Multimodal feature fusion-based respiratory system disease kidney injury risk intelligent early warning system

Abstract

The present application relates to the technical field of medical health information processing, in particular to a respiratory system disease kidney injury risk intelligent early warning system based on multi-modal feature fusion, which comprises: a feature extraction module for extracting respiratory function time domain and frequency domain features and kidney function trend features; a fusion prediction module for inputting the features and individual clinical static features into an improved multi-modal feature fusion prediction model; the model is based on a deep neural network, dynamically fuses and correlates the heterogeneous modal features through a cross-attention mechanism, and outputs an individualized probability risk value of the patient suffering from acute kidney injury in a future time window; when the risk value exceeds a threshold, a risk attribution module identifies a key feature subset driving the risk increase according to an explainability mechanism in the model; and a report generation module generates a risk early warning report containing a targeted intervention prompt based on the key feature subset and a clinical decision rule base.

Description

Technical Field

[0001] This invention relates to the field of medical and health information processing technology, and in particular to an intelligent early warning system for the risk of kidney injury in respiratory diseases based on multimodal feature fusion. Background Technology

[0002] In the clinical management of patients with respiratory diseases, acute kidney injury is a common and serious complication, and early identification of high-risk patients is crucial for improving prognosis. Existing risk warning methods mainly rely on static biochemical indicators reflecting kidney function, such as serum creatinine, for threshold judgment. These methods are not sensitive to early and subtle changes in kidney function, resulting in warning delays. With the development of medical informatization, methods have emerged that utilize electronic medical record data to build predictive models. These models have begun to incorporate multimodal information such as vital sign time-series data and laboratory test results. However, at the feature processing level, existing methods typically extract statistical features of various data independently or perform simple concatenation, failing to fully consider the complex dynamic relationship between respiratory function fluctuations (such as time-domain and frequency-domain features) and trends in kidney function. This simple feature aggregation approach cannot fully explore the deep interaction information between heterogeneous modal data, limiting the model's ability to capture early and hidden risks. Furthermore, most deep learning-based predictive models are "black boxes," and their decision-making processes lack interpretability. When the model outputs a high-risk warning, clinicians cannot identify which specific abnormal signs or indicators drove this judgment. This results in warning reports lacking targeted intervention suggestions, insufficient clinical operability, and continued confusion for doctors in decision-making. Therefore, current technology needs to address how to perform deep dynamic fusion and correlation modeling of multi-source heterogeneous temporal features such as respiratory and renal function data, and how to provide clinically targeted risk attribution explanations while outputting risk probabilities. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a multimodal feature fusion-based intelligent early warning system for kidney damage risk in respiratory diseases.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a multimodal feature fusion-based intelligent early warning system for kidney injury risk in respiratory diseases, comprising: The feature extraction module is used to extract time-domain and frequency-domain features that reflect fluctuations in respiratory function, as well as trend features that reflect subtle changes in renal function. The fusion prediction module inputs the time-domain and frequency-domain features, trend features, and static features from the individual's clinical background data into an improved multimodal feature fusion prediction model. The improved multimodal feature fusion prediction model is based on a deep neural network architecture and uses a cross-attention mechanism to dynamically fuse and model the correlation of heterogeneous modal features. The improved multimodal feature fusion prediction model outputs an individualized probability risk value for the target patient to develop acute kidney injury within a specified future time window. The risk attribution module is used to identify the key feature subset that drives the increase in the individualized probability risk value when the individualized probability risk value exceeds a preset decision threshold, based on the interpretability mechanism inside the improved multimodal feature fusion prediction model. The report generation module generates a risk warning report containing targeted clinical intervention suggestions based on the aforementioned key feature subset and a preset clinical decision rule base.

[0005] As a further aspect of the present invention, the extraction of time-domain and frequency-domain features reflecting fluctuations in respiratory function, and trend features reflecting subtle changes in renal function, includes: Acquire respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data from target patients. The respiratory physiological monitoring data includes dynamic blood oxygen saturation time-series signals and exhaled gas molecular profiles. The renal function biochemical data includes serum creatinine time-series changes and urine output monitoring data. The respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data are subjected to multi-timescale alignment and outlier filtering to generate a synchronized multimodal time series dataset. From the synchronized multimodal time series dataset, time-domain and frequency-domain features reflecting fluctuations in respiratory function, as well as trend features reflecting subtle changes in kidney function, are extracted respectively.

[0006] As a further aspect of the present invention, the respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data are subjected to multi-timescale alignment and outlier filtering to generate a synchronized multimodal time series dataset, including: High-frequency sampling data streams in the respiratory physiological monitoring data are identified, and the high-frequency sampling data streams are downsampled to the same daily recording frequency as the renal function biochemical data using a sliding window averaging method. The static information in the personal clinical background data is expanded into a sequence with constant values ​​and the same time length as the renal function biochemical data; Establish a unified timestamp index, with the minimum time unit of the timestamp index set to days; The downsampled respiratory physiological monitoring data, the original renal function biochemical data, and the expanded personal clinical background data sequence are spliced ​​and aligned according to the unified timestamp index. Outlier detection is performed on the spliced ​​and aligned data. The statistical local outlier factor method is used to identify outlier data points that deviate from the statistical distribution characteristics of adjacent data points. The abnormal value point is replaced by the linear interpolation value of the value of the abnormal value point at the previous time and the value at the next time, or by the overall median of the data stream to which the abnormal value point belongs in the respiratory physiological monitoring data, the renal function biochemical data and the personal clinical background data, to complete the outlier filtering. The data after outlier filtering is encapsulated into a synchronized multimodal time series dataset with a unified time step and feature channel.

[0007] As a further aspect of the present invention, the improved multimodal feature fusion prediction model is based on a deep neural network architecture and includes: A shared initial feature embedding layer is used to perform a preliminary nonlinear transformation on the original numerical sequence of each modality in the synchronized multimodal time series dataset to obtain the initial feature representation of each modality. A modal deep feature encoder, comprising a convolutional neural network branch for respiratory physiological monitoring data, a gated recurrent unit network branch for renal function biochemical data, and a multilayer perceptron branch for individual clinical background data, is used to extract high-order abstract feature representations of each modality of data respectively. The cross-attention mechanism module is used to calculate the dynamic correlation weight between the respiratory physiological modality feature representation and the renal function modality feature representation, and to perform weighted fusion of the features of the two modalities based on the dynamic correlation weight to generate cross-modal interactive features. A multi-level feature aggregation module is used to concatenate the cross-modal interaction features with the deep features of the individual clinical background data and aggregate them through a feedforward neural network to generate a comprehensive patient state vector. A risk prediction head composed of a multilayer sensing mechanism receives the comprehensive patient state vector and outputs the individualized probability risk value through a stack of fully connected layers and nonlinear activation functions.

