An artificial intelligence-based blood glucose change prediction system and method

Abstract

The application provides a blood glucose change prediction system and method based on artificial intelligence. Real-time blood glucose change data is collected by a wearable CGM sensor. Based on the blood glucose change data, time sequence hidden features representing blood glucose time sequence fluctuation implicit driving signals are extracted. Blood glucose fluctuation phase recognition is performed according to the blood glucose change data and the time sequence hidden features, and a fluctuation phase state is obtained. Based on the fluctuation phase state, a phase offset of the blood glucose change data relative to a standard time sequence fluctuation reference is determined. The blood glucose change data is phase calibrated according to the phase offset, and blood glucose change features are obtained. Based on the blood glucose change features, a blood glucose change prediction result in a future preset time range is generated and displayed through a monitoring terminal. The technical solution provided by the application can recognize blood glucose fluctuation phase and perform phase calibration, thereby improving the accuracy of the blood glucose change prediction result.

Description

Technical Field

[0001] This application relates to the field of blood glucose monitoring and artificial intelligence prediction technology, and more specifically, this application relates to a blood glucose change prediction system and method based on artificial intelligence. Background Technology

[0002] Continuous glucose monitoring (CGM) technology can continuously collect blood glucose change data through wearable CGM sensors and display blood glucose change trends to users through monitoring terminals. Compared with single-point detection methods, continuous glucose monitoring can provide a more complete time-series data basis, which helps to identify blood glucose rising, falling and stabilizing states.

[0003] However, data collected by wearable CGM sensors typically suffers from time lag, local noise, and short-term fluctuations, potentially leading to phase shifts between blood glucose variability data and actual time-series fluctuation benchmarks. Existing blood glucose prediction methods often rely directly on raw monitoring data or simply filtered data for trend extrapolation, making it difficult to fully identify the implicit driving signals within time-series blood glucose fluctuations and to perform phase calibration for different fluctuation phase states. This leaves room for improvement in the accuracy and stability of future blood glucose variability prediction results. Therefore, identifying and calibrating the phase of blood glucose fluctuations to enhance the accuracy of blood glucose variability prediction has become a significant challenge for the industry. Summary of the Invention

[0004] This application provides an artificial intelligence-based blood glucose change prediction system and method, which can identify the phase of blood glucose fluctuations and perform phase calibration to improve the accuracy of blood glucose change prediction results.

[0005] In a first aspect, this application provides an artificial intelligence-based method for predicting blood glucose changes, comprising the following steps: Real-time blood glucose change data is collected using a wearable CGM sensor; Based on the blood glucose change data, temporal latent features characterizing the implicit driving signal of blood glucose temporal fluctuations are extracted. Blood glucose fluctuation phase is identified according to the blood glucose change data and the temporal latent features to obtain the fluctuation phase state. Based on the fluctuation phase state, the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference is determined, and the blood glucose change data is phase-calibrated according to the phase offset to obtain blood glucose change characteristics; Based on the blood glucose change characteristics, a prediction result of blood glucose change within a preset time range is generated and displayed on the monitoring terminal.

[0006] In some embodiments, real-time acquisition of blood glucose change data via a wearable CGM sensor specifically includes: Raw glucose concentration data are collected using a wearable CGM sensor at a preset sampling period. The monitoring terminal performs preprocessing on the raw glucose concentration data to generate blood glucose change data.

[0007] In some embodiments, extracting temporal latent features characterizing the implicit driving signal of blood glucose temporal fluctuations based on the blood glucose change data specifically includes: The blood glucose change data are subjected to signal quality assessment to obtain signal quality weights that characterize the sampling integrity, local noise level, and sensor status influence. The blood glucose sampling points within the most recent preset time period are extracted from the blood glucose change data to generate a blood glucose subsequence, and the time-series fluctuation index is extracted from the blood glucose subsequence to obtain blood glucose morphology characteristics. Obtain state association data corresponding to the time of the blood glucose subsequence, and generate hidden event features based on the blood glucose morphology features, the state association data, and the current time period information; The blood glucose morphology features, the latent event features, and the signal quality weights are concatenated to obtain the temporal latent features that characterize the latent driving signal of blood glucose temporal fluctuations.

[0008] In some embodiments, blood glucose fluctuation phase identification based on the blood glucose change data and the temporal latent features to obtain the fluctuation phase state specifically includes: The blood glucose change data and the temporal latent features are input into the fluctuation phase recognition model to obtain several fluctuation phase labels and corresponding phase confidence levels. When several wave phase tags meet the preset composite phase determination conditions, the corresponding phase weights are assigned according to the phase reliability of each wave phase tag. By combining all the fluctuation phase labels and their corresponding phase confidence and phase weights, the fluctuation phase state is obtained.

[0009] In some embodiments, determining the phase offset of the blood glucose change data relative to a standard time-series fluctuation benchmark based on the fluctuation phase state specifically includes: From historical blood glucose monitoring data, obtain the historical response templates corresponding to each fluctuation phase label in the fluctuation phase state; Multiple candidate offsets are pre-set. For each candidate offset, the blood glucose change data is time-aligned with the historical response template corresponding to each fluctuation phase label to determine the phase matching degree of the candidate offset under the corresponding fluctuation phase label. For the same candidate offset, all phase matching degrees are fused to obtain the fused phase matching degree corresponding to the candidate offset; From multiple candidate offsets, select the candidate offset whose fusion phase matching degree meets the preset matching condition as the initial phase offset; By applying a continuity constraint to the initial phase offset, the phase offset of the blood glucose change data relative to the standard time-series fluctuation benchmark is obtained.

[0010] In some embodiments, performing phase calibration on the blood glucose change data based on the phase offset to obtain blood glucose change characteristics specifically includes: By combining the phase offset with the blood glucose change data, a phase calibration sequence with a unified time reference is generated; For each fluctuation phase label in the fluctuation phase state, a calibration response feature matching the fluctuation phase label is extracted from the phase calibration sequence, and a corresponding calibration reliability sub-feature is generated based on the calibration response feature; The calibration response features corresponding to each fluctuation phase label are integrated to obtain the overall composite phase response features. The calibration reliability sub-features corresponding to each fluctuation phase label are integrated to obtain the overall calibration reliability features. The composite phase response feature, the calibration reliability feature, and the phase offset are combined to obtain the blood glucose change feature.

