A Digital Assessment Method and System for Human Pancreatic Islet Function Based on AI Algorithms

Abstract

This invention discloses a digital assessment method and system for human pancreatic islet function based on AI algorithms, belonging to the field of artificial intelligence medical assessment technology. The method includes: collecting and preprocessing multimodal physiological data; extracting multi-source features to generate low-dimensional input vectors; performing dynamic weighted fusion based on a spatiotemporal attention mechanism; constructing a neural differential equation model incorporating the β-cell biphasic secretion mechanism to output pancreatic islet function assessment indicators; and generating dynamic assessment reports through visualization and early warning modules. This invention can improve the accuracy of modeling pancreatic islet function under complex behavioral disturbances, achieve individualized quantitative assessment of functional decline trends and sensitivity scores, enhance the reference value of clinical staging and model interpretability, and solve the technical problems of low fusion accuracy, poor dynamism, and lack of physiological constraints in existing methods.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence medical assessment technology, specifically to a digital assessment method and system for human pancreatic islet function based on AI algorithms. Background Technology

[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. The increasing prevalence of multimodal physiological sensing devices such as continuous glucose monitoring (CGM), smart bracelets, and mobile electrocardiographs has made it possible to acquire large-scale, multi-source, and time-series individual health data. Based on this, the integration of multi-source time-series physiological signals for individualized health status assessment has gradually become a research hotspot, especially in chronic disease management fields such as early diabetes screening, disease progression assessment, and pancreatic function monitoring. Intelligent assessment technologies are gradually replacing traditional methods that rely on single-point detection and manual interpretation. At the same time, AI methods such as neural networks, attention mechanisms, and differential modeling are increasingly widely applied in biomedical time-series modeling, providing a technological foundation for establishing more accurate, dynamic, and interpretable physiological status assessment systems.

[0003] However, existing pancreatic function assessment techniques are mostly based on static assessments using single fasting blood glucose, postprandial blood glucose, or HbA1c levels, neglecting the complex correlation between an individual's dynamic behavior, physiological state, and pancreatic response over multiple time periods. Although some studies have attempted to incorporate CGM data and exercise data for joint analysis, significant shortcomings remain in multimodal data alignment, feature fusion, and physiological interpretability modeling. For example, traditional assessment models often employ static rules or simple linear combinations, making it difficult to effectively handle differences in sampling frequencies and data noise interference between different modalities; feature fusion lacks a spatiotemporal weighting adjustment mechanism, resulting in slow model responses to key behavioral events; more critically, mainstream models cannot effectively embed the pancreatic β-cell secretion mechanism, leading to a lack of clear physiological correspondences in output indicators. The method proposed in this invention, by introducing neural differential equations and biphasic insulin kinetic constraints, simultaneously achieves high model fit and physiological interpretability during dynamic modeling; furthermore, by combining a spatiotemporal attention mechanism for adaptive feature weighting, it effectively improves the robustness of the model under different behavioral scenarios and data quality conditions, solving the problem that existing technologies cannot achieve both high accuracy and clinical interpretability in multimodal data fusion. Summary of the Invention

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

[0005] Therefore, the technical problem solved by this invention is: existing pancreatic islet function assessment methods have limited static index information, lack multimodal data fusion mechanisms, and lack physiological interpretability of model results. The problem is how to achieve personalized digital assessment of pancreatic islet function based on multimodal data, dynamic behavior perception, and interpretable output.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a digital assessment method for human pancreatic islet function based on AI algorithms, including collecting multimodal physiological data and performing synchronous preprocessing to generate a preprocessed multimodal physiological dataset;

[0008] Feature extraction is performed based on the preprocessed multimodal physiological dataset to generate a low-dimensional input feature vector;

[0009] Based on the low-dimensional input feature vector, spatiotemporal attention dynamic weighted fusion is performed to generate a dynamic weighted fusion feature vector;

[0010] Based on the dynamically weighted fusion feature vector, a hybrid evaluation model driven by neural differential equations is constructed to generate pancreatic islet function evaluation indicators.

[0011] Based on the aforementioned pancreatic islet function assessment indicators, dynamic visualization and abnormal early warning processing are performed to generate a dynamic visualization assessment report.

[0012] As a preferred embodiment of the AI-based digital assessment method for human pancreatic islet function described in this invention, the acquisition of multimodal physiological data includes acquiring glucose signal data of subcutaneous interstitial fluid through a continuous blood glucose monitoring system.

[0013] High-frequency physiological signal data collected by smart wearable devices, including triaxial acceleration data, heart rate data, and status marker data;

[0014] Biochemical data collected in the laboratory, including fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide levels.

[0015] As a preferred embodiment of the AI-based digital assessment method for human pancreatic islet function described in this invention, the method for generating aligned multimodal physiological datasets includes using the time series of glucose signal data as the main reference time axis.

[0016] The high-frequency physiological signal data is repositioned and interpolated using a nonlinear dynamic time alignment algorithm based on the reference time axis, and the aligned high-frequency physiological signal data is output.

[0017] The biochemical index data are subjected to piecewise linear interpolation and spline fitting based on the reference time axis, and the aligned biochemical index data is output.

[0018] The aligned high-frequency physiological signal data, aligned biochemical index data and glucose signal data corresponding to the reference time axis are subjected to unified time mapping, synchronous denoising and sliding window smoothing to obtain preprocessed glucose signal data, preprocessed high-frequency physiological signal data and preprocessed biochemical index data.

[0019] The preprocessed multimodal physiological dataset is output, which includes preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data.

[0020] As a preferred embodiment of the AI-based digital assessment method for human pancreatic islet function described in this invention, the generation of low-dimensional input feature vectors includes performing local extremum detection on preprocessed glucose signal data based on preprocessed multimodal physiological datasets, and outputting a set of peak temporal features.

[0021] Wavelet decomposition is performed on the preprocessed glucose signal data to output the multi-scale frequency domain energy feature set of each decomposition layer.

[0022] Numerical integration is performed on the preprocessed glucose signal data within a preset time window to output a set of piecewise curve area under feature.

[0023] The amplitude of triaxial acceleration in the preprocessed high-frequency physiological signal data is integrated, and exercise energy consumption is estimated by combining individual body weight parameters, and exercise energy feature set is output.

