COPD acute attack risk assessment method and system based on multi-parameter fusion

Abstract

The invention discloses a COPD acute attack risk assessment method and system based on multi-parameter fusion, and the method comprises the steps: S1, carrying out the preprocessing and adaptive standardization processing of multi-dimensional physiological parameter data of a COPD patient, and obtaining a standardized feature matrix; s2, based on the standardized feature matrix, constructing a bidirectional long-short-term memory network, designing a time sequence attention mechanism, constructing a time sequence feature extraction model, and obtaining a time sequence feature vector; s3, performing feature combination on the time sequence feature vector and the patient static feature to obtain a combined feature vector; s4, performing feature dimension reduction on the combined feature representation by using an auto-encoder, and extracting features after dimension reduction; and S5, constructing a COPD acute attack risk assessment model based on the dimensionality-reduced features, and performing risk assessment. According to the method, features can be automatically extracted through the deep learning model, risk assessment is carried out, and accurate assessment of the acute attack risk of the COPD patient is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment for acute exacerbation of COPD, and particularly to a method and system for risk assessment of acute exacerbation of COPD based on multi-parameter fusion. Background Art

[0002] Chronic obstructive pulmonary disease (COPD) is a common chronic respiratory disease, and its acute exacerbation can lead to a rapid deterioration of the patient's condition, seriously affecting the patient's quality of life and even threatening life. With the aggravation of population aging and the increasing prominence of environmental pollution problems, the incidence rate of COPD continues to rise, and the prevention and management of its acute exacerbation have become an important challenge in the current medical field. Accurately assessing the risk of acute exacerbation in COPD patients is of great significance for achieving early intervention and personalized treatment.

[0003] Currently, clinically, the risk of acute exacerbation of COPD is mainly assessed based on doctors' experience and single clinical indicators. This method has problems such as strong subjectivity and single assessment indicators. Although some studies have attempted to use machine learning methods for risk assessment, for example, Patent CN202210121112.7 proposes a method for improving the AUC of screened patients based on COPD risk judgment, which improves the screening accuracy through data characterization and model integration. However, this method mainly focuses on the processing of static features, fails to fully utilize the dynamic physiological parameter information of patients, and its feature extraction process is relatively simple, making it difficult to capture complex feature patterns.

[0004] The existing COPD risk assessment methods generally have the following problems: First, most methods only consider clinical indicators at a certain moment, ignoring the temporal variation characteristics of patients' physiological parameters and being unable to comprehensively reflect the disease evolution process of patients; Second, existing methods often simply combine different types of features and fail to perform effective feature engineering for the data characteristics of different features, resulting in limited feature expression ability; Third, the structure of traditional machine learning models is relatively simple, making it difficult to effectively process high-dimensional and multi-modal medical data, and the interpretability of the models is insufficient, making it difficult to gain the trust and recognition of doctors.

[0005] Therefore, there is an urgent need to develop a method for risk assessment of acute exacerbation of COPD that can comprehensively utilize patients' dynamic physiological parameters and static feature information, has strong feature learning ability and interpretability, so as to improve the accuracy and reliability of risk assessment and provide a more valuable reference basis for clinical decision-making. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion, aiming to make full use of the dynamic physiological parameters and static characteristic information of patients, automatically extract features through a deep learning model and perform risk assessment, so as to achieve an accurate assessment of the risk of acute exacerbation of COPD patients.

[0007] To achieve the above object, the method for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion provided by the present invention includes the following steps: S1: Preprocess and adaptively standardize the multi-dimensional physiological parameter data of COPD patients to obtain a standardized feature matrix; S2: Based on the standardized feature matrix, construct a bidirectional long short-term memory network and design a temporal attention mechanism to construct a temporal feature extraction model and obtain a temporal feature vector; S3: Combine the temporal feature vector and the static features of the patient to obtain a combined feature vector; specifically: perform Z-score standardization on the continuous static features to make them conform to the standard normal distribution, and perform one-hot encoding on the discrete static features to convert them into binary vector representations; splice and combine the processed continuous features, discrete features and the temporal feature vector to form a complete feature representation; S4: Use an autoencoder to perform feature dimensionality reduction on the combined feature vector and extract the reduced-dimensional features; S5: Based on the reduced-dimensional features, construct a risk assessment model for acute exacerbation of COPD and perform risk assessment.