[0011] In some embodiments, generating a predicted blood glucose level change within a preset time range based on the blood glucose change characteristics and displaying it on a monitoring terminal specifically includes: The blood glucose change characteristics are input into the blood glucose change prediction model, which outputs the predicted blood glucose values ​​and trend directions for multiple prediction times within a preset future time range. Based on the blood glucose change characteristics and the predicted blood glucose values ​​and trend directions at multiple prediction times, blood glucose change prediction results are generated and displayed through a monitoring terminal.

[0012] Secondly, this application provides an artificial intelligence-based blood glucose variability prediction system for executing an artificial intelligence-based blood glucose variability prediction method, including: The data acquisition module is used to collect blood glucose change data in real time through a wearable CGM sensor; The phase recognition module is used to extract the temporal latent features that characterize the implicit driving signal of blood glucose temporal fluctuations based on the blood glucose change data, and to perform blood glucose fluctuation phase recognition based on the blood glucose change data and the temporal latent features to obtain the fluctuation phase state. The phase calibration module is used to determine the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference based on the fluctuation phase state, and to perform phase calibration on the blood glucose change data according to the phase offset to obtain blood glucose change characteristics. The result prediction module is used to generate a prediction result of blood glucose changes within a preset time range based on the blood glucose change characteristics and display it on the monitoring terminal.

[0013] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described artificial intelligence-based blood glucose change prediction method.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned artificial intelligence-based method for predicting blood glucose changes.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This application provides an artificial intelligence-based blood glucose change prediction system and method, which collects blood glucose change data in real time using a wearable CGM sensor; based on the blood glucose change data, extracts temporal latent features characterizing the implicit driving signal of blood glucose temporal fluctuations; performs blood glucose fluctuation phase identification based on the blood glucose change data and the temporal latent features to obtain the fluctuation phase state; based on the fluctuation phase state, determines the phase offset of the blood glucose change data relative to a standard temporal fluctuation benchmark; performs phase calibration on the blood glucose change data based on the phase offset to obtain blood glucose change features; and based on the blood glucose change features, generates blood glucose change prediction results within a preset future time range and displays them on a monitoring terminal.

[0016] Therefore, this application demonstrates that, firstly, the quality weights generated through signal quality assessment can reduce the interference of distorted sampling data on phase recognition; when behavioral event records are missing, latent event features, relying on blood glucose morphology features, state-related data, and time period information to infer potential perturbation factors, can reduce recognition errors caused by missing data; and using a time-series classification model for fluctuation phase recognition avoids biases caused by relying solely on instantaneous concentration values ​​or single rates of change. Secondly, setting candidate offsets based on the initial sampling time of the blood glucose subsequence, combined with a dynamic time warping algorithm to calculate the phase matching degree, can quantify the morphological differences between the current blood glucose data and historical templates; weighted fusion of the matching degree according to the phase weights of each label reflects the differences in contribution of different fluctuation stages, avoiding the one-sidedness of single label matching; and the pairing extraction of calibration response features and calibration confidence sub-features, combined with data quality, label confidence, and matching residuals to comprehensively evaluate calibration reliability, automatically reducing the contribution of low-quality data fragments and improving prediction robustness. Finally, the calibrated blood glucose change features are input into the prediction model, and predictions are made based on time-aligned data to avoid prediction bias caused by time-series offsets and improve prediction accuracy. The model uses a long short-term memory network or a gated recurrent unit to capture the temporal dependencies of blood glucose data and adapt to the dynamic changes in blood glucose fluctuations.

[0017] In summary, the technical solution adopted in this application can identify the phase of blood glucose fluctuations and perform phase calibration to improve the accuracy of blood glucose change prediction results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an exemplary flowchart of an artificial intelligence-based method for predicting blood glucose changes according to some embodiments of this application; Figure 2 This is a schematic diagram illustrating an application scenario of an artificial intelligence-based blood glucose change prediction method according to some embodiments of this application; Figure 3 This is an exemplary flowchart illustrating the determination of phase offset according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of an artificial intelligence-based blood glucose change prediction system according to some embodiments of this application; Figure 5This is a schematic diagram of the structure of a computer device that implements an artificial intelligence-based method for predicting blood glucose changes, according to some embodiments of this application. Detailed Implementation

[0020] 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 a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] This application provides an artificial intelligence-based blood glucose variability prediction system and method. The core of this system involves real-time acquisition of blood glucose variability data using a wearable CGM sensor. Based on this data, temporal latent features characterizing the implicit driving signals of blood glucose temporal fluctuations are extracted. Blood glucose fluctuation phases are identified based on the blood glucose variability data and the temporal latent features to obtain the fluctuation phase state. Based on the fluctuation phase state, the phase offset of the blood glucose variability data relative to a standard temporal fluctuation benchmark is determined. The blood glucose variability data is then phase-calibrated based on the phase offset to obtain blood glucose variability features. Based on these features, a blood glucose variability prediction result within a preset future time range is generated and displayed on a monitoring terminal. This approach can identify blood glucose fluctuation phases and perform phase calibration, thereby improving the accuracy of blood glucose variability prediction results.

[0022] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 The figure is an exemplary flowchart of an artificial intelligence-based blood glucose change prediction method according to some embodiments of this application. The figure mainly includes the following steps: In step S101, blood glucose change data are collected in real time using a wearable CGM sensor.

[0023] In some embodiments, the real-time acquisition of blood glucose change data via a wearable CGM sensor can be achieved in the following ways: Raw glucose concentration data are collected using a wearable CGM sensor at a preset sampling period. The monitoring terminal performs preprocessing on the raw glucose concentration data to generate blood glucose change data.

