Diabetes and pre-diabetes self-management system based on ai and cgm

Abstract

The application relates to the technical field of healthcare informatization, in particular to a diabetes and pre-diabetes patient self-management system based on AI and CGM, which comprises the following steps: delaying and aligning CGM blood glucose flow and multi-modal physiological signals through a signal acquisition and synchronization module; dynamically adjusting a trigger threshold for different disease courses to identify suspected events by using an intelligent event trigger module; combining medical prior rules and an integrated classification model to perform joint discrimination by a hybrid consistency verification engine, and outputting a consistency probability; and cooperating with a hierarchical decision mechanism of an interaction module and online incremental learning of an adaptive learning module to optimize individual parameters. Therefore, the diabetes and pre-diabetes patient self-management system solves frequent false alarms caused by the inherent physiological delay, signal drift and other physical limitations of the CGM sensor, and the resulting patient alarm fatigue and decreased compliance, and provides more stable and personalized dynamic blood glucose management support for active people such as pre-diabetes patients.

Description

Technical Field

[0001] This application relates to the field of healthcare information technology, and in particular to a self-management system for patients with diabetes and prediabetes based on AI and CGM. Background Technology

[0002] Continuous glucose monitoring (CGM) technology offers the possibility of refined management for patients with diabetes and prediabetes by measuring interstitial fluid glucose concentration in real time. However, its data stream inherently suffers from insurmountable physical and physiological biases, constituting a major obstacle to current precision management. The core issue is that interstitial fluid glucose concentration lags behind blood glucose by approximately 5 to 15 minutes. This physiological delay is particularly pronounced during periods of rapid blood glucose changes (such as after meals or strenuous exercise), causing a significant deviation between CGM readings and the actual physiological state. Even more challenging is sensor drift. The sensitivity and baseline of electrochemical sensors undergo unpredictable and gradual changes over time, temperature, and local tissue reactions, generating systematic errors. These biases directly lead to two risks in clinical settings: false alarms, misinterpreting sensor noise or drift as real blood glucose events; and missed alarms, failing to capture real, dangerous fluctuations due to delays or decreased sensitivity. For patients, frequent false alarms, especially unexplained hypoglycemic alerts at night, severely damage their trust in the technology, leading to "alarm fatigue" and even a collapse in adherence.

[0003] To address these issues, relevant technologies primarily follow two paths, both of which have significant limitations. The first path involves optimizing the sensor hardware itself, aiming to improve accuracy and stability through materials science and electrochemical improvements. However, this path suffers from long iteration cycles, high costs, and cannot fundamentally eliminate the inherent physiological delay. The second path relies on algorithms to post-process CGM data. Common methods include signal smoothing based on models such as Kalman filtering, or using a single auxiliary signal (such as heart rate or motion data) for coarse anomaly labeling; for example, labeling a hyperglycemic event when the heart rate accelerates. However, these approaches generally fall into the trap of one-way verification.

[0004] Specifically, related technologies still primarily rely on a single blood glucose data stream, lacking simultaneous verification from multiple physiological states. A real hypoglycemic stress is the combined result of the synergistic effects of sympathetic nerve excitation (changes in heart rate and HRV), endocrine responses (cortisol, etc.), and peripheral vasoconstriction (decreased skin temperature). Any algorithm that ignores this physiological coupling is inherently fragile and susceptible to interference. Most related technologies employ fixed population thresholds or general models, failing to consider the significant physiological differences between individuals. Furthermore, the autonomic nervous system response patterns to blood glucose fluctuations differ drastically between young people with prediabetes and elderly patients with chronic diabetes; in other words, these technologies suffer from model rigidity. Additionally, deep black-box models may provide a "suspicious" judgment but fail to offer users or doctors a reasonable explanation consistent with medical common sense (e.g., "During this hyperglycemic alert, your heart rate variability did not decrease as expected, which may be related to sensor signals"). This is a significant deficiency in medical applications that emphasize safety and trust.

[0005] The prediabetic population represents a golden window for health intervention. Compared to diagnosed patients, their blood glucose fluctuation patterns are more complex and active, often deeply coupled with diet, high-intensity interval training, and acute stress events. A brief, transient hyperglycemia after exercise has entirely different clinical significance and management strategies than a sustained rise in postprandial hyperglycemia. Current CGM (Continuous Glucose Monitoring) systems lack the ability to finely distinguish these dynamic patterns, often issuing unnecessary alerts for the former, leading to health anxiety and potentially hindering patients from adhering to active but potentially volatile lifestyle habits (such as exercise), which contradicts the original intention of health management.

[0006] In summary, clinical practice and technological development both point to an unmet need: the urgent need for an intelligent system that can dynamically assess the credibility of CGM data itself. This system should not be another black box alarm, but rather an intelligent system that cross-validates based on multi-dimensional physiological evidence, capable of identifying and filtering noise caused by technological limitations, accurately capturing real physiological events, and providing a clinically understandable chain of evidence for its own judgment. Summary of the Invention

[0007] This application provides a self-management system for diabetic and prediabetic patients based on AI (Artificial Intelligence) and CGM (Continuous Glucose Meter) to address the frequent false alarms caused by the inherent physiological delays and signal drift of CGM sensors, as well as the resulting patient alarm fatigue and decreased compliance. This application integrates multimodal physiological signals such as heart rate variability (HRV) and skin temperature to perform real-time, interpretable cross-validation of CGM data, realizing an intelligent consistency verification system that can proactively identify and filter false alarms while accurately capturing real blood glucose events. This reduces unnecessary alarms, improves system reliability, and provides more stable and personalized dynamic blood glucose management support for active populations such as those with prediabetes.

[0008] The first aspect of this application provides a self-management system for patients with diabetes and prediabetes based on AI and CGM, including: a signal acquisition and synchronization module, an intelligent event triggering module, a feature extraction module, a hybrid consistency verification engine, a credibility determination and interaction module, and an adaptive learning module; wherein...

[0009] The signal acquisition and synchronization module is used to synchronously receive blood glucose data streams from continuous glucose monitoring (CGM) devices and multimodal physiological signals from wearable devices, and to perform time axis alignment processing on the blood glucose data streams and multimodal physiological signals according to preset physiological delay parameters.

[0010] The multimodal physiological signals include at least the photo-plethysmography (PPG) signal and the skin temperature signal used to calculate heart rate variability (HRV).

[0011] The intelligent event triggering module is connected to the signal acquisition and synchronization module and is used to monitor the blood glucose data stream in real time based on preset dynamic triggering conditions. When the current blood glucose data stream meets any triggering condition, the current blood glucose data stream is identified and marked as a suspected blood glucose event.

[0012] The feature extraction module is connected to the intelligent event triggering module and is used to extract a multimodal feature vector containing blood glucose change features, HRV index calculated based on PPG signal, heart rate change features and skin temperature change features within the time window of the suspected blood glucose event.

[0013] The hybrid consistency verification engine is connected to the feature extraction module and is used to receive the multimodal feature vector, and perform joint discrimination based on a preset medical rule model and a trained machine learning model, and output the consistency probability characterizing the suspected blood glucose event as a real physiological event.

[0014] The credibility determination and interaction module is connected to the hybrid consistency verification engine and is used to compare the consistency probability with a preset confidence threshold. If the consistency probability value is lower than the preset confidence threshold, the suspected blood glucose event is determined to be a suspicious event, and a non-mandatory secondary verification prompt is sent to the user terminal.

[0015] The adaptive learning module is used to calibrate the system using external verification data after the user provides feedback based on the secondary verification prompt, and to update the individualized parameters of the machine learning model through an online learning mechanism.

[0016] Optionally, in some embodiments, the preset dynamic triggering conditions in the intelligent event triggering module include:

[0017] Threshold triggering based on blood glucose value and slope triggering based on blood glucose rate of change;

[0018] The threshold trigger includes: the CGM blood glucose value is higher than a first hyperglycemia threshold or lower than a first hypoglycemia threshold; the slope trigger includes: the rate of blood glucose rise is higher than a first rate of rise threshold or the rate of blood glucose fall is lower than a first rate of fall threshold.

[0019] The system dynamically adjusts the first rise rate threshold and / or the first fall rate threshold based on the user's preset disease stage labels.

[0020] When the preset disease stage label is prediabetes, the absolute value of the first rate of increase threshold is lowered.

[0021] Optionally, in some embodiments, in the hybrid consistency verification engine, the preset medical rule model is set based on the coordinated physiological change pattern of the autonomic nervous system and peripheral vascular response under hyperglycemia or hypoglycemia events, and is used to perform prior logic verification on the multimodal feature vector.

[0022] The machine learning model is an ensemble classification model based on gradient boosting decision trees, used to calculate the consistency probability of the multimodal feature vectors.

[0023] Optionally, in some embodiments, the HRV index extracted by the feature extraction module includes SDNN (Standard Deviation of Normal-to-Normal Intervals) and RMSSD (Root Mean Square of Successive Differences), and the extracted features include the HRV index, heart rate, and skin temperature changes relative to the user's individual baseline within the time window.

[0024] Optionally, in some embodiments, the online learning mechanism of the adaptive learning module specifically includes:

[0025] The external validation data provided by users is compared with the CGM readings at the corresponding time points to generate training samples with authenticity labels.

[0026] An incremental learning approach is used to fine-tune the weight parameters of the machine learning model using the training samples.

[0027] The second aspect of this application provides a self-management method for patients with diabetes and prediabetes based on AI and CGM, including the following steps:

[0028] Simultaneously acquire real-time blood glucose data streams from CGM devices and multimodal physiological signals from wearable devices, including at least PPG signals and skin temperature signals, and perform time-shift alignment of the blood glucose data streams according to the physiological delay of CGM.