[0008] As a further aspect of the present invention, based on the interpretability mechanism within the improved multimodal feature fusion prediction model, a subset of key features driving the increase in the individualized probability risk value is identified, including: After the improved multimodal feature fusion prediction model completes the forward propagation calculation on the synchronized multimodal time series dataset of a specific patient, the activation value vector of the layer before the final fully connected layer of the risk prediction head is obtained. For each dimension of the activation vector, calculate its gradient with respect to each input element of the integrated patient state vector; The absolute value of the calculated gradient is taken, and the absolute value is averaged over all time steps of the synchronized multimodal time series dataset to obtain the average importance score of each input feature to the final risk prediction decision. From all features of the respiratory physiology modality, renal function modality, and personal clinical background modality, a specified number of features with the highest average importance score are selected to form the key feature subset.

[0009] As a further aspect of the present invention, based on the aforementioned subset of key features and combined with a preset clinical decision rule base, a risk warning report containing targeted clinical intervention prompts is generated, including: From the aforementioned key feature subset, features belonging to the respiratory physiological modality and features belonging to the renal function modality are extracted respectively; From the preset clinical decision rule base, search for respiratory system abnormality pattern definitions that match the extracted features belonging to respiratory physiological modalities, and obtain a set of candidate intervention suggestions associated with the respiratory system abnormality pattern from the rule base; From the preset clinical decision rule base, search for the definition of abnormal renal function patterns that match the extracted features belonging to the renal function modality, and obtain another set of candidate intervention suggestions associated with the abnormal renal function patterns from the rule base; The two sets of candidate intervention suggestions were deduplicated and merged, and then sorted according to the preset priority of intervention measures. The sorted list of intervention recommendations, along with the individualized probability risk value, the specific feature names and their values ​​in the subset of key features that drive the increase of the individualized probability risk value, are integrated into a structured risk warning report.

[0010] As a further aspect of the present invention, time-domain and frequency-domain features reflecting respiratory function fluctuations are extracted from the synchronized multimodal time-series dataset, including: For the dynamic blood oxygen saturation time series signal, calculate its mean, minimum, maximum and standard deviation in each time window to obtain a set of time domain statistical characteristics; For the dynamic blood oxygen saturation time-series signal, apply Fast Fourier Transform to calculate its power spectral density in a specific physiological frequency band to obtain frequency domain energy characteristics; As a further aspect of the present invention, trend features reflecting subtle changes in kidney function are extracted from the synchronized multimodal time series dataset, including: For the time-series changes in serum creatinine, calculate the cumulative change value, the rate of change, and the relative change multiple relative to the individual baseline level within a preset review time window; For the urine monitoring data, calculate the total daily urine volume and urine volume per unit body weight, and identify the number of consecutive hours when the urine volume is below a certain threshold; Calculate the cross-correlation coefficient between the rate of change in serum creatinine and the change in urine volume; The cumulative change in serum creatinine, the rate of change, the relative change fold, the total daily urine volume, the urine volume per unit body weight, the duration of low urine volume, and the cross-correlation coefficient are combined to form the trend characteristics reflecting subtle changes in renal function.

[0011] As a further aspect of the present invention, the specific working process of the cross-attention mechanism module includes: Linear projection is performed on the deep feature representations from the respiratory physiological monitoring data branch and the deep feature representations from the renal function biochemical data branch to generate query vector sequences, key vector sequences and value vector sequences. The query vector sequence generated from the feature mapping from the renal function branch and the key vector sequence generated from the feature mapping from the respiratory physiology branch are used to perform dot product attention calculation to obtain a set of attention weights; Using the obtained attention weights, the value vector sequence generated by the feature mapping from the respiratory physiology branch is weighted and summed to generate a context vector that incorporates respiratory physiology information. The generated context vector is residually connected to the original renal functional branch feature representation and passed through a feedforward network to output the cross-modal interaction features.

[0012] As a further aspect of the present invention, the system further includes a model update module for incrementally updating the improved multimodal feature fusion prediction model, specifically including: In clinical practice, actual prognostic data of target patients after the application of the risk warning method are continuously collected. The actual prognostic data includes whether clinically diagnosed acute kidney injury occurred within the specified future time window. The synchronized multimodal time series dataset, the corresponding individualized probability risk value prediction results, and the actual prognosis data obtained afterward are used together to form a new training sample. New training samples are added to an incremental learning sample pool. When the number of samples in the incremental learning sample pool reaches the preset batch size, the model incremental update process is started. In the incremental update process of the model, the current parameters of the improved multimodal feature fusion prediction model are used as the initial weights, and all samples in the incremental learning sample pool are used as training data to perform a retraining with a small learning rate to update the network parameters of the model. The updated model network parameters are deployed as the new default prediction model for subsequent patient risk warnings.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: An improved multimodal feature fusion prediction model based on a deep neural network architecture is introduced, the core of which lies in handling heterogeneous modal features through a cross-attention mechanism. This mechanism allows the model to dynamically calculate the interaction weights between the time-domain / frequency-domain features of respiratory function and the trend features of kidney function during training and inference, rather than simply concatenating or processing them independently. For example, the model can autonomously learn how the spectral features of respiratory rate in a specific pattern are correlated with subtle upward trends in creatinine, and adjust the contribution of different features in the final prediction accordingly. This dynamic correlation modeling enables the quantitative capture of pathophysiological coupling relationships across organ systems, and can more sensitively identify early and complex multimodal feature combinations that predict the risk of kidney injury, thereby providing more accurate individualized probability risk values ​​before a specified time window.

