Gestational diabetes postpartum follow-up monitoring method and system

Abstract

The invention relates to a postpartum follow-up monitoring method and system for gestational diabetes mellitus, in particular to the field of biochemical sensors, and aims to realize accurate cleaning and anomaly recognition of blood glucose data by fusing multi-source information and spatial-temporal characteristics, realize a personalized calibration mechanism based on context awareness and dynamically optimize data correction by utilizing reinforcement learning, and improve the accuracy of postpartum follow-up monitoring of gestational diabetes mellitus. The accuracy of single-point measurement is remarkably improved, the federal element learning framework aggregates individual experience on the premise of protecting privacy, drives a calibration model to continuously evolve, enhances generalization ability, constructs a credible quality closed loop for block chain evidence storage and intelligent contract monitoring, and automatically triggers verification and feedback, so that the accuracy of the single-point measurement is improved. High-quality data injection is obtained in the abnormal recognition and model optimization links, and a self-perfect monitoring system is formed.

Description

Technical Field

[0001] This invention relates to the field of biochemical sensors, and more specifically, to a method and system for postpartum follow-up monitoring of gestational diabetes mellitus. Background Technology

[0002] Gestational diabetes mellitus (GDM) is a common complication of pregnancy. Postpartum women are at high risk of developing diabetes and require continuous monitoring of their blood glucose levels to assess metabolic recovery and the long-term risk of developing type 2 diabetes. Traditional follow-up methods rely mainly on regular hospital visits for testing. However, postpartum women need to care for their newborns and adjust their physical and mental state, making adherence difficult to guarantee and resulting in a high rate of missed follow-up visits. Portable home blood glucose meters, as a type of biochemical sensor, allow postpartum women to monitor their own blood glucose in their home environment. They work by using enzymes on test strips to react with glucose in the blood to generate an electrical signal, which the instrument then converts into a blood glucose value. However, the monitoring process in a home environment is affected by various factors, such as fluctuations in temperature and humidity, which can affect the test strips. The activity of enzymes and the performance of electronic components in the instrument, coupled with the physical and mental fatigue of postpartum women, can easily affect the standardization of operation. For example, incomplete cleaning of the blood collection site, blood collection before the disinfectant alcohol has completely evaporated, and excessive squeezing of the fingers causing tissue fluid to dilute the blood can all introduce errors. In addition, the physiological changes during pregnancy, such as increased hematocrit, and the metabolic fluctuations during postpartum recovery place special demands on the accuracy of blood glucose meters. Ordinary blood glucose meters may not be able to automatically correct for these physiological differences. If test strips are not stored properly, such as getting damp or expiring, the chemical substances on them may become ineffective, which can also lead to measurement deviations. All these factors combined result in uncertainty in the reliability and accuracy of home blood glucose testing data, making it difficult to use as the sole basis for clinical decision-making.

[0003] The core technical challenge currently facing postpartum follow-up monitoring systems lies in the lack of efficient and automated accuracy verification and calibration mechanisms for home blood glucose monitoring data based on biochemical sensors. Most existing home blood glucose meters only provide spot measurement data, and the devices themselves may not have built-in automatic calibration functions tailored to the physiological characteristics of postpartum women. Their accuracy largely depends on users strictly adhering to operating procedures and regular manual calibration, which is difficult to guarantee consistently in a home setting lacking professional supervision. While continuous glucose monitoring and other solutions can provide more continuous data, calibration requirements and time delays still exist, and the costs are high. Existing systems typically lack mechanisms for verifying and calibrating the data of postpartum women. The ability to automatically compare and correlate home monitoring data with reference standards such as hospital laboratory venous blood test results is insufficient. If the accuracy of the data is questionable, the reliability of risk assessment, hierarchical management, and remote intervention based on it will be greatly reduced. The shortcomings of the existing technology are: data verification relies heavily on the mother's active recording, manual review, or regular return to the hospital for blood sampling for comparison. The process is cumbersome, inefficient, and lacks real-time performance, which cannot meet the needs of long-term, continuous, and reliable monitoring in postpartum follow-up. This urgently requires the implementation of an embedded intelligent calibration system at the technical level, which can combine multi-parameter information to adjust and verify the sensor output in real time or near real time. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a method and system for postpartum follow-up monitoring of gestational diabetes mellitus, thereby resolving the issues raised in the background section.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for postpartum follow-up monitoring of gestational diabetes mellitus, specifically including the following steps: Step S1: After the postpartum mother collects blood glucose data through a portable biochemical sensor, the collected blood glucose data is arranged in chronological order to form a blood glucose data sequence. Based on the spatiotemporal feature fusion and the physiological state perception vector of the mother obtained in real time by the wearable device, the abnormal data identification process is triggered. The local fluctuation pattern features of the blood glucose data sequence are extracted through a one-dimensional convolutional neural network. At the same time, the long-term trend features are captured through a gated recurrent unit network. The physiological state perception vector is introduced as the conditional input of the attention mechanism to generate a dynamic decision function to output the blood glucose data abnormality index, thereby marking suspected abnormal blood glucose data points. Step S2: Start the dynamic calibration model and integrate the suspected abnormal blood glucose data points marked in step S1, the estimated carbohydrate content obtained from the analysis of dietary images, the estimated energy consumption of the motion sensor, and the physiological state perception vector in step S1 to obtain multi-source information features. Process the multi-source information feature vector through a dual-path calibration network. The reinforcement learning agent takes the multi-source information feature vector as its state, the fine-tuning amount of the initial calibration value as its action, and the time-series physiological consistency delayed feedback as its reward function. Output the calibrated blood glucose value and calibration confidence in real time. Step S3: Aggregate the local model parameter increments generated by each terminal in the dynamic calibration model execution in step S2 through the federated meta-learning framework, execute the optimization process of the global meta-calibration model in the central server using the meta-learning update strategy, and distribute the optimized global meta-calibration model parameters to each terminal. Each terminal obtains a personalized calibration model by fine-tuning based on the received global meta-calibration model parameters. Step S4: Utilize a lightweight blockchain network to store calibration records. The calibration records include the hash value of the original blood glucose data, the calibrated blood glucose value output in step S2, the version number of the personalized calibration model obtained in step S3, the summary of the multi-source information feature vector in step S2, the timestamp, and the calibration confidence level output in step S2. The relationship between the on-chain calibration confidence level and the dynamic threshold is continuously monitored through a smart contract. When the calibration confidence level is detected to be continuously lower than the dynamic threshold or there is a statistical difference, a verification request is automatically triggered to guide the two-finger blood measurement or generate laboratory testing suggestions. The verification result is associated with the hash value of the original blood glucose data and fed back to the abnormal data identification process in step S1 and the optimization process of the global meta-calibration model executed in step S3. In a preferred embodiment, the specific operation of the abnormal data identification process in step S1 is as follows: First, a blood glucose data sequence is defined, consisting of multiple blood glucose measurements collected by a portable biochemical sensor and arranged in chronological order. This sequence includes a predetermined number of time points, with each time point corresponding to a blood glucose measurement. Simultaneously, a physiological state perception vector is defined, which is acquired in real time by a wearable device and contains multiple physiological characteristic parameters, including heart rate variability, weight change rate, and mental stress index. Secondly, a one-dimensional convolutional neural network is used to extract local fluctuation pattern features from the blood glucose data sequence to generate a local fluctuation pattern feature vector. At the same time, a gated recurrent unit network is used to extract long-term trend features from the blood glucose data sequence to generate a long-term trend feature vector. Next, a physiological state perception vector is introduced as a conditional input to the attention mechanism. A concatenated feature vector is obtained by concatenating the local fluctuation pattern feature vector with the long-term trend feature vector. This concatenated feature vector is then multiplied by the first learnable parameter matrix, and the physiological state perception vector is multiplied by the second learnable parameter matrix. The two products are then added together, and a hyperbolic tangent function is applied to the result for a nonlinear transformation. This nonlinear transformation result is multiplied by the third learnable parameter vector. Finally, a normalized exponential function is applied to the multiplication result to generate an attention weight vector. This attention weight vector is used to perform element-wise multiplication on the concatenated feature vector to obtain a fused feature vector. Finally, based on the fusion feature vector and physiological state perception vector, the abnormality index of blood glucose data is calculated through a dynamic decision function, and the suspected abnormal blood glucose data points in the blood glucose data sequence are marked according to the abnormality index.

