Diabetic complication prediction model construction method and diabetic complication prediction system

Abstract

The present application relates to the technical field of medical data analysis, in particular to a diabetes complication prediction model construction method and a diabetes complication prediction system; the system adopts Pearson correlation coefficient and mutual information to preliminarily screen the individual clinical time series data of patients, adopts a temporary DBN model to calculate the importance score of the preliminary features, determines the important features based on the importance score, calculates the partial correlation coefficient of the important features in the latter part of the importance score, confirms whether the corresponding important features are retained based on the size of the partial correlation coefficient, obtains the final features, trains and verifies the training set based on the final features by using the DBN model optimized by IWOA, and obtains a diabetes complication prediction model; the prediction accuracy reaches 94.5%, which is better than that of a traditional model, has high clinical practicability, reversely guides data collection, provides reliable support for early warning and precise intervention of complications, and has scientific research value and clinical conversion potential.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, specifically to a method for constructing a prediction model for diabetes complications and a prediction system for diabetes complications. Background Technology

[0002] Diabetes mellitus, a prevalent chronic metabolic disease worldwide, is prone to causing various serious complications such as kidney disease, retinopathy, and cardiovascular disease during its progression. These complications are characterized by their insidious nature and irreversible progression; once they reach the middle or late stages, they significantly reduce patients' quality of life and can even be life-threatening. Therefore, achieving early and accurate prediction of diabetic complications is of significant clinical value and social importance for optimizing clinical intervention programs, slowing disease progression, and reducing the medical burden.

[0003] With the development of medical informatization, a large amount of individual patient clinical time-series data, including multi-dimensional data such as blood glucose fluctuations, blood pressure changes, biochemical indicators, and medication records, has been accumulated during clinical diagnosis and treatment. This data contains potential patterns in the occurrence and development of complications, providing a foundation for building data-driven predictive models. However, existing technologies for predicting diabetes complications still have many shortcomings: On the one hand, traditional prediction methods often rely on single types of clinical indicators or static data, failing to fully explore the dynamic dependencies across time steps in the time-series data, resulting in an incomplete characterization of the risk of complications; on the other hand, clinical time-series data suffers from high dimensionality, redundant information, and complex feature correlations. Existing feature selection methods often employ single correlation analysis or simple statistical screening, making it difficult to effectively remove noisy features and retain key information, thus affecting the accuracy of the prediction model; furthermore, the parameters of deep learning models used for prediction often rely on empirical settings or simple optimization algorithms, which are prone to slow convergence and getting stuck in local optima, thus limiting the model's generalization ability and prediction accuracy.

[0004] Meanwhile, the time sensitivity of clinical data varies among different patients, and existing data collection methods lack targeted adjustments, potentially leading to missing key time-series information or redundant invalid data, further reducing the reliability of predictive models. Therefore, how to construct a diabetes complication prediction system capable of accurately processing clinical time-series data, efficiently screening key features, and optimizing model parameters to achieve early and accurate warnings of complication risks has become a pressing technical problem in the current medical and health field. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention aims to provide a method for constructing a diabetes complication prediction model and a diabetes complication prediction system, so as to analyze complex diabetes data and accurately predict the risk of diabetes complications.

[0006] To solve the above problems, the present invention adopts the following technical solution:

[0007] On the one hand, the present invention provides a diabetic complication prediction system, including a data acquisition module, a feature screening module, a model building module, and a prediction module;

[0008] The data acquisition module is used to collect individual clinical time-series data of patients;

[0009] The feature screening module is used to perform preliminary feature screening on individual patient clinical time series data using Pearson correlation coefficient and mutual information, calculate the importance score of preliminary features using a temporary DBN model, determine important features based on the importance score, calculate the partial correlation coefficient of some important features with lower importance scores, and determine whether to retain the corresponding important features based on the magnitude of the partial correlation coefficient, thus obtaining the final features.

[0010] The model building module is used to train and validate the DBN model optimized by IWOA based on the training set constructed from the final features, so as to obtain a diabetes complication prediction model.

[0011] The prediction module is used to derive prediction results for diabetes complications based on the final characteristics of the patient being tested using the diabetes complication prediction model.