[0029] The blood glucose data stream is monitored in real time, and when the current blood glucose value or the rate of change of blood glucose value meets the preset dynamic triggering conditions, it is marked as a suspected blood glucose event.

[0030] Multidimensional physiological features were extracted within the time window corresponding to the suspected blood glucose events, and a consistent feature vector containing blood glucose features, HRV index, heart rate features, and skin temperature features was constructed.

[0031] The consistency feature vector is input into the hybrid consistency verification engine, which combines medical prior rules and machine learning models for joint analysis and outputs the consistency probability of the suspected blood glucose event.

[0032] The consistency probability is compared with a preset confidence threshold. If the consistency probability is lower than the preset confidence threshold, the suspected blood glucose event is determined to be a low-confidence suspicious event and a secondary verification process for the user is triggered. Otherwise, it is determined to be a high-confidence event and a standard management response is executed.

[0033] In response to feedback data provided by users during the secondary verification process, the CGM readings are calibrated, and the individualized parameters of the machine learning model are optimized through online learning using the feedback data.

[0034] Optionally, in some embodiments, the preset dynamic triggering condition includes any of the following:

[0035] The CGM blood glucose level is greater than 10.0 mmol / L, and the rate of increase of the CGM blood glucose level is greater than 0.5 mmol / L / min;

[0036] The CGM blood glucose level is less than 3.9 mmol / L, and the rate of decrease of the CGM blood glucose level is less than -0.4 mmol / L / min;

[0037] CGM blood glucose levels enter a personalized warning zone dynamically generated based on the user's historical data.

[0038] Optionally, in some embodiments, the medical prior rules include:

[0039] For suspected hyperglycemia events, verify whether the time window is accompanied by a decrease in heart rate variability (HRV) exceeding the first proportional threshold, an increase in heart rate, and an increase in skin temperature.

[0040] For suspected hypoglycemia events, verify whether the time window is accompanied by a decrease in HRV index exceeding the second proportional threshold, a significant increase in heart rate, and a decrease in skin temperature.

[0041] Optionally, in some embodiments, comparing the consistency probability with a preset confidence threshold, if the consistency probability value is lower than the preset confidence threshold, then the suspected blood glucose event is determined as a low-confidence suspicious event, and a secondary verification process for the user is triggered; otherwise, it is determined as a high-confidence event and a standard management response is executed, specifically including:

[0042] If the consistency probability is greater than the first confidence threshold, it is determined to be a credible event, and the system executes the corresponding early warning or intervention suggestion.

[0043] If the consistency probability is greater than or equal to the second confidence threshold and less than or equal to the first confidence threshold, it is determined to be an uncertain event, and the system records and continuously observes it;

[0044] If the consistency probability is less than the second confidence threshold, it is determined to be a suspicious event. The system sends a prompt message to the user terminal, suggesting confirmation by finger-prick blood measurement, and waits for user feedback within a preset time.

[0045] Optionally, in some embodiments, the online learning optimization specifically includes:

[0046] The algorithm takes the actual blood glucose level reported by users as the target and minimizes the difference between the credibility of the event predicted by the model and the actual reality of the event as the optimization objective. It uses a gradient boosting iterative algorithm to update the weights of the weak classifiers in the machine learning model. The optimization process of the online learning is triggered during the system's idle period or after accumulating a preset number of new samples.

[0047] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the AI ​​and CGM-based self-management method for diabetes and prediabetes patients as described in the above embodiments.

[0048] A fourth aspect of this application provides a computer program product having a computer program stored thereon, which is executed by a processor to implement the AI ​​and CGM-based self-management method for diabetes and prediabetes patients as described in the above embodiments.

[0049] The beneficial effects of the embodiments of this application are as follows:

[0050] (1) This application uses a hybrid consistency verification engine to treat the medical rule model as a “defense boundary”, which can effectively identify and filter more than 30% of physical noise caused by sensor drift, local pressure or environmental temperature difference, ensuring that blood glucose warning has high biological credibility.

[0051] (2) The adaptive reduction mechanism of the "first rate of rise threshold" provided by this application for people with prediabetes can detect the signs of transient blood glucose rise earlier than the traditional fixed threshold, providing users with more dynamic adjustment space (such as immediate exercise or diet modification).

[0052] (3) The hybrid consistency verification engine of this application logically couples physiological indicators such as heart rate, heart rate variability and skin temperature with blood glucose fluctuations, providing a chain of evidence that conforms to medical common sense for each warning, thereby improving the credibility of medical AI.

[0053] (4) The adaptive learning module of this application adopts a lightweight incremental learning algorithm and uses the fingertip blood data of scattered user feedback to fine-tune the model locally, realizing the evolution from a general "average person" model to a "personalized" model for specific users, making the system more and more accurate the more it is used.

[0054] (5) The credibility judgment and interaction module of this application establishes a “high credibility, uncertain, low credibility” hierarchical decision-making system, which greatly reduces users’ health anxiety and protects their long-term self-management motivation by guiding them through non-mandatory secondary verification rather than frequent hard alarms.

[0055] (6) The “low power monitoring + event-triggered wake-up” mode adopted in the method only performs high-frequency multimodal feature extraction and analysis after a suspected event is identified. This effectively reduces the computing power consumption of wearable devices and extends the device’s battery life while ensuring accuracy.

[0056] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0057] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0058] Figure 1 This is a schematic diagram of the structure of the AI- and CGM-based self-management system for patients with diabetes and prediabetes according to an embodiment of this application;

[0059] Figure 2 This is a flowchart of a self-management method for diabetes and prediabetes patients based on AI and CGM, provided according to an embodiment of this application.

[0060] Figure 3 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0061] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0062] The following description, with reference to the accompanying drawings, describes an AI- and CGM-based self-management system for patients with diabetes and prediabetes. Addressing the frequent false alarms caused by the inherent physiological delay and signal drift of CGM sensors, as mentioned in the background art, and the resulting patient alarm fatigue and decreased compliance, this application first provides an AI- and CGM-based self-management system for patients with diabetes and prediabetes. Specifically, Figure 1 This is a schematic diagram of the structure of the AI- and CGM-based self-management system for diabetes and prediabetes patients provided in the embodiments of this application, as shown below. Figure 1 As shown, the AI- and CGM-based self-management system for patients with diabetes and prediabetes includes: a signal acquisition and synchronization module 100, an intelligent event triggering module 200, a feature extraction module 300, a hybrid consistency verification engine 400, a credibility determination and interaction module 500, and an adaptive learning module 600.

[0063] The AI- and CGM-based self-management system 10 for patients with diabetes and prediabetes (hereinafter referred to as "System 10") in this application embodiment is an intelligent health management platform based on cross-validation of multimodal physiological signals. System 10 not only passively receives CGM data but also actively and synchronously collects physiological signals such as heart rate variability and skin temperature from users. When a suspected hyperglycemia / hypoglycemia event is detected, the hybrid consistency verification engine 400 of System 10 will initiate analysis: the rule layer performs rapid screening based on medical prior knowledge such as "true hypoglycemia should be accompanied by increased heart rate and decreased skin temperature"; the model layer uses machine learning algorithms to calculate the consistency probability of the event being a true physiological event by integrating all signals. Based on this probability, System 10 executes intelligent decision-making: it issues warnings for high-confidence events and prioritizes guiding users to conduct secondary confirmation via finger-prick blood for low-confidence events. Each confirmation feedback from the user is used to calibrate data and optimize their personal AI model to achieve personalized adaptation.

[0064] Compared to related technologies, the system 10 of this application embodiment improves the accuracy and reliability of blood glucose management through upper-level intelligent algorithms without changing the existing hardware. In particular, for prediabetic individuals who are active in daily life, blood glucose fluctuations caused by exercise can be effectively distinguished from actual pathological fluctuations through the exercise-related features (such as heart rate slope) of the system 10. The various components of the system 10 will be described in detail below.

[0065] like Figure 1 As shown, the signal acquisition and synchronization module 100 is used to synchronously receive blood glucose data streams from the continuous glucose monitoring (CGM) device and multimodal physiological signals from the wearable device, and to perform time axis alignment processing on the blood glucose data streams and multimodal physiological signals according to preset physiological delay parameters.

[0066] In this embodiment of the application, the blood glucose data stream refers to the sequence data that is continuously transmitted at a preset frequency (e.g., 5 minutes / time) after the interstitial fluid glucose concentration signal sensed in real time by a subcutaneous sensor is converted into a digital signal by a transmitter. This blood glucose data stream not only includes the current blood glucose value, but also the trend information (slope) of blood glucose changes over time, and is the basic data source for the system 10 to trigger abnormal events.

[0067] The multimodal physiological signals in this application embodiment include at least the photoplethysmography (PPG) signal (photoplethysmography, which captures changes in vascular volume through a photoelectric sensor to calculate heart rate variability (HRV)) and the skin temperature signal.

[0068] The multimodal physiological signal refers to a combination of signals related to glucose metabolism stress, including PPG signals that reflect autonomic nervous activity and skin temperature signals that reflect peripheral metabolism and thermoregulation. In this embodiment, a photoelectric sensor is used to monitor the periodic changes in vascular volume under skin tissue with each heartbeat to obtain PPG signals. Then, the heart rate velocity (HRV) is obtained by extracting the subtle differences between successive heartbeat cycles from the PPG signals. It should be noted that HRV is an important indicator for measuring the balance of the autonomic nervous system (sympathetic and parasympathetic nervous systems). Under high / low glucose stress, the autonomic nervous system experiences significant fluctuations (e.g., inhibition of the parasympathetic nervous system leads to a decrease in HRV), which is one of the core medical criteria for consistency verification in System 10.

[0069] The physiological delay parameter is due to the time required for blood glucose to diffuse from blood vessels (blood) to subcutaneous tissue (interstitial fluid), resulting in a time difference between the CGM measurement value and the actual blood glucose value. The preset physiological delay parameter in this application embodiment can be between 5 and 15 minutes, depending on the individual's metabolic level, the sensor wearing location, and the severity of blood glucose changes.

