Modeling method of diabetic complication prediction model

Abstract

The invention relates to the technical field of diabetic complication prediction, and discloses a modeling method of a diabetic complication prediction model. The method comprises the following steps: collecting clinical index data of a patient, including blood glucose level, blood pressure value, blood fat content and renal function parameters, and performing standardization processing to generate a standard data set; time sequence features in the data set are extracted, the short-term fluctuation trend and the long-term change rule of each index are calculated through a sliding window algorithm, and a time sequence feature vector is generated; and inputting the vector into a deep feature extraction network formed by multilayer convolution and long and short term memory units, and extracting high-dimensional time sequence feature representation. Based on the representation, a hierarchical clustering algorithm is used for grouping patient groups, and personalized complication risk prediction sub-models are constructed for different groups. And integrating the outputs of all the personalized sub-models, and generating a final diabetic complication prediction result through a weighted voting mechanism.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of diabetes complication prediction, in particular to a modeling method of a diabetes complication prediction model. BACKGROUND

[0002] Diabetes is a complex chronic metabolic disease, which is often accompanied by multiple serious complications such as diabetic nephropathy, retinopathy, neuropathy and cardiovascular disease during the development process. These complications are the main causes of disability and death of diabetic patients, and bring heavy burden to patients, families and social medical system. At present, the monitoring and early warning of diabetes complications in clinic mainly rely on regular follow-up and threshold judgment of single physiological index. For example, the level of glycosylated hemoglobin is detected to evaluate the recent blood glucose control, and the urine microalbumin is detected to screen the risk of early diabetic nephropathy. Although this method can identify patients with obvious abnormalities to a certain extent, its early warning ability has significant lag.

[0003] Existing complication risk assessment tools are mostly based on static statistical models established by cross-sectional studies, such as using logistic regression analysis to determine the association strength between several risk factors and complications. Such models usually rely on static data at a certain time point, ignoring the fact that diabetes is a chronic progressive disease, and the occurrence and development of its complications is a dynamic and continuous process. The fluctuation pattern, change trend and interaction between blood glucose, blood pressure, blood lipid and other clinical indicators over time may contain more and earlier risk information than the absolute value of a single detection. However, traditional risk assessment methods are difficult to effectively capture and utilize these time series dynamic characteristics.

[0004] In terms of technology, some studies have tried to apply machine learning methods for disease prediction, but most models still treat multiple measurement data of patients as independent samples for processing, or only perform simple linear fitting of trends, and fail to deeply mine the complex patterns contained in high-dimensional time series data. The pathophysiological mechanism of diabetes complications involves long-term and cumulative damage to multiple metabolic pathways, which may manifest as specific nonlinear evolution rules in clinical indicators. Conventional machine learning models have limited representation ability when dealing with such long-range dependent and multi-variable interactive time series signals. In addition, the population of diabetic patients has high heterogeneity, and different individuals may have great differences in complication progression speed and type even if they have similar baseline indicators. A one-size-fits-all prediction model is difficult to adapt to such individualized differences, resulting in decreased prediction accuracy.

[0005] In the field of diabetes management, there is an urgent need for a new method of complication prediction that can deeply integrate multi-dimensional time-series clinical data, capture individual dynamic evolution rules, and adapt to patient heterogeneity, in order to achieve earlier and more accurate risk stratification and intervention guidance. SUMMARY

[0006] The purpose of the present application is to provide a modeling method for a diabetes complication prediction model to solve the problems raised in the background art.

[0007] To achieve the above-mentioned purpose, the present application provides a modeling method for a diabetes complication prediction model, which comprises:

[0008] Collecting clinical indicator data of patients, including blood glucose level, blood pressure value, blood lipid content and renal function parameters, and performing standardization processing on the data to generate a standardized clinical indicator data set;

[0009] Extracting time-series features from the standardized clinical indicator data set, calculating short-term fluctuation trends and long-term change rules of each indicator by a sliding window algorithm, and generating a time-series feature vector;

[0010] Inputting the time-series feature vector into a deep feature extraction network, extracting high-dimensional time-series feature representation through a combination structure of multiple layers of convolution and long short-term memory units;

[0011] Based on the high-dimensional time-series feature representation, using a hierarchical clustering algorithm to group the patient population, and constructing individualized complication risk prediction sub-models according to the grouping results;

[0012] Integrating the outputs of all individualized complication risk prediction sub-models, and generating the final diabetes complication prediction result through a weighted voting mechanism.

[0013] Preferably, the standardization processing of the data to generate a standardized clinical indicator data set specifically comprises:

[0014] Calculating the mean and standard deviation of blood glucose level, blood pressure value, blood lipid content and renal function parameters, respectively;

[0015] Performing Z-score standardization processing on the original data according to the mean and standard deviation to eliminate dimensional differences;

[0016] Detecting outliers in the standardized data, and identifying and removing abnormal data points outside the reasonable range using the interquartile range method;

[0017] Performing interpolation on missing data, and using a missing value filling method based on K-nearest neighbor algorithm to ensure data integrity;

[0018] Arranging the processed data in chronological order to generate a standardized clinical indicator data set.

[0019] Preferably, the time series features in the extracted standardized clinical indicator dataset are calculated by a sliding window algorithm to calculate the short-term fluctuation trend and long-term change rule of each indicator, and a time series feature vector is generated, specifically including:

[0020] The size and step length of the sliding window are set, and the standardized clinical indicator dataset is segmented and intercepted;

[0021] In each sliding window, the mean, variance and autocorrelation coefficient of the blood glucose level, blood pressure value, blood lipid content and renal function parameter are calculated;

[0022] The change rate between adjacent windows is extracted by difference operation to generate a short-term fluctuation trend feature;

[0023] The data in the sliding window is subjected to Fourier transform, and the low-frequency component is extracted as a long-term change rule feature;

[0024] The mean, variance, autocorrelation coefficient, short-term fluctuation trend feature and long-term change rule feature are spliced to generate a time series feature vector.

