Machine learning based clinical data deep mining system for acute kidney injury

Abstract

The application relates to the technical field of machine learning and clinical data processing, and particularly discloses an acute kidney injury clinical data deep mining system based on machine learning. The technical scheme of the application constructs a time sequence matrix and a semantic embedding vector by acquiring numerical and text data respectively, extracts local and trend features through multi-time scale division, fuses multi-source information by using self-attention and cross-modal attention after mapping to the same dimension, mines nonlinear correlations by means of graph attention network message passing, and performs time sequence and dilated convolution prediction in combination with a time sequence convolution network, so that the problems that a conventional method is difficult to process high-dimensional sparse clinical time sequence data, cannot capture complex nonlinear relationships and potential correlations in dynamic evolution of physiological indexes, and an existing model based on static features leads to time sequence dependence and insufficient multi-source information interaction mining are solved, and the accuracy and reliability of early warning of acute kidney injury are improved.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and clinical data processing technology, and in particular to a machine learning-based system for deep mining of clinical data on acute kidney injury. Background Technology

[0002] Acute kidney injury (AKI) is a common critical complication in clinical practice, and early identification and intervention are crucial for improving patient prognosis. With the continuous advancement of hospital information technology, electronic medical record systems have accumulated a large amount of clinical data related to AKI, covering multi-dimensional information such as laboratory tests, vital signs, medication records, and medical documents. However, existing technologies have significant limitations in deeply mining this massive amount of heterogeneous clinical data. Conventional statistical analysis methods struggle to effectively handle high-dimensional and sparse clinical time-series data, failing to fully capture the complex nonlinear relationships and potential correlation patterns inherent in the dynamic evolution of patients' physiological indicators. Simultaneously, existing machine learning models are mostly built based on static features, insufficiently exploring the temporal dependencies and interactions between multiple sources of information in clinical data, resulting in limited model prediction accuracy and generalization ability, making it difficult to meet the actual needs of early clinical warning.

[0003] Therefore, there is an urgent need to provide a technical solution to address the above problems. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a machine learning-based system for deep mining of clinical data related to acute kidney injury. The technical solution of this machine learning-based system for deep mining of clinical data related to acute kidney injury is as follows: The processing module acquires patients' numerical clinical data and text-based diagnosis and treatment data, constructs a multivariate time series matrix from the numerical clinical data, and transforms the text-based diagnosis and treatment data into semantic embedding vectors. The extraction module extracts multi-scale local features and trend features from the multivariate time series matrix and fuses fine-grained time series features to generate a multi-grained time series feature sequence. The mining module maps multi-granularity temporal feature sequences and semantic embedding vectors to the same dimension to obtain numerical feature sequences and textual feature sequences. First, it encodes them separately using a self-attention mechanism, and then uses a cross-modal attention mechanism to calculate the interaction weights between the encoded numerical features and textual features. The interaction weights are used as edge weights between nodes in the graph attention network. Through message passing, it mines the non-linear association patterns between numerical clinical data and textual diagnosis and treatment data, and generates association feature representations. The prediction module uses a temporal convolutional network to perform temporal convolution and dilated convolution on the associated feature representations, and outputs the probability of acute kidney injury.

[0005] The technical solution of this invention constructs a temporal matrix and semantic embedding vector by acquiring numerical and textual data respectively, extracts local and trend features by dividing them into multiple time scales, maps them to the same dimension, and then uses self-attention and cross-modal attention to fuse multi-source information. It also uses graph attention network message passing to mine nonlinear associations and combines temporal convolutional networks to perform temporal and dilated convolution predictions. This solves the problems that conventional methods have difficulty in handling high-dimensional sparse clinical time-series data, cannot capture the complex nonlinear relationships and potential associations in the dynamic evolution of physiological indicators, and that existing models based on static features have insufficient temporal dependence and multi-source information interaction mining. This improves the accuracy and reliability of early warning of acute kidney injury.

