuation characteristics of the treatment response; the third dimension marks the peak value in the window, identifying the high-risk or bleeding risk critical point; and the fourth dimension records the minimum value, monitoring the potential period of insufficient anticoagulation. The weight feature vector synchronously performs four-dimensional statistical calculation. The mean value of the weight change in the same time window reflects the trend, the variance reveals the metabolic stability, the maximum value warns of the risk of fluid retention, and the minimum value indicates insufficient nutrient intake. After the two feature vectors are matched by the time window identifier, they are combined into a 24-column treatment response feature matrix using a horizontal splicing method. The matrix row index corresponds to the patient's unique identifier. The construction of the gene age feature matrix starts from the gene data quality verification stage. The original gene sequencing report is analyzed by a bioinformatics pipeline. The CYP2C9 2 locus is encoded as a 0 / 1 binary variable using a dominant inheritance model. The VKORC1-1639 genotype is converted to a 0.0-2.0 continuous value according to the haplotype dose effect. Missing gene loci are filled with the high-frequency genotype of the patient's ethnic group. Patient age data is converted into three biological features: metabolic age coefficient is generated by inputting actual age into a nonlinear conversion function, age group is marked according to ten-year intervals, and deviation from standard body weight age is calculated. The encoded gene variables and age features are input into the feature fusion layer. Eight latent variables representing genetic metabolic capacity are extracted through principal component analysis. Finally, a gene age feature vector containing eleven features is formed. The vector is matched and expanded to a gene age feature matrix by the patient ID.

[0024] The dynamic update of the treatment response feature matrix follows the clinical treatment cycle, and the rate of change of the international normalized ratio is recalculated after each new test result is entered. The system retains the last thirty calculation results to form a rolling time series. The weight feature change rate is automatically updated every morning, and a data missing warning is triggered when there is no new weight data for three consecutive days. The feature vector generation module sets statistical validity check rules. When the proportion of valid data points in the time window is less than seventy percent, the window length is automatically extended to one hundred and twenty hours. Robust estimation methods are used for statistical quantity calculation, and Winsorize processing is performed on outliers. Clinical interpretability conversion rules are established for gene data processing. The conversion of CYP2C9 genotype to metabolic capacity score refers to the dose calculation table of the International Warfarin Pharmacogenomics Consortium, and the VKORC1 genotype conversion coefficient refers to the latest pharmacogenomics annotation in the PharmGKB database. Age segmentation uses a dynamic binning algorithm, with the five quintiles of the patient cohort age distribution as the boundary to divide the intervals. After principal component analysis, the gene loading matrix of each component is retained to support clinical staff to reverse analyze the contribution of genetic characteristics.