[0024] By utilizing the state markers in the preprocessed high-frequency physiological signal data, resting and active segments are distinguished, and a segmented set of motion state features is output.

[0025] The preprocessed biochemical index data are standardized to output a standardized set of biochemical features.

[0026] The peak time series feature set, multi-scale frequency domain energy feature set, piecewise area under curve feature set, motion energy feature set, piecewise motion state feature set, and standardized biochemical feature set are fused and dimensionality reduced to generate a low-dimensional input feature vector.

[0027] As a preferred embodiment of the AI-based digital assessment method for human pancreatic islet function described in this invention, the generation of the dynamic weighted fusion feature vector includes feeding the low-dimensional input feature vector into a long short-term memory network, encoding the time dimension, and outputting a time encoding matrix.

[0028] Based on the low-dimensional input feature vector, a feature attention network is used to output a feature dimension attention weight matrix.

[0029] Based on the current physiological scenario and data quality, the weight coefficients in the aforementioned time encoding matrix and feature dimension weight matrix are dynamically adjusted, and weight compensation logic is enabled for lost or abnormal channels, outputting the corrected spatiotemporal weight matrix.

[0030] The modified spatiotemporal weight matrix is ​​fused with the low-dimensional input feature vector element by element using Hadamard product to generate and output a dynamically weighted fused feature vector.

[0031] As a preferred embodiment of the AI-based digital assessment method for human pancreatic islet function described in this invention, the generation of pancreatic islet function assessment indicators 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.

[0032] The pancreatic β-cell two-phase secretion kinetic equation is embedded as a physiological constraint term into the ODE model framework. The first phase equation describes the immediate secretion response of insulin, and the second phase equation describes the cumulative and sustained secretion response. The secretion rate is dynamically adjusted by combining the kinetic energy features in the fused feature vector, and the ODE model with physiological constraints is output.

[0033] The adjoint sensitivity method is used to iteratively optimize the network parameters and the weights of the physiological constraint regularization term of the ODE model with physiological constraints, and the optimized model parameter set is output.

[0034] The optimized model parameters are used to numerically integrate the ODE model with physiological constraints to generate the trajectory of the hidden state evolution over time and output the hidden state evolution sequence.

[0035] The final state or state within the sliding window of the hidden state evolution sequence is mapped to insulin sensitivity score and β-cell function decline rate through a decoding network, and the insulin sensitivity score and the β-cell function decline rate are used as indicators for pancreatic islet function assessment.

[0036] As a preferred embodiment of the AI-based digital assessment method for human pancreatic islet function described in this invention, the generation of the dynamic visualization assessment report includes generating a multi-dimensional dynamic chart based on the pancreatic islet function assessment indicators and corresponding time series data, using time, insulin sensitivity score and β-cell function decay rate as coordinates and color mapping, and outputting an initial dynamic visualization chart.

[0037] Set abnormal thresholds for pancreatic function indicators. When any evaluation indicator exceeds its corresponding threshold, the abnormal range is automatically highlighted in the dynamic chart, and a visual chart with abnormality highlighting is output.

[0038] The interactive linkage function is enabled for the dynamic chart. When the user clicks or hovers at a certain time point or segment, the original signal, feature value and model prediction information corresponding to that time point will pop up in real time, and an interactive data details view will be output.

[0039] Based on the changing trends of the aforementioned evaluation indicators, time series forecasting methods are applied to extrapolate pancreatic function indicators over a future period, and the forecast results are overlaid onto a dynamic chart to output an interactive dynamic chart with future forecast curves.

[0040] The system monitors whether the evaluation indicators continuously exceed the threshold at multiple consecutive time points. When the continuous exceedance condition is met, an automatic early warning mechanism is triggered to generate a risk warning notification and push it to the user terminal, outputting a warning message.

[0041] The dynamic charts and early warning notifications with anomaly annotation, interactive linkage and future prediction functions are summarized to form the final dynamic visualization evaluation report and saved, and a complete dynamic visualization evaluation report is output.

[0042] Secondly, embodiments of the present invention provide a digital assessment system for human pancreatic islet function based on AI algorithms, including:

[0043] Data acquisition and preprocessing module: Acquires multimodal physiological data and performs synchronous preprocessing to generate preprocessed multimodal physiological datasets;

[0044] Feature extraction module: Performs feature extraction based on the preprocessed multimodal physiological dataset to generate a low-dimensional input feature vector;

[0045] Fusion module: Based on the low-dimensional input feature vector, perform spatiotemporal attention dynamic weighted fusion to generate a dynamically weighted fused feature vector;

[0046] Indicator generation module: Based on the dynamically weighted fusion feature vector, a hybrid assessment model driven by neural differential equations is constructed to generate pancreatic islet function assessment indicators;

[0047] Visualization and anomaly warning module: Based on the pancreatic islet function assessment indicators, perform dynamic visualization and anomaly warning processing, and generate dynamic visualization assessment reports.

[0048] The beneficial effects of this invention are as follows: By fusing continuous glucose monitoring data, high-frequency physiological signals collected by wearable devices, and laboratory biochemical indicators, this invention achieves multimodal dynamic modeling of human pancreatic islet function. By introducing spatiotemporal attention mechanisms and neural differential equation models, it effectively improves the accuracy of feature fusion and the model's responsiveness to physiological changes. Furthermore, by incorporating the pancreatic β-cell biphasic secretion mechanism as a physiological constraint, it enhances the clinical interpretability and predictive reliability of the assessment results. Compared to traditional methods, this invention significantly improves accuracy, dynamism, and individualized assessment capabilities, making it suitable for scenarios such as diabetes risk warning, disease progression assessment, and personalized intervention recommendations. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0050] Figure 1 The overall flowchart of the AI-based digital assessment method for human pancreatic islet function provided in the first embodiment of the present invention is shown below. Detailed Implementation

[0051] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0052] Example 1, referring to Figure 1 As one embodiment of the present invention, a digital assessment method for human pancreatic islet function based on AI algorithms is provided, comprising:

[0053] S1: Collect multimodal physiological data and perform synchronous preprocessing to generate a preprocessed multimodal physiological dataset.

[0054] Glucose signal data of subcutaneous interstitial fluid were collected using a continuous glucose monitoring system;

[0055] High-frequency physiological signal data is collected through smart wearable devices, including triaxial acceleration data, heart rate data, and status marker data.

