Personalized medical health service recommendation system based on big data analysis

Abstract

The invention discloses a personalized medical health service recommendation system based on big data analysis, and belongs to the technical field of medical health services. The synchronization module is used for generating a standardized data stream according to dynamic blood glucose monitoring data and an original heart rate variability signal; the analysis module is used for generating a diet time sequence data stream according to the intake event timestamp; the blood glucose feature extraction module is used for extracting data from the standardized data stream; the insulin sensitivity analysis module is used for generating an insulin sensitivity factor matrix; the nutrition analysis module is used for generating a carbohydrate equivalent time distribution vector and a dietary fiber intake intensity value; the feature fusion module is used for generating a fusion feature vector according to the blood glucose drift trend coefficient, the insulin sensitive factor matrix and the carbohydrate equivalent time distribution vector; the risk assessment module is used for generating a hypoglycemia risk probability value according to the fusion feature vector; and the recommendation generation module is used for generating a personalized service recommendation set according to the hypoglycemia risk probability value.

Description

Technical Field

[0001] The present invention relates to the field of medical and health service technology, and in particular to a personalized medical and health service recommendation system based on big data analysis. Background Art

[0002] With the advancement of technology, personalized healthcare and health management are becoming increasingly important developments in modern medicine. In particular, in the management of diabetic patients, providing precise medical recommendations based on their individual characteristics and dynamic physiological status has become a pressing technical challenge. In recent years, electronic health records (EHRs) and regular physical examination data have been widely used in diabetes management recommendation systems. However, these systems typically rely on static data and fail to fully utilize patients' dynamic monitoring information.

[0003] Currently, many existing recommendation systems generate service recommendations primarily based on patients' electronic health records and regular test data. While this data provides basic information for patient management, due to its static nature, it lacks a dynamic response to real-time changes in patients' status. Furthermore, the health status of diabetic patients is not only affected by static data, but also by multiple factors such as blood sugar fluctuations, diet, and exercise. In diabetes management in particular, patients' time-series data, such as continuous glucose monitoring (CGM), insulin injection records, and diet logs, are often stored in isolated forms across different medical devices, making information integration difficult.

[0004] Currently, traditional recommendation systems usually use fixed time windows (such as 24 hours) to aggregate data. This approach not only ignores individual differences in patients, but also leads to significant matching problems between the generated exercise and diet recommendations and the patient's real-time physiological state. Especially at night, patients may experience the risk of asymptomatic hypoglycemia. At this time, traditional systems are unable to integrate heart rate variability data and CGM trend values from wearable devices within a reasonable time delay (<30 minutes). This delay directly causes a deviation between the recommended service and the patient's actual physiological state, especially in terms of the difference in postprandial blood glucose targets, which may exceed the clinically allowed range (such as >1.8mmol / L), thereby having a negative impact on the patient's health management.

[0005] The disclosure of the above background technology content is only used to assist in understanding the concept and technical solution of the present invention. It does not necessarily belong to the prior art of this patent application. In the absence of clear evidence that the above content has been disclosed on the filing date of this patent application, the above background technology should not be used to evaluate the novelty and creativity of this application. Summary of the Invention

[0006] This application provides a personalized medical and health service recommendation system based on big data analysis, which can achieve minute-level accurate prediction and personalized intervention of hypoglycemia risk in diabetic patients.

[0007] To achieve the above objectives, the present application discloses the following technical solutions:

[0008] A personalized medical and health service recommendation system based on big data analysis, including:

[0009] an acquisition module, configured to acquire the first user's dynamic blood glucose monitoring data, insulin injection records, diet log text data, and original heart rate variability signal;

[0010] The synchronization module is used to synchronize the dynamic blood glucose monitoring data and the original heart rate variability signal with minute-level timestamps to generate a standardized data stream;

[0011] A parsing module is used to parse the intake event timestamps in the diet log text data, associate each diet event with a preset metabolic impact time window, and generate a diet time series data stream;

[0012] Blood glucose feature extraction module, used to extract blood glucose change rate curve, blood glucose drift trend coefficient and blood glucose fluctuation period feature vector from the standardized data stream;

[0013] Insulin sensitivity analysis module, used to generate an insulin sensitivity factor matrix based on the time series of insulin injection records and the blood glucose change rate curve;

[0014] Nutrition parsing module, which performs natural language processing on the dietary time series data stream to parse out the carbohydrate equivalent time distribution vector and dietary fiber intake intensity value;

[0015] A feature fusion module is used to generate a fusion feature vector based on the blood glucose drift trend coefficient, the insulin sensitivity factor matrix and the carbohydrate equivalent time distribution vector;

[0016] A risk assessment module is used to process the fused feature vector using a recurrent neural network model to generate a hypoglycemia risk probability value;

[0017] The recommendation generation module is used to generate a personalized service recommendation set when the hypoglycemia risk probability value exceeds a preset threshold.

[0018] In the embodiment of the present application, the acquisition module breaks down the data barriers between devices, and high-frequency CGM data (minute level), event-triggered insulin injection records, unstructured diet log text, and continuously collected heart rate variability signals are uniformly incorporated into the processing flow; then the synchronization module is used to align the minute-level timestamps of CGM and heart rate variability signals, eliminating the time scale misalignment caused by the traditional 24-hour fixed window, so that the physiological response correlation of data of different frequencies is improved from the coarse granularity of "day level" to the fine granularity of "minute level"; the parsing module targets the sparsity characteristics of the diet log and converts discrete diet events into continuous time series data streams through a preset metabolic impact time window, breaking through the mechanical modeling method of the traditional system that simplifies the diet impact into a fixed time interval. On this basis, the feature fusion module constructs a multi-dimensional real-time physiological state representation by extracting the dynamic change characteristics of blood glucose, insulin sensitivity factors, and nutritional metabolism curves, and finally realizes the dynamic probability prediction of hypoglycemia risk by the recurrent neural network. When the risk probability exceeds the threshold, the recommendation generation module immediately triggers the personalized intervention strategy, forming a millisecond-level closed loop from data acquisition to decision execution.

[0019] In this way, the data processing paradigm of the traditional health management system is fundamentally reconstructed, the asynchrony of isolated data sources is transformed into the synergy of time-series correlation, and static historical data analysis is upgraded to dynamic real-time status perception, thereby effectively suppressing the deviation between recommended services and actual physiological status, and achieving clinically acceptable intervention timeliness in the scenario of asymptomatic hypoglycemia.