[0025] The feature matrix construction process implements full-link quality monitoring, logical verification is performed when the international standardization ratio raw value is entered, and the data processing flow is immediately frozen if the value exceeds the medical reasonable range of 0.5-10.0. When body weight data is collected, the sensor value and the nursing record value are compared, and if the difference exceeds 3%, manual review is triggered. The gene data quality control sets three filters: sites with sequencing depth less than 30X are marked as low confidence, when the allele frequency difference with the population database exceeds two standard deviations, sample contamination is prompted, and abnormal sites are automatically isolated by linkage disequilibrium test. The treatment response feature matrix storage adopts a columnar database structure, and each column feature is attached with metadata description: record the time window parameters of statistical quantity calculation, abnormal value processing rules, and data source version. When storing the gene age feature matrix, the gene sequencing report version number and the PharmGKB database release date are associated, and all feature conversion parameters are written into the audit log to support feature reconstruction verification at any time point. The system establishes a version control system for the feature matrix, generates a new matrix version for each feature update, and retains the historical version comparison function. In the real-time computing environment, the treatment response feature matrix construction module is deployed as a distributed computing service, the international standardization ratio change rate calculation task is distributed to multiple computing nodes for parallel processing, and the body weight feature statistical quantity generation adopts a streaming computing architecture. The gene age feature matrix calculation service is configured with an automatic scaling mechanism, which automatically expands computing resources when new patient gene data is imported in batches, and the feature fusion operation is designed in incremental update mode, updating only when new data arrives. The matrix construction process implements a resource isolation mechanism to ensure that the calculation of a single patient does not affect the overall system operation. The feature generation process is embedded with a clinical decision support interface, and the treatment response feature matrix output automatically attaches a trend analysis report: an anti-coagulation instability warning is marked when the international standardization ratio variance increases in three consecutive time windows, and a nutrition status evaluation summary is generated when the body weight feature vector is output. When the gene age feature matrix is generated, a drug interaction check is triggered, and when the CYP2C9 slow metabolism genotype is detected and the patient is also using sulfonamides, a dose adjustment warning mark is output. All feature output values are accompanied by confidence scores, reflecting data quality and calculation reliability. The system establishes a feature traceability mechanism, and each feature value of the treatment response feature matrix can be located back to the time point of the original detection record, and the gene age feature matrix supports backtracking query of the original genotype sequencing data. The intermediate results of feature calculation are stored in a temporary cache area, which is retained for 72 hours for quality review, and a final consistency check is performed before the feature matrix is released: the feature value difference between the current version and the previous version is compared, and when the key feature changes more than 15%, the clinical staff is required to confirm the release.

[0026] The feature fusion process starts from the double matrix standardization preprocessing stage. The standardization of the treatment response feature matrix adopts a robust standardization method based on the median and quartile range, which can effectively resist the interference of outliers. The specific conversion process is to subtract the median of the feature column from each feature value and then divide by the quartile range. The standardization of the gene age feature matrix adopts the maximum and minimum value scaling method, which linearly transforms the original value to a certain interval. The standardized double matrix is aligned by row through the patient identifier, and then spliced horizontally by feature column to generate a merged matrix containing treatment response features and gene age features. The dimension of the matrix is equal to the sum of the feature dimensions of the two original matrices.

[0027] The merged matrix enters the feature selection stage. The system adopts a screening mechanism based on model importance. A random forest regressor is used to fit the merged matrix, and the importance score of each feature for warfarin dose prediction is calculated. The calculation formula of the importance score is: ,

[0028] wherein: represents the feature importance score, is the total number of trees in the random forest, represents the total number of split nodes in the tth tree, represents the impurity reduction amount when the nth node in the tth tree is split, is an indicator function that takes the value of 1 when the feature used for splitting the nth node in the tth tree is f, and 0 otherwise.

[0029] The feature extraction stage uses a linear discriminant analysis method to reduce the dimension of the selected features. This method finds the projection direction that maximizes the ratio of between-class variance to within-class variance. First, calculate the between-class scatter matrix and the within-class scatter matrix , where the classes are divided into three groups according to the patient's warfarin dose response level. Then solve the generalized eigenvalue problem , wherein: is the eigenvector, is the discriminant ability coefficient corresponding to the eigenvector . Take the first k largest eigenvalues corresponding to the eigenvectors to form a projection matrix, which projects the original feature space to a k-dimensional essential feature space. The value of k is determined by cross-validation, usually between five and eight.

[0030] The construction of the warfarin dose prediction model takes the features in the essence feature space as input variables and the warfarin stable dose labeled by clinical experts as the target variable. The training sample set contains the feature data and corresponding dose labels of historical patients. After random shuffling, the sample set is divided into training set and validation set in the ratio of seven to three. The multiple linear regression model uses the least squares method for parameter estimation, and the objective function is the minimization of residual sum of squares: , wherein: is the minimum value of the weight coefficient vector w, is the dose label vector, is the feature matrix, is the weight coefficient vector to be solved. The model training uses gradient descent algorithm to iteratively optimize the weight parameters, the learning rate is set to adaptive adjustment mode, the initial value is 0.01, and every 100 iterations are attenuated by 5%.

