A respiratory early warning method based on multi-source data

Abstract

This invention relates to the field of medical artificial intelligence technology, and particularly to a respiratory disease early warning method based on multi-source data. The method includes: collecting multi-source data on patients' respiratory diseases and preprocessing it to obtain single-source data feature vectors, and constructing a single-source data matrix; weighting and concatenating the single-source data matrices to obtain a fused feature matrix; based on the fused feature matrix, constructing a trend capture branch and a mutation detection branch to obtain a trend feature vector and a residual mutation value vector, quantifying the consistency between trends and mutations, and constructing a consistency score vector; based on the consistency score vector, trend feature vector, and residual mutation value vector, constructing a joint feature vector, and outputting the respiratory disease early warning level. This method solves the problems of traditional respiratory disease early warning methods, such as lack of effective modeling of the correlation between data timeliness and disease condition, insufficient memory and response capabilities for high-risk time points, and lack of a mechanism for judging the consistency between trends and mutations.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a respiratory early warning method based on multi-source data. Background Technology

[0002] In the clinical monitoring of respiratory diseases, patients are usually accompanied by complex, multimodal continuous monitoring data, such as high-frequency physiological parameters, laboratory test indicators, medical images and clinical texts. These multi-source heterogeneous data not only have significant differences in sampling frequency and expression mode, but also have different sensitivities and timeliness in reflecting the evolution of the disease. Therefore, how to fully extract key risk information from multi-source data and achieve highly reliable time-series early warning modeling has become a research hotspot in the current field of intelligent medicine. However, existing technologies still have significant shortcomings in multi-source data modeling, time-series dependency modeling and key risk identification: (1) Most studies simply splice multimodal data or use static attention mechanisms to assign weights to different data sources, which cannot effectively adapt to the changes in the clinical value of different data sources decaying over time in medical scenarios, and it is also difficult to consider the heterogeneous correlation of different data sources in the disease evolution process. This static processing method is prone to causing feature interference fusion effect with weak timeliness but large fluctuations, affecting the perception and extraction of key features of the disease; (2) In terms of time-series modeling, although bidirectional long short-term memory networks (Bi-L) STM is widely used in medical time series modeling tasks, but its standard structure, when processing continuous long sequences, defaults to equal weighting of data at all time steps, ignoring the importance of some time points (such as critical moments before disease mutation) for early warning judgment, resulting in insufficient memory of critical high-risk stages; (3) In terms of mutation risk identification, most existing models only focus on trend changes, ignoring the clinical pattern of "trend-mutation consistency" in disease mutation events. Models lacking consistency judgment mechanisms often have problems such as false alarms of mutation features and unstable early warning, which seriously affect the clinical applicability of the model. Therefore, in order to address the above shortcomings, it is urgent to provide a respiratory early warning method based on multi-source data. Summary of the Invention

[0003] This invention provides a respiratory early warning method based on multi-source data to solve the technical problems of traditional respiratory early warning methods, such as lack of effective modeling of the correlation between data timeliness and disease condition, insufficient memory and response capabilities for high-risk time points, and lack of a mechanism for judging the consistency of trends and mutations.

[0004] The present invention provides a respiratory early warning method based on multi-source data, comprising the following steps: S1. Collect multi-source data on patients' respiratory diseases, preprocess them to obtain single-source data feature vectors, and construct single-source data matrices; introduce a two-factor weighting mechanism based on time decay and disease similarity to calculate the weight coefficients of the data sources, and weight the single-source data matrices based on the weight coefficients of the data sources to obtain weighted single-source data matrices; use column-direction feature concatenation to concatenate all weighted single-source data matrices to obtain a fusion feature matrix; S2. Based on the fusion feature matrix, a bidirectional long short-term memory network is used with gating enhancement to construct a trend capture branch and output a trend feature vector; based on the fusion feature matrix, a mutation detection branch is constructed to calculate the residual mutation value and construct a residual mutation value vector; based on the trend feature vector and the residual mutation value vector, a trend-mutation consistency scoring mechanism is introduced to quantify the consistency between trends and mutations and construct a consistency scoring vector; based on the consistency scoring vector, trend feature vector, and residual mutation value vector, a joint feature vector is constructed; based on the joint feature vector, the respiratory disease warning level is output.

