Dialysis concurrent risk dynamic assessment system

Abstract

The invention relates to the field, and discloses a dialysis concurrent risk dynamic assessment system, which comprises a data acquisition module, a data processing module and a data processing module, the data preprocessing module is used for carrying out abnormal value detection and processing on the multi-source data by adopting an isolated forest algorithm, and filling missing values in the multi-source data through an Attention-GAN network; the feature fusion module is used for performing feature fusion on the cleaned multi-source data by using a cross-modal attention fusion algorithm to generate a comprehensive feature vector; the model construction module is used for constructing a ResNet-GRU hybrid model based on the comprehensive feature vector and introducing a dynamic feature attention mechanism; the risk assessment module is used for inputting to-be-assessed patient data into the ResNet-GRU hybrid model to obtain a risk assessment result; according to the invention, the timeliness and accuracy of evaluation are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of medical and health technology, specifically to a dynamic assessment system for dialysis complications. Background Technology

[0002] Dialysis is a vital means of sustaining life for patients with end-stage renal disease. However, during dialysis, patients often face various complications, such as dialysis-induced hypotension, muscle cramps, arrhythmias, and disequilibrium syndrome. These complications can not only affect the effectiveness of dialysis treatment but, in severe cases, even endanger the patient's life. Therefore, timely and accurate assessment of dialysis complication risks is crucial. Currently, existing dialysis risk assessment methods are mostly based on static physiological indicators and fixed assessment models. The data sources are limited, resulting in limited assessment dimensions. The fixed parameters of the assessment models cannot be dynamically adjusted according to the real-time data of patients during dialysis, making it difficult to adapt to the dynamic changes in patients' conditions and resulting in poor timeliness of assessment. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a dynamic risk assessment system for dialysis complications.

[0004] This invention provides a dynamic assessment system for dialysis complication risk, the system comprising: The data acquisition module is used to acquire multi-source data from patients in real time, including at least physiological indicator data, treatment parameter data, historical medical record data, and real-time monitoring data. The data preprocessing module is used to perform outlier detection and processing on multi-source data using the isolated forest algorithm, and to fill in missing values ​​in multi-source data using the Attention-GAN network; The feature fusion module is used to perform feature fusion on the cleaned multi-source data using a cross-modal attention fusion algorithm to generate a comprehensive feature vector. The model building module is used to construct a ResNet-GRU hybrid model based on comprehensive feature vectors and introduces a dynamic feature attention mechanism; The risk assessment module is used to input patient data to be assessed into the ResNet-GRU hybrid model to obtain risk assessment results.

[0005] Optionally, in a first implementation of the present invention, the data preprocessing module includes: The grouping submodule is used to group multi-source data according to data type, build a set of isolated trees for each data group, and calculate the path length of each data point in all isolated trees; The adjustment submodule is used to calculate an initial threshold for each data group based on the data distribution characteristics, and dynamically adjust the threshold size through an adaptive threshold adjustment mechanism. The data distribution characteristics are the mean, standard deviation, and data range. The replacement submodule is used to identify data points whose path length is less than the adjusted threshold as outliers and replace them with the median of the normal data within the time period in the data group.

[0006] Optionally, in a second implementation of the present invention, the data preprocessing module further includes: The input submodule is used to input multi-source data with missing values ​​into the Attention-GAN network to build the generator and discriminator; The determination submodule is used by the generator to determine the features and their weights associated with the missing values ​​by calculating the correlation between the features, and to generate imputation values ​​based on the features and their weights. The imputation submodule is used by the discriminator to distinguish between the generated imputation values ​​and the real data, and to complete the imputation of missing values ​​based on the imputation values.

[0007] Optionally, in a third implementation of the present invention, the feature fusion module includes: The convolutional submodule is used to perform convolution operations on time series data in multi-source data after cleaning using the TCN network through multiple convolutional layers to capture local features and long-term dependencies in the time series and obtain time series feature vectors. The semantic understanding submodule is used to convert text data into word vectors after cleaning multi-source data and input them into the BERT model. Semantic understanding and feature extraction are performed through a multi-layer Transformer structure to obtain text feature vectors. The computation submodule is used to construct a cross-modal attention mechanism, calculate the similarity between feature vectors of different modalities, determine the weight of each modal feature in the fusion process based on the similarity, and perform weighted summation based on the weights to generate a comprehensive feature vector.