[0012] Furthermore, the data acquisition module is also used to adjust the acquisition of individual clinical time-series data of patients based on the time-series sensitivity coefficient.

[0013] Furthermore, the preliminary feature screening of individual patient clinical time-series data using Pearson correlation coefficient and mutual information includes:

[0014] Pearson correlation coefficients and mutual information are calculated for all features in the individual patient's clinical time-series data. By comparing them with the set Pearson correlation coefficient thresholds and mutual information thresholds, Pearson feature sets and mutual information feature sets are obtained respectively. The common features in the Pearson feature sets and mutual information feature sets are taken as preliminary features.

[0015] Furthermore, the IWOA-optimized DBN model includes an input layer, a first hidden layer, a second hidden layer, and an output layer;

[0016] The input layer is used to receive the final features;

[0017] The first hidden layer is used to extract local features at a single time step and output the state of the first hidden layer;

[0018] The second hidden layer outputs the second hidden layer state by fusing the dynamic dependencies across time steps based on the state of the first hidden layer.

[0019] The output layer maps the state of the second hidden layer to the probability of complication occurrence through the Sigmoid activation function.

[0020] Furthermore, the IWOA-optimized DBN model uses the IWOA algorithm to optimize DBN parameters. The DBN parameters include the weight matrix of the input layer and the first hidden layer, the bias vector of the first hidden layer, the weight matrix of the first hidden layer and the second hidden layer, the bias vector of the second hidden layer, the weight matrix of the second hidden layer and the output layer, and the bias vector of the output layer.

[0021] Furthermore, the optimization of DBN parameters using the IWOA algorithm includes optimizing DBN parameters by simulating whale predation behavior:

[0022] Each particle corresponds to a set of DBN parameters;

[0023] Set the fitness function and initialize the particle swarm;

[0024] For each particle in the particle swarm, the initial fitness is calculated according to the fitness function, and the particle with the lowest fitness is selected as the current best particle.

[0025] The position of each particle is updated based on the prey-surrounding behavior, bubble-web predation behavior, and random search behavior. The fitness of each updated particle is calculated, and it is determined whether to replace the current best particle for iteration. Finally, the best particle is output.

[0026] The final optimal particle decoding yields the optimized DBN parameters, which are used in the DBN model to obtain a predictive model for diabetes complications.

[0027] On the other hand, the present invention provides a method for constructing a predictive model for diabetic complications, comprising:

[0028] Collect individual clinical time-series data from patients;

[0029] Pearson correlation coefficient and mutual information were used to perform preliminary feature screening on individual clinical time series data of patients. An ad hoc DBN model was used to calculate the importance score of the preliminary features. Important features were determined based on the importance score. The partial correlation coefficients of some important features with lower importance scores were calculated. Based on the magnitude of the partial correlation coefficients, it was determined whether the corresponding important features should be retained, and the final features were obtained.

[0030] The DBN model optimized by IWOA was trained and validated on the training set constructed from the final features to obtain a prediction model for diabetes complications.

[0031] Furthermore, the preliminary feature screening of individual patient clinical time-series data using Pearson correlation coefficient and mutual information includes:

[0032] Pearson correlation coefficients and mutual information are calculated for all features in the individual patient's clinical time-series data. By comparing them with the set Pearson correlation coefficient thresholds and mutual information thresholds, Pearson feature sets and mutual information feature sets are obtained respectively. The common features in the Pearson feature sets and mutual information feature sets are taken as preliminary features.

[0033] Furthermore, the IWOA-optimized DBN model includes an input layer, a first hidden layer, a second hidden layer, and an output layer;

[0034] The input layer is used to receive the final features;

[0035] The first hidden layer is used to extract local features at a single time step and output the state of the first hidden layer;

[0036] The second hidden layer outputs the second hidden layer state by fusing the dynamic dependencies across time steps based on the state of the first hidden layer.

[0037] The output layer maps the state of the second hidden layer to the probability of complication occurrence through the Sigmoid activation function.

[0038] Furthermore, the IWOA-optimized DBN model uses the IWOA algorithm to optimize DBN parameters. The DBN parameters include the weight matrix of the input layer and the first hidden layer, the bias vector of the first hidden layer, the weight matrix of the first hidden layer and the second hidden layer, the bias vector of the second hidden layer, the weight matrix of the second hidden layer and the output layer, and the bias vector of the output layer.