[0008] Optionally, in the step S1, preprocessing and standardizing the physiological parameter data of COPD patients to obtain a standardized feature matrix includes: S11: Denoise the multi-dimensional physiological parameter data of COPD patients, specifically: ; wherein, is the denoised physiological parameter data of the th dimension at the time point ; is the original physiological parameter data of the th dimension at the time point ; is the time series index; is the physiological parameter dimension index, and is the total number of physiological parameter dimensions; is the summation function; is the denoising filter coefficient, is the filter index; S12: Calculate the statistical features of each dimension of physiological parameters, specifically: ; ; Among them, and are respectively the mean and standard deviation of the physiological parameters after denoising in the th dimension; is the length of the time series; S13: Perform adaptive standardization processing on the denoised physiological parameter data to obtain a standardized feature matrix , specifically: ; Among them, is the th dimension standardized feature matrix at time point ; is the adjustment parameter of the physiological parameters after denoising in the th dimension; is the natural constant.

[0009] Optionally, in step S2, based on the standardized feature matrix, obtain a time series feature vector, including: S21: Construct a bidirectional long short-term memory network to extract initial time series features, specifically: ; ; Among them, and are respectively the forward and backward hidden state outputs at time point ; is the forward hidden state output at time point - 1; are respectively the backward hidden state outputs at time point + 1; and are respectively the forward and backward long short-term memory networks; is the row vector of the standardized feature matrix at time point ; S22: Design a time series attention mechanism to calculate the attention weights at different time points, specifically: ; ; ; ; Among them, , and are respectively at time point Query vector, key vector, and value vector at , and are weight matrices of the query vector, key vector, and value vector respectively; is a vector concatenation operation; is a time point For the time point attention weights; is a time point key vector at; is a time point key vector at, ; is a matrix transpose operation; S23: Fuse temporal features based on attention weights, specifically: ; Among them, is a time point fused temporal features at; is a time point value vector at; S24: Introduce multi-head attention operation to enhance feature extraction ability, specifically: ; Among them, is the output of the multi-head attention operation at the time point ; is the multi-head attention operation; S25: Construct a temporal feature vector, specifically: ; Among them, is the final temporal feature vector; is the hyperbolic tangent function.

[0010] Optionally, in the S3 step, the temporal feature vector and the patient's static features are combined to obtain a combined feature vector; specifically: the continuous static features are standardized by Z-score to conform to the standard normal distribution, and the discrete static features are one-hot encoded into binary vector representations; the processed continuous features, discrete features, and the temporal feature vector are concatenated and combined to form a complete feature representation, including: S31: Standardize and one-hot encode the patient's static features: Standardize the continuous static features by Z-score: ; Among them, is the The eigenvalue after standardization of the th continuous static feature of the th patient; The th eigenvalue of the th continuous static feature of the th patient; And are the mean and standard deviation of the

[0011] th continuous static feature respectively; One-hot encoding is performed on the discrete static features: Among them, is the one-hot encoding vector of the th discrete static feature of the th patient, ; Among them, is the combined feature vector of the th patient; , and are the time series feature vector, the feature after standardization of the continuous static feature, and the one-hot encoding of the discrete static feature of the th patient respectively.