[0014] The system incorporates an interpretability mechanism within the model, activating a risk attribution module when the risk exceeds a threshold. This mechanism, typically based on analysis of attention weights, feature gradients, or the importance of ensemble model features, traces and quantifies the contribution of each input feature (including respiratory fluctuations, segments of renal function change curves, and static background) to the final high-risk probability output. The system then identifies a subset of key features driving the increased risk; for example, it might pinpoint the combination of "increased blood oxygen variability in a specific frequency band over the past 24 hours" and "accelerated rate of increase in creatinine during the same period" as a key driving factor. This transforms the risk output from a black box into an understandable chain of clinical evidence. Combined with risk warning reports generated from a clinical decision rule base, these reports directly alert clinicians to these specific combinations of abnormal features, providing a clear basis for targeted respiratory support, volume management, or renal protection measures, thus enhancing the clinical guidance of the warnings. Attached Figure Description

[0015] Figure 1 The flowchart is as follows: Intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion, as described in this invention. Figure 2 A flowchart illustrating the workflow for multimodal data multi-timescale alignment and outlier filtering; Figure 3 A flowchart illustrating the workflow for identifying key feature subsets that drive risk escalation. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0018] See Figure 1 The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion of the present invention is implemented as follows: After the system is started, the feature extraction module works first. This module obtains multi-source heterogeneous clinical data from the target patient and performs a feature extraction process to extract time-domain and frequency-domain features reflecting fluctuations in respiratory function, as well as trend features reflecting subtle changes in kidney function. Subsequently, the fusion prediction module receives these feature data and static features from the patient's personal clinical background data. These heterogeneous features are used as input and transmitted to a pre-trained improved multimodal feature fusion prediction model. This model is built on a deep neural network architecture, and its core lies in dynamically fusing and deeply associating modal features from different sources such as respiratory and renal data through a built-in cross-attention mechanism. The model finally outputs a quantified value, namely the individualized probability risk value of the target patient developing acute kidney injury within a specified future time window. If the risk value exceeds a preset decision threshold, the risk attribution module is triggered. This module invokes the interpretability mechanism integrated within the improved multimodal feature fusion prediction model to perform a retrospective analysis of the model's decision-making process, identifying which input features played a key driving role in this high-risk prediction. These features are identified as a subset of key features. Finally, the report generation module is activated. It receives the subset of key features and accesses a preset clinical decision rule base storing medical knowledge and practical experience. It matches the identified key features with patterns in the rule base, automatically generating a structured risk warning report containing targeted clinical intervention suggestions, thus completing the entire intelligent processing from data to warning information.

[0019] In one embodiment of the present invention, the system acquires multi-source data from the target patient, including respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data. The respiratory physiological monitoring data specifically includes dynamic blood oxygen saturation time-series signals and exhaled breath molecular spectra, while the renal function biochemical data includes serum creatinine time-series changes and urine output monitoring data. The system performs multi-timescale alignment and outlier filtering on these data, first generating a synchronized multimodal time-series dataset. This processing involves multiple steps, as detailed in the following section. Figure 2 The system identifies high-frequency sampled data streams in respiratory physiological monitoring data and downsamples them to the same daily recording frequency as renal function biochemistry data using a sliding window averaging method. Simultaneously, it expands static information in personal clinical background data into a numerically constant sequence with the same time length as the renal function biochemistry data. The system establishes a unified timestamp index with a minimum time unit of one day. The downsampled respiratory physiological monitoring data, the original renal function biochemistry data, and the expanded personal clinical background data sequence are then concatenated and aligned according to this unified timestamp index. The system performs outlier detection on the concatenated and aligned data, using a statistically based local outlier method to identify outliers that deviate from the statistical distribution characteristics of adjacent data points. For identified outliers, the system replaces them with either the linear interpolation of their values ​​at the previous and next time points or the overall median of the data stream to which the outlier belongs, thus completing outlier filtering. The system encapsulates the data that has completed outlier filtering into a synchronized multimodal time series dataset with a unified time step and feature channel. It then extracts time-domain and frequency-domain features reflecting fluctuations in respiratory function, as well as trend features reflecting subtle changes in kidney function, from this dataset.

[0020] In its implementation, the intelligent early warning system for kidney injury risk in respiratory diseases, which integrates multimodal features, performs data preprocessing and alignment. Taking a target patient with chronic obstructive pulmonary disease as an example, the system acquires respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data from the target patient. The respiratory physiological monitoring data includes dynamic blood oxygen saturation time-series signals collected once per minute and exhaled breath molecular spectra collected twice daily. The renal function biochemical data includes serum creatinine time-series changes measured once every morning and hourly urine volume monitoring data. The personal clinical background data includes static information such as age and history of underlying diseases. In its implementation, the system performs multi-timescale alignment and outlier filtering on respiratory physiological monitoring data, renal function biochemical data, and individual clinical background data to generate a synchronized multimodal time series dataset. The processing begins by identifying high-frequency sampled data streams within the respiratory physiological monitoring data. For example, the dynamic blood oxygen saturation time-series signal collected every minute is downsampled using a sliding window averaging method and aligned with the daily recorded renal function biochemical data. The sliding window averaging method can be described as calculating the arithmetic mean of all data points within the window; its mathematical expression is: for high-frequency signals... On the target date downsampling value Calculated using the formula: in: Indicates the first The timestamp of the day It is the total number of data points within the window. It is the first one in the window The system generates a set of original high-frequency data points. In some embodiments, the system expands the static information in the individual clinical background data into a numerically constant sequence with the same time length as the renal function biochemical data. For example, the patient's age "65 years old" is expanded into a sequence containing seven identical elements "65" over a seven-day observation period. In a specific implementation, the system establishes a unified timestamp index, with the minimum time unit set to days. The index sequence consists of consecutive dates, such as "2023-10-01", "2023-10-02" to "2023-10-07". The downsampled respiratory physiological monitoring data, the original renal function biochemical data, and the expanded individual clinical background data sequence are spliced ​​and aligned according to the unified timestamp index. Data points that are not aligned are marked with specific values. In some embodiments, outlier detection is performed on the stitched and aligned data. A statistically based local outlier factor method is used to identify outliers that deviate from the statistical distribution characteristics of their neighboring data points. For example, the local reachability density ratio of each data point relative to its k nearest neighbors is calculated, and data points with a ratio significantly greater than 1 are identified as outliers. Optionally, the outlier is replaced with the linear interpolation of its value at the previous and next time steps. For example, if a daily urine output monitoring data point is identified as an outlier, and its value was 1500 ml the previous day and 1600 ml the next day, then 1550 ml is used to replace the outlier. Optionally, the outlier is replaced with the overall median of the data stream to which it belongs, based on respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data. For example, if the median blood oxygen saturation over the past seven days is 96%, then 96% is used to replace the marked outlier blood oxygen saturation value, thus completing outlier filtering. Understandably, the data after outlier filtering is encapsulated into a synchronized multimodal time series dataset with a unified time step and feature channels. This dataset is a three-dimensional array, where the dimensions represent the time step, the number of feature channels, and the number of samples. For example, an array of shape (7, 20, 1) represents data from 7 time steps, 20 feature channels, and 1 patient. From the synchronized multimodal time series dataset, time-domain and frequency-domain features reflecting fluctuations in respiratory function, as well as trend features reflecting subtle changes in renal function, are extracted. For example, the mean of the aligned daily blood oxygen saturation data is calculated, and the rate of change over three days for the serum creatinine data is calculated.

