Human islet function digital evaluation method and system based on AI algorithm

Abstract

The invention discloses a human pancreas islet function digital evaluation method and system based on an AI algorithm, and relates to the technical field of artificial intelligence medical evaluation, and the method comprises the steps: collecting and preprocessing multi-modal physiological data; extracting multi-source features to generate a low-dimensional input vector; performing dynamic weighted fusion based on a space-time attention mechanism; constructing a neural differential equation model in which a beta cell two-phase secretion mechanism is introduced, and outputting pancreas islet function evaluation indexes; and a dynamic evaluation report is generated through a visualization and early warning module. According to the method, the modeling accuracy of the pancreas islet function under the interference of complex behaviors can be improved, the individualized quantitative evaluation of function decline trend prediction and sensitivity scoring is realized, the clinical staging reference value and the model interpretation ability are enhanced, and the technical problems of low fusion precision, poor dynamic nature and lack of physiological constraints in the existing method are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence medical evaluation, and specifically provides a digital evaluation method and system for human islet function based on AI algorithms. Background Art

[0002] In recent years, with the rapid development of artificial intelligence and wearable device technologies, human health monitoring has gradually shifted from traditional static detection modes to dynamic continuous monitoring. Multimodal physiological sensing devices such as continuous glucose monitoring (CGM), smart bracelets, and mobile electrocardiographs have become increasingly popular, making it possible to obtain large-scale, multi-source, and time-series individual health data. On this basis, the fusion of multi-source time-series physiological signals for individualized health status assessment has gradually become a research hotspot. Especially in the field of chronic disease management such as early diabetes screening, disease course assessment, and islet function monitoring, intelligent assessment technologies are gradually replacing traditional methods that rely on single-point detection and manual interpretation. At the same time, the application of AI methods such as neural networks, attention mechanisms, and differential modeling in biomedical time-series modeling has become increasingly widespread, providing a technical foundation for establishing more accurate, dynamic, and interpretable physiological state assessment systems.

[0003] However, most existing islet function assessment technologies still perform static assessments based on single fasting blood glucose, postprandial blood glucose, or HbA1c indicators, ignoring the complex correlations between an individual's multi-period dynamic behaviors, physiological states, and islet responses. Although some studies have attempted to introduce CGM data, exercise data, etc. for joint analysis, there are still obvious deficiencies in multi-modal data alignment, feature fusion, and physiological interpretability modeling. For example, traditional assessment models often use static rules or simple linear combinations and are difficult to effectively handle the sampling frequency differences and data noise interference between different modalities; the feature fusion lacks a spatio-temporal weight adjustment mechanism, resulting in the model being slow to respond to key behavioral events; more critically, mainstream models cannot effectively embed the islet β-cell secretion mechanism, resulting in the output indicators lacking clear physiological correspondence relationships. The method proposed in the present invention simultaneously achieves high model fitting and physiological interpretability in the dynamic modeling process by introducing neural differential equations and two-phase insulin kinetics constraints; at the same time, it combines a spatio-temporal attention mechanism for adaptive feature weighting, effectively enhancing the robustness of the model under different behavioral scenarios and data quality conditions, and solving the problem that existing technologies cannot achieve both multi-modal data fusion accuracy and clinical interpretability. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by the present invention are as follows: the existing islet function evaluation methods have limited static index information, lack of multimodal data fusion mechanism, lack of physiological interpretability of model results, and the problem of how to achieve personalized digital evaluation of islet function for multimodal data, dynamic behavior perception, and interpretable output.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, an embodiment of the present invention provides a method for digital evaluation of human islet function based on AI algorithm, including collecting multimodal physiological data and performing synchronous preprocessing to generate a preprocessed multimodal physiological data set; Performing feature extraction based on the preprocessed multimodal physiological data set to generate a low-dimensional input feature vector; Performing spatio-temporal attention dynamic weighted fusion based on the low-dimensional input feature vector to generate a dynamically weighted fusion feature vector; Constructing a hybrid evaluation model driven by neural differential equations based on the dynamically weighted fusion feature vector to generate islet function evaluation indicators; Performing dynamic visualization and anomaly warning processing based on the islet function evaluation indicators to generate a dynamic visualization evaluation report.

[0007] As a preferred scheme of the method for digital evaluation of human islet function based on AI algorithm of the present invention, wherein: the collection of multimodal physiological data includes collecting glucose signal data of subcutaneous interstitial fluid through a continuous glucose monitoring system; Collecting high-frequency physiological signal data through intelligent wearable devices, and the high-frequency physiological signal data includes triaxial acceleration data, heart rate data, and status marker data; Collecting biochemical index data through a laboratory, and the biochemical index data includes fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide level.

[0008] As a preferred scheme of the method for digital evaluation of human islet function based on AI algorithm of the present invention, wherein: the generation of the aligned multimodal physiological data set includes using the time series of glucose signal data as the main reference time axis; The high-frequency physiological signal data is subjected to time repositioning and interpolation processing based on the reference time axis using a non-linear dynamic time alignment algorithm, and the aligned high-frequency physiological signal data is output; The biochemical index data is subjected to piecewise linear interpolation and spline fitting based on the reference time axis, and the aligned biochemical index data is output; Perform unified time mapping, synchronous denoising, and sliding window smoothing on the aligned high-frequency physiological signal data, aligned biochemical index data, and glucose signal data corresponding to the reference time axis to obtain preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data; Output the preprocessed multimodal physiological dataset, where the preprocessed multimodal physiological dataset includes preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data.

[0009] As a preferred solution of the method for digital evaluation of human islet function based on AI algorithm according to the present invention, wherein: the generation of the low-dimensional input feature vector includes performing local extreme value detection on the preprocessed glucose signal data based on the preprocessed multimodal physiological dataset, and outputting a peak timing feature set; Perform wavelet decomposition on the preprocessed glucose signal data, and output a multi-scale frequency domain energy feature set for each decomposition layer; Perform numerical integration on the preprocessed glucose signal data within a preset time window, and output a segmented curve area feature set; Perform amplitude integration on the triaxial acceleration in the preprocessed high-frequency physiological signal data, and estimate the exercise energy consumption in combination with the individual body weight parameter, and output an exercise energy feature set; Use the status markers in the preprocessed high-frequency physiological signal data to distinguish between resting and active sections, and output a segmented exercise state feature set; Perform standardization processing on the preprocessed biochemical index data, and output a standardized biochemical feature set; Fuse the peak timing feature set, multi-scale frequency domain energy feature set, segmented curve area feature set, exercise energy feature set, segmented exercise state feature set, and standardized biochemical feature set, and perform dimensionality reduction to generate a low-dimensional input feature vector.