[0012] Optionally, in the S4 step, an autoencoder is used to perform feature dimensionality reduction on the combined feature representation and extract the reduced features, including: S41: Construct an autoencoder, which includes an encoder and a decoder, specifically: The encoder maps the input combined feature vector layer by layer to the reduced feature: ; ; ; Among them, , , are the output features of the first, second, and third encoding layers respectively; , , are the weight matrices of the first, second, and third encoding layers; , , are the bias vectors of the first, second, and third encoding layers; is a linear rectifier function; The decoder reconstructs the feature with reduced dimension layer by layer into the original feature dimension: ; ; ; Among them, , , are the output features of the first, second, and third decoding layers respectively; , , are the weight matrices of the first, second, and third decoding layers; , , are the bias vectors of the first, second, and third decoding layers; S42: Train the autoencoder: The total loss function of the autoencoder is:

[0013] Among them, is the total loss function; is the L2 norm; By minimizing the total loss function , the parameters of the autoencoder are optimized using the stochastic gradient descent method to obtain the feature with reduced dimension .

[0014] Optionally, in step S5, constructing a COPD acute exacerbation risk assessment model based on the feature with reduced dimension and performing risk assessment includes: S51: Construct a multi-layer perceptron network structure, specifically: ; Among them, is the output result of the multi-layer perceptron network; and are the weight matrices of the first and second hidden layers of the multi-layer perceptron network; and are the bias vectors of the first and second hidden layers of the multi-layer perceptron network; S52: Construct a Softmax classifier to calculate the probability of COPD acute exacerbation risk: ; Among them, is the probability vector of acute exacerbation risk, ; is the probability of the level risk; is the number of risk levels; is the weight matrix of the classification layer; is the bias vector of the classification layer; S53: Construct the loss function of the COPD acute exacerbation risk assessment model, specifically: For the predicted acute exacerbation risk probability vector of each sample , compare it with the true risk label, calculate the cross-entropy loss, and use it for the training and parameter optimization of the COPD acute exacerbation risk assessment model; S54: Use the trained COPD acute exacerbation risk assessment model to evaluate risks, specifically: Input the dimensionality-reduced features of the sample to be evaluated into the trained COPD acute exacerbation risk assessment model to obtain the predicted acute exacerbation risk probability vector; select the risk level corresponding to the dimension with the largest probability in the predicted acute exacerbation risk probability vector as the evaluation result; at the same time, output the probability value corresponding to this risk level as the prediction confidence.

[0015] The present invention also discloses a COPD acute exacerbation risk assessment system based on multi-parameter fusion, including: Standardization module: Preprocess and adaptively standardize the multi-dimensional physiological parameter data of COPD patients; Time series feature extraction module: Based on the standardized feature matrix, construct a bidirectional long short-term memory network and design a time series attention mechanism, construct a time series feature extraction model, and obtain the time series feature vector; Feature combination module: Combine the time series feature vector and the patient's static features to obtain the combined feature vector; Feature dimensionality reduction module: Use an autoencoder to perform feature dimensionality reduction on the combined feature vector and extract the dimensionality-reduced features; Model construction and risk assessment module: Construct a COPD acute exacerbation risk assessment model based on the dimensionality-reduced features and perform risk assessment.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects: The present invention uses a combination of a bidirectional long short-term memory network and a multi-head attention mechanism to process time series data, which can not only capture the long-term dependence relationship of physiological parameters, but also highlight the feature expression of important time points through the multi-head attention mechanism. This structural design overcomes the limitations of traditional methods in dealing with long sequence data. At the same time, through the synergistic effect of bidirectional information flow and multi-head attention, it enhances the ability to identify key features in the process of the patient's disease evolution, providing a more reliable feature basis for risk assessment.

[0017] The present invention innovatively fuses the dynamic physiological parameters of patients (i.e., the multi-dimensional physiological parameter data of COPD patients in the present invention) and static features, and ensures the effective integration of different types of features through targeted feature processing methods. The continuous static features are standardized, the discrete features are one-hot encoded, and then dimensionality reduction is performed through an autoencoder, which not only retains the essential information of the features but also improves the efficiency of feature representation. This feature fusion strategy comprehensively considers various factors affecting the acute exacerbation of COPD, making the risk assessment results more comprehensive and accurate.