[0006] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an embodiment of a machine learning-based clinical data deep mining system for acute kidney injury according to the present invention. Detailed Implementation

[0009] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0010] Figure 1 This diagram illustrates the structure of an embodiment of a machine learning-based clinical data deep mining system for acute kidney injury provided by the present invention. Figure 1 As shown, this machine learning-based system for deep mining of clinical data on acute kidney injury includes: Processing module 110 acquires the patient's numerical clinical data and textual diagnosis and treatment data, constructs a multivariate time series matrix from the numerical clinical data, and converts the textual diagnosis and treatment data into semantic embedding vectors.

[0011] Numerical clinical data refers to time-series data reflecting a patient's physiological state and clinical indicators, recorded in numerical form. For example, a patient was admitted on May 1, 2026, and had their serum creatinine and hourly urine output measured daily for three consecutive days. On day 1, serum creatinine was 1.2 mg / dL and urine output was 50 mL / h; on day 2, serum creatinine was 1.8 mg / dL and urine output was 45 mL / h; and on day 3, serum creatinine was 2.5 mg / dL and urine output was 30 mL / h. These values, arranged chronologically, constitute numerical clinical data. Text-based diagnostic and treatment data refers to unstructured medical documentation generated during the patient's diagnosis and treatment process, recorded in natural language text format. For example, a patient's admission medical record stating "significantly decreased urine output in the past three days, pitting edema in both lower extremities, suggestive of progressive deterioration of renal function" is text-based diagnostic and treatment data.

[0012] A multivariate time series matrix is ​​a two-dimensional data structure that arranges the values ​​of multiple clinical indicator variables at different time steps in chronological order. For example, aligning a patient's serum creatinine and urine output values ​​for three consecutive days by time, a 3x2 multivariate time series matrix can be constructed. The first column corresponds to serum creatinine, and the second column corresponds to urine output. The first row of the matrix is ​​(1.2, 50), the second row is (1.8, 45), and the third row is (2.5, 30). Semantic embedding vectors utilize natural language processing techniques to map textual diagnostic data into a dense vector representation with fixed dimensions. For example, inputting the medical record text "Urine output has decreased significantly in the past three days, and pitting edema has appeared in both lower limbs, suggesting progressive deterioration of renal function" into a pre-trained language model, the output is a semantic embedding vector with a dimension of 768. The values ​​at different positions in the vector encode semantic information such as "decreased urine output," "edema," and "deterioration of renal function."

[0013] Before constructing the multivariate time series matrix, the processing module 110 performs missing value imputation and standardization on the numerical clinical data. The data is then aligned and arranged according to timestamps to form a multivariate time series matrix, where rows correspond to time steps and columns correspond to different clinical indicator variables. Specifically, missing value imputation uses a multiple imputation method based on expectation maximization. The missing data is imputed multiple times using the expectation maximization algorithm, and the results are combined to restore data integrity. For example, if the urine output value is missing at the second time step, it is imputed based on the serum creatinine value and the distribution of urine output values ​​before and after the time steps, generating multiple imputed values ​​and taking the average to obtain the imputed urine output value of 47 mL / h. Standardization uses Z-score standardization, which involves subtracting the mean from the data and dividing by the standard deviation to ensure the data follows a standard normal distribution with a mean of 0 and a standard deviation of 1, thus eliminating dimensional differences between different clinical indicators. For example, using the mean (1.83 mg / dL) and standard deviation (0.53 mg / dL) of the serum creatinine column in the multivariate time series matrix as standardization parameters, a Z-score standardization was performed on a serum creatinine value of 1.2 mg / dL, resulting in a standardized serum creatinine value of -1.19. This preprocessing operation restores data integrity while eliminating dimensional differences, ensuring the uniformity and comparability of numerical clinical data, thus laying a data foundation for subsequent deep feature extraction and association mining.

[0014] The processing module 110 transforms the original heterogeneous clinical data into a structured, standardized multivariate time-series matrix and semantic embedding vector through missing value imputation and standardization. This provides high-quality data input for subsequent multi-scale feature extraction and multimodal fusion, solving the problems of missing values ​​and inconsistent dimensions in the original electronic medical record data.

[0015] The extraction module 120 extracts multi-scale local features and trend features of the multivariate time series matrix and integrates fine-grained time series features to generate a multi-grained time series feature sequence.