[0025] Preferably, the time series feature vector is input into a deep feature extraction network, and a high-dimensional time series feature representation is extracted through a combination structure of multiple layers of convolution and long short-term memory units, specifically including:

[0026] A deep feature extraction network is constructed, the input layer of which receives the time series feature vector, and the hidden layer is composed of one-dimensional convolution layers and long short-term memory units stacked alternately;

[0027] The one-dimensional convolution layer is used to extract local features of the time series feature vector and capture the dependency between adjacent time points;

[0028] The long short-term memory unit is used to model the feature sequence output by the convolution layer and learn the long-term time series dependency pattern;

[0029] A global average pooling operation is adopted in the last layer of the network to compress the variable-length time series feature into a fixed-dimensional high-dimensional time series feature representation;

[0030] The high-dimensional time series feature representation is used as a comprehensive feature description of the patient individual, which is used for subsequent clustering and prediction.

[0031] Preferably, based on the high-dimensional time series feature representation, a hierarchical clustering algorithm is used to cluster the patient population, and an individualized complication risk prediction sub-model is constructed according to the clustering result, specifically including:

[0032] A hierarchical clustering algorithm is used to cluster analyze the high-dimensional time series feature representation, and the similarity between patients is determined by calculating the cosine similarity between features;

[0033] According to the clustering results, the patients are divided into several clusters with similar clinical characteristics, and each cluster represents a specific risk pattern;

[0034] For each cluster, its corresponding clinical indicator data and complication label are extracted to construct an independent training data set;

[0035] A personalized complication risk prediction sub-model is trained for each cluster using the gradient boosting tree algorithm, and the model parameters are optimized to minimize the prediction error;

[0036] Save all personalized complication risk prediction sub-models and their corresponding cluster identifiers for subsequent cluster matching in the prediction stage.

[0037] Preferably, the outputs of all personalized complication risk prediction sub-models are integrated to generate the final diabetes complication prediction result through a weighted voting mechanism, specifically including:

[0038] According to the high-dimensional time series feature representation of the patient, the distance between the patient and each cluster center is calculated to determine the belonging cluster or cluster combination;

[0039] Call the corresponding personalized complication risk prediction sub-model, input the patient's time series feature vector, and get the prediction probability of each sub-model;

[0040] According to the membership of the patient to each cluster, the prediction result of each sub-model is assigned a weight, and the higher the membership, the greater the weight;

[0041] Weighted average of all weighted prediction probabilities to generate a comprehensive prediction probability;

[0042] Set the probability threshold of the complication, compare the comprehensive prediction probability with the threshold, and output the final diabetes complication prediction result.

[0043] Preferably, according to the high-dimensional time series feature representation of the patient, the distance between the patient and each cluster center is calculated to determine the belonging cluster or cluster combination, specifically including:

[0044] Load the pre-trained cluster center data, and each cluster center is the mean vector of the high-dimensional time series feature representation of the corresponding cluster;

[0045] Calculate the Euclidean distance between the high-dimensional time series feature representation of the patient and each cluster center;

[0046] Select the first K clusters with the smallest distance as candidate clusters, and K is a preset hyperparameter;

[0047] If the minimum distance is less than the cluster radius threshold, it is determined that the patient belongs to the cluster; otherwise, according to the distance reciprocal, the membership of the patient to each candidate cluster is calculated;

[0048] output the cluster or cluster combination to which the patient belongs and the corresponding membership weight thereof.

[0049] Preferably, the corresponding personalized complication risk prediction sub-model is called, the time series feature vector of the patient is input, and the prediction probability of each sub-model is obtained, specifically including:

[0050] According to the cluster or cluster combination to which the patient belongs, the corresponding personalized complication risk prediction sub-model is loaded;

[0051] The time series feature vector of the patient is input into each sub-model respectively, and forward inference calculation is performed;

[0052] The original prediction probability of the complication is obtained from the output layer of the sub-model;

[0053] The original prediction probability is calibrated, the probability distribution is adjusted by using the Platt scaling method, and the prediction reliability is improved;

[0054] The calibrated prediction probability of each sub-model and the corresponding cluster identifier are output.

[0055] Preferably, the prediction result of each sub-model is assigned a weight according to the membership of the patient to each cluster, and the higher the membership, the greater the weight, specifically including:

[0056] The membership weight of the patient to each candidate cluster is obtained, and the sum of all weights is ensured to be 1;

[0057] The prediction probability of the corresponding sub-model is linearly weighted according to the membership weight;

[0058] If the patient definitely belongs to a single cluster, the prediction probability of the cluster sub-model is directly used as the final result;

[0059] If the patient belongs to multiple clusters, the weighted average prediction probability is calculated according to the membership weight;

[0060] The intermediate result in the weighting process is saved for subsequent prediction result explanation and model audit.

[0061] Preferably, the weighted average of all weighted prediction probabilities is performed to generate a comprehensive prediction probability, specifically including:

[0062] All relevant sub-models of the weighted prediction probability are summarized to form a probability set;

[0063] The arithmetic mean or weighted mean of the probability set is calculated as the comprehensive prediction probability;

[0064] The comprehensive prediction probability is smoothed by using the moving average method to eliminate random fluctuations;

[0065] The smoothed integrated prediction probability is compared with a preset complication occurrence probability threshold value;

[0066] The binary prediction result is outputted, and the integrated prediction probability is attached as a risk score.

