Blood glucose prediction method and device based on sparse feature processing and causal analysis

Abstract

The invention discloses a blood glucose prediction method and device based on sparse feature processing and causal analysis. The method comprises the following steps: acquiring blood glucose data and multi-source covariable data related to blood glucose; after the data are checked and complemented, the continuity of the time sequence is recovered; constructing a blood glucose prediction model, performing physiological formula conversion by utilizing sparse features of food intake data and drug data, then performing causal relationship analysis with blood glucose data, and determining optimal influence time lag of sparse variables on blood glucose; then, based on a time lag analysis result, time alignment processing is conducted on the multi-source covariant data and the blood glucose data; high-dimensional feature coding and multi-scale feature decoding are adopted to obtain a blood glucose prediction value in a future time period; training the blood glucose prediction model, and inputting the food intake data and the medicine data into the blood glucose prediction model after training is completed to obtain a predicted blood glucose value. The problem that in the prior art, blood glucose prediction is not accurate enough can be solved.

Description

Technical Field

[0001] The present invention belongs to the field of blood glucose prediction, and in particular relates to a blood glucose prediction method and device based on sparse feature processing and causal analysis. Background Art

[0002] Diabetes is a metabolic disease that widely exists globally and is divided into two types: type I and type II. Among them, the blood glucose level of type I diabetes patients highly depends on the supplementary management of exogenous insulin. Due to the lack or insufficiency of insulin secretion in diabetic patients, the blood glucose level is easily affected by various factors such as diet, exercise, drugs, and physiological stress, thus showing complex and unstable fluctuations. Abnormal fluctuations in blood glucose levels can lead to acute complications such as hyperglycemic crises and hypoglycemic coma, and long-term fluctuations also increase the risk of microvascular and macrovascular complications, such as retinopathy, neuropathy, and cardiovascular diseases. Therefore, accurately predicting the change trend of blood glucose levels is of great significance for the daily management of diabetic patients, insulin dose adjustment, and prevention of complications.

[0003] For example, the Chinese patent document with the publication number CN117766145A discloses a method for constructing a continuous blood glucose prediction model, a blood glucose prediction method and device. By obtaining historical data of continuously measuring blood glucose of different patients for several days to train the model, the generalization ability and prediction accuracy of the prediction model are improved. The Chinese patent document with the publication number CN107174258A discloses a blood glucose concentration prediction method. According to the physiological data, spectral data, true blood glucose concentration values, and non-blood glucose concentration data of multiple sample testers, a blood glucose concentration prediction model based on the "M + N" theory is established using a multivariate calibration algorithm, and the prediction accuracy of blood glucose concentration is effectively improved using this model.

[0004] Sparse features are a common but complex data feature category in the field of blood glucose prediction, mainly including data related to patients' dietary behaviors and drug use, such as eating time and intake, insulin injection dose and frequency. These features have an important impact on the change of blood glucose levels, but due to their sparse characteristics, many challenges are faced in the data analysis and modeling process. For example, patients' dietary behaviors are usually recorded only a few times a day, and the recording times are often uneven, appearing sparse and scattered compared with the continuously collected blood glucose data recorded every 5 minutes. Drug data, especially insulin injection data, also has time sparsity. For example, relevant information is only recorded before meals or when blood glucose is abnormal, and the entire time range is not covered. Time sparsity leads to a significant mismatch between the distribution of these features on the time axis and the continuity of blood glucose changes, which poses challenges for data fusion and modeling. This event-driven sparsity further increases the complexity of feature processing because it is difficult for the model to capture the dynamic association between sparse features and blood glucose fluctuations.

[0005] Despite the sparsity, the importance of these features in blood glucose prediction cannot be ignored. Dietary behavior is the main driving force behind blood glucose fluctuations. For example, carbohydrate intake can rapidly increase blood glucose levels. Medication data, especially insulin injections, is the core means of regulating blood glucose. The dosage, time interval, etc. of injection behavior directly determine the amplitude and speed of blood glucose decline. The profound impact of these features on blood glucose fluctuations means that they must be effectively incorporated into the prediction model. Summary of the Invention

[0006] The present invention discloses a blood glucose prediction method and device based on sparse feature processing and causal analysis, which can solve the problem of inaccurate blood glucose prediction in the prior art.