[0056] Biochemical data were collected in the laboratory, including fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide levels.

[0057] Use the time series of glucose signal data as the primary reference time axis;

[0058] The high-frequency physiological signal data is repositioned and interpolated using a nonlinear dynamic time alignment algorithm based on the reference time axis, and the aligned high-frequency physiological signal data is output.

[0059] The biochemical index data are subjected to piecewise linear interpolation and spline fitting based on the reference time axis, and the aligned biochemical index data is output.

[0060] The aligned high-frequency physiological signal data, aligned biochemical index data and glucose signal data corresponding to the reference time axis are subjected to unified time mapping, synchronous denoising and sliding window smoothing to obtain preprocessed glucose signal data, preprocessed high-frequency physiological signal data and preprocessed biochemical index data.

[0061] The preprocessed multimodal physiological dataset is output, which includes preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data.

[0062] In this embodiment of the invention, this step combines continuous blood glucose monitoring, smart wearable devices, and laboratory testing to acquire multi-source physiological signal data, and performs synchronous preprocessing on a unified time axis to ensure consistent temporal correspondence between the data. Specifically, firstly, subcutaneous tissue glucose signals are acquired in real time using a continuous blood glucose monitoring system. Taking a subcutaneous implantable sensor (e.g., the Dexcom G6 continuous blood glucose monitoring device) as an example, the sensor acquires blood glucose concentration data at fixed time intervals (e.g., every 5 minutes). The raw blood glucose signal is processed by algorithms such as Kalman filtering to eliminate high-frequency random noise, resulting in a smooth time-series data sequence of blood glucose concentration; simultaneously, the timestamp of each blood glucose sampling is recorded, forming a time-series record of the blood glucose monitoring device.

[0063] Secondly, high-frequency physiological signals, including triaxial acceleration, heart rate, and related status markers, are collected through smart wearable devices. Taking a smartwatch (such as the Apple Watch Series 8) as an example, its built-in triaxial accelerometer and photoelectric sensor can record motion acceleration and heart rate signals, respectively. The device continuously records the wearer's exercise intensity and heart rate sequences at a higher frequency (e.g., 1Hz in normal mode), and can estimate energy consumption (such as calories burned) in real time through integral calculations based on the acceleration signals. The wearable device also simultaneously records important status events or markers (such as entering exercise mode, changes in wearing status, etc.) and adds its own timestamp to each record, thus obtaining a high-frequency physiological signal data sequence containing multiple dimensions such as acceleration, heart rate, energy consumption, and status markers.

[0064] Then, low-frequency biochemical indicator data are collected periodically through laboratory testing and uploaded to the data processing center to enrich the multimodal data source. Specifically, important biochemical indicator values ​​of subjects are obtained through laboratory testing of venous blood samples, such as fasting blood glucose, postprandial blood glucose, glycated hemoglobin (HbA1c) percentage, and serum C-peptide levels. Each laboratory test result records the precise sampling time (for example, fasting blood glucose and C-peptide are usually collected in the morning on an empty stomach, postprandial blood glucose is collected at a fixed time point after a standard meal or glucose load, while HbA1c is collected monthly or quarterly). These test time points are stored as T_lab timestamps associated with the corresponding indicator values, forming a structured discrete biochemical indicator data table.

[0065] To ensure accurate timeline alignment of data from different sources, this step employs several synchronization measures. Firstly, all data acquisition devices' clocks are forcibly synchronized daily at fixed times. For example, at several fixed times each day (e.g., 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 (e.g., aligned with a standard atomic clock, ensuring a clock error of less than 1 millisecond) via Network Time Protocol (NTP) to eliminate clock discrepancies between devices. Secondly, signals from other sources are dynamically time-aligned using the timeline of continuous blood glucose monitoring data as a reference. When a significant trend is detected in the blood glucose monitoring sequence (e.g., a rate of increase exceeding 2 mg / dL / min, commonly seen during rapid postprandial blood glucose rise), the system automatically triggers the wearable device to enter a higher-frequency sampling mode (e.g., increasing from 1Hz to 10Hz) to capture more finely detailed physiological responses; it can also trigger expedited notifications to arrange additional laboratory testing (e.g., completing an HbA1c test within a short period) to obtain the latest biochemical indicators. All data acquired through these mechanisms uses the timeline of the blood glucose sensor as a reference. Next, a nonlinear dynamic time alignment algorithm is used to perform time relocation and interpolation processing on the high-frequency signal sequences collected by the wearable device. That is, based on the time reference of the blood glucose data, the time axis of the wearable signal is nonlinearly stretched or compressed to align key events with the moments of blood glucose changes, and the aligned wearable signal is then resampled and interpolated onto the blood glucose time axis. This effectively eliminates differences between different sampling frequencies, ensuring that 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, linear interpolation and spline curve fitting are used. Within the time interval between two adjacent laboratory tests, assuming the index changes continuously and gradually, linear interpolation is used to estimate the gradual transition value of the index within the interval, or spline fitting is used to generate a smooth curve passing through all discrete measurement points to approximately reflect the trend of the biochemical index over time. After the above interpolation and alignment processing, the multimodal data with different sampling frequencies are mapped to a unified time reference.

[0066] After unifying the time axis alignment, this step performs synchronous denoising and smoothing preprocessing operations on the multimodal data. Specifically, noise reduction and sliding window smoothing are performed on each channel signal. For example, for acceleration and heart rate signals recorded by wearable devices, digital filtering algorithms are applied to remove instantaneous spike noise and outliers, and then sliding window smoothing methods (such as moving averages or median filtering over a certain time period) are used to smooth short-term fluctuations and reduce the impact of random noise. For blood glucose monitoring sequences, individual outlier samples can be further removed based on the aforementioned Kalman filtering, and the blood glucose curve is moderately smoothed using a sliding window to eliminate residual small fluctuations between samples. The entire denoising and smoothing process is performed synchronously, that is, all signals are processed using the same time window and step size to ensure that the processed multiple time series remain strictly aligned and do not introduce new time offsets. Through these preprocessing measures, random errors and noise components in multi-source data are removed as much as possible, and each signal curve becomes smoother and more reliable.