[0024] In this application, Figure 2This is a schematic diagram of an application scenario for an artificial intelligence-based blood glucose change prediction method according to some embodiments of this application; as shown in the figure: the wearable CGM sensor is responsible for collecting raw glucose concentration data in real time; the monitoring terminal receives the raw glucose concentration data and performs preprocessing to generate blood glucose change data; wherein, the monitoring terminal is an electronic device with data receiving, data processing and result display capabilities, including any one of smartphones, tablets, smartwatches, and dedicated blood glucose monitors.

[0025] In practice, firstly, a wearable CGM sensor is attached to the surface of the subject's body. It uses electrochemical detection to collect electrical signals generated by the reaction of glucose and enzymes in interstitial fluid, which are then converted into glucose concentration values. The sensor's built-in processing unit collects glucose concentration values ​​at a preset sampling period, for example, once per minute. The sensor's built-in near-end communication module sends the raw glucose concentration data and its corresponding timestamp to the monitoring terminal. Then, the monitoring terminal performs a preprocessing procedure on the received raw glucose concentration data, including: removing abnormal jump values ​​caused by poor sensor contact or signal interference; interpolating and compensating for short-term missing data due to communication packet loss; and filtering and smoothing high-frequency noise in the sensor signal. After all preprocessing operations are completed, uniformly formatted, time-continuous, and reliable blood glucose change data is generated.

[0026] In step S102, based on the blood glucose change data, temporal latent features characterizing the implicit driving signal of blood glucose temporal fluctuations are extracted, and blood glucose fluctuation phase identification is performed according to the blood glucose change data and the temporal latent features to obtain the fluctuation phase state.

[0027] In some embodiments, extracting the temporal latent features characterizing the implicit driving signal of blood glucose temporal fluctuations based on the blood glucose change data can be specifically done in the following manner: The blood glucose change data are subjected to signal quality assessment to obtain signal quality weights that characterize the sampling integrity, local noise level, and sensor status influence. The blood glucose sampling points within the most recent preset time period are extracted from the blood glucose change data to generate a blood glucose subsequence, and the time-series fluctuation index is extracted from the blood glucose subsequence to obtain blood glucose morphology characteristics. Obtain state association data corresponding to the time of the blood glucose subsequence, and generate hidden event features based on the blood glucose morphology features, the state association data, and the current time period information; The blood glucose morphology features, the latent event features, and the signal quality weights are concatenated to obtain the temporal latent features that characterize the latent driving signal of blood glucose temporal fluctuations.

[0028] It should be noted that the signal quality weight is used to characterize the reliability of blood glucose change data at each sampling time, and is jointly determined by sampling integrity, local noise level and sensor state influence; the blood glucose morphology feature is used to describe the trend and fluctuation pattern of blood glucose sequence changes within the most recent preset time period; the implicit event feature refers to feature data generated based on blood glucose morphology feature, state association data and current time period information, used to characterize potential temporal influencing factors related to blood glucose fluctuation changes within the most recent preset time period.

[0029] In practice, the signal quality of the collected blood glucose change data is first assessed: for each sampling point, the percentage of valid sampling points within its local time period is calculated to obtain the sampling integrity of that sampling point. The local time period can be determined by a small window consisting of two sampling points before and after that sampling point. The criteria for a valid sampling point are that the glucose concentration value is within the normal detection range of the CGM sensor and there is no significant jump. Wavelet transform is used to estimate the local noise level of the sampling point, where the db4 wavelet basis function is selected and the decomposition level is set to 3. The impedance value corresponding to each sampling moment is obtained based on the output of the built-in impedance detection circuit of the CGM sensor. The impedance value is converted into a sensor state influence factor in the range of 0-1 according to the conventional normalization method. The higher the impedance value, the larger the influence factor, which represents a worse electrode state. The sampling integrity, local noise level, and sensor state influence factor corresponding to each sampling point are weighted. The weights of the three can be determined according to the existing blood glucose monitoring technical specifications or the empirical data accumulated from past experiments. After weighting, the signal quality weight corresponding to each sampling point is generated. The weight value ranges from 0 to 1. The closer the weight is to 1, the more reliable the data of the corresponding sampling point. Secondly, sampling points within the most recent preset time period are extracted from the blood glucose change data to generate a blood glucose subsequence. The most recent preset time period can be pre-configured in combination with the predicted time period and the device sampling cycle, and can be selected as 30 minutes, 60 minutes or 120 minutes. Temporal fluctuation indicators are extracted from the subsequence, including at least one of the following: blood glucose mean, rate of change, acceleration of change, peak value, trough value, fluctuation amplitude, duration of blood glucose rise, duration of blood glucose fall, and duration of stable maintenance. After normalizing the various indicators and converting them into dimensionless values ​​in the 0-1 range, all indicators are combined to form the blood glucose morphology characteristics.

[0030] In addition, in specific implementation, state-related data aligned with the time of the blood glucose subsequence is collected. The data sources are wearable device-collected data and terminal input interaction data. The wearable device-collected data includes heart rate, steps, activity intensity, sleep status, and body surface temperature. The terminal input interaction data includes food records, exercise records, and rest records. The state-related data is numerically processed. Continuous data is normalized to convert it into dimensionless values ​​in the 0-1 interval, and discrete data is processed using one-hot encoding to obtain state coding features with uniform dimensions. The current time period information is numerically processed. The current time period information includes meal periods, exercise periods, and nighttime rest periods, etc. The time period information is converted into dimensionless values ​​in the 0-1 interval using time encoding to obtain time coding features. Based on the blood glucose morphology features, the state coding features, and the time coding features, interaction features are generated through element-wise multiplication. Then, the state coding features, the time coding features, and the interaction features are concatenated in a preset order to obtain implicit event features. Finally, the signal quality weights corresponding to each sampling point in the blood glucose subsequence are extracted, and the signal quality weights of each sampling point are averaged to obtain a single signal quality weight value corresponding to the blood glucose subsequence. The value ranges from 0 to 1, and the closer the value is to 1, the higher the overall data reliability of the blood glucose subsequence. Blood glucose morphology features, latent event features, and the single signal quality weight value are spliced ​​together according to a preset arrangement order to construct temporal latent features. These temporal latent features correspond to a preset temporal interval of the blood glucose subsequence and are used to characterize the potential driving signals of blood glucose fluctuations within this interval, such as temporal fluctuation driving signals such as eating disturbances, exercise consumption, and nighttime stability.