[0018] The COPD acute exacerbation risk assessment model constructed by the present invention adopts a structure of a multi-layer perceptron combined with a Softmax classifier, which can not only output the risk level but also provide the confidence of the prediction, enhancing the interpretability of the model. The training process of the model is optimized by the cross-entropy loss function, ensuring the accuracy of the assessment results and effectively improving the prognosis management level of COPD patients. Brief Description of the Drawings

[0019] Figure 1 It is a schematic flow chart of the COPD acute exacerbation risk assessment method based on multi-parameter fusion according to an embodiment of the present invention; Figure 2 It is a schematic diagram of an autoencoder according to an embodiment of the present invention. Detailed Embodiments

[0020] The present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited in any way. Any transformation or replacement made based on the teachings of the present invention falls within the protection scope of the present invention.

[0021] Embodiment 1: The COPD acute exacerbation risk assessment method based on multi-parameter fusion, as Figure 1 shown, includes the following steps: S1: Preprocess and adaptively standardize the physiological parameter data of COPD patients to obtain a standardized feature matrix: S11: Denoise the multi-dimensional physiological parameter data of COPD patients, specifically: ; Wherein, is the denoised physiological parameter data of the th dimension at the time point ; is the original physiological parameter data of the th dimension at the time point ; is the summation function; is the time series index; is the physiological parameter dimension index, , is the total number of physiological parameter dimensions. In this embodiment , including blood oxygen saturation, blood pressure, heart rate, arterial blood carbon dioxide partial pressure, and bicarbonate concentration; is the denoising filter coefficient, is the filter index. In this embodiment , is the natural constant, is the pi; S12: Calculate the statistical features of physiological parameters in each dimension, specifically: ; ; Among them, and are the mean and standard deviation of the denoised physiological parameters in the th dimension respectively; is the time series length. In this embodiment ; S13: Perform adaptive standardization processing on the denoised physiological parameter data to obtain the standardized feature matrix , specifically: ; Among them, is the standardized feature matrix at time point in the th dimension; is the adjustment parameter of the denoised physiological parameters in the th dimension. In this embodiment .

[0022] It should be noted that in this step, by preprocessing and standardizing the multi-dimensional physiological parameter data of COPD patients, the noise interference and abnormal fluctuations in the original data can be effectively eliminated, and the data quality can be improved. The denoising processing method using a sliding window can smooth short-term fluctuations while retaining the long-term change trend of the data. In the standardization process, an adaptive adjustment mechanism is introduced, which not only realizes the scale unification of different-dimensional data, but also can adaptively adjust the standardization intensity according to the distribution characteristics of the data itself, avoiding the problem that the traditional fixed-parameter standardization method is overly sensitive to outliers.

[0023] S2: Based on the standardized feature matrix, construct a bidirectional long short-term memory network and design a temporal attention mechanism to construct a temporal feature extraction model to obtain a temporal feature vector: S21: Construct a bidirectional long short-term memory network to extract initial temporal features, specifically: ; ; Among them, and are the forward and backward hidden state outputs at the time point respectively; is the forward hidden state output at the time point -1; are the backward hidden state outputs at the time points +1 respectively; and are the forward and backward long short-term memory networks respectively. In this embodiment, the hidden layer dimension of the long short-term memory network is 256; is the row vector of the normalized feature matrix at the time point ; S22: Design a temporal attention mechanism to calculate the attention weights at different time points, specifically: ; ; ; ; Among them, , and are the query vector, key vector, and value vector at the time point respectively. In this embodiment, the dimension is 128; , and are the weight matrices of the query vector, key vector, and value vector respectively; is the vector concatenation operation; is the attention weight of the time point to the time point ; is the key vector at the time point ; is the key vector at the time point , ; is the matrix transpose operation; S23: Fuse the temporal features based on the attention weights, specifically: ; Among them, is the fused temporal feature at the time point ; is the value vector at the time point ; S24: Introduce multi-head attention operation to enhance the feature extraction ability, specifically: ; wherein, is the output of the multi-head attention operation at time point ; is the multi-head attention operation, and the number of heads of the multi-head attention operation in this embodiment is 8; S25: Construct a temporal feature vector, specifically: ; wherein, is the final temporal feature vector; is the hyperbolic tangent function; In this embodiment, the training method of the temporal feature extraction model is specifically: Construct the loss function of the temporal feature extraction model: ; wherein, is the total loss function; is the reconstruction loss; is the temporal consistency loss; is the regularization loss; The calculation methods of each loss term are: ; ; ; wherein, is the L2 norm; is the Frobenius norm; is the normalized feature matrix after averaging in the time dimension; is the reconstruction mapping function, specifically: ; wherein, and are the weight matrices of the reconstruction mapping. In this embodiment has a dimension of , has a dimension of ; and are and corresponding bias vectors; is the rectified linear unit function; is the sigmoid function; Use the stochastic gradient descent method to update the parameters in the temporal feature extraction model; when the difference between the loss function values of two adjacent times is less than 0.00001, stop training. In this embodiment, the initial learning rate is 0.001.