[0010] As a preferred solution of the method for digital evaluation of human islet function based on AI algorithm according to the present invention, wherein: the generation of the dynamic weighted fusion feature vector includes sending the low-dimensional input feature vector into a long short-term memory network to encode the time dimension, and outputting a time encoding matrix; Based on the low-dimensional input feature vector, output a feature dimension attention weight matrix through a feature attention network; Dynamically adjust the weight coefficients in the aforementioned time encoding matrix and feature dimension weight matrix according to the current physiological scenario and data quality status, and enable weight compensation logic for lost or abnormal channels, and output a corrected spatio-temporal weight matrix; Perform an element-wise Hadamard product fusion of the corrected spatio-temporal weight matrix and the low-dimensional input feature vector to generate and output a dynamically weighted fusion feature vector.

[0011] As a preferred embodiment of the AI algorithm-based digital evaluation method for human islet function according to the present invention, wherein: the generation of the islet function evaluation index includes defining a hidden state update function of a neural differential equation model based on the dynamically weighted fusion feature vector, and outputting an unconstrained ODE model framework; Embed the two-phase secretion kinetics equation of islet β cells as a physiological constraint term into the ODE model framework, where the first-phase equation describes the immediate secretion response of insulin, the second-phase equation describes the cumulative and continuous response of secretion, and dynamically adjusts the secretion rate in combination with the motion energy feature in the fusion feature vector, and outputs an ODE model with physiological constraints; Adopt the adjoint sensitivity method to iteratively optimize the network parameters and the weights of the physiological constraint regularization term of the ODE model with physiological constraints, and output an optimized set of model parameters; Use the optimized model parameters to numerically integrate and solve the ODE model with physiological constraints, generate a trajectory of the hidden state evolving over time, and output a hidden state evolution sequence; Map the final state or the states within a sliding window of the hidden state evolution sequence to an insulin sensitivity score and a β-cell function attenuation rate through a decoding network, and use the insulin sensitivity score and the β-cell function attenuation rate as islet function evaluation indexes.

[0012] As a preferred embodiment of the AI algorithm-based digital evaluation method for human islet function according to the present invention, wherein: the generation of the dynamic visualization evaluation report includes generating a multi-dimensional dynamic chart based on the islet function evaluation index and the corresponding time series data, using time, insulin sensitivity score, and β-cell function attenuation rate as coordinates and color mappings, and outputting an initial dynamic visualization chart; Set an abnormal threshold for the islet function index. When any evaluation index exceeds its corresponding threshold, automatically highlight the abnormal interval in the dynamic chart and output a visualization chart with abnormal highlighting; Enable an interactive linkage function for the dynamic chart. When the user clicks or hovers at a certain time point or section, the original signal, eigenvalue, and model prediction information corresponding to that moment are popped up in real time, and an interactive data details view is output; Based on the change trend of the evaluation index, apply a time series prediction method to deduce the islet function index for a period of time in the future, and superimpose the prediction result on the dynamic chart to output an interactive dynamic chart with a future prediction curve; Monitor whether the evaluation indicators at consecutive moments continuously exceed the threshold. When the continuous over-threshold condition is met, trigger the automatic early warning mechanism, generate a risk early warning notice and push it to the user terminal, and output an early warning message; Summarize the dynamic chart with abnormal annotation, interactive linkage and future prediction functions and the early warning notice to form a final dynamic visual evaluation report and save it, and output a complete dynamic visual evaluation report.

[0013] In a second aspect, an embodiment of the present invention provides a digital evaluation system for human islet function based on an AI algorithm, including: Data acquisition and preprocessing module: Collect multimodal physiological data and perform synchronous preprocessing to generate a preprocessed multimodal physiological data set; Feature extraction module: Extract features based on the preprocessed multimodal physiological data set to generate a low-dimensional input feature vector; Fusion module: Perform spatio-temporal attention dynamic weighted fusion based on the low-dimensional input feature vector to generate a dynamically weighted fusion feature vector; Index generation module: Construct a hybrid evaluation model driven by a neural differential equation based on the dynamically weighted fusion feature vector to generate an islet function evaluation index; Visualization and anomaly early warning module: Perform dynamic visualization and anomaly early warning processing based on the islet function evaluation index to generate a dynamic visual evaluation report.

[0014] Advantages of the present invention: By fusing continuous blood glucose monitoring data, high-frequency physiological signals collected by wearable devices, and laboratory biochemical indicators, the present invention realizes multimodal dynamic modeling of human islet function; by introducing a spatio-temporal attention mechanism and a neural differential equation model, the accuracy of feature fusion and the model's response ability to physiological changes are effectively improved; further combining the biphasic secretion mechanism of pancreatic islet β cells as a physiological constraint enhances the clinical interpretability and prediction reliability of the evaluation results. Compared with traditional methods, the present invention has significantly improved accuracy, dynamics and individual evaluation capabilities, and is applicable to scenarios such as diabetes risk early warning, disease course evaluation and personalized intervention recommendation. Brief Description of the Drawings

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings, where: Figure 1 It is the overall flowchart of the digital evaluation method for human islet function based on the AI algorithm provided by the first embodiment of the present invention. Detailed Embodiments

[0016] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0017] Example 1, referring to Figure 1 , which is an embodiment of the present invention, provides a digital evaluation method for human islet function based on AI algorithms, including: S1: Collect multimodal physiological data and perform synchronous preprocessing to generate a preprocessed multimodal physiological data set.

[0018] Collect glucose signal data from interstitial fluid in the subcutaneous tissue through a continuous glucose monitoring system; Collect high-frequency physiological signal data through intelligent wearable devices, and the high-frequency physiological signal data includes triaxial acceleration data, heart rate data, and status marker data; Collect biochemical index data through a laboratory, and the biochemical index data includes fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide level.

