Diabetes abnormal index early warning system and method based on recurrent neural network

Abstract

The invention provides a diabetes abnormal index early warning system and method based on a recurrent neural network. The system comprises a data acquisition module which is responsible for collecting various health data including blood sugar, heart rate, blood pressure, body weight, body fat, exercise amount, sleep quality and heart rate variability from health monitoring equipment of a patient; the data preprocessing module is used for denoising and normalizing the collected health data and detecting and correcting abnormal values; the feature extraction module is used for extracting time sequence features, statistical features and principal component features from the preprocessed health data; the convolutional gating recurrent neural network model module is used for analyzing the extracted features; and the early warning generation module generates early warning information according to the prediction result. Through the system and the method, the real-time monitoring of the multi-dimensional health data of the diabetic patient and the early warning of the abnormal indexes can be realized, and the diabetes management effect and the life quality of the patient can be improved.

Description

Technical Field

[0001] This invention mainly relates to early warning of abnormal indicators in diabetes, and more particularly to an early warning system and method for abnormal indicators in diabetes based on recurrent neural networks. Background Technology

[0002] Diabetes mellitus is a chronic metabolic disease characterized by persistently elevated blood glucose levels, primarily caused by insufficient insulin secretion or poor insulin action. Diabetes can lead to various complications such as cardiovascular disease, neuropathy, nephropathy, and retinopathy, seriously threatening patients' lives and health. According to statistics from the International Diabetes Federation (IDF), the number of people with diabetes worldwide has exceeded 400 million, and this number continues to grow. Therefore, early detection and effective management of diabetes are of great significance in controlling the occurrence of diabetes and its complications. The development of modern medical technology has diversified the methods for detecting and managing diabetes. Traditional diabetes detection methods mainly rely on fasting blood glucose and glycated hemoglobin (HbA1c) tests. While these methods can reflect a patient's blood glucose level to some extent, they suffer from time lag and the inability to monitor in real time. In recent years, with the widespread use of portable blood glucose meters and continuous glucose monitoring (CGM) devices, patients can measure their blood glucose at home, achieving real-time blood glucose monitoring. Furthermore, the widespread application of wearable devices such as smartwatches and smart bracelets has made monitoring various physiological indicators such as heart rate, blood pressure, exercise volume, and sleep quality more convenient, providing strong support for the comprehensive management of diabetes.

[0003] In current technologies, diabetes management primarily relies on blood glucose level monitoring. Portable blood glucose meters and continuous glucose monitoring (CGM) devices are the most commonly used. Portable blood glucose meters measure blood glucose concentration through finger-prick blood sampling, offering advantages such as ease of operation and low cost, but they only provide instantaneous blood glucose values ​​and cannot reflect fluctuations in blood glucose levels. CGM devices, on the other hand, continuously measure glucose concentration in interstitial fluid using subcutaneous sensors, providing dynamic trends in blood glucose changes and helping patients adjust their treatment plans promptly. Studies have shown that diabetic patients often have cardiovascular dysfunction. Besides blood glucose monitoring, indicators such as heart rate and blood pressure are also important in diabetes management. Heart rate variability (HRV) is a crucial indicator for assessing cardiovascular function. Devices such as smartwatches and smart bracelets can monitor heart rate in real time and calculate HRV, helping to assess a patient's cardiovascular health. Furthermore, hypertension is a common complication of diabetes; wireless blood pressure monitors can conveniently measure and record blood pressure, helping patients control their blood pressure levels and reduce the risk of cardiovascular disease. Weight and body fat are also important indicators for diabetes management. Obesity is one of the main contributing factors to type 2 diabetes. Smart scales and body fat scales allow patients to monitor their weight and body fat changes at any time, and by combining diet and exercise with lifestyle adjustments, they can control their weight and improve glucose metabolism. Exercise volume and sleep quality are closely related to the occurrence and control of diabetes. Appropriate exercise can improve insulin sensitivity and glucose metabolism. Smart bracelets and smartwatches can record data such as steps and exercise volume, helping patients develop scientific exercise plans. Good sleep quality helps maintain normal metabolic function. Smart bracelets can monitor sleep duration and quality, helping patients improve sleep and promote health.

[0004] While existing technologies have played a significant role in the monitoring and management of diabetes, some shortcomings and challenges remain. For example, traditional blood glucose monitoring methods can only provide single blood glucose level data, failing to comprehensively reflect a patient's health status. While devices such as smartwatches and smart bracelets can monitor multiple physiological indicators, this data is often isolated and lacks systematicity and comprehensiveness. Machine learning and artificial intelligence technologies show broad application prospects in the field of health monitoring. Existing research indicates that machine learning algorithms can be used to comprehensively analyze multiple physiological indicators and predict the risk of developing diabetes. For example, algorithms such as Support Vector Machines (SVM), Decision Trees, and Random Forests have achieved some success in diabetes prediction. However, these methods suffer from insufficient accuracy and poor real-time performance when processing large-scale, multi-dimensional time-series data. Recurrent Neural Networks (RNNs) are deep learning models suitable for processing sequential data, especially their variants, Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRUs), which have significant advantages in processing time-series data. By introducing memory units and gating mechanisms, they can capture long-term dependencies in the data, improving prediction accuracy. In recent years, Convolutional Neural Networks (CNNs) have achieved great success in the field of image processing. Their local connectivity and weight sharing characteristics give convolutional operations an advantage in extracting local features. Combining CNNs and Recurrent Neural Networks (RNNs) can fully utilize the local feature extraction capabilities of CNNs and the sequence data modeling capabilities of RNNs to form Convolutional Gated Recurrent Neural Networks (CGRNNs), which exhibit superior performance in time series data prediction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention proposes a diabetes abnormality indicator early warning system and method based on recurrent neural networks. This system integrates multiple health monitoring devices (including portable blood glucose meters, smartwatches, and wireless blood pressure monitors) to collect real-time data on various physiological indicators of patients, including blood glucose, heart rate, blood pressure, weight, body fat, exercise volume, sleep quality, and heart rate variability. After preprocessing and feature extraction, the data is input into a convolutional gated recurrent neural network model for comprehensive analysis and prediction. This model combines the local feature extraction capability of convolutional neural networks with the long-term dependency modeling capability of gated recurrent units, enabling more accurate capture of complex patterns in time-series data. The system generates personalized diabetes abnormality indicator early warning information, which is delivered to patients and doctors through a user interface module, achieving real-time monitoring and timely intervention, thus improving the effectiveness of diabetes management. The system specifically includes the following modules:

[0006] T1: Data acquisition module, used to collect health data from various health monitoring devices of patients, including blood glucose, heart rate, blood pressure, weight, body fat, exercise volume, sleep quality and heart rate variability;

[0007] T2: Data preprocessing module, connected to the data acquisition module, used to perform noise reduction and normalization processing on the acquired health data;

[0008] T3: Feature extraction module, connected to the data preprocessing module, used to extract time series features, statistical features and principal component features from the processed health data;

[0009] T4: Convolutional gated recurrent neural network model module, connected to the feature extraction module, used to predict abnormal indicators of diabetes based on the extracted features and context information;

[0010] T5: Early warning generation module, connected to the convolutional gated recurrent neural network model module, used to generate early warning information of abnormal diabetes indicators based on the prediction results;

[0011] T6: User interface module, connected to the warning generation module, used to transmit warning information and health management suggestions to patients and doctors.

[0012] As a preferred embodiment of the present invention, the health data collected in module T1 uses the following devices: a portable blood glucose meter for measuring and recording the patient's blood glucose level; a smartwatch for monitoring and recording the patient's heart rate, exercise volume, and heart rate variability; a wireless blood pressure monitor for measuring and recording the patient's blood pressure; a smart scale for measuring and recording the patient's weight and body fat; and a smart bracelet for monitoring and recording the patient's sleep quality.

[0013] As a preferred embodiment of the present invention, the convolutional gated recurrent neural network model in module T4 consists of the following modules:

[0014] K1: Input layer: Receives preprocessed time-series health data including blood glucose, heart rate, blood pressure, weight, body fat, exercise volume, sleep quality, and heart rate variability;

[0015] K2: Initial convolutional layer: Performs convolution operation on the input time series data to extract preliminary time features. The convolution kernel size is k×d, where k is the number of time steps, d is the dimension of the data, and the number of convolution kernels is n.

[0016] K3: Batch Normalization Layer: Batch normalizes the output of the initial convolutional layer to balance the distribution of input data across layers and reduce internal covariate bias.

[0017] K4: Deep convolutional layer: Contains multiple consecutive convolutional layers and batch normalization layers. Each convolutional layer performs a convolution operation on the output of the previous layer to extract higher-level temporal features. Each convolutional layer is followed by a batch normalization layer.

[0018] K5: Pooling layer: After the deep convolutional layer, the feature map is downsampled through pooling operation to reduce the feature dimension;

[0019] K6: Gated Recursive Unit Layer: Contains input gates, forget gates, and output gates to capture long-term dependencies in time series data. The gated recursive unit layer receives the output of the pooling layer, processes the sequence data step by step, and maintains the hidden state and memory unit state.

[0020] K7: Context-aware layer: Enhances the model's understanding of time series data by introducing contextual information; this layer combines the output of the gated recursive unit layer with contextual information to adjust the model's behavior;

[0021] K8: Fully connected layer: Processes the output of the context-aware layer, mapping high-dimensional features to a low-dimensional space. The fully connected layer outputs the prediction result.

[0022] K9: Output layer: Generates prediction results of abnormal indicators of diabetes. The output layer receives the results of the fully connected layer and converts them into actual health risk assessment scores or classification results.

[0023] As a preferred embodiment of the present invention, the gated recursive unit layer in module K6 includes an input gate, a forget gate, an output gate, and a context gate to capture long-term dependencies and contextual information in time series data. The calculation models are as follows:

[0024] The input gate determines the importance of the current input information, and the specific calculation formula is as follows:

[0025] i t =σ(W i ·BN n +U i ·h t-1 +C i ·c t-1 +b i )

[0026] Among them, i t W is the input gate activation value. i Given the input weight matrix, BN n U is the output of the last layer of the depthwise convolutional layer. i Let h be the hidden state weight matrix. t-1 C is the hidden state from the previous time step. i Let c be the weight matrix of the memory units. t-1 b represents the state of the memory cell at the previous time step. i σ is the input gate bias term, and σ is the Sigmoid activation function.

[0027] The forgetting gate determines the past information that needs to be forgotten; the specific calculation formula is as follows:

[0028] f t =σ(W f ·BN n +U f ·h t-1 +C f ·c t-1 +b f )

[0029] Among them, f t W is the activation value for the forget gate. f For the forgetting weight matrix, BN n U is the output of the last layer of the depthwise convolutional layer. f Let h be the hidden state weight matrix. t-1 C is the hidden state from the previous time step. f Let c be the weight matrix of the memory units. t-1 b represents the state of the memory cell at the previous time step. f This is the forget gate bias term.