[0067] Finally, after synchronization alignment, interpolation completion, and denoising and smoothing, a preprocessed multimodal physiological dataset was formed. This dataset consists of various types of physiological data sequences, all of which are synchronously correlated in time. Specifically, it includes: time series of blood glucose concentrations obtained from continuous blood glucose monitoring (reflecting continuous changes in subcutaneous tissue glucose levels); high-frequency physiological signal sequences recorded by smart wearable devices, such as processed exercise intensity (acceleration) signals, heart rate signals, energy consumption indicators, and related state-labeled events; and continuous time series generated by interpolation / fitting of discrete laboratory biochemical test values, such as data curves showing the changes in fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide levels over time. The preprocessed and integrated multimodal dataset has a unified time axis and good data quality, containing complete information on individual blood glucose dynamics, exercise physiology, and biochemical indicators within the same period, providing sufficient and reliable basic data for subsequent steps such as feature extraction, model training, and system evaluation.

[0068] S2: Based on the preprocessed multimodal physiological dataset, feature extraction is performed to generate a low-dimensional input feature vector.

[0069] Based on the preprocessed multimodal physiological dataset, local extremum detection is performed on the preprocessed glucose signal data, and a set of peak temporal features is output.

[0070] Wavelet decomposition is performed on the preprocessed glucose signal data to output the multi-scale frequency domain energy feature set of each decomposition layer.

[0071] Numerical integration is performed on the preprocessed glucose signal data within a preset time window to output a set of piecewise curve area under feature.

[0072] The amplitude of triaxial acceleration in the preprocessed high-frequency physiological signal data is integrated, and exercise energy consumption is estimated by combining individual body weight parameters, and exercise energy feature set is output.

[0073] By utilizing the state markers in the preprocessed high-frequency physiological signal data, resting and active segments are distinguished, and a segmented set of motion state features is output.

[0074] The preprocessed biochemical index data are standardized to output a standardized set of biochemical features.

[0075] The peak time series feature set, multi-scale frequency domain energy feature set, piecewise area under curve feature set, motion energy feature set, piecewise motion state feature set, and standardized biochemical feature set are fused and dimensionality reduced to generate a low-dimensional input feature vector.

[0076] In this embodiment of the invention, after the collection and preprocessing of multimodal physiological data are completed, it is necessary to further extract effective features from various types of data in order to fully describe the physiological changes and potential trends of pancreatic islet function and provide input data for subsequent model training.

[0077] First, local extremum detection is performed on the preprocessed glucose signal data. This involves detecting and marking local maxima and maxima on the continuous glucose concentration curve to obtain information on the time, amplitude, and frequency of blood glucose peaks. These local extrema reflect the fluctuation characteristics of an individual's blood glucose level, especially during periods of significant blood glucose changes triggered by postprandial blood glucose rises and falls, exercise-induced changes, or other life events, and are of significant physiological importance. Peak temporal features are extracted from these extrema and the time intervals between their occurrence to reflect the frequency and severity of blood glucose fluctuations.

[0078] Secondly, to further explore the hidden frequency domain features in glucose signals, this step employs a multi-scale wavelet decomposition method to perform multi-level decomposition processing on the preprocessed glucose signal data. Wavelet decomposition can transform the glucose signal from a single time-domain signal into a frequency-domain component with multiple scales, each scale corresponding to different physiological processes or noise components. By analyzing the signal energy distribution at different decomposition scales, a set of multi-scale frequency-domain energy features is extracted. These features can reflect the periodic or irregular changes in blood glucose levels, helping the model to more accurately identify abnormal patterns.

[0079] Next, a numerical integration method was used to extract features from the preprocessed glucose signal data. Specifically, the area under the blood glucose concentration curve was calculated within a pre-defined time window. This method is similar to integral curve analysis used clinically to evaluate overall blood glucose control, capturing the overall blood glucose exposure level and trend over a specific time period, particularly helpful in quantifying important indicators such as postprandial glycemic load or nighttime blood glucose control. This resulted in a piecewise area under the curve feature set.

[0080] On the other hand, this step also requires extracting motion-related feature data from the preprocessed high-frequency physiological signals. Specifically, firstly, the triaxial acceleration signals recorded by the wearable device are subjected to amplitude integration processing to estimate the total amount of exercise during physical activity. Then, based on the individual subject's weight information, the exercise data is converted into exercise energy consumption values ​​to intuitively reflect the body's energy metabolism during daily activities and exercise, forming a set of exercise energy features. These exercise energy features provide a quantitative basis for subsequent assessment of the impact of individual exercise on pancreatic function.

[0081] Furthermore, the system analyzes the status-labeled data collected by smart wearable devices. This data typically records the subject's physical state at various time points, such as resting, light activity, moderate activity, or vigorous activity. Based on this status-labeled data, the system automatically identifies different physiological activity segments, clearly distinguishing between resting and active states. It then calculates the temporal distribution and duration of each activity state to output a segmented set of motion state features. This status-labeled feature helps accurately differentiate the varying impacts of different activity states on blood glucose and pancreatic function.

[0082] For biochemical indicator data collected in the laboratory, this step involves data standardization to transform biochemical indicators with different units or measurement ranges into uniform dimensionless values. For example, using the mean and standard deviation of the subject population as a reference, each indicator data is converted into standardized data so that subsequent feature fusion can be performed uniformly across different indicators. This standardization process can effectively eliminate data bias and analytical errors caused by different measurement units, outputting a standardized set of biochemical features, and significantly improving the accuracy of cross-indicator and cross-individual data analysis.

[0083] Finally, all the extracted features, including peak time-series features, multi-scale frequency domain energy features, piecewise area under the curve features, motion energy features, piecewise motion state features, and the standardized biochemical feature set in the blood glucose signal, are uniformly fused. To effectively reduce data redundancy and further improve the computational efficiency and stability of the model, this step uses dimensionality reduction techniques such as principal component analysis to reduce the feature space of the fused feature set, obtaining a low-dimensional input feature vector with moderate data volume and sufficient information retention. This low-dimensional input feature vector accurately represents the core changing trend of individual pancreatic islet function with as few dimensions as possible, ensuring reliable input and computational efficiency in the subsequent spatiotemporal dynamic modeling stage.

[0084] S3: Based on the low-dimensional input feature vector, perform spatiotemporal attention dynamic weighted fusion to generate a dynamic weighted fusion feature vector.

[0085] The low-dimensional input feature vector is fed into a long short-term memory network to encode the time dimension and output a time encoding matrix.

