Multimodal data prediction model for response to diabetes gene therapy
By using a multimodal data prediction model to collect and fuse explicit and implicit data, and using convolutional networks to identify the net effect of drugs on blood glucose, the problem of low accuracy in drug efficacy prediction in existing technologies is solved, and precise medication guidance for gene therapy of diabetes is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SICHUAN TOURISM UNIV
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-07
Smart Images
Figure CN122067701B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of drug efficacy prediction technology, and more specifically, to a multimodal data prediction model for the response to gene therapy for diabetes. Background Technology
[0002] The content in this section provides only background information related to this application and may not constitute prior art.
[0003] Gene therapy for diabetes regulates glucose metabolism by targeting specific gene loci, and is highly individualized. Due to significant differences in patients' genetic backgrounds, the response of different individuals to the same gene-modifying drug is highly uncertain. This uncertainty leads to a lack of precise guidance in clinical medication, and some patients may experience delayed treatment or face the risk of side effects due to ineffective therapy.
[0004] Current methods for predicting diabetes treatment primarily rely on multi-source data analysis. These methods monitor changes in patients' physiological indicators (such as blood glucose, blood pressure, and blood lipids) and lifestyle data after medication use to construct drug efficacy prediction models. For example, Chinese patent CN116994775A proposes a drug efficacy prediction method based on multi-source data. This method uses a machine learning model to directly learn the correlation between patients' lifestyle habits, clinical examination results, and expected drug efficacy, thereby predicting the drug's regulatory effect on target indicators (such as blood glucose).
[0005] The mechanism of blood glucose regulation involves complex interactions among multiple factors, including genetics, metabolism, and environment. The same change in blood glucose may be caused by different factors such as drug effects, dietary adjustments, exercise intervention, or spontaneous fluctuations. Existing models directly use changes in physiological indicators after medication (such as the decrease in blood glucose) as labels for drug efficacy during training, making it difficult for the models to distinguish the contribution of the true drug effect from other confounding factors. This data-driven correlational learning lacks an analysis of causal mechanisms and is prone to incorrectly attributing changes in blood glucose caused by non-drug factors to drug efficacy, thereby reducing the accuracy of prediction results. Summary of the Invention
[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.
[0007] Some embodiments of this application propose a multimodal data prediction model for diabetes gene therapy response to address the technical problems mentioned in the background section above.
[0008] As a first aspect of this application, some embodiments of this application provide a multimodal data prediction model for diabetes gene therapy response, including:
[0009] The medication non-dose data collection module collects data samples of users who have not taken their medication.
[0010] The medication data collection module collects data samples after the user takes medication. The data samples include explicit data and implicit data. Explicit data is body information and medication information related to blood glucose changes, while implicit data is the blood glucose value corresponding to the explicit data.
[0011] The blood glucose prediction module trains the blood glucose prediction model with data samples from users who have not taken medication, and inputs explicit data from data samples from users who have taken medication into the blood glucose prediction model to obtain prediction residuals.
[0012] The data fusion module fuses data samples from users who have not taken medication, data samples from users who have taken medication, and prediction residuals to generate fused features;
[0013] The data prediction module inputs the fused features into the convolutional network model to generate drug efficacy prediction values.
[0014] This application generates a prediction residual corresponding to the explicit data after medication through a blood glucose prediction module. In the data fusion module, the data before medication, the data after medication, and the prediction residual are spliced together. Based on the spliced fusion features, the data prediction module uses the prediction residual to identify the blood glucose fluctuation characteristics caused by non-drug factors and separate the net effect of the drug on blood glucose, thereby improving the accuracy of drug efficacy prediction.
[0015] Furthermore, multimodal data prediction models for diabetes gene therapy responses also include:
[0016] The data population module is used to populate explicit data to obtain a standard data matrix for the user within a time period;
[0017] A standard data matrix contains data information for each point in time within a time period;
[0018] The Kalman filter module uses the Kalman filter algorithm to filter out outliers in the latent data, obtaining a numerical matrix.
[0019] The numerical matrix represents the blood glucose values for each hour within a time period.
[0020] Furthermore, the data prediction module includes:
[0021] The information fusion layer and data encoding module construct the output spatial dictionary.
[0022] The predicted output layer, based on fused features, maps the closest label from the spatial dictionary as the efficacy for the current time period.
[0023] As a second aspect of this application, the multimodal data prediction model for diabetes gene therapy response also includes:
[0024] Preprocessing module: Obtains the current time period, calculates the average variance of the standard data matrix of the current time period and the standard data matrix of the mapping period to obtain the deviation matrix;
[0025] Use the standard data matrix and deviation matrix of the current time period as the first prediction feature;
[0026] The numerical matrix of the mapping period is used as the second prediction feature.
[0027] The first and second predictive features are input into the blood glucose prediction model to predict the blood glucose value for the current time period.