[0006] In a preferred embodiment, the specific process of marking suspected abnormal blood glucose data points in a blood glucose data sequence based on a blood glucose data anomaly index is as follows: An anomaly marker sequence of the same length as the blood glucose data sequence is predefined. Each position of the anomaly marker sequence is initialized to the value zero, which represents normality. At the same time, a preset threshold value between zero and one is defined. For each time point in the blood glucose data sequence, the following operation is performed: the abnormal index of blood glucose data at that time point is calculated based on the fused feature vector and the physiological state perception vector corresponding to that time point; The specific process for calculating the blood glucose data anomaly index is as follows: Calculate the second norm of the fused feature vector to obtain the fused feature norm; calculate the second norm of the physiological state perception vector to obtain the physiological state norm; calculate the second norm of the difference between the fused feature vector and the physiological state perception vector to obtain the difference norm; multiply the fused feature norm and the physiological state norm to obtain the first product; add the difference norm to a positive coefficient with a value of 10 to the power of -6 to obtain a zero-preventing denominator; divide the first product by the zero-preventing denominator to obtain the intermediate quotient; apply the hyperbolic tangent function to the intermediate quotient for a nonlinear transformation to obtain the first transformation value; multiply the first transformation value by a scaling factor to obtain the second product; calculate the natural exponential function value of the negative second product to obtain the natural exponential result; add one to the natural exponential result to obtain the intermediate sum; take the reciprocal of the intermediate sum to obtain the blood glucose data anomaly index. The calculated abnormal blood glucose index is compared with a preset threshold. When the abnormal blood glucose index is greater than or equal to the preset threshold, the value of the corresponding time point in the abnormality marker sequence is updated to a value of one, which indicates abnormality, thereby marking the blood glucose data point at that time point as a suspected abnormal blood glucose data point; otherwise, the value of the corresponding time point in the abnormality marker sequence is marked as zero, indicating that the blood glucose data point at that time point is a normal data point.

[0007] In a preferred embodiment, in step S2, the specific triggering conditions and input data for starting the dynamic calibration model are defined as follows: When the abnormal blood glucose index output in step S1 exceeds a preset threshold, the dynamic calibration model is triggered. The input data of the dynamic calibration model includes four items: the first item is the suspected abnormal blood glucose data points marked in step S1; the second item is the carbohydrate content estimate generated by food image parsing, which is obtained by analyzing the meal image using a convolutional neural network with three convolutional layers, a kernel size of 3x3, and a stride of 1; the third item is the energy consumption estimate collected by the motion sensor, calculated based on accelerometer and heart rate data; and the fourth item is the physiological state perception vector in step S1. Before being input into the dynamic calibration model, all input data are aligned according to a unified timestamp. By processing and fusing the four input data, a multi-source information feature vector is formed. The specific process of processing the multi-source information feature vector through a dual-path calibration network is as follows: First, two parallel dynamic processing network paths are constructed. The first path processes the suspected abnormal blood glucose data points marked in step S1 and the local subsequences formed by their adjacent blood glucose measurements in the blood glucose data sequence, extracting local fluctuation features through a one-dimensional convolutional layer. The second path uniformly processes the carbohydrate content estimate obtained from dietary image parsing, the energy consumption estimate from the motion sensor, and the physiological state perception vector from step S1, mapping heterogeneous data to a unified semantic space through a fully connected layer, and outputting a global path feature vector. Based on the physiological state perception vector, a dynamic weight coefficient is calculated using an sigmoid growth function. The local path feature vector is multiplied by the dynamic weight coefficient to obtain the first weighted vector, and the global path feature vector is multiplied by one minus the dynamic weight coefficient to obtain the second weighted vector. The first weighted vector and the second weighted vector are added to obtain the multi-source information feature vector that serves as the state input of the reinforcement learning agent.

[0008] In a preferred embodiment, the action space of the reinforcement learning agent is defined as the fine-tuning amount for fine-tuning the initial blood glucose calibration value; The reward function of the reinforcement learning agent is generated by time-series physiological consistency delayed feedback, and its specific calculation process is as follows: Q1. Define the intermediate variables and parameters required for reward calculation: Define the calibrated blood glucose value at time t; define the predicted blood glucose value at time t+Δt based on the metabolic physiological model and predicted by multi-source information feature vectors; define the theoretical blood glucose value at time t derived from the physiological state perception vector through a nonlinear regression model; define a constant coefficient with a value of 10 to the power of negative 6; define a consistency reward coefficient with a default value of 0.2; define an indicator function with a value of one or zero, taking the value of one when the relative error between the calibrated blood glucose value and the laboratory venous blood reference value is less than 5%, and otherwise taking the value of zero. Q2. Calculate the consistency error term: Calculate the absolute value of the difference between the calibrated blood glucose value and the predicted blood glucose value at time t+Δt, and use it as the first difference; calculate the absolute value of the difference between the calibrated blood glucose value and the theoretical blood glucose value at time t, and add it to the constant coefficient, and use it as the second sum; divide the first difference by the second sum to obtain the intermediate ratio; apply the hyperbolic tangent function to the intermediate ratio for nonlinear transformation to obtain the consistency error term; Q3. Calculate the consistency reward: Multiply the consistency reward coefficient by the value of the indicator function to obtain the consistency reward. Q4. Calculate the total reward value: Add the negative value of the consistency error term to the consistency reward term to obtain the total reward value obtained by the reinforcement learning agent at time t. The process of outputting the calibrated blood glucose value and calibration confidence level is as follows: The fine-tuning output of the reinforcement learning agent is applied to the initial blood glucose calibration value to obtain the final calibrated blood glucose value. Simultaneously, the calibration confidence is generated through the following process: The state-action value function of the reinforcement learning agent at time t is obtained, which is evaluated by the agent based on the state of the multi-source information feature vector and the selected fine-tuning action; the negative value of the state-action value function is used as input to calculate its natural exponential function value; this is added to the natural exponential function value to obtain an intermediate sum; the reciprocal of the intermediate sum is taken to obtain the final calibration confidence.

[0009] In a preferred embodiment, step S3, which involves aggregating local model parameter increments and optimizing the global meta-calibration model using a federated meta-learning framework, is as follows: The model parameters stored locally on each terminal in the dynamic calibration model are defined as local model parameters. After executing the dynamic calibration model in step S2, each terminal updates its local model parameters using the calibrated blood glucose value generated in step S2 and the physiological state perception vector in step S1. It calculates the difference between the current local model parameters and the global meta-calibration model parameters received in the previous round, generating a local model parameter increment. Each terminal encrypts the local model parameter increment using homomorphic encryption. Simultaneously, each terminal calculates a representative value from the calibration confidence score output in step S2 as a weight index and binds this weight index to the encrypted local model parameter increment. The bound data packet containing the encrypted local model parameter increment and its weight index is then sent to the central server. After receiving data packets containing encrypted local model parameter increments and weight metrics uploaded by each terminal, the central server decrypts them to obtain the plaintext increments and weights, and initiates the optimization process for the global meta-calibration model. The specific steps of this optimization process are as follows: First, obtain the global meta-calibration model parameters for the current round, denoted as the current parameters; second, for each terminal participating in the aggregation, calculate an adaptation loss reflecting the adaptation performance of the global meta-calibration model on the local data of that terminal based on its uploaded local model parameter increments and weight metrics; next, calculate the weighted sum of the adaptation losses of all terminals based on the weight metrics of all terminals; then, calculate the gradient of this weighted sum with respect to the current parameters to obtain the meta-gradient; finally, subtract a predefined product of the meta-learning rate and the meta-gradient from the current parameters to update and obtain the optimized global meta-calibration model parameters.

[0010] In a preferred embodiment, the specific operation for calculating the adaptation loss for each terminal is as follows: The central server constructs a simulated training environment. Starting with the current global meta-calibration model parameters, it performs a gradient descent update on the starting parameters in this simulated training environment based on the parameter update direction implied by the local model parameter increment uploaded by the terminal, resulting in an adapted model parameter. Subsequently, in this simulated training environment, the performance of the adapted model parameter is evaluated using a preset simulated evaluation criterion, and the calculated evaluation loss value is the adaptation loss of the terminal. After collecting the adaptation losses of all terminals, the central server calculates the weighted average gradient of all adaptation losses with respect to the current global meta-calibration model parameters based on the weight indicators uploaded by each terminal; this is the meta-gradient. Finally, the current global meta-calibration model parameters are subtracted from the product of the meta-learning rate and the meta-gradient, thereby completing the update of the global meta-calibration model parameters and obtaining the optimized global meta-calibration model parameters. The optimized global meta-calibration model parameters are then distributed to each terminal. Each terminal receives the optimized global meta-calibration model parameters as the starting point for fine-tuning. Then, it combines the calibrated blood glucose value generated in step S2 with the historical data composed of the physiological state perception vector in step S1, and performs several rounds of iterative training on the starting point parameters. This process is called fine-tuning. Through fine-tuning, each terminal finally obtains a personalized calibration model.