[0039] Furthermore, the optimization of DBN parameters using the IWOA algorithm includes optimizing DBN parameters by simulating whale predation behavior:

[0040] Each particle corresponds to a set of DBN parameters;

[0041] Set the fitness function and initialize the particle swarm;

[0042] For each particle in the particle swarm, the initial fitness is calculated according to the fitness function, and the particle with the lowest fitness is selected as the current best particle.

[0043] The position of each particle is updated based on the prey-surrounding behavior, bubble-web predation behavior, and random search behavior. The fitness of each updated particle is calculated, and it is determined whether to replace the current best particle for iteration. Finally, the best particle is output.

[0044] The final optimal particle decoding yields the optimized DBN parameters, which are used in the DBN model to obtain a predictive model for diabetes complications.

[0045] The beneficial effects of this invention are as follows: The method for constructing a diabetic complication prediction model and the diabetic complication prediction system of this invention have outstanding prediction accuracy. By capturing temporal correlations through dynamic data in all dimensions and combining IWOA to optimize the DBN to break through local optima, the prediction accuracy reaches 94.5%, and the missed diagnosis rate and misdiagnosis rate are both less than 7.5%, which is better than traditional models. It has high clinical applicability, adaptive risk grading and personalized intervention suggestions, and reverse guidance for data collection, providing reliable support for early warning and precise intervention of complications. It has both scientific research value and clinical translation potential. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of a diabetes complication prediction system according to the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to specific embodiments.

[0048] It should be noted that these embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Simple improvements to the method under the premise of the present invention are all within the scope of protection claimed by the present invention.

[0049] See Figure 1 This is a diabetes complication prediction system, comprising a data acquisition module 100, a feature selection module 200, a model building module 300, and a prediction module 400.

[0050] The data acquisition module is used to collect individual clinical time-series data of patients.

[0051] Individual clinical time-series data, such as biochemical indicators, physiological dynamics, and behavioral data, are exemplified in Tables 1-3.

[0052] Table 1 Biochemical Indicator Data

[0053] Biochemical indicator name Indicator Definition Initial sampling frequency fasting blood glucose Venous blood glucose concentration after 8 hours of fasting Every 1.5 weeks 2-hour postprandial blood glucose venous blood glucose concentration 2 hours after eating Every 1.5 weeks Glycated hemoglobin (HbA1c) Reflects average blood glucose levels over the past 2-3 months Every 1.5 weeks Systolic blood pressure (SBP) Maximum arterial blood pressure during cardiac systole Every 1.5 weeks Diastolic blood pressure (DBP) Minimum diastolic arterial blood pressure Every 1.5 weeks Creatinine (Cr) Core indicators reflecting kidney function Every 2 weeks Blood urea nitrogen (BUN) auxiliary indicators reflecting kidney function Every 2 weeks Macular thickness foveal thickness in the retina Every 3 months Retinal nerve fiber layer thickness (RNFL) Average thickness of retinal nerve fiber layer Every 3 months

[0054] Table 2 Physiological dynamic data

[0055] Physiological dynamic index name Indicator Definition Initial sampling frequency Heart rate variability (SDNN) Standard deviation of normal RR intervals (time-domain indicator) Every 5 minutes Heart rate variability (RMSSD) Root mean square of the difference between adjacent RR intervals (time domain index) Every 5 minutes Heart rate variability (LF) Low-frequency components (0.04-0.15Hz, reflecting sympathetic nerve activity) Every 5 minutes Heart rate variability (HF) High-frequency components (0.15-0.4Hz, reflecting parasympathetic nerve activity) Every 5 minutes Nighttime sleep breathing rate Average number of breaths during sleep Every minute Daily activity level (steps) Daily cumulative steps Every 15 minutes Daily activity level (MET) Metabolic equivalent (reflecting exercise intensity) Every 15 minutes

[0056] Physiological dynamic data can be collected using devices such as smart bracelets and averaged on an hourly or daily basis.