[0021] In one embodiment of the invention, the model is built on a deep neural network architecture and includes a shared initial feature embedding layer. This layer performs a preliminary nonlinear transformation on the original numerical sequences of each modality in the synchronized multimodal time-series dataset to obtain initial feature representations for each modality. The model includes a modality-specific deep feature encoder, which has a convolutional neural network branch for respiratory physiological monitoring data, a gated recurrent unit network branch for renal function biochemical data, and a multilayer perceptron branch for individual clinical background data. These branches are used to extract high-order abstract feature representations for each modality. The core component of the model is a cross-attention mechanism module, which calculates the dynamic correlation weights between the respiratory physiological modality feature representations and the renal function modality feature representations, and performs weighted fusion of the features of the two modalities based on the dynamic correlation weights to generate cross-modal interactive features. The specific working process of the cross-attention mechanism module includes linearly projecting the deep feature representations from the respiratory physiology monitoring data branch and the deep feature representations from the renal function biochemical data branch, respectively, to generate query vector sequences, key vector sequences, and value vector sequences. A set of attention weights is obtained by performing dot product attention calculation on the query vector sequence generated from the feature mapping of the renal function branch and the key vector sequence generated from the feature mapping of the respiratory physiology branch. The obtained attention weights are then used to weight and sum the value vector sequence generated from the feature mapping of the respiratory physiology branch to generate a context vector incorporating respiratory physiology information. The generated context vector is then residually concatenated with the original renal function branch feature representation and passed through a feedforward network to output cross-modal interaction features. The model also includes a multi-level feature aggregation module, which concatenates the cross-modal interaction features with the deep features of the individual's clinical background data and aggregates them through a feedforward neural network to generate a comprehensive patient state vector. Finally, a risk prediction head composed of multi-layer perceptrons receives this comprehensive patient state vector and outputs an individualized probability risk value through a stack of fully connected layers and nonlinear activation functions.

[0022] In practical implementation, the improved multimodal feature fusion prediction model receives a synchronized multimodal time series dataset from the feature extraction module. This dataset is a three-dimensional tensor with dimensions (time step, number of feature channels, 1). For example, the shape of the input data for a specific patient is (7, 20, 1), representing 7 time steps, 20 feature channels, and 1 patient sample. The shared initial feature embedding layer performs a preliminary nonlinear transformation on the original numerical sequence of each modality in the synchronized multimodal time series dataset. This embedding layer can be a fully connected layer with an input dimension of 20 and an output dimension of 32. It maps the 20-dimensional original features of each time step to a 32-dimensional initial feature representation, thereby obtaining the initial feature representation of each modality. In specific implementation, the modality-specific deep feature encoder contains three parallel branch networks. The convolutional neural network branch for respiratory physiological monitoring data receives the part corresponding to the respiratory modality from the initial feature representation, such as a sequence of shape (7, 10). The convolutional neural network branch can be designed to contain two one-dimensional convolutional layers. The first layer uses 16 convolutional kernels of size 3 and passes through the ReLU activation function, and the second layer uses 32 convolutional kernels of size 3 and passes through the ReLU activation function. Finally, it outputs a high-order abstract feature representation of the respiratory physiological monitoring data.

[0023] In some embodiments, the gated recurrent unit network (GRN) branch for renal functional biochemical data receives a portion of the initial feature representation corresponding to the renal modality, such as a sequence of shape (7, 8). The GRN branch can be designed as a single-layer GRN network with 16 hidden units, and its hidden state at the last time step is extracted as a high-order abstract feature representation of the renal functional biochemical data. The multilayer perceptron branch for personal clinical background data receives a portion of the initial feature representation corresponding to the background modality, such as a sequence of shape (7, 2). Since background information is constant over time, the multilayer perceptron branch typically takes the average value of the sequence over time as input. Its structure can be designed as a feedforward network with 8 hidden layers, ultimately outputting a high-order abstract feature representation of the personal clinical background data. The cross-attention mechanism module calculates the dynamic correlation weights between the respiratory physiological modality feature representation and the renal functional modality feature representation, and weights and fuses the features of the two modalities based on these dynamic correlation weights to generate cross-modal interactive features. In specific implementation, the cross-attention mechanism module works by performing linear projections on the deep feature representations from the respiratory physiology monitoring data branch and the deep feature representations from the renal function biochemical data branch, respectively, to generate query vector sequences, key vector sequences, and value vector sequences. For example, the 16-dimensional feature representation of the renal function branch is projected through a linear layer into an 8-dimensional query vector sequence, and the 32-dimensional feature representation of the respiratory physiology branch is projected through two independent linear layers into an 8-dimensional key vector sequence and a value vector sequence, respectively. In some embodiments, a set of attention weights is obtained by performing a dot product attention calculation using the query vector sequence generated from the feature mapping of the renal function branch and the key vector sequence generated from the feature mapping of the respiratory physiology branch. The calculation of the attention weights can be expressed as the formula: in: Represents the query vector matrix. Represents the key vector matrix, This refers to the dimension of the key vector. The obtained attention weights are used to weight and sum the value vector sequence generated from the feature mappings of the respiratory physiology branch, producing a context vector that incorporates respiratory physiology information. This generated context vector is then residually concatenated with the original renal function branch feature representation and passed through a feedforward network to output cross-modal interaction features. This feedforward network can contain a fully connected layer with 16 neurons. A multi-level feature aggregation module concatenates the cross-modal interaction features with deep features from the individual's clinical background data and aggregates them through a feedforward neural network. For example, a 16-dimensional cross-modal interaction feature is concatenated with an 8-dimensional deep feature from the individual's clinical background data to form a 24-dimensional vector. This concatenated vector passes through a first aggregation layer with 12 neurons and a second aggregation layer with 8 neurons, ultimately generating an 8-dimensional comprehensive patient state vector. It is understandable that the risk prediction head composed of a multilayer sensing mechanism receives a comprehensive patient state vector and outputs an individualized probability risk value through a stack of fully connected layers and nonlinear activation functions. The structure of the risk prediction head can be a network containing two fully connected layers. The first layer has 4 neurons and uses the ReLU activation function, and the second layer has 1 neuron and uses the Sigmoid activation function. Finally, it outputs a probability value between 0 and 1, which represents the individualized probability risk value of the target patient developing acute kidney injury within a specified future time window.