[0020] In some possible implementations, the synchronization module includes:

[0021] The signal conversion unit is used to perform wavelet transform on the original heart rate variability signal to generate a real-time spectrum energy distribution vector, wherein the wavelet transform formula is:

[0022]

[0023] Where x(t) is the original heart rate variability signal, ψ(t) is the Morlet mother wavelet function, a is the scale factor, b is the time translation factor; t is the time variable;

[0024] The time alignment unit is used to align the continuous blood glucose monitoring data with the real-time spectrum energy distribution vector according to minute-level timestamps to generate a standardized data stream, where the alignment formula is:

[0025] S(t)={CGM(t i ),HRV(t j )||t i -t j |<∈

[0026] Where, ∈≥2.5 minutes, CGM(ti ) is the time t i Blood glucose level, HRV(t j ) is the time t j spectral energy values. The signal conversion unit performs a wavelet transform on the raw heart rate variability signal to generate a real-time spectral energy distribution vector. This generates a real-time spectral energy distribution vector, leveraging the time-frequency localization of the wavelet transform to accurately capture transient changes in autonomic nervous system regulation. Continuous glucose monitoring data is collected at fixed intervals, while the heart rate variability signal is transformed using a wavelet transform to generate a real-time spectral energy distribution vector. These two differ in data structure and temporal granularity. The time alignment unit synchronizes the continuous glucose monitoring data with the real-time spectral energy distribution vector using minute-by-minute timestamps, establishing a unified time base and ensuring real-time alignment between glucose levels and autonomic nervous system activity. The multi-scale analysis capabilities of the wavelet transform, combined with the minute-by-minute synchronization mechanism, enable a precise temporal mapping of the spectral features of the high-frequency heart rate variability signal with the relatively low-frequency glucose fluctuation data, revealing the dynamic correlation between glucose changes and autonomic nervous system stress responses. The synchronized, standardized data stream provides time-consistent multimodal physiological features for subsequent analysis, enabling the system to simultaneously track coordinated changes in glucose trends and neural regulation.

[0027] In some possible implementations, the parsing module includes:

[0028] Log extraction unit, used to extract food intake time point t from diet log text data k and description of food type;

[0029] The metabolic weight calculation unit is used to calculate the metabolic weight value of each food type within the impact time window according to the pharmacokinetic model. The specific formula is:

[0030] where t∈[t k ,t k +T]

[0031] Where σ represents the standard deviation of metabolic rate, σ≤T / 4; t k The time when the dietary event occurs; T: the duration of the preset metabolic impact time window;

[0032] The data stream generation unit is used to bind the metabolic weight value w(t) to the corresponding timestamp to generate a dietary time-series data stream containing the time weight distribution. In this way, through the synergistic effect of the log extraction unit, the metabolic weight calculation unit, and the data stream generation unit, accurate parsing of the dietary log and dynamic metabolic impact modeling can be achieved. Specifically, the log extraction unit extracts food intake time points and type descriptions from the unstructured dietary log text, providing basic data support for metabolic weight calculation. Based on the pharmacokinetic model, the metabolic weight calculation unit uses a Gaussian distribution function to quantify the dynamic weight of food type within a preset metabolic impact time window. The standard deviation parameter controls the individual differences in metabolic rate, breaking through the traditional simplified assumption of the fixed time window for the nutrient absorption process. The data stream generation unit dynamically binds the metabolic weight distribution to the timestamp, converting discrete dietary events into a continuous time series data stream, fully depicting the absorption and metabolic trajectory of nutrients such as carbohydrates.

[0033] In this way, the delayed effect, duration and individual metabolic differences of dietary events on blood glucose fluctuations can be accurately quantified, providing high-fidelity time-series input data for subsequent blood glucose prediction models, thereby effectively suppressing blood glucose prediction errors caused by deviations in dietary metabolic modeling.

[0034] In some possible embodiments, the insulin sensitivity analysis module includes:

[0035] The correlation calculation unit is used to calculate the unit dose I in the insulin injection record according to the preset sliding window step size. n The Pearson correlation coefficient of the blood glucose change rate curve in the corresponding window is as follows:

[0036]

[0037] Where ΔCGM(i) is the blood glucose change rate in the i-th window, m is the number of windows; I n (i): unit dose of the nth insulin injection within the i-th sliding window; The mean dose of the nth insulin injection within the sliding window; The global mean of the blood glucose change rate within the sliding window; r n : correlation coefficient between the nth insulin injection and blood glucose changes;

[0038] Matrix generation unit for Pearson correlation coefficient r n The sliding window mean of generates the insulin sensitivity factor matrix, and the matrix elements are defined as:

[0039]

[0040] Where, t i and tj is the endpoint of the time interval; n∈[t i ,t j ]; N is the total number of correlation coefficients in the window; M ij : The element in the i-th row and j-th column of the insulin sensitivity factor matrix. The correlation calculation unit uses a sliding window step size to calculate the real-time Pearson correlation coefficient between insulin injection dose and blood glucose change rate. This breaks through the traditional static analysis framework's reliance on single-point data and captures the differences in patients' insulin responses under different metabolic states through a continuous time window.

[0041] The matrix generation unit further converts the mean correlation coefficient within the sliding window into a multidimensional sensitivity factor matrix. By partitioning the time interval endpoints and aggregating the window data, it forms a sensitivity profile covering different time spans. The temporal resolution advantage of the sliding window mechanism complements the spatial dimensional expansion of the matrix generation. The former ensures sensitive capture of instantaneous insulin effects, while the latter enables holistic sensitivity assessment across time periods.

[0042] The dynamic calculation of the Pearson correlation coefficient reveals the nonlinear relationship between insulin dose and blood glucose change rate, while the multidimensional expression of the matrix structure provides parseable spatiotemporal feature input for the deep learning model.

[0043] In this way, the interaction between sliding window dynamic time series analysis and multidimensional matrix spatial modeling forms a spatiotemporal feature coupling mechanism, effectively overcoming the bias in dosage recommendations caused by traditional methods that ignore circadian fluctuations in individual sensitivity. This enables the system to identify significant differences in insulin sensitivity between breakfast and dinner, accurately matching dosage adjustments to human metabolic rhythms. Furthermore, the combination of sliding window and matrix generation techniques can transform discrete injection events into continuous sensitivity evolution curves, providing a dynamically updated physiological state parameter foundation for personalized medical recommendations.

[0044] In some possible implementations, the nutrition analysis module includes:

[0045] The entity recognition unit is used to extract the carbohydrate content value C from the diet log text data using the named entity recognition algorithm k , the entity recognition formula is:

[0046]

[0047] Where x i is the text sequence feature vector, y i is the entity tag sequence; y′: any tag in the set of all possible entity tags; n: the length of the text sequence;

[0048] Normalization unit, used to convert carbohydrate content value Ck Normalized calculation is performed according to the weight distribution of the influencing time window, and the carbohydrate equivalent time distribution vector is output. The formula is:

[0049]

[0050] Where, K: total number of dietary events; C k : Carbohydrate content value t∈[t k ,t k +T]. Among them, the named entity recognition algorithm accurately extracts carbohydrate content values from unstructured diet logs, solving the problem of data omission or misreporting that is prone to occur in traditional manual records, and ensuring the reliability of the original data of nutritional analysis. After the extracted carbohydrate values are input into the normalization unit, they are combined with the weight distribution function of the metabolic impact time window for dynamic weighted calculation to convert discrete dietary events into nutritional intake curves in a continuous time dimension. The accuracy of named entity recognition directly affects the basic data quality of the normalization calculation, while the time distribution model of the metabolic weight determines the modeling accuracy of the dynamic characteristics of nutrient absorption. The synergistic effect of the two technical features enables carbohydrate intake to be accurately quantified at the numerical level and to match the actual metabolic rhythm of the human body in the time dimension, forming nutritional metabolic characterization data in both time and space dimensions. The generated carbohydrate equivalent time distribution vector is used as a key input parameter, and multimodally fused with the physiological response characteristics output by the insulin sensitivity analysis module to effectively improve the spatiotemporal resolution of the blood glucose change prediction model. In this way, the system can identify the dynamic balance between carbohydrate intake and insulin action within a specific period of time, thereby accurately distinguishing between the two different causes of insufficient nutritional intake and excessive insulin injection in hypoglycemia risk prediction, and providing differentiated decision-making basis for personalized service recommendations.