[0028] Furthermore, the mapping period is several time periods adjacent to the current time period.
[0029] Furthermore, blood glucose prediction models include:
[0030] The information input layer is used to input the first predicted feature to obtain the data encoding;
[0031] Convolutional networks extract hidden features from data encoding;
[0032] The temporal feature extraction layer simultaneously receives hidden features and second predicted features to generate temporal features;
[0033] The output layer generates the predicted blood glucose value for the current time period.
[0034] This application uses a preprocessing module to calculate the average variance of the standard data matrices of the current time period and the mapping period to generate a deviation matrix. The current period data and the deviation matrix are used to construct the first prediction feature, and the original data of the mapping period are used to construct the second prediction feature. The temporal feature extraction layer simultaneously receives the hidden features output by the convolutional network and the second prediction feature. By fusing the current data deviation state with historical blood glucose change trajectories, it extracts cross-period temporal dependencies, improving the accuracy of blood glucose predictions. Furthermore, when the patient's blood glucose regulation mechanism changes, the model's accuracy rapidly decreases, resulting in specific residual information. As a third aspect of this application, the multimodal data prediction model for diabetes gene therapy responses also includes:
[0035] The comparison sample extraction module collects the data sample that is closest to the standard data matrix in the current time period from all data samples of those who have not taken medication as the calibration sample, and uses the data sample in the current time period as the prediction sample;
[0036] The calibration samples and prediction samples are input into the blood glucose prediction model to calculate the prediction residuals;
[0037] The prediction residuals include blood glucose prediction error and network output parameter error relative to the calibrated samples.
[0038] Furthermore, the prediction residuals include: numerical error, data encoding error, hidden feature error, and temporal feature error; the blood glucose prediction error is a numerical error, and the network output parameter error is a data encoding error, a hidden feature error, and a temporal feature error.
[0039] Among them, numerical error is the error data between the predicted blood glucose value and the actual blood glucose value within the current time period;
[0040] Data encoding error is the residual between the calibrated sample and the predicted sample at the information input layer;
[0041] The hidden feature error is the residual between the labeled sample and the predicted sample in the convolutional network;
[0042] The temporal feature error is the residual between the calibrated sample and the predicted sample in the temporal feature extraction layer.
[0043] This application selects the untreated calibration sample that is closest to the current period data through a comparative sample extraction module. The calibration sample and the prediction sample are simultaneously input into the blood glucose prediction model. Data encoding error, hidden feature error, temporal feature error and numerical error are generated in the information input layer, convolutional network, temporal feature extraction layer and output layer, respectively. By analyzing the transmission relationship of the residuals at each level in the neural network, the source of error inside the model is located. Thus, the subsequent data fusion module can better extract the parameter information that affects the change of drug efficacy, which makes it easier for the data prediction module to locate the actual influencing factors of drug efficacy changes.
[0044] Furthermore, the data fusion module includes:
[0045] The model feature map encoder is used to encode the prediction residuals into latent information feature maps;
[0046] The data feature map encoder is used to encode the actual blood glucose values of the predicted sample and the actual blood glucose values of the calibrated sample at the corresponding time into explicit information feature maps.
[0047] The first feature extraction network is used to extract the first labeled features from the latent information feature map;
[0048] The second feature extraction network is used to extract the second labeled features from the explicit information feature map;
[0049] The fusion unit is used to fuse the first and second labeled features to generate a fused feature.
[0050] Furthermore, the first feature extraction network and the second feature extraction network have the same structure and are set in parallel;
[0051] The first feature extraction network consists of alternating local feature extractors and global feature extractors;
[0052] Local feature extractor, used to extract local self-attention information;
[0053] A global feature extractor is used by gating mechanisms to fuse and filter global information.
[0054] This application uses a model feature map encoder to encode the predicted residuals into a latent information feature map, and a data feature map encoder to encode the actual blood glucose values into an explicit information feature map. A first feature extraction network captures the self-attention associations of the latent information feature maps through a local feature extractor, and a global feature extractor filters global residual patterns based on a gating mechanism to generate first labeled features. A second feature extraction network processes the explicit information feature maps with the same structure to generate second labeled features. A fusion unit integrates the dual-path labeled features to establish a mapping relationship between the abstract residual patterns and the physical changes in blood glucose, generating fused features. Thus, by separating the residual distribution patterns extracted from the latent information feature maps and the blood glucose fluctuation patterns from the explicit information feature maps, an explicit association rule is established in the fusion unit between the spatial distribution characteristics of the residuals and the temporal changes in blood glucose, thereby accurately determining the impact of gene therapy drugs on blood glucose. Attached Figure Description
[0055] Figure 1 A schematic diagram of the structure of a multimodal data prediction model for the response to gene therapy for diabetes.
