An ai dynamic modeling system for insulin resistance risk assessment

Abstract

The present application relates to the technical field of insulin resistance risk assessment, and particularly relates to an AI dynamic modeling system for insulin resistance risk assessment. The present application first acquires multi-modal data of each type of window through a data acquisition module; further analyzes the blood glucose, insulin and fluctuation at the falling asleep moment of the proximal window within the distal window, combines the abnormal heart rate performance, and acquires the proximal IR coefficient; further acquires the disease course stage parameters of the current distal window according to the change of the slope of the fluctuation curve of each biochemical data, combines the current BMI value distribution within the distal window, the sleep deficiency performance and the change of the clinical diagnosis record, and the distribution of the proximal IR coefficient; finally, the modeling analysis module allocates the attention weight of each distal window, and trains the deep learning model based on the attention weight, so as to improve the dynamic sensitivity and prediction accuracy of the model to the current risk.

Description

Technical Field

[0001] This invention relates to the field of insulin resistance risk assessment technology, and specifically to an AI dynamic modeling system for insulin resistance risk assessment. Background Technology

[0002] Insulin resistance (IR) is a core pathological basis for metabolic diseases such as type 2 diabetes, obesity, and non-alcoholic fatty liver disease. Early identification and intervention can significantly reduce the risk of morbidity and complications. Therefore, constructing an early, dynamic, and personalized risk assessment model for IR has significant clinical and public health value. Existing assessment methods mostly rely on static indicators or costly and invasive testing methods, which are difficult to reflect the dynamic evolution characteristics of IR. With the accumulation of multimodal time-series health data, using AI models to achieve dynamic risk assessment has become a trend.

[0003] However, insulin resistance (IR) is not a static state but a dynamic evolutionary process. Therefore, the predictive power of multimodal data collected at different time points for IR is time-dependent, resulting in significantly different risk prediction capabilities for data collected at different time points. However, most existing systems currently assign higher weight to recent data, assuming that the recent state better reflects the current risk, but neglecting the long-term cumulative effect of insulin resistance development, thus affecting the accuracy of IR risk assessment. Summary of the Invention

[0004] To address the technical problem that existing models have unreasonable attention weight settings for different time periods, making it difficult to accurately reflect the dynamic evolution of insulin resistance, the present invention aims to provide an AI dynamic modeling system for insulin resistance risk assessment. The specific technical solution adopted is as follows: An AI dynamic modeling system for assessing insulin resistance risk, the system comprising: The data acquisition module is used to divide the time hierarchy into preset near-end, mid-range, and far-end windows and acquire multimodal data for each type of window; The disease course analysis module is used to obtain the proximal IR coefficient based on the fluctuations of blood glucose, insulin, and sleep time in the proximal window within the current distal window, combined with abnormal heart rate manifestations; it also fits the slope fluctuation curves of biochemical data of each dimension in different intermediate windows, and obtains the disease course stage parameters of the current distal window based on the changes in the fluctuation curves, combined with the changes in BMI value distribution, sleep deprivation manifestations, and clinical diagnosis records within the current distal window, as well as the distribution of the proximal IR coefficient; The modeling and analysis module assigns attention weights to each remote window based on the changes in the parameters of the disease stage and the time interval between different remote windows and the current remote window; and trains a deep learning model based on the attention weights.

[0005] Furthermore, the method for obtaining the proximal IR coefficient includes: Based on the offset between the peak blood glucose and peak insulin levels within each proximal window, short-term offset parameters are obtained; the sympathetic-parasympathetic balance index of heart rate variability in each proximal window is extracted; and the evaluation feature index is obtained based on the differences between different proximal windows within the current distal window in three dimensions: peak blood glucose, time of sleep onset, and mean blood glucose. The proximal IR coefficient of each proximal window is obtained by fusing the short-term offset parameter, the sympathetic-parasympathetic balance index, and all the corresponding evaluation feature indices for each proximal window.

[0006] Furthermore, the method for obtaining the short-term offset parameter includes: The time interval between the time point of the blood glucose peak and the time point of the insulin peak within each proximal window is obtained as a short-term offset parameter.

[0007] Furthermore, the method for obtaining the evaluation feature index includes: For any two proximal windows, the absolute values ​​of the differences in the three dimensions of blood glucose peak, sleep time, and mean blood glucose are normalized respectively, and the product of the normalized results of the three dimensions is used as the initial feature index; the sum of all the initial feature indices of each proximal window in the distal window is normalized to obtain the evaluation feature index of each proximal window.

[0008] Furthermore, the method for obtaining the disease stage parameters includes: Based on the differences in the fluctuation curves of the mid-range windows in each dimension of the adjacent far-end windows in the same order, and combined with the biochemical data values ​​of the corresponding dimensions, the biochemical indicator IR stage index is obtained. Based on the changes in BMI value distribution, sleep deprivation manifestations, and clinical diagnostic records within the current remote window, obtain the remote IR coefficient; The negative correlation normalization result of the mean of the proximal IR coefficients of all proximal windows in each mid-range window is used as the IR development regulation factor; based on the IR development regulation factor, the product of the biochemical index IR stage index and the distal IR coefficient of each mid-range window in the current distal window is fused to obtain the IR progression characteristic index of the current distal window. By fusing the mean of the proximal IR coefficients of all the proximal windows and the IR progression characteristic index, the disease stage parameters are obtained.