[0057] Table 3 Behavioral Data

[0058] Behavioral indicator name Indicator Definition sampling frequency Carbohydrate percentage Daily carbohydrate intake as a percentage of total calories daily Protein percentage Daily protein intake as a percentage of total calories daily Fat percentage Daily fat intake as a percentage of total calories daily Duration of moderate to high intensity exercise Daily MET≥3 cumulative exercise time daily Medication adherence Dosage schedule / number of injections / daily dosage / number of injections daily

[0059] Data preprocessing includes missing value imputation, outlier handling, and so on.

[0060] The data acquisition module 100 is also used to adjust the acquisition of individual clinical time-series data of patients based on the time-series sensitivity coefficient.

[0061] For example: individual patient clinical time series data includes 12 months of time series data. , Let i be the i-th feature.

[0062] First, calculate the time series variance of the features. :

[0063] , ;

[0064] Where t represents time, specifically the number of months. The characteristic mean is denoted as .

[0065] Then calculate the features. Pearson correlation coefficient with complication outcome Y :

[0066] ;

[0067] in, The status of complications at time t (1 for occurrence, 0 for non-occurrence). .

[0068] Finally, the features are calculated. Time sensitivity coefficient :

[0069] ;

[0070] Where n is the total number of features.

[0071] like A value ≥0.08 can increase the sampling frequency by 20%-50%; If the value is less than 0.03, the sampling frequency is reduced by 20%-50%; the rest remain unchanged.

[0072] The feature screening module 200 is used to perform preliminary feature screening on individual patient clinical time series data using Pearson correlation coefficient and mutual information. It uses a temporary DBN model to calculate the importance score of the preliminary features, determines important features based on the importance score, calculates the partial correlation coefficient of some important features with lower importance scores, and determines whether to retain the corresponding important features based on the magnitude of the partial correlation coefficient, thus obtaining the final features.

[0073] Pearson correlation coefficient and mutual information were used to perform preliminary feature screening on individual patient clinical time-series data as follows:

[0074] By calculating the Pearson correlation coefficients between all features and complication outcomes, and by setting a Pearson correlation coefficient threshold, feature filtering is performed to obtain the Pearson feature set.

[0075] By calculating the mutual information of all features, such as for continuous features Discretize the data into four intervals (x1, x2, x3, x4) according to quartiles, and discretize the complication structure Y into two classes (0, 1). Then the mutual information... :

[0076] ;

[0077] Where k is a continuous feature The index of the value, P is the probability calculation, and l is the value of the symptom Y.

[0078] Can be retained Features with a mutual information threshold of ≥0.3 are used to obtain the mutual information feature set.

[0079] The common features in the Pearson feature set and the mutual information feature set are used as the initial features.

[0080] Following this, a temporary DBN model is used to calculate the importance score of the initial feature. The important features are determined based on the importance score, and the important features are then determined based on the magnitude of the partial correlation coefficient to determine whether to retain them, thus obtaining the final features.

[0081] for example:

[0082] The temporary DBN model parameters are set as follows:

[0083] Input layer: Number of nodes = Number of candidate features (preliminary features mentioned above, such as 25), each node corresponds to 12 months of time series data for one feature (input dimension 25×12).

[0084] Hidden layer: 1 layer, number of nodes = number of candidate features × 2 (e.g., 50), activation function ReLU: ;

[0085] Output layer: 1 node, activation function Sigmoid: ;

[0086] Training parameters: learning rate 0.01, number of iterations 50, batch size 8.

[0087] For the temporary DBN after training, calculate each input feature X i Partial derivative with respect to output O (reflecting the degree of influence of the feature on the output):

[0088] ;

[0089] Among them, H t This is the output of the hidden layer at time t. W HO For hidden layer-output layer weights, W XH For the input layer-hidden layer weights, b H The hidden layer bias is represented by ReLU', which is 1 when x>0 and 0 when x≤0.

[0090] Feature importance score:

[0091] .