[0056] Figure 2 This is a schematic diagram of the blood glucose prediction model.
[0057] Figure 3 This is a schematic diagram of the data fusion module.
[0058] Figure 4 This represents the blood glucose prediction residuals for a subset of samples. Detailed Implementation
[0059] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.
[0060] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.
[0061] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.
[0062] Gene therapy drugs for diabetes are biological agents that deliver therapeutic gene fragments to the patient's target cells via vectors (such as viral or non-viral vectors) to correct or regulate defects in the expression of diabetes-related genes. Each patient's response to gene therapy drugs for diabetes is different. Therefore, after a patient begins taking gene therapy drugs for diabetes, it is necessary to continuously monitor blood glucose levels to assess efficacy. Once efficacy is confirmed, other blood glucose control medications should be gradually discontinued to prevent recurrence of the patient's condition.
[0063] Therefore, there is a key point in this protocol: the point at which gene therapy drugs for diabetes are started. The time after starting gene therapy drugs for diabetes is defined as "after medication," and the time before starting gene therapy drugs for diabetes is defined as "before medication."
[0064] After starting diabetes gene therapy, patients will reduce their current dosage of hormone medications. The specific method and timing of discontinuation are determined by the doctor. This plan only extracts the implicit effects (efficacy) of diabetes gene therapy on blood glucose regulation from complex data.
[0065] Of course, in practice, directly increasing the intake of gene therapy drugs for diabetes while simultaneously stopping the intake of hormone-based blood sugar medications can quickly reveal the patient's sensitivity to gene therapy drugs for diabetes, or in other words, their efficacy.
[0066] However, this approach can cause irreversible damage to the patient's health, and the changes in blood sugar after a large amount of medication is discontinued in a short period of time cannot directly determine the long-term effects of gene therapy. Therefore, this application provides the following technical solution:
[0067] refer to Figure 1 Example 1: The multimodal data prediction model for diabetes gene therapy response includes: a no-drug data acquisition module, a medication data acquisition module, a blood glucose prediction module, a data fusion module, and a data prediction module.
[0068] The module for collecting data on users who have not taken medication collects data samples; the module for collecting data on users who have taken medication collects data samples.
[0069] In this plan, "medication" refers to the use of gene therapy drugs for diabetes. Data samples collected before the user begins taking the gene therapy drugs are gathered separately to create a raw database. Data collected after the user begins taking the gene therapy drugs are also gathered separately to create a prediction database. The prediction database will continuously increase its sample size as the user's medication duration increases. The sample size in the raw database remains constant. Therefore, a minimum sample size needs to be set for the raw database to ensure that sufficient data samples are collected.
[0070] The data sample structures in the original database and the prediction database are identical. Specifically, the data samples include explicit and implicit data. Explicit data consists of physical and medication information related to blood glucose changes, while implicit data consists of blood glucose values corresponding to the explicit data.
[0071] Specifically, explicit data includes:
[0072] Physiological parameters: height, weight, body fat percentage, blood pressure, heart rate, BMI;
[0073] Behavioral parameters: aerobic exercise duration, anaerobic exercise duration, carbohydrate intake, dietary fiber content, glycemic index;
[0074] Medication information: Insulin injection dosage and diabetes gene therapy drug dosage. Before starting medication, the diabetes gene therapy drug dosage was set to 0 standard reference dosage, and the insulin injection dosage was set to 1 standard reference dosage or 0.5 standard reference dosage according to the treatment plan.
[0075] After starting medication, the dosage of the diabetes gene therapy drug is set at 0.5 standard reference doses, and the insulin injection dosage is also set at 0.5 standard reference doses. The insulin injection dosage and the dosage of the diabetes gene therapy drug are determined by the doctor based on the patient's actual situation. The general idea is to adjust the patient's insulin injection dosage within safe limits before starting the gene therapy drug to obtain blood glucose feedback information.
[0076] Implicit data includes capillary blood glucose levels at various time points: fasting in the morning, before meals, 2 hours after meals, and before bedtime.
[0077] The explicit and implicit data in this plan are all data that can be directly collected from patients in their daily lives, thus ensuring data integrity as much as possible.
[0078] Furthermore, multimodal data prediction models for diabetes gene therapy responses also include:
[0079] The data population module is used to populate explicit data to obtain a standard data matrix for the user within a time period;
[0080] A standard data matrix contains data information for each point in time within a time period;
[0081] The Kalman filter module uses the Kalman filter algorithm to filter out outliers in the latent data, resulting in a numerical matrix.
[0082] Specifically, the data filling module is mainly used to fill in missing data.
[0083] For example, if a user has missed some data, this data needs to be filled in as an empty set and represented using a specific encoding.
[0084] For example, in medication information, if a user did not inject insulin, the entry for the insulin dosage would be replaced with 0. This ensures that elements at the same position in the resulting standard data matrix have the same physical meaning.