[0031] Regularization is implemented during training to prevent overfitting. The L2 regularization term coefficient is selected by cross-validation. The optimal coefficient value is determined by calculating the mean square error on the validation set. After model convergence, the weight coefficient vector and the bias term are output. Each element in the coefficient vector corresponds to the weight of a feature in the essence feature space, and the bias term represents the basic dose level. The final model is stored in a serialized form, including complete information such as weight parameters, feature name mapping table, standardization parameters and projection matrix.

[0032] The update maintenance of the essence feature space adopts a version control mechanism. Whenever new batches of patient data are added, the feature fusion process is re-executed to generate a new version of the feature space, while the old version is preserved for historical data queries. The importance threshold in the feature selection stage is dynamically adjusted according to the sample size. When the sample size is small, a stricter threshold is used to retain fewer features, avoiding the problem of dimension disaster. The model training module is deployed as a scalable distributed computing service. When the training data volume is large, multiple node parallel computing is automatically started. Each computing node processes a shard of data, and finally the gradient updates of each node are aggregated. Real-time monitoring is implemented during the training process, and the training loss and validation loss of each iteration are recorded. When the validation loss does not decrease for ten consecutive iterations, the early stopping mechanism is started. Quality checkpoints are set during the feature fusion process. The distribution of each feature is checked during the standardization stage. For features with severe skew distribution, logarithmic transformation preprocessing is performed. After feature selection, the multicollinearity between retained features is calculated. Features with a variance inflation factor greater than ten are removed. The class separability index is monitored during the feature extraction stage. When the class separability of linear discriminant analysis is lower than the threshold, the feature reorganization process is triggered. The warfarin dose prediction model outputs a feature importance report, listing the top ten features that contribute most to dose prediction and their weight coefficients. Clinicians can understand the basis of the model's decision accordingly. After the model is deployed, a continuous learning mechanism is established. The model is incrementally trained using newly generated patient data every month to keep the model's prediction performance up to date. The system generates a feature contribution analysis report for each patient, showing the specific impact of each feature value on the final dose prediction, helping doctors understand the formation process of individualized dose decision-making. All feature processing parameters and model training hyperparameters are recorded in the metadata database, supporting complete data traceability and algorithm auditing.

[0033] The definition of dose adjustment levels refers to the international anticoagulation guidelines and clinical expert consensus. The system classifies warfarin dose adjustment into five discrete levels: Level 0 means maintaining the current dose (with a change of ±0.5 mg / day), Level 1 represents a small increase (0.5-1 mg per day), Level 2 is a moderate increase (1-2 mg per day), Level 3 refers to a large increase (more than 2 mg per day), and the corresponding reduction levels are defined as Level-1 (0.5-1 mg per day), Level-2 (1-2 mg per day), and Level-3 (more than 2 mg per day). Each level is associated with clinical operation specifications, such as Level 2 adjustment requiring an increase in international normalized ratio monitoring frequency within the next 72 hours. The mapping relationship is established based on the historical patient data warehouse. The system extracts patient records within three years who have completed warfarin treatment, including the final warfarin dose indicator and the actual dose adjustment level performed clinically. In the data preprocessing stage, cases with abnormal dose indicators or incomplete adjustment records are excluded, and the remaining samples are reviewed by anticoagulation specialists to confirm the rationality of their adjustment levels. The analysis uses a density clustering algorithm to identify the natural grouping boundaries of the dose indicator and the adjustment level. Referring to Table 1, the correspondence between the dose indicator interval and the adjustment level is shown.