[0086] Based on the low-dimensional input feature vector, a feature attention network is used to output a feature dimension attention weight matrix.

[0087] Based on the current physiological scenario and data quality, the weight coefficients in the aforementioned time encoding matrix and feature dimension weight matrix are dynamically adjusted, and weight compensation logic is enabled for lost or abnormal channels, outputting the corrected spatiotemporal weight matrix.

[0088] The modified spatiotemporal weight matrix is ​​fused with the low-dimensional input feature vector element by element using Hadamard product to generate and output a dynamically weighted fused feature vector.

[0089] In this embodiment of the invention, in order to further improve the information utilization efficiency and feature expression capability of the input feature vector, after obtaining the low-dimensional input feature vector, a spatiotemporal attention dynamic weighted fusion method is adopted to dynamically weight the low-dimensional feature vector from the perspectives of time dimension and feature dimension, so as to obtain a fused feature vector that takes into account both the time series change trend and the contribution of key features.

[0090] First, the low-dimensional input feature vectors are input into a Long Short-Term Memory (LSTM) network in the order of acquisition time for temporal encoding. LSTM networks can effectively capture long-term dependencies in time-series data, making them particularly suitable for processing the multimodal physiological data features in this invention. This network structure dynamically retains or forgets historical input information through internal memory units and gating mechanisms, and continuously updates the hidden state based on the current input, thereby accurately expressing the patterns and trends of feature evolution over time. This invention preferably uses a unidirectional LSTM network to adapt to real-time processing scenarios, but a bidirectional LSTM (BiLSTM) structure can also be selected according to actual needs to further improve the ability to express the dependencies between different time series.

[0091] Secondly, this invention further introduces a feature attention network to evaluate and calculate the contribution of each feature dimension to the final evaluation target. Specifically, the output of the aforementioned LSTM network is input into the feature attention network. First, overall information aggregation is performed on each feature dimension, for example, using pooling operations or feature channel statistics calculation, to compress and extract the importance information of each feature channel. Next, a nonlinear transformation is performed on the compressed feature information using a multilayer perceptron or a fully connected network to generate weight coefficients representing the importance of each feature dimension. After appropriate normalization, these weight coefficients finally form a feature dimension attention weight matrix, used to highlight the importance of key features in the spatial dimension while suppressing redundant features or noise interference that contribute little to the evaluation target.

[0092] Furthermore, this invention incorporates a dynamic adjustment mechanism for attention weights. Specifically, to more accurately adapt to actual clinical or daily scenarios, this invention dynamically adjusts the time encoding results and feature attention weight matrix of the aforementioned LSTM output based on the current physiological scenario and data quality status. The system identifies the subject's current physiological state in real time, such as resting, exercising, or special physiological states, and monitors data quality levels, such as the presence of equipment detachment, signal interruption, or significant interference noise. When a specific scenario or data anomaly is detected, corresponding weight compensation logic is activated. For example, when abnormal or missing heart rate data is detected during strenuous exercise, the system reduces the weight of that feature channel and initiates data interpolation compensation for adjacent time windows; conversely, when the subject is at rest and data quality is good, the system increases the weight of stability features. Through this dynamic scenario awareness and data quality adaptive adjustment mechanism, this 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 results.

[0093] Finally, the dynamically adjusted time encoding matrix and feature dimension attention weight matrix are fused element-wise with the low-dimensional input feature vector. Specifically, each feature vector is multiplied element-wise according to its corresponding time point and feature dimension attention weight, thus highlighting important features and significantly reducing the impact of redundant or unimportant features on the fusion result. This element-wise weighted fusion method ensures that the generated fused feature vector maintains information richness while effectively reducing interference and redundancy between features, highlighting the most valuable key information. The dynamically weighted fused feature vector obtained through this fusion strategy serves as the input to the next step of the pancreatic islet function assessment model, guaranteeing the accuracy, stability, and interpretability of subsequent model building and evaluation analysis.

[0094] In summary, the attention fusion strategy proposed in this invention, based on LSTM encoding and feature attention mechanism and combined with dynamic scene and data quality perception, can achieve more refined adaptive fusion of multimodal physiological data features. Compared with existing static fusion methods, this invention has significant technical advantages in data robustness, model generalization and evaluation accuracy.

[0095] S4: Construct a hybrid evaluation model driven by neural differential equations based on the dynamically weighted fusion feature vectors to generate pancreatic islet function evaluation indicators.

[0096] Based on the dynamically weighted fusion feature vector, the hidden state update function of the neural differential equation model is defined, and the unconstrained ODE model framework is output.

[0097] The pancreatic β-cell two-phase secretion kinetic equation is embedded as a physiological constraint term into the ODE model framework. The first phase equation describes the immediate secretion response of insulin, and the second phase equation describes the cumulative and sustained secretion response. The secretion rate is dynamically adjusted by combining the kinetic energy features in the fused feature vector, and the ODE model with physiological constraints is output.

[0098] The adjoint sensitivity method is used to iteratively optimize the network parameters and the weights of the physiological constraint regularization term of the ODE model with physiological constraints, and the optimized model parameter set is output.

[0099] The optimized model parameters are used to numerically integrate the ODE model with physiological constraints to generate the trajectory of the hidden state evolution over time and output the hidden state evolution sequence.

[0100] The final state or state within the sliding window of the hidden state evolution sequence is mapped to insulin sensitivity score and β-cell function decline rate through a decoding network, and the insulin sensitivity score and the β-cell function decline rate are used as indicators for pancreatic islet function assessment.

[0101] In this embodiment of the invention, after obtaining the dynamically weighted fusion feature vector, a hybrid modeling method integrating neural networks and physiological mechanisms is proposed to further accurately describe the dynamic changes in pancreatic islet function and improve the interpretability and prediction accuracy of the model. The specific implementation process first involves constructing a Neural Differential Equation (Neural ODE) model framework and defining the update mechanism of the hidden state over time. This model structure essentially utilizes neural networks to define the evolution of the system state in continuous time. This continuous representation allows for a more refined simulation of the long-term evolution of pancreatic islet function.