[0067] Compared with the prior art, the present application has the beneficial effects that:

[0068] The present application can reveal the potential risk of diabetic complications earlier by deeply analyzing the time sequence dynamic characteristics of clinical indicators. The traditional method mainly relies on the absolute level of single or multiple detection values for judgment, while the present application focuses on the fluctuation trend and long-term evolution law of the indicators. These dynamic information may reflect the decompensation state of the body's metabolic control and the early damage signal of organ function earlier than the static numerical value. This deep mining of time sequence pattern moves the time point of risk assessment forward, creating more favorable conditions for early intervention.

[0069] The method uses a deep feature extraction network to process multivariate time sequence data, which can effectively capture the complex nonlinear interaction and long-range dependence relationship between indicators. The development of diabetic complications is the result of the joint action and long-term accumulation of multiple physiological parameters, and the relationship is complex. The convolution structure helps to capture the synergistic or antagonistic pattern between multiple indicators within a local time window, and the long short-term memory unit can learn the dependence relationship across a long time scale, thereby more comprehensively describing the dynamic pathophysiological process of complication occurrence.

[0070] Based on the clustering results, individualized prediction sub-models are constructed, which significantly improve the adaptability to the heterogeneity of the patient population. Diabetic patients have significant differences in complication susceptibility, progression speed, etc. The present application classifies patients with similar time sequence characteristic patterns into the same subgroup through hierarchical clustering, and constructs a customized prediction model for each subgroup. This strategy avoids using a single model to fit all patients, so that the prediction rule is more in line with the internal law of a specific subgroup, and is expected to obtain higher prediction accuracy.

[0071] The weighted voting integration decision mechanism enhances the robustness and reliability of the final prediction result. Integrating the outputs of multiple individualized sub-models can integrate the risk discrimination experience of different patient subgroups, balance the possible bias of individual models, and form a more robust group decision. This idea of ensemble learning helps to reduce the risk of overfitting and improve the generalization ability of the model on unseen data.

[0072] The prediction model constructed by the method has the potential for dynamic evolution and self-optimization. With the accumulation of more patient time sequence data, the clustering grouping and sub-model parameters can be updated regularly, so that the model can continuously learn new patterns and adapt to the changes in disease diagnosis and treatment practice. This dynamic nature enables the prediction tool to maintain its timeliness and clinical practical value.

[0073] The application of the method is expected to promote the development of diabetes management in a more refined, individualized and forward-looking direction. By providing earlier and more accurate information on the risk of complications, clinicians can develop more targeted follow-up monitoring plans and individualized treatment plans, guide medical resources to high-risk groups, and ultimately delay or reduce the occurrence of complications and improve the long-term health outcomes of diabetes patients. BRIEF DESCRIPTION OF DRAWINGS

[0074] Figure 1 The working principle diagram of the modeling method of the diabetes complication prediction model described in the present application;

[0075] Figure 2 The flowchart for generating a standardized clinical index dataset for data standardization processing;

[0076] Figure 3 The flowchart for extracting high-dimensional time series feature representation for input time series feature vectors;

[0077] Figure 4 The time series feature extraction and sliding window analysis diagram. DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0079] Please refer to Figure 1The application provides a modeling method of a diabetic complication prediction model, which comprises collecting clinical index data of patients, which covers key indexes such as blood glucose level, blood pressure value, blood lipid content and renal function parameter. The original data are subjected to standardization treatment to eliminate dimensional difference and improve data consistency, thereby generating a standardized clinical index data set. Time sequence features are extracted from the standardized data, short-term fluctuation trend and long-term change rule of each index are calculated through a sliding window algorithm, and a time sequence feature vector is formed. The time sequence feature vector is input into a deep feature extraction network, which combines the structure of multiple layers of convolution and long short-term memory unit, learns high-dimensional feature representation from time sequence data. Based on the high-dimensional features, a hierarchical clustering algorithm is used to group the patient population, each group represents a clinical mode, and an individual complication risk prediction sub-model is constructed for each group. The outputs of all sub-models are integrated to generate the final diabetic complication prediction result through a weighted voting mechanism, ensuring the accuracy and individualization of the prediction. The core is to combine time sequence analysis with machine learning, process patient heterogeneity through a grouping strategy, thereby improving the generalization ability and reliability of the prediction model.

[0080] Example 1: see Figure 2The standardization preprocessing of patient clinical indicator data and the extraction of time series features provide high-quality input data with time-dependent relationships for subsequent modeling. The data comes from the electronic health record system of medical institutions, including time series data such as blood glucose levels, blood pressure values, blood lipid content, and kidney function parameters. The original data usually has problems such as different dimensions, missing values, and outliers, so systematic cleaning and conversion are needed. The first step of standardization processing is to calculate the arithmetic mean and standard deviation of each clinical indicator at all time points. For example, calculate the mean and standard deviation of all patient test values for blood glucose levels, and process blood pressure values, blood lipid content, and other indicators in the same way. Then perform Z-score standardization, which subtracts the mean of each original data point from the corresponding indicator and divides it by the standard deviation, so that each indicator's data is converted to a distribution with a mean of zero and a standard deviation of one. This processing eliminates the incomparability of data caused by different detection units and ranges. After completing the Z-score conversion, it is necessary to identify and process outliers in the data. The interquartile range method is used for analysis. For each indicator, the data is arranged in chronological order, and the first quartile and third quartile are found, and the difference between the two is calculated to get the interquartile range. Any data point below the first quartile minus 1.5 times the interquartile range or above the third quartile plus 1.5 times the interquartile range is considered an outlier. These outliers may be due to recording errors or extreme physiological states, and are removed to avoid interference with model training. For missing values in the data set, a K-nearest neighbor-based filling method is used for processing. This method calculates the Euclidean distance between the sample with missing values and all other complete samples in the complete sample data space, and selects the K nearest samples. The value of K is usually set according to the size of the data, for example, it can be set to 10 when the sample size is large. Finally, the mean or median of the K nearest neighbor samples in that indicator is used to fill in the missing position, thus maintaining the integrity of the data set. After all processing is complete, the data is sorted by patient ID and timestamp to generate a standardized clinical indicator data set with clear structure and controllable quality.