[0070] Since heart rate (HRV) and skin temperature changes are almost synchronous with blood glucose stress (driven by neural reflexes), and CGM data is delayed, the CGM blood glucose data stream needs to be shifted forward (or the physiological signals need to be shifted backward) by a corresponding amount of time on the time axis to ensure that it is within the "event trigger window" and that the AI ​​model is comparing physiological correlation features generated at the same time. Therefore, the time axis alignment processing of blood glucose data stream and multimodal physiological signals in this application aims to compensate for the asynchrony caused by physiological delay.

[0071] Specifically, the signal acquisition and synchronization module 100, serving as the sensing hub of system 10, has the core function of achieving sub-second synchronization of heterogeneous physiological signals and alignment with the time axis of physiological dimensions. This module connects synchronously to the CGM device and wearable device via a short-range wireless communication module. The module receives interstitial fluid glucose concentration data streams from the CGM device, for example, at a sampling frequency of 5 minutes per data point. Simultaneously, it acquires high-frequency PPG signals (sampling rate not less than 1 Hz) and skin surface temperature signals from the wearable device. Furthermore, to ensure transparency and data compliance in feature extraction, the module immediately performs standardized preprocessing upon receiving the signals, such as using a 0.5 Hz to 4 Hz bandpass filter to remove noise and extract high-quality pulse wave waveforms.

[0072] To address the physiological delay issue mentioned in the background (i.e., interstitial fluid glucose changes lag behind peripheral physiological stress), the signal acquisition and synchronization module 100 incorporates a time alignment algorithm. The physiological delay parameter in this embodiment can be set according to the characteristics of the CGM device and the user's physiological features; as a preferred implementation, this parameter can be set within a range of approximately 5 to 15 minutes and can be dynamically optimized based on the user's individual historical baseline data. Its operating logic is as follows: System 10 buffers the high-frequency acquired PPG signal (used to calculate HRV) and skin temperature signal into a circular buffer. When the CGM reading at the current time t is received, module 100 automatically invokes the alignment algorithm to retrieve the physiological features corresponding to time t-τ (where τ is the physiological delay parameter) from the buffer. Through this time-shift alignment, system 10 ensures that each dimension of the feature vector F subsequently entering the feature extraction module 300 has a strong causal correlation in terms of biological timestamps.

[0073] This application solves the AI ​​misjudgment problem caused by "data misalignment" through time alignment processing performed by the signal acquisition and synchronization module 100. For example, when the CGM shows a sudden drop in blood glucose at time t, the physiological characteristics extracted at time t-τ (such as a significant increase in heart rate and a significant decrease in HRV) can confirm that the fluctuation is a genuine hypoglycemic stress response, rather than a physical drift caused by the sensor at time t. The above alignment process provides the hybrid consistency verification engine 400 with multimodal inputs with logical consistency, reduces the false alarm rate, and improves the accuracy of management for prediabetic active individuals in exercise or eating scenarios.

[0074] The intelligent event triggering module 200 is connected to the signal acquisition and synchronization module 100 and is used to monitor the blood glucose data stream in real time based on preset dynamic triggering conditions. When the current blood glucose data stream meets any triggering condition, the current blood glucose data stream is identified and marked as a suspected blood glucose event.

[0075] In this embodiment, a suspected blood glucose event refers to an abnormal fluctuation in the blood glucose data stream identified by system 10 that has clinical intervention significance or potential risk. A suspected blood glucose event is not a confirmed diagnosis, but rather serves as a start signal for the subsequent hybrid consistency verification engine 400. The dynamic triggering conditions in this embodiment are a set of judgment criteria updated in real time based on multi-dimensional features, used to capture abnormal trends in the blood glucose data stream. These dynamic triggering conditions are not fixed, rigid values, but rather a parameter matrix dynamically adjusted according to the user's "preset disease stage label" and historical baseline, including threshold triggering based on blood glucose values ​​and slope triggering based on the rate of blood glucose change.

[0076] Optionally, in some embodiments, the preset dynamic triggering conditions in the intelligent event triggering module 200 include: threshold triggering based on blood glucose values ​​and slope triggering based on the rate of blood glucose change; threshold triggering includes: CGM blood glucose value being higher than a first hyperglycemia threshold or lower than a first hypoglycemia threshold, and slope triggering includes: blood glucose rise rate being higher than a first rise rate threshold or blood glucose fall rate being lower than a first fall rate threshold; wherein, the system 10 dynamically adjusts the first rise rate threshold and / or the first fall rate threshold according to the user's preset disease stage label; when the preset disease stage label is prediabetes, the absolute value of the first rise rate threshold is lowered.

[0077] In this embodiment, the threshold triggering is an early warning mechanism that is triggered when the CGM blood glucose value enters a preset high blood glucose warning zone (above the first high blood glucose threshold) or low blood glucose warning zone (below the first low blood glucose threshold). Optionally, the first high blood glucose threshold is 10.0 mmol / L and the first low blood glucose threshold is 3.9 mmol / L.

[0078] The slope triggering mechanism in this application is a predictive triggering mechanism based on the first derivative (rate of change) of blood glucose. When the rate of increase in blood glucose exceeds a first rate of increase threshold or the rate of decrease falls below a first rate of decrease threshold, even if the current absolute blood glucose value is still within the normal range, the system 10 will still determine it as a suspected event, thereby providing a longer lead time for dealing with postprandial hyperglycemia or post-exercise hypoglycemia. Optionally, the first rate of increase threshold is 0.5 mmol / L / min, and the first rate of decrease threshold is -0.4 mmol / L / min.

[0079] It is understandable that blood glucose fluctuations in prediabetic individuals are often deeply coupled with diet and high-intensity exercise. This embodiment of the application, specifically targeting prediabetic patients, makes the "sensitivity switch" for triggering suspected event recognition more sensitive. The system 10 lowers the rise slope threshold, for example, from 0.5 mmol / L / min to 0.3 mmol / L / min, to achieve "very early detection" of slight fluctuations in prediabetic patients. Therefore, targeted adjustments help identify transient blood glucose increases caused by lifestyle changes (such as overeating or sudden stress), thereby guiding prediabetic patients to make real-time corrections through diet or exercise, preventing progression to a confirmed diabetes diagnosis.

[0080] As a specific application scenario, for users with prediabetes, System 10 will make the "sensitivity switch" for suspected event recognition more sensitive. Specifically, System 10 will lower the first rate of increase threshold from the default baseline value (e.g., 0.5 mmol / L / min) to a first specific threshold (e.g., 0.3 mmol / L / min). This quantitative adjustment allows System 10 to capture weak metabolic signals caused by diet or acute stress before the absolute blood glucose value exceeds the safe range, achieving "very early capture" of slight fluctuations in prediabetic patients, thus providing more response time for subsequent verification and user intervention.

[0081] Specifically, the intelligent event triggering module 200 establishes real-time data communication with the signal acquisition and synchronization module 100 to perform high-frequency monitoring of the aligned blood glucose data stream. This aims to accurately capture clinically significant suspected blood glucose events through multi-dimensional logical judgment. The intelligent event triggering module 200 incorporates dual triggering logic: threshold triggering based on absolute blood glucose values ​​and slope triggering based on the first derivative (rate of change). The threshold triggering logic states that when the current CGM blood glucose value enters a preset extreme range (blood glucose value greater than 10.0 mmol / L or less than 3.9 mmol / L), the system 10 determines that the current physiological state has reached a safety threshold and immediately marks it as a suspected blood glucose event. The slope triggering logic states that the system 10 calculates the instantaneous rate of change of the blood glucose data stream in real time. Even if the absolute blood glucose value is within the normal range, if the upward slope is greater than 0.5 mmol / L / min or the downward slope is less than -0.4 mmol / L / min, the system 10 also identifies it as a suspected event. Therefore, by introducing a slope triggering mechanism based on the rate of change, the system can identify abnormal trends in advance before the absolute blood glucose value exceeds the safe range, thus gaining more response time for subsequent verification and user intervention.

[0082] To meet the specific needs of prediabetic patients for managing transient fluctuations, the intelligent event triggering module 200 implements an adaptive parameter adjustment function based on "disease stage labels." The system 10 automatically retrieves the corresponding threshold matrix based on the user profile. When the preset disease stage label is identified as "prediabetes," the intelligent event triggering module 200 automatically executes a threshold reduction command. Since impaired glucose tolerance in prediabetic patients is mainly manifested as a rapid rise in postprandial blood glucose, by lowering the trigger slope, the system 10 can capture weaker abnormal metabolic signals. Therefore, the system 10 can effectively distinguish early fluctuations in prediabetic individuals under eating or stress conditions, thus providing more preventative health intervention recommendations than for diagnosed patients.

[0083] Furthermore, once any of the above conditions are triggered, the intelligent event triggering module 200 identifies and marks the current blood glucose stream as a suspected blood glucose event, and generates a timestamp index containing the current time and its preceding time window (e.g., 15 minutes). This index is sent as an instruction to the feature extraction module 300, guiding it to perform targeted multimodal feature slicing, thereby achieving a mode switch from "global low-frequency monitoring" to "local high-frequency verification," reducing the normal computational power consumption of the backend algorithm.

[0084] It should be noted that the determination of the first hyperglycemia threshold, the first hypoglycemia threshold, the first rate of increase threshold, and the first rate of decrease threshold in this application embodiment is not fixed, but is determined comprehensively based on clinical benchmarks, personalized dynamic generation, and adaptive adaptation to the disease stage. Specifically, the system 10 presets a set of initial values ​​that conform to clinical medical consensus as the starting point for system operation. The determination of the first hyperglycemia threshold and the first hypoglycemia threshold refers to the standard warning lines for diabetes management, setting the first hyperglycemia threshold at 10.0 mmol / L and the first hypoglycemia threshold at 3.9 mmol / L. The determination of the first rate of increase threshold and the first rate of decrease threshold is based on the physiological limit of interstitial fluid glucose fluctuation, with the initial value set as a rate of increase greater than 0.5 mmol / L / min and a rate of decrease less than -0.4 mmol / L / min.