[0051] In some possible implementations, the feature fusion module includes:

[0052] The first mapping unit is used to map the blood glucose drift trend coefficient α to a first characteristic dimension space vector V1. The mapping formula is:

[0053] V1=tanh(W1α+b1)

[0054] The second decomposition unit is used to decompose the insulin sensitivity factor matrix M into the second characteristic dimension space vector V2. The decomposition formula is:

[0055] V2=SVD(M) T ·U

[0056] Where SVD(M) is the result of matrix singular value decomposition, and U is the dimension reduction projection matrix;

[0057] The tensor concatenation unit is used to concatenate the carbohydrate equivalent time distribution vector C(t) with the first feature dimension space vector V1 and the second feature dimension space vector V2 to generate a fused feature vector F. The formula is:

[0058] F = [V1; V2; C(t)]. The first mapping unit uses the tanh activation function to map the blood glucose excursion trend coefficient to the feature dimensional space. This nonlinear transformation preserves the gradual characteristics of the blood glucose fluctuation amplitude and suppresses the dimensional differences of the eigenvalues. The second decomposition unit performs singular value decomposition on the insulin sensitivity factor matrix. The dimensionality reduction projection matrix extracts the principal component features of the insulin action timeliness, eliminating noise interference in the dose-response relationship. A tensor concatenation operation cross-modally fuses the blood glucose excursion eigenvector, the insulin sensitivity eigenvector, and the carbohydrate equivalent time distribution vector, constructing a composite feature space that encompasses the three-dimensional relationships among metabolic dynamics, drug effects, and nutrient intake.

[0059] The nonlinear encoding of the blood glucose excursion trend coefficient captures the second-order derivative information of the blood glucose curve, revealing the potential characteristics of the hypoglycemia inflection point; the principal component analysis of insulin sensitivity extracts the individual's differentiated response patterns to rapid-acting / long-acting insulin; and the temporal distribution of the carbohydrate equivalent vector quantifies the cumulative effect of carbohydrate intake at different time periods on blood glucose fluctuations. The composite feature vector formed by the fusion of the three not only characterizes the static parameters of the current metabolic state, but also embeds the dynamic trajectory of multi-factor interactions, providing the recurrent neural network with input data with temporal continuity and spatial completeness. During the feature fusion process, the synergistic effect of nonlinear transformation and matrix decomposition eliminates the modeling bias of a single data source, and the dimensional expansion mechanism of tensor splicing breaks through the information capacity limitations of traditional linear weighted fusion.

[0060] In some possible implementations, the recommendation generation module includes:

[0061] The stress state judgment unit is used to judge the current autonomic nervous system stress state according to the real-time spectrum energy distribution vector E(t). The state classification formula is:

[0062]

[0063] Where θ is the spectrum energy threshold, which is used to distinguish the autonomic nervous system activation state;

[0064] The blood glucose slope prediction unit is used to predict the blood glucose drop slope k in the next hour by combining the blood glucose fluctuation cycle feature vector β. The prediction formula is:

[0065] k=LSTM(β)

[0066] Where LSTM is the long short-term memory network layer in the pre-trained temporal attention recurrent neural network model;

[0067] Ratio adjustment unit, used to adjust the fiber intake intensity value F d Adjust the calorie ratio of the carbohydrate supplementation plan to generate a personalized service recommendation set R:

[0068]

[0069] Where ΔG is the target blood glucose difference, C max It is the maximum allowable value for a single supplement. Among them, the stress state judgment unit thresholds the real-time spectrum energy distribution vector, directly identifies the physiological stress signal caused by autonomic nervous activation, and provides immediate state input for subsequent regulation. The blood glucose slope prediction unit relies on the pre-trained temporal attention recurrent neural network model and combines the blood glucose fluctuation cycle feature vector to generate a continuous prediction value of the blood glucose drop slope in the next hour. Its prediction process integrates historical fluctuation patterns and current metabolic trends. The ratio adjustment unit jointly analyzes the autonomic nervous stress state classification results and the predicted blood glucose slope, and dynamically optimizes the ratio relationship between carbohydrate supplementation and insulin dosage adjustment based on the dietary fiber intake intensity value, so as to achieve real-time coordination of nutritional intervention and drug regulation.

[0070] In addition, the identification of autonomic nervous system stress states provides immediate environmental parameter calibration for blood glucose slope prediction, preventing the incorrect application of resting metabolic patterns to stressful situations. Continuous predictions of blood glucose slopes inversely constrain the calorie supplementation threshold in the ratio adjustment, preventing secondary blood glucose fluctuations caused by excessive carbohydrate intake. Dietary fiber intake intensity is included as an independent variable in the ratio calculation, ensuring that the nutritional supplementation plan is consistent with individual digestion and absorption characteristics. This enables the system to distinguish between exercise-induced stress hypoglycemia and metabolic imbalances caused by drug overdose, generating differentiated treatment plans for different triggers, maintaining blood glucose stability while avoiding the risk of excessive intervention.

[0071] In some possible implementations, the risk assessment module includes:

[0072] The interpolation processing unit is used to perform cubic spline interpolation on the missing time points in the diet time series data stream to generate a continuous time distribution vector C′(t). The interpolation formula is:

[0073] C′(t)=a(tt k ) 3 +b(tt k ) 2 +c(tt k )+dwhere t∈[t k ,t k+1 ]

[0074] Where a, b, c, d are the interpolation coefficients calculated by the cubic spline interpolation algorithm; tk : the time of occurrence of the kth dietary event; t k+1 : the occurrence time of the k+1th dietary event;

[0075] The attention calculation unit is used to input the interpolated fusion feature vector F′=[V1; V2; C′(t)] into the pre-trained temporal attention recurrent neural network. The attention weight calculation formula is:

[0076]

[0077] Where h t-1 is the hidden state of the attention recurrent neural network at time t-1; V a ,W a : Trainable parameter matrix of attention mechanism; F′ t : The interpolated fused feature vector at time t, obtained by interpolating the fused feature vector F;

[0078] The probability generation unit is used to calculate the risk probability distribution P(t) at each time point within the preset time period in the future through the hidden layer output of the temporal attention recurrent neural network. The formula is:

[0079] P(t)=σ(W p ·(a t ⊙h t )+b p )

[0080] Where σ is the sigmoid function; W p , b p : The trainable weight matrix and bias vector of the output layer; h t : The hidden state of the temporal attention recurrent neural network at time t;

[0081] The probability output unit is used to take the maximum value of the risk probability distribution P(t) as the hypoglycemia risk probability value.

[0082] In this way, the synergistic effect of the cubic spline interpolation algorithm and the attention mechanism can address the problem of hypoglycemia risk assessment bias caused by incomplete dietary data for diabetic patients. Specifically, the cubic spline interpolation algorithm fills in missing time points in the dietary time series data stream, generating a continuous time distribution vector to ensure data integrity. The interpolation coefficients are calculated using the cubic spline algorithm, constructing a smooth curve with continuous second-order derivatives within the time interval of adjacent dietary events. This effectively restores the dynamic characteristics of the actual metabolic process and avoids the prediction abrupt errors caused by traditional linear interpolation. The temporal attention recurrent neural network receives the interpolated fused feature vector and dynamically assigns feature importance to different time nodes using the attention weight calculation formula. The pre-trained attention mechanism parameter matrix extracts the correlation between the hidden layer state and the input features through nonlinear transformation, enabling the model to automatically focus on key periods of blood sugar fluctuations. The probability generation unit combines the attention-weighted hidden state to output a risk probability distribution. This is compressed to a clinically interpretable range using the sigmoid function, ultimately outputting the maximum probability value as the basis for risk assessment.