[0102] To enhance the physiological interpretability of the model, this invention innovatively embeds the biphasic secretion kinetics of pancreatic β-cells as a constraint within the framework of neural differential equations. The first-phase secretion equation describes the rapid, immediate response of insulin secretion in the initial stages of sharp blood glucose changes (such as after a meal), while the second-phase secretion equation simulates the sustained, cumulative response of insulin over a longer period. This phased kinetic mechanism is a widely recognized physiological phenomenon in the medical field. This invention deeply integrates it with an artificial intelligence model, enabling the AI ​​model to not only capture data-driven patterns but also possess clear physiological significance. Furthermore, this invention innovatively incorporates exercise energy expenditure features from dynamic fusion characteristics, dynamically adjusting the insulin secretion rate in real time based on changes in the subject's exercise state. This allows the model to adapt to different lifestyle scenarios, more closely resembling real physiological states.

[0103] To effectively optimize the aforementioned hybrid model, this invention employs the adjoint sensitivity method for iterative optimization of network parameters and physiological constraints. This optimization approach not only efficiently calculates the model's sensitivity to each parameter but also has low memory resource requirements, making it particularly suitable for processing large-scale datasets with long time series. This optimization method continuously adjusts the weights of model parameters and physiological constraints through iterative calculations to achieve an optimal balance, ensuring both accurate data fitting and the rationality of the physiological mechanisms.

[0104] Using the optimized model parameters, numerical integration is further performed to solve for the evolutionary trajectory of the latent state over time. This latent state trajectory details the continuous changes in various aspects of pancreatic islet function over time, providing a foundation for the calculation of subsequent physiological indicators. This invention constructs a specialized decoding network to map the final state of the model's latent state trajectory or the latent state information within a specific time period into two clinically significant indicators: insulin sensitivity score and β-cell function decline rate. These assessment indicators obtained through latent state decoding not only possess clinically intuitive interpretability but also significantly surpass traditional static assessment methods in sensitivity and accuracy, thereby achieving precise quantification and dynamic assessment of pancreatic islet function status, fully demonstrating the inventiveness and technical rationality of this invention.

[0105] S5: Based on the pancreatic islet function assessment indicators, perform dynamic visualization and abnormal early warning processing to generate a dynamic visualization assessment report.

[0106] Based on the pancreatic islet function assessment indicators and the corresponding time series data, a multidimensional dynamic chart is generated, using time, insulin sensitivity score and β cell function decline rate as coordinates and color mapping to output an initial dynamic visualization chart.

[0107] Set abnormal thresholds for pancreatic function indicators. When any evaluation indicator exceeds its corresponding threshold, the abnormal range is automatically highlighted in the dynamic chart, and a visual chart with abnormality highlighting is output.

[0108] The interactive linkage function is enabled for the dynamic chart. When the user clicks or hovers at a certain time point or segment, the original signal, feature value and model prediction information corresponding to that time point will pop up in real time, and an interactive data details view will be output.

[0109] Based on the changing trends of the aforementioned evaluation indicators, time series forecasting methods are applied to extrapolate pancreatic function indicators over a future period, and the forecast results are overlaid onto a dynamic chart to output an interactive dynamic chart with future forecast curves.

[0110] The system monitors whether the evaluation indicators continuously exceed the threshold at multiple consecutive time points. When the continuous exceedance condition is met, an automatic early warning mechanism is triggered to generate a risk warning notification and push it to the user terminal, outputting a warning message.

[0111] The dynamic charts and early warning notifications with anomaly annotation, interactive linkage and future prediction functions are summarized to form the final dynamic visualization evaluation report and saved, and a complete dynamic visualization evaluation report is output.

[0112] In this embodiment of the invention, a complete dynamic visualization and abnormal warning processing method is proposed to present pancreatic islet function assessment results intuitively and promptly, and to effectively monitor and warn of potential risks. The specific implementation process begins by designing and constructing multidimensional dynamic charts with clear expressive capabilities based on two main assessment indicators: calculated insulin sensitivity score and β-cell function decline rate, combined with their corresponding time-series information. These dynamic charts intuitively display time information, sensitivity score, and function decline rate through different dimensional coordinates, and use color changes to map the levels and trends of the indicators, providing doctors or subjects with a platform for quickly and intuitively observing dynamic changes in pancreatic islet function.

[0113] Secondly, this invention sets clear clinical or research reference thresholds for pancreatic islet function assessment indicators, such as abnormal ranges determined based on medical guidelines, expert consensus, or statistical standards. When any assessment indicator exceeds these predetermined thresholds, the dynamic chart automatically highlights the corresponding abnormal interval, clearly and conspicuously indicating the abnormality or potential risk area of ​​the assessed object, so that clinicians or users can pay attention and intervene in a timely manner.

[0114] To further enhance the interactivity and practicality of the assessment report, this invention innovatively introduces an interactive linkage function for dynamic charts. When a user clicks or hovers over any point in time or area on the chart, the system will display detailed data information for the corresponding time period in real time, including raw physiological signals, feature data, and relevant data predicted by the model. This interactive linkage function helps clinical experts and patients intuitively understand the causes of abnormalities and related physiological background information, greatly improving the clinical application value of the assessment report and the user experience.

[0115] Furthermore, to enhance the predictive power and timeliness of risk prevention in this invention, time series forecasting methods, such as autoregressive models or long short-term memory neural networks, are further utilized to predict and extrapolate the changing trends of pancreatic islet function indicators over a certain future period. These prediction results are then overlaid with historical monitoring data and displayed in dynamic charts, forming a dynamic visualization interface with future trend predictions. This effectively assists users in anticipating changes in pancreatic islet function and preparing intervention or treatment plans in advance.

[0116] Finally, this invention introduces a continuous monitoring and automatic early warning mechanism. Specifically, it continuously monitors pancreatic function indicators. If multiple consecutive time periods show indicators exceeding set abnormal thresholds, the system automatically identifies a high-risk state and immediately generates a risk warning notification, proactively pushing it 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 proactive intervention reminder function not only demonstrates the technical ingenuity of this invention in the evaluation process but also truly endows it with practical value for proactive medical management and personalized health management.

[0117] By integrating functions such as anomaly labeling, interactive linkage, future trend prediction, and risk warning, the complete dynamic visualization assessment report generated by this invention significantly improves the comprehensibility, reliability, and operability of the assessment results, and significantly enhances the innovation, practicality, and clinical application potential of the pancreatic function monitoring and intervention system.

