Multimodal data fusion system for predicting the risk of diabetes after acute pancreatitis

Abstract

This invention relates to the field of medical data processing technology, specifically to a multimodal data fusion system for predicting diabetes risk after acute pancreatitis. The system includes: acquiring feature data and treatment indices for each dimension of the target patient and several reference patients after acute pancreatitis; obtaining a weighted similarity index based on the differences in feature data between the target patient and each reference patient in the same dimension; obtaining a cutoff feature length based on the treatment index of each reference patient and the overall distribution differences of the reference patients, combined with the weighted similarity index; fitting the treatment indices of reference patients whose diagnosis time is greater than or less than or equal to the cutoff feature length to obtain pancreatic islet compensation function curves; and obtaining the diabetes risk prediction result for the target patient based on the deviation between the target patient's treatment index and the pancreatic islet compensation function curves. This invention shows better accuracy in predicting the risk of patients in a compensated state.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, specifically to a multimodal data fusion system for predicting the risk of diabetes after acute pancreatitis. Background Technology

[0002] Acute pancreatitis is an acute chemical inflammation caused by the abnormal activation of digestive juices produced by the pancreas, which begin to digest the pancreas's own tissues. Severe cases can lead to irreversible damage to pancreatic tissue and are a significant cause of type 3C diabetes. This type of diabetes is often missed, but its complication risks are no different from those of conventional diabetes. By predicting the risk of diabetes after acute pancreatitis, high-risk individuals can be identified before blood sugar abnormalities appear. This allows limited medical resources to be concentrated on high-risk groups, enabling early lifestyle interventions and even drug prevention to delay or prevent the onset of diabetes and reduce long-term complications such as cardiovascular, renal, and neurological complications.

[0003] In existing technologies, the random forest algorithm is generally used to construct a diabetes risk prediction model after acute pancreatitis. The specific implementation process is as follows: First, multimodal data of patients during the acute exacerbation of acute pancreatitis are systematically collected. Simultaneously, each patient sample is assigned an initial binary label to determine whether they have been newly diagnosed with diabetes within a specific time window after discharge. Then, during the model training phase, the multimodal feature matrices of all patient samples and their corresponding classification labels are used as input to train the random forest classifier. Finally, the trained model can be used to perform individualized risk prediction for newly diagnosed acute pancreatitis patients, outputting diabetes risk monitoring results.

[0004] From a pathological perspective, the onset of diabetes is not an instantaneous event, but a dynamic and continuous evolutionary process. In the prediabetic stage, pancreatic β-cell function is not immediately lost; instead, it may be in a state of compensatory enhancement. The body compensates for the decreased efficiency caused by tissue insulin resistance or early cell damage by increasing insulin secretion, thereby maintaining normal blood glucose levels for a considerable period. Only when β-cell function eventually becomes decompensated and unable to meet metabolic demands will clinically significant increases in blood glucose levels occur. Random forest-based prediction models essentially capture this final outcome of "decompensation." Therefore, this model is insufficient in predicting the large patient population in a "compensated state" and cannot accurately predict the diabetes risk of these patients. Summary of the Invention

[0005] To address the technical problem of poor accuracy in predicting the risk of patients in a compensated state using existing risk prediction models, the present invention aims to provide a multimodal data fusion-based risk prediction system for diabetes after acute pancreatitis. The specific technical solution adopted is as follows:

[0006] This invention provides a multimodal data fusion-based system for predicting the risk of diabetes following acute pancreatitis, comprising a memory and a processor, wherein the processor executes a computer program stored in the memory to perform the following steps:

[0007] Acquire characteristic data and treatment index for each dimension of the target patient and several reference patients after acute pancreatitis; among them, the target patient is a patient without a confirmed diabetes, and the reference patients are patients with a confirmed diabetes, with each reference patient corresponding to a diagnosis time length;

[0008] Based on the differences and similarities between the target patient and each reference patient in the same dimension of feature data, a weighted similarity index between the target patient and each reference patient is obtained.

[0009] Based on the treatment index of each reference patient and the distribution differences of the overall reference patients, the length of each diagnosis time is screened in combination with the weighted similarity index to obtain the boundary feature length;

[0010] Curve fitting was performed on the treatment index of reference patients whose diagnosis time was longer than the cutoff characteristic length, and linear fitting was performed on the treatment index of reference patients whose diagnosis time was shorter than or equal to the cutoff characteristic length, to obtain the pancreatic islet compensation function curve.

[0011] Based on the deviations between the treatment index and pancreatic islet compensation function curves of the target patients at different time lengths, the diabetes risk prediction results for the target patients are obtained.

[0012] Preferably, the step of filtering each diagnosis time length based on the treatment index of each reference patient and the distribution differences of the overall reference patients, combined with the weighted similarity index, to obtain the boundary feature length specifically includes:

[0013] Obtain a first time length and a second time length for a selected time length, wherein the selected time length is any confirmed time length, the first time length indicates that the confirmed time length is greater than or equal to the selected time length, and the second time length indicates that the confirmed time length is less than the selected time length;

[0014] Based on the treatment index of each reference patient corresponding to the first time length and the treatment indices of all patients corresponding to the first time length, combined with the weighted similarity index corresponding to each reference patient corresponding to the first time length, the first fluctuation coefficient corresponding to the first time length is obtained; the second fluctuation coefficient corresponding to the second time length is obtained.

[0015] The boundary probability corresponding to the selected time length is obtained based on the first fluctuation coefficient and the second fluctuation coefficient; there is a positive correlation between the first fluctuation coefficient and the boundary probability, and a negative correlation between the second fluctuation coefficient and the boundary probability;

[0016] The length of time to diagnosis corresponding to the maximum value of all boundary possibilities is taken as the boundary feature length.

[0017] Preferably, the step of obtaining the first fluctuation coefficient corresponding to the first time length based on the treatment index of each reference patient corresponding to the first time length and the treatment indices corresponding to all first time lengths, combined with the weighted similarity index corresponding to each reference patient corresponding to the first time length, specifically includes:

[0018] Based on the weighted similarity index of each reference patient and target patient corresponding to the first time length, the weighted mean of the treatment index of all reference patients corresponding to the first time length is obtained;

[0019] Using the weighted similarity index, the absolute value of the difference between the treatment index of each reference patient corresponding to the first time length and the weighted mean is averaged to obtain the first fluctuation coefficient corresponding to the first time length.

[0020] Preferably, the step of curve fitting of the treatment index for reference patients whose diagnosis time is greater than the cutoff characteristic length, and linear fitting of the treatment index for reference patients whose diagnosis time is less than or equal to the cutoff characteristic length, to obtain the pancreatic islet compensation function curve, specifically includes:

[0021] For reference patients whose diagnosis time is longer than the cutoff characteristic length, curve fitting is performed based on the characteristic distribution of diagnosis time and treatment index for each reference patient to obtain the first functional curve;

[0022] For reference patients whose diagnosis time is less than or equal to the cutoff characteristic length, a linear fit is performed based on the linear relationship between the treatment index and the diagnosis time for each reference patient to obtain the second function curve.

[0023] The first functional curve and the second functional curve constitute the pancreatic islet compensation function curve.