[0085] In fact, the structure of a standard data matrix is as follows: the horizontal axis represents the set time point, and the vertical axis represents the values of the remaining information.
[0086] The key information to focus on is the insulin injection dosage and the dosage of gene therapy drugs for diabetes; the rest of the information remains largely unchanged.
[0087] The reason for setting up the standard information matrix in this way is mainly to ensure data alignment.
[0088] Correspondingly, the implicit data is blood glucose levels (capillary blood glucose levels), which means that users upload their blood glucose values at each time point. Users may make input errors when uploading their blood glucose values, so Kalman filtering is needed to correct these errors. Kalman filtering is a common data processing method in this field, and the specific steps will not be elaborated here.
[0089] The blood glucose prediction module trains the blood glucose prediction model with data samples from users who have not taken medication, and inputs explicit data from data samples from users who have taken medication into the blood glucose prediction model to obtain prediction residuals.
[0090] The blood glucose prediction model is a common neural network model. Its input is explicit data, and its output is implicit data. That is, by inputting the explicit data of the day into the blood glucose prediction model, the model can roughly predict the blood glucose value at each time point.
[0091] Specifically, the blood glucose prediction model is first trained using untreated data from the original database, fixing the model parameters. Then, explicit data from the user's medication-treated data samples are input into the blood glucose prediction model to predict the blood glucose value at each time point. Next, the blood glucose value at each time point is actually measured, and the predicted blood glucose value is subtracted from the actual blood glucose value to obtain the prediction residual.
[0092] The data fusion module fuses data samples from users who have not taken medication, data samples from users who have taken medication, and prediction residuals to generate fused features;
[0093] The data prediction module inputs the fused features into the convolutional network model to generate drug efficacy prediction values.
[0094] The data prediction module is a convolutional network model that generates drug efficacy predictions by fusing features.
[0095] Specifically, the data prediction module includes:
[0096] The information fusion layer and data encoding module construct the output spatial dictionary.
[0097] The predicted output layer, based on fused features, maps the closest label from the spatial dictionary as the efficacy for the current time period.
[0098] This application encodes the fused features into high-dimensional vectors through an information fusion layer, constructs a spatial dictionary (storing feature vectors corresponding to different drug effects) based on historical drug effect response data, and calculates the similarity between the current fused feature and each vector in the spatial dictionary, selecting the drug effect label associated with the dictionary entry with the highest similarity as the final prediction output.
[0099] Thus, in practice, this application has good specificity (adaptability) for each patient. For each patient, it is only necessary to collect relevant data from the patient within one or two months before taking the medication, and after training the blood glucose prediction model, the doctor can reduce the patient's normal use of hormone drugs within safe limits, and then add the use of diabetes gene therapy drugs. In this way, while maintaining the patient's blood glucose within safe limits, the doctor can also discover the effect of diabetes gene therapy drugs on blood glucose.
[0100] refer to Figure 2 Example 2: Based on Example 1, Example 2 provides a blood glucose prediction model and a preprocessing module for preprocessing the data input to the blood glucose prediction model. Specifically:
[0101] Multimodal data prediction models for diabetes gene therapy responses also include:
[0102] Preprocessing module: Obtain the current time period, calculate the average variance of the standard data matrix of the current time period and the standard data matrix of the mapping period to obtain the deviation matrix; the mapping period is several time periods adjacent to the current time period.
[0103] Use the standard data matrix and deviation matrix of the current time period as the first prediction feature;
[0104] The numerical matrix of the mapping period is used as the second prediction feature.
[0105] The first and second predictive features are input into the blood glucose prediction model to predict the blood glucose value for the current time period.
[0106] The time period refers to a fixed time unit for predicting drug efficacy, or in other words, predicting blood glucose levels. It is typically set to 24 hours. The current time period refers to the time period within which the blood glucose level needs to be predicted. The mapping period is the period closest to the current time period.
[0107] For example, today is October 5th, the current time period is October 5th, and the mapping period is selected from October 1st to October 4th. That is, the mapping period is selected to be 4 time periods in length.
[0108] After October 5th, the standard data matrix for October 5th can be collected. Then, the standard data matrices extracted from each day from October 1st to October 4th are extracted, and the average variance of the standard data matrix for October 5th is calculated by comparing it with that of October 1st to October 4th.
[0109] The formula is as follows:
[0110] ;
[0111] in, This represents the d-th element in the deviation matrix. This represents the d-th element in the standard data matrix for the current time period. This represents the d-th element in the c-th standard data matrix during the mapping period, where c represents the index of the standard data matrix during the mapping period, and C represents the length of the mapping period.
[0112] Thus, all of them are calculated. Then, the deviation matrix can be obtained.