[0005] Preferably, S1 specifically includes: In the preprocessing process, time alignment is performed on the multi-source data to obtain time-aligned multi-source data; the time-aligned multi-source data is then standardized according to different data sources to obtain single-source data feature vectors, and a single-source data matrix is ​​constructed.

[0006] Preferably, S1 specifically includes: In the implementation of the two-factor weighting mechanism based on time decay and disease similarity, based on the feature vector of single-source data, the initial weight coefficients of different data sources are introduced, and the weight coefficients of different data sources are dynamically calculated by combining the disease similarity factor and the time decay factor.

[0007] Preferably, S1 specifically includes: By calculating the time difference between the current time point and the most recent data collection time from the data source, and applying an exponential decay function to the time difference, a time decay factor is generated; historical severe case data is introduced, and the cosine similarity between the feature vector of the single-source data and the feature mean vector of the historical severe case data is calculated to obtain the disease similarity factor.

[0008] Preferably, S2 specifically includes: By constructing a trend-capturing branch, the continuous evolution trend of the fused feature vector in the time dimension is captured, and the trend feature vector is output. In the process of constructing the trend-capturing branch, a risk label-guided gating enhancement mechanism is introduced based on the bidirectional long short-term memory network model. By introducing risk labels, high-risk time points are marked, and the forget gate output is enhanced to obtain the gating-enhanced forget gate output. Based on the gating-enhanced forget gate output, the hidden state at the current time is updated. The hidden state at the current time is used as the trend feature vector. The fused feature vector is taken from the fused feature matrix.

[0009] Preferably, S2 specifically includes: In the process of constructing the mutation detection branch, a historical sliding window is introduced. By calculating the residual difference between the standardized value of the feature in the fused feature vector and the feature mean of the historical sliding window, the residual mutation value is obtained, and a residual mutation value vector is constructed.

[0010] Preferably, S2 specifically includes: In the implementation of the trend-mutation consistency scoring mechanism, a trend-mutation consistency scoring scaling factor is introduced. By combining the trend change amount and the residual mutation value, the consistency between the trend change direction and the mutation direction is judged, a trend-mutation consistency score is generated, and a consistency score vector is constructed. The trend change amount is calculated based on the trend feature components.

[0011] Preferably, S2 specifically includes: The consistency score vector is multiplied element-wise by the trend feature vector and the residual mutation value vector, and then concatenated to obtain the joint feature vector. The joint feature vector is then input into the fully connected layer and the classification layer in sequence to output the probability distribution of respiratory disease risk level. The category corresponding to the highest probability is taken as the respiratory disease warning level.

[0012] The beneficial effects of the technical solution of the present invention are: 1. This invention proposes a two-factor weighting mechanism based on time decay and disease similarity. On the basis of the initial weight coefficient, the weight coefficients of different data sources at different time points are dynamically calculated, which can enhance the perception of time changes of data sources, strengthen key disease features, weaken the interference of outdated information and key disease pseudo-correlation features, thereby improving the sensitivity and robustness of multi-source data feature fusion for patients' respiratory diseases.

[0013] 2. This invention introduces a risk-label-guided gating enhancement mechanism into the existing bidirectional long short-term memory network model. Without explicitly increasing the structural complexity of the Bi-LSTM model, it can dynamically improve the Bi-LSTM model's ability to remember data at high-risk time points, solve the problem of insufficient memory of key disease stages in traditional Bi-LSTM models, and improve early warning performance.

[0014] 3. This invention introduces a trend-mutation consistency scoring mechanism, which can effectively identify high-risk features such as "the trend is deteriorating and accompanied by mutations in the same direction", while suppressing false mutation features caused by stable trends or inconsistent mutation directions, thereby improving the ability to identify the truly high-risk moment of the disease and reducing the mutation false alarm rate. Attached Figure Description

[0015] Figure 1 This is a flowchart of a respiratory early warning method based on multi-source data as described in this invention. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] The following description, in conjunction with the accompanying drawings, details a specific scheme for a respiratory early warning method based on multi-source data provided by the present invention.