[0024] Preferably, the step of performing curve fitting based on the characteristic distribution of the diagnosis time and treatment index of each reference patient to obtain the first functional curve specifically includes:

[0025] For reference patients whose diagnosis time is longer than the cutoff feature length, a curve fitting expression is constructed using an exponential decay model with diagnosis time as the x-axis and treatment index as the y-axis.

[0026] By using a weighted similarity index between each reference patient and the target patient, the difference between the treatment index and the predicted treatment index for each reference patient is weighted to obtain the curve error;

[0027] Under the condition of minimizing the curve error, the curve fitting expression is solved to obtain the first function curve; where the predictive treatment index is the ordinate obtained by substituting the diagnosis time of each reference patient into the curve fitting expression.

[0028] Preferably, the step of performing linear fitting based on the linear relationship between the treatment index and the time to diagnosis for each reference patient to obtain the second functional curve specifically includes:

[0029] For reference patients whose diagnosis time is less than or equal to the threshold feature length, a second function curve is obtained by fitting a straight line with the diagnosis time of each reference patient as the x-axis and the corresponding treatment index of each reference patient as the y-axis.

[0030] Preferably, the step of obtaining a weighted similarity index between the target patient and each reference patient based on the differences and similarities between the feature data of the target patient and each reference patient in the same dimension specifically includes:

[0031] Based on the differences between the target patient and each reference patient in the feature data of each same dimension, the availability factor of the feature data of each dimension is obtained;

[0032] The similarity between the target patient and each reference patient in each same dimension of feature data is obtained. The similarity is then weighted and summed using the proportion of the availability factor to obtain a weighted similarity index between the target patient and each patient.

[0033] Preferably, the step of obtaining the usability factor of the feature data in each dimension based on the differences between the target patient and each reference patient in the feature data of each same dimension specifically includes:

[0034] For any dimension, obtain the absolute value of the difference between the feature data of the target patient and each reference patient in that dimension, and sum the proportions of the negative correlation coefficients of the absolute values ​​of the differences for all reference patients to obtain the availability factor of the feature data for that dimension.

[0035] Preferably, obtaining the diabetes risk prediction result for the target patient based on the deviation between the treatment index and the pancreatic islet compensation function curve at different time lengths specifically includes:

[0036] The compensatory index of the target patient at each time length is obtained by comparing the difference between the treatment index of the target patient at each time length and the predicted treatment index of the corresponding time length on the pancreatic islet compensatory function curve.

[0037] Compensatory indicators are used as labels for the diabetes risk prediction model for target patients, and are used to train the diabetes risk prediction model.

[0038] Preferably, the step of obtaining the compensatory index of the target patient at each time length based on the difference between the treatment index of the target patient at each time length and the predicted treatment index at the corresponding time length on the pancreatic islet compensation function curve specifically includes:

[0039] By negatively correlating the absolute values ​​of the differences between the treatment index of the target patient at each time length and the predicted treatment index at the corresponding time length on the pancreatic islet compensation function curve, the compensatory index of the target patient at each time length is obtained.

[0040] The embodiments of the present invention have at least the following beneficial effects:

[0041] This invention first collects characteristic data from various dimensions of target patients who have not been diagnosed with diabetes after acute pancreatitis, as well as reference patients who have been diagnosed with diabetes, providing a data foundation for subsequent compensatory feature analysis. Secondly, it quantifies the multimodal feature similarity between the target patients and each reference patient, providing a precise reference sample selection basis for subsequent assessment of the pancreatic islet compensatory function curve of the target patients. Then, it uses the distribution differences of the treatment index to identify time boundaries, combines similarity weights to improve boundary accuracy, and finally selects the boundary feature length from multiple candidate time periods to locate the inflection point of pancreatic islet function transitioning from a slow decline phase to a rapid decline phase. Furthermore, it selects appropriate fitting methods to match different stages with different pathological features, and then comprehensively covers the functional decline cycle through segmented fitting, providing a visualized and quantitative functional evolution model for subsequent assessment of insulin compensation in target patients and prediction of diabetes risk. Finally, for patient samples without a confirmed diabetes diagnosis, the deviation between the actual measured treatment index and the predicted trajectory fully considers the compensatory status of the patient samples before diagnosis, resulting in better accuracy in risk prediction for patients in a compensated state. Attached Figure Description

[0042] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0043] Figure 1 This is a flowchart of the steps of a method for predicting the risk of diabetes after acute pancreatitis based on multimodal data fusion provided by the present invention.

[0044] Figure 2 This is a flowchart of the steps for obtaining the weighted similarity index between the target patient and the reference patient provided by the present invention;

[0045] Figure 3This is a flowchart of the steps for obtaining the boundary feature length provided by the present invention;

[0046] Figure 4 This is a flowchart of the steps for obtaining the pancreatic islet compensation function curve provided by the present invention. Detailed Implementation

[0047] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a multimodal data fusion-based diabetes risk prediction system based on the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0049] The following describes in detail, with reference to the accompanying drawings, a specific scheme of the multimodal data fusion-based risk prediction system for diabetes after acute pancreatitis provided by the present invention. The multimodal data fusion-based risk prediction system for diabetes after acute pancreatitis provided by this embodiment of the invention includes a memory and a processor. The processor executes a computer program stored in the memory to implement the steps of a multimodal data fusion-based risk prediction method for diabetes after acute pancreatitis. The steps of this multimodal data fusion-based risk prediction method for diabetes after acute pancreatitis are described in detail below.

[0050] The specific implementation scenario targeted by this invention is as follows:

[0051] Pancreatic compensatory function refers to the body's ability to maintain blood glucose levels within the normal range by increasing insulin secretion when insulin resistance (decreased tissue sensitivity to insulin) or early, minor damage to pancreatic β-cells occurs. This function is an important defense mechanism for the body in the early stages of abnormal glucose metabolism. It is not an instantaneous response but a dynamic and continuous process, which can be divided into two key stages: the compensatory stage and the decompensatory stage. Ultimately, if the compensatory capacity is exhausted, diabetes will develop.

[0052] In the context of predicting diabetes risk following acute pancreatitis, the significance of pancreatic islet compensation function is particularly prominent: Acute pancreatitis can cause irreversible damage to pancreatic tissue, affecting the function of pancreatic β cells. Patients may initially enter a "compensatory islet function state" (normal blood sugar but β cells are already overworked), rather than directly manifesting as diabetes. Traditional methods based solely on static labels such as whether diabetes has been diagnosed may miss these high-risk individuals, while accurate identification of pancreatic islet compensation function status enables earlier risk intervention.

[0053] Please see Figure 1 The diagram illustrates a flowchart of a method for predicting the risk of diabetes after acute pancreatitis based on multimodal data fusion, according to an embodiment of the present invention. The method includes the following steps:

[0054] Step S100: Obtain the characteristic data and treatment index of each dimension of the target patient and several reference patients after acute pancreatitis; wherein, the target patient is a patient without a confirmed diabetes, the reference patients are patients with a confirmed diabetes, and each reference patient corresponds to a diagnosis time length.