[0011] In a preferred embodiment, the specific process of storing calibration records using a lightweight blockchain network in step S4 is as follows: First, the blood glucose data sequence in step S1 is processed using the secure hash algorithm 256 to generate a fixed-length original blood glucose data hash value. Next, a calibration record is constructed, which contains the following six fields: the first field is the hash value of the original blood glucose data; the second field is the calibrated blood glucose value output in step S2; the third field is the version number of the personalized calibration model obtained by the terminal in step S3; the fourth field is the summary of the multi-source information feature vector generated in step S2; the fifth field is the timestamp of the record generation time; and the sixth field is the calibration confidence output in step S2. Then, the calibration record data containing the above six fields is broadcast as a transaction among the nodes of the consortium blockchain network, which consists of hospital servers, community service center servers, and certification laboratory servers. Finally, network nodes verify and package the transaction through a consensus mechanism, and append a new block containing the calibration record to the blockchain.

[0012] In a preferred embodiment, the specific process by which the smart contract continuously monitors the relationship between the on-chain calibration confidence level and the dynamic threshold and triggers a verification request is as follows: Define the length of the monitoring time window and the preset threshold: Define and maintain a dynamically adjusted confidence threshold, called the dynamic threshold, within the smart contract; define a fixed-length time interval, called the sliding time window; define a first preset proportional threshold and a second preset exponential threshold; Acquiring and Calculating Monitoring Metrics: For newly stored calibration records on the blockchain, the smart contract reads the calibration confidence level of the sixth field and determines the terminal corresponding to the calibration record. Based on the terminal, it acquires all historical calibration records generated by that terminal within the sliding time window, forming a historical record set. It calculates the ratio of the number of records with calibration confidence levels below the dynamic threshold in the historical record set to the total number of records in the set, obtaining the low confidence level ratio. Simultaneously, the consistency difference index is calculated through the following steps: W1. Extract all calibration confidence scores of the terminal within the sliding time window, arrange them in chronological order, and form the confidence score sequence of the terminal. W2. Obtain the calibration confidence scores of all other terminal devices grouped with the same clinical characteristics as the terminal within the same sliding time window, calculate the average of these values, called the population mean, and calculate the standard deviation of these values, called the population standard deviation. W3. For each calibration confidence level in the confidence level sequence of the terminal, perform the following calculations: subtract the population mean from the value to obtain the difference; divide the difference by the sum of the population standard deviation and a normal term coefficient to obtain the standardized ratio; calculate the fourth power of the standardized ratio. W4. Calculate the average of the fourth power results obtained in step 3 for all calibrated confidence levels in the confidence sequence; W5. Calculate the arithmetic square root of the average value obtained in step four to finally obtain the consistency difference index. Trigger Judgment and Execution: When the calculated low confidence ratio is greater than the first preset ratio threshold, or the calculated consistency difference index is greater than the second preset index threshold, the smart contract determines to trigger a verification request; the smart contract generates a structured verification request, which includes the terminal that triggered the request, the type of triggering condition, the hash value of the calibration record on which it is based, and the generation timestamp; the verification request is written into the blockchain as a new transaction and sent to the preset processing module.

[0013] This application also provides a postpartum follow-up monitoring system for gestational diabetes mellitus, specifically including: a data anomaly identification module, a personalized data calibration module, a model federated optimization module, and a quality monitoring and feedback module, wherein; Data anomaly identification module: Based on blood glucose data sequences collected by portable biochemical sensors and physiological state perception vectors obtained by wearable devices, it extracts local fluctuation and long-term trend features of the sequence through convolutional neural networks and gated recurrent unit networks, respectively, and introduces physiological state perception vectors for attention fusion. The blood glucose data anomaly index is output through dynamic decision function to mark suspected abnormal blood glucose data points. Personalized data calibration module: When the abnormal index of blood glucose data exceeds the preset threshold, the dynamic calibration model is activated. It integrates suspected abnormal blood glucose data points, carbohydrate content estimates from dietary image analysis, energy consumption estimates from motion sensors, and physiological state perception vectors to form a multi-source information feature vector. This vector is then processed by a dual-path calibration network and a reinforcement learning agent to output the calibrated blood glucose value and calibration confidence. The model federated optimization module is used to aggregate the local model parameter increments generated by each terminal executing the dynamic calibration model through the federated meta-learning framework, optimize the global meta-calibration model at the central server using a meta-learning update strategy, and distribute the optimized parameters to each terminal for each terminal to fine-tune and obtain a personalized calibration model. Quality monitoring and feedback module: This module uses a lightweight blockchain network to store calibration records containing the hash value of the original blood glucose data, the calibrated blood glucose value, the personalized calibration model version number, the multi-source information feature vector summary, the timestamp, and the calibration confidence level. It monitors the relationship between the on-chain calibration confidence level and the dynamic threshold through smart contracts. When the confidence level is consistently below the threshold or there is a statistical difference, it automatically triggers a verification request, guiding two-finger prick blood tests or laboratory tests. The verification results are then sent to the data anomaly identification module and the model federated optimization module to generate feedback data that is associated with the hash value of the original blood glucose data, thereby driving model updates.

[0014] The beneficial effects of this invention are as follows: by integrating multi-source information and spatiotemporal features, it achieves accurate cleaning and anomaly identification of blood glucose data; based on a context-aware personalized calibration mechanism, it utilizes reinforcement learning to dynamically optimize data correction, significantly improving the accuracy of single-point measurements; the federated meta-learning framework aggregates individual experience while protecting privacy, driving the calibration model to continuously evolve and enhancing generalization capabilities, thus constructing a reliable quality closed loop for blockchain notarization and smart contract monitoring; and through automatic triggering of verification and feedback, it ensures that the anomaly identification and model optimization stages receive high-quality data injection, forming a self-improving monitoring system. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1 The method for postpartum follow-up monitoring of gestational diabetes mellitus, as shown, specifically includes the following steps: Step S1: After the postpartum mother collects blood glucose data through a portable biochemical sensor, the collected blood glucose data is arranged in chronological order to form a blood glucose data sequence. Based on the spatiotemporal feature fusion and the physiological state perception vector of the mother obtained in real time by the wearable device, the abnormal data identification process is triggered. The local fluctuation pattern features of the blood glucose data sequence are extracted through a one-dimensional convolutional neural network. At the same time, the long-term trend features are captured through a gated recurrent unit network. The physiological state perception vector is introduced as the conditional input of the attention mechanism to generate a dynamic decision function to output the blood glucose data abnormality index, thereby marking suspected abnormal blood glucose data points and completing the preliminary cleaning and quality assessment of blood glucose data. Step S2: Start the dynamic calibration model and integrate the suspected abnormal blood glucose data points marked in step S1, the estimated carbohydrate content obtained from the analysis of dietary images, the estimated energy consumption of the motion sensor, and the physiological state perception vector in step S1 to obtain multi-source information features. The multi-source information feature vector is processed by a dual-path calibration network. The reinforcement learning agent takes the multi-source information feature vector as its state, the fine adjustment of the initial calibration value as its action, and the time-series physiological consistency delay feedback as its reward function. The calibrated blood glucose value and calibration confidence are output in real time to achieve context-aware personalized data correction. Step S3: Aggregate the local model parameter increments generated by each terminal in the dynamic calibration model execution in step S2 through the federated meta-learning framework, execute the optimization process of the global meta-calibration model in the central server using the meta-learning update strategy, and distribute the optimized global meta-calibration model parameters to each terminal. Each terminal obtains a personalized calibration model by fine-tuning based on the received global meta-calibration model parameters. Step S4: Utilize a lightweight blockchain network to store calibration records. The calibration records include the hash value of the original blood glucose data, the calibrated blood glucose value output in step S2, the version number of the personalized calibration model obtained in step S3, the summary of the multi-source information feature vector in step S2, the timestamp, and the calibration confidence level output in step S2. The relationship between the on-chain calibration confidence level and the dynamic threshold is continuously monitored through a smart contract. When the calibration confidence level is detected to be continuously lower than the dynamic threshold or there is a statistical difference, a verification request is automatically triggered, guiding the two-finger blood measurement or generating laboratory testing suggestions. The verification result is associated with the hash value of the original blood glucose data and fed back to the abnormal data identification process in step S1 and the optimization process of the global meta-calibration model executed in step S3.