[0008] Optionally, in a fourth implementation of the present invention, the model building module includes: The permutation submodule is used to arrange the comprehensive feature vectors in chronological order to form a time-series feature sequence, and to standardize the time-series feature sequence. The submodule is used to build the ResNet network part and perform static feature extraction, and to build the GRU network part and perform dynamic temporal feature extraction. The weighted fusion submodule is used to perform weighted fusion of the static feature vector extracted by ResNet and the dynamic temporal feature vector extracted by GRU through a dynamic feature attention mechanism.

[0009] Optionally, in the fifth implementation of the present invention, each residual block of the ResNet network contains two convolutional layers and one skip connection. The first convolutional layer performs preliminary feature extraction and dimensionality transformation, and the second convolutional layer further extracts deep features. The skip connection adds the original features of the input residual block to the features processed by the two convolutional layers. After multiple residual blocks are processed in sequence, the extracted features are converted into a fixed-length static feature vector through a global average pooling layer.

[0010] Optionally, in the sixth implementation of the present invention, the GRU network includes multiple GRU units, each GRU unit controls the flow of information through update gates and reset gates, and the GRU network captures the dynamic pattern of feature changes over time and outputs a dynamic temporal feature vector.

[0011] Optionally, in a seventh implementation of the present invention, the risk assessment module includes: The first extraction submodule is used to input the patient data to be evaluated into the ResNet-GRU hybrid model, and extract static features through the ResNet network to obtain static feature vectors. A submodule is formed to concatenate the static feature vector output by the ResNet network with the patient data to be evaluated to form a new temporal feature sequence; The second extraction submodule is used to input the spliced ​​new temporal feature sequence into the GRU network for dynamic feature extraction. Each GRU unit updates the state at the current time and passes the state to the GRU unit at the next time. After processing at all time steps, the GRU network outputs a dynamic temporal feature vector containing the dynamic change law of time. The generation submodule is used to input the current dialysis stage, real-time indicators, and dynamic temporal feature vector output by GRU of the patient to be evaluated into the dynamic feature attention mechanism to generate the initial weight of each feature. Then, the weight ratio is adjusted according to the dialysis stage, and the weights corresponding to static features and dynamic features are obtained after normalization. The weighting submodule is used to multiply each feature value in the static feature vector by the corresponding static feature weight and each feature value in the dynamic time series feature vector by the corresponding dynamic feature weight, and obtain the joint feature after weighting. The regression calculation submodule is used to input the joint features into the fully connected layer for regression calculation and output the risk assessment results of the patient to be evaluated.

[0012] Optionally, in the eighth implementation of the present invention, the step of inputting the joint features into the fully connected layer for regression calculation and outputting the risk assessment result of the patient to be assessed includes: Set up a multi-layer fully connected network, each layer of the fully connected network including multiple hidden layers, ReLU activation function and one output layer, with the number of neurons in the hidden layers ranging from 10 to 100; The joint feature vector is sequentially transformed and nonlinearly activated through each layer of the fully connected network, and the output layer uses the Softmax activation function to output the risk assessment results of the patient to be evaluated.

[0013] Optionally, in the ninth implementation of the present invention, a method for implementing a dynamic risk assessment system for dialysis complications includes the following steps: Real-time acquisition of multi-source patient data, including at least physiological indicator data, treatment parameter data, historical medical record data, and real-time monitoring data; The Isolation Forest algorithm is used to detect and process outliers in multi-source data, and the Attention-GAN network is used to fill in missing values ​​in the multi-source data. A cross-modal attention fusion algorithm is used to fuse features from cleaned multi-source data to generate a comprehensive feature vector. A ResNet-GRU hybrid model is constructed based on comprehensive feature vectors, and a dynamic feature attention mechanism is introduced; The patient data to be evaluated is input into the ResNet-GRU hybrid model to obtain the risk assessment results.