[0055] This embodiment specifically targets a patient who has been diagnosed with acute pancreatitis but not yet diagnosed with diabetes as the target patient for compensatory feature analysis. Then, a patient who has been diagnosed with both acute pancreatitis and diabetes is selected as a reference patient for reference feature analysis.

[0056] To construct the training set, it is necessary to systematically collect longitudinal, multi-dimensional data for each patient sample. The data content covers the acute pancreatitis (AP) attack period, namely the two key stages before and after the patient's discharge. Specifically, the data collection in this embodiment involves two time dimensions.

[0057] Firstly, for both the target patient and each reference patient, feature data across various dimensions are collected before discharge from the hospital for pancreatitis, with each dimension corresponding to one type of data. Secondly, for the target patient, a management index is collected after discharge from the hospital for pancreatitis, and the time interval between the time point at which the management index is detected and the time after discharge from the hospital for pancreatitis is recorded. In this embodiment, the time interval for collecting the management index after the target patient is diagnosed with pancreatitis and discharged from the hospital can be used as the detection time length. For the reference patient, the management index at the time the patient is diagnosed with diabetes after discharge from the hospital for pancreatitis is collected, and the time interval between the time point at which the management index is detected and the time after discharge from the hospital for pancreatitis is recorded as the diagnosis time length for the reference patient.

[0058] It should be understood that the actual time length represents the time interval between the target patient's diagnosis of pancreatitis and the acquisition of the current treatment index, while the diagnosis time length represents the time interval between the reference patient's diagnosis of pancreatitis and the diagnosis of diabetes. In other embodiments, the implementer may also use the time length between the target patient's diagnosis of pancreatitis and their readmission for a follow-up visit as the actual time length, and use the time length between the reference patient's diagnosis of pancreatitis and their diagnosis of diabetes as the diagnosis time length.

[0059] As a concrete example, multimodal feature data can include: structured clinical data, such as AP severity score (modified CT severity index); demographic indicators, such as age, BMI, and sex; and diagnostic indicators, such as pancreatic injury indicators: serum amylase (U / L, normal reference range 25-125), lipase (U / L, normal reference range 0-80); glucose metabolism-related indicators: fasting blood glucose (mmol / L), glycated hemoglobin (HbA1c, %); and inflammatory indicators: white blood cell count (…). ), C-reactive protein (CRP, mg / L). High-throughput omics data, such as serum concentrations of glucose metabolism-related metabolites (e.g., insulin, glucagon). Unstructured imaging data, such as enhanced CT / MRI scans. Unstructured pathology reports, typically including pancreatic histopathological features: degree of fibrosis, type of inflammatory cell infiltration, and number of β cells.

[0060] It should be noted that the core assessment indicator for pancreatic islet compensatory function is the dispensing index (DI value). The DI value comprehensively reflects the secretory function of pancreatic β cells and insulin sensitivity, and its trend directly reflects the dynamic evolution of compensatory function. Specific data varies significantly with different stages of compensation. Therefore, this embodiment mainly focuses on the comparative analysis of the dispensing index to assess the compensatory characteristics of patients.

[0061] It should be further clarified that multimodal feature data corresponds to one dimension. For certain dimensions of feature data that are numerical, such as age, BMI, structured clinical data, various examination indicators, metabolite concentrations, DI values, etc., the differences in the same dimension of feature data between two different patients can be directly assessed using feature analysis methods such as Euclidean distance and Manhattan distance. For parameters that cannot be directly calculated using feature analysis methods such as absolute difference or Euclidean distance, such as unstructured data, imaging reports, pathology reports, and categorical parameters, implementers can choose appropriate methods for difference and similarity assessment based on the specific implementation scenario.

[0062] As a concrete example, for categorized feature data, such as gender and smoking history, numerical encoding can be used. For instance, males are assigned a value of 1, females a value of 0, those with a smoking history a value of 1, and those without a smoking history a value of 0. For image reports, images can be converted into computable numerical vectors, such as statistical features (describing the overall distribution of image pixel values, such as mean and variance); shape features (describing the geometry of the pancreas, such as volume, surface area, and longest diameter); and texture features (describing the spatial distribution of image pixel values, such as contrast and energy of the image's gray-level co-occurrence matrix). Furthermore, the multiple types of image feature values ​​collected from the image can be combined into a numerical vector for feature difference and similarity analysis.

[0063] Pathology reports are text data and need to be converted into numerical vectors through text feature extraction before subsequent feature analysis can be performed. As a concrete example, Natural Language Processing (NLP) techniques can be used to structure the pathology report text, such as using word embedding to convert the text into vectors.

[0064] It should be understood that the methods for obtaining feature data for each dimension are well-known techniques and will not be elaborated upon here. To avoid the impact of different dimensions of multimodal data on data analysis results, the feature data for each dimension can be standardized using the z-score standardization method. It should also be noted that for categorical feature data, such as binary data of 0 / 1, and vector data corresponding to image reports and pathology reports, the data format is fixed and standardization is not required.

[0065] Step S200: Based on the differences and similarities between the target patient and each reference patient in the same dimension of feature data, obtain the weighted similarity index between the target patient and each reference patient.

[0066] The main purpose of this step is to analyze the predictive value of the feature data in each dimension, integrate multi-dimensional similarities to obtain the overall correlation, and quantify the multimodal feature correlation strength between the target patient and each reference patient, so as to provide an accurate reference sample selection basis for subsequent evaluation of the pancreatic islet compensation function curve of the target patient.

[0067] Firstly, different feature dimensions exhibit varying discriminative abilities in predicting diabetes risk. It is necessary to quantify the predictive value of each dimension using an availability factor to avoid irrelevant features interfering with similarity calculations. In predicting diabetes risk after acute pancreatitis, the predictive value of multimodal feature dimensions differs significantly: some dimensions are highly correlated with pancreatic function decline, while others have no significant association with risk. Assuming all feature dimensions are equally important for similarity calculations would dilute the impact of key features with irrelevant features. Therefore, it is necessary to first quantify the actual predictive value of each dimension for the target patient's risk, providing feature importance weights for subsequent weighted similarity calculations.

[0068] Secondly, multi-dimensional similarity needs to be weighted and integrated with availability factors to accurately reflect the overall correlation strength between the target patient and the reference patient. A single-dimensional similarity cannot fully reflect the pathophysiological consistency between the two; therefore, it is necessary to integrate the similarity of all dimensions. However, directly averaging the similarity of all dimensions ignores the differences in predictive value between different dimensions. Therefore, the similarity of each dimension should be weighted using the proportion of availability factors. Dimensions with high availability have a larger similarity weight, contributing more significantly to the overall similarity index; dimensions with low availability have a smaller similarity weight, reducing their interference with the correlation judgment. The similarity index obtained through this weighting method can focus on feature dimensions highly correlated with risk, accurately screening reference patients with similar pathological trajectories to the target patient, providing a reliable data foundation for subsequent inflection point determination and compensation curve fitting.

[0069] In this regard, such as Figure 2 As shown, the method for obtaining the weighted similarity index between the target patient and the reference patient can be implemented by steps S201 and S202.

[0070] Step S201: Based on the differences between the feature data of the target patient and each reference patient in each same dimension, obtain the availability factor of the feature data for each dimension.