[0092] Filter by important features: such as by Importance (X) i Sort in descending order and take the first 20 features;

[0093] Calculate the partial correlation coefficients for the more important features with lower importance scores, such as features ranked 15-20:

[0094] For X 15 Perform multiple linear regression on Y and Y respectively:

[0095] ;

[0096] ;

[0097] in, For X 15 The corresponding regression intercept term, For X 15 In regression, the explanatory variable X j The corresponding regression coefficients, For X 15 The residual of the regression, The intercept term of the regression corresponding to Y. In the regression of Y, the explanatory variable X j The corresponding regression coefficients, Let Y be the regression residual;

[0098] Calculate residuals and The Pearson correlation coefficient is the partial correlation coefficient.

[0099] If the partial correlation coefficient is not less than the threshold set for the partial correlation coefficient, the feature is retained; otherwise, it is removed.

[0100] Finally, the features retained here are combined with the earlier important features (i.e., the first 1-14 features) to form the final features.

[0101] The final features are normalized and incorporated into the subsequent model training and prediction process.

[0102] The model building module 300 is used to train and validate the DBN model optimized by IWOA based on the training set constructed from the final features, so as to obtain a prediction model for diabetic complications.

[0103] The IWOA-optimized DBN model consists of an input layer, a first hidden layer, a second hidden layer, and an output layer.

[0104] The input layer receives the final features, such as the temporal matrix of the final features. The input vector at each time step t is .

[0105] The first hidden layer extracts local features at a single time step and outputs the state of the first hidden layer. :

[0106] ;

[0107] This is the weight matrix between the input layer and the first hidden layer. This is the bias vector of the first hidden layer. This is the time dependence coefficient;

[0108] The second hidden layer, based on the state of the first hidden layer, fuses the dynamic dependencies across time steps and outputs the state of the second hidden layer. :

[0109] ;

[0110] in, This is the weight matrix for the first hidden layer and the second hidden layer. This is the bias vector for the second hidden layer. For dynamic dependency coefficients;

[0111] The output layer uses a sigmoid activation function to map the state of the second hidden layer to the probability of complication in the interval [0,1]. :

[0112] ;

[0113] in, This is the weight matrix between the second hidden layer and the output layer. This is the bias vector for the output layer.

[0114] If the probability P of complications occurring at multiple time steps is to be predicted, a weighted approach is used to calculate the probability of complications occurring at multiple time steps (e.g., 12). Processing:

[0115] ;

[0116] The weights are determined by time steps and are linearly increasing to ensure that recent data contributes more to the final prediction.

[0117] The IWOA-optimized DBN model uses the IWOA algorithm to optimize the DBN parameters. The DBN parameters include the aforementioned weight matrix of the input layer and the first hidden layer, the bias vector of the first hidden layer, the weight matrix of the first hidden layer and the second hidden layer, the bias vector of the second hidden layer, the weight matrix of the second hidden layer and the output layer, and the bias vector of the output layer.

[0118] The specific process involves optimizing DBN parameters by simulating whale predation behavior:

[0119] Each particle corresponds to a set of DBN parameters. All parameters are initialized with a uniform distribution in the range [−0.5, 0.5] to avoid gradient explosion caused by excessively large initial values.

[0120] Define the fitness function:

[0121] ;

[0122] Where x is a single particle, , , is the weighting coefficient, determined through 5-fold cross-validation; TC is the temporal consistency; MES is the prediction error; and Reg is the parameter regularization.

[0123] ;

[0124] Where M is the number of patients in the training set. This represents the actual complication status of patient m in month t. Let be the predicted probability of the m-th patient in month t.

[0125] ;

[0126] in, The sign function is 1 when z>0, 0 when z=0, and -1 when z<0.

[0127] ;

[0128] in, Let D be the Euclidean norm and D be the particle dimension.

[0129] The iterative process is as follows:

[0130] Step 1: Initialize the particle swarm

[0131] Generate 50 particles, each with 3457 dimensions following a uniform distribution in the range [−0.5, 0.5]. ;in, This represents the d-th dimension value of the n-th particle.

[0132] Step 2: Calculate the initial fitness

[0133] For each particle, decode it into DBN parameters (W1, b1, W2, b2, W3, b3), train the DBN using the training set data and calculate the fitness, and record the current best particle x. ∗ (The particle with the lowest fitness).