[0081] The time series feature extraction stage takes the above standardized dataset as input and uses a sliding window algorithm. The size of the sliding window is set according to the frequency of clinical monitoring and the requirements of the prediction task. For example, for daily monitoring data, the window size can be set to 30 days, and the sliding step is set to 7 days. This means that every 7 days, a window is slid forward, and each window covers 30 consecutive days of data. The algorithm sequentially extracts the data segment within each window for analysis. Within each sliding window, the blood glucose, blood pressure, blood lipid, and renal function parameters, and other indicators are calculated for multi-dimensional features. The average value of each indicator in the window is calculated to reflect the overall level in that time period; the variance is calculated to represent the fluctuation degree of the indicator; the first-order autocorrelation coefficient is calculated to measure the linear correlation between the current value and the previous time value, which is used to evaluate the short-term autocorrelation characteristics of the indicator. In addition, first-order difference operation is also performed, which subtracts the value at the previous time from the value at the next time to obtain a series of difference values. The statistical characteristics (such as mean value) of these difference values reflect the short-term trend of the indicator in the window.

[0082] To capture the long-term change pattern of the indicator, Fourier transform is performed on the time series data within each window. Fourier transform decomposes the time domain signal into a sum of sinusoidal components of different frequencies, where low-frequency components correspond to slow-changing long-term trends, and high-frequency components correspond to short-term rapid fluctuations. The energy or amplitude of the low-frequency components (e.g., components with frequencies below a certain threshold) is extracted as the long-term change pattern feature. Finally, all the features calculated within a window - including the mean, variance, autocorrelation coefficient, short-term trend feature, and long-term pattern feature of each indicator - are concatenated into a high-dimensional vector in a predetermined order. This vector is the time series feature vector of this time window. The feature extraction process needs to traverse all the time series data of all patients to generate the feature vector corresponding to each window. These feature vectors not only contain the static statistical information of the indicators, but more importantly, capture the patterns of their dynamic evolution over time. Through this detailed preprocessing and feature engineering, the original clinical data is transformed into a standardized and feature-based form suitable for complex model processing.

[0083] Example 2: see Figure 3The deep feature extraction network is designed as a hierarchical neural network, whose input is the time series feature vector generated by the above embodiment, the length of which is determined by the number of features extracted within the sliding window. The first part of the network is usually composed of several one-dimensional convolutional layers, which perform convolution operations along the time dimension, with the convolution kernel sliding over the time series to detect local patterns, such as rapid rise patterns of blood glucose or sustained high pressure segments of blood pressure. A non-linear activation function, such as the ReLU function, is usually connected after each convolutional layer to introduce non-linear transformations and enable the network to learn more complex feature relationships. A pooling layer, such as a max pooling layer, is usually added after the convolutional part to reduce the computational load and enhance the translation invariance of the features by reducing the time dimension of the feature sequence. The local feature sequence extracted by the convolutional part is then sent to the network part composed of long short-term memory units. Long short-term memory units are a special structure of recurrent neural networks that contain input gates, forget gates, and output gates, which together determine which information should be retained and which should be forgotten. This structure is particularly suitable for handling long-term dependencies in time series data, such as identifying the relationship between the slow changing trend of blood lipid levels over months and the occurrence of complications. Multiple long short-term memory units can be stacked to form a deep network to learn more abstract temporal dynamic patterns. The output of the last long short-term memory unit is a feature sequence that contains the context information of the entire time series.

[0084] To convert the variable-length feature sequence into a fixed-dimensional representation for subsequent classification or clustering tasks, a global average pooling operation is used at the end of the network. This operation takes the average value of the feature sequence in the time dimension, thereby compressing the time series information of each feature channel into a single numerical value. The pooling results of all feature channels are concatenated together to form a fixed-dimensional high-dimensional time series feature representation, which is a highly abstracted representation of the overall clinical condition of the patient over a period of time. Based on the above high-dimensional time series feature representation, the next step is to cluster the patient population. A hierarchical clustering algorithm is used, which is a distance-based clustering method that starts by considering each patient as an independent cluster. The algorithm needs to define a method to measure the similarity between patients, in this case, the cosine similarity is used to calculate the similarity between the high-dimensional feature vectors of two patients, which focuses on the direction of the vector rather than the size. The clustering process is an iterative merging process, in which the distance between all clusters is calculated at each step (e.g., using the group average linkage method, which calculates the average similarity between all members of two clusters), and the two closest or most similar clusters are merged into a new cluster. This process is repeated until all patients are merged into several large clusters, or a pre-set number of clusters is reached. By analyzing the cluster tree diagram, a suitable number of clusters can be determined, so that the differences within the clusters are small and the differences between the clusters are large.