[0083] The joint application of interpolation processing and attention mechanism forms a complementary enhancement effect: the interpolation algorithm ensures the spatiotemporal continuity of the input data, providing a reliable feature basis for the attention mechanism; the dynamic adjustment of attention weights enhances the model's ability to capture the implicit physiological patterns in the interpolated data.

[0084] In some possible implementations, the personalized medical and health service recommendation system based on big data analysis further includes:

[0085] An interaction module, configured to convert the personalized service recommendation set into interactive instructions executable by the mobile terminal and send the instructions to the user device;

[0086] The feedback optimization module is used to collect the real-time blood sugar changes after the user executes the recommendation, generate the model parameter correction vector and feed it back to the recurrent neural network model.

[0087] The interactive module ensures the immediate actionability of recommendations, shortening the response chain from decision-making to execution, and ensuring the clinical implementation efficiency of recommendation services. The feedback optimization module leverages real-world blood glucose data to capture the dynamic evolution of individual metabolic characteristics and gradually eliminate model prediction bias. Physiological feedback data generated by executing recommendations is reinjected into the training process, enabling the neural network model to adapt to long-term variables such as fluctuations in a patient's insulin sensitivity and changes in dietary habits, continuously improving the spatiotemporal accuracy of risk prediction.

[0088] In some possible implementations, the feedback optimization module includes:

[0089] Residual calculation unit, used to calculate real-time blood sugar change ΔG real The predicted change before the recommendation is executed ΔGpred The residual value e between is:

[0090] e=ΔG real -ΔG pred

[0091] Gradient update unit, used to update the hidden layer weight W of the recurrent neural network model according to the residual value e h Perform back propagation gradient calculation to generate the model parameter correction vector ΔW, the formula is:

[0092]

[0093] Where η is the learning rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 A schematic diagram of the structure of a personalized medical and health service recommendation system based on big data analysis provided in some embodiments of the present application;

[0095] Figure 2 for Figure 1 The diagram shows the structure of the synchronization module in the personalized medical and health service recommendation system based on big data analysis;

[0096] Figure 3 for Figure 1 The schematic diagram of the structure of the parsing module in the personalized medical and health service recommendation system based on big data analysis is shown;

[0097] Figure 4 for Figure 1 The diagram shows the structure of the blood glucose feature extraction module in the personalized medical and health service recommendation system based on big data analysis;

[0098] Figure 5 for Figure 1 The schematic diagram of the structure of the insulin sensitivity analysis module in the personalized medical and health service recommendation system based on big data analysis is shown;

[0099] Figure 6 for Figure 1 The diagram shows the structure of the nutrition analysis module in the personalized medical and health service recommendation system based on big data analysis;

[0100] Figure 7 for Figure 1 The diagram shows the structure of the feature fusion module in the personalized medical and health service recommendation system based on big data analysis;

[0101] Figure 8 for Figure 1 The diagram shows the structure of the risk assessment module in the personalized medical and health service recommendation system based on big data analysis;

[0102] Figure 9 for Figure 1 The diagram shows the structure of the recommendation generation module in the personalized medical and health service recommendation system based on big data analysis;

[0103] Figure 10 for Figure 1 The figure shows the structural diagram of the feedback optimization module in the personalized medical and health service recommendation system based on big data analysis. DETAILED DESCRIPTION

[0104] Specific embodiments of the present invention will now be mentioned in detail. Although the present invention is described in conjunction with these specific embodiments, it should be appreciated that the present invention is not intended to be limited to these specific embodiments. On the contrary, these embodiments are intended to cover substitutions, changes or equivalent embodiments that may be included in the spirit and scope of the invention defined by the claims. In the following description, a large amount of specific details are set forth to provide a comprehensive understanding of the present invention. The present invention can be implemented without some or all of these specific details. In other cases, in order not to make the present invention unnecessarily obscure, well-known process operations are not described in detail.

[0105] When used in conjunction with "including," "methods comprising," or similar language in this specification and the appended claims, the singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0106] See also Figure 1 The present invention provides a personalized medical and health service recommendation system based on big data analysis, including:

[0107] Acquisition module 1, for acquiring the first user (diabetic patient)'s dynamic blood glucose monitoring data, insulin injection records, diet log text data and original heart rate variability signal;

[0108] Among them, dynamic blood glucose monitoring data can monitor the interstitial fluid glucose concentration in real time through a subcutaneously implanted continuous blood glucose monitoring sensor, automatically generating a set of blood glucose values every 5 minutes, and the original heart rate variability signal can be collected through wearable devices; insulin injection records can be manually entered by mobile applications; diet log text data can be recorded by patients independently through mobile phone applications or photographed for identification.

[0109] Synchronization module 2 is used to synchronize the dynamic blood glucose monitoring data and the original heart rate variability signal with minute-level timestamps to generate a standardized data stream;

[0110] Parsing module 3 is used to parse the intake event timestamps in the diet log text data, associate each diet event with a preset metabolic impact time window, and generate a diet time series data stream;

[0111] Blood glucose feature extraction module 4, used to extract blood glucose change rate curve, blood glucose drift trend coefficient and blood glucose fluctuation period feature vector from the standardized data stream;

[0112] Insulin sensitivity analysis module 5, used to generate an insulin sensitivity factor matrix based on the time series of insulin injection records and the blood glucose change rate curve;

[0113] Nutritional analysis module 6, used to perform natural language processing on the dietary time series data stream to parse out the carbohydrate equivalent time distribution vector and dietary fiber intake intensity value;

[0114] A feature fusion module 7 is used to generate a fusion feature vector based on the blood glucose drift trend coefficient, the insulin sensitivity factor matrix and the carbohydrate equivalent time distribution vector;

[0115] The risk assessment module 8 is used to process the fused feature vector using a recurrent neural network model to generate a hypoglycemia risk probability value;

[0116] The recommendation generating module 9 is configured to generate a personalized service recommendation set when the hypoglycemia risk probability value exceeds a preset threshold.

[0117] The system's acquisition module 1 breaks down data barriers between devices, integrating high-frequency CGM data (minute-level), event-triggered insulin injection records, unstructured dietary log text, and continuously collected heart rate variability signals into a unified processing workflow. Synchronization module 2 then aligns CGM and heart rate variability signals with minute-level timestamps, eliminating the timescale misalignment caused by the traditional 24-hour fixed window. This allows the correlation of physiological responses from different frequency data to be refined from a coarse "day-level" to a fine-grained "minute-level" level. Parsing module 3 exploits the sparsity of dietary logs by converting discrete dietary events into a continuous time-series data stream using a pre-set metabolic impact time window, breaking away from the traditional mechanistic modeling approach of simplifying dietary impact into fixed time intervals. Based on this foundation, feature fusion module 7 extracts dynamic blood glucose characteristics, insulin sensitivity factors, and nutrient metabolism profiles to construct a multidimensional, real-time representation of physiological status. Finally, a recurrent neural network implements dynamic probabilistic prediction of hypoglycemia risk. When the risk probability exceeds a threshold, recommendation generation module 9 immediately triggers a personalized intervention strategy, forming a millisecond-level closed loop from data acquisition to decision execution.

