Interpretable ordered classifier and system for predicting prognosis of AECOPD patient

Abstract

The invention relates to an interpretable ordered classifier and system for predicting prognosis of an AECOPD patient, and belongs to the technical field of medical data processing, and the system comprises a data processing module which is used for preprocessing medical data and completing multi-modal and nonlinear data fusion; the model construction module is used for capturing implied nonlinear modes in the medical data and processing components of class imbalance and sparse data; the model training module constructs the loss function optimization model into a convex quadratic optimization model, learns explainable parameters, and creatively explains a nonlinear compensation effect through a power average operator; the prediction module is connected with the model training module, predicts the prognosis of the AECOPD patient by using the trained interpretable ordered classifier, and establishes a prognosis risk dynamic mapping model based on a piecewise linear value function; the method has the beneficial effects that class imbalance and sparsity of real medical data are processed, the influence of nonlinear features on prognosis risks is captured, and the clinical credibility and application value of a prediction result are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical data processing technology, and specifically relates to an interpretable ordered classifier and system for predicting the prognosis of AECOPD patients. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a major global public health problem, and acute exacerbations of COPD (AECOPD) significantly increase the risk of death in patients. Effective prognostic management of AECOPD patients is crucial, but existing scoring systems have limitations in predicting prognosis and struggle to accurately identify high-risk patients for optimal healthcare resource allocation.

[0003] Data-driven intelligent clinical decision support systems are typically based on black-box models to capture nonlinearities in data, thus achieving extremely high prediction accuracy. However, black-box models lack interpretability, which contradicts the interpretability requirements of high-risk decisions in the medical field, limiting the deployment of data-driven intelligent clinical decision support in medical practice. Summary of the Invention

[0004] This invention constructs an interpretable ordered classifier and system for predicting the prognosis of patients with acute exacerbation of coronary heart disease (AECOPD), aiming to address the problem of insufficient interpretability in existing medical data modeling. The system effectively captures the nonlinear characteristics of medical data by introducing an interpretable parameter structure and performs structured processing to address the inherent class imbalance and sparsity issues of the data, thereby improving the stability and clinical applicability of the model.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] An interpretable ordered classifier for predicting the prognosis of patients with acute exacerbation of COPD includes:

[0007] The data processing module is used to preprocess the multimodal electronic health records of AECOPD patients, realize the fusion of structured and unstructured data, and generate medical data with nonlinear characteristics.

[0008] The model building module, connected to the data processing module, is based on interpretable parameter structure modeling to identify potential nonlinear patterns in medical data and integrates mechanisms to handle class imbalance and data sparsity. Specifically, the class imbalance problem is handled by selective undersampling using a combination of K-means clustering and density-aware hierarchical clustering, and the data sparsity problem is addressed by introducing a factorization machine structure to model high-order feature interaction relationships.

[0009] The model training module is connected with the model construction module, trains the designed interpretable learning model, adopts a power averaging operator to realize quantitative modeling of a nonlinear compensation effect, and converts an overall optimization problem into a convex quadratic programming model to improve training stability.

[0010] The prediction module is connected with the model training module, uses the trained interpretable ordinal classifier to predict the prognosis risk level of the AECOPD patient, and constructs a dynamic risk mapping mechanism based on a segmented linear value function to realize continuous representation of risk grading.

[0011] Optionally, when the data processing module pre-processes the multi-modal electronic health records, specifically includes: filling the missing values by using an interpolation method based on medical field knowledge, identifying and correcting the abnormal values by Z-score standardization combined with clinical diagnosis standards, and extracting features from the text health records by using a BERT model and converting them into structured data.

[0012] Optionally, in the model construction module, the under-sampling strategy of K-means and hierarchical density clustering is specifically: first, divide the majority class samples into K clusters by K-means clustering, then select core samples from each cluster by hierarchical density clustering and delete edge samples to realize under-sampling, wherein the value of K is dynamically adjusted according to the ratio of the number of majority class samples to the number of minority class samples.