[0030] The output gate determines the output at the current time step, and the specific calculation formula is as follows:

[0031] o t =σ(W o ·BN n +U o ·h t-1 +C o ·c t-1 +b o )

[0032] Among them, o t W is the output gate activation value. o To output the weight matrix, U o Let h be the hidden state weight matrix. t-1 C is the hidden state from the previous time step. o c is the weight matrix of the memory units; t-1 b represents the state of the memory cell at the previous time step. o This is the output gate bias term.

[0033] Context gates adjust the behavior of the gates by introducing context information. The specific calculation formula is as follows:

[0034] φ t =tanh(W φ ·BN n +U φ ·h t-1 +b φ )

[0035] Where, φ t W is the context gate activation value.φ U is the context weight matrix. φ Let b be the hidden state weight matrix. φ is the context gate bias term, and tanh is the hyperbolic tangent activation function.

[0036] As a preferred embodiment of the present invention, a method for early warning of abnormal indicators of diabetes based on recurrent neural networks is implemented using the aforementioned early warning system for abnormal indicators of diabetes based on recurrent neural networks, specifically including the following steps:

[0037] S1: Health Data Collection: Through the data collection module, health data including blood glucose, heart rate, blood pressure, weight, body fat, exercise volume, sleep quality and heart rate variability are collected in real time from the patient's portable blood glucose meter, smartwatch, wireless blood pressure monitor, smart scale and smart bracelet, forming a health data time series;

[0038] S2: Data Preprocessing: The data preprocessing module preprocesses the collected health data, specifically including the following steps:

[0039] S2-1: Denoising: The Kalman filter is used to denoise the time series data including BG, HR, BP, WT, BF, EX, SL and HRV, where BG represents blood glucose, HR represents heart rate, BP represents blood pressure, WT represents weight, BF represents body fat, EX represents exercise volume, SL represents sleep quality and HRV represents heart rate variability.

[0040] S2-2: Normalization processing: Normalize the denoised time series data to convert BG, HR, BP, WT, BF, EX, SL and HRV data to the same scale;

[0041] S2-3: Outlier Detection and Correction: Perform outlier detection, identify and correct outliers in BG, HR, BP, WT, BF, EX, SL and HRV time series data;

[0042] S3: Feature Extraction: The feature extraction module extracts features from the preprocessed health data, specifically including the following steps:

[0043] S3-1: Time Series Characteristics: Analyze the time series characteristics of BG, HR, BP, WT, BF, EX, SL, and HRV data;

[0044] S3-2: Statistical characteristics: Calculate the mean, standard deviation, maximum and minimum values ​​of BG, HR, BP, WT, BF, EX, SL and HRV data;

[0045] S3-3: Principal Component Features: Principal component analysis was performed on BG, HR, BP, WT, BF, EX, SL, and HRV data to extract the main feature components;

[0046] S4: Model Prediction: The extracted features and contextual information are input into the model through the convolutional gated recurrent neural network model module to predict abnormal indicators of diabetes.

[0047] S5: Warning Generation: The warning generation module generates warning information for abnormal diabetes indicators based on the prediction results of the convolutional gated recurrent neural network model module.

[0048] S6: Information Transmission: Through the user interface module, the generated early warning information and health management suggestions are transmitted to patients and doctors, so as to carry out corresponding interventions and management.

[0049] As a preferred embodiment of the present invention, the Kalman filter in step S2 performs noise reduction processing, specifically including the following steps:

[0050] H1: State Prediction: State prediction is performed for each health data point, calculated using the following formula:

[0051]

[0052] in, Let be the predicted state value at time k, representing the predicted values ​​of various health data; A is the state transition matrix; Let B be the state estimate at time k-1; B is the control matrix; u k To control the quantity.

[0053] H2: Covariance Prediction: Covariance prediction is performed for each type of health data. The calculation formula is as follows:

[0054] P k|k-1 =AP k-1|k-1 A T +Q

[0055] Among them, P k|k-1 Let P be the prediction error covariance matrix at time k; Q is the process noise covariance matrix; P is the prediction error covariance matrix at time k; Q is the process noise covariance matrix at time k; P ... k-1|k-1 Let be the error covariance matrix at time k-1.

[0056] H3: Kalman Gain Calculation: Calculate the Kalman gain using the following formula:

[0057] K k =P k|k-1 H T HP k|k-1 H T +R) -1

[0058] Among them, K k Let K be the Kalman gain at time k; H be the observation matrix; and R be the observation noise covariance matrix.

[0059] H4: Status Update: Updates the status estimate of the health data, calculated using the following formula:

[0060]

[0061] in, z is the updated state estimate at time k; k Let k be the actual health data value at time k.

[0062] H5: Covariance Update: Updates the covariance matrix. The calculation formula is as follows:

[0063] P k|k =(IK k H)P k∣k-1

[0064] Among them, P k|k Let be the updated error covariance matrix at time k; I is the identity matrix.

[0065] Through the above steps, the Kalman filter performs noise reduction processing on health data including blood glucose, heart rate, blood pressure, weight, body fat, exercise volume, sleep quality, and heart rate variability.

[0066] As a preferred embodiment of the present invention, the convolutional gated recurrent neural network model module in step S4 specifically includes the following steps:

[0067] S4-1: Input data preparation: Organize the time series features, statistical features and principal component features obtained from the feature extraction module to form an input data matrix;

[0068] S4-2: Input layer processing: The input data matrix is ​​passed to the input layer, which normalizes the data and prepares it for transmission to subsequent layers;

[0069] S4-3: Initial convolutional layer processing: Convolution operation is performed on the input data in the initial convolutional layer to extract preliminary temporal features. One-dimensional convolution is performed on the input data through multiple convolutional kernels to generate feature maps.