[0071] Specifically, for any dimension, the absolute value of the difference between the feature data of the target patient and each reference patient in that dimension is obtained, and the proportion of the negative correlation coefficients of the absolute values ​​of the differences corresponding to all reference patients is accumulated to obtain the availability factor of the feature data of the arbitrary dimension.

[0072] It should be noted that if the proportion of diagnosed diabetes is significantly higher among reference patients with similar values ​​on a certain characteristic dimension to the target patient who has not been diagnosed with diabetes, this provides strong statistical evidence demonstrating a high correlation between the value of that characteristic dimension and the risk of developing diabetes after acute pancreatitis. This can be explained in two ways: First, this characteristic dimension possesses strong risk predictive power and can serve as a core predictor. Its different values ​​can clearly distinguish patient groups with significantly different risks of developing diabetes; that is, the distribution of values ​​on this characteristic dimension can effectively differentiate between high-risk and low-risk groups. Second, and more importantly, this statistical result can indirectly infer the risk status of the target patient who has not been diagnosed with diabetes. Given that the target patient's value on this characteristic dimension is similar to that of the already identified high-risk group, and that this characteristic is highly correlated with the risk of developing diabetes, it can be determined that the target patient is also highly likely to be at high risk for diabetes.

[0073] Based on the above logic, the following formula can be constructed to characterize the first digit in multimodal data. l The usability of feature data from any one dimension to the target patient is illustrated in this embodiment, which uses feature data from any one dimension as an example. l The method for obtaining the availability factor of feature data in each dimension can be expressed by the formula:

[0074] ;

[0075] in, Indicates the middle l The usability factor of the feature data in each dimension for the target patient, where i represents the target patient. Indicates the target patient in the first l Feature data in each dimension This indicates that the nth reference patient is in the nth... l Feature data in each dimension This indicates the total number of reference patients. This represents an exponential function with the natural constant e as its base.

[0076] Indicates the target patient and reference patient at the 1st l The differences between feature data in each dimension, and thus The negative correlation coefficient represents the difference in characteristic data. This represents the cumulative difference in feature data between the target patient and all reference patients on the same dimension. This represents the proportion of negative correlation coefficients between the target patient and the nth reference patient in the same dimension of feature data.

[0077] For each dimension of feature data, the smaller the difference between the reference patient and the target patient's data, the greater the usability of the reference patient's feature data in that dimension for the analysis of the target patient's compensatory characteristics. By comprehensively considering the data differences among all reference patients, an usability factor is used to assess the usability of the feature data for each dimension. To a certain extent, the usability factor characterizes the discriminative ability of the feature data for each dimension in differentiating the diabetes outcomes of the target patients.

[0078] Step S202: Obtain the similarity between the target patient and each reference patient in each same dimension of feature data. Using the proportion of the availability factor, perform a weighted summation on the similarity to obtain a weighted similarity index between the target patient and each patient.

[0079] It should be understood that the higher the usability factor value corresponding to the feature data of each dimension, the greater the usability, importance, or reference value of the feature data of that dimension. In order to prevent the similarity calculation from being "diluted" or "interfered" by a large number of irrelevant features, the usability factor of the feature data of all dimensions of multimodal data for the target patients who have not been diagnosed with diabetes is used as the weight. The similarity between each reference patient sample and the target patient under all feature dimensions is weighted and processed, and the weighted similarity index can be obtained by the following calculation method.

[0080] As a concrete example, taking any reference patient as an illustration, the method for obtaining the weighted similarity index between the target patient and the nth reference patient can be expressed by the formula:

[0081] ;

[0082] in, This represents the weighted similarity index between the target patient and the nth reference patient. Indicates the middle l The usability factor of the feature data in each dimension for the target patient, where i represents the target patient. Indicates the number of dimensions in the feature data. This indicates that the target patient and the nth reference patient are at the 1st... l Similarity between feature data in each dimension.

[0083] It should be noted that well-known techniques can be used to quantify similarity. For example, cosine similarity is not restricted here, and implementers can choose according to the specific implementation scenario.

[0084] Indicates the first lThe initial similarity is weighted by the proportion of availability factors in each dimension of feature data. Dimensions with high availability have a larger similarity weight and contribute more significantly to the overall similarity index; dimensions with low availability have a smaller similarity weight to reduce their interference with the correlation judgment, so that the final weighted similarity index can more accurately assess the risk correlation between each reference patient and the target patient.

[0085] In other embodiments, the implementer may also have relevant medical personnel set weight values ​​for the feature data of each dimension, and then use the weights to perform a weighted summation of the initial similarity to obtain a weighted similarity index. This is not the focus of this invention and will not be elaborated here.

[0086] Step S300: Based on the treatment index of each reference patient and the distribution differences of the overall reference patients, the length of each diagnosis time is screened in combination with the weighted similarity index to obtain the boundary feature length.

[0087] The main purpose of this step is to identify time boundaries by utilizing the distribution differences of the treatment index, improve the accuracy of the boundaries by combining similarity weights, and finally select the optimal inflection point from multiple candidate times. That is, the screening process of decomposing feature length aims to locate the time inflection point when pancreatic function enters the rapid decline phase from the slow decline phase.

[0088] Considering that the compensatory process of insulin typically presents two distinct phases: 1. Slow decline phase: In this phase, although insulin secretion function may have been slightly impaired due to long-term high load, β cells compensate by increasing secretion, allowing overall pancreatic function (represented by the DI value) to remain relatively stable or only show a slow, near-linear decline. 2. Rapid decline phase: As compensatory potential is exhausted, secretory function begins to deteriorate rapidly. Coupled with the continued existence or further aggravation of insulin resistance, the DI value enters a non-linear phase of accelerated decline, exhibiting an exponential decay trend. (There is usually a third plateau-like low-level phase, but since most patients have already been diagnosed with diabetes before this, it is not included in this discussion).

[0089] Based on the aforementioned dynamic characteristics, significant statistical differences emerge on either side of the inflection point at the boundary between the two periods: the DI value changes rapidly and fluctuates greatly on the right side of the inflection point (rapid decline phase), while the DI value changes more gradually on the left side of the inflection point (slow decline phase). Based on these differences, the time inflection point at the boundary between the two periods can be accurately located by assessing the data fluctuations corresponding to each confirmed diagnosis time period.

[0090] In this regard, such as Figure 3 As shown, the method for obtaining the boundary feature length can be implemented by steps S301 to S304.

[0091] Step S301: Obtain the first time length and the second time length of the selected time length. The selected time length is any confirmed time length. The first time length indicates that the confirmed time length is greater than or equal to the selected time length, and the second time length indicates that the confirmed time length is less than the selected time length.

[0092] First, using the length of each confirmed diagnosis as a time boundary, the length of the confirmed diagnosis for all reference patients is divided into two data sets: one set represents the slow decline period, and the other set represents the rapid decline period.

[0093] Assuming the selected time length is the optimal period boundary, the reference patients corresponding to the first time length represent the reference sample in the rapid decline phase, and the reference patients corresponding to the second time length represent the reference sample in the slow decline phase.