[0118] In this way, the data processing paradigm of the traditional health management system is fundamentally reconstructed, the asynchrony of isolated data sources is transformed into the synergy of time-series correlation, and static historical data analysis is upgraded to dynamic real-time status perception, thereby effectively suppressing the deviation between recommended services and actual physiological status, and achieving clinically acceptable intervention timeliness in the scenario of asymptomatic hypoglycemia.

[0119] See also Figure 2 In some embodiments, the synchronization module 2 includes:

[0120] The signal conversion unit 21 is used to perform wavelet transform on the original heart rate variability signal to generate a real-time spectrum energy distribution vector, wherein the wavelet transform formula is:

[0121]

[0122] Where x(t) is the original heart rate variability signal, ψ(t) is the Morlet mother wavelet function, a is the scale factor, b is the time translation factor; t is the time variable;

[0123] The time alignment unit 22 is used to align the continuous blood glucose monitoring data with the real-time spectrum energy distribution vector according to minute-level timestamps to generate a standardized data stream, wherein the alignment formula is:

[0124] S(t)={CGM(t i ),HRV(t j )||t i -t j |<∈

[0125] Where, ∈≥2.5 minutes, CGM(t i ) is the time t i Blood glucose level, HRV(t j ) is the time t jspectral energy values. Thus, the signal conversion unit 21 performs a wavelet transform on the raw heart rate variability signal to generate a real-time spectral energy distribution vector. The time-frequency localization properties of the wavelet transform are leveraged to accurately capture transient changes in the state of autonomic nervous system regulation. Continuous blood glucose monitoring data is collected at fixed intervals, while the heart rate variability signal is transformed through a wavelet transform to generate a real-time spectral energy distribution vector. The two differ in data structure and temporal granularity. The time alignment unit 22 synchronizes the dynamic blood glucose monitoring data with the real-time spectral energy distribution vector using minute-level timestamps, establishing a unified time base and ensuring real-time matching of blood glucose levels with autonomic nervous system activity. The multi-scale analysis capabilities of the wavelet transform, combined with the minute-level synchronization mechanism, enable a precise temporal mapping of the spectral features of the high-frequency heart rate variability signal with the relatively low-frequency blood glucose fluctuation data, revealing the dynamic correlation between blood glucose changes and autonomic nervous system stress responses. The synchronized, standardized data stream provides time-consistent multimodal physiological features for subsequent analysis, enabling the system to simultaneously track the coordinated changes in blood glucose trends and neural regulation. For example, when the spectral energy of the heart rate variability signal increases abnormally, combined with the rate of blood glucose decline at the synchronized time point, the risk of hypoglycemia caused by autonomic hyperactivity can be identified in advance. The synergistic effect of signal transformation and time alignment overcomes the limitations of traditional systems in modeling physiological associations due to data timeline deviations, establishing a highly accurate time synchronization foundation for real-time risk assessment.

[0126] See also Figure 3 In some embodiments, the parsing module 3 includes:

[0127] The log extraction unit 31 is used to extract the food intake time point t in the diet log text data. k and description of food type;

[0128] The metabolic weight calculation unit 32 is used to calculate the metabolic weight value of each food type within the impact time window according to the pharmacokinetic model. The specific formula is:

[0129] where t∈[t k ,t k +T]

[0130] Where σ represents the standard deviation of metabolic rate, σ≤T / 4; t k The time when the dietary event occurs; T: the duration of the preset metabolic impact time window; illustratively, T can be but is not limited to 4 hours.

[0131] The data stream generation unit 33 is used to bind the metabolic weight value w(t) to the corresponding timestamp to generate a dietary time series data stream containing the time weight distribution. In this way, the precise analysis of the dietary log and the dynamic metabolic impact modeling can be achieved through the synergistic effect of the log extraction unit 31, the metabolic weight calculation unit 32 and the data stream generation unit 33. Specifically, the log extraction unit 31 extracts the food intake time point and type description from the unstructured dietary log text to provide basic data support for the metabolic weight calculation. The metabolic weight calculation unit 32 is based on the pharmacokinetic model and uses a Gaussian distribution function to quantify the dynamic weight of the food type within the preset metabolic impact time window. The standard deviation parameter is used to control the individual differences in metabolic rate, breaking through the traditional fixed time window's simplified assumption of the nutrient absorption process. The data stream generation unit 33 dynamically binds the metabolic weight distribution to the timestamp, converts discrete dietary events into a continuous time series data stream, and fully depicts the absorption and metabolic trajectory of nutrients such as carbohydrates.

[0132] In this way, the delayed effect, duration and individual metabolic differences of dietary events on blood glucose fluctuations can be accurately quantified, providing high-fidelity time-series input data for subsequent blood glucose prediction models, thereby effectively suppressing blood glucose prediction errors caused by deviations in dietary metabolic modeling.

[0133] See also Figure 4 In some embodiments, the blood glucose feature extraction module 4 specifically includes:

[0134] The blood glucose rate calculation unit 41 is used to extract a sequence of blood glucose monitoring values with continuous time stamps from the standardized data stream, calculate the blood glucose difference between adjacent time points and perform secondary difference processing to generate a blood glucose change rate curve representing the change in squared acceleration per unit time;

[0135] The drift trend analysis unit 42 is used to perform exponential weighted integration processing on the blood glucose change rate curve, implement dynamic attenuation weighting on the historical data based on a preset attenuation factor, and generate a blood glucose drift trend coefficient reflecting the cumulative effect of blood glucose drift;

[0136] The periodic feature generating unit 43 is used to perform frequency domain transformation on the blood glucose monitoring value in the standardized data stream, extract the energy distribution characteristics within the preset physiological related frequency band, and generate a blood glucose fluctuation periodic feature vector containing the main fluctuation frequency component.

[0137] See also Figure 5 In some embodiments, the insulin sensitivity analysis module 5 includes:

[0138] The correlation calculation unit 51 is used to calculate the unit dose I in the insulin injection record according to the preset sliding window step size. nThe Pearson correlation coefficient of the blood glucose change rate curve in the corresponding window is as follows:

[0139]

[0140] Where ΔCGM(i) is the blood glucose change rate in the i-th window, m is the number of windows; I n (i): unit dose of the nth insulin injection within the i-th sliding window; The mean dose of the nth insulin injection within the sliding window; The global mean of the blood glucose change rate within the sliding window; r n : correlation coefficient between the nth insulin injection and blood glucose changes;

[0141] Exemplarily, the preset sliding window step size may be, but is not limited to, 30 minutes.

[0142] The matrix generating unit 52 is used to generate the matrix according to the Pearson correlation coefficient r n The sliding window mean of generates the insulin sensitivity factor matrix, and the matrix elements are defined as:

[0143]

[0144] Where, t i and t j is the endpoint of the time interval; n∈[t i ,t j ]; N is the total number of correlation coefficients in the window; M ij : The element in the i-th row and j-th column of the insulin sensitivity factor matrix. Correlation calculation unit 51 uses a sliding window step size to perform real-time Pearson correlation coefficient calculations between insulin injection dose and blood glucose change rate. This breaks through the traditional static analysis framework's reliance on single-point time data and captures the differences in patients' insulin responses under different metabolic states through a continuous time window.