[0019] Take the time series of the glucose signal data as the main reference time axis; The high-frequency physiological signal data is subjected to time repositioning and interpolation processing based on the reference time axis using a non-linear dynamic time alignment algorithm, and the aligned high-frequency physiological signal data is output; The biochemical index data is subjected to piecewise linear interpolation and spline fitting based on the reference time axis, and the aligned biochemical index data is output; Perform unified time mapping, synchronous denoising, and sliding window smoothing processing on the aligned high-frequency physiological signal data, the aligned biochemical index data, and the glucose signal data corresponding to the reference time axis to obtain preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data; Output the preprocessed multimodal physiological data set, and the preprocessed multimodal physiological data set includes preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data.

[0020] In an embodiment of the present invention, in this step, multi-source physiological signal data is obtained through a method combining continuous blood glucose monitoring, smart wearable devices, and laboratory tests, and synchronous preprocessing is performed on a unified time axis to ensure the consistency of the time correspondence relationship between the data. Specifically, first, subcutaneous tissue glucose signals are collected in real time through a continuous blood glucose monitoring system. Taking a subcutaneous implantable sensor (such as the Dexcom G6 continuous blood glucose monitoring device) as an example, the sensor collects blood glucose concentration data at fixed time intervals (such as once every 5 minutes). The original blood glucose signal is processed by algorithms such as Kalman filtering to eliminate high-frequency random noise, obtaining a smooth time series data sequence of blood glucose concentration; at the same time, the timestamp of each blood glucose sampling is recorded to form a time series record of the blood glucose monitoring device.

[0021] Secondly, high-frequency physiological signals are collected through smart wearable devices, including three-axis acceleration, heart rate, and related status markers. Taking a smart watch (such as the Apple Watch Series 8) as an example, its built-in three-axis accelerometer and photoelectric sensor can record the motion acceleration signal and heart rate signal respectively. The device continuously records the motion intensity sequence and heart rate sequence of the wearer at a high frequency (such as the 1Hz normal mode), and can estimate the energy consumption (such as the calories consumed in kcal) in real time through integral operation based on the acceleration signal. The wearable device also synchronously records important status events or markers that occur (such as entering the exercise mode, wearing state changes, etc.), and attaches the device's own timestamp to each record, thus obtaining a multi-dimensional high-frequency physiological signal data sequence including acceleration, heart rate, energy consumption, and status markers.

[0022] Then, low-frequency biochemical index data is collected through laboratory tests regularly and uploaded to the data processing center to enrich the multi-modal data source. Specifically, important biochemical index values of the subject are obtained through laboratory tests of venous blood samples, such as fasting blood glucose, postprandial blood glucose, glycated hemoglobin (HbA1c) percentage, and serum C-peptide level. The precise sampling time is recorded for each laboratory test result (for example, fasting blood glucose and C-peptide are usually collected on an empty stomach in the early morning, postprandial blood glucose is collected at fixed time points after a standard meal or glucose load, and HbA1c is collected monthly or quarterly). These test time points are associated with the corresponding index values as T_lab timestamps and stored to form a structured discrete biochemical index data table.

[0023] To accurately align the data from the above different sources on the time axis, multiple synchronization measures are taken in this step. On the one hand, the clocks of all data acquisition devices are forcibly synchronized at a fixed time every day. For example, at multiple fixed time points every day (such as 7:00, 12:00, 19:00), the clocks of the CGM device, wearable device, and laboratory information system are calibrated to a unified standard time (such as aligning with the standard atomic clock time to ensure the clock error is less than 1 millisecond) through the Network Time Protocol (NTP) to eliminate the clock deviation between devices. On the other hand, taking the time axis of continuous glucose monitoring data as the benchmark, dynamic time alignment processing is performed on signals from other sources. When a significant change trend is detected in the blood glucose monitoring sequence (for example, the blood glucose concentration rises at a rate exceeding 2 mg / dL / min, which is common in the rapid postprandial blood glucose rise stage), the system will automatically trigger the wearable device to enter a higher sampling frequency mode (such as increasing from 1Hz to 10Hz) to capture the details of physiological responses more precisely; at the same time, it can also trigger the sending of an urgent notice to arrange additional laboratory tests (such as completing an HbA1c test in a short time) to obtain the latest biochemical indicators. All the data obtained through the above mechanism uses the time axis of the blood glucose sensor as the reference benchmark. Then, a non-linear dynamic time alignment algorithm is used to perform time repositioning and interpolation processing on the high-frequency signal sequence collected by the wearable device: that is, according to the time benchmark of the blood glucose data, the time axis of the wearable signal is non-linearly stretched or compressed so that its key events are aligned with the blood glucose change moments, and the aligned wearable signal is resampled and interpolated onto the blood glucose time axis. This can effectively eliminate the differences between different sampling frequencies, making high-frequency data such as acceleration and heart rate correspond and match with blood glucose data on the same timeline. For low-frequency discrete biochemical test data, methods such as linear interpolation and spline curve fitting are used for processing: within the time interval between two adjacent laboratory tests, assuming that the indicator changes continuously and smoothly, the gradual transition value of the indicator within the interval is estimated through linear interpolation, or a smooth curve is generated by spline fitting passing through all discrete measurement points to approximately reflect the change trend of the biochemical indicator over time. After the above interpolation and alignment processing, the multi-modal data with different original sampling frequencies are all mapped to a unified time benchmark.

[0024] After alignment on the unified timeline, this step performs preprocessing operations such as synchronous denoising and smoothing on the multimodal data. Specifically, noise reduction and sliding window smoothing are performed on the signals of each channel respectively. For example, for the acceleration and heart rate signals recorded by wearable devices, digital filtering algorithms are applied to remove instantaneous spike noise and outliers, and then the sliding window smoothing method (such as moving average or median filtering of a certain time length) is used to smooth short-term fluctuations and reduce the influence of random noise. For the blood glucose monitoring sequence, on the basis of the aforementioned Kalman filtering, individual abnormal sampling values can be further removed, and the blood glucose curve can be moderately smoothed through a sliding window to eliminate the small residual jitter between samplings. The entire denoising and smoothing process is carried out synchronously, that is, all signals are processed using a consistent time window and step size to ensure that the multiple processed time series are still strictly aligned and no new time offsets are introduced. Through these preprocessing measures, the random errors and noise components in the multi-source data are removed as much as possible, and each signal curve becomes smoother and more reliable.