[0013] Optionally, when the factorization machine depicts the interaction of features, a combination of second-order cross-features and high-order features is adopted; wherein the high-order feature combination gives different feature interactions different weights on the prediction results by introducing an attention mechanism.

[0014] Optionally, in the model training module, the parameter value range of the power averaging operator is [0.5, 2], and the optimal parameter value is determined by grid search method to maximize the quantitative precision of the nonlinear compensation effect.

[0015] Optionally, when calculating the power averaging operator, the time decay factor of different features is considered, and higher weight is given to the patient data collected in the near future. The time decay factor is calculated by an exponential function, and the decay rate is determined according to the update period of clinical data.

[0016] Optionally, the constraint conditions of the convex quadratic model include the non-negativity constraint of feature weights and the probability constraint that the sum of sample prediction probabilities is 1, and the optimal solution of the model is solved by the Lagrange multiplier method.

[0017] Optionally, the segmented linear value function in the prediction module takes patient age, disease duration and inflammation index level as segmentation nodes, and fits the mapping relationship between prognosis risk value and input features in each segmentation interval by a linear regression model.

[0018] Optionally, the multi-modal electronic health record includes image data of the patient, and when the data processing module pre-processes the image data, a deep learning-based image segmentation algorithm is used to extract lung key region features.

[0019] An interpretable ordinal classification system for predicting the prognosis of AECOPD patients, comprising:

[0020] A data processing unit for acquiring nonlinear data or medical data;

[0021] A model unit connected to the data processing unit, comprising a model construction module and a model training module;

[0022] A prediction output unit connected to the model unit and capable of outputting a prediction result.

[0023] The beneficial effects of the present application are:

[0024] The present application aims at the problems of nonlinear structure, class imbalance and feature sparsity commonly existing in the electronic health record of AECOPD patients, and constructs an ordinal classification model with clinical interpretability. On the one hand, the structured undersampling of the majority class samples is realized by introducing K-means combined with density-aware hierarchical clustering method, the core case samples representative in medicine are effectively retained, and the redundant marginal information is eliminated, so as to alleviate the class imbalance problem without sacrificing the expression of clinical heterogeneity. On the other hand, the model introduces factorization machine structure, fuses second-order cross features and high-order feature combinations as the basis of nonlinear modeling, dynamically gives different weights to the interaction of key features through attention mechanism, and strengthens the recognition ability of the model to the potential relationship between sparse features, so as to improve the expression ability and robustness of the model in the high-dimensional and low-sample medical scene. The overall method not only ensures that the model output has clear parameter interpretability, but also enhances the application stability and doctor trust in the clinical decision support system. BRIEF DESCRIPTION OF DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0026] Figure 1 A structure diagram of an interpretable ordinal classification system for predicting the prognosis of AECOPD patients according to the present application;

[0027] Figure 2 A structure diagram of an interpretable ordinal classifier for predicting the prognosis of AECOPD patients according to the present application. DETAILED DESCRIPTION

[0028] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0029] Embodiment 1

[0030] As shown in the Figure 1 embodiment, the present embodiment provides an interpretable ordinal classification system for predicting the prognosis of AECOPD patients, comprising:

[0031] a data processing unit for acquiring nonlinear data or medical data, including a data processing module;

[0032] a model unit connected to the data processing unit, including a model construction module and a model training module;

[0033] a prediction output unit including a prediction module connected to the model unit, capable of outputting a prediction result.

[0034] The data processing unit acquires nonlinear data or medical data, the model unit trains the acquired nonlinear data or medical data, and the prediction output unit outputs the prediction result.