[0007] A blood glucose prediction method based on sparse feature processing and causal analysis includes the following steps:

[0008] (1) Obtain blood glucose data and multi-source covariate data related to blood glucose. The multi-source covariate data includes eating data and medication data; after checking and data completion for the blood glucose data and the corresponding multi-source data, restore the continuity of the time series.

[0009] (2) Construct a personalized blood glucose prediction model. The blood glucose prediction model uses the sparse features of eating data and medication data, performs physiological formula conversion, and then captures the influence patterns of sparse features at different time scales through a multi-scale segmentation method in the time dimension to distinguish short-term and long-term effects; then conducts a causal relationship analysis with the blood glucose data to determine the optimal time lag of the sparse variables on blood glucose; then, based on the time lag analysis results, perform time alignment processing on the multi-source covariate data and the blood glucose data; learn personal blood glucose characteristics from the processed data; during the learning process, the model uses high-dimensional feature encoding and multi-scale learning to extract individualized key features, and finally decodes to obtain the blood glucose prediction value for the future time period.

[0010] (3) Use the data processed in step (1) to train the blood glucose prediction model. After training is completed, input the eating data and medication data into the blood glucose prediction model to obtain the predicted blood glucose value.

[0011] Using the present invention, through accurate blood glucose prediction, patients can achieve personalized blood glucose management. For example, adjust diet, exercise amount, or insulin dosage according to the predicted blood glucose level, thereby avoiding unnecessary health risks. At the same time, the prediction model can also help medical staff monitor the changes in the condition and provide data support for the optimization of treatment plans.

[0012] In step (1), the following formula is used for data completion:

[0013]

[0014] Among them, x i and x j are known adjacent data points, at times t i and t j respectively, and t k is the time point of the missing data x k .

[0015] In step (2), physiological formula conversion is performed to convert the single-point values of feeding data and drug data into a curve over a period of time.

[0016] For the feeding data, the following formula is used to convert it into a curve over a period of time:

[0017]

[0018] where t s is the sampling time, t meal is the feeding time, Carb represents the effective carbohydrates at a given time, C meal is the total amount of carbohydrates ingested in a meal, t peak is the time when Carb reaches the maximum value, at this time a 1 is the growth rate, and a 2 is the decay rate.

[0019] For the drug data, the following formula is used to convert it into a curve over a period of time:

[0020]

[0021] where t s is the sampling time, t d is the duration of insulin activity, T is the time of exponential decay, a is the growth coefficient, and S is the proportionality coefficient.

[0022] In step (2), causal relationship analysis is performed for each transformed sparse variable and blood glucose variable, and the specific process is as follows:

[0023] First, the historical data of the blood glucose variable itself needs to be considered, and the following formula is used:

[0024]

[0025] where X t is the blood glucose value at the current moment, a i is the regression coefficient of the accident blood glucose's own historical value, p is the size of the blood glucose historical value window, is the residual value after modeling the blood glucose value;

[0026] Considering the historical data of blood glucose values and other variables simultaneously, the following formula is used:

[0027]

[0028] where Y t is the candidate causal variable, b j is the regression coefficient of the candidate causal variable, q is the window size of the historical values of the candidate causal variable, is the residual value after modeling the historical data of blood glucose values and other variables simultaneously;

[0029] To determine whether other variable Y t has a significant causal effect on blood glucose X t , calculate the F-test statistic using the following formula:

[0030]

[0031] where N is the total number of samples in the time series. Find the critical value F critical of the F-distribution based on the F statistic. If F > F critical , it indicates that Y t has a significant causal relationship with X t .

[0032] To determine the optimal time lag of the sparse variable's effect on blood glucose, the formula is as follows:

[0033]

[0034] Select the historical value window q that minimizes the sum of squared residuals. When q = τ, the causal relationship is most significant at this time, and τ is the optimal time lag, which reflects the main time of the effect of other variables on blood glucose.