[0019] See attached document Figure 1 The diagram illustrates a flowchart of a respiratory early warning method based on multi-source data according to an embodiment of the present invention. The method includes the following steps: S1. Collect multi-source data on patients' respiratory diseases, preprocess them to obtain single-source data feature vectors, and construct single-source data matrices; introduce a two-factor weighting mechanism based on time decay and disease similarity to calculate the weight coefficients of the data sources, and weight the single-source data matrices based on the weight coefficients of the data sources to obtain weighted single-source data matrices; use column-direction feature concatenation to concatenate all weighted single-source data matrices to obtain a fusion feature matrix; Multi-source data on patients' respiratory diseases were collected through the hospital's HIS system, including physiological parameters such as heart rate, blood oxygen saturation, respiratory rate, and body temperature; laboratory test data such as white blood cell count, C-reactive protein, and procalcitonin; imaging data such as chest CT scans; and clinical text data such as ward round records and nursing records. Further, the multi-source data was preprocessed to obtain single-source data feature vectors, and a single-source data matrix was constructed. Specifically, firstly, the multi-source data underwent time alignment processing, a fixed time step was set according to the multi-source data sampling frequency, and a time series window was constructed. T represents the number of time steps contained in the time series window, and k∈[1,T] represents the k-th time step. As a specific example, a fixed time step of 5 minutes and an observation window length of 1 hour can be set, then the time series window contains 12 time steps; for time points Based on time series window The physiological parameter data were time-aligned using a forward sliding window mean method with a length of 3 time steps. Laboratory test data and imaging data were time-aligned using a forward filling method. Clinical text data were time-aligned based on recorded timestamps, resulting in time-aligned multi-source data. Then, the time-aligned multi-source data were standardized. For the physiological parameter data, a forward sliding window mean method with a length of 3 time steps was used for filtering to reduce noise, and then the data was scaled to the [0,1] interval using a min-max normalization method to obtain normalized physiological parameter data. Data from the same time point were then... The normalized physiological parameter data are horizontally stitched together to form a single-source data feature vector of physiological parameter data. For laboratory test data, outliers were removed using the 3σ criterion, and then the Z-score standardization method was used to transform it into a standard normal distribution, resulting in standardized laboratory test data. Data from the same time point were then analyzed. The standardized laboratory test data are horizontally stitched together to form a single-source data feature vector of laboratory test data. For the imaging data, pixels with CT values ​​below a preset CT threshold (e.g., -400 HU) in chest CT images are first marked as candidate lung regions. Based on these candidate lung regions, morphological opening operations are performed using circular structuring elements with a radius of 5 pixels to remove noise, and 8-connected domains are extracted. The two connected domains with the largest areas are retained as the left and right lung regions, resulting in a lung region mask. The CT values ​​of the chest CT images are linearly normalized to the range [0, 255] using window width and window level mapping (e.g., window level -600 HU, window width 1500 HU), resulting in a normalized CT image. The normalized CT image is then input into a U-Net segmentation network for lesion segmentation, outputting a lesion probability map (pixel value range 0~1). The U-Net segmentation network uses a binary classification cross-entropy loss function and is trained on chest CT image data with pixel-level annotations, preferably including 4 The system consists of a 1-layer encoder and a 4-layer decoder. The encoder layers have 64, 128, 256, and 512 channels respectively, with a bottleneck layer having 1024 channels. The decoder layers have 512, 256, 128, and 64 channels respectively. Each encoder and decoder layer includes two 3×3 convolutions and a ReLU activation function. The encoder achieves downsampling through 2×2 max pooling, while the decoder achieves upsampling through 2×2 upsampling or deconvolution. Skip connections are used to fuse shallow features at the corresponding scale. The lesion probability map is binarized, and pixels with a probability value greater than a preset lesion probability threshold are marked as lesion regions. The lesion probability threshold is preferably 0.5. Based on the lung region mask, the number of lesion pixels is counted only within the lung region, and the ratio of lesion region pixels to the total number of lung region pixels and the average gray value of the lesion region are calculated. The results are then displayed at the same time point. The ratio of the number of pixels in the lesion area to the total number of pixels in the lung area and the average gray value of the lesion area are horizontally stitched together to form a single-source data feature vector of the imaging data. For clinical text data, the first step is text cleaning, including removing invalid symbols such as punctuation marks and using a general stop word list, such as the Harbin Institute of Technology stop word list, to remove stop words. Then, the maximum matching segmentation method based on the Chinese translation of the Medical Subject Headings (MeSH) is used for word segmentation. Finally, based on the segmentation results, the TF-IDF method is used to calculate keyword weights, and the top-ranked keywords are selected. By horizontally concatenating the keywords, we can obtain the single-source data feature vector of the clinical text data. Based on the feature vectors of the above-mentioned single-source data, a single-source data matrix is ​​constructed. Where i∈{1,2,3,4} represents the i-th type of data source, This represents the feature dimension of the i-th type of data source. Let T represent the set of real numbers, and let T represent the number of time steps contained in the time series window; store the feature vectors of multi-source and single-source data of patients' respiratory diseases in the respiratory department database; To avoid the problem of traditional attention mechanisms neglecting time sensitivity and disease correlation in multi-source data fusion, a time decay factor and a disease similarity factor are introduced. A two-factor weighting mechanism based on time decay and disease similarity is designed. Based on the initial weight coefficients of the data sources, the weighting factor at time point i for the i-th data source is dynamically calculated. The weighting coefficients, where the time decay factor is corrected for time points through an exponential function. The disease similarity factor is calculated based on the time difference between data collection and time points. The cosine similarity between the feature vector of the single-source data and the mean feature vector of historical severe cases is determined. These historical severe cases are patient samples from the existing respiratory database that meet preset severity criteria, including but not limited to those requiring transfer to the intensive care unit, experiencing respiratory failure, or requiring mechanical ventilation. The initial weighting coefficient is multiplied by a time decay factor and a disease similarity factor to obtain the data from the i-th data source at time point. The weighting coefficients are calculated using the following formula: ; in, ∈[0,1] indicates at time point The weight coefficient of the i-th type of data source; ∈[0.3,1] represents the initial weight coefficient of the i-th type of data source. The larger the value, the more important the i-th type of data source. It is set by professional technicians. The time decay factor, ∈[0.01,0.2] represents the time decay coefficient of the i-th type of data source, in hours. -1 , used to control the rate of decline of the exponential decay function, is obtained by taking the reciprocal of the average sampling period of the i-th type of data source; Indicates a point in time The most recent collection time with the i-th type of data source Time difference, in hours; This indicates the most recent collection time of the i-th type of data source, obtained through the timestamps of multi-source data on patient respiratory diseases; ∈[0,1] represents the disease similarity factor, used to measure the similarity of diseases, calculated at time points. Feature vector of the i-th type of data source The mean vector of features of historical severe cases The cosine similarity between the cases is obtained, and the closer the disease similarity factor is to 1, the more similar the current case is to the severe case. ∈ Indicates at a point in time The feature vector of the i-th type of data source, i.e., the feature vector of single-source data; ∈ This represents the mean vector of historical severe cases from the i-th data source, obtained by averaging the feature vectors of historical severe cases.