[0024] It should be noted that in this step, by constructing a time series feature extraction model based on a bidirectional long short-term memory network and a multi-head attention mechanism, the time series dynamic features of the physiological parameter data of COPD patients can be effectively captured. The bidirectional long short-term memory network can consider both historical and future information simultaneously, overcoming the limitation of traditional unidirectional networks that can only utilize historical information. The introduced time series attention mechanism can adaptively assign importance weights to features at different time points, highlighting the feature expressions at key time points. The multi-head attention mechanism further enhances the feature extraction ability of the model by learning feature relationships in different representation subspaces.

[0025] S3: Combine the time series feature vector and the patient's static features to obtain a combined feature vector; perform Z-score normalization on the continuous static features to make them conform to the standard normal distribution, and perform one-hot encoding on the discrete static features to convert them into binary vector representations; concatenate and combine the processed continuous features, discrete features, and the time series feature vector to form a complete feature representation: S31: Perform normalization processing and one-hot encoding on the patient's static features. The static features used in this embodiment include age, gender, smoking history, height, weight, daily oxygen inhalation duration, COPD assessment test, and chronic nasal symptom score: Perform Z-score normalization on the continuous static features: ; where is the normalized feature value of the th continuous static feature of the th patient; is the feature value of the th continuous static feature of the th patient; and are the mean and standard deviation of the th continuous static feature respectively; Perform one-hot encoding on the discrete static features: ; where is the one-hot encoding vector of the th discrete feature of the th patient; is the one-hot encoding function; S32: Feature combination: Construct a combined feature vector: ; where is the combined feature vector of the th patient; , and are the time series feature vector, the standardized features of continuous static features, and the one-hot encoding of discrete static features of the th patient, respectively.

[0026] It should be noted that in this step, through the effective combination of the patient's static features and time series features, the integration of comprehensive feature information of COPD patients is achieved. For static features that do not change over time, targeted standardization processing methods are adopted: standardize continuous static features to eliminate the influence of dimension, and perform one-hot encoding on discrete static features to maintain the categorical attributes of the features. This classification processing method fully considers the data characteristics of different types of features and ensures the complete retention of feature information.

[0027] S4: Use an autoencoder to perform feature dimensionality reduction on the combined feature representation and extract the reduced features: S41: Construct an autoencoder, which includes two parts: an encoder and a decoder, as Figure 2 shown. Specifically: The encoder maps the input combined feature vector layer by layer into the reduced features: ; ; ; where , , are the output features of the first, second, and third encoding layers, respectively; , , are the weight matrices of the first, second, and third encoding layers; , , are the bias vectors of the first, second, and third encoding layers; The decoder reconstructs the reduced features layer by layer into the original feature dimension: ; ; ; where , , are the output features of the first, second, and third decoding layers, respectively; , , are the weight matrices of the first, second, and third decoding layers; , , are the bias vectors of the first, second, and third decoding layers; S42: Train the autoencoder: The total loss function of the autoencoder is: ;

[0028] By minimizing the total loss function , use the stochastic gradient descent method to optimize the parameters of the autoencoder, and obtain the dimensionality-reduced features . In this embodiment, the training strategy is the same as that in step S2.