[0035] Embodiment 2

[0036] Based on embodiment 1, as shown in the Figure 2 embodiment, the present embodiment provides an interpretable ordinal classifier for predicting the prognosis of AECOPD patients, comprising:

[0037] a data processing module for preprocessing the multi-modal electronic health records of AECOPD patients, realizing the fusion of structured and unstructured data, and generating medical data with nonlinear features;

[0038] a model construction module connected to the data processing module, based on an interpretable parameter structure modeling, for identifying potential nonlinear patterns in medical data and integrating mechanisms for handling class imbalance and data sparsity; wherein the class imbalance problem is handled by selective undersampling through the combination of K-means clustering and density-aware hierarchical clustering method, and the data sparsity problem is addressed by introducing a factorization machine structure to model high-order feature interaction relationships;

[0039] a model training module connected to the model construction module, for training the designed interpretable learning model, using a power averaging operator to realize the quantitative modeling of nonlinear compensation effect, and converting the overall optimization problem into a convex quadratic programming model to improve the training stability;

[0040] The prediction module is connected with the model training module, uses the trained explainable ordinal classifier to predict the prognosis risk level of the AECOPD patient, and constructs a dynamic risk mapping mechanism based on a segmented linear value function, so as to realize continuous representation of risk grading.

[0041] Specifically, when the data processing module pre-processes the multi-modal electronic health record, the following steps are included: for missing values, a medical field knowledge-based interpolation method is used for filling; for abnormal values, Z-score standardization is combined with clinical diagnosis standards for identification and correction; for text health records, a BERT model is used for feature extraction and conversion into structured data.

[0042] For missing values, a medical field knowledge-based interpolation method is used for filling. Because medical data has professionalism and particularity, the medical field knowledge-based interpolation method can more reasonably estimate the missing values and ensure the effectiveness of the data. For missing values in laboratory examination data, such as missing of some indicators in blood gas analysis, based on medical field knowledge, if the indicator has a strong clinical correlation with other related indicators of the same examination, an interpolation method based on these related indicators is used for filling. For example, when arterial oxygen partial pressure (PaO2) is missing, but arterial carbon dioxide partial pressure (PaCO2) and other indicators are complete, interpolation is performed according to the common correlation between the two in the clinic.

[0043] For abnormal values, Z-score standardization is combined with clinical diagnosis standards for identification and correction. Z-score standardization can convert data into a standard normal distribution, making it easier to identify data that deviates from the normal range, and then combining clinical diagnosis standards for judgment and correction can ensure the accuracy of abnormal value processing. First, the data is standardized by Z-score, and the calculation formula is: ; wherein, is the original data, is the mean, is the standard deviation. When the absolute value of Z-score is greater than 3, it is preliminarily judged as an abnormal value. Then, further confirmation is made in combination with the clinical diagnosis standard, such as a patient's white blood cell count being abnormally high, if combined with the patient's symptoms, signs and other examination results, it is consistent with the clinical diagnosis of infection, then the value may be a reasonable value and is not corrected; if not, it is corrected according to clinical experience and the data distribution of similar patients.

[0044] For text-based health records, the BERT model is used for feature extraction and conversion to structured data. BERT model has excellent performance in natural language processing, which can effectively extract key information from text and convert it into structured data for subsequent processing and analysis. The pre-trained BERT model is used for feature extraction of text-based medical records, first the text is preprocessed, including word segmentation and stop word removal. Then the processed text is input into the BERT model to obtain the feature vector of the text, and then it is converted into structured data. For example, for the diagnosis description in the medical record, the key diagnostic terms and related information are extracted and converted into corresponding structured features.

[0045] In addition, the multi-modal electronic health record includes the patient's image data, and when the data processing module pre-processes the image data, a deep learning-based image segmentation algorithm is used to extract lung key region features. Deep learning-based image segmentation algorithms can accurately locate and segment lung key regions, providing valuable image features for subsequent prognosis prediction. For the patient's chest CT image data, a deep learning-based U-Net image segmentation algorithm is used to extract lung key region features. First, the image data is pre-processed, including image normalization and noise reduction. Then, the trained U-Net model is used to segment the lung, and the mask of the lung region is obtained, and then the texture features, morphological features and other key features are extracted from the mask region.

[0046] The interpretable parameter structure is used to capture non-linear data, which has the characteristics of class imbalance and data sparsity. The class imbalance and data sparsity of non-linear data need to be handled.