[0009] Furthermore, the method for obtaining the biochemical indicator IR stage index includes: Biochemical data for each dimension is selected as target data, and the derivative of the time series position corresponding to each mid-range window on the fluctuation curve is obtained as the fluctuation stability index of the target data at the corresponding time point. Based on the difference in the fluctuation stability index of the intermediate windows in the same order of adjacent far windows, obtain the reference weight of all intermediate windows in the current far window; based on the reference weight, fuse the target data of all intermediate windows in the current far window as the biochemical stage index. The mean of the biochemical stage indices across all dimensions is used as the biochemical indicator IR stage index.

[0010] Furthermore, the method for obtaining the far-end IR coefficient includes: Based on the changes in the clinical diagnosis records of adjacent remote windows, the number of newly added symptoms in the current remote window is extracted; the sum of the differences between the preset recommended sleep duration and the patient's actual daily sleep duration in the remote window is taken as sleep debt; the number of newly added symptoms, the sleep debt, and the average BMI value are fused to obtain the remote IR coefficient.

[0011] Furthermore, the method for obtaining the attention weights includes: Obtain the first-order difference sequence of the disease stage parameters, segment the time domain according to the positive or negative sign of the difference values, and obtain the initial weight of each remote window based on the difference value of each remote window, combined with the length of the segment and the overall characteristics of the disease stage parameters. The initial weights are adjusted based on the time interval between each of the remote windows and the current remote window to obtain the attention weights for each remote window.

[0012] Furthermore, the method for obtaining the initial weights includes: The absolute value of the difference value of each of the distal windows is used as the disease progression coefficient; the product of the number of segmented windows and the disease progression coefficient of each of the distal windows is used as the disease progression correlation coefficient of each of the distal windows. The normalized result of the mean of all the disease stage parameters included in the segment is used as the IR acceleration coefficient. The disease progression coefficient is weighted based on the IR acceleration coefficient. The disease progression correlation coefficient is weighted using the difference between the constant 1 and the IR acceleration coefficient. The weighted sum is used as the initial weight of the corresponding distal window.

[0013] Further, the method for correcting the initial weights based on the time interval between each of the far windows and the current far window to obtain the attention weights of each far window includes: The negative correlation mapping result between each of the far windows and the current far window is used as a correction coefficient, and the normalized result of the product of the correction coefficient of each far window and the initial weight is used as the attention weight.

[0014] The present invention has the following beneficial effects: This invention first acquires multimodal data for each type of window through a data acquisition module, providing a foundation for subsequent analysis. Next, a disease progression analysis module analyzes fluctuations in blood glucose, insulin, and sleep onset time within the proximal window of the distal window, combined with abnormal heart rate manifestations, to obtain the proximal IR coefficient. This coefficient serves as an assessment characteristic index for early IR development, providing a basis for subsequent assessment of disease progression parameters and attention weights within the distal window. Furthermore, based on the changes in the slope fluctuation curves of each biochemical data point, the invention analyzes the trend fluctuations of biochemical data across different distal windows. Combined with the current BMI distribution and sleep deprivation manifestations within the distal window, the invention analyzes the impact of patients' unhealthy lifestyle habits, and analyzes changes in clinical diagnostic records and the distribution of the proximal IR coefficient to obtain disease progression parameters. These parameters characterize the staged state of insulin resistance development and provide a basis for subsequent attention weight allocation. Finally, a modeling analysis module analyzes changes in disease progression parameters to analyze the dynamic evolution trend of IR development. Combined with the time interval between the distal window and the present, attention weights are allocated to each distal window. This approach maintains long-term memory while suppressing interference from outdated information, improving the model's dynamic sensitivity to current risks and predictive accuracy. A deep learning model is then trained based on these attention weights. Multimodal data with multiple time windows are acquired, and proximal IR features are extracted and combined with mid-term biochemical slope fluctuations and distal BMI, sleep and diagnostic information to generate disease stage parameters. Attention weights are dynamically allocated based on parameter changes to train a deep learning model and improve the model's dynamic sensitivity to current risk and prediction accuracy. Attached Figure Description

[0015] Figure 1 A system block diagram of an AI dynamic modeling system for assessing insulin resistance risk, provided in one embodiment of the present invention; Figure 2 A flowchart illustrating a method for obtaining near-end IR coefficients according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating a method for obtaining disease stage parameters according to an embodiment of the present invention. Detailed Implementation

[0016] The following description, in conjunction with the accompanying drawings, details the specific scheme of the AI ​​dynamic modeling system for insulin resistance risk assessment provided by this invention.

[0017] Please see Figure 1 The diagram illustrates a system block diagram of an AI dynamic modeling system for assessing insulin resistance risk according to an embodiment of the present invention. The system includes: a data acquisition module 101, a disease course analysis module 102, and a modeling analysis module 103.

[0018] The data acquisition module 101 is used to divide the time hierarchy into preset near-end, mid-range and far-end windows and acquire multimodal data for each type of window.

[0019] The dynamic evolution of insulin resistance (IR) requires support from two core types of data: first, short-term high-frequency data, which reflects transient pancreatic responses, such as postprandial insulin secretion and nocturnal blood glucose fluctuations, corresponding to the load sensitivity and rhythmic correlation characteristics of IR; second, long-term low-frequency data, which reflects cumulative pathological changes, such as HbA1c changes and annual lipid trends, corresponding to the insidious progression and staged characteristics of IR. To improve the accuracy of the model in assessing IR risk, a data acquisition scheme should be designed according to both short-term high-frequency and long-term low-frequency dimensions, with standardized data formats and timestamps to avoid losing key features during subsequent alignment.

