CN-121641459-B - Diabetes and early-stage patient self-management system based on AI and CGM

CN121641459BCN 121641459 BCN121641459 BCN 121641459BCN-121641459-B

Abstract

The application relates to the technical field of medical care informatization, in particular to an AI and CGM-based diabetes and early-stage patient self-management system, which comprises the steps of carrying out delay alignment on CGM blood glucose flow and multi-mode physiological signals through a signal acquisition and synchronization module, dynamically adjusting trigger thresholds for different disease process stages by utilizing an intelligent event trigger module to identify suspected events, carrying out joint discrimination by combining a medical priori rule and an integrated classification model through a hybrid consistency check engine to output consistency probability, and optimizing individual parameters by matching with a hierarchical decision mechanism of an interaction module and online increment learning of a self-adaptive learning module. Therefore, the frequent false alarm caused by physical limitations such as physiological delay, signal drift and the like inherent to the CGM sensor and the fatigue and compliance reduction of the patient alarm caused by the frequent false alarm are solved, and more stable and personalized dynamic blood glucose management support is provided for active people such as pre-diabetes and the like.

Inventors

WEN JINHUI
LI SHIYUN
WANG SHAN
XIONG PING
ZHANG HAO

Assignees

成都大学附属医院(成都市创伤骨科研究所)

Dates

Publication Date: 20260505
Application Date: 20260203

Claims (8)