[0145] Matrix generation unit 52 further converts the mean correlation coefficient within the sliding window into a multidimensional sensitivity factor matrix. By dividing the time interval endpoints and aggregating the window data, it forms a sensitivity feature map covering different time spans. The temporal resolution advantage of the sliding window mechanism complements the spatial dimensional expansion of the matrix generation. The former ensures sensitive capture of instantaneous insulin effects, while the latter enables holistic sensitivity assessment across time periods.

[0146] The dynamic calculation of the Pearson correlation coefficient reveals the nonlinear relationship between insulin dose and blood glucose change rate, while the multidimensional expression of the matrix structure provides parseable spatiotemporal feature input for the deep learning model.

[0147] In this way, the interaction between sliding window dynamic time series analysis and multidimensional matrix spatial modeling forms a spatiotemporal feature coupling mechanism, effectively overcoming the bias in dosage recommendations caused by traditional methods that ignore circadian fluctuations in individual sensitivity. This enables the system to identify significant differences in insulin sensitivity between breakfast and dinner, accurately matching dosage adjustments to human metabolic rhythms. Furthermore, the combination of sliding window and matrix generation techniques can transform discrete injection events into continuous sensitivity evolution curves, providing a dynamically updated physiological state parameter foundation for personalized medical recommendations.

[0148] See also Figure 6 In some embodiments, the nutrition analysis module 6 includes:

[0149] The entity recognition unit 61 is used to extract the carbohydrate content value C from the diet log text data using a named entity recognition algorithm. k , the entity recognition formula is:

[0150]

[0151] Where x i is the text sequence feature vector, y i is the entity tag sequence; y′: any tag in the set of all possible entity tags; n: the length of the text sequence;

[0152] Normalization unit 62 is used to convert the carbohydrate content value C k Normalized calculation is performed according to the weight distribution of the influencing time window, and the carbohydrate equivalent time distribution vector is output. The formula is:

[0153]

[0154] Where, K: total number of dietary events; C k : Carbohydrate content value t∈[t k ,t k+T]. Among them, the named entity recognition algorithm accurately extracts carbohydrate content values from unstructured diet logs, solves the problem of data omission or misreporting that is prone to occur in traditional manual records, and ensures the reliability of the original data of nutritional analysis. After the extracted carbohydrate values are input into the normalization unit 62, they are dynamically weighted and calculated in combination with the weight distribution function of the metabolic impact time window to convert discrete dietary events into nutritional intake curves in a continuous time dimension. The accuracy of named entity recognition directly affects the basic data quality of the normalization calculation, while the time distribution model of the metabolic weight determines the modeling accuracy of the dynamic characteristics of nutrient absorption. The synergistic effect of the two technical features enables carbohydrate intake to be accurately quantified at the numerical level and to match the actual metabolic rhythm of the human body in the time dimension, forming nutritional metabolic characterization data in both time and space dimensions. The generated carbohydrate equivalent time distribution vector is used as a key input parameter, and is multimodally fused with the physiological response characteristics output by the insulin sensitivity analysis module 5 to effectively improve the spatiotemporal resolution of the blood glucose change prediction model. In this way, the system can identify the dynamic balance between carbohydrate intake and insulin action within a specific period of time, thereby accurately distinguishing between the two different causes of insufficient nutritional intake and excessive insulin injection in hypoglycemia risk prediction, and providing differentiated decision-making basis for personalized service recommendations.

[0155] See also Figure 7 In some embodiments, the feature fusion module 7 includes:

[0156] The first mapping unit 71 is used to map the blood glucose drift trend coefficient α to a first characteristic dimension space vector V1. The mapping formula is:

[0157] V1=tanh(W1α+b1)

[0158] The second decomposition unit 72 is used to decompose the insulin sensitivity factor matrix M into a second characteristic dimension space vector V2. The decomposition formula is:

[0159] V2=SVD(M) T ·U

[0160] Where SVD(M) is the result of matrix singular value decomposition, and U is the dimension reduction projection matrix;

[0161] The tensor concatenation unit 73 is used to concatenate the carbohydrate equivalent time distribution vector C(t) with the first feature dimension space vector V1 and the second feature dimension space vector V2 to generate a fused feature vector F. The formula is:

[0162] F = [V1; V2; C(t)]. The first mapping unit 71 uses the tanh activation function to map the blood glucose excursion trend coefficient to the feature dimensional space. The nonlinear transformation process preserves the gradual characteristics of the blood glucose fluctuation amplitude and suppresses the dimensional differences of the eigenvalues. The second decomposition unit 72 performs singular value decomposition on the insulin sensitivity factor matrix, extracting the principal component characteristics of the insulin action timeliness through the dimensionality reduction projection matrix, and eliminating noise interference in the dose-response relationship. The tensor splicing operation cross-modally fuses the blood glucose excursion eigenvector, the insulin sensitivity eigenvector, and the carbohydrate equivalent time distribution vector to construct a composite feature space that contains the three-dimensional correlation between metabolic dynamics, drug effects, and nutritional intake.

[0163] The nonlinear encoding of the blood glucose excursion trend coefficient captures the second-order derivative information of the blood glucose curve, revealing the potential characteristics of the hypoglycemia inflection point; the principal component analysis of insulin sensitivity extracts the individual's differentiated response patterns to rapid-acting / long-acting insulin; and the temporal distribution of the carbohydrate equivalent vector quantifies the cumulative effect of carbohydrate intake at different time periods on blood glucose fluctuations. The composite feature vector formed by the fusion of the three not only characterizes the static parameters of the current metabolic state, but also embeds the dynamic trajectory of multi-factor interactions, providing the recurrent neural network with input data with temporal continuity and spatial completeness. During the feature fusion process, the synergistic effect of nonlinear transformation and matrix decomposition eliminates the modeling bias of a single data source, and the dimensional expansion mechanism of tensor splicing breaks through the information capacity limitations of traditional linear weighted fusion.

[0164] See also Figure 8 In some embodiments, the risk assessment module 8 includes:

[0165] The interpolation processing unit 81 is used to perform cubic spline interpolation on the missing time points in the diet time series data stream to generate a continuous time distribution vector C′(t). The interpolation formula is:

[0166] C′(t)=a(tt k ) 3 +b(tt k ) 2 +c(tt k )+dwhere t∈[t k ,t k+1 ]

[0167] Where a, b, c, d are the interpolation coefficients calculated by the cubic spline interpolation algorithm; t k : the time of occurrence of the kth dietary event; t k+1 : the occurrence time of the k+1th dietary event;

[0168] The attention calculation unit 82 is used to input the interpolated fusion feature vector F′=[V1; V2; C′(t)] into the pre-trained temporal attention recurrent neural network. The attention weight calculation formula is:

[0169]

[0170] Where h t-1 is the hidden state of the attention recurrent neural network at time t-1; V a ,W a : Trainable parameter matrix of attention mechanism; F′ t : The interpolated fused feature vector at time t, obtained by interpolating the fused feature vector F;

[0171] The probability generation unit 83 is used to calculate the risk probability distribution P(t) at each time point within a preset time period in the future through the hidden layer output of the temporal attention recurrent neural network. The formula is:

[0172] P(t)=σ(W p ·(a t ⊙h t )+b p )

[0173] Where σ is the sigmoid function; W p , b p : The trainable weight matrix and bias vector of the output layer; h t : The hidden state of the temporal attention recurrent neural network at time t;

[0174] Specifically, the future preset duration may be, but is not limited to, 2 hours.