[0029] In this step, the dimensionality reduction process of the combined features is performed by constructing a multi-layer autoencoder, effectively solving the problems of data redundancy and computational efficiency that may be brought by high-dimensional features. A symmetric structure of three-layer encoder and decoder is adopted, and the essential representation of features is extracted layer by layer through non-linear activation functions, which not only retains the key information of the original features but also significantly reduces the feature dimension. During the encoding process, high-dimensional features are gradually compressed into a more compact low-dimensional representation, while the decoding process ensures the minimum loss of information during the dimensionality reduction process through the reconstruction task.

[0030] S5: Construct a COPD acute exacerbation risk assessment model based on the dimensionality-reduced features and perform risk assessment: S51: Construct a multi-layer perceptron network structure, specifically: ; where is the output result of the multi-layer perceptron network; and are the weight matrices of the first and second hidden layers of the multi-layer perceptron network; and are the bias vectors of the first and second hidden layers of the multi-layer perceptron network; S52: Construct a Softmax classifier to calculate the probability of COPD acute exacerbation risk: ; where is the acute exacerbation risk probability vector, ; is the probability of the th level of risk; is the number of risk levels. In this embodiment , which are divided into low risk, medium risk, high risk, and extremely high risk; is the weight matrix of the classification layer; is the bias vector of the classification layer; S53: Construct the loss function of the COPD acute exacerbation risk assessment model, specifically: The predicted acute exacerbation risk probability vector for each sample is compared with the true risk label, and the cross-entropy loss is calculated for the training and parameter optimization of the COPD acute exacerbation risk assessment model. In this embodiment, the Adam optimization method is used to train the COPD acute exacerbation risk assessment model, with an initial learning rate of 0.0003 and 100 training epochs; S54: Use the trained COPD acute exacerbation risk assessment model to evaluate the risk, specifically: Input the dimensionality-reduced features of the sample to be evaluated into the trained COPD acute exacerbation risk assessment model to obtain the predicted acute exacerbation risk probability vector; select the risk level corresponding to the dimension with the highest probability in the predicted acute exacerbation risk probability vector as the evaluation result; at the same time, output the probability value corresponding to this risk level as the prediction confidence.

[0031] It should be noted that in this step, by constructing a deep learning model based on a multi-layer perceptron and a Softmax classifier, the accurate assessment of the acute exacerbation risk of COPD patients is realized. The multi-layer perceptron network structure can perform deep non-linear transformation on the dimensionality-reduced features, fully explore the complex relationships between features, and extract more discriminative feature representations. The Softmax classifier maps the extracted features into probability distributions of different risk levels, not only giving the risk assessment result, but also providing the prediction confidence, making the assessment result more interpretable and credible.

[0032] To verify the effectiveness of the method of the present invention, 3000 cases of COPD patient data are used for verification in the present invention, of which 2400 cases are used for model training and 600 cases are used for testing.

[0033] The evaluation metrics used in the experiment include accuracy, precision, and recall. To prove the superiority of the method of the present invention, it is compared with three existing methods: the random forest method, the single LSTM network method, and the static feature fusion method. The experimental results are shown in Table 1: Table 1 Performance comparison of different methods

[0034] Embodiment 2: The present invention also discloses a COPD acute exacerbation risk assessment system based on multi-parameter fusion, including the following five modules: Normalization module: Perform preprocessing and adaptive normalization on the multi-dimensional physiological parameter data of COPD patients; Temporal feature extraction module: Based on the normalized feature matrix, construct a bidirectional long short-term memory network and design a temporal attention mechanism to construct a temporal feature extraction model to obtain a temporal feature vector; Feature combination module: combines the time-series feature vector and the patient's static features to obtain a combined feature vector; Feature dimensionality reduction module: uses an autoencoder to perform feature dimensionality reduction on the combined feature vector and extracts the features after dimensionality reduction; Model construction and risk assessment module: constructs a COPD acute exacerbation risk assessment model based on the features after dimensionality reduction and performs risk assessment.