[0020] In this embodiment, the specific operation of the abnormal data identification process in step S1 is as follows: First, a blood glucose data sequence is defined, consisting of multiple blood glucose measurements collected by a portable biochemical sensor and arranged in chronological order. This sequence includes a predetermined number of time points, with each time point corresponding to a blood glucose measurement. Simultaneously, a physiological state perception vector is defined, which is acquired in real time by a wearable device and contains multiple physiological characteristic parameters, including heart rate variability, weight change rate, and mental stress index. Secondly, a one-dimensional convolutional neural network is used to extract local fluctuation pattern features from the blood glucose data sequence, generating a local fluctuation pattern feature vector. The convolutional kernel size of the one-dimensional convolutional neural network is three, the stride is one, and the activation function is a linear rectified function. This design can effectively capture the abrupt changes and short-term trends between adjacent points in the blood glucose sequence and filter out high-frequency noise. At the same time, a gated recurrent unit network is used to extract long-term trend features from the blood glucose data sequence, generating a long-term trend feature vector. The gated recurrent unit network has two layers and sixty-four hidden units. This structure is suitable for memorizing and modeling the periodic changes and slow drift patterns of blood glucose in the postpartum period from several days to several weeks. Next, a physiological state perception vector is introduced as a conditional input to the attention mechanism. This is achieved by concatenating the local fluctuation pattern feature vector with the long-term trend feature vector. The concatenated feature vector is then multiplied by the first learnable parameter matrix, and the physiological state perception vector is multiplied by the second learnable parameter matrix. The two products are then added together, and a hyperbolic tangent function is applied to the result for a nonlinear transformation. This nonlinear transformation result is multiplied by the third learnable parameter vector. Finally, a normalized exponential function is applied to the multiplication result to generate the attention weight vector. Here, the number of rows in the first learnable parameter matrix represents the attention hiding dimension, and the number of columns represents the dimension of the concatenated feature vector. The second learnable parameter matrix... The number of rows in the first learnable parameter matrix is ​​the same as the number of rows in the first learnable parameter matrix, and the number of columns is the dimension of the physiological state perception vector. The dimension of the third learnable parameter vector is the same as the number of rows in the first learnable parameter matrix. The dimension of the attention weight vector is the same as the dimension of the concatenated feature vector, and the sum of its elements is one. This attention mechanism enables the model to dynamically adjust the attention to features of different time scales of the blood glucose sequence according to the real-time physiological state of the mother. For example, when an increase in the mental stress index is detected, more attention is paid to long-term trend features to distinguish between physiological stress response and sensor abnormalities, thereby improving the contextual relevance of feature fusion. The attention weight vector is used to perform element-wise multiplication on the concatenated feature vector to obtain the fused feature vector. Finally, based on the fusion feature vector and physiological state perception vector, the abnormal blood glucose data index is calculated through a dynamic decision function, and the suspected abnormal blood glucose data points in the blood glucose data sequence are marked according to the abnormal blood glucose data index. The specific process of marking suspected abnormal blood glucose data points in a blood glucose data sequence based on the blood glucose data anomaly index is as follows: An anomaly marker sequence of the same length as the blood glucose data sequence is predefined. Each position of the anomaly marker sequence is initialized to the value zero, which represents normality. At the same time, a preset threshold value between zero and one is defined. For each time point in the blood glucose data sequence, the following operation is performed: the abnormal index of blood glucose data at that time point is calculated based on the fused feature vector and the physiological state perception vector corresponding to that time point; The specific process for calculating the blood glucose data anomaly index is as follows: Calculate the second norm of the fused feature vector to obtain the fused feature norm; calculate the second norm of the physiological state perception vector to obtain the physiological state norm; calculate the second norm of the difference between the fused feature vector and the physiological state perception vector to obtain the difference norm; multiply the fused feature norm and the physiological state norm to obtain the first product; add the difference norm to a positive coefficient of 10 to the power of -6 to obtain a zero-preventing denominator; divide the first product by the zero-preventing denominator to obtain the intermediate quotient. In this calculation logic, the product of the fused feature norm and the physiological state norm measures the combined signal of feature strength and the degree of physiological anomaly, while the difference norm introduces a penalty term in the denominator. When the extracted blood glucose features are severely inconsistent with the current physiological state, this quotient will decrease, aiming to distinguish true sensor anomalies. The system identifies patterns that appear abnormal but conform to physiological laws due to drastic physiological changes. A hyperbolic tangent function is applied to the intermediate quotient for a nonlinear transformation, limiting the output value to between -1 and +1, resulting in the first transformed value. This first transformed value is multiplied by a scaling factor of two to obtain the second product. The natural exponential function value of the negative second product is calculated to obtain the natural exponential result. The result of one is added to the natural exponential result to obtain the intermediate sum. The reciprocal of the intermediate sum is taken to obtain the blood glucose data anomaly index. After the hyperbolic tangent, exponential, and reciprocal transformations, the final output blood glucose data anomaly index is compressed to between 0 and 1. A larger value indicates a higher probability of anomaly. The scaling factor is used to adjust the sensitivity of the index to changes in the input intermediate quotient; for example, a value of two can provide a good discriminative gradient within the common data range. The calculated abnormal blood glucose index is compared with a preset threshold. When the abnormal index is greater than or equal to the preset threshold, the value of the corresponding time point in the abnormality marker sequence is updated to a value of one, indicating an abnormality, thus marking the blood glucose data point at that time point as a suspected abnormal blood glucose data point. Otherwise, the value of the corresponding time point in the abnormality marker sequence is marked as zero, indicating that the blood glucose data point at that time point is a normal data point. A typical value for the preset threshold is 0.7. This threshold is a balance point selected based on historical data verification. It can effectively reduce false alarms caused by normal physiological fluctuations in the context of large blood glucose fluctuations in postpartum women, while ensuring a high capture rate for real sensor failures or significant measurement deviations. The threshold can be fine-tuned according to the tolerance for false alarms and missed alarms in actual application scenarios.