[0025] Finally, after synchronous alignment, interpolation completion, denoising and smoothing, a preprocessed multimodal physiological data set is formed. This data set consists of multiple types of physiological data sequences, and all sequences are synchronized in time one by one. Specifically, it includes: the time series of blood glucose concentration obtained from continuous blood glucose monitoring (reflecting the continuous change of subcutaneous tissue glucose level); the high-frequency physiological signal sequences recorded by intelligent wearable devices, such as the processed exercise intensity (acceleration) signal, heart rate signal, energy consumption index, and related status marker events; and the continuous time series generated by interpolation / fitting of discrete laboratory biochemical test values, such as the data curves of fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide level changing with time. The preprocessed and integrated multimodal data set has a unified timeline and good data quality, contains complete information such as blood glucose dynamics, exercise physiology, and biochemical indicators of an individual in the same period, and provides sufficient and reliable basic data for subsequent feature extraction, model training, and system evaluation.

[0026] S2: Feature extraction is performed based on the preprocessed multimodal physiological data set to generate a low-dimensional input feature vector.

[0027] Based on the preprocessed multimodal physiological data set, local extreme value detection is performed on the preprocessed glucose signal data, and a set of peak timing features is output; Wavelet decomposition is performed on the preprocessed glucose signal data, and a set of multi-scale frequency domain energy features of each decomposition layer is output; Numerical integration is performed on the preprocessed glucose signal data within a preset time window, and a set of area features under the segmented curve is output; Perform amplitude integration on the three-axis acceleration in the preprocessed high-frequency physiological signal data, estimate the exercise energy consumption in combination with the individual weight parameters, and output the exercise energy feature set; Using the state markers in the preprocessed high-frequency physiological signal data, the rest and activity segments are distinguished, and a segmented motion state feature set is output; Standardize the preprocessed biochemical index data and output a standardized biochemical feature set; The peak time series feature set, multi-scale frequency domain energy feature set, segmented area under the curve feature set, motion energy feature set, segmented motion state feature set and standardized biochemical feature set are fused and dimensionally reduced to generate a low-dimensional input feature vector.

[0028] In the embodiment of the present invention, after completing the collection and preprocessing of multimodal physiological data, it is necessary to further extract effective features from various types of data to fully describe the physiological changes and potential trends of the pancreatic islet functional state and provide input data for subsequent model training.

[0029] First, perform local extreme value detection on the preprocessed glucose signal data, that is, detect and mark the local highest and lowest points on the continuous glucose concentration curve to obtain information about the time, amplitude and frequency of the peak value of blood glucose changes. These local extreme points can reflect the fluctuation characteristics of individual blood glucose levels, especially in the stage of significant changes in blood glucose after meals, exercise-induced changes, or other life events. It has important physiological significance. Extract peak timing features from these extreme points and their time intervals to reflect the frequency and severity of blood glucose fluctuations.

[0030] Secondly, in order to further explore the implicit frequency domain features in the glucose signal, this step uses a multi-scale wavelet decomposition method to perform multi-level decomposition processing on the preprocessed glucose signal data. Wavelet decomposition can convert the glucose signal from a single time domain signal into a frequency domain component with multiple scales, each scale corresponding to a different physiological process or noise component. By analyzing the signal energy distribution at different decomposition scales, a multi-scale frequency domain energy feature set is extracted. These features can reflect the periodic or irregular changes in blood sugar changes, which can help the model identify abnormal patterns more accurately.

[0031] Again, the numerical integration method is used to extract features from the preprocessed glucose signal data. The specific method is: in the pre-set time window, the area under the curve of the blood glucose concentration curve is calculated respectively. This method is similar to the integral curve analysis used clinically to evaluate the overall blood glucose control level. It can capture the overall exposure level and change trend of blood glucose in a specific time period, especially helping to quantify important indicators such as postprandial blood glucose load or nighttime blood glucose control. This forms a segmented area under the curve feature set.

[0032] On the other hand, this step also needs to extract motion-related feature data from the preprocessed high-frequency physiological signals. Specifically: First, perform amplitude integration on the three-axis acceleration signals recorded by the wearable device, and then estimate the total amount of exercise during body movement. Then, according to the weight information of individual subjects, convert the exercise amount data into exercise energy consumption values to intuitively reflect the energy metabolism of the human body during daily activities and exercise, and form a set of exercise energy features. These exercise energy features provide a quantitative basis for subsequent evaluation of the impact of individual exercise on islet function.

[0033] In addition, the status marker data collected by the smart wearable device is also analyzed. The status marker data usually records the body status of the subject in each time period, such as resting, mild activity, moderate activity or vigorous activity, etc. According to these status marker data, the system automatically identifies different physiological activity segments, clearly divides the physiological data into resting state and active state segments, and further calculates the time distribution ratio and continuous duration in each active state to output a set of segmented motion state features. This status marker feature helps to accurately distinguish the differential effects of different activity states on blood glucose and islet function.

[0034] For the biochemical index data collected in the laboratory, in this step, through data standardization processing, the biochemical indexes with different units or different measurement ranges are converted into unified dimensionless values. For example, with reference to the group average value and standard deviation of the subjects, each index data is converted into standardized data so that subsequent feature fusion can be unified among different indexes. This standardization process can effectively eliminate data deviation and analysis error caused by different measurement dimensions, output a set of standardized biochemical features, and significantly improve the data analysis accuracy across indexes and individuals.

[0035] Finally, fuse all the above-mentioned extracted features, including the peak timing features, multi-scale frequency domain energy features, segmented curve area features, exercise energy features, segmented motion state features, and standardized biochemical feature set in the blood glucose signal, through feature fusion processing. In order to effectively reduce data redundancy and further improve the operation efficiency and stability of the model, this step uses dimensionality reduction techniques such as principal component analysis to reduce the dimensionality of the fused feature set in the feature space to obtain a low-dimensional input feature vector with a moderate amount of data and sufficient information retention. This low-dimensional input feature vector accurately represents the core change trend of individual islet function with as few dimensions as possible, ensuring reliable input and calculation efficiency in the subsequent spatio-temporal dynamic modeling stage.

[0036] S3: Perform spatio-temporal attention dynamic weighted fusion based on the low-dimensional input feature vector to generate a dynamically weighted fusion feature vector.