[0094] It should be understood that there may be multiple second time lengths to the left of the selected time length (the side where the confirmed time length is less than the selected time length), and there may be multiple first time lengths to the right of the selected time length (the side where the confirmed time length is greater than or equal to the selected time length).

[0095] Step S302: Based on the treatment index of each reference patient corresponding to the first time length and all treatment indices corresponding to the first time length, and combined with the weighted similarity index corresponding to each reference patient corresponding to the first time length, obtain the first fluctuation coefficient corresponding to the first time length; obtain the second fluctuation coefficient corresponding to the second time length.

[0096] The two-stage nature of pancreatic islet function decline (i.e., a slow decline phase and a rapid decline phase) manifests as fluctuations in the insulin resistance index (DI) value across different time intervals. This difference needs to be quantified using a fluctuation coefficient to pinpoint potential boundaries. During the slow decline phase, β-cells compensate to maintain a relatively stable DI value, resulting in small fluctuations. During the rapid decline phase, compensation is exhausted, leading to accelerated DI decay and large fluctuations. If a certain time since diagnosis is used as a boundary, dividing reference patients into pre-boundary (diagnosis time ≤ this time) and post-boundary (diagnosis time > this time) groups, the ideal boundary should satisfy the condition that the pre-boundary group has small DI value fluctuations and the post-boundary group has large fluctuations. The fluctuation coefficient for both groups needs to be calculated. This fluctuation coefficient integrates the dispersion of DI values ​​across all reference patients within a group, and assigns higher weights to reference patients with high similarity to the target patient through a weighted similarity index. This ensures that the fluctuation coefficient more closely reflects the actual evolution of pancreatic islet function in the target patient, providing a quantitative basis for subsequent assessment of boundary probability.

[0097] Specifically, the first step is to obtain the weighted mean of the treatment index of all reference patients corresponding to the first time length, based on the weighted similarity index of each reference patient and the target patient corresponding to the first time length.

[0098] In this embodiment, in order to make the feature analysis more referential, the proportion of weighted similarity indicators between each reference patient and the target patient corresponding to the first time length is used as the weight of each reference patient corresponding to the first time length. The weights are then used to calculate the weighted average of the treatment indices of all reference patients corresponding to the first time length to obtain the weighted average of the treatment indices of all reference patients corresponding to the first time length.

[0099] The second step involves using the weighted similarity index to calculate the weighted average of the absolute values ​​of the differences between the treatment index of each reference patient corresponding to the first time length and the weighted mean, thereby obtaining the first fluctuation coefficient corresponding to the first time length.

[0100] As a concrete example, the method for obtaining the first fluctuation coefficient corresponding to the first time length can be expressed by the formula:

[0101] ;

[0102] in, This represents the first fluctuation coefficient corresponding to the first time length. Indicates the length of time since diagnosis. Indicates the selected time length. This indicates the first time length, which is the time length for diagnosis when the diagnosis time is greater than or equal to the selected time length. This represents the weighted similarity index between the m-th reference patient and the target patient in the first patient set. This represents the treatment index, or DI value, for the m-th reference patient in the first patient set. This represents the weighted mean of the treatment indices of all reference patients in the first patient set. This represents the number of reference patients included in the first patient set. The first patient set is the set of all reference patients whose diagnosis time is longer than the selected time length.

[0103] It should be noted that if the denominator A value of 0 indicates that all weighted similarity indicators are 0, which is an extreme phenomenon, meaning there are no similar reference patients. The first volatility coefficient can be set to 0 to indicate extremely low volatility.

[0104] This indicates the difference between the reference patients in the first patient set and the overall distribution. The weighted difference between the two is represented by the treatment index. For reference patients with high similarity, the characteristic difference results of the treatment index are more valuable for reference.

[0105] It should be noted that the second fluctuation coefficient for the second time length can be obtained by using the same calculation method as the first fluctuation coefficient for the first time length. That is, all the relevant data of the reference patients corresponding to the first time length in the above calculation method can be replaced with the relevant data of the reference patients corresponding to the second time length. This will not be elaborated further here.

[0106] Ultimately, the first fluctuation coefficient of the first time length characterizes the fluctuation of the DI value of the treatment index for all reference patients whose diagnosis time length is greater than or equal to the selected time length. The larger the value, the greater the possibility that the selected time length is the time inflection point of the stage, and the more consistent it is with the fluctuation characteristics of the data.

[0107] The second fluctuation coefficient of the second time length characterizes the fluctuation of the treatment index (DI) value of all reference patients whose diagnosis time length is less than the selected time length. The smaller the value, the greater the possibility that the selected time length is the time inflection point of the stage, and the more consistent it is with the fluctuation characteristics of the data.

[0108] Step S303: Obtain the boundary probability corresponding to the selected time length based on the first fluctuation coefficient and the second fluctuation coefficient; the first fluctuation coefficient and the boundary probability are positively correlated, and the second fluctuation coefficient and the boundary probability are negatively correlated.

[0109] The comparison between the fluctuation coefficients corresponding to the first and second time lengths directly determines the rationality of using a certain diagnostic time length as a cutoff point. This relationship needs to be transformed into a comparable indicator through cutoff probability. According to pathological logic, a high-quality cutoff point should have the characteristics of small fluctuations before the cutoff (low first fluctuation coefficient) and large fluctuations after the cutoff (high second fluctuation coefficient). This combination of low and high fluctuations precisely corresponds to the transition of pancreatic islet function from stable compensation to rapid decompensation. Therefore, it is necessary to construct a cutoff probability index, coupling the first and second fluctuation coefficients to obtain decomposition probability. The greater the cutoff probability, the more the difference in fluctuations between the two sets of DI values ​​conforms to the two-stage characteristics when the current diagnostic time length is used as the cutoff point, and the more likely it is to be a true time inflection point.

[0110] As a concrete example, the formula for calculating the probability of a boundary corresponding to a selected time length can be expressed as:

[0111] ;

[0112] in, Indicates the selected time length The corresponding dividing probability, This represents the first fluctuation coefficient corresponding to the first time length. This represents the second fluctuation coefficient corresponding to the second time length. Indicates the length of time since diagnosis. Indicates the selected time length. This indicates the first time length, which is the time length for diagnosis when the diagnosis time is greater than or equal to the selected time length. This indicates the second time length, which is the time length for diagnosis when the diagnosis time is shorter than the selected time length. It is a very small positive number, and in this embodiment it is taken as 0.1, in order to avoid the second fluctuation coefficient being 0, which would make the fraction meaningless and thus affect the calculation results.

[0113] Step S304: The length of time for diagnosis corresponding to the maximum value of all boundary probabilities is taken as the boundary feature length.

[0114] The larger the first fluctuation coefficient and the smaller the second fluctuation coefficient, the greater the likelihood of a boundary. This indicates that when the current confirmed diagnosis time is used as the boundary point, the difference in fluctuations between the two sets of DI values ​​more closely matches the two-stage characteristics, and is more likely to be a true time inflection point. Conversely, the smaller the first fluctuation coefficient and the larger the second fluctuation coefficient, the smaller the likelihood of a boundary. This indicates that when the current confirmed diagnosis time is used as the boundary point, the difference in fluctuations between the two sets of DI values ​​less closely matches the two-stage characteristics, and is less likely to be a true time inflection point.