[0021] In this embodiment, it is specifically necessary to explain that the specific triggering conditions and input data for starting the dynamic calibration model in step S2 are defined as follows: When the abnormality index of the blood glucose data output in step S1 exceeds the preset threshold of 0.7, the dynamic calibration model is triggered. This threshold of 0.7 is a balance point determined based on the analysis of historical data of postpartum women, aiming to ensure that the computationally intensive calibration process is only initiated when there is a high probability of abnormality in the blood glucose data, thereby optimizing the balance between detection sensitivity and system energy efficiency. The input data of the dynamic calibration model includes four items: the first item is the suspected abnormal blood glucose data point marked in step S1, which is in the form of a sequence of zero and one values ​​of the same length as the blood glucose data sequence, where the one value indicates that the blood glucose data at that time point is marked as abnormal; the second item is the estimated carbohydrate content generated by dietary image analysis, in grams. The first term is obtained by analyzing meal images using a convolutional neural network with three convolutional layers, a kernel size of 3x3, and a stride of 1. This network was pre-trained on a public image dataset containing various common foods and fine-tuned for the typical meal structure of postpartum women. The second term is the energy consumption estimate collected by motion sensors, in kilocalories, calculated based on accelerometer and heart rate data. The third term is the physiological state perception vector in step S1, which includes parameters in three dimensions: heart rate variability, weight change rate, and mental stress index. The parameter values ​​of each dimension have been normalized to between zero and one. Before being input into the dynamic calibration model, all input data are aligned according to a unified timestamp to ensure temporal consistency. By processing and fusing the four input data, a multi-source information feature vector is formed. The specific process of processing the multi-source information feature vector through a dual-path calibration network is as follows: First, two parallel dynamic processing network paths are constructed. The first path processes the suspected abnormal blood glucose data points marked in step S1 and the local subsequences formed by their adjacent blood glucose measurements in the blood glucose data sequence. Local fluctuation features are extracted through a one-dimensional convolutional layer with 32 kernels of size 3, a stride of 1, and a gated linear unit activation function, outputting a 128-dimensional local path feature vector. The second path uniformly processes the carbohydrate content estimate obtained from dietary image analysis, the energy consumption estimate from the motion sensor, and the physiological state perception vector from step S1. A fully connected layer maps the heterogeneous data to a unified semantic space, outputting a 128-dimensional global path feature vector. Based on the physiological state perception vector, a dynamic weight coefficient ranging from zero to one is calculated using an sigmoid growth function. The local path feature vector is then multiplied by the dynamic weight coefficient. The model obtains a first weighted vector by multiplying the global path feature vector by one and subtracting the dynamic weight coefficient to obtain a second weighted vector. The first weighted vector and the second weighted vector are then added together to obtain a multi-source information feature vector that serves as the state input for the reinforcement learning agent. This fusion mechanism enables the model to adaptively adjust its dependence on sensor data and external context information based on the mother's real-time physiological state. For example, when the physiological state perception vector shows an abnormally high mental stress index, the dynamic weight coefficient will decrease, allowing the model to focus more on finding a reasonable explanation for blood glucose fluctuations from global information such as diet and exercise, thereby avoiding misjudging physiological stress responses as sensor errors requiring calibration. The dynamic weight fusion mechanism ensures that when the physiological state perception vector indicates that the mother is experiencing severe physiological fluctuations, the global context features of the second path receive higher weights, while when the physiological state is stable, the local blood glucose features of the first path dominate the fusion result. The action space of a reinforcement learning agent is defined as the fine-tuning amount for fine-tuning the initial blood glucose calibration value. The fine-tuning amount ranges from -0.5 mmol / L to +0.5 mmol / L. This range is set based on the typical allowable error of blood glucose meters in clinical practice and the safe fluctuation range of postpartum blood glucose, aiming to allow sufficient correction space while preventing over-correction from introducing new risks. The reward function of the reinforcement learning agent is generated by time-series physiological consistency delayed feedback, and its specific calculation process is as follows: Q1. Define the intermediate variables and parameters required for reward calculation: Define the calibrated blood glucose value at time t; define the predicted blood glucose value at time t+Δt based on the metabolic physiological model and predicted by multi-source information feature vectors, where Δt is a predefined delay time window (typically 30 minutes, corresponding to the critical observation period for postprandial blood glucose changes); define the theoretical blood glucose value at time t derived from the physiological state perception vector through a nonlinear regression model; define a zero-prevention constant coefficient with a value of 10 to the power of negative 6; define a consistency reward coefficient with a default value of 0.2 (this coefficient is used to adjust the strength of this sparse but high-confidence reward signal in laboratory verification); define an indicator function with a value of one or zero, taking the value of one when the relative error between the calibrated blood glucose value and the laboratory venous blood reference value is less than 5%, and otherwise taking the value of zero; Q2. Calculate the consistency error term: Calculate the absolute value of the difference between the calibrated blood glucose value and the predicted blood glucose value at time t+Δt, as the first difference; calculate the absolute value of the difference between the calibrated blood glucose value and the theoretical blood glucose value at time t, and add it to the zero constant coefficient, as the second sum; divide the first difference by the second sum to obtain the intermediate ratio; apply the hyperbolic tangent function to the intermediate ratio for nonlinear transformation, limiting the output range to between zero and one, to obtain the consistency error term. Standardize the prediction deviation with the theoretical blood glucose value as a reference, so that the penalty intensity of the reward for the deviation is adapted to the expected fluctuation level under the current physiological state. Q3. Calculate the consistency reward: Multiply the consistency reward coefficient by the value of the indicator function to obtain the consistency reward. Q4. Calculate the total reward value: Add the negative value of the consistency error term to the consistency reward term to obtain the total reward value obtained by the reinforcement learning agent at time t. The process of outputting the calibrated blood glucose value and calibration confidence level is as follows: The fine-tuning output of the reinforcement learning agent is applied to the initial blood glucose calibration value to obtain the final calibrated blood glucose value. Simultaneously, calibration confidence is generated through the following process: The state-action value function of the reinforcement learning agent at time t is obtained. This value is assessed by the agent based on the state of the multi-source information feature vector and the selected fine-tuning action, representing the expected long-term cumulative reward under a given state and action. The negative value of the state-action value function is used as input to calculate its natural exponential function value. One is added to the natural exponential function value to obtain an intermediate sum. The reciprocal of the intermediate sum is taken to obtain the calibration confidence, which ranges from zero to one. A higher value indicates a more reliable calibration result. This confidence is essentially the result of mapping the state-action value function value through a sigmoid function. It transforms the agent's long-term value assessment of the current decision into an intuitive probability estimate of reliability, thus providing a quantitative basis for downstream steps to adopt this calibration result.

[0022] In this embodiment, the specific process of aggregating local model parameter increments and optimizing the global meta-calibration model through the federated meta-learning framework in step S3 is as follows: Each computing device executing the postpartum follow-up monitoring method for gestational diabetes mellitus is defined as a terminal. The model parameters stored locally on each terminal in the dynamic calibration model are defined as local model parameters. After executing the dynamic calibration model in step S2, each terminal updates its local model parameters using the calibrated blood glucose value generated in step S2 and the physiological state perception vector in step S1. It calculates the difference between the current local model parameters and the global meta-calibration model parameters received in the previous round to generate the local model parameter increment. To protect data privacy, each terminal encrypts the local model parameter increment using homomorphic encryption. Simultaneously, each terminal calculates a representative value from the calibration confidence score output in step S2 as a weight index and binds this weight index to the encrypted local model parameter increment. The representative value is usually the average or median of multiple calibration confidence scores generated by the terminal over a period of time, used to comprehensively evaluate the overall reliability of the terminal's recent data calibration. The bound data packet containing the encrypted local model parameter increment and its weight index is sent to the central server. After receiving data packets containing encrypted local model parameter increments and weight metrics uploaded by each terminal, the central server decrypts them to obtain the plaintext increments and weights, and initiates the optimization process for the global meta-calibration model. The specific steps of this optimization process are as follows: First, obtain the global meta-calibration model parameters for the current round, denoted as the current parameters; second, for each terminal participating in the aggregation, calculate an adaptation loss reflecting the adaptation performance of the global meta-calibration model on the local data of that terminal based on the local model parameter increments and weight metrics uploaded by that terminal; next, calculate the weighted sum of the adaptation losses of all terminals based on the weight metrics of all terminals; then, calculate the gradient of this weighted sum with respect to the current parameters to obtain the meta-gradient; finally, subtract a predefined product of the meta-learning rate and the meta-gradient from the current parameters to update and obtain the optimized global meta-calibration model parameters. The specific steps for calculating the adaptation loss for each terminal are as follows: The central server constructs a simulated training environment. Starting with the current global meta-calibrated model parameters, and based on the parameter update direction implied by the local model parameter increments uploaded by the terminal, a gradient descent update is performed on the starting parameters in this simulated training environment to obtain adapted model parameters. Subsequently, in this simulated training environment, the performance of the adapted model parameters is evaluated using a preset simulation evaluation criterion. The calculated evaluation loss value is the adaptation loss of the terminal. The simulation evaluation criterion is a preset loss function in the simulated training environment, used to quantify the prediction error of the adapted model parameters. The objective optimized by the reinforcement learning agent in step S2 is consistent with that of the target. After the central server collects the adaptation loss of all terminals, it calculates the weighted average gradient of all adaptation losses with respect to the current global meta-calibration model parameters based on the weight index uploaded by each terminal. This is the meta-gradient. Finally, the current global meta-calibration model parameters are subtracted from the product of the meta-learning rate and the meta-gradient to complete the update of the global meta-calibration model parameters and obtain the optimized global meta-calibration model parameters. The meta-learning rate is a preset positive hyperparameter used to control the step size of each global model parameter update. The typical value range is between one-thousandth and one-hundredth. The optimized global meta-calibration model parameters are then distributed to each terminal. Each terminal receives the optimized global meta-calibration model parameters as the starting point for fine-tuning. Then, combining the calibrated blood glucose value generated in step S2 and the historical data composed of the physiological state perception vector in step S1, it performs several rounds of iterative training on the starting point parameters. This process is called fine-tuning. Several rounds of iterative training usually consist of five to twenty rounds, aiming to achieve rapid adaptation from global knowledge to local personalized needs with limited computational overhead. Through fine-tuning, each terminal finally obtains a personalized calibration model. This personalized calibration model is a version of the dynamic calibration model optimized for the user data distribution of the terminal. Its version number will be stored as part of the calibration record in the subsequent step S4.