[0113] Thus, the first predictive feature includes a standard data matrix and a deviation matrix. The standard data matrix describes the user's exercise, diet, and medication on that day. The deviation matrix describes the deviation of the user's behavior on that day from their behavior over the past few days.
[0114] like Figure 4 As shown, Figure 4 It can be seen that the residual information of patients changed significantly before and after taking the medication.
[0115] The second predictive feature includes a numerical matrix mapping the period. This numerical matrix describes the changes in blood glucose levels. Therefore, the second predictive feature provides information on blood glucose changes over the previous few days.
[0116] Thus, the first and second prediction features are input into the blood glucose prediction model to predict the blood glucose value for the current time period.
[0117] The key aspect of this application is not predicting blood glucose levels, but rather analyzing the causes of blood glucose fluctuations using the perturbations in a neural network model. To this end, after the current time period ends, not only a standard data matrix can be obtained, but also a numerical matrix for the current time period (i.e., the actual blood glucose levels at different time points). This blood glucose value is used to generate prediction residuals, which are crucial for subsequently predicting the accuracy of drug efficacy.
[0118] The following is a specific blood glucose prediction model:
[0119] Blood glucose prediction models include:
[0120] Information input layer, used to input the first predicted feature Obtain data encoding ;
[0121] ;
[0122] in, Encode the data. Indicates encoding operation. This represents the weight matrix of the information input layer. This represents the bias term of the information input layer;
[0123] The dimension is T represents the total number of time points. The total number of all elements in the explicit data includes height, weight, body fat percentage, blood pressure, heart rate, BMI, aerobic exercise duration, anaerobic exercise duration, carbohydrate intake, dietary fiber content, glycemic index, insulin injection dosage, and dosage of gene therapy drugs for diabetes; the time points are: fasting in the morning, before lunch, 2 hours after lunch, and before bedtime.
[0124] Convolutional networks, from data encoding Extract hidden features ;
[0125] ;
[0126] in, Indicates hidden features, This represents the convolution kernel, and the dimension of the convolution kernel is: ; This represents the total number of each type of element in the explicit data. Indicates the number of hidden layers; Represents the bias vector of the convolutional network; This represents the activation function. Indicates the convolution operation;
[0127] The temporal feature extraction layer simultaneously receives hidden features and second predicted features to generate temporal features;
[0128] ;
[0129] ;
[0130] ;
[0131] in, Indicates the second predictive feature. The projected features representing the second predicted features, Indicates the projection weights; Indicates feature splicing, Indicates splicing characteristics, This indicates temporal feature extraction. This represents the gating parameters, which include updating the gate, resetting the gate, and candidate states. Indicates the initial hidden state;
[0132] The output layer generates the predicted blood glucose value for the current time period.
[0133] ;
[0134] Indicates the predicted blood glucose value. This represents the output layer weight matrix. This represents the output layer bias vector. Indicates the activation function;
[0135] The key to this approach is utilizing the first predictive feature to obtain the standard data features for the current day, and the difference (deviation matrix) between these standard data features and recent standard data features. After extracting these hidden features in a convolutional network, a temporal feature extraction layer is used to simultaneously process the hidden features and the second predictive feature. The resulting temporal features reflect the deviation relationship of blood glucose values. Therefore, using these temporal features for blood glucose prediction yields high accuracy. Furthermore, this blood glucose prediction model highly values the dynamic relationships between data points, ensuring accurate regression to the actual predicted values even after data changes. When a user takes gene therapy, because the gene therapy alters the user's blood glucose metabolism from a fundamental level, it restructures the user's blood glucose change logic. Consequently, the previously learned relationships between data in the blood glucose prediction model will exhibit significant errors. Therefore, extracting the residual information from the blood glucose prediction model can be used to analyze whether the gene therapy is effective in treating diabetes.
[0136] Example 3: Based on Example 1, Example 3 provides a new data fusion model.
[0137] The data prediction model is a convolutional network classifier that maps fused features to corresponding labels based on similarity calculations within a pre-defined classification space. Therefore, to accurately determine drug efficacy, the final fused features need to contain more implicit information related to drug efficacy. Based on this, this application provides the following technical solution:
[0138] Multimodal data prediction models for diabetes gene therapy responses also include:
[0139] The comparison sample extraction module collects the data sample that is closest to the standard data matrix in the current time period from all data samples of those who have not taken medication as the calibration sample, and uses the data sample in the current time period as the prediction sample;
[0140] The calibration samples and prediction samples are input into the blood glucose prediction model to calculate the prediction residuals;
[0141] The prediction residuals include blood glucose prediction error and network output parameter error relative to the calibrated samples.
[0142] The calibration samples were obtained by retrieving historical samples from the untreated original database that were most similar to the standard data matrix of the current time period (after medication) as calibration samples. Specifically, the Euclidean distance between each untreated sample and the dominant data (physiological parameters, behavioral parameters, and drug information) of the current period was calculated, and the sample with the smallest distance was selected as the calibration sample.