[0020] The above formula proposes a two-factor weighting mechanism based on time decay and disease similarity, which can dynamically calculate the weight coefficients of different data sources at different time points to enhance the perception of time changes in data sources. This strengthens key disease features during feature fusion while weakening the interference of outdated information and spurious correlation features of key diseases, thereby improving the sensitivity and robustness of feature fusion of multi-source data on respiratory diseases in patients.

[0021] Furthermore, the single-source data matrix With weighting coefficients Perform element-wise multiplication to obtain the weighted single-source data matrix. The weighted single-source data matrix is ​​constructed using a column-direction feature concatenation method. Fusion feature matrix fused into a unified dimension , This represents the feature dimension after fusion, with each row of the fused feature matrix F representing a time point. Fusion feature vectors from multiple sources Each column represents the sequence of changes of the j-th feature at all time points.

[0022] S2. Based on the fusion feature matrix, a bidirectional long short-term memory network is used with gating enhancement to construct a trend capture branch and output a trend feature vector; based on the fusion feature matrix, a mutation detection branch is constructed to calculate the residual mutation value and construct a residual mutation value vector; based on the trend feature vector and the residual mutation value vector, a trend-mutation consistency scoring mechanism is introduced to quantify the consistency between trends and mutations and construct a consistency scoring vector; based on the consistency scoring vector, trend feature vector, and residual mutation value vector, a joint feature vector is constructed; based on the joint feature vector, the respiratory disease warning level is output.