[0115] Therefore, when taking the maximum value of the boundary probability corresponding to all diagnosis time lengths, using the diagnosis time length as the time boundary point is the optimal choice. That is, the boundary characteristic length is the time node that best conforms to the two-stage transition law of pancreatic islet function at the population level, which can ensure to the greatest extent that the subsequent fitted slow decline curve and rapid decline curve fit the pathological evolution trajectory of the target patient.

[0116] Step S400: Perform curve fitting on the treatment index of reference patients whose diagnosis time is greater than the threshold characteristic length, and perform linear fitting on the treatment index of reference patients whose diagnosis time is less than or equal to the threshold characteristic length to obtain the pancreatic islet compensation function curve.

[0117] The core of the pancreatic islet compensatory function curve is to reconstruct the dynamic evolution trajectory of pancreatic islet function from compensated stability to decompensated deterioration. The fitting process of the pancreatic islet compensatory function curve mainly involves matching the pathological characteristics of different stages and selecting appropriate fitting methods. Then, by segmenting the fitting, the functional decline cycle is fully covered, providing a visualized and quantitative functional evolution model for subsequent assessment of insulin compensation in target patients and prediction of diabetes risk.

[0118] The two-stage pathological characteristics of pancreatic islet function decline necessitate differentiated fitting methods for different time intervals to accurately reflect the patterns of functional changes. A complete pancreatic islet compensatory function curve needs to be constructed through a combination of piecewise fitting and curve integration to avoid gaps in stage coverage caused by a single fitting method.

[0119] In this regard, such as Figure 4 As shown, the method for obtaining the pancreatic islet compensation function curve can be implemented by steps S401 to S403.

[0120] Step S401: For reference patients whose diagnosis time is longer than the threshold feature length, curve fitting is performed based on the characteristic distribution of diagnosis time and treatment index for each reference patient to obtain the first functional curve.

[0121] The core pathological feature of the rapid decline phase is the depletion of the compensatory potential of pancreatic β cells, with the DI value exhibiting an exponentially accelerating decline (distinct from the linear, gradual decline phase). The functional evolution trajectory during this phase can only be reflected through real data from diagnosed diabetic patients. Therefore, the principle behind constructing the curve for the rapid decline phase can be understood as follows: based on the assumption that patients with similar multimodal characteristics have similar pathophysiological trajectories, the DI values ​​and time data of reference patients with diagnosed diabetes highly similar to the target patient are weighted and fitted to transform them into an individualized exponential curve for the target patient, predicting the functional decline path if the target patient enters the decompensated phase. It can also be simply understood as a reuse process of the DI value change trajectory of diagnosed patients similar to the target patient. Therefore, a nonlinear model can be used to capture the "slow in the early stage, rapid in the later stage" decay pattern of the DI value, accurately reconstructing the functional collapse process in the decompensated phase and ensuring that the fitting results are consistent with the pathological mechanism of the decompensated phase.

[0122] As described above, during the rapid decline phase, with the depletion of compensatory potential, secretory function begins to deteriorate rapidly. Coupled with persistent or further aggravated insulin resistance, the insulin resistance index (DI) enters a non-linear phase of accelerated decline, exhibiting an exponential decay trend. Therefore, the following functional expression can be constructed to represent the rapid decline phase in the target patient. The pancreatic islet compensation function curve, also known as the first function curve, in which... Indicates the length of time since diagnosis. Indicates the length of the boundary feature.

[0123] Specifically, the first step is to construct a curve fitting expression using an exponential decay model for reference patients whose diagnosis time is longer than the threshold characteristic length, with the diagnosis time as the x-axis and the treatment index as the y-axis.

[0124] As a concrete example, when At that time, the curve fitting expression corresponding to the first function curve adopts an exponential decay model, which can be expressed by the formula:

[0125] ;

[0126] in, This represents the fitting function corresponding to the first function curve of the target patient, which is also the curve fitting expression; Indicates the length of time since diagnosis. All parameters are unknown. It is a natural constant.

[0127] It should be understandable that These are all curve parameters to be solved. Indicates the initial attenuation magnitude. Indicates the decay rate. Indicates the time offset. The lower limit of the DI value can be obtained from subsequent steps.

[0128] The second step involves using a weighted similarity index between each reference patient and the target patient to weight the difference between the treatment index and the predicted treatment index for each reference patient, thus obtaining the curve error. The predicted treatment index is the ordinate obtained by substituting the diagnosis time of each reference patient into the curve fitting expression.

[0129] By using weighted error, we can ensure that reference samples with high similarity have a greater impact on curve parameters. As a concrete example, the method for obtaining curve error can be expressed by the formula:

[0130] ;

[0131] in, Indicates curve error. This represents the weighted similarity index between the m-th reference patient and the target patient in the first patient set. This represents the treatment index, or DI value, for the m-th reference patient in the first patient set. This represents the predictive treatment index for the m-th reference patient in the first patient set. This represents the number of reference patients included in the first patient set.

[0132] The greater the similarity, the greater the contribution of the corresponding reference patient's data to the error. After weighting, the error of highly similar reference patients (whose pathological trajectory is closer to that of the target patient) is amplified, while the error of low-similarity reference patients is suppressed, ensuring that the curve more closely matches the potential decline trend of the target patient. This represents the length of time from diagnosis for the m-th reference patient. Substituting the above curve fitting expression, we get, which is... .

[0133] The third step is to solve the curve fitting expression under the condition of minimizing the curve error, and obtain the first functional curve.

[0134] Specifically, the sum of errors, i.e., the curve error, is minimized through a numerical optimization algorithm. Solving for the curve error Minimum parameter The numerical optimization algorithm can be gradient descent, L-BFGS, etc., and the implementer can choose according to the specific implementation scenario. This solution process is a well-known technology and will not be described in detail here.

[0135] It should be noted that when solving for the curve function, constraints can also be set: the known pancreatic islet compensation function curve must satisfy the characteristic of continuity at the inflection point, that is, the first function curve (rapid decline phase) and the second function curve (slow decline phase). The DI values ​​at each location are consistent, therefore they can be... Substituting into the curve fitting expression, we obtain , The value of is known and can be derived from the linear model of the second function curve (slow descent period). This constraint reduces the degrees of freedom in parameter solving and improves the stability of the results.

[0136] Thus, the construction of the first functional curve is based on the pathological trajectory of similar confirmed patients. The accelerated decline of the DI value is characterized by an exponential decay model. The error calculation adopts a weighted sum of squares, which not only ensures the dominance of highly similar samples, but also effectively punishes abnormal data. Finally, a complete logical closed loop of pathological features, model selection, error constraints and parameter solution is formed, providing an accurate quantitative tool for predicting pancreatic function decline in undiagnosed patients.

[0137] Step S402: For reference patients whose diagnosis time is less than or equal to the cutoff characteristic length, a linear fit is performed based on the linear relationship between the treatment index and diagnosis time for each reference patient to obtain the second function curve.