1. A diabetes and early-stage patient self-management system based on AI and CGM is characterized by comprising a signal acquisition and synchronization module, an intelligent event triggering module, a feature extraction module, a hybrid consistency check engine, a credibility judgment and interaction module and an adaptive learning module, wherein, The signal acquisition and synchronization module is used for synchronously receiving the blood glucose data stream from the continuous blood glucose monitoring CGM equipment and the multi-mode physiological signal from the wearable equipment, and carrying out time axis alignment processing on the blood glucose data stream and the multi-mode physiological signal according to a preset physiological delay parameter; wherein the multi-modal physiological signal comprises at least a photoplethysmogram PPG signal and a skin temperature signal for calculating heart rate variability HRV; The intelligent event triggering module is connected with the signal acquisition and synchronization module and is used for monitoring the blood glucose data stream in real time based on preset dynamic triggering conditions, and 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: a threshold trigger based on the blood glucose value and a slope trigger based on the rate of change of blood glucose; the threshold trigger comprises that the CGM blood sugar value is higher than a first hyperglycemia threshold value or lower than a first hypoglycemia threshold value, and the slope trigger comprises that the blood sugar rising speed is higher than a first rising speed threshold value or the blood sugar falling speed is lower than a first falling speed threshold value; The system dynamically adjusts the first rising rate threshold and/or the first falling rate threshold according to a preset disease stage label of a user; When the preset disease stage label is a pre-diabetes mellitus stage, the absolute value of the first rising speed threshold value is regulated down; The feature extraction module is connected with the intelligent event triggering module and is used for extracting multi-mode feature vectors containing blood sugar change features, HRV indexes calculated based on PPG signals, heart rate change features and skin temperature change features in a time window of occurrence of the suspected blood sugar event; the mixed consistency check engine is connected with the feature extraction module and is used for receiving the multi-mode feature vector, carrying out joint discrimination based on a preset medical rule model and a trained machine learning model and outputting consistency probability representing that the suspected blood glucose event is a real physiological event; the preset medical rule model is set based on a collaborative physiological change mode of the autonomic nervous system and peripheral vascular response under a hyperglycemia or hypoglycemia event, and is used for carrying out priori logic verification on the multi-mode feature vector; The machine learning model is an integrated classification model based on a gradient lifting decision tree and is used for calculating the multi-modal feature vector and outputting the consistency probability; the credibility judging and interacting module is connected with the mixed consistency verification engine and is used for comparing the consistency probability with a preset confidence threshold value, judging the suspected blood sugar event as a suspicious event if the consistency probability value is lower than the preset confidence threshold value, and sending an optional secondary verification prompt to a user terminal; and the self-adaptive learning module is used for calibrating the system by utilizing the verification data after the user feeds back the external verification data based on the secondary verification prompt, and updating the individuation parameters of the machine learning model through an online learning mechanism.
2. The AI-and CGM-based diabetes and pre-patient self-management system according to claim 1, wherein the HRV indicator extracted by the feature extraction module comprises a standard deviation, SDNN, of normal sinus cardiac intervals and a root mean square, RMSSD, of adjacent normal sinus cardiac interval differences, and the extracted features comprise the HRV indicator, heart rate, and variation of skin temperature over the time window relative to individual user baseline.
3. The AI-and CGM-based diabetes and pre-patient self-management system according to claim 1, wherein the online learning mechanism of the adaptive learning module is specifically: comparing the external verification data fed back by the user with the CGM reading at the corresponding moment to generate a training sample with an authenticity label; and fine tuning the weight parameters of the machine learning model by using the training sample in an incremental learning mode.
4. A method for self-management of diabetes mellitus and pre-patients based on AI and CGM, characterized in that it uses the system according to any one of claims 1 to 3, said method comprising the steps of: synchronously acquiring a real-time blood glucose data stream from a CGM device and a multi-mode physiological signal at least comprising a PPG signal and a skin temperature signal from a wearable device, and performing time-shift alignment on the blood glucose data stream according to physiological delay of the CGM; Monitoring the blood glucose data flow in real time, and marking as a suspected blood glucose event when the current blood glucose value or the change rate of the blood glucose value meets a preset dynamic trigger condition; extracting multidimensional physiological characteristics in a time window corresponding to the suspected blood glucose event, and constructing a consistency characteristic vector comprising blood glucose characteristics, HRV indexes, heart rate characteristics and skin temperature characteristics; inputting the consistency feature vector into a hybrid consistency check engine, and carrying out joint analysis by the hybrid consistency check engine in combination with a medical priori rule and a machine learning model to output the consistency probability of the suspected blood glucose event; Comparing the consistency probability with a preset confidence threshold, if the consistency probability value is lower than the preset confidence threshold, judging the suspected blood sugar event as a low-credibility suspicious event, triggering a user-oriented secondary verification process, otherwise, judging the suspected blood sugar event as a high-credibility event and executing standard management response; And responding to feedback data provided by a user in the secondary verification process, calibrating the CGM reading, and optimizing the individuation parameters of the machine learning model by online learning by utilizing the feedback data.
5. The AI-and CGM-based diabetes and pre-patient self-management method according to claim 4, wherein the predetermined dynamic trigger condition comprises any one of the following conditions: the CGM blood sugar value is more than 10.0mmol/L, and the rising slope of the CGM blood sugar value is more than 0.5mmol/L/min; The CGM blood sugar value is smaller than 3.9mmol/L, and the descending slope of the CGM blood sugar value is smaller than-0.4 mmol/L/min; the CGM blood glucose level enters an individualized alert zone dynamically generated based on user history data.
6. The AI-and CGM-based diabetes and pre-patient self-management method according to claim 4, wherein the medical prior rule comprises: For a hyperglycemia suspected event, checking whether the decreasing amplitude of the Heart Rate Variability (HRV) index in the time window exceeds a first proportional threshold, the rising of the heart rate and the rising of the skin temperature; for a hypoglycemic suspected event, checking whether the decrease in HRV indicator is accompanied by a decrease in the amplitude of the HRV indicator exceeding a second proportional threshold, a significant increase in heart rate, and a decrease in skin temperature within the time window.
7. The AI-CGM based diabetes mellitus and pre-patient self-management method according to claim 4, wherein the comparing the coincidence probability with a preset confidence threshold value, if the coincidence probability value is lower than the preset confidence threshold value, determining the suspected glycemic event as a low confidence suspicious event and triggering a user-oriented secondary verification process, otherwise, determining as a high confidence event and performing a standard management response, specifically comprising: If the consistency probability is larger than a first confidence threshold, judging that the event is a trusted event, and executing corresponding early warning or intervention suggestion by the system; if the consistency probability is greater than or equal to a second confidence threshold and less than or equal to the first confidence threshold, judging that the event is uncertain, and recording and continuously observing by a system; If the consistency probability is smaller than the second confidence threshold, judging that the event is suspicious, sending prompt information to the user terminal by the system, suggesting to confirm through fingertip blood measurement, and waiting for user feedback within preset time.
8. The AI-and CGM-based diabetes and pre-patient self-management method according to claim 4, wherein the online learning optimization specifically comprises: And updating the weak classifier weight in the machine learning model by adopting a gradient lifting iterative algorithm by taking a fingertip blood sugar true value fed back by a user as a target and taking the difference between the minimum model prediction event credibility and the event actual authenticity as an optimization target, wherein the optimization process of online learning optimization is triggered after a system idle period or a new sample with a preset number is accumulated.