[0134] Step 3: Iterative optimization (iteration number t=1 to T) max =100)

[0135] (1) Update control parameters

[0136] Linear decrease coefficient: ;

[0137] Dynamic weighting factor: , This represents the initial maximum value of the dynamic weights. This represents the initial minimum value of the dynamic weights;

[0138] Random vector: , which follows a uniform distribution U(0,1) (each dimension of each particle is generated independently);

[0139] Behavioral probability: .

[0140] (2) Particle position update

[0141] For each particle x n (t), update the position according to the following rules:

[0142] 1) Encircling prey behavior (triggering condition: p < 0.5 and < 1)

[0143] Direction-step control parameters: ,

[0144] Random weight parameters: ,

[0145] Relative distance with random weights: ,

[0146] Final position update formula: , Let be the current optimal particle position at the t-th iteration. This represents the new position of the nth particle in the (t+1)th iteration.

[0147] The simulation simulates a whale approaching its optimal prey (the current optimal particle). The dynamic weight w(t) controls the approach speed, with a fast approach in the early stage (large w) and precise fine-tuning in the later stage (small w).

[0148] 2) Bubble web predation behavior (triggering condition: p < 0.5 and |A|≥1)

[0149] Original relative distance: ,

[0150] Helix coefficient: ,

[0151] Position update formula: b is the helical constant;

[0152] Simulates a whale spiraling upwards around its prey and exhaling bubbles to surround it, expanding the search range through the spiral trajectory and avoiding getting trapped in local optima.

[0153] 3) Random search behavior (triggering condition: p≥0.5)

[0154] Random non-optimal particles: Randomly select a non-optimal particle.

[0155] Relative distance with random weights: ,

[0156] Position update formula: ;

[0157] The new position of the particle is updated towards a random non-optimal particle (rather than towards the current optimal position), while simultaneously... Adjust the step size and w(t) to control the update amplitude.

[0158] Calculate the fitness of each updated particle: ;like Then update the optimal particle. .

[0159] When the number of iterations reaches 100, the optimal particle x is output. ∗ The optimized DBN parameters are decoded, the model is trained, and a diabetes complication prediction model is obtained.

[0160] The prediction module 400 is used to derive a prediction result of diabetes complications based on the final characteristics of the patient being tested using the diabetes complication prediction model.

[0161] To verify the optimization effect of IWOA on DBN, traditional optimization algorithms (Stochastic Gradient Descent (SGD), Particle Swarm Optimization (PSO), and traditional Whale Optimization (WOA)) were selected for comparison. The experiment was based on a dataset of 1000 diabetic patients (800 for training and 200 for testing). The results are shown in Table 4.

[0162] Table 4 Model Comparison Results

[0163] algorithm Prediction accuracy (%) Mean Squared Error (MSE) Convergence iterations Local optimal trap rate (%) SGD 82.3 0.156 200 35.7 PSO 86.7 0.121 150 22.4 Traditional WOA 89.5 0.098 100 15.2 This plan 94.2 0.063 100 4.8

[0164] After IWOA optimization, the model prediction accuracy improved by 4.7%-11.9%, the MSE decreased by 34.4%-60.3%, and the local optimum trap rate dropped to 4.8%, verifying the effectiveness of the improved strategy.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described with reference to preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A system for predicting diabetic complications, characterized in that, It includes a data acquisition module, a feature selection module, a model building module, and a prediction module; The data acquisition module is used to collect individual clinical time-series data of patients; The feature screening module is used to perform preliminary feature screening on individual patient clinical time series data using Pearson correlation coefficient and mutual information, calculate the importance score of preliminary features using a temporary DBN model, determine important features based on the importance score, calculate the partial correlation coefficient of some important features with lower importance scores, and determine whether to retain the corresponding important features based on the magnitude of the partial correlation coefficient, thus obtaining the final features. The model building module is used to train and validate the DBN model optimized by IWOA based on the training set constructed from the final features, so as to obtain a diabetes complication prediction model. The prediction module is used to derive prediction results for diabetes complications based on the final characteristics of the patient being tested using the diabetes complication prediction model.

2. The diabetic complication prediction system according to claim 1, characterized in that, The data acquisition module is also used to adjust the acquisition of individual clinical time-series data of patients based on the time-series sensitivity coefficient.