[0023] In this embodiment, the specific process of storing calibration records using a lightweight blockchain network in step S4 needs to be explained as follows: First, the blood glucose data sequence in step S1 is processed using the secure hash algorithm 256 to generate a fixed-length original blood glucose data hash value. Next, a calibration record is constructed, which contains the following six fields: the first field is the hash value of the original blood glucose data; the second field is the calibrated blood glucose value output in step S2; the third field is the version number of the personalized calibration model obtained by the terminal in step S3; the fourth field is a summary of the multi-source information feature vector generated in step S2. The summary can be obtained by extracting the values ​​of the first 128 dimensions of the multi-source information feature vector or by calculating its hash value. Using the summary form can significantly reduce the amount of data stored on the chain while ensuring that key information can be traced back to the feature context later; the fifth field is the timestamp of the record generation time; and the sixth field is the calibration confidence level output in step S2. Then, the calibration record data containing the above six fields is broadcast as a transaction among the nodes of the consortium blockchain network, which consists of hospital servers, community service center servers, and certification laboratory servers. Finally, network nodes verify and package the transaction through a consensus mechanism, and append a new block containing the calibration record to the blockchain to achieve immutable data storage. Through distributed consensus storage, it is ensured that once the calibration record is on the chain, it cannot be tampered with by any single participant, providing a credible foundation for medical data auditing and accountability. The specific process by which the smart contract continuously monitors the relationship between on-chain calibration confidence and dynamic thresholds and triggers verification requests is as follows: Define the length of the monitoring time window and the preset threshold: Define and maintain a dynamically adjusted confidence threshold, called the dynamic threshold, within the smart contract. Its initial value can be set to 0.6, and it can be periodically adjusted according to the statistical distribution of the recent on-chain calibration confidence. Define a fixed-length time interval, called the sliding time window, with a typical length of 24 hours. Define a first preset proportional threshold and a second preset exponential threshold. For example, the first preset proportional threshold can be 0.6, and the second preset exponential threshold can be 2.0. Acquiring and Calculating Monitoring Metrics: For newly stored calibration records on the blockchain, the smart contract reads the calibration confidence level of the sixth field and determines the terminal corresponding to the calibration record. Based on the terminal, it acquires all historical calibration records generated by that terminal within the sliding time window, forming a historical record set. It calculates the ratio of the number of records with calibration confidence levels below the dynamic threshold in the historical record set to the total number of records in the set, obtaining the low confidence level ratio. Simultaneously, the consistency difference index is calculated through the following steps: W1. Extract all calibration confidence scores of the terminal within the sliding time window, arrange them in chronological order, and form the confidence score sequence of the terminal. W2. Obtain the calibration confidence scores of all other terminal devices grouped with the same clinical characteristics as the terminal within the same sliding time window, calculate the average of these values, called the population mean, and calculate the standard deviation of these values, called the population standard deviation. W3. For each calibration confidence level in the confidence level sequence of the terminal, perform the following calculations: Subtract the population mean from the value to obtain the difference; Divide the difference by the sum of the population standard deviation and a normalization coefficient, where the normalization coefficient is a very small positive number, such as 10 to the power of negative 8, to prevent the denominator from being zero and to ensure calculation stability, to obtain the standardized ratio; Calculate the fourth power of the standardized ratio. W4. Calculate the average of the fourth power results obtained in step 3 for all calibrated confidence levels in the confidence sequence; W5. Calculate the arithmetic square root of the mean obtained in step four to finally obtain the consistency difference index. This index, by amplifying the tail differences between individual and group distributions, can more sensitively detect situations where the terminal calibration confidence level continuously deviates from the normal pattern of the group, and is more robust than traditional variance detection. Trigger Judgment and Execution: When the calculated low confidence ratio exceeds the first preset ratio threshold, or the calculated consistency difference index exceeds the second preset index threshold, the smart contract determines to trigger a verification request. The smart contract generates a structured verification request, which includes the terminal that triggered the request, the trigger condition type, the hash value of the calibration record on which it is based, and the generation timestamp. The verification request is written into the blockchain as a new transaction and sent to the preset processing module. The processing module performs one or two of the following operations based on the content of the verification request: sends a prompt message to the user through the terminal application, guiding the user to perform a finger-prick blood glucose measurement and verifying the blood glucose value obtained from the measurement; or generates a laboratory testing recommendation form, suggesting that the user go to a designated medical institution for venous blood glucose testing, and the medical institution verifies the blood glucose value obtained from the test. Regardless of how the verification result blood glucose value is obtained, it is associated with the hash value of the original blood glucose data stored in the original calibration record that triggered the verification request, and is stored in the blockchain as a new record containing the verification result blood glucose value, the associated hash value of the original blood glucose data, and the timestamp. This associated storage establishes a complete, tamper-proof audit trail from the original data and calibration process to the authoritative verification result. The verification result blood glucose value and its associated raw blood glucose data hash value are encapsulated into a feedback data record, which includes at least the verification result blood glucose value field, the associated raw blood glucose data hash value field, and the timestamp field. The feedback data record is sent to the abnormal data identification process in step S1. This process uses the verification result blood glucose value in the feedback data record as a reference standard to adjust and update the internal model parameters on which the dynamic decision function used to generate the blood glucose data abnormality index in step S1 depends. Specifically, the verification result blood glucose value is used as the real label of whether the blood glucose data at this time point is abnormal. Together with the fusion feature vector and physiological state perception vector generated in step S1, it constitutes a training sample, which drives the neural network parameters in the dynamic decision function to optimize, thereby improving the accuracy of subsequent identification of similar abnormal patterns and reducing false alarms and false negatives. Simultaneously, the feedback data record is sent to the optimization process of the global meta-calibration model in step S3. This process combines the verification result blood glucose value in the feedback data record with the fusion feature vector and physiological state perception vector generated in step S1, which are associated with the hash value of the original blood glucose data, as new training sample data for subsequent optimization training of the global meta-calibration model parameters. Under the federated meta-learning framework, the central server can prioritize scheduling terminals containing such high-quality verification samples to participate in aggregation, or treat this part of the data as a high-weight task during the global model training stage, thereby efficiently injecting the "gold standard" verification information into global knowledge and systematically improving the accuracy of personalized calibration models for all terminals.

[0024] Example 2 This embodiment provides, for example Figure 2 The method for postpartum follow-up monitoring of gestational diabetes mellitus shown includes: a data anomaly identification module, a personalized data calibration module, a model federated optimization module, and a quality monitoring and feedback module, wherein; Data anomaly identification module: Based on blood glucose data sequences collected by portable biochemical sensors and physiological state perception vectors obtained by wearable devices, it extracts local fluctuation and long-term trend features of the sequence through convolutional neural networks and gated recurrent unit networks, respectively, and introduces physiological state perception vectors for attention fusion. The blood glucose data anomaly index is output through dynamic decision function to mark suspected abnormal blood glucose data points. Personalized data calibration module: When the abnormal index of blood glucose data exceeds the preset threshold, the dynamic calibration model is activated. It integrates suspected abnormal blood glucose data points, carbohydrate content estimates from dietary image analysis, energy consumption estimates from motion sensors, and physiological state perception vectors to form a multi-source information feature vector. This vector is then processed by a dual-path calibration network and a reinforcement learning agent to output the calibrated blood glucose value and calibration confidence. The model federated optimization module is used to aggregate the local model parameter increments generated by each terminal executing the dynamic calibration model through the federated meta-learning framework, optimize the global meta-calibration model at the central server using a meta-learning update strategy, and distribute the optimized parameters to each terminal for each terminal to fine-tune and obtain a personalized calibration model. Quality monitoring and feedback module: This module uses a lightweight blockchain network to store calibration records containing the hash value of the original blood glucose data, the calibrated blood glucose value, the personalized calibration model version number, the multi-source information feature vector summary, the timestamp, and the calibration confidence level. It monitors the relationship between the on-chain calibration confidence level and the dynamic threshold through smart contracts. When the confidence level is consistently below the threshold or there is a statistical difference, it automatically triggers a verification request, guiding two-finger prick blood tests or laboratory tests. The verification results are then sent to the data anomaly identification module and the model federated optimization module to generate feedback data that is associated with the hash value of the original blood glucose data, thereby driving model updates.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of postpartum follow-up monitoring of gestational diabetes, characterized in that, Specifically, the following steps are included: Step S1: After the postpartum mother collects blood glucose data through a portable biochemical sensor, the collected blood glucose data is arranged in chronological order to form a blood glucose data sequence. Based on the spatiotemporal feature fusion and the physiological state perception vector of the mother obtained in real time by the wearable device, the abnormal data identification process is triggered. The local fluctuation pattern features of the blood glucose data sequence are extracted through a one-dimensional convolutional neural network. At the same time, the long-term trend features are captured through a gated recurrent unit network. The physiological state perception vector is introduced as the conditional input of the attention mechanism to generate a dynamic decision function to output the blood glucose data abnormality index, thereby marking suspected abnormal blood glucose data points. Step S2: Start the dynamic calibration model and integrate the suspected abnormal blood glucose data points marked in step S1, the estimated carbohydrate content obtained from the analysis of dietary images, the estimated energy consumption of the motion sensor, and the physiological state perception vector in step S1 to obtain multi-source information features. Process the multi-source information feature vector through a dual-path calibration network. The reinforcement learning agent takes the multi-source information feature vector as its state, the fine-tuning amount of the initial calibration value as its action, and the time-series physiological consistency delay feedback as its reward function to output the calibrated blood glucose value and calibration confidence in real time. Step S3: Aggregate the local model parameter increments generated by each terminal in the dynamic calibration model execution in step S2 through the federated meta-learning framework, execute the optimization process of the global meta-calibration model in the central server using the meta-learning update strategy, and distribute the optimized global meta-calibration model parameters to each terminal. Each terminal obtains a personalized calibration model by fine-tuning based on the received global meta-calibration model parameters. Step S4: Utilize a lightweight blockchain network to store calibration records. The calibration records include the hash value of the original blood glucose data, the calibrated blood glucose value output in step S2, the version number of the personalized calibration model obtained in step S3, the summary of the multi-source information feature vector in step S2, the timestamp, and the calibration confidence level output in step S2. The relationship between the on-chain calibration confidence level and the dynamic threshold is continuously monitored through a smart contract. When the calibration confidence level is detected to be continuously lower than the dynamic threshold or there is a statistical difference, a verification request is automatically triggered, guiding the two-finger blood measurement or generating laboratory testing suggestions. The verification result is associated with the hash value of the original blood glucose data and fed back to the abnormal data identification process in step S1 and the optimization process of the global meta-calibration model executed in step S3.