[0085] Furthermore, the system 10 adjusts according to each user's specific physiological performance: the intelligent event triggering module 200 analyzes the user's historical CGM data stream and identifies the user's personal blood glucose fluctuation baseline. The system 10 dynamically generates an individualized warning interval based on the user's average blood glucose and coefficient of variation over a period of time (such as 24 hours or longer). In other words, for users with high or low baseline blood glucose, the numerical threshold for triggering a suspected event will fluctuate accordingly.

[0086] In the intelligent event triggering module 200, the system 10 dynamically adjusts the first rise rate threshold and / or the first fall rate threshold according to preset disease stage labels. When the label indicates that the user is prediabetic, the system 10 adopts a more sensitive monitoring strategy, that is, lowers the value of the first rise rate threshold. Specifically, the threshold adjustment can be achieved by querying a preset threshold mapping table or according to a predefined adjustment function. As a preferred embodiment, the adjustment strategy of this application embodiment is: setting the first rise rate threshold to a dynamic value that is negatively correlated with the user's historical blood glucose fluctuation coefficient. For example, the system 10 calculates the average peak value of the user's blood glucose rise rate over the past week (S). mean ) and standard deviation (S std And determine the individualized threshold according to the following formula: First rate of ascent threshold = Base value - k*S stdThe baseline value is the default threshold (e.g., 0.5 mmol / L / min), and k is the sensitivity coefficient. When the disease course label is early stage, k takes a larger value (e.g., 1.5), otherwise it takes a smaller value (e.g., 0.5), thus ensuring that the system can trigger monitoring with a lower threshold for the early stage population with greater fluctuations.

[0087] It should be noted that "k*S" in the above formula std "To adjust the offset. Because the blood glucose regulation ability of prediabetic patients is in the early stage of impairment, the dispersion of their blood glucose rise rate (S)" std The value of k often reflects the instantaneous characteristics of pancreatic islet stress; by increasing the value of k, the calculated first rate of ascent threshold is made to be closer to 0 in value, thereby improving the system 10’s sensitivity to early abnormal climbing fluctuations.

[0088] The feature extraction module 300, connected to the intelligent event triggering module 200, is used to extract a multimodal feature vector containing blood glucose change features, HRV index calculated based on PPG signal, heart rate change features, and skin temperature change features within the time window of a suspected blood glucose event.

[0089] In this embodiment of the application, the time window refers to a specific time interval for data extraction that is based on the time of triggering a suspected blood glucose event and extends forward or backward. The time window can be set from 15 minutes before the suspected blood glucose event to the current time. By setting the time window, the feature extraction module 300 can capture the entire physiological stress process before and after the occurrence of drastic blood glucose fluctuations, providing dynamic background data for subsequent consistency verification.

[0090] The multimodal feature vectors in this application are used to characterize the user's overall physiological profile within a specific time window, including blood glucose change features (such as ΔGlucose, blood glucose slope), heart rate change features (such as ΔHR (the difference between the current real-time heart rate value and the user's individual baseline heart rate value), heart rate trend slope), HRV index, and skin temperature change features.

[0091] Optionally, in some embodiments, the HRV indicators extracted by the feature extraction module 300 include SDNN (reflecting total autonomic activity) and RMSSD (reflecting parasympathetic tone), and the extracted features include the HRV indicators, heart rate and skin temperature changes relative to the user's individual baseline within a time window.

[0092] In this embodiment, the change relative to the user's individual baseline refers to the difference between the current feature value and its preset long-term individual baseline value (such as the average value of the past 24 hours). Since there are significant differences in the baseline heart rate, baseline skin temperature and glucose tolerance among different individuals, directly using the absolute value will lead to poor model generalization ability. This embodiment of the application achieves feature normalization and deindividuation by extracting the change.

[0093] Specifically, after receiving a suspected blood glucose event notification sent by the intelligent event triggering module 200 containing a specific timestamp index, the feature extraction module 300 immediately extracts multi-dimensional quantitative physiological indicators within the corresponding time window (e.g., from 15 minutes before the event to the current moment) from the raw data stream cached by the signal acquisition and synchronization module 100 and which has been time-aligned, and constructs a standardized multimodal feature vector for subsequent analysis by the hybrid consistency verification engine.

[0094] The feature extraction module 300 follows this technical path for extraction and construction: First, raw signal localization and slicing. Based on the trigger timestamp, the feature extraction module 300 accurately locates and extracts three sets of aligned data within the time window from the annular buffer: CGM blood glucose sequence, preprocessed photoplethysmography (PPG) signal, and skin temperature signal.

[0095] It should be noted that, since the sampling rates of CGM (5 minutes / time) and PPG (not less than 1Hz) are different, the feature extraction module 300 will perform linear interpolation or equal-interval resampling after locating the slice to ensure that the data of each modality have consistency in the time dimension in the multimodal feature vector.

[0096] The second step involves parallel computation and extraction of multi-dimensional physiological features. The feature extraction module 300 processes the aforementioned data slices in parallel, calculating eight key features across four categories to form a quantitative description of the physiological state of the event. Specifically, these include the absolute blood glucose value at the trigger moment (Glucose_t), the change relative to the user's individual historical baseline (ΔGlucose), and the instantaneous change slope (Slope_Glucose) obtained by linear fitting of the data within the window. The relative change (Δ) is used to eliminate differences in baseline levels between individuals, focusing on the abnormal fluctuations themselves. Furthermore, peak detection is performed on the PPG signal to generate a heartbeat interval sequence, based on which two core time-domain indicators are calculated: SDNN (standard deviation of all normal heartbeat intervals), reflecting the total activity of the autonomic nervous system, and RMSSD (root mean square difference of adjacent heartbeat intervals), specifically reflecting parasympathetic nerve tension. Before calculation, the feature extraction module 300 performs automatic artifact removal based on morphology, and only adopts the calculation results when the effective data accounts for more than 85%, ensuring feature reliability. Next, the mean heart rate within the window (HR_mean) and its change in heart rate relative to the user's resting baseline (ΔHR) are calculated as a direct quantitative indicator of the degree of sympathetic nerve activation. Finally, the mean skin temperature within the window (Temp_mean) and its change in skin temperature relative to baseline (ΔTemp) are calculated to characterize the vasomotor response that may be triggered by glycemic events.

[0097] The third step is the assembly and output of the standardized feature vector. The nine feature values ​​calculated above—Glucose_t, ΔGlucose, Slope_Glucose, SDNN, RMSSD, HR_mean, ΔHR, Temp_mean, and ΔTemp—are assembled into a fixed-dimensional feature vector in a preset order. This vector constitutes a complete and computable physiological snapshot of the suspected blood glucose event, serving as the data foundation for System 10's AI-powered intelligent verification.

[0098] The hybrid consistency verification engine 400, connected to the feature extraction module 300, is used to receive multimodal feature vectors and perform joint discrimination based on a preset medical rule model and a trained machine learning model, outputting the consistency probability that the suspected blood glucose event is a real physiological event.

[0099] In this application, the hybrid consistency verification is a dual verification mechanism that combines deterministic medical logic and probabilistic statistical models. It is used to determine whether sensor data fluctuations match the actual physiological response of the human body. The hybrid consistency verification engine 400 uses medical prior knowledge as a defense boundary. Its core objective is to identify and filter mechanical errors caused by sensor drift and pressure, so as to ensure that the output blood glucose warning has a high degree of biological credibility.

[0100] The medical rule model in this application is a set of logical judgments that is based on clinical medical consensus and has strong interpretability, serving as a "filter" before feature vectors enter the model.

[0101] Optionally, in some embodiments, in the hybrid consistency verification engine 400, the preset medical rule model is set based on the coordinated physiological change pattern of the autonomic nervous system and peripheral vascular response under hyperglycemia or hypoglycemia events, and is used to perform prior logic verification on the multimodal feature vectors; the machine learning model is an ensemble classification model based on gradient boosting decision trees, which is used to calculate the multimodal feature vectors and output the consistency probability.

[0102] The co-physiological change pattern refers to the accompanying responses of the autonomic nervous system and microcirculation system when blood glucose fluctuates. For example, true hyperglycemia is usually accompanied by parasympathetic inhibition (significant decrease in HRV indicators such as RMSSD) and sympathetic compensation (increased heart rate); hypoglycemia, as a severe physiological stress, triggers a strong release of catecholamines, manifested as a sharp increase in heart rate and peripheral vasoconstriction (decreased skin temperature, i.e., the "cold sweat" effect). The prior logic verification in this application means that if the feature vector completely violates the above physiological laws (e.g., CGM shows extremely low blood glucose but heart rate and HRV show no fluctuation), the rule layer will downgrade its consistency rating and directly block possible sensor false alarms.

[0103] The ensemble classification model based on gradient boosting decision tree in this application refers to a lightweight XGBoost model. This model supports an online learning mechanism. When the user provides the actual value of the fingertip blood, the model can iterate through gradient boosting and fine-tune the parameter weights for the specific physiological response pattern of the individual (such as some users being insensitive to hypoglycemia). This enables the model to evolve from "general" to "personalized". Furthermore, the lightweight implementation of the gradient boosting decision tree can adapt to the low power consumption requirements of terminal devices.