[0085] Each of the final formed clusters represents a subgroup of patients with similar patterns of time-series characteristics, such as the "large blood glucose fluctuation and slow kidney function decline" group or the "stable blood pressure but high blood lipid" group. For each cluster, the time-series data of clinical indicators and the label data of whether complications occur for all patients in the cluster are extracted from the corresponding original data to form an independent training subset. For each training subset of each cluster, a complication risk prediction sub-model specific to the cluster is trained using the gradient boosting tree algorithm. Gradient boosting tree is an ensemble learning algorithm that sequentially trains a series of decision trees, each of which attempts to correct the prediction errors of the previous tree. During training, techniques such as cross-validation are used to adjust hyperparameters such as tree depth and learning rate to minimize the prediction error of the model on the validation set, for example, using a log loss function. After training, each cluster corresponds to an optimized personalized prediction sub-model for the specific characteristics of the patient group, and all these sub-models and their corresponding cluster identification information (such as cluster center feature vectors) are persistently saved. This divide-and-conquer strategy allows the model to capture the risk patterns of different subgroups of patients more finely, and is expected to achieve better prediction performance compared to using a single global model.

[0086] Referring to Figure 4 , the core visualization of the time-series feature extraction stage in the diabetes complication prediction model reflects the application of the sliding window algorithm in the dynamic analysis of clinical indicators. The chart uses a double-y-axis design, with the left y-axis showing the absolute value of blood glucose level and the right y-axis possibly corresponding to other relevant clinical indicators or statistical values, and the x-axis representing continuous time points. The chart contains three key curves: the original blood glucose value curve truly reflects the actual measurement data of the patient's blood glucose level, showing the characteristic fluctuation of diabetes; the 6-month moving average line smooths the original data through the sliding window algorithm, effectively filtering out short-term random fluctuations and clearly revealing the long-term trend of blood glucose control; the fluctuation range shadow area shows the variability of blood glucose values through standard deviation calculation. The technical value of this chart lies in its ability to translate complex time-series data processing into intuitive clinical insights. The sliding window algorithm sets appropriate window size and step length to retain enough time-series information to capture long-term patterns while avoiding the response lag caused by excessively long windows. This analysis method can identify hidden patterns of deterioration in blood glucose control that cannot be detected by traditional single detection, providing a quantitative basis for early complication risk warning.

[0087] Example 3: The outputs of multiple personalized predictor models trained for different patient subgroups are effectively integrated to generate a comprehensive and reliable final complication risk prediction result, the core of the process lies in a set of weighted decision mechanisms. The whole process begins with the determination of the patient's subgroup affiliation, the system reads the high-dimensional time-series feature representation of the patient generated via a deep feature extraction network, which is a numerical vector that condenses the patient's clinical indicator change pattern over a period of time. The vector is compared with all the subgroup centers pre-stored in the system, which are obtained through clustering analysis during the model training phase, each center point represents the typical feature profile of a specific patient subgroup. The similarity comparison is usually calculated using the Euclidean distance, and the numerical distance from the patient's feature vector to each subgroup center is obtained. The system will preset a parameter K to filter out the K closest subgroups as candidate subgroups, and each subgroup is set with a radius threshold when defined, which reflects the dispersion of the features within the subgroup. If the distance from the patient to a subgroup center is less than the radius threshold of the subgroup, the patient can be clearly identified as belonging to the subgroup; if the distance from the patient to all candidate subgroups is greater than the corresponding threshold, it is considered that the patient's features are between multiple subgroups, and the membership weight of each candidate subgroup needs to be calculated, the weight is inversely proportional to the distance, the closer the distance, the higher the weight obtained by the subgroup, and the sum of all weights is 1. The output of this step clearly defines the strength of the patient's association with each subgroup.

[0088] After determining the patient's subgroup or subgroup combination and their corresponding membership weights, the system calls the corresponding personalized complication risk predictor models, which are independently trained for different subgroups' features and risk patterns. The patient's original time-series feature vector is input into all related submodels in parallel, and each submodel independently performs forward inference operations to output an initial, uncalibrated complication occurrence probability value. Due to differences in training data and internal structure of different submodels, their direct output of raw probability values may be biased and of different scales, so probability calibration is needed to improve the comparability and reliability of the results. Platt scaling method is used to calibrate the raw probability output by each submodel, which uses a binary classification logistic regression model trained on the validation set to map the original output score to a more accurate actual probability interval. The calibration process adjusts the distribution of probabilities so that the model's output probability values can better reflect the true observation frequency. After calibration, a set of more reliable prediction probability values representing risk are obtained from each related submodel.

[0089] The calibrated prediction probability of each sub-model is multiplied by its corresponding membership weight, and all these weighted probability values are summed up to obtain a preliminary integrated prediction probability. If a patient is definitely assigned to only one cluster, the membership weight is 1, and the calibrated probability of the sub-model corresponding to this cluster is directly used as the integrated probability without weighting calculation. This weighted average process is essentially an ensemble learning that considers the opinions of multiple "experts" (sub-models) and allocates the right to speak according to the matching degree of the patient with each "expert" field. Considering that the prediction at a single time point may be affected by random fluctuations, sometimes the integrated prediction probability calculated at consecutive multiple time points is smoothed, for example, using a moving average algorithm. Smoothing helps to filter out short-term noise and reveal the long-term trend of risk changes, making the prediction result more stable. Finally, the smoothed integrated prediction probability is compared with a preset decision threshold, which is set based on clinical needs and tolerance to false positives and false negatives. If the probability is higher than the threshold, a positive prediction result (high risk) is generated, otherwise a negative prediction result (low risk) is generated. The final output of the system is not only a binary classification conclusion (yes / no), but also a specific integrated prediction probability value, which can be used as a quantitative reference for risk classification and clinical decision making. The entire integration mechanism respects the characteristics of different patient subgroups through weighted voting, and also collects the wisdom of the crowd, aiming to achieve a balance between personalized prediction and model robustness.

[0090] ;

[0091] wherein: represents the membership weight of the patient to the i-th cluster, and the symbol represents the Euclidean distance from the feature vector of the patient to the i-th cluster center, and the symbol is the distance to the j-th cluster center, and the symbol represents the total number of pre-selected candidate clusters, which is a positive integer. This formula is used to calculate the membership weight of the i-th candidate cluster when the patient does not belong to any single cluster.