[0034] Table 1: Mapping table of warfarin dose indicator and adjustment level

[0035] The storage of the mapping relationship adopts a dual-track architecture. The main mapping table is stored in the dose_adjustment_rules table of the relational database, which includes fields such as dose_range_lower (dose lower limit), dose_range_upper (dose upper limit), adjustment_level (adjustment level), and clinical_protocol (clinical operation specification). At the same time, to ensure high-performance queries, the complete mapping dictionary is cached in the in-memory database, and a red-black tree structure is used to achieve O(logn) time complexity interval queries. The system automatically checks the validity of the mapping table every quarter, and triggers the mapping table revision process when the conflict rate between newly accumulated adjustment cases and existing mappings exceeds five percent. The logical implementation of the dose adjustment level includes a multi-level judgment mechanism. First, the warfarin dose indicator is queried to determine the basic adjustment level, and then the level is corrected in combination with the patient's real-time clinical status: when the international normalized ratio is detected to fluctuate sharply, the adjustment level is automatically increased by one level; if the patient has recently shown clinical signs of bleeding tendency, the adjustment level is reduced by one level. The final determined adjustment level is accompanied by a confidence score, reflecting the reliability of the mapping relationship and the consistency degree of the clinical status.

[0036] The storage of clinical operation specifications adopts structured document format, and each adjustment level is associated with a set of operation instruction set: Level 3 adjustment contains twelve operation items such as dose change amplitude, review time requirement, patient education points, etc., and Level 0 maintenance dose contains five operation items such as continue to observe indicators, appointment next detection, etc. These specification contents are regularly updated by the drug therapy committee, and the system records the effective time and version number of each specification clause. The establishment process of the mapping relationship implements quality monitoring, and the sample representativeness check is set in the historical data extraction stage: the sample size of each dose interval is required to reach at least thirty cases, and the gender and age distribution has no significant difference with the total patient population. The profile coefficient threshold is set in the clustering analysis process, and when the clustering quality is lower than 0.7, manual intervention is required to adjust the parameters. The final mapping relationship needs to be audited back-to-back by more than three anticoagulation specialists, and only after reaching a consensus can it be put into system use. The system establishes a version management system for the mapping relationship, generates a new version and keeps the historical version every time the mapping table is updated, and supports clinical personnel to query the dose adjustment rules in a specific time period. When the mapping relationship changes, the affected patient group is automatically marked, and the dose adjustment scheme update suggestion report is generated. All modification operations of the mapping relationship record audit logs, including modification time, operator, modification reason and other complete information.

[0037] In actual clinical application, the output of the dose adjustment module contains structured adjustment suggestions: the adjustment level is clearly marked as "Level2 increase 1-2mg", accompanied by specific clinical operation guidelines: "review international normalized ratio within 48 hours, evaluate patient's dietary changes". The output result is pushed to the doctor workstation and nursing system through the hospital information system interface, and after the doctor confirms, the medical order execution item is automatically generated. The maintenance of the mapping relationship establishes a feedback mechanism, and the system records the actual clinical effect after each dose adjustment is executed, and when the international normalized ratio does not reach the target range within the expected time, the case of poor adjustment effect is automatically marked. Regularly analyze the matching degree of adjustment effect and mapping relationship, and when the poor effect case exceeds ten percent, start the mapping relationship optimization process. The system provides a mapping relationship visualization analysis tool, and clinical personnel can intuitively view the heat distribution map of the adjustment effect of each dose interval. The mapping relationship of special patient groups has exception rules, for patients with renal dysfunction (eGFR<30ml / min), the dose change amplitude of all adjustment levels is automatically reduced by thirty percent; the adjustment level of elderly patients (>80 years old) is automatically lowered by one level to avoid excessive anticoagulation. These exception rules are stored in a separate exception mapping table, which is executed after the application of the basic mapping relationship. The dose adjustment module is integrated with the hospital alarm system, and when the predicted adjustment level reaches Level3 or Level-3, it automatically triggers a pharmacy consultation application, and sends a high-risk drug adjustment warning message to the attending physician. All adjustment decision support information is saved in the patient's medication history, forming a complete warfarin dose adjustment trajectory record, supporting the retrospective analysis and quality improvement of the treatment process.