[0014] The technical solution provided by this invention involves real-time acquisition of multi-source patient data, employing an isolated forest algorithm for outlier detection and processing, and using an Attention-GAN network to fill in missing values ​​in the multi-source data. A cross-modal attention fusion algorithm is used to fuse features in the cleaned multi-source data, generating a comprehensive feature vector. A ResNet-GRU hybrid model is constructed based on the comprehensive feature vector, and a dynamic feature attention mechanism is introduced. The patient data to be evaluated is input into the ResNet-GRU hybrid model to obtain the risk assessment results. This invention integrates multi-source data such as physiological indicators, treatment parameters, and historical medical records, overcoming the limitation of single data sources in existing technologies and improving the comprehensiveness of the assessment. The ResNet-GRU hybrid model and dynamic feature attention mechanism achieve dynamic and personalized risk assessment, solving the problem of fixed model parameters in existing technologies and significantly improving the timeliness and accuracy of the assessment. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1This is a schematic diagram of the structure of the dynamic assessment system for dialysis complications risk provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the data preprocessing module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the feature fusion module provided in an embodiment of the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 A schematic diagram of the structure of the dynamic assessment system for dialysis complication risk provided in this embodiment of the invention. The system includes: The data acquisition module is used to acquire multi-source data from patients in real time, including at least physiological indicator data, treatment parameter data, historical medical record data, and real-time monitoring data. The data preprocessing module is used to perform outlier detection and processing on multi-source data using the isolated forest algorithm, and to fill in missing values ​​in multi-source data using the Attention-GAN network; The feature fusion module is used to perform feature fusion on the cleaned multi-source data using a cross-modal attention fusion algorithm to generate a comprehensive feature vector. The model building module is used to construct a ResNet-GRU hybrid model based on comprehensive feature vectors and introduces a dynamic feature attention mechanism; The risk assessment module is used to input patient data to be assessed into the ResNet-GRU hybrid model to obtain risk assessment results.

[0019] In this embodiment, please refer to Figure 2 The data preprocessing module includes: The grouping submodule is used to group multi-source data according to data type, build a set of isolated trees for each data group, and calculate the path length of each data point in all isolated trees; The adjustment submodule is used to calculate an initial threshold for each data group based on the data distribution characteristics, and dynamically adjust the threshold size through an adaptive threshold adjustment mechanism. The data distribution characteristics are the mean, standard deviation, and data range. The replacement submodule is used to identify data points whose path length is less than the adjusted threshold as outliers and replace the outliers with the median of the normal data within the time period in the data group. The input submodule is used to input multi-source data with missing values ​​into the Attention-GAN network to build the generator and discriminator; The determination submodule is used by the generator to determine the features and their weights associated with the missing values ​​by calculating the correlation between the features, and to generate imputation values ​​based on the features and their weights. The imputation submodule is used by the discriminator to distinguish between the generated imputation values ​​and the real data, and to complete the imputation of missing values ​​based on the imputation values.

[0020] In this embodiment, please refer to Figure 3 The feature fusion module includes: The convolutional submodule is used to perform convolution operations on time series data in multi-source data after cleaning using the TCN network through multiple convolutional layers to capture local features and long-term dependencies in the time series and obtain time series feature vectors. The semantic understanding submodule is used to convert text data into word vectors after cleaning multi-source data and input them into the BERT model. Semantic understanding and feature extraction are performed through a multi-layer Transformer structure to obtain text feature vectors. The computation submodule is used to construct a cross-modal attention mechanism, calculate the similarity between feature vectors of different modalities, determine the weight of each modal feature in the fusion process based on the similarity, and perform weighted summation based on the weights to generate a comprehensive feature vector.

[0021] In this embodiment, the model building module includes: The permutation submodule is used to arrange the comprehensive feature vectors in chronological order to form a time-series feature sequence, and to standardize the time-series feature sequence. The submodule is used to build the ResNet network part and perform static feature extraction, and to build the GRU network part and perform dynamic temporal feature extraction. The weighted fusion submodule is used to perform weighted fusion of the static feature vector extracted by ResNet and the dynamic temporal feature vector extracted by GRU through a dynamic feature attention mechanism.