[0047] In the model construction module, the K-means and hierarchical density clustering undersampling strategy is used to handle class imbalance. Specifically, first, the majority class samples are divided into K clusters by K-means clustering, and then the core samples are selected from each cluster by hierarchical density clustering, and the edge samples are deleted to achieve undersampling. The value of K is dynamically adjusted according to the ratio of the number of majority class samples to the number of minority class samples. This undersampling strategy can reduce the number of majority class samples while retaining the key information of majority class samples, effectively alleviating the class imbalance problem.

[0048] The K-means and hierarchical density clustering undersampling strategy is as follows: assuming that the number of majority class samples is N and the number of minority class samples is M, the value of K is dynamically adjusted according to the ratio of N to M, for example, when N / M=5, K takes 5. First, the majority class samples are divided into K clusters by K-means clustering, and the Euclidean distance is used as the similarity measure in the clustering process. Then, hierarchical density clustering is used for each cluster, with appropriate neighborhood radius and minimum sample size, to select core samples and delete edge samples to achieve undersampling.

[0049] In the factor decomposition machine, the second-order cross feature and the high-order feature combination are used to depict the feature interaction. The second-order cross feature is a common feature combination method, which uses an explicit or implicit method to construct the interaction information between features to improve the model expression ability. The high-order feature combination assigns different weights to different feature interactions by introducing an attention mechanism, considers feature interactions of different orders, and assigns different weights to them, which can better mine the potential information in the data and improve the model's processing ability for sparse data.

[0050] The factor decomposition machine is used to depict feature interaction. Specifically, for the second-order cross feature, the interaction term between features is calculated. For high-order feature combination, an attention mechanism is introduced to determine the weight of different feature interactions on the prediction result through training. For example, for the two features of age and disease duration, not only their second-order interaction is considered, but also their high-order interaction with other features such as inflammatory indicators, and different weights are assigned, which can better mine the potential information in the data and improve the model's processing ability for sparse data.

[0051] In the model training module, the parameter value range of the power averaging operator is [0.5, 2], and the optimal parameter value is determined by grid search to maximize the quantization precision of the nonlinear compensation effect. In calculating the power averaging operator, the time decay factor of different features is considered, and higher weight is given to the patient data collected in the near future. The time decay factor is calculated by an exponential function, and its decay rate is determined according to the update period of clinical data. Considering the time decay factor, the model can pay more attention to the influence of recent data on the prognosis, which is consistent with the clinical actual situation.

[0052] The step size of the power averaging operator grid search is set to 0.1, and all parameter values in this range are traversed to calculate the quantization precision of the nonlinear compensation effect under different parameter values, and the parameter value with the highest precision is selected. In calculating the power averaging operator, the time decay factor of different features is considered, and the exponential function of the time decay factor is: ; wherein is the interval between the data collection time and the current time, is the decay rate, which is determined according to the update period of clinical data, such as daily inspection data, The value is 0.1.

[0053] The constraint conditions of the convex quadratic model include the non-negativity constraint of feature weights and the probability constraint that the sum of sample prediction probabilities is 1, and the optimal solution of the model is solved by the Lagrange multiplier method.

[0054] The objective function of the convex quadratic model is to minimize the prediction error, and the constraint conditions include the non-negativity constraint of feature weights and the probability constraint that the sum of sample prediction probabilities is 1. The constraint conditions are integrated into the objective function by the Lagrange multiplier method, a Lagrange function is constructed, and then the partial derivative of the Lagrange function is solved, and the optimal solution of the model is obtained by setting the partial derivative to 0.

[0055] The segmented linear value function in the prediction module takes patient age, disease duration and inflammation index level as segmentation nodes, and the mapping relationship between prognosis risk value and input features in each segmentation interval is fitted by a linear regression model. Using these key clinical indicators as segmentation nodes can make the prediction results of the model more in line with clinical practice, and the use of the linear regression model enhances the interpretability of the model.