[0104] The consistency probability in this embodiment is a value within the closed interval [0, 1] output by the hybrid consistency verification engine 400, which quantifies the credibility of a suspected blood glucose event as a real physiological event. This consistency probability is the sole basis for subsequent grading decisions. For example, when the consistency probability P ≥ 0.7, the system 10 determines it as a "real event" and triggers intervention; when the consistency probability P < 0.7, it determines it as a "suspicious event" and initiates a secondary verification process, thereby eliminating alarm fatigue caused by false alarms.

[0105] The hybrid consistency verification engine 400 receives a standardized multimodal feature vector F from the feature extraction module 300 and performs joint discrimination based on medical logic and machine learning to output the consistency probability P of a suspected blood glucose event. Specifically, the hybrid consistency verification engine 400 first calls a preset medical rule model to perform a priori logic scan on the feature vector F. This rule layer is set based on the coordinated physiological response mode of the human autonomic nervous system and peripheral blood vessels under blood glucose fluctuations. Specifically, if the suspected event is hyperglycemia, the rule layer checks whether the HRV index (such as RMSSD) in the feature vector shows a significant decrease (e.g., a decrease of more than 20% from the baseline) and whether the heart rate change ΔHR is positive. If the above physiological consistency is not met, the rule layer will output a lower prior weight. If the suspected event is hypoglycemia, the rule layer checks whether it is accompanied by a severe increase in heart rate (ΔHR>10bpm) and a decrease in skin temperature (ΔTemp<-0.5℃, representing the cold sweat effect). If the physiological manifestations violate this logic, the system initially marks it as an "atypical physiological reaction".

[0106] In the parallel path, the hybrid consistency verification engine 400 inputs the feature vector F into the trained machine learning model. This embodiment employs an ensemble classification model based on gradient boosting decision trees (specifically, a lightweight XGBoost model). This model learns from the nonlinear features of massive clinical samples, enabling it to capture the complex coupling relationship between blood glucose slope, HRV fluctuation amplitude, and skin temperature change rate. Specifically, the model layer performs weighted calculations on multiple dimensions of the feature vector, outputting a preliminary score P reflecting the true physiological probability of the event. ml Compared to traditional linear regression, the model in this application embodiment is more robust to individual differences among prediabetic users in an active lifestyle.

[0107] Furthermore, the hybrid consistency verification engine 400 calculates the final consistency probability P using the following fusion function:

[0108] P=w1*P rule +w2*P ml ;

[0109] Where P is the consistency probability, w1 is the rule layer weight, and P rule w2 represents the probability / prior consistency score of the rule layer, w2 represents the model layer weights, and P represents the probability / prior consistency score of the rule layer. ml This represents the probability / machine learning prediction score of the model layer.

[0110] It should be noted that w1+w2=1. In this embodiment of the application, by adjusting the weighting coefficients w1 and w2, the system 10 can achieve dynamic balance of the judgment logic. For example, in the initial stage where the sensor is prone to drift, w1 can be appropriately increased to exert the hard-core interception role of medical rules. As the adaptive learning module 600 collects more individualized finger-prick blood calibration samples from users, the model weight corresponding to w2 will be continuously optimized through online learning, thereby achieving more individualized and accurate verification.

[0111] Furthermore, if the output probability value P≥0.7, the hybrid consistency verification engine 400 determines the suspected event as a "real physiological event" and triggers subsequent warnings and intervention suggestions; if P<0.7, it is determined as a "suspicious event", and the hybrid consistency verification engine 400 sends the instruction to the credibility determination and interaction module 500 to start the secondary verification process.

[0112] Therefore, by combining medical logic and AI probability, the embodiments of this application effectively solve the problem of "uninterpretable black box model" mentioned in the background technology. Through the interception of the rule layer, the system 10 can filter out more than 30% of the erroneous readings caused by sensor drift or local pressure at the upper logic level, thereby improving the reliability of the system 10.

[0113] The credibility determination and interaction module 500 is connected to the hybrid consistency verification engine 400. It is used to compare the consistency probability with a preset confidence threshold. If the consistency probability value is lower than the preset confidence threshold, the suspected blood glucose event is judged as a suspicious event, and a non-mandatory secondary verification prompt is sent to the user terminal.

[0114] The credibility determination and interaction module 500 in this application compares the consistency probability with a preset credibility threshold and performs hierarchical decision-making and user interaction based on the comparison result. Its core function is to transform the quantified event credibility (consistency probability P) output by the hybrid consistency verification engine 400 into clear, hierarchical, and ethically sound clinical action instructions.

[0115] In this embodiment, the preset confidence threshold is a set of decision boundary parameters that are dynamically adjusted based on the event type (high / low blood sugar), the user's risk level, and the current situation. The confidence determination and interaction module 500 compares the consistency probability P with these thresholds to classify events into different risk levels such as "confidential," "needs observation," or "suspicious," thereby triggering differentiated system 10 responses.

[0116] Specifically, the confidence determination and interaction module 500 receives the consistency probability P (range [0, 1]) from the hybrid consistency verification engine 400. The system 10 presets at least two confidence thresholds, forming a multi-level decision interval, such as a high-confidence interval, an uncertain interval, and a low-confidence interval. Specifically, when P ≥ T... high (For example, T) high =0.75), the credibility determination and interaction module 500 determines that the event is a credible event, meaning that the multimodal evidence is highly consistent and the event is highly likely to be a real physiological fluctuation. The credibility determination and interaction module 500 will immediately trigger a standardized clinical intervention protocol, such as pushing a high-level alert to the user terminal (e.g., "Rapidly rising hyperglycemia detected, attention recommended"), or, with user authorization and in accordance with clinical guidelines, sending a temporary basal rate adjustment suggestion to the associated insulin pump (this function requires additional authorization and security verification). When T low ≤P <T high (For example, T) low =0.4, T high =0.75), the credibility determination and interaction module 500 determines the event to be an uncertain event. This range indicates that there is contradictory or insufficient physiological evidence, which cannot be confirmed as true nor can it be categorically denied. For such events, the system 10 adopts a "close monitoring, no disturbance" strategy. The credibility determination and interaction module 500 will record the details of the event and slightly increase the sensitivity to subsequent continuous monitoring, while not sending alarms that may cause anxiety to the user, only marking and prompting when the user actively views the application log. When P <T low (For example, T) low =0.4), the credibility determination and interaction module 500 determines that the event is a suspicious event. This determination strongly suggests that the current CGM reading is seriously inconsistent with the synchronized physiological response, which is very likely due to sensor drift, pressure or other technical interference. For such events, the core innovative interaction logic of the credibility determination and interaction module 500 is activated.

[0117] Furthermore, for cases deemed "suspicious events," the credibility assessment and interaction module 500 initiates a secondary verification process. The credibility assessment and interaction module 500 generates a non-mandatory, guiding verification request and sends it via push notification or in-app message to the user's terminal (such as a smartphone). The prompt text aims to clarify concerns and guide action rather than incite panic, for example: "The system has noticed abnormal fluctuations in your blood glucose readings, but other bodily signals (such as heart rate) have not shown typical changes. This may be due to interference with the sensor signal. To ensure accuracy, we recommend that you perform a measurement confirmation using a finger-prick blood glucose meter at your convenience."

[0118] After receiving the prompt, users have ample time to make their own decisions (e.g., setting a 30-minute response window). Users can choose to verify immediately by measuring with a fingertip blood glucose meter and manually entering the result through the application interface, or by having a smart blood glucose meter that supports NFC (Near Field Communication) / Bluetooth automatically transmit the result; they can also choose to ignore or postpone, but the system 10 will not continuously urge them, but will mark the event as an "unverified suspicious event" in the background, and its data will still be used for subsequent statistical analysis.

[0119] Once the user submits their fingertip blood glucose value (BG) finger Finger-prick blood glucose levels, used as the "gold standard" input for System 10 (for model evolution and real-time calibration), and the credibility determination and interaction module 500, i.e., the collaborative adaptive learning module 600, initiate the following key operations: First, real-time calibration, calculating BG. finger CGM reading (BG) at the time of triggering the event cgm The deviation Δ=BG finger -BG cgm This deviation value is used in real time to perform short-term offset compensation on the current CGM data stream, immediately improving the instantaneous accuracy of subsequent monitoring. Then, standard labels are generated by combining the complete multimodal feature vector F of this event with the data according to the BG (Browser Gauge). finger The true label ("real event" or "false positive event") of the judgment is bound to form a high-quality training sample (F, Label) of supervised information. This triggers model evolution, and the sample (F, Label) is sent to the adaptive learning module 600 in real time as input for online learning, which is used to individually fine-tune the machine learning model in the hybrid consistency verification engine 400.

[0120] Regardless of the type of interaction triggered, the credibility determination and interaction module 500 will generate a structured decision log, which includes: event time, consistency probability P, threshold used, judgment result, user interaction behavior (such as whether to verify, verification value) and subsequent actions taken by system 10.

[0121] It should be noted that the system 10 in this embodiment does not use a single global threshold, but maintains a multi-level threshold matrix for fine-grained differentiation of events with different risk levels. This matrix includes at least a high-confidence threshold (T). high ) and low confidence threshold (T) lowThe high confidence threshold is used to determine "credible events," triggering proactive alerts or automatic intervention. This threshold is set high to ensure extremely high specificity and minimize false alarms. The low confidence threshold is used to determine "suspicious events," triggering secondary verification. This threshold is set low to maintain sufficient sensitivity and avoid missing real events. The interval between the two is the buffer zone for "uncertain events."

[0122] The determination of the preset threshold is an evolutionary process from general to personalized, and from static to dynamic, which is divided into the following three stages.