[0092] Example 4: For a specific patient, determine which risk cluster he belongs to, and call the corresponding personalized sub-model to make prediction. This process relies on pre-computed cluster center data and trained sub-models. Suppose there is a diabetes patient Wang, after pre-processing and feature extraction, his high-dimensional time-series features are represented as a numerical vector, now we need to determine which risk cluster he belongs to and calculate his complication risk. Load the pre-trained cluster center data, which comes from the clustering analysis results of historical patient population. Suppose the system currently defines three main clusters, labeled as cluster A, cluster B and cluster C, each cluster center is a vector with the same dimension as the patient feature vector. Refer to Table 1, which shows the numerical values of these three cluster centers in certain key feature dimensions (in actual application, the dimension will be much higher).

[0093] Table 1: Partial feature values of cluster centers

[0094] Calculate the Euclidean distance between the feature vector of patient Wang and each cluster center vector, the Euclidean distance is calculated by subtracting the numerical values of the corresponding positions of the two vectors, then squaring the difference, summing all the squared differences and taking the square root. The calculation results show that Wang's distance to cluster A is 2.1, to cluster B is 3.6, and to cluster C is 4.8. The system presets the parameter K = 2, which means selecting the top 2 clusters with the smallest distance as candidate clusters, so cluster A and cluster B are selected. Compare the distance with the cluster radius threshold, which is set according to the average distance of the members in the cluster to the center point plus several times the standard deviation during clustering. Suppose the radius threshold of cluster A is 2.5 and that of cluster B is 2.0. Since Wang's distance to cluster A is 2.1, which is less than its threshold 2.5, it can be determined that Wang belongs to cluster A. If Wang's distance to all clusters is greater than the respective radius threshold, the membership weight needs to be calculated, for example, the weight can be set as the normalized value of the inverse distance, the smaller the distance, the greater the weight. After determining that Wang belongs to cluster A, the system loads the personalized complication risk prediction sub-model trained for cluster A from storage. This sub-model is a gradient boosting tree model, which has learned the complex mapping relationship between the features specific to cluster A patients and the complication risk. Input the original time-series feature vector of Wang (which may be different from the high-dimensional summary features used in distance calculation) into this cluster A sub-model, the sub-model performs forward calculation and produces an original prediction value at the output layer, such as a score between 0 and 1, the initial value is 0.65, indicating the risk probability originally judged by the model.

[0095] Since the probabilities directly output by the machine learning model may not be sufficiently calibrated, i.e., 0.65 does not necessarily mean a true 65% risk of occurrence, it is necessary to calibrate the original prediction probabilities. The Platt scaling method is adopted, which uses a logistic regression model to adjust the original output. The system applies the Platt scaling parameters (including scaling factor and offset) trained specifically for the subgroup A sub-model to transform the original output 0.65 and map it to a more accurate probability value, assuming that the calibrated probability becomes 0.58. Finally, the system outputs the result: patient Wang is determined to belong to subgroup A, and his calibrated complication prediction probability is 0.58. This result contains both the subgroup identification and the prediction probability, providing a basis for subsequent decision-making. The entire process from distance calculation, subgroup matching to model invocation and probability calibration forms an automated risk assessment chain, aiming to provide more targeted risk prediction for different types of patients. Through this refined process, the model can better adapt to individual differences among patients, improving the accuracy and practicality of prediction. It is worth noting that in the actual system, the number of subgroups may be much more than three, and the feature dimension can reach hundreds of dimensions, but the core logic is consistent with this example.

[0096] Example 5: Process of assigning weights to results from different prediction sub-models and generating final prediction values, which determines the final output form of individualized risk assessment. Take a fictional diabetic patient Zhang as an example, whose clinical data has been converted into high-dimensional time-series features after pre-processing and has completed subgroup matching. Assume that the system determines that Zhang is significantly associated with subgroup Alpha and subgroup Beta, with membership weights of 0.7 and 0.3 respectively, which means that Zhang's clinical features are closer to the typical pattern of subgroup Alpha. The system obtains the membership weights of Zhang for these two candidate subgroups, which are calculated and normalized in advance and sum up to 1. According to these weight values, the system assigns influence to the prediction results of the corresponding sub-models, with the prediction result of the subgroup Alpha sub-model taking the dominant position. Next, the individualized complication risk prediction sub-models trained for subgroup Alpha and subgroup Beta are invoked, and Zhang's feature data is input into these two models. Assume that the calibrated prediction probability output by the subgroup Alpha sub-model is 0.62, and the calibrated prediction probability output by the subgroup Beta sub-model is 0.45. Since Zhang does not belong to any single subgroup, the system enters the multi-subgroup weighting calculation process.