[0022] In this embodiment, each residual block of the ResNet network contains two convolutional layers and one skip connection. The first convolutional layer performs preliminary feature extraction and dimensionality transformation, and the second convolutional layer further extracts deep features. The skip connection adds the original features of the input residual block to the features processed by the two convolutional layers. After multiple residual blocks are processed in sequence, the extracted features are converted into a fixed-length static feature vector through a global average pooling layer.

[0023] In this embodiment, the GRU network includes multiple GRU units. Each GRU unit controls the flow of information through update gates and reset gates. The GRU network captures the dynamic patterns of feature changes over time and outputs a dynamic temporal feature vector.

[0024] In this embodiment, the risk assessment module includes: The first extraction submodule is used to input the patient data to be evaluated into the ResNet-GRU hybrid model, and extract static features through the ResNet network to obtain static feature vectors. A submodule is formed to concatenate the static feature vector output by the ResNet network with the patient data to be evaluated to form a new temporal feature sequence; The second extraction submodule is used to input the spliced ​​new temporal feature sequence into the GRU network for dynamic feature extraction. Each GRU unit updates the state at the current time and passes the state to the GRU unit at the next time. After processing at all time steps, the GRU network outputs a dynamic temporal feature vector containing the dynamic change law of time. The generation submodule is used to input the current dialysis stage, real-time indicators, and dynamic temporal feature vector output by GRU of the patient to be evaluated into the dynamic feature attention mechanism to generate the initial weight of each feature. Then, the weight ratio is adjusted according to the dialysis stage, and the weights corresponding to static features and dynamic features are obtained after normalization. The weighting submodule is used to multiply each feature value in the static feature vector by the corresponding static feature weight and each feature value in the dynamic time series feature vector by the corresponding dynamic feature weight, and obtain the joint feature after weighting. The regression calculation submodule is used to input the joint features into the fully connected layer for regression calculation and output the risk assessment results of the patient to be evaluated.

[0025] In this embodiment, a multi-layer fully connected network is set up. Each layer of the fully connected network includes multiple hidden layers, a ReLU activation function, and an output layer. The number of neurons in the hidden layer is 10-100. The joint feature vector is sequentially passed through each layer of the fully connected network for linear transformation and non-linear activation. The output layer uses the Softmax activation function to output the risk assessment results of the patient to be evaluated. The number of network layers is determined based on actual needs, typically 3-5 layers. Each fully connected layer contains multiple hidden layers and one output layer. For each hidden layer, the number of neurons is set between 10-100, and the number of neurons can be gradually adjusted from input to output according to changes in feature dimension, such as decreasing from 64 to 32, to achieve feature dimensionality reduction and abstraction. In each hidden layer, the input features are first linearly transformed by multiplying the weight matrix with the input feature vector and adding a bias term. Then, a ReLU activation function is applied to the result of the linear transformation, performing non-linear processing to filter out negative values ​​and retain positive features, enhancing the network's ability to fit non-linear relationships. The final layer is the output layer, with the number of neurons corresponding to the number of risk assessment categories, such as 3 neurons for low risk, medium risk, and high risk. The joint feature vector obtained after preprocessing is input into the first hidden layer of a pre-configured multilayer fully connected network. A linear transformation is first performed to map the feature vector to the neuron dimension of this hidden layer. Then, a ReLU activation function is applied for non-linear activation, yielding the output feature of the first hidden layer. This output feature serves as the input to the next hidden layer, undergoing the same linear transformation and ReLU activation operations repeatedly. After multiple layers of processing, the features are continuously refined and abstracted before finally being passed to the output layer. The output layer performs a linear transformation on the output features of the previous layer and then applies a Softmax activation function, converting the output values ​​into a probability distribution. Each probability value corresponds to a probability of a risk level, with the category with the highest probability representing the risk assessment result for the patient being evaluated.