[0037] Feed the low-dimensional input feature vector into a long short-term memory network to encode the time dimension and output a time encoding matrix; Based on the low-dimensional input feature vector, output a feature dimension attention weight matrix through a feature attention network; According to the current physiological scenario and data quality status, dynamically adjust the weight coefficients in the aforementioned time encoding matrix and feature dimension weight matrix, and enable weight compensation logic for lost or abnormal channels, and output a corrected spatio-temporal weight matrix; Perform Hadamard product fusion element-wise on the corrected spatio-temporal weight matrix and the low-dimensional input feature vector to generate and output a dynamically weighted fusion feature vector.

[0038] In the embodiments of the present invention, in order to further improve the information utilization efficiency and feature expression ability of the input feature vector, after obtaining the low-dimensional input feature vector, a spatio-temporal attention dynamic weighting fusion method is adopted to perform dynamic weighting processing on the low-dimensional feature vector from two perspectives of the time dimension and the feature dimension, so as to obtain a fusion feature vector that takes into account the change trend of the time series and the contribution of key features.

[0039] First, input the low-dimensional input feature vector into a long short-term memory network (Long Short-Term Memory, LSTM) in the order of acquisition time for time dimension encoding processing. The LSTM network can effectively capture the long-term dependencies in time series data and is particularly suitable for processing the multi-modal physiological data features in the present invention. This network structure dynamically retains or forgets historical input information through internal memory units and gating mechanisms, and continuously updates the hidden state according to the current input, so as to accurately express the patterns and trends of feature evolution over time. The present invention preferably uses a unidirectional LSTM network to adapt to real-time processing scenarios, and a bidirectional LSTM structure (Bidirectional LSTM, BiLSTM) can also be selected according to actual needs to further improve the expression ability of the forward and backward dependencies of the time series.

[0040] Secondly, the present invention further introduces a feature attention network to evaluate and calculate the contribution degree of each feature dimension to the final evaluation target. In the specific implementation process, the output result of the above LSTM network is input into the feature attention network. First, the overall information of each feature dimension is aggregated respectively, such as using pooling operations or calculating the statistical quantities of feature channels, to compress and extract the importance information of each feature channel. Next, the compressed feature information is non-linearly transformed through a multi-layer perceptron or a fully connected network to generate weight coefficients representing the importance of each feature dimension. After appropriate normalization processing of these weight coefficients, a feature dimension attention weight matrix is finally formed to highlight the importance of key features in the spatial dimension, while suppressing redundant features or noise interference with low contribution to the evaluation target.

[0041] In addition, the present invention further designs a dynamic adjustment mechanism for attention weights. Specifically, in order to more precisely adapt to actual clinical or daily scenarios, the present invention dynamically adjusts the time encoding result and the feature attention weight matrix output by the aforementioned LSTM according to the current physiological scenario and data quality status. The system real-time identifies the current physiological state of the subject, such as resting, exercising, or special physiological states, as well as the monitoring data quality level, such as whether there is device detachment, signal interruption, or obvious interference noise. When a specific scenario or data anomaly is detected, the corresponding weight compensation logic will be enabled. For example, when abnormal or missing heart rate data is found during intense exercise, the system will reduce the weight of this feature channel and initiate data interpolation compensation for adjacent time windows; while when in a resting state and the data quality is good, the system will increase the weight of stability features. Through this dynamic scenario perception and data quality adaptive adjustment mechanism, the present invention significantly improves the robustness and effectiveness of feature fusion, effectively avoiding the adverse effects of abnormal data or non-critical features on the fusion result.

[0042] Finally, using the above-mentioned dynamically adjusted time encoding matrix and feature dimension attention weight matrix, they are fused with the low-dimensional input feature vectors element by element. The specific operation is to perform element-by-element multiplication operations on each feature vector according to the attention weights corresponding to the time points and feature dimensions, so as to highlight important features and significantly weaken the influence of redundant or unimportant features on the fusion result. This element-by-element weighted fusion method enables the generated fusion feature vectors to maintain information richness while effectively reducing the interference and redundancy between features and highlighting the key information with the most evaluation value. The dynamically weighted fusion feature vectors obtained through this fusion strategy are used as the input for the next-step islet function evaluation model of the present invention, ensuring the accuracy, stability, and interpretability of subsequent model modeling and evaluation analysis.

[0043] In summary, the attention fusion strategy proposed by the present invention based on LSTM encoding and feature attention mechanism, combined with dynamic scenario and data quality perception, can more precisely achieve the adaptive fusion of multi-modal physiological data features. Compared with the existing static fusion methods, the present invention has obvious technical advantages in data robustness, model generalization, and evaluation accuracy.

[0044] S4: Construct a hybrid evaluation model driven by a neural differential equation based on the dynamically weighted fusion feature vectors to generate islet function evaluation indicators.

[0045] Define the hidden state update function of the neural differential equation model based on the dynamically weighted fusion feature vectors, and output an unconstrained ODE model framework; Embed the two-phase secretion kinetic equation of pancreatic islet β cells into the ODE model framework as a physiological constraint term, where the first-phase equation describes the immediate secretion response of insulin, the second-phase equation describes the cumulative and sustained response of secretion, and dynamically adjust the secretion rate in combination with the motion energy feature in the fused feature vector, and output an ODE model with physiological constraints; Use the adjoint sensitivity method to iteratively optimize the network parameters of the ODE model with physiological constraints and the weights of the physiological constraint regularization term, and output the optimized model parameter set; Use the optimized model parameters to numerically integrate and solve the ODE model with physiological constraints, generate the trajectory of the hidden state evolving over time, and output the hidden state evolution sequence; Map the final state of the hidden state evolution sequence or the state within the sliding window to the insulin sensitivity score and the β-cell function attenuation rate through the decoding network, and use the insulin sensitivity score and the β-cell function attenuation rate as the pancreatic islet function evaluation indicators.

[0046] In the embodiment of the present invention, after obtaining the dynamically weighted fused feature vector, in order to further accurately describe the dynamic change process of pancreatic islet function and improve the interpretability and prediction accuracy of the model, a hybrid modeling method integrating neural network and physiological mechanism is proposed. The specific implementation process is first to construct a Neural ODE model framework and define the update mechanism of the hidden state evolving over time. The essence of this model structure is to use a neural network to define the evolution mode of the system state in continuous time, and with the help of this continuous representation method, the long-term evolution process of pancreatic islet function can be simulated more precisely.