[0035] Align the multi-source covariate data with the blood glucose data using the following formula:

[0036]

[0037] where is the covariate value after alignment with X t , and Y t-τ is the covariate value lagged by τ; through this adjustment, ensure that the historical information Y t of the covariate Y t-τ makes a direct contribution to the prediction of the current blood glucose X t .

[0038] A blood glucose prediction device based on sparse feature processing and causal analysis, comprising a memory and one or more processors. Executable code is stored in the memory. When the one or more processors execute the executable code, it is used to implement the above-mentioned blood glucose prediction method.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. By processing sparse covariate features, the present invention converts the original single-valued sparse features into function curves, significantly improving the utilization rate of sparse features and providing more complete data input for blood glucose prediction.

[0041] 2. By introducing causal analysis, the present invention quantifies the causal relationship between covariates and blood glucose fluctuations, determines the optimal time lag, and performs time series alignment, ensuring the scientificity and reliability of the model.

[0042] 3. Through the processing of sparse features and multi-feature encoding, the model of the present invention can provide personalized blood glucose prediction results, supporting precise diabetes management.

[0043] 4. Through multi-index evaluation, the present invention can achieve a smaller error and a higher proportion of the risk-free medical area in the prediction, better supporting clinical decisions and ensuring safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flowchart of a blood glucose prediction method based on sparse feature processing and causal analysis according to the present invention.

[0045] Figure 2 It is a schematic diagram of the conversion of eating data provided by an embodiment of the present invention.

[0046] Figure 3 It is a schematic diagram of the conversion of drug data provided by an embodiment of the present invention.

[0047] Figure 4 It is a graph of the model prediction results provided by an embodiment of the present invention.

[0048] Figure 5 It is a comparison graph between the model of the present invention and other models.

[0049] Figure 6 It is a comparison graph between the model of the present invention and other models in terms of medical index evaluation. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The present invention will be further described in detail below with reference to the drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0051] AsFigure 1 As shown in Figure 1 , a blood glucose prediction method based on sparse feature processing and causal analysis is mainly implemented through three modules, namely: a data preprocessing module 101, a model construction module 102, and a blood glucose prediction performance evaluation and comparison module 103.

[0052] The data preprocessing module 101 is used to check and complete the blood glucose data. The check is to determine the missing values of the blood glucose and set them to zero uniformly. The completion is to perform interpolation through data features to restore the continuity of the time series.

[0053] The model construction module 102 is used to train and construct a personalized blood glucose prediction model. The model adopts sparse feature processing for eating, basal metabolism, and drug sparse variables, designs a systematic sparse feature processing method, and captures the influence patterns of sparse features at different time scales through a multi-scale segmentation method in the time dimension to distinguish short-term and long-term effects. Combining blood glucose features, it conducts causal relationship analysis and optimal time lag learning to determine the optimal influence time lag of sparse variables on blood glucose. Based on the time lag analysis results, it performs time alignment processing on multi-source covariate data and blood glucose data to ensure the consistency of all features in the time series and eliminate time mismatch interference. It performs high-dimensional feature encoding on the aligned eating, basal metabolism, drug, and blood glucose data to extract individualized key features. Adopting a multi-scale learning framework, it decodes the multi-scale features into blood glucose prediction values for future time periods and generates the final prediction result by combining a deep decoding structure.

[0054] The blood glucose prediction performance evaluation and comparison module 103 is used to predict blood glucose values with different time spans, compare the prediction results of different models and actual data, and the significance of the predicted values in the medical field. A series of metrics can be calculated to evaluate the prediction accuracy of the model. The evaluation metrics include the mean absolute error (MAE), the root mean square error (RMSE), which are used to measure the accuracy of the predicted blood glucose values, and the medical index evaluation (Clarke), which is used to measure the risk of the model in predicting blood glucose values in the medical field.

[0055] The data completion uses the following formula:

[0056]

[0057] where x i and x j are known adjacent data points at times t i and t j respectively, and t k is the time point of the missing data x k .