[0024] In one embodiment of the present invention, when the individualized probability risk value exceeds a preset decision threshold, refer to... Figure 3The risk attribution module identifies a subset of key features driving increased risk values ​​based on the interpretability mechanism within the improved multimodal feature fusion prediction model. This process is performed after the model completes forward propagation computation on the synchronized multimodal time-series dataset for a specific patient. The system obtains the activation value vector of the layer preceding the final fully connected layer of the risk prediction head. For each dimension of the neuron in this activation value vector, the gradient of each input element with respect to the integrated patient state vector is calculated. The absolute value of the calculated gradient is taken and averaged over all time steps of the synchronized multimodal time-series dataset to obtain the average importance score of each input feature for the final risk prediction decision. The system selects a specified number of features with the highest average importance scores from all features of the respiratory physiology modality, renal function modality, and personal clinical background modality to form a subset of key features. The report generation module generates a risk warning report based on a subset of key features combined with a pre-defined clinical decision rule base. The report generation process extracts features belonging to the respiratory physiology modality and features belonging to the renal function modality from the subset of key features. It then searches the pre-defined clinical decision rule base for definitions of respiratory system abnormalities that match the extracted respiratory physiology modality features and obtains a set of candidate intervention suggestions associated with that modality. Simultaneously, it searches the pre-defined clinical decision rule base for definitions of renal function abnormalities that match the extracted renal function modality features and obtains another set of candidate intervention suggestions associated with that modality. The system deduplicates and merges the two sets of candidate intervention suggestions and sorts them according to a pre-defined intervention priority. Finally, it integrates the sorted list of intervention suggestions with the individualized probability risk value, the specific feature names and values ​​from the key feature subset driving the risk increase, and other relevant information into a structured risk warning report.

[0025] In practical implementation, the risk attribution module initiates its workflow after the improved multimodal feature fusion prediction model outputs a high-risk probability value for an individual. For example, if the model calculates an individualized probability risk value of 0.85 for a patient, and the preset decision threshold is 0.7, since 0.85 is greater than 0.7, the system triggers the risk attribution process. After the improved multimodal feature fusion prediction model completes forward propagation calculations on the synchronized multimodal time-series dataset for a specific patient, the system obtains the activation value vector of the layer preceding the final fully connected layer of the risk prediction head. Assume this activation value vector is a 4-dimensional vector [0.2, -0.5, 1.1, 0.8]. In practical implementation, the gradient of each neuron in each dimension of the activation value vector with respect to each input element of the comprehensive patient state vector is calculated. The comprehensive patient state vector is an 8-dimensional vector. The system automatically differentiates and calculates the partial derivatives of each activation neuron with respect to these 8 input elements, obtaining a 4x8 gradient matrix. In some embodiments, the absolute value of the calculated gradient is taken and averaged over all time steps of the synchronized multimodal time series dataset to obtain the average importance score of each input feature for the final risk prediction decision. The calculation of the average importance score can be expressed as follows: for the ... Each input feature has an average importance score. Calculated by the formula: in: This is the total number of time steps (e.g., 7 days). It is the dimension of the activation value vector of the layer preceding the final fully connected layer of the risk prediction head (e.g., 4). Indicates the first At the next time step, the first The activation value neuron regarding the first activation value neuron The gradient value of each input feature. The system selects a specified number of features with the highest average importance score from all features of the respiratory physiology modality, renal function modality, and personal clinical background modality to form a key feature subset. For example, if the system selects the top 5 features by importance, and assumes that the importance scores of 20 features are calculated and ranked (see Table 1), then the key feature subset includes "minimum blood oxygen saturation", "three-day change rate of serum creatinine", "age", "duration of hypoxemia", and "urine volume per unit body weight".

[0026] Table 1: Feature Importance Score Table The report generation module generates risk warning reports containing targeted clinical intervention suggestions based on a subset of key features and a pre-defined clinical decision rule base. In specific implementation, the report generation module extracts features belonging to the respiratory physiological modality and features belonging to the renal function modality from the subset of key features. For example, it extracts "minimum blood oxygen saturation" and "duration of low blood oxygen" as key features of the respiratory physiological modality, and "three-day change rate of serum creatinine" and "urine volume per unit body weight" as key features of the renal function modality. In some embodiments, a pre-defined clinical decision rule base is used to search for respiratory system abnormality pattern definitions that match the extracted features belonging to the respiratory physiological modality. A set of candidate intervention suggestions associated with the respiratory system abnormality pattern is then obtained from the rule base. For example, the rule base defines "minimum blood oxygen saturation below 92% for three consecutive days" as matching the "persistent hypoxemia" pattern. The candidate intervention suggestions associated with this pattern include "assess oxygen therapy needs" and "re-examine arterial blood gas analysis". Simultaneously, a pre-defined clinical decision rule base is used to search for renal function abnormality pattern definitions that match the extracted features belonging to the renal function modality. Another set of candidate intervention suggestions associated with the renal function abnormality pattern is then obtained from the rule base. For example, the rule base defines "three-day change rate of serum creatinine greater than 25%" as matching the "risk of acute kidney injury" pattern. The candidate intervention suggestions associated with this pattern include "assess volume status" and "avoid nephrotoxic drugs". It is understandable that the two sets of candidate intervention recommendations are deduplicated and merged, and then sorted according to the preset priority of intervention measures. For example, "assess oxygen therapy needs" has a priority of 1, "assess volume status" has a priority of 2, "avoid nephrotoxic drugs" has a priority of 3, and "re-examine arterial blood gas analysis" has a priority of 4. After merging and deduplication, a sorted list is generated ["assess oxygen therapy needs", "assess volume status", "avoid nephrotoxic drugs", "re-examine arterial blood gas analysis"]. It is also understandable that the sorted list of intervention recommendations, along with the individualized probability risk value and the specific feature names and values ​​in the key feature subset that drive the increase of the individualized probability risk value, are integrated into a structured risk warning report. The report is presented in structured text or table form, including patient identification, calculated risk value of 0.85, decision threshold of 0.7, list of key features and their specific values, and the sorted list of intervention recommendations.