[0023] Because multi-source data on respiratory diseases exhibits both significant long-term time-series dependencies and localized mutations, existing Bidirectional Long Short-Term Memory (Bi-LSTM) models typically apply equal weights to each time point during time-series modeling. This makes it difficult to highlight responsiveness to critical time periods such as rapid deterioration of the condition, thus affecting the accuracy of disease early warning. Therefore, a time-series early warning modeling method based on a dual-branch structure of trend capture and mutation detection is proposed. A trend capture branch is constructed to capture the continuous evolution trend of the fused feature vector from multiple sources over time, outputting the trend feature vector. Simultaneously, a mutation detection branch is constructed to accurately locate drastic fluctuation events within a short period. The specific implementation process of the time-series early warning modeling method based on this dual-branch structure is as follows: A trend-capturing branch is constructed. Based on the existing Bi-LSTM model, a risk-label-guided gating enhancement mechanism is introduced. By introducing a gating gain coefficient, the forget gate output is enhanced to amplify the memory capacity during high-risk periods, resulting in a gating-enhanced forget gate output. The Bi-LSTM model consists of two layers, with each layer having a forward and backward hidden layer dimension of 128. The specific formula for the gating-enhanced forget gate output is as follows: ; in, Indicates the time point after gating enhancement The output of the forget gate; Indicates a point in time The gated gain coefficient, For time points The risk label, if =0 means If it is a low-risk time point, then =1, does not affect the output of the forget gate, if =1, then it means If it is a high-risk time point, then =1.5, used to enhance the forgetting gate output; the risk label is obtained by analyzing the collected multi-source data of patient respiratory diseases, specifically, if the patient at a certain time point If a significant deterioration event occurs, such as a sharp drop in blood oxygen saturation or a sudden increase in procalcitonin, then the time point will be [not specified]. All time points within the preceding consecutive ∆t-hour time window are marked as high-risk time points, i.e., if ,but =1, and the rest of the time points are marked as 0, where ∆t∈[1,2] is the adjustable warning window width in hours, which is set according to the type of respiratory disease and modeling requirements; σ(·) is the Sigmoid activation function; , The weight matrix and bias term of the forgetting gate are obtained through backpropagation. Specifically, the fused feature matrix is ​​fed into the gate-enhanced Bi-LSTM model for forward propagation to obtain a trend feature vector. This trend feature vector is then input into a hidden layer consisting of a fully connected layer and a ReLU nonlinear activation function, as well as a softmax classification layer, to output the probability distribution of risk levels. The actual respiratory disease risk levels are converted into one-hot encoded vectors. Based on these one-hot encoded vectors and the probability distribution of risk levels, a multi-class cross-entropy loss function is applied to calculate the prediction error. The actual respiratory disease risk levels are derived from the condition status of historical multi-source data annotated by professional technicians in the respiratory department database, with values ​​{0, 1, 2}, where 0 represents normal (no intervention required), 1 represents mild warning (observation recommended), and 2 represents severe warning (immediate intervention recommended). The prediction error is calculated using the chain rule to perform gradient calculation on the forgetting gate parameters, and backpropagation is performed layer by layer. The Adam optimizer is used to update the forgetting gate weight matrix according to a preset learning rate. and bias terms And other parameters of the Bi-LSTM model, until the multi-class cross-entropy loss function converges. Preferably, the initial learning rate of the Adam optimizer is set to... Set the batch size to 32 and the number of training rounds to 50 to 200. Indicates a point in time The hidden state is calculated by the Bi-LSTM model during the forward propagation process of the previous time step; Indicates a point in time Input features, i.e., time points The fused feature vector is taken from the k-th row of the fused feature matrix F; In the gated-enhanced Bi-LSTM model, the output of the gated-enhanced forget gate is... Used for adjustment The weight of the influence of the hidden state at time step on the current state, thus participating in the current time step. Hidden state The update of the hidden state; This is used to characterize the dynamic evolution information of fused features over time, and is taken as a time point. The trend feature vector, denoted as ; The aforementioned trend-capturing branch proposes a clinical risk label-guided gating enhancement mechanism in existing Bi-LSTM models, without explicitly increasing the structural complexity of the Bi-LSTM model, and outputs the enhanced forgetting gate. It can be directly used in the state update process of the Bi-LSTM model, which can dynamically improve the Bi-LSTM model's ability to remember data at high-risk time points, thus making the trend feature vector output by the Bi-LSTM model more accurate. It can more accurately capture the temporal patterns of high-risk stages, solve the problem of insufficient memory of key disease stages in traditional Bi-LSTM models, and improve early warning performance; Furthermore, a mutation detection branch is constructed, a historical sliding window is introduced, the standardized value of the j-th feature of the fused feature vector within the historical sliding window is extracted, and the feature mean of the historical sliding window is calculated; the value of the j-th feature of the fused feature vector at time point is calculated. The residual between the standardized value and the feature mean of the historical sliding window is used to obtain the j-th feature of the fused feature vector at time point. The residual mutation value is calculated using the following formula: ; in, For the j-th feature of the fused feature vector at time point The residual mutation value at time point represents the value at time point 1. The degree and direction of the abrupt change of the j-th feature relative to the feature mean of the historical sliding window, if Then the j-th feature is an abrupt increase, such as a sudden increase in respiratory rate or a rapid increase in procalcitonin. Then the j-th feature is a sudden decrease, such as a rapid drop in blood oxygen saturation or a sudden drop in body temperature. This indicates that the j-th feature has no significant mutation. The threshold for judging small fluctuations of the j-th feature in the fused feature vector can be the 10th quantile of the absolute value distribution of the residual mutation value of the j-th feature. This indicates that the j-th feature of the fused feature vector is at time point [time value missing]. The standardized value is taken from the k-th row and j-th column of the fusion feature matrix F; This represents the length of the historical sliding window for the j-th feature, which is adaptively set according to the data source type to which the feature belongs. Preferably, the value range is [3, 24]. Indicates the index of the historical sliding window time step; This indicates that the j-th feature of the fused feature vector is at time point [time value missing]. The standardized value is taken from the F-th row and s-th column of the fusion feature matrix; This represents the characteristic mean of the historical sliding window.