[0070] S4-4: Batch normalization layer processing: Batch normalize the output of the initial convolutional layer to balance the distribution of input data in each layer and reduce the internal covariate bias;

[0071] S4-5: Deep convolutional layer processing: The batch-normalized data is passed to the deep convolutional layer, which contains multiple consecutive convolutional layers and batch normalization layers. Each convolutional layer further extracts high-level temporal features, and each convolutional layer is followed by a batch normalization layer to stabilize the training process.

[0072] S4-6: Pooling layer processing: After the deep convolutional layer, the feature map is downsampled through pooling operation to reduce the feature dimension and retain the main features;

[0073] S4-7: Gated Recursive Unit Layer Processing: The pooled feature map is passed to the gated recursive unit layer, which contains input gate, forget gate, output gate and context gate. It processes the sequence data step by step, captures long-term dependencies and context information in the time series data, and maintains the hidden state and memory unit state.

[0074] S4-8: Context-aware layer processing: Through the context-aware layer, contextual information is introduced to enhance the model's understanding of time series data;

[0075] S4-9: Fully connected layer processing: The output of the context-aware layer is passed to the fully connected layer, which processes the output, maps the high-dimensional features to the low-dimensional space, and generates preliminary prediction results;

[0076] S4-10: Output layer generates prediction results: The preliminary prediction results of the fully connected layer are passed to the output layer, which converts the preliminary prediction results into the final health risk assessment score or classification result, generating prediction results for abnormal diabetes indicators.

[0077] Compared with the existing technologies, the present invention has the following main technical advantages, and the beneficial effects of the present invention are:

[0078] Comprehensive monitoring: The system can collect a variety of health data, including blood glucose, heart rate, blood pressure, weight, body fat, exercise volume, sleep quality, and heart rate variability, to comprehensively reflect the patient's health status.

[0079] Real-time alerts: By collecting and analyzing health data in real time, the system can promptly detect abnormal indicators, generate early warning information, and help patients and doctors take timely intervention measures.

[0080] High-precision prediction: By utilizing a convolutional gated recurrent neural network model, the system can effectively capture long-term dependencies and contextual information in time series data, thereby improving the accuracy of predicting abnormal indicators of diabetes.

[0081] Personalized management: The system combines a patient's historical health data with their current status to provide personalized health management suggestions, helping patients optimize their lifestyle and control their diabetes risk.

[0082] Convenient operation: Through smart devices and mobile applications, patients can easily collect and manage data, and doctors can also obtain patients' health data and early warning information through remote platforms for remote diagnosis and guidance. Attached Figure Description

[0083] Figure 1 This is a structural diagram of a diabetes abnormality indicator early warning system based on a recurrent neural network as described in this invention;

[0084] Figure 2 This is a flowchart of a method for early warning of abnormal indicators of diabetes based on recurrent neural networks, as described in this invention.

[0085] Figure 3 This is a structural diagram of a convolutional gated recurrent neural network model module provided in an embodiment of the present invention. Detailed Implementation

[0086] The present invention will be further described below with reference to the accompanying drawings and embodiments. However, the present invention can be implemented in many different ways and should not be construed as limited to the embodiments shown; rather, these embodiments provide those skilled in the art with implementation methods that meet applicable legal requirements.

[0087] Example 1: According to Figure 1 As shown, this embodiment provides a diabetes abnormality indicator early warning system based on a recurrent neural network. The system includes the following main modules: a data acquisition module (T1); a data preprocessing module (T2); a feature extraction module (T3); a convolutional gated recurrent neural network model module (T4); an early warning generation module (T5); and a user interface module (T6). The data acquisition module is responsible for collecting health data from various health monitoring devices of the patient, including blood glucose (BG), heart rate (HR), blood pressure (BP), weight (WT), body fat (BF), exercise volume (EX), sleep quality (SL), and heart rate variability (HRV). Specifically, it includes the following devices:

[0088] Portable blood glucose meters: Used to measure and record a patient's blood glucose levels. Through finger-prick blood sampling, portable blood glucose meters provide instantaneous blood glucose values, helping patients monitor daily blood glucose fluctuations.

[0089] Smartwatches: Used to monitor and record a patient's heart rate, activity level, and heart rate variability. Smartwatches use optical sensors and accelerometers to monitor heart rate and activity levels in real time and calculate heart rate variability.

[0090] Wireless blood pressure monitors: used to measure and record a patient's blood pressure. Wireless blood pressure monitors measure blood pressure via an arm cuff or wrist cuff and transmit the data to a system.

[0091] Smart scales: Used to measure and record a patient's weight and body fat. Smart scales provide weight and body fat data through bioelectrical impedance analysis (BIA) technology.

[0092] Smart bracelets: Used to monitor and record a patient's sleep quality. Smart bracelets assess sleep duration and quality by monitoring movement and heart rate during sleep.

[0093] The data preprocessing module is responsible for denoising and normalizing the collected health data. For example... Figure 2 As shown, the specific steps include:

[0094] I. Denoising Processing: A Kalman filter is used to denoise time series data, including BG, HR, BP, WT, BF, EX, SL, and HRV, reducing the impact of measurement noise. The Kalman filter processes each type of health data through five steps: state prediction, covariance prediction, Kalman gain calculation, state update, and covariance update, thereby improving data accuracy.

[0095] 2. Normalization Processing: The denoised time series data is normalized to transform data from different sources to the same scale. Min-max normalization and Z-Score normalization methods are used to ensure that all data points are within the same range, facilitating subsequent feature extraction and model training.