[0123] The first phase is initialization based on clinical risk consensus. Specifically, when System 10 is first launched or a profile is created for a new user, it is initialized using universally accepted safety thresholds derived from large-scale clinical studies. For hypoglycemic events, due to their high risk and rapid progression, System 10 adopts a conservative strategy, with T... high It will be set relatively low (e.g., 0.65), T low The threshold was also lowered accordingly (e.g., by 0.25) to ensure that even slightly weaker evidence would trigger a high level of concern. For hyperglycemic events, given that the risk is typically chronic and cumulative, System 10 employs a prudent strategy, with T... high The value will be set relatively high (e.g., 0.80) to require stronger evidence support, thereby avoiding disturbing users with frequent false alarms.

[0124] For example, the initial threshold can be expressed as: T init =f(glucose event), where f outputs a lower value (e.g., 0.65) when the glucose event is "hypoglycemia" and a higher value (e.g., 0.80) when the glucose event is "hyperglycemia". T low With T high Maintain a fixed interval (e.g., 0.4).

[0125] The second phase is static optimization based on individual user historical data. Specifically, after the system has accumulated individual user data for a certain period (e.g., 7-14 days), the static optimization algorithm is initiated. This involves analyzing the user's historical event data and calculating two key indicators: the user's individual false positive rate (FPR). user The percentage of historical "suspicious events" that were verified as false positives: the percentage of historical "suspicious events" that were confirmed as false positives after verification via finger-prick blood; the individual user's false negative rate (FNR). user The proportion of risks that System 10 did not alarm for but actually existed: In retrospective analysis, the proportion of real risk events that System 10 did not alarm for but were later found to exist (discovered through retrospective CGM graph analysis or user logs).

[0126] For example, embodiments of this application may employ simple feedback control logic: if FPR userToo high (greater than the preset target, such as 30%) indicates that the current T low If the value is too low, excessive noise will be sent to the verification process; therefore, T should be increased proportionally. low and T high If FNR user Too high (greater than the preset target, such as 5%) indicates that the current T high If the value is too high, it may lead to underreporting of real events, so T will be reduced proportionally. high and T low .

[0127] The third stage is continuous dynamic optimization based on online feedback. After the system reaches stable operation, the threshold enters a state of continuous fine-tuning. Each user's feedback on a suspicious event via finger-prick blood verification becomes an optimization data point. This application embodiment employs Bayesian optimization or stochastic gradient descent for online learning. Specifically, the current threshold vector T = [T...] low T high [] is considered as a parameter to be optimized; the loss function is defined as:

[0128] L(T) = α* false positive loss + β* false negative loss;

[0129] The false positive loss is calculated by weighting the verified false positive events, while the false negative loss can be calculated by the unreported events found in a very small number of user reports or clinical re-examinations. The weighting coefficient β is usually much larger than α to reflect the extreme importance attached to safety.

[0130] It should be noted that, due to the frequent fluctuations in patients in the early stages, System 10 will further reduce daily disturbances to users and improve management compliance by increasing α (false positive loss weight).

[0131] After each new validation feedback is obtained, the impact of this feedback on the gradient of the loss function is calculated, and the threshold is updated with a very small learning rate η: T new =T old -η*∇L(T)

[0132] Where η is the online learning rate, which controls the step size of the threshold update to ensure the stability of the system.

[0133] To ensure safety, the system 10 in this embodiment of the application has strict safety boundary constraints on the adjustment range of the threshold, for example, T high For hypoglycemic events, the level should not exceed 0.7. low It must not be lower than 0.2 to prevent the optimization process from deviating from the safe range.

[0134] To further illustrate the ability of the Hybrid Consistency Verification Engine 400 to distinguish between physical noise and real physiological events, this application provides the following two atypical scenarios commonly encountered in the daily management of diabetic and prediabetic patients.

[0135] Example 1: Stress-induced pseudohypoglycemia (physical deviation caused by pressure).

[0136] In the scenario of Example 1, when the user is sleeping at night, the arm wearing the CGM sensor is pressed while lying on its side, which causes local tissue microcirculation to be blocked. The CGM reading drops sharply due to the restricted glucose diffusion, triggering the "slope trigger" condition (suspected hypoglycemia event) of the intelligent event triggering module 200.

[0137] Features extracted by the feature extraction module 300 showed that the CGM blood glucose level plummeted from 6.5 mmol / L to 3.2 mmol / L, but the simultaneously acquired heart rate change ΔHR was only +2 bpm (no significant sympathetic activation), and the HRV index (RMSSD) remained at the individual baseline level (no stress-induced decrease). The rule layer of the hybrid consistency verification engine 400 identified that true hypoglycemia, as a severe physiological stress, must be accompanied by a significant increase in heart rate and autonomic nervous system imbalance. Because the physiological characteristics severely violated the "hypoglycemic synergistic physiological change pattern," the rule layer output P rule The probability of final consistency was extremely low, resulting in a final consistency probability P of only 0.25. The credibility judgment and interaction module 500 judged it as a "suspicious event" and reminded the user to adjust their sleeping position by vibration instead of directly issuing a low blood sugar alarm, successfully avoiding false alarms in the middle of the night caused by "false hypoglycemia".

[0138] Example 2: Reading deviation caused by ambient temperature difference (temperature drift interference).

[0139] In the scenario of Example 2, a prediabetic user moves from a warm indoor environment to a cold outdoor environment in winter, or takes a cold shower. The drastic change in ambient temperature causes the sensitivity of the electrochemical sensor to drift, and the CGM shows a rapid rise in blood glucose, triggering the "first rate of rise threshold" (a suspected hyperglycemic event).

[0140] The feature vector extracted by the feature extraction module 300 shows a significant upward trend in blood glucose slope, but the synchronously collected skin temperature change ΔTemp shows a sharp decrease (e.g., -1.5℃), and the heart rate and HRV indicators do not show the typical metabolic fluctuations of postprandial hyperglycemia. The engine rule layer judges that real hyperglycemic events (such as postprandial hyperglycemia) are usually accompanied by increased heat production or stable metabolic rate caused by digestion, and should be accompanied by inhibition of the parasympathetic nervous system. The current combination of "hyperglycemic slope + sudden drop in skin temperature" is extremely mismatched in physiological logic. System 10 identifies this event as an "atypical physiological reaction", and the consistency probability P drops to 0.38 (in the uncertain range). System 10 guides the user through the credibility judgment and interaction module 500: "Ambient temperature fluctuations have been detected, which may interfere with sensor readings. It is recommended that you observe after your body temperature returns to stability, or perform finger-prick blood calibration."

[0141] The adaptive learning module 600 is used to calibrate the system 10 using external verification data after the user provides feedback based on secondary verification prompts, and to update the individualized parameters of the machine learning model through an online learning mechanism.

[0142] The adaptive learning module 600 in this embodiment uses calibration data actively fed back by the user to dynamically optimize and individually adapt the system 10.

[0143] In this embodiment, the individualized parameters refer to a set of weights and structural parameters within the machine learning model (such as lightweight XGBoost) of the hybrid consistency verification engine 400. During the general model phase, these parameters are trained based on population data and represent the physiological response pattern of the "average person." The goal of the adaptive learning module 600 is to adjust these parameters through online learning to better reflect the actual physiological characteristics of a specific user. For example, it may strengthen the recognition weight of the heart rate response pattern specific to a user's hypoglycemia or weaken the dependence on skin temperature characteristics, which are less sensitive to hypoglycemia.

[0144] Optionally, in some embodiments, the online learning mechanism of the adaptive learning module 600 specifically involves: comparing the external verification data fed back by the user with the CGM reading at the corresponding time to generate training samples with authenticity labels; and using an incremental learning method to fine-tune the weight parameters of the machine learning model with the training samples.

[0145] The online learning mechanism of this application embodiment refers to a technical process in which, during the deployment of system 10 on a user terminal and continuous provision of services, the existing machine learning model is fine-tuned and its performance optimized using real-time generated [user feedback, real blood glucose values] data pairs without interrupting the service or requiring centralized retraining. Unlike offline training, which requires massive amounts of data and powerful computing capabilities, the online learning mechanism of this application embodiment emphasizes lightweight, immediacy, and individualized targeting.

[0146] Specifically, when the credibility determination and interaction module 500 receives the fingertip blood glucose value (BG) submitted by the user... finger After that, the adaptive learning module 600 is activated. The adaptive learning module 600 first precisely extracts the following data from the system cache based on the timestamp of the verification event: the original multimodal feature vector F that triggered the verification. original (i.e., the characteristics of the input hybrid consistency verification engine 400), the corresponding CGM raw reading BG at that time. cg User-submitted actual fingertip blood glucose values ​​(BG) finger The adaptive learning module 600 is based on BG. finger Based on the clinical diagnostic threshold, a binary true label Y is generated.true For example, if the verification is for a suspected hypoglycemic event, and BG finger <3.9mmol / L, then Y true =1 (represents "true hypoglycemia"), otherwise Y true =0 (representing a "false positive"), thus transforming sparse, dotted blood glucose values ​​into labeled samples required for supervised learning. To further improve learning efficiency, the adaptive learning module 600 utilizes BG... finger For F original The blood glucose-related features are calibrated in real time, for example, the deviation Δ=BG is calculated. finger -BG cgm And use the calibrated blood glucose value (BG) cgm _calibrated=BG cgm +Δ) Recalculate F original The ΔGlucose and Slope_Glucose values ​​are used to generate a calibrated feature vector F. calibrated F calibrated A sample that more closely resembles the user's current physiological state can more effectively guide the model's learning. Ultimately, a standard training sample (X, Y) true The construction is complete, where X is F. original or even better F calibrated .

[0147] Furthermore, the adaptive learning module 600 uses an incremental learning algorithm to update the machine learning model in the hybrid consistency verification engine 400, rather than retraining it, to ensure real-time learning within the limited computing resources of the terminal device. The core objective of the fine-tuning in this embodiment is to improve the model's prediction output P for the current sample X. model (That is, the probability that the model considers the event to be true) should be as close as possible to the true label Y. true This can be achieved by minimizing a loss function (such as cross-entropy loss).