[0058] Blood glucose data is usually incomplete, and complete data is crucial for model training and prediction. Through the completion process of the data preprocessing module 101, the continuity and integrity of the data can be improved, thereby enhancing the prediction accuracy and reliability of the model.

[0059] The model construction module 102 is a model for feature learning and prediction of personal blood glucose. Its core idea is to utilize sparse features, perform physiological formula conversion on them, and capture the influence patterns of sparse features at different time scales through a multi-scale segmentation method in the time dimension to distinguish short-term and long-term effects. Then, perform causal relationship analysis with the blood glucose data to determine the optimal influence time lag. Based on the time lag analysis results, perform time alignment processing on the multi-source covariate data and the blood glucose data to ensure the consistency of all features in the time series and eliminate time mismatch interference. Learn personal blood glucose features based on the processed data. During the learning process, the model adopts high-dimensional feature encoding and multi-scale learning to extract individualized key features, and finally performs decoding to obtain the blood glucose prediction values for future time periods.

[0060] The model prediction module 102 includes physiological formula conversion, causal relationship analysis, optimal time lag determination, and time step alignment, etc. Among them, physiological formula conversion is to convert the single-point values of eating and medications into a curve for a period of time. For eating, the following formula is used:

[0061]

[0062] where t s is the sampling time, t meal is the eating time, Carb represents the effective carbohydrates at a given time, C meal is the total amount of carbohydrates ingested in a meal, t peak is the time when Carb reaches the maximum value, at this time a 1 is the growth rate, a 2 is the decay rate.

[0063] The specific conversion method is as follows Figure 2 as shown, where the dotted line represents the original eating data, a value of -1 indicates no eating event, and a value other than -1 indicates the total amount of carbohydrates eaten at that time point. The solid line represents the converted eating data. The values for 3 time points after an eating event occur are 0, then it starts to rise, reaches the extreme value at the 12th time point, then starts to decline, and returns to 0 at the 48th time point. It should be noted that the effects of each eating event will be superimposed.

[0064] For medications, the following formula is used:

[0065]

[0066] where, t s is the sampling time, t d is the duration of insulin activity, T is the time of exponential decay, a is the growth coefficient, and S is the proportionality coefficient.

[0067] Since the effect of insulin is time-dependent, it takes a long time for insulin to affect blood glucose levels: typically, it reaches its peak after 1 hour and then gradually decays within 6 hours. The percentage of remaining insulin activity or active insulin after postprandial insulin injection can be modeled by an exponential decay curve to show the effect of insulin from the perspective of active insulin.

[0068] The specific conversion method is as follows Figure 3 as shown, where the dashed line represents the original drug injection data, a value of -1 indicates no drug injection, and a value other than -1 indicates the total number of drugs injected at that time point. The solid line represents the converted drug injection data. The peak is reached at the 12th time point after drug injection, and it returns to 0 at the 72nd time point. It should be noted that the effects of each drug injection event are additive.

[0069] Causal relationship analysis is of great significance in blood glucose prediction. It can constitute the mutual influence mechanism between blood glucose levels and various sparse variables. By analyzing the causal relationship, not only can the degree of influence of these variables on blood glucose be quantified, but also the time delay of their effects can be determined, thereby optimizing the input structure of the prediction model. This analysis helps to eliminate irrelevant or weakly related variables, avoid interference from data noise to the model, and at the same time can capture the long-term and short-term influence patterns of important variables on blood glucose. First, consider the historical data of the blood glucose variable itself, using the following formula:

[0070]

[0071] where, X t is the blood glucose value at the current moment, a i is the regression coefficient of the accident blood glucose's own historical value, p is the size of the blood glucose historical value window, is the residual value after modeling the blood glucose value. At the same time, consider the historical data of the blood glucose value and other variables, using the following formula:

[0072]

[0073] where, Y t is the candidate causal variable, such as drugs, diet, etc., b j is the regression coefficient of the candidate causal variable, q is the size of the historical value window of the candidate causal variable, is the residual value after modeling considering the historical data of the blood glucose value and other variables. To judge other variables Y tWhether there is a significant causal effect on blood glucose X t To calculate the F-test statistic, the following formula is used:

[0074]

[0075] where N is the total number of samples in the time series. The critical value F of the F-distribution is found according to the F statistic critical , if F > F critical , it indicates that Y t has a significant causal relationship with X t .