[0138] The core pathophysiological characteristic of the slow decline phase is that pancreatic β cells compensate by increasing insulin secretion to maintain normal blood glucose levels. However, prolonged high insulin load leads to slight impairment of secretory function, and the DI value shows an approximately linear, slow decline, lacking both the nonlinear characteristics of accelerated decline (distinguishing it from the rapid decline phase) and dramatic fluctuations (distinguishing it from detection errors). Therefore, a univariate linear model is chosen as the curve form.

[0139] It should be noted that the uniform decrease characteristic of the linear model is completely consistent with the clinically observed pattern of a small decrease in the DI value per unit time during the compensatory period, which can accurately reflect the process of slow depletion of β-cell function; the linear model contains only 2 unknown parameters, does not require complex nonlinear optimization, and the physical meaning of the parameters is clear, which is convenient for clinical interpretation.

[0140] Specifically, for reference patients whose diagnosis time is less than or equal to the threshold feature length, a second function curve is obtained by fitting a straight line with the diagnosis time of each reference patient as the x-axis and the corresponding treatment index of each reference patient as the y-axis.

[0141] As a concrete example, when When the second function curve is in its second state, its functional expression can be expressed as:

[0142] ;

[0143] in, The function expression representing the second function curve of the target patient. Indicates the length of time since diagnosis. and For parameters to be solved, The slope This is the intercept.

[0144] The main parameters for solving this function expression come from two key anchor data points: one is the DI value of undiagnosed target patients during an acute pancreatitis attack, which is... The first set of data is the DI value at the time of inflection, reflecting the initial pancreatic islet function state; the second set is the inflection point data, which is the DI value of the reference patient corresponding to the boundary characteristic length, reflecting the critical functional state during the transition from the compensated to the decompensated stage. Substituting these two sets of data into the above function expression yields the second functional curve.

[0145] The core of constructing the second functional curve, also known as the slow decline curve, lies in its independence from post-discharge follow-up data of the target patient. Instead, the curve is fitted using the patient's own baseline data combined with inflection points derived from similar diagnosed patients. This approach ensures individualization through the target patient's baseline data, ensures the reasonableness of the trend through inflection point data from similar diagnosed patients, and balances fitting accuracy and computational efficiency through a linear model. This logic aligns with the pathophysiological laws of pancreatic islet compensation function and addresses the practical problem of insufficient clinical data, providing a precise quantitative tool for the early prediction of diabetes risk after acute pancreatitis.

[0146] Step S403: The first functional curve and the second functional curve constitute the pancreatic islet compensation function curve.

[0147] By concatenating the first and second functional curves, the pancreatic islet compensation function curve is obtained. This curve fully reflects the evolutionary trajectory of pancreatic islet function in the target patient from the compensated to the decompensated stage. The above process no longer treats undiagnosed diabetes patients as isolated "points" (i.e., whether diabetes is diagnosed at a certain point in time), but rather as a "trajectory" evolving over time, constructing a pancreatic islet compensation function curve for diagnosed diabetes patients. For any undiagnosed patient, their actual monitored treatment indices are compared with this predicted trajectory. If the changes in these indices closely match the high-risk decay curve, it strongly suggests that the patient is progressing along a typical decompensated path; although their physiological state has not reached the diagnostic threshold for diabetes, it has already shown a clear trend of deterioration.

[0148] Step S500: Based on the deviation between the treatment index and the pancreatic islet compensation function curve of the target patient at different time lengths, obtain the diabetes risk prediction results for the target patient.

[0149] A new quantitative label can be assigned to patient samples without a confirmed diagnosis of diabetes. This label is no longer a simple "0" or "1", but rather the degree of match between their actual data and the predicted trajectory (this degree of match is the compensatory nature of insulin). Specifically, patients with high insulin compensatory nature (i.e., high trajectory fit) are assigned a high-risk label, meaning that their pancreatic function is following a declining trajectory; while patient samples with low insulin compensatory nature are assigned a low-risk label, indicating that their function is relatively stable.

[0150] Based on the above description, the compensatory index of the target patient at each time length can be obtained by comparing the difference between the treatment index of the target patient at each time length and the predicted treatment index at the corresponding time length on the pancreatic islet compensatory function curve.

[0151] As a concrete example, the following formula can be constructed to represent the insulin compensation in target patients without a confirmed diagnosis of diabetes, that is, a compensatory indicator:

[0152] ;

[0153] in, Indicators representing compensatory mechanisms in the target patient. This indicates the target patient's time duration (in this example, the actual time duration). The disposal index (DI value) under ) This represents the predictive intervention index (which represents the actual time length) for the target patient in the pancreatic islet compensation function curve. (Originally obtained by substituting into the pancreatic islet compensation function curve) This represents an exponential function with the natural constant e as its base.

[0154] This represents the difference between the actual DI value and the predicted DI value. The smaller the value, the better the DI value for the target patient within that time period. The high insulin compensatory rate indicates that the target patient's index change trajectory closely matches that of a diagnosed diabetic patient, suggesting that the target patient is following a decline trajectory that would lead to a diagnosis of diabetes, and thus the target patient has a higher risk.

[0155] Compensatory indices are used as labels in a diabetes risk prediction model for target patients, and are used to train the model. In this embodiment, compensatory indices are used as labels in the training process of the random forest regression model. A higher label value indicates a higher risk of diabetes for the corresponding patient, while a lower label value indicates a lower risk. It should be understood that the label value for patients already diagnosed with diabetes is fixed at 1.

[0156] It's important to note that the goal of the random forest regressor model is to minimize the mean squared error between the predicted hazard value and the label value, ensuring that the model's predicted hazard value is as close as possible to the patient's actual pancreatic compensation. The model's input consists of feature data across various dimensions, and its output is a hazard value, representing the patient's risk of developing diabetes. Relevant healthcare professionals can then use the results for further assessments and other actions.

[0157] It is understandable that the existence and decline of compensatory indicators provide crucial signals for the early identification and intervention of diabetes and its complications. This not only helps individuals pay attention to their own glucose metabolism health earlier but also provides important evidence for precise clinical prevention and control of diabetes. By capturing compensatory signals, the focus of disease prevention and control can be shifted from when blood sugar rises to when compensation begins, achieving early detection, early intervention, and early benefits.

[0158] It should be noted that the training process of the random forest regressor model is a well-known technique, and will not be elaborated on further here.

[0159] In summary, compared to existing technologies where the training of the original model relies on a static label—that is, whether a patient is diagnosed with diabetes at an uncertain time point after an acute pancreatitis (AP) attack—monitoring during the patient's pancreatic islet function compensation period can incorrectly label them as "non-diabetic patients," thus failing to identify high-risk individuals whose pancreatic islet function is declining but whose blood sugar is not yet excessive. This invention, in its embodiments, for patient samples without a confirmed diabetes diagnosis, determines insulin compensation as the label for these patient samples to correct for the impact of pancreatic islet function compensation, and then trains a random forest regression model to predict the risk of diabetes after acute pancreatitis in newly diagnosed patients.