[0038] The real-time prediction module activates the data listening service when it starts, continuously collects the international normalized ratio real-time stream data transmitted by the patient's bedside monitoring equipment, and the sampling frequency is set to every four hours, while accessing the electronic medical record system to obtain the daily updated body weight record and medication log. When the patient receives blood transfusion or dialysis treatment, the system automatically captures the time point and parameter details of the related treatment event, and these real-time treatment parameters are temporarily stored in the distributed cache area after data cleaning, waiting for the feature conversion engine to process. The construction of the weight adjustment coefficient mapping relationship relies on the clinical rule knowledge base, which is maintained by a team of pharmacology experts and contains twenty-seven weight correction rules in various clinical scenarios, as shown in Table 2.

[0039] Table 2: Core rule table of weight adjustment coefficient mapping relationship

[0040] Each rule is converted into a triple structure (trigger condition, impact feature, adjustment coefficient) and stored in a hash map table in the in-memory database, supporting millisecond-level rule matching. The calculation of real-time weight adjustment coefficients uses a multi-level triggering mechanism. The system scans the real-time treatment parameter queue and initiates a first-level processing flow when an international normalized ratio (INR) abnormal fluctuation is detected: queries the rule mapping table to determine the impact feature dimension, and generates a temporary adjustment coefficient array. If a significant weight change occurs simultaneously, a second-level processing is activated: combines the adjustment coefficients generated by multiple rules to determine the final coefficient value using the maximum logic. Conflict detection is implemented during the calculation process. When different rules propose contradictory adjustments for the same feature, the latest coefficient scheme recommended by the clinical guidelines is prioritized. Dynamic adjustment of weight coefficients is performed before model reasoning. The system loads the weight vector of the base warfarin dose prediction model from the persistent storage, which contains the initial coefficients of twelve feature dimensions. The real-time weight adjustment coefficient array is multiplied element by element with the initial coefficient vector to generate the adjusted weight coefficient vector. The multiplication operation uses a segmented processing strategy. Scalar multiplication is performed for gene-related features, and time-decaying multiplication is performed for treatment response features to ensure that recent data are given higher weights.

[0041] The calculation of real-time adaptive warfarin dose indicators is deployed as a microservices architecture. The feature extraction service converts real-time treatment parameters into a feature vector in the refined feature space. This vector is dot multiplied with the adjusted weight coefficient vector, and the dose prediction value is output after adding the model bias term. Real-time monitoring is implemented during the calculation process. When the feature vector contains missing values, the nearest neighbor feature imputation algorithm is automatically triggered. The imputed feature values are marked with a confidence degradation flag. The implementation process is illustrated using a specific patient as an example. Patient ID0731 had international normalized ratio (INR) monitoring values that rose to 3.8 (target range 2.0-3.0) for two consecutive days on the seventh day after heart valve replacement surgery, and weight records showed a three-day weight loss of 2.3 kg. After the system captured the abnormal parameters in real time, it matched two relevant rules from the rule library: rule R15 (INR exceeding the upper limit triggers a 30% increase in gene feature weight) and rule R22 (sudden weight loss triggers a 20% decrease in age feature weight). The weight adjustment engine adjusted the initial gene feature weight of 0.78 to 1.014 and the age feature weight from 0.65 to 0.52. The feature conversion service converted the patient's current treatment parameters into a feature vector [0.87, 1.02, 0.34, …, 0.56], which was dot multiplied with the adjusted weight vector to output a real-time dose indicator of 4.1 mg / day. The system establishes a traceability mechanism for weight adjustment, recording complete operation logs for each coefficient modification: storing the original coefficient value, adjustment rule number, adjusted coefficient, and operation timestamp. The clinical interface provides a weight impact visualization tool, allowing doctors to view the historical change curve of each feature weight and the triggering reason. When the number of daily weight adjustments exceeds three, the system automatically generates a weight stability report, prompting the examination of the patient's clinical status.