[0096] III. Outlier Detection and Correction: Outlier detection is performed to identify and correct outliers in the data. The Z-Score method is used to detect outliers, which are then replaced with nearest-neighbor values ​​or the median to ensure the rationality and continuity of the data.

[0097] The feature extraction module extracts time series features, statistical features, and principal component features from the preprocessed health data, specifically including the following steps:

[0098] I. Time Series Feature Extraction: Analyze the time series features of BG, HR, BP, WT, BF, EX, SL, and HRV data. Extract trend features using methods such as moving average and exponential smoothing, and calculate the periodicity of the data.

[0099] II. Statistical Feature Extraction: Calculate the statistical features of each data point, such as mean, standard deviation, maximum value, and minimum value, to provide a description of the overall distribution of the data.

[0100] III. Principal Component Feature Extraction: Perform Principal Component Analysis (PCA) on each data point to extract the main feature components, reduce data dimensionality, decrease redundant information, and improve data processing efficiency.

[0101] The Convolutional Gated Recurrent Neural Network (CGRNN) model module is the core of the system, responsible for predicting abnormal indicators of diabetes based on extracted features and contextual information. Its specific structure is as follows:

[0102] Input layer: Receives preprocessed time-series health data, including BG, HR, BP, WT, BF, EX, SL, and HRV data.

[0103] Initial convolutional layer: Performs convolution operations on the input data to extract preliminary temporal features. The kernel size is k×d, where k is the number of time steps, d is the dimension of the data, and the number of kernels is n.

[0104] Batch normalization layer: The output of the initial convolutional layer is batch normalized to balance the distribution of input data in each layer, reduce internal covariate bias, and improve training stability.

[0105] Deep convolutional layers: contain multiple consecutive convolutional layers and batch normalization layers. Each convolutional layer further extracts high-level temporal features, and each convolutional layer is followed by a batch normalization layer.

[0106] Pooling layer: The feature map is downsampled through pooling operations to reduce the feature dimension and retain the main features.

[0107] The gated recursive unit layer contains input gates, forget gates, output gates, and context gates to capture long-term dependencies and contextual information in time-series data. The gated recursive unit layer receives the output of the pooling layer, processes the sequence data step-by-step, and maintains the hidden state and memory unit state.

[0108] Context-aware layer: Introduces contextual information to enhance the model's understanding of time-series data. This layer combines the output of the gated recursive unit layer with contextual information to adjust the model's behavior.

[0109] Fully connected layer: processes the output of the context-aware layer, maps high-dimensional features to a low-dimensional space, and generates preliminary prediction results.

[0110] Output layer: Generates prediction results of abnormal indicators of diabetes. The output layer receives the results of the fully connected layer and converts them into actual health risk assessment scores or classification results.

[0111] The early warning generation module generates early warning information for abnormal diabetes indicators based on the prediction results of the convolutional gated recurrent neural network model module. This includes the following steps:

[0112] Threshold setting: Set warning thresholds for health indicators based on medical standards and individual differences. For example, set normal and abnormal ranges for indicators such as blood sugar, heart rate, and blood pressure.

[0113] Anomaly detection: When a health indicator exceeds the warning threshold, an early warning message is generated to alert the patient and doctor.

[0114] Personalized recommendations: By combining a patient's historical health data and current status, personalized health management recommendations are generated to help patients adjust their lifestyles and control their diabetes risk.

[0115] The user interface module is responsible for delivering alerts and health management recommendations to patients and doctors. Specific methods include:

[0116] Mobile Application: A mobile application sends alerts and health management suggestions to patients, allowing them to view and manage their health status in real time. The application also provides visualizations of health data, helping patients understand changes in their health more intuitively.

[0117] Doctor-side platform: This platform provides doctors with patient health monitoring data and early warning information, facilitating remote diagnosis and guidance. Doctors can view patients' historical data and real-time monitoring results through the platform to develop personalized treatment plans.

[0118] Example 2: According to Figure 3 As shown, this embodiment provides a convolutionally gated recurrent neural network model module, specifically including the following: The convolutionally gated recurrent neural network (CGRNN) model combines the advantages of convolutional neural networks (CNN) and gated recurrent neural networks (GRNN) for processing multidimensional time series data. The model includes the following main layers: Input Layer; Initial Convolutional Layer; Batch Normalization Layer; Deep Convolutional Layers; Pooling Layer; Gated Recurrent Unit Layer; Contextual Awareness Layer; Fully Connected Layer; and Output Layer.

[0119] The input layer receives preprocessed time-series health data. The input data has a dimension of T×D, where T is the number of time steps and D is the feature dimension. Specifically, the health data received by the input layer includes blood glucose (BG), heart rate (HR), blood pressure (BP), weight (WT), body fat (BF), exercise volume (EX), sleep quality (SL), and heart rate variability (HRV). The input layer is primarily responsible for formatting and normalizing this multidimensional data, preparing it for subsequent convolutional operations.

[0120] The loss function is used to evaluate the difference between the model's predicted results and the actual labels, guiding the update of model parameters. This model uses the cross-entropy loss function, calculated as follows:

[0121]

[0122] Where L is the cross-entropy loss; N is the number of samples; C is the number of classes; y ij The actual label for sample i is the value (0 or 1) of the j-th class. Let be the predicted probability of sample i, and be the value of class j.

[0123] The following is the specific implementation process of the model:

[0124] Health Data Collection: Using devices such as portable blood glucose meters, smartwatches, wireless blood pressure monitors, smart scales, and smart bracelets, real-time health data including blood glucose, heart rate, blood pressure, weight, body fat, exercise volume, sleep quality, and heart rate variability is collected. This data is then wirelessly transmitted to the system to create a time-series health data dataset.