[0016] Multiscale local features are features extracted by calculating local statistics on a multivariate time series matrix within time windows of different lengths. For example, using a 2-day time window to perform a sliding analysis on a multivariate time series matrix, calculating the mean serum creatinine (1.5 mg / dL) and mean urine output (47.5 mL / h) for the first two days, these statistics constitute multiscale local features at one time scale. Trend features reflect the direction and rate of change of clinical indicator variables within a local time window. For example, linear fitting of serum creatinine value sequences of 1.2, 1.8, and 2.5 within a 3-day time window yields a slope of 0.65 mg / dL / day. A slope of 0.65 indicates that serum creatinine is increasing at a relatively rapid rate; this slope value is a trend feature.

[0017] Fine-grained time series features are the original feature representations of each time step in a multivariate time series matrix that are not aggregated through time windows and retain the original time step granularity. For example, the feature vector of the first time step in a multivariate time series matrix is ​​obtained by encoding the serum creatinine value of 1.2 mg / dL and the urine volume value of 50 mL / h, which retains the precise value information of a single time step and is a fine-grained time series feature.

[0018] Multi-granularity temporal feature sequences are feature sequences containing information at different time scales obtained by fusing multi-scale local features, trend features, and fine-granular temporal features. For example, local statistical features extracted with a 2-day window, trend features extracted with a 3-day window, and fine-granular temporal features at the first time step are fused by cross-scale attention to obtain the fused feature vector at the first time step. This process is repeated for all time steps to generate multi-granularity temporal feature sequences.

[0019] The extraction module 120 uses multiple time windows of different lengths to segment the multivariate time series matrix and extracts local statistical and trend features under each window through a one-dimensional dilated convolutional network. The dilation rate increases with the window length. Specifically, multiple time windows of different lengths are used to slide the multivariate time series matrix. For example, a time window with a length of 2 time steps and a step size of 1 time step is used to slide the 3x2 multivariate time series matrix. The first segmentation yields the sub-matrices for the 1st and 2nd time steps, and the second segmentation yields the sub-matrices for the 2nd and 3rd time steps. The one-dimensional dilated convolutional network introduces a dilation factor in the one-dimensional convolution operation to increase the receptive field of the convolution kernel. The dilation rate increases with the window length. For example, a one-dimensional dilated convolutional network is applied to the local time series segments obtained by sliding segmentation. The dilation rate is set to 2, the convolution kernel size is 3, and the actual time steps involved in the operation are time steps with an interval of 2. The feature values ​​output by the convolution integrate information on changes in serum creatinine and urine volume over a wider time span.

[0020] Extraction module 120 further fuses fine-grained temporal features with multi-scale local features and trend features through a cross-scale attention mechanism to generate the first... The fused feature vector at each time step : ; in, For fine-grained temporal features in the first Feature vectors at each time step For the first The first time scale The feature vector is obtained by concatenating the local statistical features and trend features of each time step. The total number of time scales. , , The first Learnable projection matrices of queries, keys, and values ​​corresponding to each time scale. For each attention head, the feature dimensions, For the first The formula assigns a bias vector corresponding to each time scale. By allocating independent query, key, and value projection matrices to each time scale, it enables fine-grained time-series features to interact with local statistical and trend features at each scale through attention interaction. It calculates the semantic similarity score between fine-grained features and features at different scales, and weights and aggregates the features at each scale based on the similarity score, then superimposes the scale bias vector, thus adaptively integrating short-term fine-grained fluctuation information with long-term coarse-grained trend information.

[0021] The technical solution of extraction module 120 uses a cross-scale attention fusion mechanism to enable the fused feature sequence to retain the accurate values ​​of the original time steps and the statistical regularity and trend of change under different time window scales. This generates a multi-granular temporal feature sequence that combines detailed accuracy with macro trends, enhancing the integrity and discriminative ability of temporal feature expression, and providing feature inputs with rich temporal multi-scale expression for subsequent multi-source information association mining.