[0026] In this embodiment, real-time acquisition of multi-source patient data is the foundation for the system's dynamic assessment. Physiological indicators include key metrics monitored in real-time by various sensors during dialysis, such as blood pressure, heart rate, blood oxygen saturation, body temperature, respiratory rate, electrolyte levels, blood glucose, hemoglobin, and platelet count. This data directly reflects the patient's current physiological state. Treatment parameter data comes from real-time records of the dialysis equipment, including parameters related to treatment operations such as dialysate flow rate, concentration, temperature, blood flow rate, ultrafiltration rate, ultrafiltration volume, dialyzer type and area, and anticoagulant type and dosage, reflecting the implementation of the treatment plan. Historical medical record data is extracted from the hospital's electronic medical record system, including the patient's age, gender, height, weight, duration of illness, underlying diseases, number and duration of previous dialysis treatments, records of complications, and long-term accumulated information such as drug allergy history and surgical history, providing historical reference for risk assessment.

[0027] In this embodiment, the real-time acquired multi-source data is first sorted and divided into different categories based on the essential attributes of the data. For example, blood pressure and heart rate are classified into the physiological indicator data group, dialysis fluid flow rate and ultrafiltration rate are classified into the treatment parameter data group, patient age and underlying diseases are classified into the historical medical record data group, and limb activity and facial expression are classified into the real-time monitoring data group. Then, for each classified data group, a set of multiple isolated trees is independently constructed. During the construction of each isolated tree, a feature and a segmentation value of that feature are randomly selected, and the data is recursively divided until each leaf node contains only one data point or a preset depth is reached. Finally, each data point is traversed, and the path length from the root node to the leaf node in all isolated trees of its data group is calculated. This path length is used as the basic indicator for judging whether the data point is abnormal. For each grouped data set, the mean, standard deviation, and data range (the difference between the maximum and minimum values) are first calculated statistically. These data distribution characteristics become the key basis for calculating the initial threshold. For example, the approximate range of the initial threshold can be determined based on the difference between the mean and three times the standard deviation. Then, an adaptive threshold adjustment mechanism is activated. Combining the distribution pattern of outliers in the historical data of the data set and the fluctuation characteristics of the data in the current dialysis stage, the initial threshold is dynamically fine-tuned. For example, when the data of a data set fluctuates greatly in the early stage of dialysis, the threshold is appropriately increased to reduce false positives. When the data tends to stabilize, the threshold is lowered to improve the sensitivity of anomaly detection, ensuring that the threshold can adapt to the dynamic changes of the data. The path length of each data point in the isolated tree set of its data group is compared with the threshold finally determined by the adjustment submodule. If the path length is less than the threshold, the data point is identified as an outlier. Then, all normal data that are in the same time period as the outlier (i.e., data with a path length greater than or equal to the threshold) are selected from the data group. The median of these normal data is calculated, and finally, the corresponding outlier is replaced with the median. This ensures the continuity and reliability of the data within the data group and avoids outliers from interfering with subsequent data processing steps. We collect multi-source data that still contains missing values ​​after outlier handling. We then perform standardized preprocessing on this data to convert data of different types and magnitudes into a uniform scale. Subsequently, we input the preprocessed data containing missing values ​​into the Attention-GAN network and construct two core components of the network: a generator and a discriminator. The generator is responsible for learning the distribution patterns of the data and generating possible imputation values, while the discriminator focuses on distinguishing the imputation values ​​generated by the generator from the real data. The two components continuously optimize their performance through adversarial training, laying the foundation for accurate imputation of missing values. After receiving data containing missing values, the generator calculates the correlation coefficients between various features using its built-in attention mechanism module, identifying features that are logically closely related to the missing values ​​(for example, when blood pressure data is missing, the correlation with features such as heart rate and blood oxygen saturation will be emphasized). Based on the strength of the correlation, these related features are assigned corresponding weights, with higher-relevance features having greater weights. Then, based on the information contained in these high-weighted related features and combined with the overall distribution trend of the data, the generator generates imputation values ​​that match the location of the missing values ​​and the data patterns, ensuring that the imputation values ​​are reasonable. The discriminator receives the imputed value output by the generator and the existing real data in the data set. It then uses a multi-layer neural network to extract and compare features from both, determining whether the input data is real data or the generated imputed value. The discriminator feeds back the judgment result to the generator in the form of a loss value. The generator adjusts its parameters based on the loss value, continuously optimizing the imputed value generation strategy to make the generated imputed value increasingly closer to the distribution of the real data. When the discriminator can no longer effectively distinguish between the imputed value and the real data, i.e., when the difference between the two reaches a preset minimum threshold, training stops. At this point, the imputed value generated by the generator is used to replace the missing parts in the data, completing the final imputed value.