[0148] Regarding incremental learning strategies, this application provides two preferred strategies: For tree models (such as XGBoost), system 10 adopts a residual-based tree addition method, treating the existing model's prediction of X as the basic prediction, and calculating the prediction residual (Y). true -P modelThen, a new, shallow decision tree ("weak learner") is trained to fit the residual. This new tree is added to the original model set with a small weight (learning rate, such as 0.05). Simultaneously, a slight "pruning" process is triggered to prevent the model from becoming overly complex due to continuous learning. Through adaptive learning, System 10 dynamically adjusts the contribution of different modal features in the decision tree. For example, if validation data shows that a user's skin temperature decreases more stably than their heart rate increases during hypoglycemia, the learning process automatically increases the weight of ΔTemp-related split nodes to achieve personalized management.

[0149] For neural networks or generalized linear models, this application uses an online gradient descent method to calculate the gradient of the loss function with respect to the weights of key layers (such as the last few layers) of the model. The weights are updated in the opposite direction of the gradient with a very small learning rate (such as 1e-4). In order to prevent catastrophic forgetting (i.e., learning new samples but forgetting old patterns), elastic weight consolidation technology can be used to apply protective constraints to important old weights.

[0150] It should be noted that the above learning process incorporates multiple safeguards. Specifically, firstly, there are the learning trigger conditions. System 10 does not learn immediately after each validation. It only initiates a batch of fine-tuning when a certain number (e.g., 5) of new samples are accumulated, or during idle periods of System 10 (e.g., when charging overnight). Secondly, there is the screening of validation samples. In this embodiment, obviously abnormal or potentially erroneous user inputs (e.g., extremely unreasonable fingertip blood glucose values) are filtered out and not used for learning. There is also a performance rollback mechanism. After each learning, the model performance is quickly validated on a small, retained dataset. If the performance drops significantly, it automatically rolls back to the model version before learning.

[0151] After the adaptive learning module 600 detects a model update, the distribution of the consistency probability P for subsequent events for the user changes, along with the accuracy confirmed by secondary verification. For example, if a sensor drift pattern that previously frequently caused false alarms is assigned a lower P value and correctly guided to secondary verification after the model update, then the learning is proven effective. The updated individualized model parameters will be securely stored locally on the user's device and optionally encrypted and synchronized to the user's personal cloud account to ensure a consistent experience across different terminals.

[0152] The AI- and CGM-based self-management system for prediabetes patients proposed in this application achieves physiological delay alignment between CGM blood glucose flow and multimodal physiological signals through a signal acquisition and synchronization module. An intelligent event triggering module dynamically adjusts trigger thresholds for different disease stages in prediabetes patients. A hybrid consistency verification engine combines medical prior rules with a lightweight XGBoost model for joint discrimination. Combined with a hierarchical decision-making mechanism in the credibility judgment and interaction module and online incremental learning optimization in the adaptive learning module, a closed-loop intelligent management system is constructed that can identify and filter sensor physical deviations, actively capture real physiological events, and possess individualized continuous evolution capabilities. This effectively solves the alarm fatigue problem caused by false alarms and improves the management accuracy and compliance of prediabetes patients in complex and dynamic scenarios.

[0153] Next, referring to the accompanying drawings, a self-management method for diabetes and prediabetes patients based on AI and CGM is described according to an embodiment of this application.

[0154] Figure 2 This is a flowchart of a self-management method for diabetes and prediabetes patients based on AI and CGM, according to an embodiment of this application.

[0155] like Figure 2 As shown, the self-management method for patients with diabetes and prediabetes based on AI and CGM includes the following steps:

[0156] In step S201, real-time blood glucose data stream from the CGM device and multimodal physiological signals from the wearable device, including at least PPG signals and skin temperature signals, are acquired synchronously, and the blood glucose data stream is time-shifted and aligned according to the physiological delay of the CGM.

[0157] In step S202, the blood glucose data stream is monitored in real time. When the current blood glucose value or the rate of change of the blood glucose value meets the preset dynamic triggering conditions, it is marked as a suspected blood glucose event.

[0158] Optionally, in some embodiments, the preset dynamic triggering conditions include any of the following: CGM blood glucose level is greater than 10.0 mmol / L, and the rate of increase of CGM blood glucose level is greater than 0.5 mmol / L / min; CGM blood glucose level is less than 3.9 mmol / L, and the rate of decrease of CGM blood glucose level is less than -0.4 mmol / L / min; CGM blood glucose level enters an individualized warning range dynamically generated based on user historical data.

[0159] In step S203, multi-dimensional physiological features within the time window corresponding to the suspected blood glucose event are extracted, and a consistent feature vector containing blood glucose features, HRV index, heart rate features, and skin temperature features is constructed.

[0160] It should be noted that during the transition from step S202 to S203, system 10 executes an on-demand feature extraction strategy. That is, system 10 maintains real-time background blood glucose monitoring, and only after a suspected event is triggered does it perform specific window extraction on the high-sampling-rate multimodal signal synchronously cached in step S201. This effectively reduces the average power consumption of system 10, extends the battery life of the wearable device, and ensures that the extracted feature vectors have strong event specificity.

[0161] In step S204, the consistency feature vector is input into the hybrid consistency verification engine 400. The hybrid consistency verification engine 400 combines medical prior rules and machine learning models for joint analysis and outputs the consistency probability of suspected blood glucose events.

[0162] Optionally, in some embodiments, the medical prior rules include: for suspected hyperglycemia events, verifying whether the time window is accompanied by a decrease in heart rate variability (HRV) exceeding a first proportional threshold, an increase in heart rate, and an increase in skin temperature; for suspected hypoglycemia events, verifying whether the time window is accompanied by a decrease in HRV exceeding a second proportional threshold, a significant increase in heart rate, and a decrease in skin temperature.

[0163] Understandably, under metabolic stress caused by drastic fluctuations in blood glucose, time-domain indicators of HRV (such as RMSSD), which reflect parasympathetic tone, typically show a significant transient decrease. In this application, the embodiments refer to the statistical distribution of publicly available clinical datasets (such as large-scale physiological stress studies based on diabetic populations) and select a 20% decrease from individual baseline as the initial first / second proportional threshold (i.e., the initial default value). This threshold has been validated to ensure the capture of sympathetic activation signals while maximally filtering out normal physiological fluctuations (typically less than 10%-15%) caused by changes in body position or minor daily activities.

[0164] In step S205, the consistency probability is compared with a preset confidence threshold. If the consistency probability is lower than the preset confidence threshold, the suspected blood glucose event is judged as a low-confidence suspicious event and a secondary verification process for users is triggered. Otherwise, it is judged as a high-confidence event and a standard management response is executed.

[0165] Optionally, in some embodiments, the consistency probability is compared with a preset confidence threshold. If the consistency probability is lower than the preset confidence threshold, the suspected blood glucose event is judged as a low-confidence suspicious event, and a secondary verification process for the user is triggered. Otherwise, it is judged as a high-confidence event and a standard management response is executed. Specifically, if the consistency probability is greater than or equal to the first confidence threshold, it is judged as a credible event, and the system 10 executes the corresponding warning or intervention suggestion; if the consistency probability is greater than or equal to the second confidence threshold and less than or equal to the first confidence threshold, it is judged as an uncertain event, and the system 10 records and continuously observes it; if the consistency probability is less than the second confidence threshold, it is judged as a suspicious event, and the system 10 sends a prompt message to the user terminal, suggesting confirmation by finger-prick blood measurement, and waits for user feedback within a preset time.

[0166] It is understandable that the first confidence threshold (T) for determining credibility is... high ) and the second confidence threshold (T) low ), its initial preset value (such as T) high =0.75, T low =0.4) was obtained through a sensitivity-specificity trade-off analysis of labeled samples in the standard clinical database.

[0167] Specifically, the first confidence threshold is set to ensure that System 10 has extremely high specificity (typically >95%) when classifying events as "real events." In other words, when the multimodal evidence score reaches 0.75 or higher, the false alarm rate is strictly controlled at an extremely low level, meeting the ethical requirement in clinical practice that "strong interventions require strong evidence support." The second confidence threshold serves as the trigger boundary for "suspicious events." Its goal is to ensure that the system has sufficient sensitivity. Based on Receiver Operating Characteristic (ROC) curve analysis, the T... low Setting it to around 0.4 ensures that the vast majority of abnormal data caused by physical interference (whose model output probability is usually distributed between 0.1 and 0.3) are accurately intercepted into the verification process without the risk of missed detection.

[0168] It should be noted that in step S205, if the "uncertain event" judgment interval is entered, the system 10 will automatically activate the state machine self-locking mechanism: the system 10 will not push verification prompts to the user for the time being, but will start an observation countdown for 15-30 minutes at the same time, and dynamically lower the confidence threshold boundary of the hybrid consistency verification engine 400 in step S204, so as to achieve the "drift convergence" of probability P through the accumulation of more sampling points in the future, and finally automatically classify it as credible or suspicious.

[0169] In step S206, in response to the feedback data provided by the user in the secondary verification process, the CGM reading is calibrated, and the individualized parameters of the machine learning model are optimized through online learning using the feedback data.

[0170] Optionally, in some embodiments, online learning optimization specifically includes: taking the actual value of fingertip blood glucose reported by the user as the target, minimizing the difference between the credibility of the event predicted by the model and the actual reality of the event as the optimization objective, updating the weights of the weak classifiers in the machine learning model using a gradient boosting iterative algorithm, and the optimization process of online learning optimization is triggered during the idle period of the system 10 or after accumulating a certain preset number (e.g., no less than 5) of new samples.