[0031] It should be noted that the weights generated through signal quality assessment can reduce the interference of distorted sampling data on subsequent phase recognition. In actual blood glucose monitoring scenarios, the behavioral event records entered by the terminal, such as eating, exercise, and rest records, may be missing or incomplete. In this case, the implicit event features can rely on the patterns of blood glucose morphology features, state-related data, and current time period information to deduce the potential disturbance factors that induce blood glucose fluctuations, effectively reducing the recognition error caused by missing or incomplete behavioral event records.

[0032] In some embodiments, blood glucose fluctuation phase identification based on the blood glucose change data and the temporal latent features to obtain the fluctuation phase state can be specifically performed in the following manner: The blood glucose change data and the temporal latent features are input into the fluctuation phase recognition model to obtain several fluctuation phase labels and corresponding phase confidence levels. When several wave phase tags meet the preset composite phase determination conditions, the corresponding phase weights are assigned according to the phase reliability of each wave phase tag. By combining all the fluctuation phase labels and their corresponding phase confidence and phase weights, the fluctuation phase state is obtained.

[0033] It should be noted that the fluctuation phase label is used to identify the type of blood glucose fluctuation phase to which the current blood glucose change data belongs; the phase confidence is used to characterize the confidence level of each fluctuation phase label, which is determined by the output probability of the fluctuation phase recognition model; the composite phase determination condition is used to determine whether the current blood glucose change data is in a composite state of multiple fluctuation phases superimposed; the fluctuation phase state is structured data containing phase label, phase confidence and phase weight, used to characterize the fluctuation phase attribute of blood glucose changes within the most recent preset time period.

[0034] In specific implementation, firstly, blood glucose change data and temporal latent features are fed into the fluctuation phase recognition model as input. The input data is a single feature vector corresponding to a blood glucose subsequence, which is formed by concatenating blood glucose change data and temporal latent features. The input dimension can be pre-configured based on preset duration, sampling period, and number of features. The model is built on a temporal Transformer network, which includes a multi-head self-attention layer, a feedforward neural network layer, and a Softmax classification output layer. The functions of each layer are as follows: The multi-head self-attention layer has 4 attention heads to capture long-distance correlation features between time-series sampling points in the blood glucose change data. Simultaneously, it combines temporal latent features to mine the correlation between blood glucose fluctuations and potential influencing factors within the most recent preset duration, achieving fusion analysis of window-type features and time-series sampling point features, and outputting a fused feature vector with the same feature dimension as the input dimension. The feedforward neural network layer has 128 hidden nodes and uses the ReLU activation function. The model employs a multi-head self-attention layer to perform nonlinear mapping on the fused features output by the multi-head self-attention layer, further extracting high-dimensional features and improving the model's ability to fit complex blood glucose fluctuation patterns, outputting a high-dimensional feature vector. The Softmax classification output layer transforms the high-dimensional features output by the feedforward neural network layer into probability distributions for each preset fluctuation phase, completing the classification and determination of blood glucose fluctuation phases. The output dimension is 1×5, corresponding to the number of preset fluctuation phases. Among them, the preset fluctuation phases include the postprandial rapid rise interval, the high-value stable interval, the insulin-induced decrease interval, the post-exercise decrease interval, and the baseline stable interval. A corresponding fluctuation phase label is configured for each type of preset fluctuation phase to mark the fluctuation phase type to which the blood glucose subsequence belongs. The model outputs the probability values ​​corresponding to the above five types of fluctuation phases. These probability values ​​are the phase confidence of the corresponding fluctuation phase label. Each probability value corresponds one-to-one with a type of fluctuation phase label. The higher the probability value, the higher the confidence that the time interval to which the blood glucose subsequence belongs is the corresponding fluctuation phase.

[0035] It should be noted that the fluctuation phase recognition model needs to be pre-trained. The core training parameters and requirements are as follows: The training dataset uses historical samples containing blood glucose change data, temporal latent features, and corresponding real fluctuation phase labels, with a sample size of no less than 5000 groups. The training objective is to minimize the cross-entropy loss between the model's output probability distribution and the real label. The core parameters are set as follows: 100 iterations, batch size of 32, learning rate of 0.001, cross-entropy loss function, and Adam optimizer. After the model is trained, it is validated on a validation set to ensure that the classification accuracy is no less than 85%. Only after successful validation can the model be put into use.

[0036] In addition, in specific implementation, labels whose output probability is higher than the preset confidence threshold are selected as candidate phase labels, and the corresponding probability is the phase confidence. When there is only a single candidate phase label, the phase weight of that label is assigned a value of 1. When there are two or more candidate phase labels, and the difference in phase confidence between the labels is less than the preset confidence difference threshold, it is determined to be a composite fluctuation state. The ratio of the phase confidence of each candidate phase label to the sum of the phase confidence of all candidate phase labels is used as the phase weight of that candidate phase label. The sum of the weights of all labels in the composite state is equal to 1. Labels not included in the composite determination are uniformly assigned a phase weight of 0. The preset confidence threshold and the preset confidence difference threshold are both determined by batch standard sample calibration. It should be noted that the batch standard sample is blood glucose change data and corresponding labeled fluctuation phase labels covering various scenarios such as dining, exercise, and nighttime rest. The calibration method is to statistically analyze the accuracy of phase determination in various scenarios and select the optimal threshold as the preset confidence threshold and confidence difference threshold. For example, the preset confidence threshold is set to 0.7 and the preset confidence difference threshold is set to 0.1. Finally, the five types of fluctuation phase labels output by the model, the phase confidence and phase weights corresponding to each label are structurally integrated according to the fluctuation phase label category. The integrated form is a single multi-dimensional vector corresponding to the blood glucose subsequence, with a dimension of 1×15, thereby generating the fluctuation phase state corresponding to the time interval to which the blood glucose subsequence belongs.