[0027] In one embodiment of the present invention, time-domain and frequency-domain features reflecting fluctuations in respiratory function are extracted from the synchronized multimodal time-series dataset. This process calculates the mean, minimum, maximum, and standard deviation of the dynamic blood oxygen saturation time-series signal within each time window to obtain a set of time-domain statistical features. Simultaneously, a fast Fourier transform is applied to the dynamic blood oxygen saturation time-series signal to calculate its power spectral density within a specific physiological frequency band to obtain frequency-domain energy features. Trend features reflecting subtle changes in renal function are extracted from the synchronized multimodal time-series dataset. This process calculates the cumulative change value, rate of change, and relative change fold relative to the individual baseline level for serum creatinine time-series change data within a preset review time window. For urine volume monitoring data, the daily total urine volume and urine volume per unit body weight are calculated, and the number of hours with urine volume below a specific threshold is identified. The cross-correlation coefficient between the rate of change in serum creatinine and the change in urine volume is calculated. The cumulative change value, rate of change, relative change fold of serum creatinine, daily total urine volume, urine volume per unit body weight, duration of low urine volume, and cross-correlation coefficient are combined to constitute trend features reflecting subtle changes in renal function.

[0028] In practice, the feature extraction module performs specific feature calculations from the synchronized multimodal time series dataset to obtain the input required by the model. The synchronized multimodal time series dataset has undergone multi-timescale alignment and outlier filtering. The dataset contains seven consecutive days of aligned data from a hospitalized patient with community-acquired pneumonia. The feature extraction process first targets features reflecting fluctuations in respiratory function. In practice, this involves extracting time-domain and frequency-domain features reflecting fluctuations in respiratory function from the synchronized multimodal time series dataset, including processing the dynamic blood oxygen saturation time-series signal. The dynamic blood oxygen saturation time series signal is a sequence of one data point per day that has been downsampled and aligned. For example, the blood oxygen saturation values ​​of a patient over seven days are [95, 94, 92, 93, 91, 94, 95]. The system calculates the mean, minimum, maximum, and standard deviation of the dynamic blood oxygen saturation time series signal within each time window to obtain a set of time-domain statistical characteristics. The time window can be set as a three-day sliding window. For the third day, the window covers the data from the second, third, and fourth days [94, 92, 93]. The calculated mean is 93.0, the minimum is 92, the maximum is 94, and the standard deviation is approximately 1.0. In some embodiments, the power spectral density within a specific physiological frequency band is calculated using Fast Fourier Transform (FFT) on the dynamic blood oxygen saturation time-series signal to obtain frequency domain energy characteristics. The system treats the seven-day blood oxygen saturation sequence as a discrete-time signal and transforms it to the frequency domain using FFT. Assuming a sampling frequency of once per day and a total signal length of 7, the transformation yields complex representations of each frequency component. The specific physiological frequency band can be defined as a fluctuation band corresponding to a period between 3 and 1 day. The system calculates the sum of the power spectral densities corresponding to all frequency points within this frequency band as the frequency domain energy characteristics. The calculation can be expressed as a formula: in: It is the frequency of the signal after it undergoes a Fast Fourier Transform. The complex range at the point, This represents the set of all frequency points contained within a defined specific physiological frequency band. Optionally, time-domain statistical features and frequency-domain energy features are combined into a feature vector. For example, for each day, a vector containing five feature values—mean, minimum, maximum, standard deviation, and frequency band energy—can be generated.

[0029] In practical implementation, extracting trend features reflecting subtle changes in renal function from the synchronized multimodal time-series dataset includes processing serum creatinine time-series change data. This data is a daily recorded sequence; for example, a patient's serum creatinine values ​​over seven days are [0.8, 0.9, 1.1, 1.4, 1.6, 1.3, 1.2]. The system calculates the cumulative change, rate of change, and relative change fold relative to the individual baseline level within a preset review time window. The preset review time window can be set to three days. For the fifth day, the window covers data from the third, fourth, and fifth days [1.1, 1.4, 1.6]. The cumulative change is 1.6 - 1.1 = 0.5 mg / dL. The rate of change can be obtained by calculating the slope of a linear regression. The relative change fold needs to be referenced to the individual baseline level, which can be defined as the initial measurement upon admission of 0.8 mg / dL. Therefore, the relative change fold on the fifth day is 1.6 / 0.8 = 2.0. In some embodiments, the system calculates the total daily urine volume and urine volume per unit body weight from urine volume monitoring data and identifies the number of consecutive hours with urine volume below a specific threshold. The urine volume monitoring data is a sequence recorded hourly. The total daily urine volume is obtained by summing the urine volumes of all hours on that day, and the urine volume per unit body weight is obtained by dividing the total daily urine volume by the patient's body weight. The low urine volume threshold can be set to 0.5 ml / kg body weight per hour. The system identifies hourly sequences with urine volume below this threshold and calculates the number of consecutive hours they last. It is understood that the cross-correlation coefficient between the rate of change in serum creatinine and the change in urine volume is calculated; for example, the Pearson correlation coefficient is calculated between the sequence of daily changes in serum creatinine over the past three days and the sequence of daily changes in total urine volume during the same period to quantify the degree of association between the two changes.

[0030] As an embodiment of the present invention, to enhance the early warning capability for the risk of kidney injury, the renal function biochemical data may further include time-series monitoring data of cystatin C. When extracting trend features reflecting subtle changes in renal function, in addition to the aforementioned calculations for serum creatinine and urine output, the system performs the same trend analysis on the time-series cystatin C data in parallel. Specifically, the system calculates the cumulative change value, rate of change, and relative change fold relative to the individual baseline level for the time-series cystatin C change data within the same preset review time window. The calculation formula is similar to that for serum creatinine, and is as follows: Cumulative change value: Rate of change: Relative change factor: in, This indicates the current concentration of cystatin C. express Concentration a certain time unit ago, To review the window length, This represents an individual's baseline level. Therefore, the set of trend features reflecting subtle changes in renal function will be expanded to include a comprehensive set encompassing serum creatinine-derived features, cystatin C-derived features, urine volume-derived features, and cross-correlation features between them. The improved multimodal feature fusion prediction model will simultaneously process these expanded multidimensional renal function features, dynamically modeling their association with respiratory physiological features through their cross-attention mechanism. Correspondingly, the clinical decision rule base should pre-set rules and associated intervention recommendations matching abnormal cystatin C change patterns. When a cystatin C-related feature is identified by the risk attribution module as a key feature driving increased risk, the report generation module will generate a more forward-looking risk warning report containing indications of abnormal cystatin C.