[0097] The prediction probability of the subgroup Alpha model 0.62 is multiplied by its weight 0.7 to obtain the weighted probability 0.434; the prediction probability of the subgroup Beta model 0.45 is multiplied by its weight 0.3 to obtain the weighted probability 0.135. Adding the two weighted probabilities, the preliminary comprehensive prediction probability value is 0.569. This weighting calculation process is automatically completed inside the system, and the intermediate results including the original output of each sub-model, the weight distribution value and the weighted calculation process are recorded in the log for subsequent result tracing and model auditing. Considering that the prediction value at a single time point may be affected by accidental factors, the system performs smoothing processing on the comprehensive prediction probability generated in continuous multiple time windows. Assuming that the comprehensive prediction probability values at the last three time points are 0.551, 0.569 and 0.548 respectively, the arithmetic mean is 0.556. This moving average method effectively reduces the influence of random fluctuations on the prediction results, making the risk assessment more stable and reliable. The smoothed comprehensive prediction probability value 0.556 will be output as the final risk score. The smoothed comprehensive prediction probability is compared with the preset complication occurrence probability threshold, which is usually determined by clinical experts according to epidemiological data and actual application scenarios. The establishment of the threshold is a process of integrating clinical medicine, epidemiology and actual diagnosis and treatment needs. Clinical experts will review a large number of historical case data, analyze the corresponding relationship between the actual incidence of complications and the model prediction probability. They may consider the baseline risk of the population, the severity of different complications, and the clinical consequences of false judgments (such as false negatives may delay treatment, and false positives may lead to excessive medical treatment), and thus determine a critical point that balances sensitivity and specificity. This threshold is not fixed, and it may be set differently for different types of complications (such as diabetic nephropathy, retinopathy, etc.) or different clinical application scenarios (such as general screening and high-risk population monitoring). When this threshold is defined as 0.55, it represents a balanced clinical decision boundary.

[0098] When the patient Zhang's smooth posterior predictive probability value 0.556 is ready to be compared with the threshold, a simple logical judgment is performed internally by the system. This comparison operation is essentially a transformation of a continuous probability variable into a binary classification signal. The probability value 0.556 is greater than the threshold 0.55, and this tiny difference triggers the system decision logic to shift to the "at risk" state. This comparison process is automated, but the setting of the threshold behind it involves deep clinical knowledge, enabling the machine's judgment to align with the needs of medical practice. Based on the comparison result, the system begins to generate the final output content. The binary prediction result is the primary output, which is presented in the clear and unambiguous form of "at risk" or "not at risk". This black-or-white conclusion facilitates rapid risk screening and patient diversion. After generating this classification conclusion, the system simultaneously records the specific probability value that triggered the conclusion and the threshold used, providing data support for the traceability of the conclusion. That specific synthetic predictive probability value 0.556 is retained as a risk score, which, together with the binary classification conclusion, constitutes the complete output. This risk score contains more information than just a simple classification, as it quantifies the degree of risk. In the clinical interface, doctors can not only see the "at risk" warning, but also the 0.556 value. This score can be incorporated into a broader management perspective, and doctors can interpret this score in combination with the patient's age, disease course, comorbidities, lifestyle habits, and other factors. For patients whose scores just exceed the threshold, doctors may recommend increasing the frequency of follow-up; while for patients whose scores are much higher than the threshold, it may mean that more aggressive intervention measures need to be initiated immediately. The score also helps to achieve dynamic management of patient risk, as tracking the trend of the score over time can assess the progress of the disease or the effectiveness of intervention measures. The final presentation to the doctor may be a structured report interface, which prominently displays the "complication risk: positive" or "at risk" conclusion, and below it lists the "risk score: 0.556" and possible related subgroup information. This output method not only provides clear action guidance, but also preserves the subtle differences in probability, combining the judgment of the automated model with the professional discretion of the clinician, providing efficient and in-depth decision support for subsequent individualized diagnosis and treatment plan development.

[0099] The entire weight distribution and result generation process embodies the idea of ensemble learning of the model, combining the wisdom of multiple specialized sub-models and adjusting their influence according to the matching degree of individual characteristics and the professional field of each sub-model, ultimately obtaining a more personalized and robust prediction result. This mechanism is particularly suitable for complex scenarios such as diabetes complication prediction, which involves multiple pathological mechanisms and high heterogeneity in clinical manifestations, and can adapt to the risk assessment needs of different characteristic patient groups. The model output not only provides a simple yes-or-no judgment, but also provides a quantitative risk probability value, providing more rich decision information for hierarchical diagnosis and individualized intervention.

[0100] It has to be noted that, in the present document, the terms "first", "second", etc. merely serve to identify a subject or action, without necessarily requiring or implying any such actual relationship or order between such subjects or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0101] While embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, combinations, and variations of the embodiments can be undertaken without departing from the spirit and scope of the present application, which is defined by the appended claims and their equivalents.