3. The diabetic complication prediction system according to claim 1, characterized in that, The preliminary feature screening of individual patient clinical time-series data using Pearson correlation coefficient and mutual information includes: Pearson correlation coefficients and mutual information are calculated for all features in the individual patient's clinical time-series data. By comparing them with the set Pearson correlation coefficient thresholds and mutual information thresholds, Pearson feature sets and mutual information feature sets are obtained respectively. The common features in the Pearson feature sets and mutual information feature sets are taken as preliminary features.

4. The diabetic complication prediction system according to claim 1, characterized in that, The IWOA-optimized DBN model includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer is used to receive the final features; The first hidden layer is used to extract local features at a single time step and output the state of the first hidden layer; The second hidden layer outputs the second hidden layer state by fusing the dynamic dependencies across time steps based on the state of the first hidden layer. The output layer maps the state of the second hidden layer to the probability of complication occurrence through the Sigmoid activation function.

5. The diabetic complication prediction system according to claim 4, characterized in that, The IWOA-optimized DBN model uses the IWOA algorithm to optimize DBN parameters. The DBN parameters include the weight matrix of the input layer and the first hidden layer, the bias vector of the first hidden layer, the weight matrix of the first hidden layer and the second hidden layer, the bias vector of the second hidden layer, the weight matrix of the second hidden layer and the output layer, and the bias vector of the output layer.

6. The diabetic complication prediction system according to claim 5, characterized in that, The optimization of DBN parameters using the IWOA algorithm includes optimizing DBN parameters by simulating whale predation behavior: Each particle corresponds to a set of DBN parameters; Set the fitness function and initialize the particle swarm; For each particle in the particle swarm, the initial fitness is calculated according to the fitness function, and the particle with the lowest fitness is selected as the current best particle. The position of each particle is updated based on the prey-surrounding behavior, bubble-web predation behavior, and random search behavior. The fitness of each updated particle is calculated, and it is determined whether to replace the current best particle for iteration. Finally, the best particle is output. The final optimal particle decoding yields the optimized DBN parameters, which are used in the DBN model to obtain a predictive model for diabetes complications.

7. A method for constructing a predictive model for diabetic complications, characterized in that, include: Collect individual clinical time-series data from patients; Pearson correlation coefficient and mutual information were used to perform preliminary feature screening on individual clinical time series data of patients. An ad hoc DBN model was used to calculate the importance score of the preliminary features. Important features were determined based on the importance score. The partial correlation coefficients of some important features with lower importance scores were calculated. Based on the magnitude of the partial correlation coefficients, it was determined whether the corresponding important features should be retained, and the final features were obtained. The DBN model optimized by IWOA was trained and validated on the training set constructed from the final features to obtain a prediction model for diabetes complications.

8. The method for constructing a predictive model for diabetic complications according to claim 7, characterized in that, The preliminary feature screening of individual patient clinical time-series data using Pearson correlation coefficient and mutual information includes: Pearson correlation coefficients and mutual information are calculated for all features in the individual patient's clinical time-series data. By comparing them with the set Pearson correlation coefficient thresholds and mutual information thresholds, Pearson feature sets and mutual information feature sets are obtained respectively. The common features in the Pearson feature sets and mutual information feature sets are taken as preliminary features.

9. The method for constructing a predictive model for diabetic complications according to claim 8, characterized in that, The IWOA-optimized DBN model includes an input layer, a first hidden layer, a second hidden layer, and an output layer. The input layer is used to receive the final features; The first hidden layer is used to extract local features at a single time step and output the state of the first hidden layer; The second hidden layer outputs the second hidden layer state by fusing the dynamic dependencies across time steps based on the state of the first hidden layer. The output layer maps the state of the second hidden layer to the probability of complication occurrence through the Sigmoid activation function.

10. The method for constructing a predictive model for diabetic complications according to claim 4, characterized in that, The IWOA-optimized DBN model uses the IWOA algorithm to optimize DBN parameters. The DBN parameters include the weight matrix of the input layer and the first hidden layer, the bias vector of the first hidden layer, the weight matrix of the first hidden layer and the second hidden layer, the bias vector of the second hidden layer, the weight matrix of the second hidden layer and the output layer, and the bias vector of the output layer.

Images