[0160] Based on patients with similar acute-phase multimodal characteristics (such as clinical indicators and imaging manifestations), the severity of pancreatic damage is similar to their baseline physiological state. Therefore, the trajectory of pancreatic function decline following acute pancreatitis (AP) should also be similar. A dynamic labeling method is proposed to predict the pancreatic compensatory function curve for patients without a confirmed diabetes diagnosis. The compensatory process of insulin generally consists of two phases: 1. Slow decline phase: In this phase, although insulin secretion function may begin to show slight impairment due to prolonged high load, the overall insulin resistance (DI) value may remain relatively stable or only decline slowly and approximately linearly through compensatory increases in secretion. 2. Accelerated decline phase: In this phase, indicators representing secretory function decline sharply, while insulin resistance may still exist or even worsen. Therefore, the DI value will show an exponential trend of accelerated decline. Therefore, firstly, based on multimodal data matching, the matching degree between all other patient samples and each patient sample's multimodal data is obtained. Then, the curve of DI value decreasing over time is obtained using all matched patient sample data. Finally, by the degree of matching with the curve, the compensatory rate of each insulin can be obtained, and this can be used as a label for patient samples without diabetes diagnosis.

[0161] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system comprising a memory and a processor, characterized in that, The processor executes the computer program stored in the memory to implement the following steps: Obtaining feature data of each dimension and a treatment index of a target patient and a plurality of reference patients after acute pancreatitis; wherein the target patient is a patient without diagnosed diabetes, and the reference patients are patients with diagnosed diabetes, and each reference patient corresponds to a diagnosed time length; According to the difference and similarity between the feature data of the target patient and each reference patient in the same dimension, a weighted similarity index between the target patient and each reference patient is obtained; According to the treatment index of each reference patient and the distribution difference of the overall reference patients, and in combination with the weighted similarity index, each diagnosed time length is screened to obtain a boundary feature length; The treatment index of the reference patient with the diagnosed time length greater than the boundary feature length is curve fitted, and the treatment index of the reference patient with the diagnosed time length less than or equal to the boundary feature length is linearly fitted to obtain an islet compensation function curve; According to the deviation between the treatment index of the target patient at different time lengths and the islet compensation function curve, a diabetes risk prediction result of the target patient is obtained; The screening of each diagnosed time length according to the treatment index of each reference patient and the distribution difference of the overall reference patients, in combination with the weighted similarity index, to obtain a boundary feature length, specifically includes: Obtaining a first time length and a second time length of a selected time length, the selected time length being any one of the diagnosed time lengths, the first time length representing the diagnosed time length greater than or equal to the selected time length, and the second time length representing the diagnosed time length less than the selected time length; According to the treatment index of each reference patient corresponding to the first time length and the treatment index of all the first time lengths, in combination with the weighted similarity index corresponding to each reference patient corresponding to the first time length, a first fluctuation coefficient corresponding to the first time length is obtained; and a second fluctuation coefficient corresponding to the second time length is obtained; According to the first fluctuation coefficient and the second fluctuation coefficient, a boundary possibility corresponding to the selected time length is obtained; the first fluctuation coefficient and the boundary possibility have a positive correlation, and the second fluctuation coefficient and the boundary possibility have a negative correlation; The diagnosed time length corresponding to the maximum value of all the boundary possibilities is taken as the boundary feature length; The method for obtaining the first fluctuation coefficient corresponding to the first time length includes: Based on the weighted similarity index of each reference patient corresponding to the first time length and the target patient, a weighted mean value of the treatment index of all the reference patients corresponding to the first time length is obtained; The absolute value of the difference between the treatment index of each reference patient corresponding to the first time length and the weighted mean value is weighted and averaged to obtain the first fluctuation coefficient corresponding to the first time length.

2. The multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system as claimed in claim 1, wherein, The method for obtaining the islet compensation function curve includes: For the reference patients with a confirmed time length greater than the boundary feature length, a first function curve is obtained according to the feature distribution of the confirmed time length and the treatment index of each reference patient; For the reference patients with a confirmed time length less than or equal to the boundary feature length, a second function curve is obtained according to the linear relationship between the treatment index and the confirmed time length of each reference patient; The first function curve and the second function curve constitute an islet compensation function curve.

3. The multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system as claimed in claim 2, wherein, The first function curve is obtained by curve fitting according to the feature distribution of the confirmed time length and the treatment index of each reference patient, and specifically includes: For the reference patients with a confirmed time length greater than the boundary feature length, the confirmed time length is taken as the abscissa, the treatment index is taken as the ordinate, and an exponential decay model is used to construct a curve fitting expression; The difference between the treatment index and the predicted treatment index of each reference patient is weighted by using the weighted similarity index between each reference patient and the target patient, to obtain a curve error; Under the condition of the minimum curve error, the curve fitting expression is solved to obtain the first function curve; wherein the predicted treatment index is the ordinate obtained by inputting the confirmed time length of each reference patient into the curve fitting expression.

4. The multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system as claimed in claim 2, wherein, The second function curve is obtained by linear fitting according to the linear relationship between the treatment index and the confirmed time length of each reference patient, and specifically includes: For the reference patients with a confirmed time length less than or equal to the boundary feature length, the confirmed time length of each reference patient is taken as the abscissa, and the corresponding treatment index of each reference patient is taken as the ordinate, to obtain the second function curve by linear fitting.

5. The multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system as claimed in claim 1, wherein, The weighted similarity index between the target patient and each reference patient is obtained according to the difference and similarity between the feature data of the same dimension of the target patient and each reference patient, and specifically includes: The availability factor of the feature data of each dimension is obtained according to the difference between the feature data of the same dimension of the target patient and each reference patient; The similarity between the feature data of each same dimension of the target patient and each reference patient is obtained, the similarity is weighted and summed by using the proportion of the availability factor, to obtain the weighted similarity index between the target patient and each patient.

6. The multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system as claimed in claim 5, wherein, The availability factor of the feature data of each dimension is obtained according to the difference and similarity between the feature data of the same dimension of the target patient and each reference patient, and specifically includes: For any one dimension, the absolute value of the difference between the feature data of the target patient and each reference patient in the dimension is obtained, and the proportion of the negative correlation coefficient of the absolute value of the difference corresponding to all reference patients is accumulated to obtain the availability factor of the feature data of the arbitrary dimension.

7. The multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system as claimed in claim 1, wherein, The diabetes risk prediction result of the target patient is obtained according to the deviation between the treatment index of the target patient at different time lengths and the islet compensation function curve, and specifically includes: The compensatory index of the target patient at each time length is obtained according to the difference between the treatment index of the target patient at each time length and the predicted treatment index of the corresponding time length on the islet compensation function curve; The compensatory index is used as a label of a diabetes risk prediction model of the target patient, and is used for training the diabetes risk prediction model.

8. The multi-modal data fusion based post-acute pancreatitis diabetes risk prediction system as claimed in claim 7, wherein, The compensatory index of the target patient at each time length is obtained according to the difference between the treatment index of the target patient at each time length and the predicted treatment index of the islet compensation function curve corresponding to the time length, and specifically comprises: The compensatory index of the target patient at each time length is obtained by performing negative correlation processing on the absolute value of the difference between the treatment index of the target patient at each time length and the predicted treatment index of the islet compensation function curve corresponding to the time length.

Images