[0020] In one embodiment of the present invention, the process of acquiring multimodal data includes: First, we collect the raw data: (1) Short-term high-frequency data is collected through wearable devices and medical instruments, including: a) Continuous glucose monitoring (CGM) data (5 minutes / time, recording fasting, postprandial, and nighttime blood glucose levels and fluctuation curves); b) Dynamic insulin monitoring data (15 minutes / time, collected only within 2 hours postprandial, recording the peak insulin secretion time and concentration); c) Physiological data (per minute, including heart rate variability (HRV), sleep structure, and steps).

[0021] The data collection period is 72 consecutive hours, repeated once a month on a fixed date, which can cover the early postprandial load response and mid-term diurnal rhythm disorder characteristics of IR.

[0022] (2) Long-term low-frequency data are collected through clinical testing and electronic health records, including: a) Biochemical indicators, specifically including fasting plasma glucose (FPG), fasting insulin (FINS), C-peptide, triglycerides (TG), glycated hemoglobin (HbA1c), glycated albumin (GA), etc. b) Clinical diagnostic information, such as the presence of impaired glucose tolerance (IGT), abdominal obesity, or obstructive sleep apnea (OSA).

[0023] The data collection cycle is: biochemical indicators once a week and clinical diagnosis once a quarter, which can cover the long-term cumulative effect of IR from subclinical to clinical to β-cell failure.

[0024] (3) Preprocess the collected data: a) Mark the acquisition time of all modal data; b) Clean the collected data, including handling missing values ​​and filtering noise; c) Use the MedBERT model to extract keywords from the patient's clinical diagnosis text information and convert them into word vectors for easier subsequent analysis and processing.

[0025] The collected data is divided according to time periods, specifically including: The proximal window, measured in days, mainly includes short-term, high-frequency data such as CGM data and dynamic insulin secretion data, reflecting timely metabolic responses; Mid-term windows measured in months primarily observe biochemical trends such as HbA1c, TG, and GA to reveal metabolic regulatory disorders. The remote window, measured in quarters, primarily combines clinical diagnosis, complications, and long-term behavioral models to reflect the long-term cumulative effect and phase shift of IR.

[0026] Further extraction of data features from different windows, specifically including: 1) Analyze short-term, high-frequency data from the proximal window to assess the risk characteristics of early progression of IR in current patients: a) In the early stages of insulin resistance (IR), although insulin can be secreted, its effect is weakened, resulting in a higher postprandial blood glucose peak and a delayed recovery, leading to increased volatility. Intraday CGM parameters (postprandial peak, mean amplitude, coefficient of variation) were extracted. b) Under normal conditions, the insulin peak should occur slightly earlier than the blood glucose peak to clear glucose in a timely manner; in IR (insulin-induced insulin regurgitation), a delayed peak or excessively high concentration indicates compensatory oversecretion. Calculate the offset of the postprandial insulin peak time and concentration relative to the blood glucose peak. c) Autonomic nervous system disorders are closely related to decreased insulin sensitivity, and abnormal HRV and irregular sleep are accompanying signals in the early stages of insulin resistance (IR). Sleep cycle integrity and the low-frequency / high-frequency ratio of HRV were extracted by the proportion of deep sleep duration.

[0027] 2) For biochemical indicator data in the mid-range window: a) For time series of biochemical indicators such as HbA1c and GA, data trends are more indicative of disease progression than isolated data anomalies. Calculate the slope of these data at weekly time points to represent the trend direction of the biochemical indicator.

[0028] 3) For clinical diagnostic information data in the remote window: a) Some diagnoses indicate that IR has entered the clinically visible stage, which is a milestone event that accelerates the risk of IR. Integrate relevant information on IR development from clinical diagnoses, such as IGT, OSA, abdominal obesity, etc., and record the diagnosis time.

[0029] (b) Both chronic sleep deprivation and progressive obesity can become chronic drivers of sleep debt (IR). The sum of the differences between the National Health Commission's recommended sleep duration and the patient's actual daily sleep duration within the distal window is calculated as the patient's current sleep debt; the average BMI of the patient within the distal window is also calculated.

[0030] It should be noted that the data collection methods are based on existing technologies, and the length of various time windows can be adjusted by the implementer, so it will not be elaborated further.

[0031] The disease course analysis module 102 is used to obtain the proximal IR coefficient based on the fluctuations of blood glucose, insulin, and sleep time in the proximal window within the current distal window, combined with abnormal heart rate manifestations; it fits the slope fluctuation curves of biochemical data of each dimension in different intermediate windows, and obtains the disease course stage parameters of the current distal window based on the changes in the fluctuation curves, combined with the changes in BMI value distribution, sleep deprivation manifestations, and clinical diagnosis records within the current distal window, as well as the distribution of the proximal IR coefficient.

[0032] As known from existing technologies, the development of insulin resistance (IR) is not an instantaneous event, but a dynamic, cumulative, and phased evolutionary process. The dynamic evolution of IR exhibits clear phase divisions, with different core pathological mechanisms and significantly different key data types with high indicative ability in each phase: In the early stage, the compensatory function of insulin secretion can maintain normal blood glucose, and abnormalities are mainly manifested as delayed or excessively high postprandial insulin peaks; in the middle stage, diurnal blood glucose fluctuations, sleep rhythm disorders, and lipid metabolism abnormalities gradually appear; in the late stage, it manifests as persistent hyperglycemia, long-term elevation of HbA1c, and decline in β-cell function.

[0033] In order to integrate near-term short-term dynamic signals, mid-term steady-state trends and far-term cumulative effects, and to perform time allocation and modal attention weighting analysis based on disease stages, so as to obtain more accurate IR risk assessment in the later stage, it is necessary to analyze data at different time levels.