[0171] After the update is performed in step S206, the AI-based and CGM-based self-management method for diabetes and prediabetes patients also includes a model robustness verification step: using a set of locally stored standard historical samples to perform a "simulated backtest" on the updated model. If the accuracy of the updated model in recognizing the historical standard samples decreases by more than a preset threshold (e.g., 5%), the system 10 determines that the current sample contains abnormal interference or user misoperation, performs model parameter rollback, and rejects this weight update to ensure the correct direction of model evolution.

[0172] It should be noted that the foregoing explanation of the AI ​​and CGM-based self-management system for patients with diabetes and prediabetes also applies to the AI ​​and CGM-based self-management method for patients with diabetes and prediabetes in this embodiment, and will not be repeated here.

[0173] According to the AI- and CGM-based self-management method for diabetes and prediabetes patients proposed in this application, a closed-loop logic from signal synchronization to adaptive evolution is constructed through steps S201 to S206: First, time-shift alignment is used to eliminate CGM physiological delay bias. Then, the trigger threshold is dynamically adjusted based on the disease stage to capture suspected blood glucose events. Next, multimodal consistency feature vectors including HRV and skin temperature are extracted and input into a verification engine that combines medical prior rules and an integrated classification model for joint discrimination. Through a probability-driven hierarchical decision-making mechanism, sensor physical noise is effectively filtered and targeted interactions are triggered. Finally, incremental learning is performed to optimize individual parameters in response to real user feedback. Thus, without changing the hardware, the accuracy, reliability, and compliance of the system for managing and warning active populations such as prediabetes are improved.

[0174] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0175] The memory 301, the processor 302, and the computer program stored on the memory 301 and capable of running on the processor 302.

[0176] When the processor 302 executes the program, it implements the AI- and CGM-based self-management method for diabetes and prediabetes patients provided in the above embodiments.

[0177] Furthermore, electronic devices also include:

[0178] Communication interface 303 is used for communication between memory 301 and processor 302.

[0179] The memory 301 is used to store computer programs that can run on the processor 302.

[0180] The memory 301 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0181] If the memory 301, processor 302, and communication interface 303 are implemented independently, then the communication interface 303, memory 301, and processor 302 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0182] Optionally, in a specific implementation, if the memory 301, processor 302, and communication interface 303 are integrated on a single chip, then the memory 301, processor 302, and communication interface 303 can communicate with each other through an internal interface.

[0183] Processor 302 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0184] This application also provides a computer program product on which a computer program is stored, which, when executed by a processor, implements the above-mentioned AI- and CGM-based self-management method for patients with diabetes and prediabetes.

[0185] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0186] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0187] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0188] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0189] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0190] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A self-management system for patients with diabetes and prediabetes based on AI and CGM, characterized in that, include: The system comprises a signal acquisition and synchronization module, an intelligent event triggering module, a feature extraction module, a hybrid consistency verification engine, a credibility determination and interaction module, and an adaptive learning module; among which, The signal acquisition and synchronization module is used to synchronously receive blood glucose data streams from continuous glucose monitoring (CGM) devices and multimodal physiological signals from wearable devices, and to perform time axis alignment processing on the blood glucose data streams and multimodal physiological signals according to preset physiological delay parameters. The multimodal physiological signals include at least the photoplethysmography (PPG) signal and skin temperature signal used to calculate heart rate variability (HRV). The intelligent event triggering module is connected to the signal acquisition and synchronization module and is used to monitor the blood glucose data stream in real time based on preset dynamic triggering conditions. When the current blood glucose data stream meets any triggering condition, the current blood glucose data stream is identified and marked as a suspected blood glucose event. The preset dynamic triggering conditions include: Threshold triggering based on blood glucose value and slope triggering based on blood glucose rate of change; The threshold trigger includes: the CGM blood glucose value is higher than a first hyperglycemia threshold or lower than a first hypoglycemia threshold; the slope trigger includes: the rate of blood glucose rise is higher than a first rate of rise threshold or the rate of blood glucose fall is lower than a first rate of fall threshold. The system dynamically adjusts the first rise rate threshold and / or the first fall rate threshold based on the user's preset disease stage labels. When the preset disease stage label is prediabetes, the absolute value of the first rate of increase threshold is lowered. The feature extraction module is connected to the intelligent event triggering module and is used to extract a multimodal feature vector containing blood glucose change features, HRV index calculated based on PPG signal, heart rate change features and skin temperature change features within the time window of the suspected blood glucose event. The hybrid consistency verification engine is connected to the feature extraction module and is used to receive the multimodal feature vector, and perform joint discrimination based on a preset medical rule model and a trained machine learning model, and output the consistency probability characterizing the suspected blood glucose event as a real physiological event. The preset medical rule model is based on the coordinated physiological change pattern of the autonomic nervous system and peripheral vascular response under hyperglycemia or hypoglycemia events, and is used to perform prior logic verification on the multimodal feature vector. The machine learning model is an ensemble classification model based on gradient boosting decision trees, used to calculate the consistency probability of the multimodal feature vectors. The credibility determination and interaction module is connected to the hybrid consistency verification engine and is used to compare the consistency probability with a preset confidence threshold. If the consistency probability value is lower than the preset confidence threshold, the suspected blood glucose event is determined to be a suspicious event, and a non-mandatory secondary verification prompt is sent to the user terminal. The adaptive learning module is used to calibrate the system using external verification data after the user provides feedback based on the secondary verification prompt, and to update the individualized parameters of the machine learning model through an online learning mechanism.

2. The AI- and CGM-based self-management system for patients with diabetes and prediabetes as described in claim 1, characterized in that, The HRV index extracted by the feature extraction module includes the standard deviation of normal sinus interval (SDNN) and the root mean square (RMSSD) of the difference between adjacent normal sinus intervals. The extracted features include the HRV index, heart rate, and skin temperature changes relative to the user's individual baseline within the time window.

3. The AI- and CGM-based self-management system for patients with diabetes and prediabetes as described in claim 1, characterized in that, The online learning mechanism of the adaptive learning module is specifically as follows: The external validation data provided by users is compared with the CGM readings at the corresponding time points to generate training samples with authenticity labels. An incremental learning approach is used to fine-tune the weight parameters of the machine learning model using the training samples.

4. A self-management method for patients with diabetes and prediabetes based on AI and CGM, characterized in that, The method using the system according to any one of claims 1 to 3 includes the following steps: Simultaneously acquire real-time blood glucose data streams from CGM devices and multimodal physiological signals from wearable devices, including at least PPG signals and skin temperature signals, and perform time-shift alignment of the blood glucose data streams according to the physiological delay of CGM. The blood glucose data stream is monitored in real time, and when the current blood glucose value or the rate of change of blood glucose value meets the preset dynamic triggering conditions, it is marked as a suspected blood glucose event. Multidimensional physiological features were extracted within the time window corresponding to the suspected blood glucose events, and a consistent feature vector containing blood glucose features, HRV index, heart rate features, and skin temperature features was constructed. The consistency feature vector is input into the hybrid consistency verification engine, which combines medical prior rules and machine learning models for joint analysis and outputs the consistency probability of the suspected blood glucose event. The consistency probability is compared with a preset confidence threshold. If the consistency probability is lower than the preset confidence threshold, the suspected blood glucose event is determined to be a low-confidence suspicious event and a secondary verification process for the user is triggered. Otherwise, it is determined to be a high-confidence event and a standard management response is executed. In response to feedback data provided by users during the secondary verification process, the CGM readings are calibrated, and the individualized parameters of the machine learning model are optimized through online learning using the feedback data.

5. The self-management method for diabetes and prediabetes patients based on AI and CGM according to claim 4, characterized in that, The preset dynamic triggering conditions include any of the following situations: The CGM blood glucose level is greater than 10.0 mmol / L, and the rate of increase of the CGM blood glucose level is greater than 0.5 mmol / L / min; The CGM blood glucose level is less than 3.9 mmol / L, and the rate of decrease of the CGM blood glucose level is less than -0.4 mmol / L / min; CGM blood glucose levels enter a personalized warning zone dynamically generated based on the user's historical data.

6. The self-management method for diabetes and prediabetes patients based on AI and CGM according to claim 4, characterized in that, The medical a priori rules include: For suspected hyperglycemia events, verify whether the time window is accompanied by a decrease in heart rate variability (HRV) exceeding the first proportional threshold, an increase in heart rate, and an increase in skin temperature. For suspected hypoglycemia events, verify whether the time window is accompanied by a decrease in HRV index exceeding the second proportional threshold, a significant increase in heart rate, and a decrease in skin temperature.

7. The self-management method for diabetes and prediabetes patients based on AI and CGM according to claim 4, characterized in that, The process involves comparing the consistency probability with a preset confidence threshold. If the consistency probability is lower than the preset confidence threshold, the suspected blood glucose event is classified as a low-confidence suspicious event, triggering a secondary verification process for the user. Otherwise, it is classified as a high-confidence event, and a standard management response is executed. Specifically, this includes: If the consistency probability is greater than the first confidence threshold, it is determined to be a credible event, and the system executes the corresponding early warning or intervention suggestion. If the consistency probability is greater than or equal to the second confidence threshold and less than or equal to the first confidence threshold, it is determined to be an uncertain event, and the system records and continuously observes it; If the consistency probability is less than the second confidence threshold, it is determined to be a suspicious event. The system sends a prompt message to the user terminal, suggesting confirmation by finger-prick blood measurement, and waits for user feedback within a preset time.

8. The self-management method for diabetes and prediabetes patients based on AI and CGM according to claim 4, characterized in that, The aforementioned online learning optimizations specifically include: The algorithm takes the actual blood glucose level reported by users as the target and minimizes the difference between the credibility of the event predicted by the model and the actual reality of the event as the optimization objective. It uses a gradient boosting iterative algorithm to update the weights of the weak classifiers in the machine learning model. The optimization process of the online learning is triggered during the system's idle period or after accumulating a preset number of new samples.

Images