[0076] For the optimal impact time lag, the following formula is used:

[0077]

[0078] Select the historical value window q that minimizes the sum of squared residuals. When q = τ, the causal relationship is the most significant at this time, and τ is the optimal impact time lag, which reflects the main impact time of other variables on blood glucose.

[0079] For time alignment processing, it is characterized by using the following formula:

[0080]

[0081] where is the covariate value after alignment with X t , and Y t-τ is the covariate value lagged by τ. Through this adjustment, it is ensured that the historical information Y t of the covariate Y t-τ has a direct contribution to the prediction of the current blood glucose X t .

[0082] In model optimization, through multi-feature encoding, data such as blood glucose, eating, basal metabolism, and drugs are fused, and key features are extracted as inputs. Subsequently, through multi-scale learning, the blood glucose change laws of short-term fluctuations and long-term trends are captured, and the information of these different time scales is fused. Finally, in the multi-feature decoding stage, the multi-scale features are transformed into predictions of future blood glucose values. Combining the data of the first 6 time points, the comparison between the model prediction results and the real data is as follows Figure 4 as shown

[0083] In the blood glucose prediction performance evaluation module, various evaluation metrics can be used to objectively evaluate the prediction accuracy of the model. Among them, the mean absolute error (MAE) and the root mean square error (RMSE) are two of the most commonly used evaluation metrics. The MAE represents the average of the absolute errors between the predicted values and the actual values, reflecting the deviation degree of the model's prediction results and being able to intuitively reflect the average error size of the model. The RMSE takes the square root of the average of the squared errors, which can assign higher weights to larger prediction errors and highlight the influence of outliers.

[0084] In addition, the significance of the medical index evaluation (Clarke) lies in its closer proximity to actual applications. For diabetes management, the goal of the prediction model is not only numerical accuracy but also to ensure that the prediction errors do not lead to clinically incorrect decisions. Through Clarke analysis, the clinical reliability of the prediction model can be intuitively judged to ensure that the model has practical value in real scenarios.

[0085] The mean absolute error described in the blood glucose prediction performance evaluation comparison module adopts the following formula:

[0086]

[0087] where n is the number of predicted values, y i ′ is the predicted value, and y i is the actual value.

[0088] The root mean square error adopts the following formula:

[0089]

[0090] where n is the number of predicted values, y i ′ is the predicted value, and y i is the actual value.

[0091] The medical index evaluation adopts the following formula:

[0092]

[0093] where G ref is the actual blood glucose value, and G pred is the predicted blood glucose value. Among them, area A represents no clinical risk, area B represents acceptable deviation, area C represents high-risk prediction, area D represents dangerous prediction, and area E represents extremely dangerous prediction that is very likely to cause medical accidents.

[0094] Figure 5This is a comparison chart between the model of the present invention and other models. It can be seen that in tasks with different prediction time spans, the present model is superior to other models in terms of evaluation metrics such as mean absolute error (MAE) and root mean square error (RMSE). This indicates that the present model has high prediction accuracy and stability in blood glucose prediction tasks and can better reflect the true trend of blood glucose fluctuations.

[0095] Figure 6 This is a comparison chart between the model of the present invention and other models in medical index evaluation. Through the analysis of the proportion of the area without clinical risk, it can be seen that the proportion of the area without clinical risk in the present model is higher than that of other models, indicating that it can effectively avoid high-risk blood glucose prediction and verifies the reliability and safety of the present model in medical scenarios.