2. The method of claim 1, wherein the method is a postpartum follow-up monitoring method for gestational diabetes. In step S1, the specific operations of the abnormal data identification process are as follows: First, a blood glucose data sequence is defined, consisting of multiple blood glucose measurements collected by a portable biochemical sensor and arranged in chronological order. This sequence includes a predetermined number of time points, with each time point corresponding to a blood glucose measurement. Simultaneously, a physiological state perception vector is defined, which is acquired in real time by a wearable device and contains multiple physiological characteristic parameters, including heart rate variability, weight change rate, and mental stress index. Secondly, a one-dimensional convolutional neural network is used to extract local fluctuation pattern features from the blood glucose data sequence to generate a local fluctuation pattern feature vector. At the same time, a gated recurrent unit network is used to extract long-term trend features from the blood glucose data sequence to generate a long-term trend feature vector. Next, the physiological state perception vector is introduced as the conditional input of the attention mechanism. By concatenating the local fluctuation pattern feature vector with the long-term trend feature vector, the concatenated feature vector is obtained. The concatenated feature vector is multiplied by the first learnable parameter matrix, and the physiological state perception vector is multiplied by the second learnable parameter matrix. The two products are then added together, and a hyperbolic tangent function is applied to the sum for a nonlinear transformation. The result of the nonlinear transformation is multiplied by the third learnable parameter vector, and finally, a normalized exponential function is applied to the multiplication result to generate an attention weight vector. The attention weight vector is used to perform element-wise multiplication on the concatenated feature vector to obtain the fused feature vector. Finally, based on the fusion feature vector and physiological state perception vector, the abnormality index of blood glucose data is calculated through a dynamic decision function, and the suspected abnormal blood glucose data points in the blood glucose data sequence are marked according to the abnormality index.

3. The method for postpartum follow-up monitoring of gestational diabetes mellitus according to claim 2, characterized in that: The specific process of marking suspected abnormal blood glucose data points in a blood glucose data sequence based on the blood glucose data anomaly index is as follows: An anomaly marker sequence of the same length as the blood glucose data sequence is predefined. Each position of the anomaly marker sequence is initialized to the value zero, which represents normality. At the same time, a preset threshold value between zero and one is defined. For each time point in the blood glucose data sequence, the following operation is performed: the abnormal index of blood glucose data at that time point is calculated based on the fused feature vector and the physiological state perception vector corresponding to that time point; The specific process for calculating the abnormal index of blood glucose data is as follows: calculate the second norm of the fused feature vector to obtain the fused feature norm; Calculate the second norm of the physiological state perception vector to obtain the physiological state norm; calculate the second norm of the difference between the fused feature vector and the physiological state perception vector to obtain the difference norm; multiply the fused feature norm and the physiological state norm to obtain the first product; add the difference norm to a positive coefficient with a value of 10 to the power of negative 6 to obtain a zero-preventing denominator; divide the first product by the zero-preventing denominator to obtain the intermediate quotient; apply the hyperbolic tangent function to the intermediate quotient for a nonlinear transformation to obtain the first transformation value; Multiply the first transformed value by a scaling factor to obtain the second product; calculate the natural exponential function value of the negative second product to obtain the natural exponential result; add one to the natural exponential result to obtain the median sum; take the reciprocal of the median sum to obtain the blood glucose data abnormality index; The calculated abnormal blood glucose data index is compared with a preset threshold. When the abnormal blood glucose data index is greater than or equal to the preset threshold, the value of the corresponding time point in the abnormal labeling sequence is updated to the value of one, which indicates abnormality, thereby marking the blood glucose data point at that time point as a suspected abnormal blood glucose data point. Otherwise, the value of the corresponding time point in the abnormal marker sequence is marked as zero, indicating that the blood glucose data point at that time point is a normal data point.

4. The method for postpartum follow-up monitoring of gestational diabetes mellitus according to claim 3, characterized in that: In step S2, the specific triggering conditions and input data for starting the dynamic calibration model are defined as follows: When the abnormal blood glucose index output in step S1 exceeds a preset threshold, the dynamic calibration model is triggered. The input data of the dynamic calibration model includes four items: the first item is the suspected abnormal blood glucose data points marked in step S1; the second item is the carbohydrate content estimate generated by food image parsing, which is obtained by analyzing the meal image using a convolutional neural network with three convolutional layers, a kernel size of 3x3, and a stride of 1; the third item is the energy consumption estimate collected by the motion sensor, calculated based on accelerometer and heart rate data; and the fourth item is the physiological state perception vector in step S1. Before being input into the dynamic calibration model, all input data are aligned according to a unified timestamp. By processing and fusing the four input data, a multi-source information feature vector is formed. The specific process of processing the multi-source information feature vector through a dual-path calibration network is as follows: First, two parallel dynamic processing network paths are constructed. The first path processes the suspected abnormal blood glucose data points marked in step S1 and the local subsequences formed by their adjacent blood glucose measurements in the blood glucose data sequence, extracting local fluctuation features through a one-dimensional convolutional layer. The second path uniformly processes the carbohydrate content estimate obtained from dietary image parsing, the energy consumption estimate from the motion sensor, and the physiological state perception vector from step S1, mapping heterogeneous data to a unified semantic space through a fully connected layer, and outputting a global path feature vector. Based on the physiological state perception vector, a dynamic weight coefficient is calculated using an sigmoid growth function. The local path feature vector is multiplied by the dynamic weight coefficient to obtain the first weighted vector, and the global path feature vector is multiplied by one minus the dynamic weight coefficient to obtain the second weighted vector. The first weighted vector and the second weighted vector are added to obtain the multi-source information feature vector that serves as the state input of the reinforcement learning agent.

5. A method for postpartum follow-up monitoring of gestational diabetes mellitus according to claim 4, characterized in that: The action space of the reinforcement learning agent is defined as the fine-tuning amount for fine-tuning the initial blood glucose calibration value; The reward function of the reinforcement learning agent is generated by time-series physiological consistency delayed feedback, and its specific calculation process is as follows: Q1. Define the intermediate variables and parameters required for reward calculation: Define the calibrated blood glucose value at time t; define the predicted blood glucose value at time t+Δt based on the metabolic physiological model and predicted by the feature vector of multi-source information; define the theoretical blood glucose value at time t derived from the physiological state perception vector through a nonlinear regression model; define a constant coefficient with a value of 10 to the power of negative 6; define a consistency reward coefficient with a default value of 0.

2. Define an indicator function that takes the value of one or zero. The function takes the value of one when the relative error between the calibrated blood glucose value and the laboratory venous blood reference value is less than five percent, and takes the value of zero otherwise. Q2. Calculate the consistency error term: Calculate the absolute value of the difference between the calibrated blood glucose value and the predicted blood glucose value at the future time t+Δt, and use it as the first difference; Calculate the absolute value of the difference between the calibrated blood glucose value and the theoretical blood glucose value at time t, add it to the constant coefficient, and use it as the second sum; divide the first difference by the second sum to obtain the intermediate ratio. A nonlinear transformation of the intermediate ratio using the hyperbolic tangent function is performed to obtain the consistency error term. Q3. Calculate the consistency reward: Multiply the consistency reward coefficient by the value of the indicator function to obtain the consistency reward. Q4. Calculate the total reward value: Add the negative value of the consistency error term to the consistency reward term to obtain the total reward value obtained by the reinforcement learning agent at time t. The process of outputting the calibrated blood glucose value and calibration confidence level is as follows: The fine-tuning output of the reinforcement learning agent is applied to the initial blood glucose calibration value to obtain the final calibrated blood glucose value. Simultaneously, the calibration confidence is generated through the following process: The state-action value function of the reinforcement learning agent at time t is obtained, which is evaluated by the agent based on the state of the multi-source information feature vector and the selected fine-tuning action; the negative value of the state-action value function is used as input to calculate its natural exponential function value; this is added to the natural exponential function value to obtain an intermediate sum; the reciprocal of the intermediate sum is taken to obtain the final calibration confidence.