[0037] It should be noted that by using a time-series classification model for fluctuation phase identification, the current fluctuation stage can be automatically determined based on blood glucose change data and time-series latent features, avoiding the identification bias caused by determining the fluctuation stage solely based on instantaneous concentration values ​​or a single rate of change. The composite phase determination condition can handle the ambiguous state where blood glucose changes are in the transition zone between two stages. The weight ratio is positively correlated with the determination confidence of each label, which can quantify the proportion of each fluctuation stage in the mixed state, effectively improving the accuracy of phase identification.

[0038] In step S103, based on the fluctuation phase state, the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference is determined, and the blood glucose change data is phase-calibrated according to the phase offset to obtain blood glucose change characteristics.

[0039] Preferably, in some embodiments, reference is made to Figure 3 As shown in the figure, this is an exemplary flowchart illustrating the determination of phase offset according to some embodiments of this application. In this embodiment, the determination of the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference based on the fluctuation phase state can be achieved by the following steps: In step S1031, historical response templates corresponding to each fluctuation phase label in the fluctuation phase state are obtained from historical blood glucose monitoring data. In step S1032, multiple candidate offsets are preset. For each candidate offset, the blood glucose change data and the historical response template corresponding to each fluctuation phase label are time-aligned to determine the phase matching degree of the candidate offset under the corresponding fluctuation phase label. In step S1033, for the same candidate offset, all phase matching degrees are fused to obtain the fused phase matching degree corresponding to the candidate offset; In step S1034, from multiple candidate offsets, the candidate offset whose fusion phase matching degree meets the preset matching condition is selected as the initial phase offset. In step S1035, the initial phase offset is subjected to a continuity constraint to obtain the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference.

[0040] It should be noted that the historical response template is used to characterize the typical fluctuation pattern of historical blood glucose change data under the corresponding fluctuation phase label; the candidate offset refers to multiple candidate phase offset values ​​set within a preset offset range according to a preset step size, wherein the phase matching degree is used to characterize the degree of matching between the blood glucose change data and the corresponding historical response template under a specific candidate offset.

[0041] In practice, firstly, from the historical blood glucose monitoring database, for each fluctuation phase label, multiple historical blood glucose change data segments corresponding to that label are extracted according to a preset filtering rule. The filtering rule is: select historical blood glucose change data segments from the historical blood glucose monitoring database that carry the fluctuation phase label after phase determination and whose label phase confidence is not less than 0.7. Then, perform time normalization, amplitude normalization, and statistical aggregation on all extracted historical data segments according to the time axis to generate historical response templates corresponding to the fluctuation phase labels. Among them, time normalization uses linear interpolation to unify each data segment to the same sampling frequency as the current blood glucose change data; amplitude normalization uses maximum-minimum value normalization to convert the data into dimensionless values ​​in the 0-1 range; and statistical aggregation uses mean calculation to integrate the fluctuation characteristics of each data segment, ensuring that each fluctuation phase label corresponds to a set of historical response templates, which is consistent with the logic of all labels corresponding to the current blood glucose change data. Secondly, within the preset offset search range, multiple candidate offsets are generated according to a preset step size. The offset search range can be set in conjunction with the blood glucose monitoring scenario. For example, it can be based on the starting sampling time of the currently collected blood glucose subsequence, ranging from 30 minutes in advance to 30 minutes in retrospect, with the preset step size set to 1 minute. For each candidate offset, the blood glucose change data is shifted along the time axis in either the positive or negative direction by the offset. Then, the time axis of the blood glucose change data and the historical response template is flexibly aligned using a dynamic time warping algorithm. The absolute difference of blood glucose concentration values ​​at each aligned time point is calculated. All absolute differences are summed, averaged, and normalized to the 0-1 interval to obtain the morphological similarity. This morphological similarity is used as the phase matching degree of the candidate offset under the corresponding fluctuation phase label.

[0042] In addition, in specific implementation, for the same candidate offset, its phase matching degree under each fluctuation phase label is weighted and fused according to the phase weight corresponding to each label to obtain the fused phase matching degree of the candidate offset. Next, from all candidate offsets, the candidate offset with the highest fused phase matching degree is selected as the initial phase offset. If there are multiple candidate offsets with the same highest fused phase matching degree, the candidate offset with the smallest absolute value is selected as the initial phase offset; where the absolute value of the candidate offset refers to the absolute value of the time offset value corresponding to the candidate offset. Finally, the initial phase offset of the current time series interval is compared with the phase offset of the previous time series interval, where the time series interval is consistent with the most recent preset duration of the blood glucose subsequence mentioned above; if the difference between the two exceeds the preset maximum rate of change threshold, the current initial phase offset is attenuated and corrected to move closer to the phase offset of the previous time series interval, resulting in the phase offset after continuity constraint. The preset maximum rate of change threshold can be set to 5 minutes / time series interval in combination with blood glucose fluctuation characteristics, and the attenuation correction adopts a linear attenuation method with a correction magnitude of 50% of the difference.

[0043] It should be noted that by setting candidate offsets based on the initial sampling time of the blood glucose subsequence and combining them with a dynamic time warping algorithm to compare blood glucose concentration values ​​at each time point to calculate the phase matching degree, the reference benchmark and comparison object of the time offset are clearly defined. Furthermore, the problem of inconsistent time series lengths is solved through flexible alignment, which can effectively quantify the morphological differences between the current blood glucose data and historical templates. Weighted fusion of the phase matching degree according to the phase weight of each label can reflect the different contributions of different fluctuation stages in the current blood glucose subsequence, avoiding the one-sidedness of single-label matching and improving the accuracy of the fused phase matching degree in representing the overall fluctuation pattern of blood glucose. By comparing the phase offsets of adjacent time series intervals and applying continuity constraints, the abrupt changes in offsets caused by blood glucose data fluctuations and model output jitter are effectively suppressed, ensuring the coherence of phase offsets between adjacent time series intervals and improving the temporal consistency of phase offset estimation.