[0031] Optionally, the cumulative change value, rate of change, relative change fold, total daily urine volume, urine volume per unit body weight, duration of low urine volume, and cross-correlation coefficient of serum creatinine obtained above can be combined to form a trend feature reflecting subtle changes in renal function. These calculations can form an example of feature extraction, see Table 2.

[0032] Table 2: Calculation Table of Trend Characteristics of Kidney Function In one embodiment of the present invention, the system continuously collects actual prognostic data of target patients after the application of the risk warning method in clinical practice. The actual prognostic data includes whether the target patient developed clinically diagnosed acute kidney injury within a specified future time window. The synchronized multimodal time series dataset used for prediction, the corresponding individualized probability risk value prediction results, and the actual prognostic data obtained afterward are combined to form a new training sample. The system adds the new training sample to an incremental learning sample pool. When the number of samples in the incremental learning sample pool reaches a preset batch size, the model incremental update process is initiated. In the model incremental update process, the system uses the current parameters of the improved multimodal feature fusion prediction model as initial weights and performs a small-learning-rate retraining to update the model's network parameters using all samples in the incremental learning sample pool as training data. The system deploys the updated model network parameters as a new default prediction model for subsequent patient risk warnings.

[0033] In practical implementation, the model update module continuously collects actual prognostic data of target patients after applying the risk warning method in clinical practice. Actual prognostic data includes whether the target patient developed clinically diagnosed acute kidney injury within a specified future time window. For example, the system performs a risk assessment on a patient hospitalized for pneumonia, predicting an individualized probability risk value of 0.72 for the next three days based on a synchronized multimodal time series dataset from seven days after admission. If, during the subsequent three-day observation period, the clinician diagnoses the patient with acute kidney injury according to the KDIGO criteria, the actual prognostic data is recorded as "yes". In practical implementation, the synchronized multimodal time series dataset used for prediction, the corresponding individualized probability risk value prediction, and the subsequently obtained actual prognostic data are combined to form a new training sample. This new training sample contains an input tensor of shape (7, 20, 1) (representing data from 20 feature channels over seven days), a corresponding predicted probability value of 0.72, and a binary label 1 (indicating the occurrence of acute kidney injury). The system adds new training samples to an incremental learning sample pool, which is a cache queue located in server memory or a database for temporarily storing newly generated training samples. In some embodiments, the model incremental update process is initiated when the number of samples in the incremental learning sample pool reaches a preset batch size. The preset batch size can be set to 100, meaning that the system automatically triggers the model update process when the 100th new training sample is added to the pool. In the model incremental update process, the system uses the current parameters of the improved multimodal feature fusion prediction model as initial weights and performs a retraining round with a small learning rate using all samples in the incremental learning sample pool as training data. The current parameters of the improved multimodal feature fusion prediction model refer to all network weights of the model currently deployed online. The value of the small learning rate can be set to one-tenth of the initial training learning rate, for example, 0.0001. It can be understood that the standard binary classification cross-entropy loss function is used during the retraining process. The calculation formula is: in: It is the total number of samples in the incremental learning sample pool. It is the first The true label (0 or 1) of each sample. It is an improved multimodal feature fusion prediction model for the first The individualized probability risk value is calculated for each sample. This loss function is minimized using the backpropagation algorithm to update the network parameters of the improved multimodal feature fusion prediction model. In practice, the updated model network parameters are deployed as the new default prediction model for subsequent patient risk warnings. After retraining, the system replaces the old weight file in the online service with the updated model weight file. Subsequent patient data accessed by the system will be used for risk prediction by this updated model, thus completing a full incremental update cycle.

[0034] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A multimodal feature fusion-based intelligent early warning system for kidney injury risk in respiratory diseases, characterized in that, include: The feature extraction module is used to extract time-domain and frequency-domain features that reflect fluctuations in respiratory function, as well as trend features that reflect subtle changes in renal function. The fusion prediction module inputs the time-domain and frequency-domain features, trend features, and static features from the individual's clinical background data into an improved multimodal feature fusion prediction model. The improved multimodal feature fusion prediction model is based on a deep neural network architecture and uses a cross-attention mechanism to dynamically fuse and model the correlation of heterogeneous modal features. The improved multimodal feature fusion prediction model outputs an individualized probability risk value for the target patient to develop acute kidney injury within a specified future time window. The risk attribution module is used to identify the key feature subset that drives the increase in the individualized probability risk value when the individualized probability risk value exceeds a preset decision threshold, based on the interpretability mechanism inside the improved multimodal feature fusion prediction model. The report generation module generates a risk warning report containing targeted clinical intervention suggestions based on the aforementioned key feature subset and a preset clinical decision rule base.

2. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 1, characterized in that, The extraction of time-domain and frequency-domain features reflecting fluctuations in respiratory function, as well as trend features reflecting subtle changes in renal function, includes: Acquire respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data from target patients. The respiratory physiological monitoring data includes dynamic blood oxygen saturation time-series signals and exhaled gas molecular profiles. The renal function biochemical data includes serum creatinine time-series changes and urine output monitoring data. The respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data are subjected to multi-timescale alignment and outlier filtering to generate a synchronized multimodal time series dataset. From the synchronized multimodal time series dataset, time-domain and frequency-domain features reflecting fluctuations in respiratory function, as well as trend features reflecting subtle changes in kidney function, are extracted respectively.

3. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 2, characterized in that, The respiratory physiological monitoring data, renal function biochemical data, and personal clinical background data are subjected to multi-timescale alignment and outlier filtering to generate a synchronized multimodal time series dataset, including: High-frequency sampling data streams in the respiratory physiological monitoring data are identified, and the high-frequency sampling data streams are downsampled to the same daily recording frequency as the renal function biochemical data using a sliding window averaging method. The static information in the personal clinical background data is expanded into a sequence with constant values ​​and the same time length as the renal function biochemical data; Establish a unified timestamp index, with the minimum time unit of the timestamp index set to days; The downsampled respiratory physiological monitoring data, the original renal function biochemical data, and the expanded personal clinical background data sequence are spliced ​​and aligned according to the unified timestamp index. Outlier detection is performed on the spliced ​​and aligned data. The statistical local outlier factor method is used to identify outlier data points that deviate from the statistical distribution characteristics of adjacent data points. The abnormal value point is replaced by the linear interpolation value of the value of the abnormal value point at the previous time and the value at the next time, or by the overall median of the data stream to which the abnormal value point belongs in the respiratory physiological monitoring data, the renal function biochemical data and the personal clinical background data, to complete the outlier filtering. The data after outlier filtering is encapsulated into a synchronized multimodal time series dataset with a unified time step and feature channel.

4. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 1, characterized in that, The improved multimodal feature fusion prediction model is based on a deep neural network architecture and includes: A shared initial feature embedding layer is used to perform a preliminary nonlinear transformation on the original numerical sequence of each modality in the synchronized multimodal time series dataset to obtain the initial feature representation of each modality. A modal deep feature encoder, comprising a convolutional neural network branch for respiratory physiological monitoring data, a gated recurrent unit network branch for renal function biochemical data, and a multilayer perceptron branch for individual clinical background data, is used to extract high-order abstract feature representations of each modality of data respectively. The cross-attention mechanism module is used to calculate the dynamic correlation weight between the respiratory physiological modality feature representation and the renal function modality feature representation, and to perform weighted fusion of the features of the two modalities based on the dynamic correlation weight to generate cross-modal interactive features. A multi-level feature aggregation module is used to concatenate the cross-modal interaction features with the deep features of the individual clinical background data and aggregate them through a feedforward neural network to generate a comprehensive patient state vector. A risk prediction head composed of a multilayer sensing mechanism receives the comprehensive patient state vector and outputs the individualized probability risk value through a stack of fully connected layers and nonlinear activation functions.

5. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 4, characterized in that, Based on the interpretability mechanism within the improved multimodal feature fusion prediction model, a subset of key features driving the increase in the individualized probability risk value is identified, including: After the improved multimodal feature fusion prediction model completes the forward propagation calculation on the synchronized multimodal time series dataset of a specific patient, the activation value vector of the layer before the final fully connected layer of the risk prediction head is obtained. For each neuron in the activation value vector, calculate its gradient with respect to each input element of the integrated patient state vector; The absolute value of the calculated gradient is taken, and the absolute value is averaged over all time steps of the synchronized multimodal time series dataset to obtain the average importance score of each input feature to the final risk prediction decision. From all features of the respiratory physiology modality, renal function modality, and personal clinical background modality, a specified number of features with the highest average importance score are selected to form the key feature subset.

6. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 1, characterized in that, Based on the aforementioned subset of key features, and in conjunction with a pre-defined clinical decision rule base, a risk warning report containing targeted clinical intervention suggestions is generated, including: From the aforementioned key feature subset, features belonging to the respiratory physiological modality and features belonging to the renal function modality are extracted respectively; From the preset clinical decision rule base, search for respiratory system abnormality pattern definitions that match the extracted features belonging to respiratory physiological modalities, and obtain a set of candidate intervention suggestions associated with the respiratory system abnormality pattern from the rule base; From the preset clinical decision rule base, search for the definition of abnormal renal function patterns that match the extracted features belonging to the renal function modality, and obtain another set of candidate intervention suggestions associated with the abnormal renal function patterns from the rule base; The two sets of candidate intervention suggestions were deduplicated and merged, and then sorted according to the preset priority of intervention measures. The sorted list of intervention recommendations, along with the individualized probability risk value, the specific feature names and their values ​​in the subset of key features that drive the increase of the individualized probability risk value, are integrated into a structured risk warning report.

7. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 2, characterized in that, From the synchronized multimodal time series dataset, time-domain and frequency-domain features reflecting fluctuations in respiratory function are extracted, including: For the dynamic blood oxygen saturation time series signal, calculate its mean, minimum, maximum and standard deviation in each time window to obtain a set of time domain statistical characteristics; The power spectral density within a specific physiological frequency band is calculated by applying Fast Fourier Transform to the dynamic blood oxygen saturation time-series signal, thus obtaining its frequency domain energy characteristics.

8. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 2, characterized in that, From the synchronized multimodal time series dataset, trend features reflecting subtle changes in kidney function are extracted, including: For the time-series changes in serum creatinine, calculate the cumulative change value, the rate of change, and the relative change multiple relative to the individual baseline level within a preset review time window; For the urine monitoring data, calculate the total daily urine volume and urine volume per unit body weight, and identify the number of consecutive hours when the urine volume is below a certain threshold; Calculate the cross-correlation coefficient between the rate of change in serum creatinine and the change in urine volume; The cumulative change in serum creatinine, the rate of change, the relative change fold, the total daily urine volume, the urine volume per unit body weight, the duration of low urine volume, and the cross-correlation coefficient are combined to form the trend characteristics reflecting subtle changes in renal function.

9. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 4, characterized in that, The specific working process of the cross-attention mechanism module includes: Linear projection is performed on the deep feature representations from the respiratory physiological monitoring data branch and the deep feature representations from the renal function biochemical data branch to generate query vector sequences, key vector sequences and value vector sequences. The query vector sequence generated from the feature mapping from the renal function branch and the key vector sequence generated from the feature mapping from the respiratory physiology branch are used to perform dot product attention calculation to obtain a set of attention weights; Using the obtained attention weights, the value vector sequence generated by the feature mapping from the respiratory physiology branch is weighted and summed to generate a context vector that incorporates respiratory physiology information; The generated context vector is residually connected to the original renal functional branch feature representation and passed through a feedforward network to output the cross-modal interaction features.

10. The intelligent early warning system for kidney injury risk in respiratory diseases based on multimodal feature fusion as described in claim 1, characterized in that, The system also includes a model update module for incrementally updating the improved multimodal feature fusion prediction model, specifically including: In clinical practice, actual prognostic data of target patients after the application of the risk warning method are continuously collected. The actual prognostic data includes whether clinically diagnosed acute kidney injury occurred within the specified future time window. The synchronized multimodal time series dataset, the corresponding individualized probability risk value prediction results, and the actual prognosis data obtained afterward are used together to form a new training sample. New training samples are added to an incremental learning sample pool. When the number of samples in the incremental learning sample pool reaches the preset batch size, the model incremental update process is started. In the incremental update process of the model, the current parameters of the improved multimodal feature fusion prediction model are used as the initial weights, and all samples in the incremental learning sample pool are used as training data to perform a retraining with a small learning rate to update the network parameters of the model. The updated model network parameters are deployed as the new default prediction model for subsequent patient risk warnings.

Images