6. A method for postpartum follow-up monitoring of gestational diabetes mellitus according to claim 5, characterized in that: In step S3, the specific process of aggregating local model parameter increments and optimizing the global meta-calibration model through the federated meta-learning framework is as follows: The model parameters stored locally on each terminal in the dynamic calibration model are defined as local model parameters. After executing the dynamic calibration model in step S2, each terminal updates the local model parameters using the calibrated blood glucose value generated in step S2 and the physiological state perception vector in step S1. The difference between the current local model parameters and the global meta-calibration model parameters received in the previous round is calculated to generate the local model parameter increment. Each terminal encrypts the local model parameter increment using homomorphic encryption. At the same time, each terminal calculates a representative value from the calibration confidence output in step S2 as a weight index and binds the weight index to the encrypted local model parameter increment. The bound data packet containing encrypted local model parameter increments and their weight metrics is sent to the central server. After receiving data packets containing encrypted local model parameter increments and weight metrics uploaded by each terminal, the central server decrypts them to obtain the plaintext increments and weights, and initiates the optimization process for the global meta-calibration model. The specific steps of this optimization process are as follows: First, obtain the global meta-calibration model parameters for the current round, denoted as the current parameters; second, for each participating terminal, calculate an adaptation loss reflecting the adaptation performance of the global meta-calibration model on the local data of that terminal based on its uploaded local model parameter increments and weight metrics; next, calculate the weighted sum of the adaptation losses of all terminals based on their weight metrics; then, calculate the gradient of this weighted sum with respect to the current parameters to obtain the meta-gradient; finally, subtract the product of a predefined meta-learning rate and the meta-gradient from the current parameters to update the optimized global meta-calibration model parameters.

7. A method for postpartum follow-up monitoring of gestational diabetes mellitus according to claim 6, characterized in that: The specific operation for calculating the adaptation loss for each terminal is as follows: The central server constructs a simulated training environment. Starting with the current global meta-calibration model parameters, it performs a gradient descent update on the starting parameters in this simulated training environment based on the parameter update direction implied by the local model parameter increment uploaded by the terminal, resulting in an adapted model parameter. Subsequently, in this simulated training environment, the performance of the adapted model parameter is evaluated using a preset simulated evaluation criterion, and the calculated evaluation loss value is the adaptation loss of the terminal. After collecting the adaptation losses of all terminals, the central server calculates the weighted average gradient of all adaptation losses with respect to the current global meta-calibration model parameters based on the weight indicators uploaded by each terminal; this is the meta-gradient. Finally, the current global meta-calibration model parameters are subtracted from the product of the meta-learning rate and the meta-gradient, thereby completing the update of the global meta-calibration model parameters and obtaining the optimized global meta-calibration model parameters. The optimized global meta-calibration model parameters are then distributed to each terminal. Each terminal receives the optimized global meta-calibration model parameters as the starting point for fine-tuning. Then, it combines the calibrated blood glucose value generated in step S2 with the historical data composed of the physiological state perception vector in step S1, and performs several rounds of iterative training on the starting point parameters. This process is called fine-tuning. Through these fine-tuning steps, each terminal ultimately obtains a personalized calibration model.

8. A method for postpartum follow-up monitoring of gestational diabetes mellitus according to claim 7, characterized in that: In step S4, the specific process of storing calibration records using a lightweight blockchain network is as follows: First, the blood glucose data sequence in step S1 is processed using the secure hash algorithm 256 to generate a fixed-length original blood glucose data hash value. Next, a calibration record is constructed, which contains the following six fields: the first field is the hash value of the original blood glucose data; The second field is the calibrated blood glucose value output in step S2; The third field is the version number of the personalized calibration model obtained by the terminal in step S3; the fourth field is a summary of the multi-source information feature vector generated in step S2. The fifth field is the timestamp of when the record was generated; The sixth field is the calibration confidence level output from step S2; Then, the calibration record data containing the above six fields is broadcast as a transaction among the nodes of the consortium blockchain network, which consists of hospital servers, community service center servers, and certification laboratory servers. Finally, network nodes verify and package the transaction through a consensus mechanism, and append a new block containing the calibration record to the blockchain.

9. A method for postpartum follow-up monitoring of gestational diabetes mellitus according to claim 8, characterized in that: The specific process by which the smart contract continuously monitors the relationship between on-chain calibration confidence and dynamic threshold and triggers verification requests is as follows: Define the length of the monitoring time window and the preset threshold: Define and maintain a dynamically adjusted confidence threshold within the smart contract, called the dynamic threshold; define a fixed-length time interval, called the sliding time window; Define a first preset ratio threshold and a second preset index threshold; Acquiring and Calculating Monitoring Metrics: For newly stored calibration records on the blockchain, the smart contract reads the calibration confidence level of the sixth field and determines the terminal corresponding to the calibration record. Based on the terminal, it acquires all historical calibration records generated by that terminal within the sliding time window, forming a historical record set. It calculates the ratio of the number of records with calibration confidence levels below the dynamic threshold in the historical record set to the total number of records in the set, obtaining the low confidence level ratio. Simultaneously, the consistency difference index is calculated through the following steps: W1. Extract all calibration confidence scores of the terminal within the sliding time window, arrange them in chronological order, and form the confidence score sequence of the terminal. W2. Obtain the calibration confidence scores of all other terminal devices grouped with the same clinical characteristics as the terminal within the same sliding time window, calculate the average of these values, called the population mean, and calculate the standard deviation of these values, called the population standard deviation. W3. For each calibration confidence level in the confidence level sequence of the terminal, perform the following calculations: subtract the population mean from the value to obtain the difference; divide the difference by the sum of the population standard deviation and a normal term coefficient to obtain the standardized ratio. Calculate the fourth power of the standardized ratio; W4. Calculate the average of the fourth power results obtained in step 3 for all calibrated confidence levels in the confidence sequence; W5. Calculate the arithmetic square root of the average value obtained in step four to finally obtain the consistency difference index. Trigger judgment and execution: When the calculated low confidence ratio is greater than the first preset ratio threshold, or the calculated consistency difference index is greater than the second preset index threshold, the smart contract determines to trigger a verification request. The smart contract generates a structured verification request, which includes the terminal that triggered the request, the type of triggering condition, the hash value of the calibration record on which it is based, and the generation timestamp; the verification request is written into the blockchain as a new transaction and sent to the preset processing module.

10. A postpartum follow-up monitoring system for gestational diabetes mellitus, applied to a postpartum follow-up monitoring method for gestational diabetes mellitus as described in any one of claims 1-9, characterized in that: Specifically, it includes: The system includes a data anomaly identification module, a personalized data calibration module, a model federated optimization module, and a quality monitoring and feedback module. Data anomaly identification module: Based on blood glucose data sequences collected by portable biochemical sensors and physiological state perception vectors obtained by wearable devices, it extracts local fluctuation and long-term trend features of the sequence through convolutional neural networks and gated recurrent unit networks, respectively, and introduces physiological state perception vectors for attention fusion. The blood glucose data anomaly index is output through dynamic decision function to mark suspected abnormal blood glucose data points. Personalized data calibration module: When the abnormal index of blood glucose data exceeds the preset threshold, the dynamic calibration model is activated. It integrates suspected abnormal blood glucose data points, carbohydrate content estimates from dietary image analysis, energy consumption estimates from motion sensors, and physiological state perception vectors to form a multi-source information feature vector. This vector is then processed by a dual-path calibration network and a reinforcement learning agent to output the calibrated blood glucose value and calibration confidence. The model federated optimization module is used to aggregate the local model parameter increments generated by each terminal executing the dynamic calibration model through the federated meta-learning framework, optimize the global meta-calibration model at the central server using a meta-learning update strategy, and distribute the optimized parameters to each terminal for each terminal to fine-tune and obtain a personalized calibration model. Quality monitoring and feedback module: This module uses a lightweight blockchain network to store calibration records containing the hash value of the original blood glucose data, the calibrated blood glucose value, the personalized calibration model version number, the multi-source information feature vector summary, the timestamp, and the calibration confidence level. It monitors the relationship between the on-chain calibration confidence level and the dynamic threshold through smart contracts. When the confidence level is consistently below the threshold or there is a statistical difference, it automatically triggers a verification request, guiding two-finger prick blood tests or laboratory tests. The verification results are then sent to the data anomaly identification module and the model federated optimization module to generate feedback data that is associated with the hash value of the original blood glucose data, thereby driving model updates.

Images