[0044] In some embodiments, the blood glucose change data is phase-calibrated based on the phase offset to obtain blood glucose change characteristics, which can be specifically achieved in the following manner: By combining the phase offset with the blood glucose change data, a phase calibration sequence with a unified time reference is generated; For each fluctuation phase label in the fluctuation phase state, a calibration response feature matching the fluctuation phase label is extracted from the phase calibration sequence, and a corresponding calibration reliability sub-feature is generated based on the calibration response feature; The calibration response features corresponding to each fluctuation phase label are integrated to obtain the overall composite phase response features. The calibration reliability sub-features corresponding to each fluctuation phase label are integrated to obtain the overall calibration reliability features. The composite phase response feature, the calibration reliability feature, and the phase offset are combined to obtain the blood glucose change feature.

[0045] It should be noted that the phase calibration sequence refers to the sequence obtained by shifting the blood glucose change data along the time axis according to the phase offset, with the shift direction consistent with the sign of the phase offset, and the absolute value of the shift amount equal to that of the phase offset; the composite phase response feature is used to characterize the overall fluctuation response pattern of the current blood glucose subsequence after phase calibration; the calibration reliability feature is used to characterize the overall reliability of the composite phase response feature.

[0046] In practice, firstly, the phase offset is used as the time axis translation amount to perform an overall translation of the blood glucose change data, so that the temporal position of the blood glucose change data is aligned with the standard time series fluctuation benchmark, generating a phase calibration sequence with a unified time benchmark. During the translation process, linear interpolation is used to supplement missing sampling points to ensure the continuity of data time series. Secondly, for each fluctuation phase label, combined with the time characteristics of the blood glucose fluctuation stage corresponding to each fluctuation phase label, a data segment within the time interval of the fluctuation stage corresponding to the label is extracted from the phase calibration sequence. The extraction rule is as follows: the time interval corresponding to each fluctuation phase label is a specific sub-interval within the time range of the current blood glucose subsequence that corresponds to the fluctuation stage represented by the label. The time interval corresponding to the label can be determined according to the typical fluctuation stage duration and start and end time characteristics of the label in the historical response template, and the time intervals corresponding to all labels cover the complete time range of the current blood glucose subsequence. The time-series fluctuation index of the data segment is extracted to generate calibration response features, wherein the time-series fluctuation index has the same dimension as the blood glucose morphology features mentioned above. After normalizing various indicators and converting them into dimensionless values ​​in the 0-1 interval, all indicators are combined to form the calibration response features of the fluctuation phase label. Simultaneously, based on the signal quality weights of each sampling point within the time interval corresponding to the label, the average signal quality of the data segment is calculated as the label signal quality weight corresponding to the fluctuating phase label. The mean square error is used to determine the calibration residual between the calibration response feature corresponding to the label and the historical response template corresponding to the label. Specifically, each time-series fluctuation index in the calibration response feature is compared with the same type of time-series fluctuation index in the historical response template corresponding to the label, and the square of the difference between each pair of indices is calculated. The sum of the squared differences corresponding to all indices is divided by the total number of indices, and the result is the calibration residual corresponding to the label. Then, the reciprocal of the calibration residual is multiplied by the phase confidence and label signal quality weight corresponding to the label, and the multiplication result is normalized to the 0-1 interval to generate the calibration confidence sub-feature corresponding to the fluctuating phase label.

[0047] It should be noted that the calibration residual reflects the degree of matching between the calibration response characteristics and the historical template, the phase confidence reflects the confidence of the tag determination, and the tag signal quality weight reflects the reliability of the data itself. The combination of the three can comprehensively characterize the reliability of the calibration results.

[0048] In addition, in specific implementation, according to the phase weight corresponding to each fluctuation phase label, the calibration response features are weighted and fused to obtain a composite phase response feature. Specifically, for each fluctuation phase label, each time-series fluctuation index in its corresponding calibration response feature is multiplied by the phase weight of that label. Then, the multiplication results of the same type of time-series fluctuation index under all fluctuation phase labels are accumulated to obtain the final value of that type of time-series fluctuation index in the composite phase response feature. The final values ​​of all types of time-series fluctuation indices are integrated in a fixed order to form a composite phase response feature, and the integration order is consistent with the order of the time-series fluctuation indices in the calibration response feature. According to the same weighted fusion rule, each calibration reliable sub-feature is fused to obtain a calibration reliable feature, that is, each calibration reliable sub-feature is multiplied by the phase weight of the corresponding label, and the multiplication results are accumulated and normalized to the 0-1 interval to obtain the calibration reliable feature. Finally, the composite phase response feature, the calibration reliable feature, and the phase offset are concatenated and combined in a fixed order to form a single multidimensional feature vector, and this feature vector is used as the blood glucose change feature of the current time series interval.

[0049] It should be noted that by shifting the blood glucose change data along the time axis according to the phase offset, the temporal deviation of the blood glucose change data relative to the actual physiological driving state is eliminated, achieving alignment with the standard time-series fluctuation benchmark. The paired extraction of calibration response features and calibration confidence sub-features, combined with data quality, label confidence, and matching residuals, comprehensively evaluates the calibration reliability, ensuring that the contribution of low-quality data fragments in subsequent predictions is automatically reduced, thus improving the robustness of predictions. The splicing and combination of multi-dimensional features integrates the calibrated fluctuation pattern, reliability assessment, and offset information, improving the input feature dimensions of subsequent models.

[0050] In step S104, based on the blood glucose change characteristics, a blood glucose change prediction result within a preset time range is generated and displayed on the monitoring terminal.

[0051] In some embodiments, based on the blood glucose change characteristics, generating a predicted blood glucose change result within a preset time range and displaying it on a monitoring terminal can be achieved in the following ways: The blood glucose change characteristics are input into the blood glucose change prediction model, which outputs the predicted blood glucose values ​​and trend directions for multiple prediction times within a preset future time range. Based on the blood glucose change characteristics and the predicted blood glucose values ​​and trend directions at multiple prediction times, blood glucose change prediction results are generated and displayed through a monitoring terminal.