[0118] Example 2 is the second embodiment of the present invention, which differs from the previous embodiment in that:

[0119] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art or the current technical solution, can be embodied in the form of a software product. This current computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0120] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0121] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, 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 medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0122] Example 3, an embodiment of the present invention, provides a digital assessment system for human pancreatic islet function based on AI algorithms, including a data acquisition and preprocessing module, a feature extraction module, a fusion module, an indicator generation module, and a visualization and anomaly warning module.

[0123] Data acquisition and preprocessing module: Acquires multimodal physiological data and performs synchronous preprocessing to generate preprocessed multimodal physiological datasets;

[0124] Feature extraction module: Performs feature extraction based on the preprocessed multimodal physiological dataset to generate a low-dimensional input feature vector;

[0125] Fusion module: Based on the low-dimensional input feature vector, perform spatiotemporal attention dynamic weighted fusion to generate a dynamically weighted fused feature vector;

[0126] Indicator generation module: Based on the dynamically weighted fusion feature vector, a hybrid assessment model driven by neural differential equations is constructed to generate pancreatic islet function assessment indicators;

[0127] Visualization and anomaly warning module: Based on the pancreatic islet function assessment indicators, perform dynamic visualization and anomaly warning processing, and generate dynamic visualization assessment reports.

[0128] Example 4 is an embodiment of the present invention, which provides a digital assessment method for human pancreatic islet function based on AI algorithm. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation / comparative experiments.

[0129] The study was conducted in a professional medical institution, selecting seven healthy adults with varying health conditions to ensure the sample included patients with normal blood glucose, prediabetes, and mild to moderate diabetes. Before the trial, each participant was provided with a continuous glucose monitoring system (e.g., Dexcom G6) and a smart bracelet (e.g., Fitbit Charge 5) with built-in triaxial accelerometer and heart rate monitoring functions, which they wore continuously for at least 72 hours. During this period, participants maintained their daily routines, recording continuous blood glucose signals and real-time triaxial accelerometer, heart rate, and activity status markers. In addition, a professional laboratory conducted periodic biochemical tests on the participants, collecting multiple indicators including fasting blood glucose, 2-hour postprandial blood glucose, and glycated hemoglobin (HbA1c), and periodically measuring C-peptide levels to obtain comprehensive biochemical data. Subsequently, using the timeline of the continuous glucose monitoring system as a reference, a nonlinear dynamic time alignment method was used to accurately reposition and interpolate the high-frequency physiological data collected by the smart bracelet. Simultaneously, the low-frequency biochemical data from the laboratory were mapped to the same timeline through linear interpolation and smoothing fitting, uniformly generating a preprocessed multimodal physiological dataset. Next, feature extraction was performed on the dataset, including peak temporal feature extraction, wavelet frequency domain energy feature extraction, and curve integral feature extraction of blood glucose data. Simultaneously, exercise energy consumption was calculated and exercise state intervals were defined. Combined with the standardized results of biochemical indicators, a low-dimensional input feature vector was obtained. Furthermore, the features were temporally encoded using a Long Short-Term Memory (LSTM) network, and spatiotemporal attention networks were used to calculate and dynamically adjust feature weights, thereby generating a dynamically weighted fusion feature vector. Based on this feature vector, a neural differential equation hybrid model constrained by the pancreatic β-cell biphasic secretion mechanism was constructed to generate pancreatic islet function assessment indicators, including insulin sensitivity scores and β-cell function decay rates. Finally, dynamic visualization and anomaly warning processing were applied to the generated assessment indicators, resulting in a detailed and interactive dynamic visualization assessment report. Experimental reference data are shown in Table 1.

[0130] Table 1 Experimental Data Record

[0131]

[0132] The comparative analysis of the data in the table above shows that the method of this invention exhibits significant advantages over traditional methods in terms of accuracy and stability in pancreatic function assessment. Specifically, regarding insulin sensitivity scores, the sensitivity scores calculated using the method of this invention for all subjects were higher than those calculated using traditional methods. This indicates that the dynamic data fusion and neural differential equation hybrid model of this invention can more accurately assess individual insulin sensitivity, capture more potential physiological changes, and avoid assessment biases caused by static indicators in traditional methods. Furthermore, regarding the β-cell function decline rate, the decline trend calculated by the method of this invention for each subject was more obvious and sensitive than that calculated by traditional methods, which was particularly evident in high-risk individuals such as Subject 2 and Subject 4. This indicates that the dynamic weighted fusion mechanism and biphasic physiological constraint model proposed in this invention can more sensitively capture early or latent decline trends in pancreatic function, which is of great value for clinical early warning and intervention. In addition, the prediction error rate index shows that the error rate of the method of this invention is generally lower than that of traditional methods. For example, the prediction error of Subject 4 using traditional methods reached 11.9%, while that of this invention was reduced to 4.8%, indicating that the model structure and optimization method of this invention can significantly improve prediction accuracy and reliability. This improvement in accuracy is attributed to the invention's dynamic perception of data quality and weight adjustment strategy, as well as the rational integration of physiological mechanisms.

Claims

1. A digital assessment method for human pancreatic islet function based on AI algorithms, characterized in that, include: Multimodal physiological data were collected and preprocessed simultaneously to generate a preprocessed multimodal physiological dataset. Feature extraction is performed based on the preprocessed multimodal physiological dataset to generate a low-dimensional input feature vector; Based on the low-dimensional input feature vector, spatiotemporal attention dynamic weighted fusion is performed to generate a dynamic weighted fusion feature vector; Based on the dynamically weighted fusion feature vector, a hybrid evaluation model driven by neural differential equations is constructed to generate pancreatic islet function evaluation indicators. Based on the pancreatic islet function assessment indicators, dynamic visualization and abnormal early warning processing are performed to generate a dynamic visualization assessment report. The generated pancreatic islet function assessment index includes defining the hidden state update function of the neural differential equation model based on the dynamically weighted fusion feature vector, and outputting an unconstrained ODE model framework. The pancreatic β-cell two-phase secretion kinetic equation is embedded as a physiological constraint term into the ODE model framework. The first phase equation describes the immediate secretion response of insulin, and the second phase equation describes the cumulative and sustained secretion response. The secretion rate is dynamically adjusted by combining the kinetic energy features in the fused feature vector, and the ODE model with physiological constraints is output. The adjoint sensitivity method is used to iteratively optimize the network parameters and the weights of the physiological constraint regularization term of the ODE model with physiological constraints, and the optimized model parameter set is output. The optimized model parameters are used to numerically integrate the ODE model with physiological constraints to generate the trajectory of the hidden state evolution over time and output the hidden state evolution sequence. The final state or state within the sliding window of the hidden state evolution sequence is mapped to insulin sensitivity score and β-cell function decline rate through a decoding network, and the insulin sensitivity score and the β-cell function decline rate are used as indicators for pancreatic islet function assessment.