[0143] Thus, the calibration sample and the predicted sample are essentially the two samples that are most similar before and after medication. If the blood glucose prediction model's prediction accuracy for these two samples is very close, that is, the residual of the predicted sample is small, it indicates that the user's body's blood glucose metabolism has not been substantially affected by the medication.
[0144] To further reveal the impact of prediction residuals on drug action, a more advanced technical solution is provided:
[0145] The prediction residuals include: numerical error, data encoding error, hidden feature error, and temporal feature error; the blood glucose prediction error is a numerical error, and the network output parameter error is a data encoding error, hidden feature error, and temporal feature error.
[0146] Among them, numerical error is the error data between the predicted blood glucose value and the actual blood glucose value within the current time period;
[0147] ;
[0148] Indicates numerical error. This represents the matrix of predicted blood glucose values output by the blood glucose prediction model. Represents the matrix of actual blood glucose values. express The dimension;
[0149] Data encoding error is the residual between the calibrated sample and the predicted sample at the information input layer;
[0150] ;
[0151] Indicates data encoding error. Indicates the predicted sample, Indicates the calibration sample. This represents the output of the predicted sample at the information input layer. This represents the output of the calibration sample at the information input layer. express Dimensions This represents the total number of each type of element in the explicit data;
[0152] The hidden feature error is the residual between the labeled sample and the predicted sample in the convolutional network;
[0153] ;
[0154] This represents the hidden feature error. Indicates the predicted sample, Indicates the calibration sample. This represents the output of the predicted sample in the convolutional network. This represents the output of the calibration sample in the convolutional network. express The dimension; Indicates the number of hidden layers in a convolutional network;
[0155] The temporal feature error is the residual between the calibrated sample and the predicted sample in the temporal feature extraction layer.
[0156] ;
[0157] ;
[0158] in, This represents the output of the predicted sample in the temporal feature extraction layer. This represents the output of the calibration sample in the temporal feature extraction layer. express Dimensions.
[0159] refer to Figure 3 The data fusion module includes:
[0160] The model feature map encoder is used to encode the prediction residuals into latent information feature maps;
[0161] ;
[0162] in, Represents a feature map containing latent information. Represents the residual projection matrix. Represents the residual feature extraction kernel. This represents the residual feature extraction bias vector. This represents the activation function. Indicates feature splicing, Indicates the convolution operation;
[0163] The data feature map encoder is used to encode the actual blood glucose values of the predicted sample and the actual blood glucose values of the calibrated sample at the corresponding time into explicit information feature maps.
[0164] ;
[0165] in, This represents the explicit information feature map, where G represents the actual blood glucose value of the predicted sample at the corresponding time and the actual blood glucose value of the calibrated sample at the corresponding time. This represents the convolution operation. This represents the convolution kernel of the data feature map encoder. Represents the bias vector of the data feature map encoder;
[0166] The first feature extraction network is used to extract latent information from feature maps. Extract the first labeled feature ;
[0167] The second feature extraction network is used to extract features from explicit information maps. Extract the second annotation feature ;
[0168] A fusion unit is used to fuse the first labeled features. Second annotation features Generate fusion features .
[0169] ;
[0170] in, This represents the activation function. This represents the weight matrix of the fusion layer. This represents the bias vector of the fusion layer.
[0171] The first and second feature extraction networks have the same structure and are set up in parallel; therefore, the first and second labeled features are two separately extracted features, and the extraction methods are the same. Therefore, only the extraction method of the first labeled feature will be described here. Specifically:
[0172] The first feature extraction network consists of alternating local feature extractors and global feature extractors;
[0173] Local feature extractor, used to extract local self-attention information;
[0174] ;
[0175] ;
[0176] ;
[0177] ;
[0178] ;
[0179] in, This represents the input to the local feature extractor. In the first local feature extractor, the input is the latent information feature map. In subsequent local feature extractors, the input is the output of the previous global feature extractor. , , These represent the query projection matrix, key projection matrix, and value projection matrix, respectively, with A representing the self-attention weights. Indicates local features, This represents the local bias vector. Represents the key dimension. Represents the local weight matrix. Indicates the activation function;
[0180] A global feature extractor is used by gating mechanisms to fuse and filter global information.
[0181] ;
[0182] ;
[0183] ;
[0184] ;
[0185] in, This serves as the input to the global feature extractor. , These are the gating weight and the gating bias, respectively. , , These represent the query projection matrix, key projection matrix, and value projection matrix, respectively. This represents the activation function. This indicates element-wise multiplication. Represents the key dimension. Indicates the output projection matrix; Indicates residual connection, This represents the output of the global feature extractor;
[0186] The output of the last global feature extractor is the first labeled feature.