[0052] It should be noted that the blood glucose change prediction model is a prediction model built based on a temporal neural network, used to predict blood glucose values ​​at multiple future prediction times based on the blood glucose change characteristics of the current time interval; the trend direction is used to characterize the change trend of blood glucose values ​​at each prediction time relative to the previous prediction time, including an upward trend, a downward trend, and a flat trend. An upward trend indicates that the blood glucose value at the current prediction time is higher than the blood glucose value at the previous prediction time, a downward trend indicates that the blood glucose value at the current prediction time is lower than the blood glucose value at the previous prediction time, and a flat trend indicates that the absolute value of the difference between the blood glucose value at the current prediction time and the blood glucose value at the previous prediction time does not exceed 0.1 mmol / L; the blood glucose change prediction result includes the predicted blood glucose values ​​at multiple prediction times within a preset future time range, the trend direction corresponding to each prediction time, and the reliability indication information of the prediction result.

[0053] In practice, firstly, the blood glucose change features are input into a pre-trained blood glucose change prediction model. This model is constructed using a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU). The core parameters and requirements for model training are as follows: The training dataset uses historical blood glucose monitoring data with the same dimension as the current blood glucose change features. The samples must include the blood glucose change features and the actual blood glucose values ​​at each predicted time in the future. The training objective is to minimize the mean squared error between the model's output predicted value and the actual blood glucose value. The core parameters are set as follows: the number of hidden layer nodes in the LTM network or GRU is 64, the number of training iterations is 100, the batch size is 32, the learning rate is 0.001, the activation function is the ReLU function, the loss function is the mean squared error loss function, and the optimizer is the Adam optimizer. Based on the input blood glucose change features, the model predicts the blood glucose values ​​at multiple predicted times within a preset time range in the future, such as 30 minutes, 60 minutes, and 120 minutes in the future, and outputs the predicted blood glucose values ​​at each predicted time. At the same time, the model outputs the trend direction corresponding to each predicted time based on the change in the predicted blood glucose value at each predicted time relative to the previous time. Furthermore, the calibration confidence feature value in the blood glucose change characteristics is directly used as the overall confidence of the prediction result. Then, according to the prediction time sequence, the predicted blood glucose value, trend direction, and overall confidence are integrated to generate the blood glucose change prediction result. Finally, the blood glucose change prediction result is sent to the monitoring terminal according to a preset communication protocol, which can be Bluetooth, Wi-Fi, or mobile network communication protocol. After receiving the data, the monitoring terminal generates the blood glucose change prediction result through its own computing module. The specific display format is as follows: the change of the predicted blood glucose value within the future preset time range is presented as a trend curve graph, and the trend direction of each prediction time is marked with arrows, where upward is rising, downward is falling, and horizontal is flat. The prediction confidence is displayed with text and icons, where confidence > 0.8 is represented by a green icon, 0.5 ≤ confidence ≤ 0.8 by a yellow icon, and confidence < 0.5 by a red icon, which is easy to view intuitively.

[0054] It should be noted that by inputting the calibrated blood glucose change features into the prediction model, the model can make predictions based on time-aligned data, effectively avoiding prediction bias caused by temporal shifts between the original monitoring sequence and the actual physiological state, thus improving prediction accuracy. The model uses a long short-term memory network or gated recurrent unit to capture the temporal dependencies of blood glucose data and adapt to the dynamic changes in blood glucose fluctuations. The reliability indicator information can intuitively reflect the reference value of the prediction results, making it easier for users to make reasonable blood glucose control decisions based on their own circumstances.

[0055] Therefore, this application demonstrates that, firstly, the quality weights generated through signal quality assessment can reduce the interference of distorted sampling data on phase recognition; when behavioral event records are missing, latent event features, relying on blood glucose morphology features, state-related data, and time period information to infer potential perturbation factors, can reduce recognition errors caused by missing data; and using a time-series classification model for fluctuation phase recognition avoids biases caused by relying solely on instantaneous concentration values ​​or single rates of change. Secondly, setting candidate offsets based on the initial sampling time of the blood glucose subsequence, combined with a dynamic time warping algorithm to calculate the phase matching degree, can quantify the morphological differences between the current blood glucose data and historical templates; weighted fusion of the matching degree according to the phase weights of each label reflects the differences in contribution of different fluctuation stages, avoiding the one-sidedness of single label matching; and the pairing extraction of calibration response features and calibration confidence sub-features, combined with data quality, label confidence, and matching residuals to comprehensively evaluate calibration reliability, automatically reducing the contribution of low-quality data fragments and improving prediction robustness. Finally, the calibrated blood glucose change features are input into the prediction model, and predictions are made based on time-aligned data to avoid prediction bias caused by time-series offsets and improve prediction accuracy. The model uses a long short-term memory network or a gated recurrent unit to capture the temporal dependencies of blood glucose data and adapt to the dynamic changes in blood glucose fluctuations.

[0056] In summary, the technical solution adopted in this application can identify the phase of blood glucose fluctuations and perform phase calibration to improve the accuracy of blood glucose change prediction results.

[0057] Furthermore, in another aspect of this application, in some embodiments, this application provides an artificial intelligence-based blood glucose change prediction system, with reference to... Figure 4 The figure is a schematic diagram of the structure of an artificial intelligence-based blood glucose variability prediction system according to some embodiments of this application. The artificial intelligence-based blood glucose variability prediction system includes: Data acquisition module 201 is used to acquire blood glucose change data in real time through a wearable CGM sensor; Phase recognition module 202 is used to extract the temporal latent features that characterize the implicit driving signal of blood glucose temporal fluctuations based on the blood glucose change data, and to perform blood glucose fluctuation phase recognition based on the blood glucose change data and the temporal latent features to obtain the fluctuation phase state. The phase calibration module 203 is used to determine the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference based on the fluctuation phase state, and to perform phase calibration on the blood glucose change data according to the phase offset to obtain blood glucose change characteristics. The result prediction module 204 is used to generate a prediction result of blood glucose change within a preset time range based on the blood glucose change characteristics and display it on the monitoring terminal.

[0058] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described artificial intelligence-based blood glucose change prediction method.

[0059] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing an artificial intelligence-based blood glucose variability prediction method according to some embodiments of this application. The artificial intelligence-based blood glucose variability prediction method in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0060] The processor 301 may be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the artificial intelligence-based blood glucose change prediction method in this application.

[0061] The communication bus 302 can be used to transmit information between the aforementioned components.