[0125] Data preprocessing: The collected health data is denoised using a Kalman filter to reduce the impact of measurement noise. The denoised data is then normalized to convert data from different sources to the same scale. Outlier detection is performed to identify and correct abnormal values ​​in the data, ensuring the data's rationality and continuity.

[0126] Feature extraction: Time series features, statistical features, and principal component features are extracted from the preprocessed health data. Trend features are extracted using methods such as moving average and exponential smoothing, and the periodicity of the data is calculated. Statistical features such as mean, standard deviation, maximum, and minimum values ​​are calculated for each data point. Principal component analysis (PCA) is performed on each data point to extract the main feature components and reduce data dimensionality.

[0127] Model prediction: The extracted features are input into a convolutional gated recurrent neural network model. The model processes the data layer by layer through the input layer, initial convolutional layer, batch normalization layer, deep convolutional layer, pooling layer, gated recurrent unit layer, context-aware layer, fully connected layer, and output layer. The output layer generates the predicted results of abnormal indicators of diabetes.

[0128] Warning Generation: Generates warning information for abnormal diabetes indicators based on model predictions. Set warning thresholds for health indicators; when an indicator exceeds the threshold, a warning message is generated. Provides personalized health management recommendations by combining the patient's historical data and current status.

[0129] Information Delivery: Warning information and health management suggestions are delivered to patients and doctors through the user interface module. Patients can view and manage their health status in real time via the mobile application. Doctors access patient health data and warning information through the doctor's platform for remote diagnosis and guidance.

[0130] Through the above process, the convolutional gated recurrent neural network model can integrate multidimensional features to accurately predict abnormal indicators of diabetes, providing a foundation for subsequent early warning generation.

[0131] To achieve optimal performance, specific parameters of the model need to be set and tuned. Below are some key parameters and their setting methods:

[0132] Kernel size (k): The kernel size for the initial convolutional layer and the deep convolutional layer needs to be set according to the time step and feature dimension of the data. Generally, smaller kernels can capture local features, while larger kernels can capture global features. Common settings include 3, 5, and 7.

[0133] Number of convolutional kernels (n): The number of convolutional kernels determines the complexity of feature extraction at each layer. More convolutional kernels can extract more features, but also increase the computational cost. Common settings range from 32 to 256.

[0134] Pooling window size (p): The pooling window size determines the downsampling rate of the feature map. A larger pooling window can significantly reduce the feature dimensionality, but may lose some information. Common settings include 2, 3, and 5.

[0135] Number of GRU cells (g): The number of GRU cells determines the memory capacity of the gated recursive cell layer. More GRU cells can capture more complex time dependencies, but also increase computational cost. Common settings range from 32 to 256.

[0136] Learning rate: The learning rate is a key parameter affecting the training speed and stability of a model. A higher learning rate can speed up training but may lead to training instability; a lower learning rate can stabilize training but may lead to training that is too slow. Common settings range from 0.001 to 0.1.

[0137] Optimizer: Choosing the right optimizer can improve the efficiency and effectiveness of model training. Common optimizers include SGD, Adam, and RMSprop.

[0138] Regularization parameters: To prevent overfitting, regularization terms can be added. Common regularization methods include L1 regularization, L2 regularization, and Dropout.

[0139] Example 3: To verify the effectiveness of the convolutional gating recurrent neural network model of the present invention, a simulation experiment was conducted. The experimental data came from a real health monitoring dataset, including long-term health monitoring data from multiple diabetic patients. The specific steps and calculation results of the experiment are as follows:

[0140] For data preparation, long-term health monitoring data of 50 diabetic patients were collected. The data for each patient included the following indicators: blood glucose (BG), heart rate (HR), blood pressure (BP), weight (WT), body fat (BF), exercise volume (EX), sleep quality (SL), and heart rate variability (HRV). Data for each indicator was collected once a day for a period of one year.

[0141] The collected health data underwent preprocessing, including noise reduction, normalization, and outlier detection and correction. The specific calculation results for each step are as follows:

[0142] The Kalman filter is used to denoise the data for each indicator to reduce the impact of measurement noise, as shown in Table 1.

[0143] Table 1. Noise Reduction Results of Patient Blood Glucose Data

[0144]

[0145] The denoised data were normalized to convert data from different sources to the same scale. The min-max normalization method was used, as shown in Table 2.

[0146] Table 2. Results of Normalization of Patient Health Data

[0147]

[0148] Outlier Detection and Correction: Outlier detection is performed to identify and correct outliers in the data. The Z-Score method is used to detect outliers, as shown in Table 3.

[0149] Table 3 Results of outlier detection and correction for patient blood pressure data

[0150]

[0151]

[0152] Feature extraction involves extracting time-series features, statistical features, and principal component features from the preprocessed health data. The following are the specific calculation results for each step:

[0153] Time series feature extraction: Analyze the time series features of each data point, and extract trend features through moving average and exponential smoothing, as shown in Table 4.

[0154] Table 4. Results of time-series feature extraction of patient weight data

[0155]

[0156] Statistical feature extraction: Calculate statistical features such as mean, standard deviation, maximum, and minimum values ​​for each data point. See Table 5 for example.

[0157] Table 5. Statistical feature extraction results of a patient's sleep quality data.

[0158]

[0159] The extracted features are input into a convolutional gated recurrent neural network model module for predicting abnormal indicators of diabetes. The gated recurrent unit layer contains an input gate, a forget gate, and an output gate to capture long-term dependencies in time-series data, progressively process the sequence data, and maintain hidden states and memory unit states. Specific intermediate results are shown in Table 6.