[0024] After the trend capture branch and mutation detection branch are constructed, trend feature vectors are obtained respectively. and residual mutation value vector To further improve the early warning capability for truly high-risk disease changes and reduce misjudgments under pseudo-mutations, a trend-mutation consistency scoring mechanism is introduced: the product of trend change, residual mutation value, and trend-mutation consistency score scaling factor is mapped to a trend-mutation consistency score using the Sigmoid function, as shown in the following formula: ; in, This indicates that the j-th feature is at time point The trend-mutation consistency score reflects whether the direction of trend change is consistent with the direction of mutation. The closer it is to 1, the more consistent the direction of the trend change is with the direction of the sudden change. The closer to 0, the more inconsistent the two directions are, which is a pseudo-mutation; λ∈[1,5] is the scaling factor of the trend-mutation consistency score, which is used to control the influence of the product of the trend change and the residual mutation value on the trend-mutation consistency score. It is obtained by the five-fold cross-validation method. This indicates that the j-th feature is at time point The upward trend change measure is used to assess the degree and direction of trend changes. This indicates that the trend is improving. This indicates that the trend is deteriorating. This indicates that the trend is stable. The threshold for determining the trend stability of the j-th feature can be taken as the 10th quantile of the absolute value distribution of the trend change of the j-th feature. To capture the j-th feature of the trend branch at time point Trend feature components; To capture the j-th feature of the trend branch at time point Trend feature components; For the j-th feature at time point The residual mutation value.

[0025] The above formula introduces a trend-mutation consistency scoring mechanism, which can effectively identify high-risk characteristics of "the trend is deteriorating and accompanied by mutations in the same direction", while suppressing false mutation characteristics caused by stable trends or inconsistent mutation directions, improving the ability to identify the truly high-risk moment of the disease and reducing the mutation false alarm rate.

[0026] Constructing a consistency score vector and the consistency score vector respectively with trend feature vector and residual mutation value vector Element-wise multiplication followed by concatenation yields the joint feature vector. , joint feature vector The input consists of a fully connected layer, a hidden layer with a ReLU nonlinear activation function, and a softmax classification layer. The output is the probability distribution of respiratory disease risk levels. The respiratory disease risk level corresponding to the highest probability is taken as the time point. The respiratory disease warning levels are used to guide intervention recommendations; the above respiratory disease risk levels are {0, 1, 2}, where 0 indicates normal and no intervention is needed; 1 indicates mild warning and observation is recommended; and 2 indicates severe warning and immediate intervention is recommended.