Claims

1. A modeling method of a diabetic complication prediction model, characterized by, The method comprises the following steps: Collecting clinical index data of patients, including blood glucose level, blood pressure value, blood lipid content and renal function parameters, standardizing the data, and generating a standardized clinical index data set; Extracting time sequence features in the standardized clinical index data set, calculating short-term fluctuation trends and long-term change rules of each index by a sliding window algorithm, and generating a time sequence feature vector; Inputting the time sequence feature vector into a deep feature extraction network, extracting high-dimensional time sequence feature representation through a combination structure of multiple layers of convolution and long short-term memory units; Based on the high-dimensional time sequence feature representation, using a hierarchical clustering algorithm to group the patient population, and constructing an individualized complication risk prediction sub-model according to the grouping result; Integrating the outputs of all individualized complication risk prediction sub-models, and generating the final diabetes complication prediction result through a weighted voting mechanism. 2.The modeling method of a diabetic complication prediction model according to claim 1, wherein, The standardization processing of the data to generate the standardized clinical index data set specifically comprises: For blood glucose level, blood pressure value, blood lipid content and renal function parameters, the mean and standard deviation are calculated respectively; The original data is standardized by Z-score to eliminate the dimensional difference; Detecting outliers in the standardized data, identifying and removing abnormal data points outside the reasonable range by the interquartile range method; Imputing missing data by using a missing value filling method based on K-nearest neighbor algorithm to ensure data integrity; Arranging the processed data in chronological order to generate the standardized clinical index data set. 3.The modeling method of a diabetic complication prediction model according to claim 2, wherein, The extraction of time sequence features in the standardized clinical index data set, the calculation of short-term fluctuation trends and long-term change rules of each index by a sliding window algorithm, and the generation of a time sequence feature vector specifically comprise: Setting the size and step length of the sliding window, and segmenting and intercepting the standardized clinical index data set; In each sliding window, the mean, variance and autocorrelation coefficient of blood glucose level, blood pressure value, blood lipid content and renal function parameters are calculated; The change rate between adjacent windows is extracted by difference operation to generate short-term fluctuation trend features; The Fourier transform is performed on the data in the sliding window to extract low-frequency components as long-term change rule features; The mean, variance, autocorrelation coefficient, short-term fluctuation trend features and long-term change rule features are spliced to generate a time sequence feature vector. 4.The modeling method of a diabetic complication prediction model according to claim 3, wherein, The input of the time sequence feature vector into the deep feature extraction network, the extraction of high-dimensional time sequence feature representation through the combination structure of multiple layers of convolution and long short-term memory units, specifically comprises: Building a deep feature extraction network, the input layer of which receives the time sequence feature vector, the hidden layer of which is composed of one-dimensional convolution layers and long short-term memory units stacked alternately; Local feature extraction of the time sequence feature vector is performed by the one-dimensional convolution layer to capture the dependence between adjacent time points; The long short-term memory unit is used to model the feature sequence output by the convolution layer to learn the long-term time sequence dependence pattern; Global average pooling operation is adopted in the last layer of the network to compress the variable-length time sequence feature into a fixed-dimensional high-dimensional time sequence feature representation; The high-dimensional time sequence feature representation is used as a comprehensive feature description of the patient individual for subsequent grouping and prediction. 5.The modeling method of a diabetic complication prediction model according to claim 4, wherein, The high-dimensional time sequence feature representation is clustered by using a hierarchical clustering algorithm, and the similarity between patients is determined by calculating the cosine similarity between features. According to the clustering result, the patients are divided into several subgroups with similar clinical characteristics, and each subgroup represents a specific risk pattern. For each subgroup, the corresponding clinical indicator data and complication label are extracted to construct an independent training data set. A personalized complication risk predictor model is trained for each subgroup using the gradient boosting tree algorithm, and the model parameters are optimized to minimize the prediction error. All personalized complication risk predictor models and their corresponding subgroup identifiers are saved for subsequent subgroup matching in the prediction stage. The outputs of all personalized complication risk predictor models are integrated to generate the final diabetes complication prediction result through a weighted voting mechanism, which includes:

6. The modeling method of a diabetic complication prediction model according to claim 5, wherein, According to the high-dimensional time sequence feature representation of the patient, the distance between the patient and each subgroup center is calculated to determine the belonging subgroup or subgroup combination. The corresponding personalized complication risk predictor model is called, and the time sequence feature vector of the patient is input to obtain the prediction probability of each submodel. According to the membership degree of the patient to each subgroup, the prediction result of each submodel is assigned a weight, and the higher the membership degree, the greater the weight. The weighted average of all weighted prediction probabilities is calculated to generate a comprehensive prediction probability. Set the probability threshold of the complication, compare the comprehensive prediction probability with the threshold, and output the final diabetes complication prediction result. According to the high-dimensional time sequence feature representation of the patient, the distance between the patient and each subgroup center is calculated to determine the belonging subgroup or subgroup combination, which includes: 7.The modeling method of a diabetic complication prediction model according to claim 6, wherein, Load the pre-trained subgroup center data, and each subgroup center is the mean vector of the high-dimensional time sequence feature representation of the corresponding subgroup. Calculate the Euclidean distance between the high-dimensional time sequence feature representation of the patient and each subgroup center. Select the top K subgroups with the smallest distance as candidate subgroups, where K is a pre-set hyperparameter. If the minimum distance is less than the subgroup radius threshold, the patient belongs to the subgroup; otherwise, calculate the membership degree of the patient to each candidate subgroup according to the reciprocal of the distance. Output the subgroup or subgroup combination to which the patient belongs and the corresponding membership weight. The corresponding personalized complication risk predictor model is called according to the subgroup or subgroup combination to which the patient belongs. 8.The modeling method of a diabetic complication prediction model according to claim 7, wherein, The time sequence feature vector of the patient is input into each submodel for forward reasoning calculation. The original prediction probability of the complication is obtained from the output layer of the submodel. The original prediction probability is calibrated, and the probability distribution is adjusted by using the Platt scaling method to improve the prediction reliability. Output the calibrated prediction probability of each submodel and its corresponding subgroup identifier. The prediction result of each submodel is assigned a weight according to the membership degree of the patient to each subgroup, and the higher the membership degree, the greater the weight. ​ 9.The modeling method of a diabetic complication prediction model according to claim 8, wherein, ​ Obtain membership weights of the patient to each candidate cluster, ensuring that the sum of all weights is 1; Linearly weight the prediction probabilities of the corresponding sub-models according to the membership weights; If the patient definitely belongs to a single cluster, directly use the prediction probability of the cluster sub-model as the final result; If the patient belongs to multiple clusters, calculate the weighted average prediction probability according to the membership weights; Save the intermediate results in the weighting process for subsequent prediction result interpretation and model audit. 10.The modeling method of a diabetic complication prediction model according to claim 9, wherein, The weighted average of all weighted prediction probabilities is generated to generate a comprehensive prediction probability, specifically including: Summarize the weighted prediction probabilities of all related sub-models to form a probability set; Calculate the arithmetic mean or weighted mean of the probability set as the comprehensive prediction probability; Smooth the comprehensive prediction probability, and use the moving average method to eliminate random fluctuations; Compare the smoothed comprehensive prediction probability with the preset complication occurrence probability threshold value; Output the binary prediction result, and attach the comprehensive prediction probability as the risk score.

Images