2. The method for digital assessment of human pancreatic islet function based on AI algorithm as described in claim 1, characterized in that, The acquisition of multimodal physiological data includes acquiring glucose signal data of subcutaneous interstitial fluid through a continuous blood glucose monitoring system; High-frequency physiological signal data collected by smart wearable devices, including triaxial acceleration data, heart rate data, and status marker data; Biochemical data collected in the laboratory, including fasting blood glucose, postprandial blood glucose, glycated hemoglobin, and C-peptide levels.

3. The method for digital assessment of human pancreatic islet function based on AI algorithm as described in claim 2, characterized in that, The generation of the preprocessed multimodal physiological dataset includes using the time series of glucose signal data as the main reference time axis; The high-frequency physiological signal data is time-repositioned and interpolated based on the main reference time axis using a nonlinear dynamic time alignment algorithm, and the aligned high-frequency physiological signal data is output. The biochemical index data are subjected to piecewise linear interpolation and spline fitting based on the main reference time axis, and the aligned biochemical index data is output. The aligned high-frequency physiological signal data, aligned biochemical index data and glucose signal data corresponding to the main reference time axis are subjected to unified time mapping, synchronous denoising and sliding window smoothing to obtain preprocessed glucose signal data, preprocessed high-frequency physiological signal data and preprocessed biochemical index data. The preprocessed multimodal physiological dataset is output, which includes preprocessed glucose signal data, preprocessed high-frequency physiological signal data, and preprocessed biochemical index data.

4. The method for digital assessment of human pancreatic islet function based on AI algorithm as described in claim 3, characterized in that, The generation of low-dimensional input feature vectors includes performing local extremum detection on preprocessed glucose signal data based on preprocessed multimodal physiological datasets, and outputting a set of peak temporal features; Wavelet decomposition is performed on the preprocessed glucose signal data to output the multi-scale frequency domain energy feature set of each decomposition layer. Numerical integration is performed on the preprocessed glucose signal data within a preset time window to output a set of piecewise curve area under feature. The amplitude of triaxial acceleration in the preprocessed high-frequency physiological signal data is integrated, and exercise energy consumption is estimated by combining individual body weight parameters, and exercise energy feature set is output. By utilizing the state markers in the preprocessed high-frequency physiological signal data, resting and active segments are distinguished, and a segmented set of motion state features is output. The preprocessed biochemical index data are standardized to output a standardized set of biochemical features. The peak time series feature set, multi-scale frequency domain energy feature set, piecewise area under curve feature set, motion energy feature set, piecewise motion state feature set, and standardized biochemical feature set are fused and dimensionality reduced to generate a low-dimensional input feature vector.

5. The method for digital assessment of human pancreatic islet function based on AI algorithm as described in 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, encoding the time dimension, and outputting a time encoding matrix. Based on the low-dimensional input feature vector, a feature attention network is used to output a feature dimension attention weight matrix. Based on the current physiological scenario and data quality, the weight coefficients in the aforementioned time encoding matrix and feature dimension weight matrix are dynamically adjusted, and weight compensation logic is enabled for lost or abnormal channels, outputting the corrected spatiotemporal weight matrix. The modified spatiotemporal weight matrix is ​​fused with the low-dimensional input feature vector element by element using Hadamard product to generate and output a dynamically weighted fused feature vector.

6. The method for digital assessment of human pancreatic islet function based on AI algorithm as described in claim 5, characterized in that, The generation of the dynamic visualization assessment report includes generating a multi-dimensional dynamic chart based on the pancreatic islet function assessment indicators and the corresponding time series data, using time, insulin sensitivity score and β cell function decline rate as coordinates and color mapping, and outputting an initial dynamic visualization chart. Set abnormal thresholds for pancreatic function indicators. When any evaluation indicator exceeds its corresponding threshold, the abnormal range is automatically highlighted in the multidimensional dynamic chart, and a visual chart with abnormality highlighting is output. The interactive linkage function is enabled for the multidimensional dynamic chart. When the user clicks or hovers at a certain time point or segment, the original signal, feature value and model prediction information corresponding to that time point will pop up in real time, and an interactive data details view will be output. Based on the changing trends of the aforementioned evaluation indicators, time series forecasting methods are applied to extrapolate pancreatic function indicators over a future period, and the forecast results are superimposed onto a multidimensional dynamic chart to output an interactive dynamic chart with future forecast curves. The system monitors whether the evaluation indicators continuously exceed the threshold at multiple consecutive time points. When the continuous exceedance condition is met, an automatic early warning mechanism is triggered to generate a risk warning notification and push it to the user terminal, outputting a warning message. The visualization charts with abnormal highlights, interactive data detail views, interactive dynamic charts with future prediction curves, and early warning messages are summarized to form the final dynamic visualization evaluation report, which is then saved and output as a complete dynamic visualization evaluation report.

7. A digital assessment system for human pancreatic islet function based on an AI algorithm, used to implement the digital assessment method for human pancreatic islet function based on an AI algorithm as described in any one of claims 1 to 6, characterized in that, include: Data acquisition and preprocessing module: Acquires multimodal physiological data and performs synchronous preprocessing to generate preprocessed multimodal physiological datasets; Feature extraction module: Performs feature extraction based on the preprocessed multimodal physiological dataset to generate a low-dimensional input feature vector; Fusion module: Based on the low-dimensional input feature vector, perform spatiotemporal attention dynamic weighted fusion to generate a dynamically weighted fused feature vector; Indicator generation module: Based on the dynamically weighted fusion feature vector, a hybrid assessment model driven by neural differential equations is constructed to generate pancreatic islet function assessment indicators; Visualization and anomaly warning module: Based on the pancreatic islet function assessment indicators, perform dynamic visualization and anomaly warning processing, and generate dynamic visualization assessment reports.

Images