[0096] The above-described embodiments have detailed the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, supplements, and equivalent replacements made within the scope of the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A blood sugar prediction method based on sparse feature processing and causal analysis, characterized in that: The following steps are involved: (1) Obtain blood glucose data and multi-source covariate data related to blood glucose, including food intake data and drug data; after checking and completing the blood glucose data and the corresponding multi-source data, restore the continuity of the time series; (2) constructing a personalized blood glucose prediction model, which utilizes the sparse features of food intake data and drug data to perform physiological formula conversion, and then uses a multi-scale segmentation method in the time dimension to capture the impact pattern of the sparse features at different time scales and distinguish short-term and long-term effects; Then, a causal relationship analysis is performed with the blood glucose data to determine the optimal time lag of the sparse variables on blood glucose. Then, based on the results of the time lag analysis, the multi-source covariate data is time-aligned with the blood glucose data. The individual blood sugar characteristics are learned based on the processed data. During the learning process, the model uses high-dimensional feature encoding and multi-scale learning to extract individual key features, and finally decodes to obtain the blood sugar prediction value for the future time period; (3) The data processed in step (1) is used to train the blood glucose prediction model. After the training is completed, the eating data and the drug data are input into the blood glucose prediction model to obtain the predicted blood glucose value.

2. The blood glucose prediction method based on sparse feature processing and causal analysis according to claim 1, characterized in that: In step (1), the following formula is used for data completion: Among them, x i and x j are known adjacent data points, respectively at time t i and t j t k is the missing data x k time point.

3. The blood sugar prediction method based on sparse feature processing and causal analysis according to claim 1, characterized in that: In step (2), the physiological formula conversion converts the single-point values ​​of the eating data and the drug data into a curve over a period of time.

4. The blood sugar prediction method based on sparse feature processing and causal analysis according to claim 3, characterized in that: For the eating data, the following formula is used to convert it into a curve over a period of time: Among them, t s is the sampling time, t meal is the time of eating, Carb represents the available carbohydrates at a given time, C meal is the total amount of carbohydrates consumed in a meal, peak This is when Carb reaches its maximum value. a1 is the growth rate and a2 is the decline rate.

5. The blood sugar prediction method based on sparse feature processing and causal analysis according to claim 3, characterized in that: For drug data, the following formula is used to convert it into a curve over a period of time: Among them, t s is the sampling time, t d is the duration of insulin activity, T is the exponential decay time, a is the growth coefficient, and S is the proportional coefficient.

6. The blood sugar prediction method based on sparse feature processing and causal analysis according to claim 1, characterized in that: In step (2), the causal relationship analysis is performed on each transformed sparse variable and blood glucose variable. The specific process is as follows: First, we need to consider the historical data of the blood sugar variable itself, using the following formula: Among them, X t is the blood sugar value at the current moment, a i The regression coefficient of the accident blood sugar historical value, p is the size of the blood sugar historical value window, is the residual value after blood glucose value modeling; Taking into account the historical data of blood sugar values ​​and other variables, the following formula is used: Among them, Y t is a candidate causal variable, b j is the regression coefficient of the candidate causal variable, q is the window size of the historical value of the candidate causal variable, It is the residual value after modeling the historical data of blood sugar value and other variables; In order to determine other variables Y t Is it effective for blood sugar? t There is a significant causal effect, and the F test statistic is calculated using the following formula: Where N is the total number of samples in the time series, and the critical value F of the F distribution is found based on the F statistic. critical , if F>F critical , then it means that Y t X t There is a significant causal relationship.

7. The blood sugar prediction method based on sparse feature processing and causal analysis according to claim 6, characterized in that: In step (2), the optimal time lag of the sparse variable on blood glucose is determined, and the formula is as follows: Select the historical value window q that minimizes the residual sum of squares. When q = τ, the causal relationship is most significant, and τ is the optimal impact lag, which reflects the main impact time of other variables on blood sugar.

8. The blood sugar prediction method based on sparse feature processing and causal analysis according to claim 7, characterized in that: The multi-source covariate data and blood glucose data are time-aligned using the following formula: in, Is with X t Covariate values ​​after alignment, Y t-τ is the covariate value at lag τ; through this adjustment, the covariate Y t Historical information of Y t-τ For current blood sugar X t directly contribute to the prediction.

9. A blood sugar prediction device based on sparse feature processing and causal analysis, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement the blood glucose prediction method according to any one of claims 1 to 8.

Images