[0028] In this embodiment, time-series data from multi-source data that has been cleaned, including outlier handling and missing value imputation, such as data on the changes in blood pressure, heart rate, and dialysate flow over time, are organized into a sequence form with fixed time intervals. Subsequently, this time-series data is input into a TCN network, which consists of multiple convolutional layers. Each convolutional layer uses a convolutional kernel of a different size and performs convolution operations on the time-series data using a sliding window. The features output by the previous convolutional layer are used as the input to the next convolutional layer. Through the superposition of multiple convolutions, it can capture both local features between adjacent data points in the time series, such as short-term blood pressure fluctuations, and long-term dependencies, such as the overall trend of heart rate changes throughout the dialysis process, by expanding the receptive field. Finally, after processing by the last convolutional layer, a time-series feature vector with uniform dimensions is output. For the text data in the multi-source data after cleaning, such as the historical medical record descriptions, diagnosis conclusions, chief complaint symptom records of patients, etc., first perform text preprocessing, including removing punctuation marks, stop words such as meaningless words like "of" and "in", and performing word form reduction, etc.; then, convert the processed text data into word vectors, where each word corresponds to a vector of a fixed dimension, representing the semantic information of the word in numerical form; next, input the word vector sequence into the BERT model, which contains multiple layers of Transformer structures. Each layer of Transformer calculates the degree of association between each word and all other words through the self-attention mechanism, so as to deeply understand the context semantics of the text; after the feature extraction and semantic understanding of multiple layers of Transformer, take the vector corresponding to the [CLS] token output by the last layer of the model as the semantic representation of the entire text, that is, obtain the text feature vector; Construct a cross-modal attention mechanism, which contains three matrices of query, key, and value, respectively used to transform the feature vectors of different modalities; for the time series feature vectors, text feature vectors, and feature vectors of other modalities, take the feature vectors of one modality as the query vector, and the feature vectors of other modalities as the key vector and value vector respectively, and calculate the similarity between the query vector and each key vector such as cosine similarity through matrix multiplication; perform weighted summation on the value vectors according to the calculated similarity to obtain the attention output corresponding to the feature vector of this modality; then, take the feature vectors of other modalities as the query vector in turn, repeat the above process, and obtain the attention outputs of all modality feature vectors; after that, determine their weights in the fusion process according to the average similarity between each modality feature vector and other modalities. The higher the similarity, the greater the weight of the modality; finally, perform weighted summation on the attention outputs of all modalities according to the corresponding weights to generate a comprehensive feature vector with a unified dimension, realizing the effective fusion of multi-modal features.