Description

Diabetes and early-stage patient self-management system based on AI and CGM Technical Field The application relates to the technical field of medical care informatization, in particular to an AI and CGM-based diabetes and early-stage patient self-management system. Background The continuous blood glucose monitoring (Continuous Glucose Monitoring, CGM for short) technology provides possibility for the fine management of diabetes and early-stage patients by measuring the interstitial fluid glucose concentration in real time. However, the physical and physiological deviations that are inherently insurmountable in their data flow constitute a major obstacle to current precise management. The most central problem is that the interstitial fluid glucose concentration in subcutaneous tissue lags the blood glucose by about 5 to 15 minutes, which is especially pronounced during rapid changes in blood glucose (e.g. postprandial, after strenuous exercise), resulting in a significant shift in CGM readings from the true physiological state. More troublesome is sensor drift, the sensitivity and the base line of the electrochemical sensor can be subjected to unpredictable slow changes along with time, temperature, local tissue reaction and the like, systematic errors are generated, and the errors directly promote two risks in clinical scenes, namely false alarm, misjudgment of sensor noise or drift as a real blood sugar event, and false alarm, and failure to timely capture real dangerous fluctuation due to delay or sensitivity reduction. Frequent false positives for patients, especially night-time, endless, low blood glucose alarms, can seriously compromise their trust in technology, leading to "alarm fatigue" or even compliance collapse. To solve the above problems, the related art mainly follows two paths, but there are obvious limitations. The first path is to optimize the sensor hardware itself, aimed at improving accuracy and stability through material science and electrochemical improvements. However, this path has a long iteration period, is costly, and cannot fundamentally eliminate the inherent property of physiological delay. The second path is to rely on algorithms to post-process the CGM data, and common methods include signal smoothing based on a model such as kalman filtering, or rough anomaly tagging with a single auxiliary signal (e.g. heart rate, motion data), e.g. tagging a hyperglycemic event at heart rate acceleration. However, such schemes are generally put into thinking of one-way verification. In particular, the related art still relies mainly on a single data stream of blood glucose, lacking a synchronous proof of a multidimensional physiological state. A true hypoglycemic stress is a combined result of the synergistic effects of sympathetic excitation (heart rate, HRV (HEART RATE Variability, heart rate variability) changes), endocrine response (cortisol, etc.), and peripheral vasoconstriction (skin temperature drop). Any algorithm that ignores this physiological coupling, the discriminant logic of which is fragile and susceptible to interference. In the related art, most algorithms adopt fixed population threshold values or general models, huge physiological differences among individuals cannot be considered, and the autonomic nerve response modes of early-stage diabetic population and long-term aged diabetic patients to blood sugar fluctuation are quite different, that is, the related art has the problem of model rigidification. In addition, the deep black box model may give a "suspicious" judgment, but fails to provide a reasonable interpretation that is medically common to the user or doctor (e.g. "this hyperglycemia alert, no predicted decrease in your heart rate variability, possibly related to sensor signals"), which is a significant drawback in medical applications where security and confidence are emphasized. The pre-diabetic population is the golden window of healthy intervention. Compared to diagnosed patients, their patterns of blood glucose fluctuations are more complex and active, often coupled with deep diets, high intensity intermittent exercise, acute stress events, etc. A transient, post-exercise transient hyperglycemia is quite different in clinical significance and management strategy from a sustained rising postprandial hyperglycemia. The lack of the ability of related art CGM systems to finely distinguish such dynamic patterns often gives unnecessary alarms to the former, resulting in health anxiety, which may instead prevent the patient from adhering to a positive lifestyle (such as exercise) that may cause blood glucose fluctuations, contrary to the original intent of health management. In summary, clinical practice and technical development are commonly directed to an unmet need for an intelligent system capable of dynamically evaluating the credibility of CGM data itself, which should not be another black box alarm, but should be a multi-dimensional physiological evidence-