[0034] In the early stages of insulin resistance (IR) development, the core physiological factor is decreased insulin sensitivity in peripheral tissues (muscle and fat), which is compensated for by excessive insulin secretion from pancreatic β-cells. At this time, the body can still maintain mid-range biochemical homeostasis through β-cell compensation, and there are usually no clinical abnormalities in distal regions. However, under postprandial glucose load, insulin cannot promote glucose uptake in a timely manner, which easily leads to an increased postprandial peak CGM and a delayed peak insulin level. Meanwhile, the autonomic nervous system regulates metabolic abnormalities, and sympathetic nerve excitation leads to abnormal heart rate, which is also an effective early warning sign of IR. Patients' sleep conditions also vary in different states, and fluctuations in sleep onset time reflect changes in the patient's biochemical status. Therefore, based on the fluctuations in blood glucose, insulin, and sleep onset time within the current distal window, combined with abnormal heart rate manifestations, the proximal IR coefficient is obtained as an indicator of the early development characteristic coefficient of IR, providing a basis for subsequent assessment of the disease stage parameters of the distal window and analysis of attention weights.

[0035] Preferably, in the first embodiment of the present invention, please refer to Figure 2 The flowchart illustrates a method for obtaining near-end IR coefficients according to an embodiment of the present invention, specifically including: Step S201: Obtain short-term offset parameters based on the offset between blood glucose peak and insulin peak within each proximal window; extract the sympathetic-parasympathetic balance index of heart rate variability for each proximal window; and obtain the evaluation feature index based on the differences between different proximal windows within the current distal window in three dimensions: blood glucose peak, sleep onset time, and mean blood glucose.

[0036] Considering that under normal physiological conditions, the rise in blood glucose is quickly contained, and the insulin peak is almost synchronous with or slightly ahead of the blood glucose peak, the two are closely coupled; while in the early stage of insulin resistance (IR), the deviation between the blood glucose peak and the insulin peak reflects the decrease in insulin sensitivity and the delay in the insulin peak, so this is used to obtain short-term deviation parameters. As an example, the time interval between the time points of blood glucose peaks and insulin peaks within each proximal window is used as a short-term offset parameter. When there is data from multiple diets within the proximal window, the average of these multiple diets is taken.

[0037] Considering that the autonomic nervous system regulates metabolic abnormalities, and that sympathetic nerve excitation leads to an increase in the sympathetic-parasympathetic balance index, i.e. an increase in the low-frequency / high-frequency ratio of HRV, the sympathetic-parasympathetic balance index is extracted to reflect abnormal heart rate manifestations.

[0038] Considering the differences in patients' eating and sleeping conditions under different conditions: outlier eating and sleeping times, or abnormal data peaks, may indicate that the patient's life status is relatively special in the proximal window. In this case, the data at the time series point has low reference value for dynamic modeling of IR development. Therefore, we compare different proximal windows to obtain the evaluation characteristic index.

[0039] As an example, considering that the three dimensions of peak blood glucose, time of falling asleep, and mean blood glucose have different dimensions, in order to integrate the differences of the three dimensions, the differences of each dimension are linearly normalized in the corresponding data dimension, so as to integrate them. Therefore, for any two proximal windows, after linearly normalizing the absolute values ​​of the differences of the three dimensions of peak blood glucose, time of falling asleep, and mean blood glucose, the product of the normalization results of the three dimensions is used as the initial feature index, which comprehensively reflects the fluctuations of blood glucose, insulin, and time of falling asleep. At this time, each proximal window within the distal window can be matched with multiple other proximal windows, corresponding to multiple initial characteristic indices. If the sum of the multiple initial characteristic indices corresponding to the proximal window is larger, it indicates that the patient's life status in the proximal window is more special. At this time, the reference value of the short-term offset parameter and the sympathetic-parasympathetic balance index is lower, providing a basis for adjustment to obtain the assessment characteristic index. Based on this, the sum of all initial feature indices of each near window within the far window is normalized to obtain the evaluation feature index of each near window.

[0040] Specifically, the sum of all initial feature indices for each near-end window is first obtained, and these sums are set together. The elements in the set are then normalized using the softmax function, and the normalization result is used as the evaluation feature index for each near-end window.

[0041] Step S202: In this step, the short-term offset parameters, sympathetic-parasympathetic balance index, and all corresponding evaluation feature indices of each proximal window are fused to obtain the proximal IR coefficient of each proximal window.

[0042] Considering that a larger short-term offset parameter in the proximal window indicates a more significant decrease in insulin sensitivity and a more pronounced risk of early-stage insulin resistance (IR), and that a larger sympathetic-parasympathetic balance index indicates a higher probability of neurometabolic abnormalities in the patient, a more pronounced risk of early-stage IR, and that a larger assessment characteristic index indicates lower reference value, the proximal IR coefficient is obtained by fusing these parameters.

[0043] As an example, the product of the short-term offset parameter of each proximal window and the sympathetic-parasympathetic balance index is used as the numerator, and the evaluation feature index is used as the denominator. After the fractional ratio is linearly normalized in the corresponding data dimension, the normalized result is used as the proximal IR coefficient of the corresponding proximal window.

[0044] The softmax function outputs a positive number, so the denominator is not zero. In other embodiments of this invention, if linear normalization is used when normalizing the sum of all initial feature indices for each near-end window, a preset positive parameter (divisible by zero) can be added to the denominator, such as... The dimension of the preset division-to-zero positive parameter is the same as that of the denominator to prevent the denominator from being zero; the division-to-zero operation used on the denominator in the embodiments of the present invention can all adopt this method.