[0047] To enhance the physiological interpretability of the model, the present invention innovatively embeds the two-phase secretion kinetic mechanism of pancreatic islet β cells into the framework of the neural differential equation as a constraint condition. Among them, the first-phase secretion equation is used to describe the rapid immediate response of insulin secretion at the initial stage of rapid blood glucose change (such as after a meal), while the second-phase secretion equation simulates the continuous cumulative response of insulin over a long time. This phased kinetic mechanism is a widely recognized physiological phenomenon in the medical field. The present invention deeply integrates it with the artificial intelligence model, so that the artificial intelligence model can not only capture the pattern rules driven by data, but also has clear physiological significance. In addition, the present invention further innovatively combines the motion energy consumption feature in the dynamic fusion feature, and dynamically adjusts the insulin secretion rate in real time according to the change of the subject's exercise state, so that the model can adapt to different life behavior scenarios and is closer to the real physiological state.

[0048] To effectively optimize the above hybrid model, the present invention uses the adjoint sensitivity method for iterative optimization of network parameters and physiological constraint terms. This optimization method can not only efficiently calculate the sensitivity information of the model to each parameter, but also has low memory resource requirements, especially suitable for processing large-scale data sets with long time series. Through repeated iterative calculations, this optimization method continuously adjusts the parameters of the model and the weights of physiological constraint terms to achieve an optimal balance state, that is, to ensure the accuracy of data fitting while maintaining the rationality of physiological mechanisms.

[0049] Using the optimized model parameters above, further numerical integration operations are carried out to solve the evolution trajectory of the hidden state over time. This hidden state trajectory details the continuous changes of various aspects of islet function over time and provides a basis for the calculation of the next physiological indicators. The present invention constructs a dedicated decoding network to map the final state of the model hidden state trajectory or the hidden state information within a specific time period to two major indicators with clear clinical significance: insulin sensitivity score and β-cell function attenuation rate. The evaluation indicators obtained through hidden state decoding not only have intuitive clinical interpretability, but also are significantly higher than traditional static evaluation methods in terms of sensitivity and accuracy, thus realizing the precise quantification and dynamic evaluation of islet function status, fully demonstrating the creativity and technical rationality of the present invention.

[0050] S5: Based on the islet function evaluation indicators, perform dynamic visualization and abnormal warning processing to generate a dynamic visualization evaluation report.

[0051] Based on the islet function evaluation indicators and the corresponding time series data, generate a multi-dimensional dynamic chart, use time, insulin sensitivity score, and β-cell function attenuation rate as coordinates and color mappings, and output an initial dynamic visualization chart; Set the abnormal threshold of the islet function index. When any evaluation index exceeds its corresponding threshold, automatically highlight the abnormal interval in the dynamic chart and output a visualization chart with abnormal highlights; Enable the interactive linkage function for the dynamic chart. When the user clicks or hovers at a certain time point or section, the original signal, eigenvalue, and model prediction information corresponding to that moment are popped up in real time, and an interactive data details view is output; Based on the change trend of the evaluation indicators, apply the time series prediction method to deduce the islet function indicators for a period of time in the future, and superimpose the prediction results on the dynamic chart to output an interactive dynamic chart with a future prediction curve; Monitor whether the evaluation indicators at multiple consecutive moments continuously exceed the threshold. When the continuous over-threshold condition is met, trigger an automatic warning mechanism, generate a risk warning notice and push it to the user terminal, and output a warning message; Summarize the dynamic charts with anomaly annotation, interactive linkage, and future prediction functions and warning notifications to form and save the final dynamic visualization assessment report, and output the complete dynamic visualization assessment report.

[0052] In the embodiments of the present invention, to intuitively and timely present the results of islet function assessment and effectively monitor and warn of potential risks, a complete set of dynamic visualization and anomaly warning processing methods is proposed. The specific implementation process is first to design and construct multi-dimensional dynamic charts with clear expression capabilities based on the two assessment indicators of the calculated insulin sensitivity score and β-cell function attenuation rate, combined with their corresponding time series information. These dynamic charts visually display time information, sensitivity scores, and function attenuation rates through different-dimensional coordinates, and map the high and low of the indicators and trend changes through color changes, providing a platform for doctors or subjects to quickly and intuitively observe the dynamic changes of islet function.

[0053] Secondly, the present invention sets clear clinical or research reference thresholds for islet function assessment indicators, such as the abnormal range determined according to medical guidelines, expert consensus, or statistical criteria. When any assessment indicator exceeds these predetermined thresholds, the dynamic chart automatically highlights the corresponding abnormal interval, clearly and prominently indicating the abnormal conditions or potential risk areas existing in the assessment object, so as to facilitate clinical personnel or users to pay attention and intervene in a timely manner.

[0054] To further enhance the interactivity and practicality of the assessment report, the present invention also innovatively introduces the interactive linkage function of the dynamic chart. When the user clicks or stays at any time point or area on the chart, the system will pop up the detailed data information of the corresponding time period in real time, including the original physiological signals, characteristic data, and relevant data predicted by the model. This interactive linkage function helps clinical experts and patients intuitively understand the reasons for the generation of abnormal points and the relevant physiological background information, greatly improving the clinical application value and user experience of the assessment report.

[0055] In addition, to enhance the prediction ability of the present invention and the timeliness of preventing risks, time series prediction methods, such as autoregressive models or long short-term memory neural networks, are further used to predict and deduce the change trends of islet function indicators in a certain future period. The prediction results are superimposed and displayed on the dynamic chart with the historical monitoring data to form a dynamic visualization interface with future trend prediction, effectively assisting users to foresee the change trends of islet function in advance and prepare intervention or treatment plans in advance.

[0056] Finally, the present invention introduces a continuous monitoring and automatic warning mechanism, that is, continuously monitor the islet function indicators. If it is found that the indicators in multiple consecutive time periods exceed the set abnormal threshold, the system will automatically determine it as a high-risk state, and immediately generate a risk warning notice, which is actively pushed to the user or clinical intervention personnel to remind them to take medical or lifestyle intervention measures as early as possible. This intelligent risk identification and active intervention reminder function not only reflects the technical creativity of the present invention in the evaluation process, but also makes the present invention truly possess the practical value of active medical management and personalized health management.

[0057] By integrating the functions such as abnormal annotation, interactive linkage, future trend prediction and risk warning, the complete dynamic visualization evaluation report finally generated by the present invention significantly improves the comprehensibility, reliability and operability of the evaluation results, and significantly enhances the innovation, practicality and clinical application potential of the islet function monitoring and intervention system.