[0035] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments. And the term "including", "comprising" or any other variant thereof in this article is intended to cover a non-exclusive inclusion, so that a process, device, article or method including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, device, article or method. Without more limitations, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, device, article or method including that element.

[0036] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium as described above (such as ROM / RAM, magnetic disk, optical disc), and includes several instructions to enable a terminal device (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in various embodiments of the present invention.

[0037] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion, characterized in that, It includes the following steps: S1: Preprocess and adaptively standardize the multi-dimensional physiological parameter data of COPD patients to obtain a standardized feature matrix; S2: Based on the standardized feature matrix, construct a bidirectional long short-term memory network and design a temporal attention mechanism to construct a temporal feature extraction model and obtain a temporal feature vector; S3: Combine the temporal feature vector and the patient's static features to obtain a combined feature vector; specifically: perform Z-score standardization on the continuous static features to make them conform to the standard normal distribution, and perform one-hot encoding on the discrete static features to convert them into binary vector representations; splice and combine the processed continuous static features, discrete static features and the temporal feature vector to form a complete combined feature representation; S4: Use an autoencoder to perform feature dimensionality reduction on the combined feature vector and extract the reduced-dimensional features; S5: Construct a COPD acute exacerbation risk assessment model based on the reduced-dimensional features and perform risk assessment.

2. The method for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion according to claim 1, wherein The step S1 includes the following steps: S11: Denoise the multi-dimensional physiological parameter data of COPD patients, specifically: ； wherein, is the time point at the dimensional denoised physiological parameter data; is the time point at the dimensional original physiological parameter data; is the time series index; is the physiological parameter dimension index, , is the total number of physiological parameter dimensions; is the summation function; is the denoising filter coefficient, is the filter index; S12: Calculate the statistical features of each dimension of physiological parameters, specifically: ； ； wherein, and are respectively the mean and standard deviation of the physiological parameters after denoising in the th dimension; is the length of the time series; S13: Perform adaptive normalization processing on the denoised physiological parameter data to obtain a normalized feature matrix , specifically: ； Among them, is the time point at the dimensional standardized feature matrix; is the adjustment parameter of the dimensional physiological parameter after denoising; is the natural constant.

3. The method for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion according to claim 2, wherein The step S2 includes the following steps: S21: Construct a bidirectional long short-term memory network to extract initial temporal features, specifically: ； ； wherein, and are the forward and backward hidden state outputs at the time point respectively; is the forward hidden state output at the time point -1; are the backward hidden state outputs at the time point +1 respectively; and are the forward and backward long short-term memory networks respectively; is the row vector of the normalized feature matrix at the time point ; S22: Design a temporal attention mechanism to calculate the attention weights at different time points, specifically: ； ； ； ； Among them, , and are the query vector, key vector, and value vector at time point respectively; , and are the weight matrices of the query vector, key vector, and value vector respectively; is the vector concatenation operation; is the attention weight of time point to time point ; is the key vector at time point ; is the key vector at time point , ; is the matrix transpose operation; S23: Fuse the temporal features based on the attention weights, specifically: ； Among them, is the fused temporal feature at the time point ; is the value vector at the time point ; S24: Introduce multi-head attention operations to enhance the feature extraction ability, specifically: ； Among them, is the output of the multi-head attention operation at the time point; is the multi-head attention operation; S25: Construct a temporal feature vector, specifically: ； Among them, is a time series feature vector; is the hyperbolic tangent function.