[0022] The mining module 130 maps multi-granularity temporal feature sequences and semantic embedding vectors to the same dimension to obtain numerical feature sequences and textual feature sequences. First, it encodes them separately using a self-attention mechanism, and then uses a cross-modal attention mechanism to calculate the interaction weights between the encoded numerical features and textual features. The interaction weights are used as edge weights between nodes in the graph attention network. Through message passing, it mines the nonlinear association patterns between numerical clinical data and textual diagnosis and treatment data to generate association feature representations.

[0023] Numerical feature sequences are obtained by mapping multi-granularity temporal feature sequences to a predetermined uniform dimension through linear projection. For example, inputting a multi-granularity temporal feature sequence into a fully connected projection layer for linear transformation maps the fused feature vector at each time step to a feature vector of dimension 256, resulting in a numerical feature sequence. Textual feature sequences are obtained by mapping semantic embedding vectors to a predetermined uniform dimension through linear projection. For example, inputting a semantic embedding vector of dimension 768 into a fully connected projection layer for linear transformation maps it to a textual feature sequence of dimension 256, ensuring that the feature dimensions of the textual feature sequence are consistent with those of the numerical feature sequence.

[0024] The mining module 130 utilizes independent fully connected projection layers to perform linear projections on multi-granularity temporal feature sequences and semantic embedding vectors, eliminating dimensional differences in the feature space between the two types of heterogeneous data and providing an alignment basis for cross-modal attention computation. Then, a multi-head self-attention mechanism is employed to capture temporal dependencies within numerical feature sequences and contextual dependencies within textual feature sequences, with both modalities sharing the same number of attention heads. Temporal dependencies are the sequential relationships between features at different time steps in a temporal sequence. For example, in a numerical feature sequence, the feature of increased serum creatinine at the third time step depends on the feature of progressively increasing serum creatinine at the first and second time steps. Contextual dependencies are the sequential relationships between words or semantic units at different positions in a textual sequence. For example, in a textual feature sequence, there is a modifying and modified contextual dependency between the semantic features corresponding to "kidney function" and "progressive deterioration." The number of attention heads is a parameter setting for the number of heads used in parallel computation of attention weights in the multi-head attention mechanism, for example, 8 heads for both.

[0025] The mining module 130 performs a cross-modal bilinear interaction between the query vector of the encoded numerical features and the key vector of the encoded text features. It calculates a cross-modal attention score using a learnable parameter matrix and normalizes the cross-modal attention score to obtain the interaction weight. Specifically, the interaction weight... The calculation method is as follows: ; in, For the encoded numerical feature sequence, the first... Query vectors at each time step For the encoded text feature sequence, the first... The key vector at each time step , , These are the learnable weight matrices for numerical query vectors, text key vectors, and element-wise interaction terms, respectively. This represents element-wise multiplication. For the transpose of the learnable high-level fusion weight vector, This is the bias term. The formula captures the linear contribution of numerical features to the interaction through the weight matrix of the numerical query vector, the linear contribution of textual features to the interaction through the weight matrix of the textual key vector, and the nonlinear product interaction between the two types of features through the weight matrix of the element-wise interaction term. After summing the three contributions, semantic fusion is performed using a hyperbolic tangent activation function and a high-level fusion weight vector. Then, a normalization function converts the weights into interaction weights between 0 and 1, quantifying the multidimensional association strength between different modalities from both linear and nonlinear product contribution perspectives. For example, the interaction weight calculated between the feature at the third time step in the numerical feature sequence and the feature corresponding to "progressive deterioration of kidney function" in the textual feature sequence is 0.72, indicating a strong association between significantly elevated serum creatinine and the description of deteriorating kidney function.