[0045] In another embodiment of the present invention, considering that the time interval between the time points of multiple blood glucose peaks and the time points of insulin peaks throughout the day, as well as the sympathetic-parasympathetic balance index, are of greater concern and sensitivity to abnormal situations, the time interval and the sympathetic-parasympathetic balance index can be respectively formed into corresponding sets. The elements in the sets are normalized by the softmax function to obtain weighted weights. The elements in these two dimensions are weighted and summed respectively using a weighted summation method. The summation results are used as the short-term offset parameter and the sympathetic-parasympathetic balance index, respectively.

[0046] It should be noted that, unless otherwise specified, linear normalization is performed on the data dimensions consisting of all similar data from all patients. The maximum and minimum values ​​are selected in each data dimension for linear normalization, which is a well-known technique and will not be elaborated further.

[0047] When the compensatory capacity of β cells drops sharply, and they move from partial compensation to compensatory failure, the risk of IR may accelerate. This is reflected in the data, where biochemical indicators within the mid-range window often show corresponding trend fluctuations. Therefore, fitting the slope fluctuation curves of biochemical data of various dimensions in different mid-range windows makes it easier to analyze the trend fluctuations of biochemical data in different distant windows based on the changes in the fluctuation curves. Meanwhile, if there is no clinical diagnosis in the previous remote window, and the current remote window adds a diagnostic event of IGT (clinical stage) or OSA (failure stage), the corresponding physiological change is that IR progresses from metabolic abnormality to organ damage. At this time, the remote self-diagnostic features of the current remote window need to be significantly weighted. Therefore, the addition of a diagnosis is a sign of the jump in IR risk. The disease stage parameters are analyzed in combination with the change process of the clinical diagnosis record. In addition to direct examination results, patients' own unhealthy lifestyle habits may also play a role in promoting the evolution of IR: insufficient sleep activates the sympathetic nervous system, and obesity aggravates fat inflammation. These lifestyle habits may promote the evolution of IR. Therefore, we also combined the current distribution of BMI values ​​within the distal window and the manifestations of insufficient sleep for analysis. In addition, the distribution of proximal IR coefficients in the proximal window included in the current distal window can reflect the overall trend of changes in insulin sensitivity and metabolic homeostasis in the body at this stage, reflecting the progression status and risk level of the disease at the current stage. Therefore, the distribution of proximal IR coefficients is also combined to obtain the disease stage parameters of the current distal window, which are used to characterize the stage status of insulin resistance development and provide a basis for subsequent attention weight allocation.

[0048] Preferably, in one embodiment of the present invention, please refer to Figure 3 The flowchart illustrates a method for obtaining disease stage parameters according to an embodiment of the present invention, specifically including: Step S211: Based on the differences in the fluctuation curves of the mid-range windows of the same order in adjacent far-end windows in each dimension, and combined with the biochemical data values ​​of the corresponding dimensions, obtain the biochemical indicator IR stage index.

[0049] By comparing the biochemical data fluctuation curves of intermediate windows with the same sequence in adjacent distal windows, we can analyze their changing trends and differences, capture the changes in biochemical signals of insulin resistance from compensation to decompensation, and identify latent progression trends. In addition, by combining the biochemical data values ​​of various dimensions, we can reflect the physiological response characteristics of the disease at different stages, so as to obtain the biochemical indicator IR stage index.

[0050] As an example, biochemical data for each dimension is selected as the target data and analyzed dimension by dimension. The slope of the target data at each time point of the week is extracted to obtain the average slope of each mid-range window. The average slope is then fitted into a fluctuation curve of the slope of the target data. At this time, the slope on the fluctuation curve can be regarded as the "acceleration" of the target data at each mid-range window, reflecting the stability of IR development. Therefore, the derivative of the fluctuation curve at the corresponding time position of each mid-range window is obtained as the fluctuation stability index of the target data at the corresponding time point. The mid-range windows are sorted in chronological order within the far-end windows. If the volatility stability index of the mid-range windows in the current far-end window is larger than that of the mid-range windows in the same order in the previous far-end window, it indicates that the trend itself is very unstable and may be in a state of violent fluctuation, turning point or noise interference. It cannot stably reflect the progress of IR, and the reference weight is smaller. Therefore, the reference weight of all mid-range windows in the current far-end window is obtained based on the difference in volatility stability index of the mid-range windows in the same order in adjacent far-end windows. Specifically, the volatility stability index of each intermediate window within the current far window is calculated, and the difference is obtained by subtracting the volatility stability index of the intermediate window with the same order corresponding to the previous far window. The difference is used to represent the change in the volatility curve. The difference is then inverted and normalized using the softmax function, and used as the reference weight for each intermediate window within the current far window. The dataset corresponding to the softmax function is the inverted result of the difference in volatility stability indices of all intermediate windows. Further, based on reference weights, target data from all mid-range windows within the current remote window are integrated to form a biochemical stage index; Specifically, the mean of all target data within the mid-range window is used to represent its overall data characteristics. The mean of all target data within the mid-range window in the far-end window is weighted and summed using reference weights. The summation result is used as the biochemical stage index.

[0051] Finally, since there are multiple dimensions of biochemical data, the mean of the biochemical stage indices of all dimensions is used as the biochemical indicator IR stage index.