[0042] The real-time computing service implements resource isolation policies, assigning critical patients to dedicated computing containers that guarantee dose prediction response times below 200 milliseconds. General ward patients share a common pool of computing resources, with a priority scheduling algorithm ensuring critical values are processed first. All real-time prediction results are written to read-only database replicas, avoiding data write conflicts in high-concurrency scenarios. The safety control module sets weight adjustment boundary conditions, limiting the variation of any feature dimension's weight coefficient to ±50%, preventing extreme adjustments that could lead to prediction distortion. When adjusted coefficients exceed reasonable intervals, the system automatically switches to a safety mode: freezing the weight adjustment mechanism, reverting to the base model prediction, and simultaneously sending a system alert to clinical engineers.

[0043] Prediction results are output with dual-channel verification, with the main computing channel generating dose indicators, and the auxiliary channel using a simplified model for result reasonableness checking. When the difference between the two-channel results exceeds 15%, a review process is triggered: patient feature data is reloaded, and the full process is recalculated on independent computing nodes. The final output result is accompanied by a quality marker: green for consistent results, yellow for differences that pass review, and red for unreliable results requiring human intervention. The system establishes a real-time prediction quality feedback loop, comparing each dose prediction result with actual administration records, calculating prediction bias values, and storing them in a quality database. When consecutive five prediction biases exceed the allowed range, an automatic weight adjustment rule optimization process is triggered: the clinical applicability of related rules is re-evaluated, and the pharmacology team is recommended to revise the adjustment coefficients. All real-time prediction request records are audited for operational traces, meeting the regulatory compliance requirements of medical AI systems.

[0044] It should be noted that the relational terms, such as first and second, and the like, are used solely to distinguish one from another entity or action, without necessarily requiring or implying any actual such relationship or order between or among such entities or actions. Moreover, the terms "comprises," "comprising," or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0045] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, combinations, and variations can be made by those skilled in the art without departing from the spirit and scope of the present application, which is defined by the appended claims and their equivalents.

Claims

1. A warfarin dose prediction and adjustment system based on multiple linear regression, characterized in that, The system includes: The patient data acquisition module is used to synchronously collect multidimensional parameter data of patients under different treatment states and construct a patient treatment dataset; The gene data acquisition module is used to acquire the patient's gene polymorphism data and patient age data, and to construct a refined patient feature space based on the patient treatment dataset, the gene polymorphism data, and the patient age data. The model building module, based on the refined patient feature space, constructs and trains a warfarin dose prediction model and extracts warfarin dose indicators; The dosage adjustment module establishes warfarin dosage adjustment levels and constructs a mapping relationship between the warfarin dosage index and the warfarin dosage adjustment levels. The real-time prediction module acquires the patient's real-time treatment parameters and obtains a real-time adaptive warfarin dose index based on the patient's real-time treatment parameters and the warfarin dose prediction model. The adjustment execution module obtains the predicted dose adjustment level based on the real-time adaptive warfarin dose index and the mapping relationship between the warfarin dose index and the dose adjustment level; it determines whether the predicted dose adjustment level exceeds a preset threshold, and if it exceeds the preset threshold, it triggers the dose adjustment mechanism. The real-time prediction module acquires the patient's real-time treatment parameters and, based on these parameters and the warfarin dose prediction model, obtains real-time adaptive warfarin dose indices, including: The real-time treatment parameters of the patient are obtained, and a mapping relationship between the real-time treatment parameters and the weight adjustment coefficients is constructed. The construction of the weight adjustment coefficient mapping relationship is based on the clinical rule knowledge base, which contains weight correction rules under twenty-seven clinical scenarios. Each rule is converted into a triplet structure, including triggering conditions, influencing features and adjustment coefficients, and stored in the hash mapping table of the in-memory database, supporting millisecond-level rule matching. Calculate the real-time weight adjustment coefficient based on the patient's real-time treatment parameters and the mapping relationship; Based on the real-time weight adjustment coefficient, the weight coefficient of each patient feature in the warfarin dose prediction model is adjusted to obtain the adjusted weight coefficient. The adjusted weighting coefficients are used to calculate the real-time adaptive warfarin dosage index under the current operating conditions.