[0029] In this embodiment, each comprehensive feature vector is labeled with a corresponding risk label, such as whether complications have occurred, the type of complication, and its severity. The dataset is divided into a training set and a validation set in a 7:3 ratio, where the training set is used for model parameter learning and the validation set is used to monitor the training effect. The weights and biases of each convolutional layer and skip connection layer in the ResNet part are randomly initialized using a normal distribution to generate initial values. The update gate, reset gate, and weight matrix and bias vector of the hidden state in the GRU part are also initialized using a random normal distribution. The parameters of the fully connected layer used to calculate the weights in the dynamic feature attention mechanism are initialized to ensure that the parameters of each part are within a reasonable initial range. The total number of training rounds and the batch size of each round are set, and comprehensive feature vectors are extracted from the training set in batches and input into the model. For each batch of data, it is first input into the ResNet network, and static features are extracted through convolution operations and skip connections of multiple residual blocks to output static feature vectors. Then, the static feature vectors are concatenated with the original temporal feature sequence and input into the GRU network. Through the update gate and reset gate at each time step, dynamic temporal feature vectors are output. After feature extraction for each batch of data, the dynamic temporal feature vector output by GRU, the patient dialysis stage information corresponding to the batch of data such as the dialysis period extracted from the data label, and the key real-time indicator features are input into the attention weight calculation module. The initial weight of each feature is calculated through a fully connected layer, and the weight ratio is adjusted according to the dialysis stage. Finally, the dynamic feature weight is obtained by normalization. The static feature vector output by ResNet and the dynamic temporal feature vector output by GRU are weighted and fused according to the dynamic feature weights to obtain a fused feature vector. The fused feature vector is then input into the output layer to obtain the model's risk prediction result for this batch of data. The cross-entropy loss function is used to compare the prediction result with the true risk label to calculate the loss value for this batch. Based on the calculated loss value, the gradient descent optimization algorithm is used to calculate the gradient of each parameter. The gradient is then backpropagated from the output layer to the fully connected layer of the dynamic feature attention mechanism, the weight matrices of the GRU network, and the convolutional layer of the ResNet network. All parameters are adjusted according to the gradient direction to reduce the loss value. After each training epoch, the model is evaluated using a validation set. The comprehensive feature vector of the validation set is input into the model, and the accuracy, recall, and other evaluation metrics are calculated after obtaining the prediction results. If the validation set metrics of the current epoch are better than the historical best values, the current model parameters are saved as a candidate optimal model. If the validation set metrics do not improve for several consecutive epochs, an early stopping mechanism is triggered to terminate the training. When the training reaches the preset number of epochs or the early stopping mechanism is triggered, the iteration stops, and the model parameters with the best performance on the validation set are selected as the final training result, resulting in a trained ResNet-GRU hybrid model.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic risk assessment system for dialysis complications, characterized in that, The system includes: The data acquisition module is used to acquire multi-source data from patients in real time, including at least physiological indicator data, treatment parameter data, historical medical record data, and real-time monitoring data. The data preprocessing module is used to perform outlier detection and processing on multi-source data using the isolated forest algorithm, and to fill in missing values ​​in multi-source data using the Attention-GAN network; The feature fusion module is used to perform feature fusion on the cleaned multi-source data using a cross-modal attention fusion algorithm to generate a comprehensive feature vector. The model building module is used to construct a ResNet-GRU hybrid model based on comprehensive feature vectors and introduces a dynamic feature attention mechanism; The risk assessment module is used to input patient data to be assessed into the ResNet-GRU hybrid model to obtain risk assessment results.

2. The dynamic risk assessment system for dialysis complications as described in claim 1, characterized in that, The data preprocessing module includes: The grouping submodule is used to group multi-source data according to data type, build a set of isolated trees for each data group, and calculate the path length of each data point in all isolated trees; The adjustment submodule is used to calculate an initial threshold for each data group based on the data distribution characteristics, and dynamically adjust the threshold size through an adaptive threshold adjustment mechanism. The data distribution characteristics are the mean, standard deviation, and data range. The replacement submodule is used to identify data points whose path length is less than the adjusted threshold as outliers and replace them with the median of the normal data within the time period in the data group.

3. The dynamic risk assessment system for dialysis complications as described in claim 1, characterized in that, The data preprocessing module further includes: The input submodule is used to input multi-source data with missing values ​​into the Attention-GAN network to build the generator and discriminator; The determination submodule is used by the generator to determine the features and their weights associated with the missing values ​​by calculating the correlation between the features, and to generate imputation values ​​based on the features and their weights. The imputation submodule is used by the discriminator to distinguish between the generated imputation values ​​and the real data, and to complete the imputation of missing values ​​based on the imputation values.

4. The dynamic risk assessment system for dialysis complications as described in claim 1, characterized in that, The feature fusion module includes: The convolutional submodule is used to perform convolution operations on time series data in multi-source data after cleaning using the TCN network through multiple convolutional layers to capture local features and long-term dependencies in the time series and obtain time series feature vectors. The semantic understanding submodule is used to convert text data into word vectors after cleaning multi-source data and input them into the BERT model. Semantic understanding and feature extraction are performed through a multi-layer Transformer structure to obtain text feature vectors. The computation submodule is used to construct a cross-modal attention mechanism, calculate the similarity between feature vectors of different modalities, determine the weight of each modal feature in the fusion process based on the similarity, and perform weighted summation based on the weights to generate a comprehensive feature vector.