[0052] It should be noted that when the system is in the first evaluation cycle and there are no adjacent historical remote windows, the system remains silent and accumulates data. In other embodiments of the present invention, the implementer may also set weights for each biochemical data dimension and perform fusion by weighted summation. Alternatively, for the first remote window, the difference in its fluctuation stability index may be set to a preset value (set by a professional physician), and all extracted clinical diagnosis records may be used as the number of new symptoms in the current remote window.

[0053] Step S212: Obtain the remote IR coefficient based on the changes in BMI value distribution, sleep deprivation manifestations, and clinical diagnostic records within the current remote window.

[0054] Considering that chronic sleep deprivation and continued obesity can both become chronic drivers of IR, and that changes in clinical diagnostic records can extract events that accelerate IR risk, the distal IR coefficient is obtained in this way.

[0055] As an example, the number of newly added symptoms in the current remote window is first extracted based on the changes in the clinical diagnosis records of adjacent remote windows. The number of newly added symptoms is a direct indicator for quantifying the intensity of "clinical event occurrence / disease progression". The sum of the difference between the preset recommended sleep duration in the remote window and the patient's actual daily sleep duration is taken as the sleep debt, which directly quantifies the chronic burden of long-term sleep deprivation on neuroendocrine, sympathetic activation and metabolic homeostasis. Then, by combining the number of new symptoms, sleep debt, and average BMI, the remote IR coefficient is obtained.

[0056] Among them, the average BMI value represents the distribution characteristics of BMI value, and sleep debt represents the manifestation of sleep deficiency. The preset recommended sleep duration can be assessed and guided by the patient's clinician and is no longer limited. The number of new symptoms represents the changes in clinical diagnosis records. Before fusion, the three types of feature parameters are standardized, such as by linear normalization in the corresponding data dimensions to avoid differences in units. The product of the standardized number of new symptoms, sleep debt and average BMI value is used as the remote IR coefficient.

[0057] Step S213: The negative correlation normalization result of the mean of the proximal IR coefficients of all proximal windows in each intermediate window is used as the IR development regulation factor; based on the IR development regulation factor, the product of the biochemical index IR stage index of each intermediate window in the current distal window and the distal IR coefficient is fused to obtain the IR progression characteristic index of the current distal window.

[0058] In the dynamic modeling of the distal window, the IR features of the proximal window (such as insulin delay and HRV abnormality) have been used to calculate the proximal IR coefficient, which itself represents the early risk indication capability of this stage.

[0059] If these proximal windows are still given higher analytical weights in subsequent analyses (such as at the distal window level or across windows), it will cause early signals to be repeatedly amplified, masking key information in later stages (such as β-cell compensation failure or continued biochemical deterioration). Therefore, the negative correlation normalization result of the mean proximal IR coefficient of all proximal windows in each mid-range window is used as a regulator of IR development.

[0060] As an example, the mean of the proximal IR coefficients of all proximal windows within each mid-range window is inverted to form a dataset. The elements in this dataset are then normalized using the softmax function to obtain the IR development adjustment factor for each mid-range window. The inverse of the value corresponds to a negative correlation logical adjustment.

[0061] Using the IR development regulation factor as the weighting factor, the product of the biochemical index of the IR stage in the mid-range window and the IR coefficient of the distal window is weighted and summed. The weighted sum is used as the IR progression characteristic index of the current distal window.

[0062] The distribution of the proximal IR coefficients is represented by the mean of the proximal IR coefficients across all proximal windows.

[0063] Step S214: Combine the mean of the proximal IR coefficients and the IR progression characteristic index of all proximal windows to obtain the disease stage parameters.

[0064] The mean of the near-end IR coefficient is fused with the IR progression characteristic index to comprehensively reflect short-term risk characteristics and long-term disease progression trends, and output the final disease stage parameters.

[0065] As an example, the mean of the proximal IR coefficients and the sum of the IR progression characteristic index corresponding to the current remote window are linearly normalized in the corresponding data dimension, and the results are used as the disease stage parameters of the current remote window to achieve multi-time level fusion analysis.

[0066] The modeling and analysis module 103 assigns attention weights to each remote window based on the changes in parameters at different disease stages and the time interval between different remote windows and the current remote window; and trains a deep learning model based on these attention weights.

[0067] Considering that changes in disease course parameters reflect the progression rate and metabolic homeostasis fluctuations of insulin resistance at different stages, and can reflect the dynamic evolution trend of IR development, and that the time interval between different distal windows and the current distal window reflects the time correlation and reference value of historical and current states decaying over time, attention weights are assigned to each distal window in combination with these two factors to improve the model's dynamic sensitivity to current risk and prediction accuracy. Finally, deep learning models are trained based on attention weights, which can suppress outdated information interference while maintaining long-term memory, and achieve adaptive modeling of the evolution of insulin resistance risk.

[0068] Preferably, in one embodiment of the present invention, considering that the positive and negative changes of the first-order difference of the disease stage parameters reflect the different trends of IR disease development, the first-order difference sequence of the disease stage parameters is obtained, and the time domain is segmented according to the positive and negative signs of the difference values, and the disease stage parameters corresponding to the continuous difference values ​​with the same sign are grouped into one segment. Based on the difference value of each remote window, and combined with the overall characteristics of the segment length and disease stage parameters, the initial weight of each remote window is obtained.