2. The warfarin dose prediction and adjustment system based on multiple linear regression according to claim 1, characterized in that, The multidimensional parameter data includes internationally normalized ratio data and patient weight characteristic data; The construction of the refined patient feature space based on the patient treatment dataset, the gene polymorphism data, and the patient age data includes: Based on the patient treatment dataset, a treatment response feature matrix is ​​obtained; Based on the gene polymorphism data and the patient age data, a gene age feature matrix is ​​obtained; The essence patient feature space is constructed by fusing the treatment response feature matrix and the gene age feature matrix.

3. The warfarin dose prediction and adjustment system based on multiple linear regression according to claim 2, characterized in that, The treatment response feature matrix obtained based on the patient treatment dataset includes: The rate of change is calculated on the patient treatment dataset, and multi-parameter rate of change data is extracted; the multi-parameter rate of change data includes the international normalized ratio rate of change and the rate of change of patient weight characteristics. The treatment response feature matrix is ​​obtained based on the multi-parameter change rate data.

4. The warfarin dose prediction and adjustment system based on multiple linear regression according to claim 3, characterized in that, The treatment response feature matrix obtained based on the multi-parameter rate of change data includes: Based on the rate of change of the international normalized ratio, the statistical characteristics of the international normalized ratio data in different time periods are calculated to obtain the international normalized ratio feature vector; Based on the rate of change of the patient's weight characteristics, the statistical characteristics of the patient's weight characteristics data in different time periods are calculated to obtain the weight characteristic vector. The internationally standardized ratio feature vector and the weight feature vector are concatenated column-wise to form the treatment response feature matrix.

5. The warfarin dose prediction and adjustment system based on multiple linear regression according to claim 2, characterized in that, The gene-age feature matrix obtained based on the gene polymorphism data and the patient age data includes: The gene polymorphism data is encoded to obtain gene coding data; Features are extracted from the gene coding data and the patient age data to obtain a gene age feature vector; The gene age feature matrix is ​​constructed from the gene age feature vector.

6. The warfarin dose prediction and adjustment system based on multiple linear regression according to claim 2, characterized in that, The process of fusing the treatment response feature matrix and the gene age feature matrix to construct the refined patient feature space includes: The treatment response feature matrix and the gene age feature matrix are subjected to data standardization processing; The standardized treatment response feature matrix and the gene age feature matrix are concatenated column-wise to form a multidimensional patient feature matrix. Feature selection is performed on the multidimensional patient feature matrix; Feature extraction is performed on the multidimensional patient feature matrix after feature selection to obtain the refined patient feature space.

7. The warfarin dose prediction and adjustment system based on multiple linear regression according to claim 6, characterized in that, The construction and training of the warfarin dose prediction model based on the refined patient feature space includes: A training sample set is constructed using patient features in the refined patient feature space as sample features and manually labeled warfarin dosage indicators as sample labels. An initial warfarin dose prediction model was constructed using a multiple linear regression model. Using the training sample set, the initial warfarin dose prediction model is trained to obtain the weight coefficients and bias terms for each patient feature; The final warfarin dose prediction model is obtained based on the weighting coefficients and bias terms of each patient's characteristics.

8. The warfarin dose prediction and adjustment system based on multiple linear regression according to claim 1, characterized in that, The dosage adjustment module establishes warfarin dosage adjustment levels and constructs a mapping relationship between the warfarin dosage index and the warfarin dosage adjustment levels, including: Define multiple warfarin dose adjustment levels; Based on historical data, a correspondence between the warfarin dosage index and the warfarin dosage adjustment level is constructed; Store the mapping relationship.

9. A warfarin dose prediction and adjustment method based on multiple linear regression, characterized in that, It includes all modules and method flows of the warfarin dose prediction and adjustment system based on multiple linear regression as described in any one of claims 1 to 8.

Images