5. The dynamic risk assessment system for dialysis complications as described in claim 1, characterized in that, The model building module includes: The permutation submodule is used to arrange the comprehensive feature vectors in chronological order to form a time-series feature sequence, and to standardize the time-series feature sequence. The submodule is used to build the ResNet network part and perform static feature extraction, and to build the GRU network part and perform dynamic temporal feature extraction. The weighted fusion submodule is used to perform weighted fusion of the static feature vector extracted by ResNet and the dynamic temporal feature vector extracted by GRU through a dynamic feature attention mechanism.

6. The dynamic risk assessment system for dialysis complications as described in claim 5, characterized in that, Each residual block in the ResNet network contains two convolutional layers and one skip connection. The first convolutional layer performs preliminary feature extraction and dimensionality transformation, and the second convolutional layer further extracts deeper features. The skip connection adds the original features of the input residual block to the features processed by the two convolutional layers. After multiple residual blocks are processed in sequence, the extracted features are converted into a fixed-length static feature vector through a global average pooling layer.

7. The dynamic risk assessment system for dialysis complications as described in claim 5, characterized in that, The GRU network comprises multiple GRU units. Each GRU unit controls the flow of information through update gates and reset gates. The GRU network captures the dynamic patterns of feature changes over time and outputs a dynamic temporal feature vector.

8. The dynamic risk assessment system for dialysis complications as described in claim 1, characterized in that, The risk assessment module includes: The first extraction submodule is used to input the patient data to be evaluated into the ResNet-GRU hybrid model, and extract static features through the ResNet network to obtain static feature vectors. A submodule is formed to concatenate the static feature vector output by the ResNet network with the patient data to be evaluated to form a new temporal feature sequence; The second extraction submodule is used to input the spliced ​​new temporal feature sequence into the GRU network for dynamic feature extraction. Each GRU unit updates the state at the current time and passes the state to the GRU unit at the next time. After processing at all time steps, the GRU network outputs a dynamic temporal feature vector containing the dynamic change law of time. The generation submodule is used to input the current dialysis stage, real-time indicators, and dynamic temporal feature vector output by GRU of the patient to be evaluated into the dynamic feature attention mechanism to generate the initial weight of each feature. Then, the weight ratio is adjusted according to the dialysis stage, and the weights corresponding to static features and dynamic features are obtained after normalization. The weighting submodule is used to multiply each feature value in the static feature vector by the corresponding static feature weight and each feature value in the dynamic time series feature vector by the corresponding dynamic feature weight, and obtain the joint feature after weighting. The regression calculation submodule is used to input the joint features into the fully connected layer for regression calculation and output the risk assessment results of the patient to be evaluated.

9. The dynamic risk assessment system for dialysis complications as described in claim 8, characterized in that, The process of inputting joint features into a fully connected layer for regression calculation and outputting risk assessment results for the patient to be evaluated includes: Set up a multi-layer fully connected network, each layer of the fully connected network including multiple hidden layers, ReLU activation function and one output layer, with the number of neurons in the hidden layers ranging from 10 to 100; The joint feature vector is sequentially transformed and nonlinearly activated through each layer of the fully connected network, and the output layer uses the Softmax activation function to output the risk assessment results of the patient to be evaluated.

10. A method for implementing a dynamic risk assessment system for dialysis complications as described in claim 1, characterized in that, The method includes the following steps: Real-time acquisition of multi-source patient data, including at least physiological indicator data, treatment parameter data, historical medical record data, and real-time monitoring data; The Isolation Forest algorithm is used to detect and process outliers in multi-source data, and the Attention-GAN network is used to fill in missing values ​​in the multi-source data. A cross-modal attention fusion algorithm is used to fuse features from cleaned multi-source data to generate a comprehensive feature vector. A ResNet-GRU hybrid model is constructed based on comprehensive feature vectors, and a dynamic feature attention mechanism is introduced; The patient data to be evaluated is input into the ResNet-GRU hybrid model to obtain the risk assessment results.

Images