[0027] In summary, a respiratory early warning method based on multi-source data has been developed.

[0028] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0029] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0030] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A respiratory department early warning method based on multi-source data, characterized in that, Includes the following steps: S1. Collect multi-source data on patients' respiratory diseases, preprocess them to obtain single-source data feature vectors, and construct single-source data matrices. Introduce a two-factor weighting mechanism based on time decay and disease similarity. Based on the single-source data feature vectors, introduce initial weight coefficients from different data sources. Combine disease similarity factors and time decay factors to dynamically calculate the weight coefficients of different data sources. The time decay factor is obtained by applying an exponential decay function to the time difference between the current time point and the most recent collection time of the data source, and the time decay coefficient of the exponential decay function is determined based on the average sampling period of the data source. Based on the weight coefficients of the data sources, weight the single-source data matrix to obtain a weighted single-source data matrix. Use column-direction feature concatenation to concatenate all weighted single-source data matrices to obtain a fused feature matrix. S2. Based on the fusion feature matrix, a bidirectional long short-term memory network is adopted, and the forgetting gate output is gated and enhanced based on risk labels to construct a trend capture branch and output a trend feature vector. The risk label is marked as follows: when a patient's condition worsens, all time points within a consecutive preset time window before the occurrence of the worsening event are marked as high-risk time points, and the remaining time points are marked as low-risk time points. The forgetting gate output corresponding to the high-risk time points is enhanced. Based on the fusion feature matrix, a mutation detection branch is constructed to calculate the residual mutation value and construct a residual mutation value vector. Based on the trend feature vector and the residual mutation value vector, a trend-mutation consistency scoring mechanism is introduced to quantify the consistency between the trend and the mutation and construct a consistency scoring vector. Based on the consistency scoring vector, the trend feature vector, and the residual mutation value vector, a joint feature vector is constructed. Based on the joint feature vector, the respiratory disease warning level is output.

2. The respiratory department early warning method based on multi-source data according to claim 1, characterized in that, S1 specifically includes: In the preprocessing process, time alignment is performed on the multi-source data to obtain time-aligned multi-source data; the time-aligned multi-source data is then standardized according to different data sources to obtain single-source data feature vectors, and a single-source data matrix is ​​constructed.

3. The respiratory department early warning method based on multi-source data according to claim 1, characterized in that, S1 specifically includes: By introducing historical severe case data and calculating the cosine similarity between the feature vector of the single-source data and the feature mean vector of the historical severe case data, the disease similarity factor is obtained.

4. The respiratory department early warning method based on multi-source data according to claim 1, characterized in that, S1 specifically includes: The time decay coefficient is obtained by taking the reciprocal of the average sampling period of the data source.

5. The respiratory department early warning method based on multi-source data according to claim 1, characterized in that, S2 specifically includes: In the process of constructing the mutation detection branch, a historical sliding window is introduced. By calculating the residual difference between the standardized value of the feature in the fused feature vector and the feature mean of the historical sliding window, the residual mutation value is obtained, and a residual mutation value vector is constructed.

6. The respiratory department early warning method based on multi-source data according to claim 1, characterized in that, S2 specifically includes: In the implementation of the trend-mutation consistency scoring mechanism, a trend-mutation consistency scoring scaling factor is introduced. By combining the trend change amount and the residual mutation value, the consistency between the trend change direction and the mutation direction is judged, a trend-mutation consistency score is generated, and a consistency score vector is constructed. The trend change amount is calculated based on the trend feature components.

7. A respiratory early warning method based on multi-source data according to claim 6, characterized in that, S2 specifically includes: The consistency score vector is multiplied element-wise by the trend feature vector and the residual mutation value vector, and then concatenated to obtain the joint feature vector. The joint feature vector is then input into the fully connected layer and the classification layer in sequence to output the probability distribution of respiratory disease risk level. The category corresponding to the highest probability is taken as the respiratory disease warning level.

Images