[0058] Example 2 is the second example of the present invention, which is different from the previous example in that: If the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the current technical solution can be embodied in the form of a software product. The current computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. And the foregoing storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks or optical disks and other various media that can store program codes.

[0059] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus or device and execute the instructions), or used in combination with these instruction execution systems, apparatus or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus or device.

[0060] More specific examples (a non-exhaustive list) of computer-readable media include the following: electrical connections (electronic devices) having one or more wirings, portable computer disk cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0061] Example 3, an embodiment of the present invention, provides a digital evaluation system for human islet function based on an AI algorithm, including a data acquisition and preprocessing module, a feature extraction module, a fusion module, an index generation module, and a visualization and anomaly warning module.

[0062] Data acquisition and preprocessing module: Acquire multimodal physiological data and perform synchronous preprocessing to generate a preprocessed multimodal physiological data set. Feature extraction module: Extract features based on the preprocessed multimodal physiological data set to generate a low-dimensional input feature vector. Fusion module: Perform spatio-temporal attention dynamic weighted fusion based on the low-dimensional input feature vector to generate a dynamically weighted fusion feature vector. Index generation module: Construct a hybrid evaluation model driven by a neural differential equation based on the dynamically weighted fusion feature vector to generate an islet function evaluation index. Visualization and anomaly warning module: Perform dynamic visualization and anomaly warning processing based on the islet function evaluation index to generate a dynamic visualization evaluation report.

[0063] Example 4, an embodiment of the present invention, provides a method for digital evaluation of human islet function based on an AI algorithm. To verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / contrast experiments.

[0064] The implementation examples were carried out in professional medical institutions. Seven adults with different health statuses were selected to participate in the study to ensure that the subject sample covered individuals with normal blood glucose, prediabetes, and mild to moderate diabetes patients. Before the experiment, each subject was provided with a continuous glucose monitoring system (such as Dexcom G6) and a smart bracelet with built-in three-axis accelerometer and heart rate monitoring functions (such as Fitbit Charge 5), and the subjects wore them continuously for at least 72 hours. During this period, the subjects maintained their daily life status unchanged, and the continuous blood glucose signal, as well as the real-time three-axis acceleration, heart rate, and activity status markers, were recorded. In addition, during the experiment, professional laboratories were arranged to conduct periodic biochemical tests on the subjects, collecting multiple indicators including fasting blood glucose, 2-hour postprandial blood glucose, glycated hemoglobin (HbA1c), and regularly measuring the C-peptide level to obtain comprehensive biochemical data. Subsequently, based on the time axis of the continuous glucose monitoring system, the high-frequency physiological data collected by the smart bracelet was precisely time-repositioned and interpolated using the non-linear dynamic time alignment method. At the same time, the low-frequency biochemical index data from the laboratory was mapped to the same time axis through linear interpolation and smooth fitting processing to uniformly generate a preprocessed multi-modal physiological data set. Next, feature extraction processing was performed on the above data set, including peak time series feature extraction, wavelet frequency domain energy feature extraction, and curve integral feature extraction of blood glucose data. At the same time, the exercise energy consumption was calculated and the exercise state interval was divided. Combining the standardized processing results of biochemical indicators, a low-dimensional input feature vector was obtained. Further, the features were time-encoded through a long short-term memory network (LSTM), and a spatio-temporal attention network was used to calculate and dynamically adjust the feature weights to generate a dynamically weighted fusion feature vector. Based on this feature vector, a hybrid model of neural differential equations with the constraint of the two-phase secretion mechanism of pancreatic islet β cells was constructed to generate islet function evaluation indicators, including insulin sensitivity score and β cell function attenuation rate. Finally, dynamic visualization and anomaly warning processing were performed on the generated evaluation indicators to form a detailed and interactive dynamic visualization evaluation report. The experimental reference data is shown in Table 1.

[0065] Table 1 Experimental Data Record

[0066] From the comparative analysis of the data in the above table, it can be seen that the method of the present invention shows significant advantages in the accuracy and stability of islet function assessment compared with the traditional method. Specifically, from the perspective of insulin sensitivity score, the sensitivity scores calculated by using the method of the present invention for all subjects are higher than those of the traditional method. This indicates that through the dynamic data fusion and neural differential equation hybrid model of the present invention, the individual insulin sensitivity can be evaluated more accurately, more potential physiological feature changes can be captured, and the assessment deviation caused by static indicators in the traditional method can be avoided. In addition, from the perspective of the decay rate of β-cell function, the decay trend calculated by the method of the present invention for each subject is more obvious and sensitive than that of the traditional method, which is particularly prominent in high-risk individuals such as Subject 2 and Subject 4. This shows that the dynamic weighted fusion mechanism and two-phase physiological constraint model proposed by the present invention can capture the early or latent decline trend of islet function more sensitively, which has important value for clinical warning and early intervention. In addition, the prediction error rate index shows that the error rate of the method of the present invention is generally lower than that of the traditional method. For example, the prediction error of the traditional method for Subject 4 reaches 11.9%, while the present invention reduces it to 4.8%, indicating that the model structure and optimization method of the present invention can greatly improve the prediction accuracy and reliability. This improvement in accuracy is attributed to the dynamic perception and weight adjustment strategy of the data quality of the present invention, as well as the reasonable integration of physiological mechanisms.

Claims

1. A digital evaluation method for human islet function based on AI algorithms, characterized in that, Including: Collecting multimodal physiological data and performing synchronous preprocessing to generate a preprocessed multimodal physiological data set; Performing feature extraction based on the preprocessed multimodal physiological data set to generate a low-dimensional input feature vector; Performing spatio-temporal attention dynamic weighted fusion based on the low-dimensional input feature vector to generate a dynamically weighted fusion feature vector; Constructing a hybrid evaluation model driven by a neural differential equation based on the dynamically weighted fusion feature vector to generate an islet function evaluation index; Performing dynamic visualization and abnormal warning processing based on the islet function evaluation index to generate a dynamic visualization evaluation report.

2. The digital evaluation method for human islet function based on AI algorithm according to claim 1, characterized in that, The collecting of multimodal physiological data includes collecting glucose signal data of subcutaneous interstitial fluid through a continuous glucose monitoring system; High-frequency physiological signal data collected by an intelligent wearable device, where the high-frequency physiological signal data includes triaxial acceleration data, heart rate data, and status marker data; Biochemical index data collected in a laboratory, where the biochemical index data includes fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide level.