[0175] The probability output unit 84 is configured to use the maximum value of the risk probability distribution P(t) as the hypoglycemia risk probability value.

[0176] In this way, the synergistic effect of the cubic spline interpolation algorithm and the attention mechanism can address the problem of hypoglycemia risk assessment bias caused by incomplete dietary data for diabetic patients. Specifically, the cubic spline interpolation algorithm fills in missing time points in the dietary time series data stream, generating a continuous time distribution vector to ensure data integrity. The interpolation coefficients are calculated using the cubic spline algorithm, constructing a smooth curve with continuous second-order derivatives within the time interval of adjacent dietary events, effectively restoring the dynamic characteristics of the actual metabolic process and avoiding the prediction mutation errors caused by traditional linear interpolation. The temporal attention recurrent neural network receives the interpolated fused feature vector and dynamically assigns the feature importance of different time nodes using the attention weight calculation formula. The pre-trained attention mechanism parameter matrix extracts the correlation between the hidden layer state and the input features through nonlinear transformation, allowing the model to automatically focus on key periods of blood sugar fluctuations. The probability generation unit 83 combines the attention-weighted hidden state to output a risk probability distribution. This is compressed to a clinically interpretable range using the sigmoid function, and the maximum probability value is ultimately output as the basis for risk assessment.

[0177] The joint application of interpolation processing and attention mechanism forms a complementary enhancement effect: the interpolation algorithm ensures the spatiotemporal continuity of the input data, providing a reliable feature basis for the attention mechanism; the dynamic adjustment of attention weights enhances the model's ability to capture the implicit physiological patterns in the interpolated data.

[0178] See also Figure 9 In some embodiments, the recommendation generation module 9 includes:

[0179] The stress state judgment unit 91 is used to judge the current autonomic nervous system stress state according to the real-time spectrum energy distribution vector E(t). The state classification formula is:

[0180]

[0181] Where θ is the spectrum energy threshold, which is used to distinguish the autonomic nervous system activation state;

[0182] The blood glucose slope prediction unit 92 is used to predict the blood glucose drop slope k in the next hour based on the blood glucose fluctuation period characteristic vector β. The prediction formula is:

[0183] k=LSTM(β)

[0184] Where LSTM is the long short-term memory network layer in the pre-trained temporal attention recurrent neural network model;

[0185] The ratio adjustment unit 93 is used to adjust the ratio of the dietary fiber intake intensity value F d Adjust the calorie ratio of the carbohydrate supplementation plan to generate a personalized service recommendation set R:

[0186]

[0187] Where ΔG is the target blood glucose difference, C max It is the maximum allowable value for a single supplement. Among them, the stress state judgment unit 91 performs thresholding processing on the real-time spectrum energy distribution vector, directly identifies the physiological stress signal caused by autonomic nervous activation, and provides immediate state input for subsequent regulation. The blood glucose slope prediction unit 92 relies on the pre-trained temporal attention recurrent neural network model and combines the blood glucose fluctuation cycle feature vector to generate a continuous prediction value of the blood glucose drop slope in the next hour. Its prediction process integrates the historical fluctuation pattern and the current metabolic trend. The ratio adjustment unit 93 jointly analyzes the autonomic nervous stress state classification result and the predicted blood glucose slope, and dynamically optimizes the ratio relationship between the carbohydrate supplement amount and the insulin dose adjustment amount based on the dietary fiber intake intensity value, so as to realize the real-time coordination of nutritional intervention and drug regulation.

[0188] In addition, the identification of autonomic nervous system stress states provides immediate environmental parameter calibration for blood glucose slope prediction, preventing the incorrect application of resting metabolic patterns to stressful situations. Continuous predictions of blood glucose slopes inversely constrain the calorie supplementation threshold in the ratio adjustment, preventing secondary blood glucose fluctuations caused by excessive carbohydrate intake. Dietary fiber intake intensity is included as an independent variable in the ratio calculation, ensuring that the nutritional supplementation plan is consistent with individual digestion and absorption characteristics. This enables the system to distinguish between exercise-induced stress hypoglycemia and metabolic imbalances caused by drug overdose, generating differentiated treatment plans for different triggers, maintaining blood glucose stability while avoiding the risk of excessive intervention.

[0189] See also Figure 1 In some embodiments, the personalized medical and health service recommendation system based on big data analysis further includes:

[0190] Interaction module 10, used to convert the personalized service recommendation set into interactive instructions executable by the mobile terminal and send it to the user device;

[0191] The feedback optimization module 11 is used to collect the real-time blood sugar change after the user executes the recommendation, generate a model parameter correction vector and feed it back to the recurrent neural network model.

[0192] The interactive module 10 ensures the immediate actionability of recommendations, shortening the response chain from decision-making to execution, and ensuring the clinical implementation efficiency of the recommendation service. The feedback optimization module 11 leverages real-world blood glucose data to capture the dynamic evolution of individual metabolic characteristics and gradually eliminate model prediction bias. The physiological feedback data generated by executing recommendations is reinjected into the training process, enabling the neural network model to adapt to long-term variables such as fluctuations in the patient's insulin sensitivity and changes in dietary habits, continuously improving the spatiotemporal matching accuracy of risk prediction.

[0193] See also Figure 10 In some embodiments, the feedback optimization module 11 includes:

[0194] Residual calculation unit 111, used to calculate the real-time blood sugar change ΔG real The predicted change before the recommendation is executed ΔG pred The residual value e between is:

[0195] e=ΔG real -ΔG pred

[0196] Gradient update unit 112, used to update the hidden layer weight W of the recurrent neural network model according to the residual value e h Perform back propagation gradient calculation to generate the model parameter correction vector ΔW, the formula is:

[0197]

[0198] Where η is the learning rate.

[0199] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, persons skilled in the art will appreciate that modifications to the specific embodiments of the present invention or substitutions of some of the technical features may be made without departing from the spirit of the present invention and are intended to be encompassed within the scope of the technical solutions claimed herein.

Claims

1. A personalized medical and health service recommendation system based on big data analysis, characterized by: include: an acquisition module, configured to acquire the first user's dynamic blood glucose monitoring data, insulin injection records, diet log text data, and original heart rate variability signal; A synchronization module, configured to synchronize the dynamic blood glucose monitoring data and the original heart rate variability signal with minute-level timestamps to generate a standardized data stream; a parsing module, configured to parse intake event timestamps in the diet log text data, associate each diet event with a preset metabolic impact time window, and generate a diet time series data stream; A blood glucose feature extraction module, configured to extract a blood glucose change rate curve, a blood glucose drift trend coefficient, and a blood glucose fluctuation period feature vector from the standardized data stream; An insulin sensitivity analysis module, configured to generate an insulin sensitivity factor matrix based on the time series of insulin injection records and the blood glucose change rate curve; a nutrition analysis module, configured to perform natural language processing on the diet time series data stream to parse out a carbohydrate equivalent time distribution vector and a dietary fiber intake intensity value; a feature fusion module, configured to generate a fusion feature vector based on the blood glucose drift trend coefficient, the insulin sensitivity factor matrix, and the carbohydrate equivalent time distribution vector; a risk assessment module, configured to process the fused feature vector using a recurrent neural network model to generate a hypoglycemia risk probability value; The recommendation generation module is used to generate a personalized service recommendation set when the hypoglycemia risk probability value exceeds a preset threshold.