[0187] The key to this approach is to utilize two identical feature extraction networks to extract more latent information that characterizes drug efficacy through multi-layered, alternating local and global feature extraction.
[0188] The data prediction module includes:
[0189] The information fusion layer and data encoding module construct the output spatial dictionary.
[0190] The predicted output layer, based on fused features, maps the closest label from the spatial dictionary as the efficacy for the current time period.
[0191] In this way, the data fusion module will generate a fusion feature, which will be compared with the labels in the spatial dictionary to determine the efficacy of the drug in the current time period.
[0192] Therefore, if the efficacy gradually increases over time, it indicates that the gene therapy drug has a certain therapeutic effect on the user. If the efficacy does not increase significantly, it indicates that the gene therapy drug is not effective in the user.
[0193] Therefore, the data fusion module and the data prediction module actually constitute a neural network model. The two are trained synchronously. After providing sufficient data samples, the data fusion module and the data prediction module can accurately extract hidden information related to the regulation of blood sugar by gene drugs.
[0194] In essence, this hidden information primarily originates from the prediction residuals, which are the residual information output layer by layer by the blood glucose prediction model. However, within each layer, the residuals of all neurons are also provided. In particular, the hidden feature error contains the error information of each neuron or hidden layer in the convolutional network. Although a convolutional network is a black box model, and it's impossible to determine exactly what information each hidden layer processes, each hidden layer in a convolutional network has a different sensitivity to different types of information. Delving into the residual information at this level allows for accurate capture of the dynamic characteristics of the convolutional network.
[0195] The implicit information feature map contains the actual error information of the blood glucose prediction model, as well as the error information of each level in each blood glucose prediction model. It is used to describe the accuracy of the blood glucose prediction model and reflect whether there have been any changes in the user's blood glucose variability. The explicit information feature map describes the changes in blood glucose; better blood glucose variability indicates that the user's diabetes is under control.
[0196] In other words, explicit information describes whether the user's blood sugar levels are trending towards normalization and showing signs of improvement. Implicit information analyzes whether the user's blood sugar conversion characteristics have changed. Combining these two types of information yields a fused feature, enabling accurate prediction of drug efficacy.
[0197] The key to this solution is the need for accurate data fusion and prediction modules. Therefore, loss functions for these modules are provided. Joint training of the data fusion and prediction modules using these loss functions can increase the accuracy of drug efficacy prediction. For the blood glucose prediction model, the cross-entropy loss function can be directly used.
[0198] Training samples are required when training the data fusion and data prediction modules. The training samples are obtained as follows: first, complete multimodal data (overt physical / medication information + latent blood glucose levels) from the patient's non-medication period is collected as the basic sample library;
[0199] Multimodal data of the same patients at the same time point during the medication cycle are collected simultaneously, and the closest untreated calibration sample is matched by the comparison sample extraction module.
[0200] The above describes the inputs and outputs of the data fusion and prediction modules. How to construct the training samples for these modules is based on existing technology and will not be described further here. The loss functions for the joint training of the data fusion and prediction modules are given below:
[0201] ;
[0202] in, This represents the joint loss of the data fusion module and the data prediction module. This indicates a loss of drug efficacy. This represents a dictionary regular expression. Indicates residual consistency loss;
[0203] ;
[0204] Where N represents the total number of samples used to train the data fusion module and the data prediction module, and i represents the index of the sample. This represents the fused feature vector of the i-th sample. This represents the true efficacy label of the i-th sample. This represents the dictionary vector corresponding to the actual labels. Indicates the magnitude of the vector;
[0205] ;
[0206] Where Kg represents the number of drug efficacy categories, for example, ineffective / ineffective / medium-effective / highly effective; g and j represent dictionary vector indices, respectively. , Let g and j be the learnable vectors in the spatial dictionary, respectively. Let represent the cosine similarity, m represent the minimum threshold, and max(0,) represent the Hinge function;
[0207] The spatial dictionary is a learnable pharmacodynamic prototype library that stores feature prototypes for different pharmacodynamic categories. Each prototype corresponds to a pharmacodynamic response state (e.g., ineffective / inefficient / medium-efficacy / high-efficacy). During training, the feature prototypes are continuously adjusted until each vector converges to the typical feature center of the corresponding pharmacodynamic category, thereby accurately distinguishing different fusion features. How to construct the dictionary space is a current technique, and the specific method will not be described here.
[0208] ;
[0209] Represents a feature map containing latent information. Represents explicit information feature maps. The gradient tensor represents the latent information feature map. The gradient tensor represents the explicit information feature map. Indicates the dominant factor of the residual. express The L2 norm.