3. The digital evaluation method for human islet function based on AI algorithm according to claim 2, characterized in that, The generating of the aligned multimodal physiological data set includes using the time series of glucose signal data as the main reference time axis; Performing time repositioning and interpolation processing on the high-frequency physiological signal data based on the reference time axis using a non-linear dynamic time alignment algorithm, and outputting the aligned high-frequency physiological signal data; Performing piecewise linear interpolation and spline fitting on the biochemical index data based on the reference time axis, and outputting the aligned biochemical index data; Performing unified time mapping, synchronous denoising, and sliding window smoothing processing on the aligned high-frequency physiological signal data, the aligned biochemical index data, and the glucose signal data corresponding to the reference time axis to obtain preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data; Outputting the preprocessed multimodal physiological data set, where the preprocessed multimodal physiological data set includes preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data.

4. The digital evaluation method for human islet function based on AI algorithm according to claim 3, characterized in that The generating of the low-dimensional input feature vector includes performing local extreme value detection on the preprocessed glucose signal data based on the preprocessed multimodal physiological data set, and outputting a peak time series feature set; Performing wavelet decomposition on the preprocessed glucose signal data, and outputting a multi-scale frequency domain energy feature set for each decomposition layer; Performing numerical integration on the preprocessed glucose signal data within a preset time window, and outputting a segmented curve under area feature set; Performing amplitude integration on the triaxial acceleration in the preprocessed high-frequency physiological signal data, and estimating the exercise energy consumption in combination with the individual body weight parameter, and outputting an exercise energy feature set; Using the status markers in the preprocessed high-frequency physiological signal data to distinguish between resting and active sections, and outputting a segmented exercise state feature set; Performing standardization processing on the preprocessed biochemical index data, and outputting a standardized biochemical feature set; Fuse the peak timing feature set, multi-scale frequency domain energy feature set, segmented curve area feature set, motion energy feature set, segmented motion state feature set, and standardized biochemical feature set, and perform dimensionality reduction to generate a low-dimensional input feature vector.

5. The digital evaluation method of human islet function based on AI algorithm according to claim 4, characterized in that The generation of the dynamic weighted fusion feature vector includes feeding the low-dimensional input feature vector into a long short-term memory network to encode the time dimension and output a time encoding matrix; Based on the low-dimensional input feature vector, output a feature dimension attention weight matrix through a feature attention network; According to the current physiological scenario and data quality status, dynamically adjust the weight coefficients in the aforementioned time encoding matrix and feature dimension weight matrix, and enable a weight compensation logic for lost or abnormal channels to output a corrected spatio-temporal weight matrix; Perform Hadamard product fusion of the corrected spatio-temporal weight matrix and the low-dimensional input feature vector element by element to generate and output a dynamic weighted fusion feature vector.

6. The digital evaluation method for human islet function based on AI algorithm according to claim 5, characterized in that, The generation of the islet function evaluation index includes defining a hidden state update function of a neural differential equation model based on the dynamic weighted fusion feature vector and outputting an unconstrained ODE model framework; Embed the islet β-cell biphasic secretion kinetics equation as a physiological constraint term into the ODE model framework, where the first-phase equation describes the immediate secretion response of insulin, the second-phase equation describes the cumulative and sustained response of secretion, and dynamically adjust the secretion rate in combination with the motion energy feature in the fusion feature vector to output an ODE model with physiological constraints; Use the adjoint sensitivity method to iteratively optimize the network parameters and physiological constraint regularization term weights of the ODE model with physiological constraints to output an optimized model parameter set; Use the optimized model parameters to numerically integrate and solve the ODE model with physiological constraints to generate a trajectory of the hidden state evolving over time and output a hidden state evolution sequence; Map the final state or the states within a sliding window of the hidden state evolution sequence to an insulin sensitivity score and a β-cell function decay rate through a decoding network, and use the insulin sensitivity score and the β-cell function decay rate as islet function evaluation indicators.

7. The digital evaluation method for human islet function based on AI algorithm according to claim 6, characterized in that, The generation of the dynamic visualization evaluation report includes generating a multi-dimensional dynamic chart based on the islet function evaluation indicators and the corresponding time series data, using time, insulin sensitivity score, and β-cell function decay rate as coordinates and color mappings to output an initial dynamic visualization chart; Set the abnormal threshold of the islet function index. When any evaluation index exceeds its corresponding threshold, automatically highlight the abnormal interval in the dynamic chart to output a visualization chart with abnormal highlights; Enable an interactive linkage function for the dynamic chart. When the user clicks or hovers at a certain time point or section, the original signal, feature values, and model prediction information corresponding to that moment are popped up in real time to output an interactive data details view; Based on the change trend of the evaluation indicators, apply a time series prediction method to deduce the islet function indicators for a period of time in the future and superimpose the prediction results on the dynamic chart to output an interactive dynamic chart with a future prediction curve; Monitor whether the evaluation indicators at multiple consecutive moments continuously exceed the threshold. When the continuous over-threshold condition is met, trigger the automatic early warning mechanism, generate a risk early warning notice and push it to the user terminal, and output an early warning message; Summarize the dynamic chart with abnormal annotation, interactive linkage and future prediction functions and the early warning notice to form a final dynamic visual evaluation report and save it, and output a complete dynamic visual evaluation report.

8. A digital evaluation system for human islet function based on AI algorithms, which is used to implement the digital evaluation method for human islet function based on AI algorithms as described in any one of claims 1 to 7, characterized in that, Including: Data acquisition and preprocessing module: Collect multimodal physiological data and perform synchronous preprocessing to generate a preprocessed multimodal physiological data set; Feature extraction module: Extract features based on the preprocessed multimodal physiological data set to generate a low-dimensional input feature vector; Fusion module: Perform spatio-temporal attention dynamic weighted fusion based on the low-dimensional input feature vector to generate a dynamically weighted fusion feature vector; Index generation module: Construct a hybrid evaluation model driven by a neural differential equation based on the dynamically weighted fusion feature vector to generate an islet function evaluation index; Visualization and abnormal early warning module: Perform dynamic visualization and abnormal early warning processing based on the islet function evaluation index to generate a dynamic visual evaluation report.

Images