4. The method for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion according to claim 3, wherein The step S3 includes the following steps: S31: Perform standardization processing and one-hot encoding processing on the patient's static features: Perform Z-score standardization on the continuous static features: ； Among them, is the feature value after standardization of the th continuous static feature of the th patient; is the feature value of the th continuous static feature of the th patient; and are the mean and standard deviation of the th continuous static feature respectively; Perform one-hot encoding on the discrete static features: ； Among them, is the -th one-hot encoded vector of the discrete static feature of the -th patient, and is the one-hot encoding function; S32: Feature combination to obtain a combined feature vector: ； Among them, is the combined feature vector of the -th patient; , and are the time series feature vector, the feature after normalization of the continuous static feature, and the one-hot encoding of the discrete static feature of the -th patient, respectively.

5. The method for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion according to claim 4, wherein The step S4 includes the following steps: S41: Construct an autoencoder, and the autoencoder includes two parts: an encoder and a decoder, specifically: The encoder maps the input combined feature vectors layer by layer into the features after dimensionality reduction: ； ； ； Among them, , , are the output features of the first, second, and third coding layers respectively; , , are the weight matrices of the first, second, and third coding layers; , , are the bias vectors of the first, second, and third coding layers; is the rectified linear unit function; The decoder reconstructs the reduced-dimensional features layer by layer into the original feature dimension: ； ； ； Among them, , , are the output features of the first, second, and third decoding layers respectively; , , are the weight matrices of the first, second, and third decoding layers; , , are the bias vectors of the first, second, and third decoding layers; S42: Train the autoencoder: The total loss function of the autoencoder is: ； Among them, is the total loss function; is the L2 norm; By minimizing the total loss function , the parameters of the autoencoder are optimized using the stochastic gradient descent method to obtain the dimensionality-reduced features .

6. The method for assessing the risk of acute exacerbation of COPD based on multi-parameter fusion according to claim 5, wherein The step S5 includes the following steps: S51: Construct a multi-layer perceptron network structure, specifically: ； Among them, is the output result of the multi-layer perceptron network; and are the weight matrices of the first and second hidden layers of the multi-layer perceptron network; and are the bias vectors of the first and second hidden layers of the multi-layer perceptron network; S52: Construct a Softmax classifier to calculate the risk probability of COPD acute exacerbation: ； Among them, is the risk probability vector of acute exacerbation of COPD, ; is the probability of the -th level risk; is the number of risk levels; is the weight matrix of the classification layer; is the bias vector of the classification layer; S53: Construct the loss function of the COPD acute exacerbation risk assessment model, specifically: The predicted acute exacerbation risk probability vector for each sample is compared with the true risk label, and the cross-entropy loss is calculated for the training and parameter optimization of the COPD acute exacerbation risk assessment model; S54: Use the trained COPD acute exacerbation risk assessment model to evaluate the risk, specifically: Input the reduced-dimensional features of the sample to be evaluated into the trained COPD acute exacerbation risk assessment model to obtain a predicted acute exacerbation risk probability vector; select the risk level corresponding to the dimension with the largest probability in the predicted acute exacerbation risk probability vector as the evaluation result; at the same time, output the probability value corresponding to this risk level as the prediction confidence.

7. A COPD acute exacerbation risk assessment system based on multi-parameter fusion, characterized in that, It includes: A standardization module: Preprocess and adaptively standardize the multi-dimensional physiological parameter data of COPD patients; Temporal Feature Extraction Module: Based on the standardized feature matrix, construct a bidirectional long short-term memory network and design a temporal attention mechanism to build a temporal feature extraction model and obtain a temporal feature vector; Feature Combination Module: Combine the temporal feature vector and the patient's static features to obtain a combined feature vector; Feature Dimensionality Reduction Module: Use an autoencoder to perform feature dimensionality reduction on the combined feature vector and extract the reduced features; Model Construction and Risk Assessment Module: Build a COPD acute exacerbation risk assessment model based on the reduced features and conduct risk assessment; To implement the COPD acute exacerbation risk assessment method based on multi-parameter fusion as described in any one of claims 1-6.

Images