[0062] The memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 303 may exist independently and be connected to the processor 301 via the communication bus 302. The memory 303 may also be integrated with the processor 301.

[0063] The memory 303 stores program code for executing the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the determination of the blood glucose change prediction method based on artificial intelligence can be implemented by the processor 301 and one or more software modules in the program code in the memory 303.

[0064] Communication interface 304 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0065] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0066] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0067] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described artificial intelligence-based method for predicting blood glucose changes.

[0068] Although preferred embodiments of this application 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 the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0069] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting blood glucose changes based on artificial intelligence, characterized in that, Includes the following steps: Real-time blood glucose change data is collected using a wearable CGM sensor; Based on the blood glucose change data, temporal latent features characterizing the implicit driving signal of blood glucose temporal fluctuations are extracted. Blood glucose fluctuation phase is identified according to the blood glucose change data and the temporal latent features to obtain the fluctuation phase state. Based on the fluctuation phase state, the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference is determined, and the blood glucose change data is phase-calibrated according to the phase offset to obtain blood glucose change characteristics; Based on the blood glucose change characteristics, a prediction result of blood glucose change within a preset time range is generated and displayed on the monitoring terminal.

2. The method as described in claim 1, characterized in that, Real-time blood glucose data collection via wearable CGM sensors specifically includes: Raw glucose concentration data are collected using a wearable CGM sensor at a preset sampling period. The monitoring terminal performs preprocessing on the raw glucose concentration data to generate blood glucose change data.

3. The method as described in claim 1, characterized in that, Based on the blood glucose change data, extracting the temporal latent features characterizing the implicit driving signal of blood glucose temporal fluctuations specifically includes: The blood glucose change data are subjected to signal quality assessment to obtain signal quality weights that characterize the sampling integrity, local noise level, and sensor status influence. The blood glucose sampling points within the most recent preset time period are extracted from the blood glucose change data to generate a blood glucose subsequence, and the time-series fluctuation index is extracted from the blood glucose subsequence to obtain blood glucose morphology characteristics. Obtain state association data corresponding to the time of the blood glucose subsequence, and generate hidden event features based on the blood glucose morphology features, the state association data, and the current time period information; The blood glucose morphology features, the latent event features, and the signal quality weights are concatenated to obtain the temporal latent features that characterize the latent driving signal of blood glucose temporal fluctuations.

4. The method as described in claim 1, characterized in that, Based on the blood glucose change data and the temporal latent features, blood glucose fluctuation phase identification is performed to obtain the fluctuation phase state, which specifically includes: The blood glucose change data and the temporal latent features are input into the fluctuation phase recognition model to obtain several fluctuation phase labels and corresponding phase confidence levels. When several wave phase tags meet the preset composite phase determination conditions, the corresponding phase weights are assigned according to the phase reliability of each wave phase tag. By combining all the fluctuation phase labels and their corresponding phase confidence and phase weights, the fluctuation phase state is obtained.

5. The method as described in claim 1, characterized in that, Based on the aforementioned fluctuation phase state, determining the phase offset of the blood glucose change data relative to the standard time-series fluctuation benchmark specifically includes: From historical blood glucose monitoring data, obtain the historical response templates corresponding to each fluctuation phase label in the fluctuation phase state; Multiple candidate offsets are pre-set. For each candidate offset, the blood glucose change data is time-aligned with the historical response template corresponding to each fluctuation phase label to determine the phase matching degree of the candidate offset under the corresponding fluctuation phase label. For the same candidate offset, all phase matching degrees are fused to obtain the fused phase matching degree corresponding to the candidate offset; From multiple candidate offsets, select the candidate offset whose fusion phase matching degree meets the preset matching condition as the initial phase offset; By applying a continuity constraint to the initial phase offset, the phase offset of the blood glucose change data relative to the standard time-series fluctuation benchmark is obtained.

6. The method as described in claim 1, characterized in that, Phase calibration of the blood glucose change data based on the phase offset yields blood glucose change characteristics, specifically including: By combining the phase offset with the blood glucose change data, a phase calibration sequence with a unified time reference is generated; For each fluctuation phase label in the fluctuation phase state, a calibration response feature matching the fluctuation phase label is extracted from the phase calibration sequence, and a corresponding calibration reliability sub-feature is generated based on the calibration response feature; The calibration response features corresponding to each fluctuation phase label are integrated to obtain the overall composite phase response features. The calibration reliability sub-features corresponding to each fluctuation phase label are integrated to obtain the overall calibration reliability features. The composite phase response feature, the calibration reliability feature, and the phase offset are combined to obtain the blood glucose change feature.

7. The method as described in claim 1, characterized in that, Based on the aforementioned blood glucose change characteristics, the generation of predicted blood glucose changes within a preset time range and the display of these predictions on the monitoring terminal specifically includes: The blood glucose change characteristics are input into the blood glucose change prediction model, which outputs the predicted blood glucose values ​​and trend directions for multiple prediction times within a preset future time range. Based on the blood glucose change characteristics and the predicted blood glucose values ​​and trend directions at multiple prediction times, blood glucose change prediction results are generated and displayed through a monitoring terminal.

8. An artificial intelligence-based blood glucose variability prediction system, used to execute the artificial intelligence-based blood glucose variability prediction method as described in any one of claims 1 to 7, characterized in that, The system includes: The data acquisition module is used to collect blood glucose change data in real time through a wearable CGM sensor; The phase recognition module is used to extract the temporal latent features that characterize the implicit driving signal of blood glucose temporal fluctuations based on the blood glucose change data, and to perform blood glucose fluctuation phase recognition based on the blood glucose change data and the temporal latent features to obtain the fluctuation phase state. The phase calibration module is used to determine the phase offset of the blood glucose change data relative to the standard time-series fluctuation reference based on the fluctuation phase state, and to perform phase calibration on the blood glucose change data according to the phase offset to obtain blood glucose change characteristics. The result prediction module is used to generate a prediction result of blood glucose changes within a preset time range based on the blood glucose change characteristics and display it on the monitoring terminal.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the artificial intelligence-based blood glucose change prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the artificial intelligence-based blood glucose change prediction method as described in any one of claims 1 to 7.

Images