[0160] Table 6 Calculation results of the gated recursive unit layer

[0161]

[0162] The predicted results of abnormal indicators of diabetes are generated, and the results of the fully connected layer are converted into actual health risk assessment scores or classification results. The specific results are shown in Table 7.

[0163] Table 7 Model Output Prediction Results

[0164]

[0165] The above simulation experiments verified the effectiveness of the convolutional gating recurrent neural network model of this invention in early warning of abnormal indicators in diabetes. Experimental results show that, after data preprocessing, feature extraction, and model prediction, this system can accurately predict the health status of diabetic patients, provide timely early warning information, and effectively improve the effectiveness of diabetes management. The table shows the detailed calculation results in the data preprocessing and model prediction processes, ensuring the transparency and traceability of each step.

[0166] The above embodiments are merely illustrative of several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A recurrent neural network based early warning system for diabetes abnormal indicators, characterized in that: The system comprises the following modules: T1: a data collection module for collecting health data including blood glucose, heart rate, blood pressure, weight, body fat, exercise amount, sleep quality and heart rate variability from various health monitoring devices of a patient; T2: a data preprocessing module connected with the data collection module for denoising and normalizing the collected health data; T3: a feature extraction module connected with the data preprocessing module for extracting time series features, statistical features and principal component features from the processed health data; T4: a convolutional gated recurrent neural network model module connected with the feature extraction module for predicting diabetes abnormal indicators based on the extracted features and context information; T5: an early warning generation module connected with the convolutional gated recurrent neural network model module for generating diabetes abnormal indicator early warning information based on the prediction results; T6: a user interface module connected with the early warning generation module for delivering the early warning information and health management suggestions to the patient and the doctor.

2. The diabetes anomaly indicator early warning system based on a recurrent neural network according to claim 1, characterized in that: The health data in module T1 is collected by the following devices respectively: a portable blood glucose meter for measuring and recording the patient's blood glucose level; a smart watch for monitoring and recording the patient's heart rate, exercise amount and heart rate variability; a wireless blood pressure meter for measuring and recording the patient's blood pressure; a smart weight scale for measuring and recording the patient's weight and body fat; a smart bracelet for monitoring and recording the patient's sleep quality. 3.The diabetes abnormal index early warning system based on recurrent neural network according to claim 1, characterized in that: The convolutional gated recurrent neural network model in module T4 is composed of the following modules: K1: an input layer receiving preprocessed time series health data including blood glucose, heart rate, blood pressure, weight, body fat, exercise amount, sleep quality and heart rate variability; K2: an initial convolutional layer performing convolution operation on the input time series data to extract preliminary time features, with a convolution kernel size of k x d, where k is the time step and d is the dimension of the data, and the number of convolution kernels is n; K3: a batch normalization layer balancing the input data distribution of each layer and reducing internal covariate shift by performing batch normalization on the output of the initial convolutional layer; K4: a deep convolutional layer containing multiple consecutive convolutional layers and batch normalization layers, where each convolutional layer performs convolution operation on the output of the previous layer to extract higher-level time features, and each convolutional layer is followed by a batch normalization layer; K5: a pooling layer performing down-sampling on the feature map through pooling operation after the deep convolutional layer to reduce the feature dimension; K6: a gated recurrent unit layer containing input gates, forget gates, output gates and context gates to capture long-term dependencies in time series data, which receives the output of the pooling layer, processes the sequence data step by step, and maintains the hidden state and memory cell state; K7: a context-aware layer that enhances the model's understanding of time series data by introducing context information; this layer combines the output of the gated recurrent unit layer with the context information to adjust the behavior of the model; K8: a fully connected layer processing the output of the context-aware layer to map high-dimensional features to a low-dimensional space, where the fully connected layer outputs the prediction results; K9: output layer, generating the prediction result of the diabetes abnormal index, the output layer receives the result of the full connection layer and converts it into an actual health risk assessment score or a classification result.