[0026] The mining module 130 further constructs a heterogeneous graph with numerical features and textual features at each time step as heterogeneous nodes. The interaction weights are used as edge weights connecting the two types of nodes. A multi-layer graph attention network is used for message passing and feature updates between nodes, generating an associated feature representation containing updated features of both numerical and textual nodes. A heterogeneous graph is a graph structure containing different types of nodes and edges. In this embodiment, numerical nodes are connected internally by self-attention edges, textual nodes are connected internally by self-attention edges, and numerical and textual nodes are connected by cross-modal attention edges. Message passing is the process by which nodes in a graph structure update their own feature representation based on the feature information and edge weights of their neighbors. For example, a textual node in the graph structure passes the semantic information of "deteriorating kidney function" to an adjacent numerical node with edge weights of 0.72 and 0.58. The numerical node then aggregates the messages from the textual node and updates its own features, fusing the textual semantic information. The numerical node update features, updated after message passing, simultaneously include information on the upward trend of serum creatinine and semantic information on the deterioration of renal function. The textual node update features, on the other hand, integrate the numerical changes in serum creatinine and urine output at each time step. The association feature representation is a comprehensive feature representation obtained by fusing or concatenating the numerical node features and textual node features updated by the graph attention network.

[0027] Specifically, the mining module 130 uses a cascaded structure of cross-modal attention mechanism and graph attention network to use the nonlinear interaction weights of numerical features and textual features as edge weights in the graph structure, thereby realizing semantic guidance message passing between heterogeneous nodes. This enables numerical features to dynamically integrate textual semantic information and generate associated feature representations that combine the evolution law of physiological parameters with clinical semantic descriptions.

[0028] The technical solution of the data mining module 130 utilizes a cascaded structure of cross-modal attention and graph attention networks to uncover complex nonlinear correlation patterns between numerical clinical data and textual diagnostic data. For example, as serum creatinine levels gradually increase from 1.2 mg / dL to 2.5 mg / dL, the semantic strength of the description "deteriorating renal function" in the medical record also increases accordingly, and the mapping relationship exhibits complex nonlinear characteristics. This module enables the model to fully utilize multi-source heterogeneous information in electronic medical records, overcoming the performance bottleneck of traditional single-modal modeling.

[0029] The prediction module 140 uses a temporal convolutional network to perform temporal convolution and dilated convolution on the associated feature representations and outputs the probability of acute kidney injury.

[0030] Temporal convolutional networks are neural network models that use causal convolution and dilated convolution structures to model temporal feature sequences. For example, when related feature representations are input into a temporal convolutional network, the convolution operation only depends on the information of the current time step and previous time steps, and abstract features in the time dimension are extracted layer by layer by stacking multiple residual blocks.

[0031] Temporal convolution is a one-dimensional convolution operation performed along the time dimension on temporal data. The convolution kernel only operates on features at the current and historical time steps. For example, the first layer of a temporal convolutional network uses a kernel of size 3 to perform a sliding convolution on the temporal dimension of the associated feature representations, calculating a weighted sum of the associated feature representations from the first to the third time steps, and outputting the convolutional feature values. Dilated convolution is a convolution operation that introduces a dilation factor into temporal convolution to increase the receptive field of the convolution kernel without increasing the number of parameters. For example, the second residual block of a temporal convolutional network uses a dilated causal convolution with a dilation rate of 2. The convolution kernel skips adjacent positions and directly operates on features at time step intervals of 2, allowing the output features to cover information over a longer time span.

[0032] The probability of acute kidney injury is the risk value predicted by the model for a patient to develop acute kidney injury within a preset time window in the future; for example, if the hidden state output from the last layer of a temporal convolutional network is input into a fully connected layer and a sigmoid activation function, the probability of acute kidney injury is 0.83, which means that the patient has an 83% risk of developing acute kidney injury in the next 48 hours.

[0033] The prediction module 140 employs multiple stacked residual blocks to perform temporal convolution and dilated convolution on the associated feature representations. Each residual block contains a dilated causal convolutional layer, a gated linear unit, and residual connections, with the dilation rate increasing exponentially with the residual block depth. For example, the prediction process includes three stacked residual blocks: the first residual block receives the associated feature representations and outputs the first layer residual features; the second layer receives the first layer residual features and outputs the second layer residual features; and the third layer receives the second layer residual features and outputs the final hidden state used for prediction. The dilation rate of the first layer residual block is 1, the second layer is 2, and the third layer is 4. The larger the dilation rate, the wider the historical time span covered by the convolutional kernel. The gated linear unit selectively filters information from the convolutional output through a gating mechanism, dividing the convolutional output into two parts along the feature dimension: one part serves as the information channel, and the other part serves as the gate channel. The gate channel generates gating coefficients through a sigmoid activation function. The information channel and the gating coefficients are multiplied element-wise to obtain the gated feature output. The probability of acute kidney injury (AKI) is the risk value predicted by the model for a patient to develop AKI within a preset time window. For example, if the hidden state of the last layer of a temporal convolutional network is input into a fully connected layer and a sigmoid activation function, the output probability of AKI is 0.83, indicating that the patient has an 83% risk of developing AKI within the next 48 hours.