2. The personalized medical and health service recommendation system based on big data analysis according to claim 1 is characterized in that: The synchronization module includes: The signal conversion unit is used to perform wavelet transform on the original heart rate variability signal to generate a real-time spectrum energy distribution vector, wherein the wavelet transform formula is: Where x(t) is the original heart rate variability signal, ψ(t) is the Morlet mother wavelet function, a is the scale factor, b is the time translation factor; t is the time variable; A time alignment unit is configured to align the dynamic blood glucose monitoring data with the real-time spectrum energy distribution vector according to minute-level timestamps to generate the standardized data stream, wherein the alignment formula is: S(t)={CGM(t i ),HRV(t j )||t i -t j |<∈} Where, ∈≥2.5 minutes, CGM(t i ) is the time t i Blood glucose level, HRV(t j ) is the time t j spectral energy value.

3. The personalized medical and health service recommendation system based on big data analysis according to claim 2 is characterized in that: The parsing module includes: A log extraction unit is used to extract the food intake time point t in the diet log text data. k and description of food type; The metabolic weight calculation unit is used to calculate the metabolic weight value of each food type within the impact time window according to the pharmacokinetic model. The specific formula is: where t∈[t k ,t k +T] Where σ represents the standard deviation of metabolic rate, σ≤T / 4; t k The time when the dietary event occurs; T: the duration of the preset metabolic impact time window; The data stream generating unit is used to bind the metabolic weight value w(t) with the corresponding timestamp to generate the dietary time series data stream containing the time weight distribution.

4. The personalized medical and health service recommendation system based on big data analysis according to claim 3 is characterized in that: The insulin sensitivity analysis module includes: The correlation calculation unit is used to calculate the unit dose I in the insulin injection record according to the preset sliding window step size. n The Pearson correlation coefficient of the blood glucose change rate curve in the corresponding window is as follows: Where ΔCGM(i) is the blood glucose change rate in the i-th window, m is the number of windows; I n (i): unit dose of the nth insulin injection within the i-th sliding window; The mean dose of the nth insulin injection within the sliding window; The global mean of the blood glucose change rate within the sliding window; r n : correlation coefficient between the nth insulin injection and blood glucose changes; Matrix generation unit for the Pearson correlation coefficient r n The sliding window mean of generates the insulin sensitivity factor matrix, and the matrix elements are defined as: Where, t i and t j is the endpoint of the time interval; n∈[t i ,t j ]; N is the total number of correlation coefficients in the window; M ij : The element in the i-th row and j-th column of the insulin sensitivity factor matrix.

5. The personalized medical and health service recommendation system based on big data analysis according to claim 4 is characterized in that: The nutrition analysis module includes: An entity recognition unit is used to extract the carbohydrate content value C from the diet log text data using a named entity recognition algorithm. k , the entity recognition formula is: Where x i is the text sequence feature vector, y i is the entity tag sequence; y′: any tag in the set of all possible entity tags; n: the length of the text sequence; Normalization unit, used to convert the carbohydrate content value C k Normalized calculation is performed according to the weight distribution of the influencing time window, and the carbohydrate equivalent time distribution vector is output. The formula is: Where, K: total number of dietary events; C k : Carbohydrate content value t∈[t k ,t k +T].

6. The personalized medical and health service recommendation system based on big data analysis according to claim 5 is characterized in that: The feature fusion module includes: The first mapping unit is used to map the blood glucose drift trend coefficient α into a first characteristic dimension space vector V1, and the mapping formula is: V1=tanh(W1a+b1) The second decomposition unit is used to decompose the insulin sensitivity factor matrix M into a second characteristic dimension space vector V2, and the decomposition formula is: Where SVD(M) is the result of matrix singular value decomposition, and U is the dimension reduction projection matrix; The tensor splicing unit is used to splice the carbohydrate equivalent time distribution vector C(t) with the first feature dimension space vector V1 and the second feature dimension space vector V2 to generate the fused feature vector F. The formula is: F=[V1;V2;C(t)].

7. The personalized medical and health service recommendation system based on big data analysis according to claim 6, characterized in that: The recommendation generation module includes: The stress state judgment unit is used to judge the current autonomic nervous system stress state according to the real-time spectrum energy distribution vector E(t). The state classification formula is: Where θ is the spectrum energy threshold, which is used to distinguish the autonomic nervous system activation state; The blood glucose slope prediction unit is used to predict the blood glucose drop slope k in the next hour based on the blood glucose fluctuation period characteristic vector β. The prediction formula is: k=LSTM(β) Where LSTM is the long short-term memory network layer in the pre-trained temporal attention recurrent neural network model; The ratio adjustment unit is used to adjust the dietary fiber intake intensity value F d The calorie ratio of the carbohydrate supplementation program is adjusted to generate the personalized service recommendation set R: Where ΔG is the target blood glucose difference, C max This is the maximum allowable value for a single replenishment.

8. The personalized medical and health service recommendation system based on big data analysis according to claim 7 is characterized in that: The risk assessment module includes: The interpolation processing unit is used to perform cubic spline interpolation on the missing time points in the diet time series data stream to generate a continuous time distribution vector C′(t). The interpolation formula is: C′(t) = a(t - t k ) 3 + b(t - t k ) 2 + c(t - t k ) + d where t ∈ [t k , t k+1 ​ Where a, b, c, d are the interpolation coefficients calculated by the cubic spline interpolation algorithm; t k : the time of occurrence of the kth dietary event; t k+1 : the occurrence time of the k+1th dietary event; The attention calculation unit is used to input the interpolated fusion feature vector F′=[V1; V2; C′(t)] into the pre-trained temporal attention recurrent neural network. The attention weight calculation formula is: Where h t-1 is the hidden state of the attention recurrent neural network at time t-1; V a ,W a : Trainable parameter matrix of attention mechanism; F′ t : The interpolated fused feature vector at time t, obtained by interpolating the fused feature vector F; The probability generation unit is used to calculate the risk probability distribution P(t) at each time point within a preset time period in the future through the hidden layer output of the temporal attention recurrent neural network. The formula is: P(t)=σ(W p ·(a t ⊙h t )+b p ) Where σ is the sigmoid function; W p ,b p : The trainable weight matrix and bias vector of the output layer; h t : The hidden state of the temporal attention recurrent neural network at time t; The probability output unit is used to use the maximum value of the risk probability distribution P(t) as the hypoglycemia risk probability value.

9. The personalized medical and health service recommendation system based on big data analysis according to any one of claims 1 to 8, characterized in that: Also includes: An interaction module, configured to convert the personalized service recommendation set into an interactive instruction executable by a mobile terminal and send the instruction to a user device; The feedback optimization module is used to collect the real-time blood sugar change after the user executes the recommendation, generate a model parameter correction vector and feed it back to the recurrent neural network model.

10. The personalized medical and health service recommendation system based on big data analysis according to claim 9, characterized in that: The feedback optimization module includes: Residual calculation unit, used to calculate the real-time blood sugar change ΔG real The predicted change before the recommendation is executed ΔG pred The residual value e between is: e=ΔG real -ΔG pred A gradient updating unit is used to update the hidden layer weight W of the recurrent neural network model according to the residual value e h Perform back propagation gradient calculation to generate the model parameter correction vector ΔW, the formula is: Where η is the learning rate.

Images