[0069] Considering that the larger the absolute value of the difference value corresponding to the far window, the more critical it is to be a risk transition point, and the more important the dynamic information of the progression it carries is for assessing future risks, the greater the weight of attention it should receive; when the number of windows (number of parameters of the disease stage) contained in the segment is greater, it indicates that the continuity of IR evolution is stronger, the correlation of disease progression is stronger, and the attention should be greater. Meanwhile, the ultimate importance of a single distal window should consider both its own pathological information density (i.e., disease stage characteristics) and dynamic progression value. Window with high intrinsic value but slow progression should have different weights than window with high intrinsic value and rapid progression. When the mean of all disease stage parameters included in the segment is larger, it indicates a higher degree of overall metabolic imbalance, a more significant decrease in insulin sensitivity, and a more rapid acceleration of IR disease progression. More attention should be paid to the dynamic trend of risk accumulation and disease mutation within this stage (correlation of disease progression). Therefore, the initial weights should be obtained based on this.

[0070] As an example, the absolute value of the difference value of each remote window is used as the disease progression coefficient; the product of the number of segments and the disease progression coefficient of each included remote window is used as the disease progression association coefficient of each remote window. The mean of all disease stage parameters included in the segment is linearly normalized to the corresponding data dimension as the IR acceleration coefficient. The disease progression coefficient is weighted based on the IR acceleration coefficient. The disease progression correlation coefficient is weighted using the difference between constant 1 and the IR acceleration coefficient. The weighted sum is used as the initial weight of the corresponding remote window.

[0071] Furthermore, considering that patients' physiological patterns and disease stages may change over time, the reference value of disease information from historical remote windows further away from the present is lower. Therefore, the initial weights are adjusted based on the time interval between each remote window and the current remote window to obtain the attention weight of each remote window.

[0072] As an example, the negative correlation mapping result between each remote window and the time interval of the current remote window is used as a correction coefficient; Specifically, taking the first moment of each remote window as the benchmark, the time interval between different remote windows is the interval between their respective first moments. The time interval is negatively correlated by passing it through the negative exponential function exp(-x) with the natural constant e as the base, and the mapping result is used as the corresponding correction coefficient. The product of the correction coefficients and the initial weights of each far window is used to form a set. The normalized result of the elements in the set using the softmax function is used as the attention weight of the corresponding far window.

[0073] It should be noted that in the first-order difference sequence, the difference between each disease stage parameter and the parameter of the adjacent previous disease stage is used as the difference value corresponding to each disease stage parameter. The difference is set to 0 and the same as the positive sign, so as to perform time-domain segmentation. The first disease stage parameter is directly assigned to the first segment, and the initial weights of the first two remote windows are the same.

[0074] For example, the sequence of disease stage parameters is [0.2,0.5,0.7,0.6,0.5,0.5,0.8,0.9,0.7,0.6], and the first-order difference sequence is [+0.3,+0.2,-0.1,-0.1,0,+0.3,+0.1,-0.2,-0.1], which are divided into [0.2,0.5,0.7], [0.6,0.5], [0.5,0.8,0.9] and [0.7,0.6].

[0075] It should be noted that if there is only one far window, there is no need to assign attention weights. The analysis process for each far window is the same each time. Only one example is described here.

[0076] Finally, a deep learning model is trained based on attention weights. In one embodiment of the invention, the deep learning model is used to assess the patient's insulin resistance risk: The system employs a three-module deep learning architecture: multi-window feature encoding, attention fusion, and risk prediction. The input layer consists of multimodal features from each remote window (near-end monitoring data, mid-range biochemical slope, and remote diagnostic / behavioral features). The encoding layer uses a lightweight Transformer with positional encoding to independently encode each window feature into a window vector. The attention layer weights and fuses the window vectors with their corresponding attention weights. The output layer outputs an IR risk value (0-1, with higher values ​​indicating higher risk) through a fully connected network. Input feature construction: Concatenate the encoding vectors of each far window in chronological order, with each vector accompanied by its attention weight, to form a sequence of "feature-weight" pairs; Training data preparation: Label the dataset (including window features, weights, and clinical gold standard GIR values), and divide the training / test sets in a 7:3 ratio; Training parameter settings: AdamW optimizer (learning rate 1e-4), MSE (mean squared error of risk value to GIR mapping value) loss function, 100 iterations, early stopping mechanism to prevent overfitting; After training, input the remote window features and weights of new patients, and the model outputs a risk value after encoding and attention fusion, thus completing the IR risk assessment.

[0077] It should be noted that the output IR risk value is for reference only and is intended to assist professionals in analysis or remind users to pay attention to changes in their physiological state.

[0078] In summary, to address the technical problem of existing models' unreasonable attention weight settings for different time periods, making it difficult to accurately reflect the dynamic evolution of insulin resistance, the present invention aims to provide an AI dynamic modeling system for insulin resistance risk assessment. This invention first acquires multimodal data for each window type through a data acquisition module; further, it analyzes fluctuations in blood glucose, insulin, and sleep onset time within the proximal window of the distal window through a disease progression analysis module, combining this with abnormal heart rate manifestations to obtain the proximal IR coefficient; further, based on the changes in the slope fluctuation curves of each biochemical data, combined with changes in BMI value distribution, sleep deprivation manifestations, and clinical diagnostic records within the current distal window, as well as the distribution of the proximal IR coefficient, it obtains the disease progression stage parameters for the current distal window; finally, it assigns attention weights to each distal window through a modeling and analysis module, and trains a deep learning model based on these attention weights. This scheme acquires multimodal data through multiple time windows, extracts proximal IR features, and combines mid-term biochemical slope fluctuations with distal BMI, sleep, and diagnostic information to generate disease progression stage parameters; it dynamically assigns attention weights based on parameter changes, improving the model's dynamic sensitivity and prediction accuracy to current risks.