[0034] The technical solution of prediction module 140 uses multiple stacked residual blocks and an exponentially increasing expansion rate to expand the receptive field range and enhance gradient flow, effectively capturing long-term temporal dependencies in the evolution of clinical indicators and improving the prediction accuracy of early warning of acute kidney injury.

[0035] The technical solution of this embodiment constructs a temporal matrix and semantic embedding vector by acquiring numerical and textual data respectively, extracts local and trend features by dividing them into multiple time scales, maps them to the same dimension, and then uses self-attention and cross-modal attention to fuse multi-source information. It also uses graph attention network message passing to mine nonlinear associations and combines temporal convolutional networks to perform temporal and dilated convolution predictions. This solves the problems that conventional methods have difficulty in handling high-dimensional sparse clinical time-series data, cannot capture the complex nonlinear relationships and potential associations in the dynamic evolution of physiological indicators, and that existing models based on static features have insufficient temporal dependence and multi-source information interaction mining. This improves the accuracy and reliability of early warning of acute kidney injury.

[0036] In this implementation, the cross-scale attention fusion calculation formula assigns independent query, key, and value projection matrices to each time scale. This allows fine-grained temporal features to interact with local statistical and trend features at each scale as query vectors, calculating semantic similarity scores between fine-grained features and features at different scales. Based on these similarity scores, features at each scale are weighted and aggregated, and a scale bias vector is superimposed to generate a multi-granularity temporal feature sequence. This cross-scale attention fusion calculation formula adaptively integrates short-term fine-grained fluctuation information with long-term coarse-grained trend information, ensuring that the fused feature sequence retains both the precise values ​​of the original time steps and the statistical regularities and trends at different time window scales. This provides rich multi-scale temporal expression-rich feature inputs for subsequent multi-source information association mining.

[0037] In this implementation, the cross-modal interaction weight calculation formula performs bilinear interaction between the encoded numerical feature query vector and the encoded textual feature key vector. The learnable weight matrix of the numerical query vector captures the linear contribution of numerical features to the interaction, the learnable weight matrix of the textual key vector captures the linear contribution of textual features to the interaction, and the learnable weight matrix of the element-wise interaction term captures the nonlinear product interaction between numerical and textual features. The three contributions are summed and then semantically fused using a hyperbolic tangent activation function and a high-level fusion weight vector. Finally, a normalization function converts the weights into interaction weights between 0 and 1. This cross-modal interaction weight calculation formula quantifies the multidimensional correlation strength between numerical clinical data and textual diagnostic data from both linear and nonlinear product contribution perspectives. It uncovers the potential complex correspondence between changes in physiological indicator values ​​and medical text descriptions, enabling the subsequent graph attention network to deliver messages with emphasis based on the interaction weights, thus enhancing the accuracy of multi-source information interaction modeling.

[0038] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A machine learning-based system for deep mining of clinical data on acute kidney injury, characterized in that, include: The processing module acquires patients' numerical clinical data and text-based diagnosis and treatment data, constructs a multivariate time series matrix from the numerical clinical data, and transforms the text-based diagnosis and treatment data into semantic embedding vectors. The extraction module extracts multi-scale local features and trend features from the multivariate time series matrix and fuses fine-grained time series features to generate a multi-grained time series feature sequence. The mining module maps multi-granularity temporal feature sequences and semantic embedding vectors to the same dimension to obtain numerical feature sequences and textual feature sequences. First, it encodes them separately using a self-attention mechanism, and then uses a cross-modal attention mechanism to calculate the interaction weights between the encoded numerical features and textual features. The interaction weights are used as edge weights between nodes in the graph attention network. Through message passing, it mines the non-linear association patterns between numerical clinical data and textual diagnosis and treatment data, and generates association feature representations. The prediction module uses a temporal convolutional network to perform temporal convolution and dilated convolution on the associated feature representations, and outputs the probability of acute kidney injury.

2. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 1, characterized in that, The processing module is specifically used for: The numerical clinical data is imputed and standardized. The imputed and standardized numerical clinical data is then aligned and arranged according to timestamps into a multivariate time series matrix. The rows of the multivariate time series matrix correspond to time steps, and the columns of the multivariate time series matrix correspond to different clinical indicator variables.

3. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 2, characterized in that, The processing module is also used for: The missing values ​​are filled using a multiple imputation method based on expectation maximization, and the standardization process is performed using Z-score standardization.

4. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 1, characterized in that, The extraction module is specifically used for: The multivariate time series matrix is ​​divided into sliding segments using multiple time windows of different lengths. Local statistical features and trend features under each time window are extracted by a one-dimensional dilated convolutional neural network. The dilation rate of the one-dimensional dilated convolutional neural network increases with the length of the time window.

5. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 4, characterized in that, The extraction module is also used for: Fine-grained temporal features are fused with multi-scale local features and trend features through cross-scale attention to generate the first-order feature in the multi-grained temporal feature sequence. The fused feature vectors at each time step are calculated using the following cross-scale attention fusion method: ； in, Represents the first rank in a multi-granularity temporal feature sequence. The fused feature vector at each time step Representing fine-grained temporal features in the first... Feature vectors at each time step Indicates the first The first time scale The feature vector obtained by concatenating local statistical features and trend features at each time step. Represents the total number of time scales. , , The first Learnable projection matrices of queries, keys, and values ​​corresponding to each time scale. This represents the feature dimension of each attention head. For the first The bias vector corresponding to each time scale. Represents the normalization function. This indicates the transpose operation.

6. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 1, characterized in that, The mining module is specifically used for: The multi-granularity temporal feature sequence and the semantic embedding vector are linearly projected using an independent fully connected projection layer, so that the feature dimensions of the projected numerical feature sequence and textual feature sequence are both preset uniform dimensions.

7. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 6, characterized in that, The mining module is also used for: A multi-head self-attention mechanism is used to capture the temporal dependencies within numerical feature sequences and the contextual dependencies within text feature sequences, respectively. The multi-head self-attention mechanism shares the same number of attention heads for both numerical and text feature sequences.

8. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 7, characterized in that, The mining module is also used for: The query vector of the encoded numerical features and the key vector of the encoded text features are subjected to cross-modal bilinear interaction. A cross-modal attention score is calculated using a learnable parameter matrix, and the interaction weight is obtained by normalizing the cross-modal attention score. The interaction weight is calculated as follows: ； in, Indicates the first The encoded numerical features at the first time step and the first time step Interaction weights between encoded text features at each time step Indicates the first element in the encoded numerical feature sequence. Query vector at each time step This represents the first element in the encoded text feature sequence. The key vector at each time step , , These are the learnable weight matrices for numerical query vectors, text key vectors, and element-wise interaction terms, respectively. This represents element-wise multiplication. For the transpose of the learnable high-level fusion weight vector, For bias terms, This represents the hyperbolic tangent activation function. This represents the normalization function.

9. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 8, characterized in that, The mining module is also used for: A heterogeneous graph is constructed with the numerical features and text features of each time step as heterogeneous nodes. The interaction weights are used as edge weights connecting the numerical nodes and text nodes. A multi-layer graph attention network is used to perform message passing and node feature updates between nodes on the heterogeneous graph, generating an associated feature representation that includes the updated features of both numerical and text nodes.

10. The machine learning-based clinical data deep mining system for acute kidney injury according to claim 1, characterized in that, The prediction module is specifically used for: The associated feature representation is subjected to temporal convolution and dilated convolution through multiple stacked residual blocks. Each residual block contains a dilated causal convolutional layer, a gated linear unit, and residual connections. The dilation rate of the dilated causal convolutional layer increases exponentially with the stacking depth of the residual blocks.

Images