[0210] The residual consistency loss is used to force the model to prioritize the drug action mechanism reflected in the prediction residual when predicting drug efficacy, rather than simply relying on changes in blood glucose levels. By constraining the gradient strength of the drug efficacy loss on the latent feature map (encoding the prediction residual) to be α times that on the dominant feature map (encoding the actual blood glucose level), the model is ensured to focus on "separating the net effect of the drug on blood glucose", that is, filtering out blood glucose fluctuations caused by non-drug factors such as diet and exercise, and accurately extracting the metabolic response pattern unique to gene therapy.
[0211] The above are merely preferred embodiments of this application and are not intended to limit the application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A multimodal data prediction model for diabetes gene therapy response, characterized in that, include: The medication non-dose data collection module collects data samples of users who have not taken their medication. The medication data collection module collects data samples after the user takes medication. The data samples include explicit data and implicit data. Explicit data is body information and medication information related to blood glucose changes, while implicit data is the blood glucose value corresponding to the explicit data. The blood glucose prediction module trains the blood glucose prediction model with data samples from users who have not taken medication, and inputs explicit data from data samples from users who have taken medication into the blood glucose prediction model to obtain prediction residuals. The data fusion module fuses data samples from users who have not taken medication, data samples from users who have taken medication, and prediction residuals to generate fused features; The data prediction module inputs the fused features into the convolutional network model to generate drug efficacy prediction values; Multimodal data prediction models for diabetes gene therapy responses also include: Preprocessing module: Obtains the current time period, calculates the average variance of the standard data matrix of the current time period and the standard data matrix of the mapping period to obtain the deviation matrix; Use the standard data matrix and deviation matrix of the current time period as the first prediction feature; The numerical matrix of the mapping period is used as the second prediction feature; The first and second prediction features are input into the blood glucose prediction model to predict the blood glucose value for the current time period. Blood glucose prediction models include: The information input layer is used to input the first predicted feature to obtain the data encoding; Convolutional networks extract hidden features from data encoding; The temporal feature extraction layer simultaneously receives hidden features and second predicted features to generate temporal features; The output layer generates the predicted blood glucose value for the current time period. The prediction residuals include: numerical error, data encoding error, hidden feature error, and temporal feature error; the blood glucose prediction error is a numerical error, and the network output parameter error is a data encoding error, hidden feature error, and temporal feature error. Among them, numerical error is the error data between the predicted blood glucose value and the actual blood glucose value within the current time period; Data encoding error is the residual between the calibrated sample and the predicted sample at the information input layer; The hidden feature error is the residual between the labeled sample and the predicted sample in the convolutional network; The temporal feature error is the residual between the calibrated sample and the predicted sample in the temporal feature extraction layer; The data fusion module includes: The model feature map encoder is used to encode the prediction residuals into latent information feature maps; The data feature map encoder is used to encode the actual blood glucose values of the predicted sample and the actual blood glucose values of the calibrated sample at the corresponding time into explicit information feature maps. The first feature extraction network is used to extract the first labeled features from the latent information feature map; The second feature extraction network is used to extract the second labeled features from the explicit information feature map; The fusion unit is used to fuse the first and second labeled features to generate a fused feature.
2. The multimodal data prediction model for diabetes gene therapy response according to claim 1, characterized in that, Multimodal data prediction models for diabetes gene therapy responses also include: The data population module is used to populate explicit data to obtain a standard data matrix for the user within a time period; A standard data matrix contains data information for each point in time within a time period; The Kalman filter module uses the Kalman filter algorithm to filter out outliers in the latent data, obtaining a numerical matrix. The numerical matrix represents the blood glucose values for each hour within a time period.
3. The multimodal data prediction model for diabetes gene therapy response according to claim 1, characterized in that, The mapping period is several time periods adjacent to the current time period.
4. The multimodal data prediction model for diabetes gene therapy response according to claim 1, characterized in that, Multimodal data prediction models for diabetes gene therapy responses also include: The comparison sample extraction module collects the data sample that is closest to the standard data matrix in the current time period from all data samples of those who have not taken medication as the calibration sample, and uses the data sample in the current time period as the prediction sample; The calibration samples and prediction samples are input into the blood glucose prediction model to calculate the prediction residuals; The prediction residuals include blood glucose prediction error and network output parameter error relative to the calibrated samples.
5. The multimodal data prediction model for diabetes gene therapy response according to claim 1, characterized in that, The first and second feature extraction networks have the same structure and are set up in parallel. The first feature extraction network consists of alternating local feature extractors and global feature extractors; Local feature extractor, used to extract local self-attention information; A global feature extractor is used by gating mechanisms to fuse and filter global information.
6. The multimodal data prediction model for diabetes gene therapy response according to claim 1, characterized in that, The data prediction module includes: The information fusion layer and data encoding module construct the output spatial dictionary. The predicted output layer, based on fused features, maps the closest label from the spatial dictionary as the efficacy for the current time period.