4. The diabetes anomaly indicator early warning system based on recurrent neural network according to claim 3, characterized in that: The gated recurrent unit layer in module K6 includes an input gate, a forget gate, an output gate, and a context gate to capture long-term dependencies and context information in time series data, and the calculation model is as follows: The input gate determines the importance of the current input information, and the specific calculation formula is: i t = σ(W i · BN n + U i · h t-1 + C i · c t-1 + b i ) where i t is the input gate activation value, W i is the input weight matrix, BN n is the output of the last layer of the deep convolutional layer, U i is the hidden state weight matrix, h t-1 is the hidden state of the previous time step, C i is the memory cell weight matrix, c t-1 is the memory cell state of the previous time step, b i is the input gate bias term, and σ is the Sigmoid activation function. The forget gate determines the past information that needs to be forgotten, and the specific calculation formula is: f t = σ(W f · BN n + U f · h t-1 + C f · c t-1 + b f ) wherein f t is the forget gate activation, W f is the forget gate weight matrix, BN n is the output of the last layer of the deep convolutional layer, U f is the hidden state weight matrix, h t-1 is the hidden state of the previous time step, C f is the memory cell weight matrix, c t-1 is the memory cell state of the previous time step, b f is the forget gate bias term; The output gate determines the output of the current time step, and the specific calculation formula is: o t = σ(W o · BN n + U o · h t-1 + C o · c t-1 + b o ) where o t is the output gate activation, W o is the output weight matrix, U o is the hidden state weight matrix, h t-1 is the hidden state of the previous time step, C o is the memory cell weight matrix, c t-1 is the memory cell state of the previous time step, b o is the output gate bias term; The context gate adjusts the behavior of the above gates by introducing context information, and the specific calculation formula is as follows: φ t = tanh(W φ · BN n + U φ · h t-1 + b φ ) where φ t is the contextual gate activation, W φ is the contextual weight matrix, U φ is the hidden state weight matrix, b φ is the contextual gate bias term, and tanh is the hyperbolic tangent activation function. 5.A method for early warning of diabetes abnormal indicators based on a recurrent neural network, characterized in that: The method is implemented by using the system of claim 1, and specifically includes the following steps: S1: health data collection: through the data collection module, real-time collection of health data including blood glucose, heart rate, blood pressure, body weight, body fat, exercise amount, sleep quality and heart rate variability from the patient's portable blood glucose meter, smart watch, wireless blood pressure meter, smart body scale and smart bracelet and other devices, forming a health data time series; S2: data preprocessing: through the data preprocessing module, the collected health data is preprocessed, specifically including the following steps: S2-1: denoising processing: using Kalman filter to denoise the BG, HR, BP, WT, BF, EX, SL and HRV time series data, wherein BG represents blood glucose, HR represents heart rate, BP represents blood pressure, WT represents body weight, BF represents body fat, EX represents exercise amount, SL represents sleep quality, and HRV represents heart rate variability; S2-2: normalization processing: normalizing the denoised time series data, converting BG, HR, BP, WT, BF, EX, SL and HRV data to the same scale; S2-3: outlier detection and correction: performing outlier detection, identifying and correcting outliers in BG, HR, BP, WT, BF, EX, SL and HRV time series data; S3: feature extraction: through the feature extraction module, extracting features from the preprocessed health data, specifically including the following steps: S3-1: time series feature: analyzing the time series features of BG, HR, BP, WT, BF, EX, SL and HRV data; S3-2: statistical feature: calculating the mean, standard deviation, maximum and minimum of BG, HR, BP, WT, BF, EX, SL and HRV data; S3-3: principal component feature: principal component analysis of BG, HR, BP, WT, BF, EX, SL and HRV data to extract main feature components; S4: model prediction: through the convolutional gated recurrent neural network model module, inputting the extracted features and context information into the model for diabetes abnormal index prediction; S5: early warning generation: through the early warning generation module, generating diabetes abnormal index early warning information according to the prediction result of the convolutional gated recurrent neural network model module; S6: Information transmission: Through the user interface module, the generated early warning information and health management suggestions are transmitted to the patient and the doctor, so as to carry out corresponding intervention and management.

6. The method of claim 5, wherein the method comprises: The Kalman filter in step S2 carries out denoising processing, specifically Comprising the following steps: H1: State prediction: The state of each health data is predicted, and the calculation formula is as follows: wherein, is the state prediction value at time k, representing the predicted value of each health data; A is a state transition matrix; is the state estimation value at time k-1; B is a control matrix; u k is the control amount; H2: Covariance prediction: The covariance of each health data is predicted, and the calculation formula is as follows: P k|k-1 = AP k-1|k-1 A T + Q where P k|k-1 is the prediction error covariance matrix at time k; Q is the process noise covariance matrix; P k-1|k-1 is the error covariance matrix at time k - 1. H3: Kalman gain calculation: The Kalman gain is calculated, and the calculation formula is as follows: K k = P k|k-1 H T (HP k|k-1 H T + R) -1 Among them, K k Let K be the Kalman gain at time k; H be the observation matrix; and R be the observation noise covariance matrix. H4: State update: The state estimation value of the health data is updated, and the calculation formula is as follows: wherein, is the updated state estimate at time k; z k is the actual health data value at time k; H5: Covariance update: The covariance matrix is updated, and the calculation formula is as follows: P kk = (I - K k H)P k∣k-1 where P kk is the updated error covariance matrix at time k; I is the identity matrix.

7. The method of claim 5, wherein the method comprises: The convolutional gated recurrent neural network model module in step S4 specifically comprises the following steps: S4-1: Input data preparation: The time series features, statistical features and principal component features obtained from the feature extraction module are arranged to form an input data matrix; S4-2: Input layer processing: The input data matrix is transmitted to the input layer, and the input layer normalizes the data and prepares to transmit to the subsequent layer; S4-3: Initial convolution layer processing: The input data is convolved in the initial convolution layer to extract preliminary time features, and a feature map is generated by one-dimensional convolution of the input data through multiple convolution kernels; S4-4: Batch normalization layer processing: The output of the initial convolution layer is batch normalized to balance the input data distribution of each layer and reduce the internal covariate shift; S4-5: Deep convolution layer processing: The batch-normalized data is transmitted to the deep convolution layer, which includes multiple consecutive convolution layers and batch normalization layers. Each convolution layer further extracts high-level time features, and each convolution layer is followed by a batch normalization layer to stabilize the training process; S4-6: Pooling layer processing: After the deep convolution layer, the feature map is down-sampled by the pooling operation to reduce the feature dimension and retain the main features; S4-7: Gated recurrent unit layer processing: The pooled feature map is transmitted to the gated recurrent unit layer, which includes an input gate, a forget gate, an output gate and a context gate. The sequence data is processed step by step to capture long-term dependencies and context information in the time series data, and to maintain hidden state and memory cell state; S4-8: Context awareness layer processing: Through the context awareness layer, context information is introduced to enhance the model's understanding of time series data; S4-9: Fully connected layer processing: The output of the context awareness layer is transmitted to the fully connected layer, which processes the output and maps high-dimensional features to a low-dimensional space to generate preliminary prediction results; S4-10: Output layer generates prediction results: The preliminary prediction results of the fully connected layer are transmitted to the output layer, which converts the preliminary prediction results into the final health risk assessment score or classification result to generate the prediction results of the diabetes abnormal indicators.

Images