Claims

1. An AI dynamic modeling system for assessing insulin resistance risk, characterized in that, The system includes: The data acquisition module is used to divide the time hierarchy into preset near-end, mid-range, and far-end windows and acquire multimodal data for each type of window; The disease course analysis module is used to obtain the proximal IR coefficient based on the fluctuations of blood glucose, insulin, and sleep time in the proximal window within the current distal window, combined with abnormal heart rate manifestations; it also fits the slope fluctuation curves of biochemical data of each dimension in different intermediate windows, and obtains the disease course stage parameters of the current distal window based on the changes in the fluctuation curves, combined with the changes in BMI value distribution, sleep deprivation manifestations, and clinical diagnosis records within the current distal window, as well as the distribution of the proximal IR coefficient; The modeling and analysis module assigns attention weights to each remote window based on the changes in the parameters of the disease stage and the time interval between different remote windows and the current remote window; and trains a deep learning model based on the attention weights.

2. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 1, characterized in that, The method for obtaining the near-end IR coefficient includes: Based on the offset between the peak blood glucose and peak insulin levels within each proximal window, short-term offset parameters are obtained; the sympathetic-parasympathetic balance index of heart rate variability in each proximal window is extracted; and the evaluation feature index is obtained based on the differences between different proximal windows within the current distal window in three dimensions: peak blood glucose, time of sleep onset, and mean blood glucose. The proximal IR coefficient of each proximal window is obtained by fusing the short-term offset parameter, the sympathetic-parasympathetic balance index, and all the corresponding evaluation feature indices for each proximal window.

3. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 2, characterized in that, The method for obtaining the short-term offset parameter includes: The time interval between the time point of the blood glucose peak and the time point of the insulin peak within each proximal window is obtained as a short-term offset parameter.

4. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 2, characterized in that, The method for obtaining the evaluation feature index includes: For any two proximal windows, the absolute values ​​of the differences in the three dimensions of blood glucose peak, sleep time, and mean blood glucose are normalized respectively, and the product of the normalized results of the three dimensions is used as the initial feature index; the sum of all the initial feature indices of each proximal window in the distal window is normalized to obtain the evaluation feature index of each proximal window.

5. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 1, characterized in that, The methods for obtaining the disease stage parameters include: Based on the differences in the fluctuation curves of the mid-range windows in each dimension of the adjacent far-end windows in the same order, and combined with the biochemical data values ​​of the corresponding dimensions, the biochemical indicator IR stage index is obtained. Based on the changes in BMI value distribution, sleep deprivation manifestations, and clinical diagnostic records within the current remote window, obtain the remote IR coefficient; The negative correlation normalization result of the mean of the proximal IR coefficients of all proximal windows in each mid-range window is used as the IR development regulation factor; based on the IR development regulation factor, the product of the biochemical index IR stage index and the distal IR coefficient of each mid-range window in the current distal window is fused to obtain the IR progression characteristic index of the current distal window. By fusing the mean of the proximal IR coefficients of all the proximal windows and the IR progression characteristic index, the disease stage parameters are obtained.

6. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 5, characterized in that, The method for obtaining the biochemical indicator IR stage index includes: Biochemical data for each dimension is selected as target data, and the derivative of the time series position corresponding to each mid-range window on the fluctuation curve is obtained as the fluctuation stability index of the target data at the corresponding time point. Based on the difference in the fluctuation stability index of the intermediate windows in the same order of adjacent far windows, obtain the reference weight of all intermediate windows in the current far window; based on the reference weight, fuse the target data of all intermediate windows in the current far window as the biochemical stage index. The mean of the biochemical stage indices across all dimensions is used as the biochemical indicator IR stage index.

7. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 5, characterized in that, The method for obtaining the far-end IR coefficient includes: Based on the changes in the clinical diagnosis records of adjacent remote windows, the number of newly added symptoms in the current remote window is extracted; the sum of the differences between the preset recommended sleep duration and the patient's actual daily sleep duration in the remote window is taken as sleep debt; the number of newly added symptoms, the sleep debt, and the average BMI value are fused to obtain the remote IR coefficient.

8. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 1, characterized in that, The method for obtaining the attention weights includes: Obtain the first-order difference sequence of the disease stage parameters, segment the time domain according to the positive or negative sign of the difference values, and obtain the initial weight of each remote window based on the difference value of each remote window, combined with the length of the segment and the overall characteristics of the disease stage parameters. The initial weights are adjusted based on the time interval between each of the remote windows and the current remote window to obtain the attention weights for each remote window.

9. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 8, characterized in that, The method for obtaining the initial weights includes: The absolute value of the difference value of each of the distal windows is used as the disease progression coefficient; the product of the number of segmented windows and the disease progression coefficient of each of the distal windows is used as the disease progression correlation coefficient of each of the distal windows. The normalized result of the mean of all the disease stage parameters included in the segment is used as the IR acceleration coefficient. The disease progression coefficient is weighted based on the IR acceleration coefficient. The disease progression correlation coefficient is weighted using the difference between the constant 1 and the IR acceleration coefficient. The weighted sum is used as the initial weight of the corresponding distal window.

10. The AI ​​dynamic modeling system for assessing insulin resistance risk according to claim 8, characterized in that, The method for obtaining the attention weight of each remote window by correcting the initial weight based on the time interval between each remote window and the current remote window includes: The negative correlation mapping result between each of the far windows and the current far window is used as a correction coefficient, and the normalized result of the product of the correction coefficient of each far window and the initial weight is used as the attention weight.

Images