[0056] The segmented linear value function takes patient age, disease duration and inflammation index level as segmentation nodes. For example, age is divided into two intervals with 60 years as the boundary, disease duration is divided into two intervals with 5 years as the boundary, and inflammation index level is divided into four intervals of normal, mild elevation, moderate elevation and severe elevation according to the clinical normal range. The mapping relationship between prognosis risk value and input features in each segmentation interval is fitted by a linear regression model, and the parameters of the linear regression model are fitted by training data. When new patient data is input, the segmentation interval of the patient is determined according to the patient's age, disease duration and inflammation index level, and then the prognosis risk value is calculated by using the linear regression model corresponding to the interval.

[0057] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An interpretable ordinal classifier for predicting the prognosis of an AECOPD patient, characterized in that, The application relates to an AECOPD prognosis risk prediction method based on interpretable learning, which comprises the following steps: A data processing module is used for preprocessing multi-modal electronic health records of AECOPD patients, realizing fusion of structured and unstructured data, and generating medical data with nonlinear characteristics; A model construction module is connected with the data processing module, and is used for identifying potential nonlinear patterns in medical data and integrating mechanisms for processing class imbalance and data sparsity based on an interpretable parameter structure modeling; A model training module is connected with the model construction module, and is used for training the designed interpretable learning model, realizing quantitative modeling of nonlinear compensation effect by using a power averaging operator, and converting an overall optimization problem into a convex quadratic programming model to improve training stability; A prediction module is connected with the model training module, and is used for predicting the prognosis risk grade of AECOPD patients by using the trained interpretable ordinal classifier, and constructing a dynamic risk mapping mechanism based on a piecewise linear value function to realize continuous representation of risk grading.

2. The interpretable ordered classifier for predicting the prognosis of an AECOPD patient according to claim 1, wherein, When the data processing module preprocesses the multi-modal electronic health records, the following steps are specifically included: missing values are filled by using an interpolation method based on medical field knowledge, abnormal values are identified and corrected by using Z-score standardization combined with clinical diagnosis standards, and text health records are converted into structured data by using a BERT model for feature extraction.

3. The explainable ordinal classifier for predicting the prognosis of an AECOPD patient according to claim 1, characterized in that, In the model construction module, the undersampling strategy of K-means and hierarchical density clustering is specifically as follows: first, the majority class samples are divided into K clusters by K-means clustering, then core samples are selected from each cluster by hierarchical density clustering, and edge samples are deleted to realize undersampling, wherein the value of K is dynamically adjusted according to the ratio of the number of majority class samples to the number of minority class samples.

4. The explainable ordinal classifier for predicting the prognosis of an AECOPD patient according to claim 3, characterized in that, When the factorization machine describes the interaction of features, a combination of second-order cross features and high-order features is adopted; wherein the high-order feature combination gives different feature interactions different weights on the prediction result by introducing an attention mechanism.

5. The explainable ordinal classifier for predicting the prognosis of an AECOPD patient according to claim 1, wherein, In the model training module, the parameter value range of the power averaging operator is [0.5, 2], and the optimal parameter value is determined by a grid search method to maximize the quantitative precision of the nonlinear compensation effect.

6. The explainable ordinal classifier for predicting the prognosis of an AECOPD patient according to claim 5, characterized in that, When the power averaging operator is calculated, the time decay factor of different features is considered, and higher weight is given to the patient data collected in the near future; the time decay factor is calculated by an exponential function, and the decay rate is determined according to the update period of clinical data.

7. The interpretable ordered classifier for predicting the prognosis of an AECOPD patient according to claim 1, wherein, The constraint conditions of the convex quadratic model include the non-negativity constraint of feature weights and the probability constraint that the sum of sample prediction probabilities is 1, and the optimal solution of the model is solved by a Lagrange multiplier method.

8. The interpretable ordered classifier for predicting the prognosis of an AECOPD patient according to claim 1, wherein, In the prediction module, the piecewise linear value function takes patient age, disease duration and inflammation index level as segmentation nodes, and the mapping relationship between prognosis risk value and input features in each segmentation interval is fitted by a linear regression model.

9. The interpretable ordered classifier for predicting the prognosis of an AECOPD patient according to claim 1, wherein, The multi-modal electronic health record includes image data of a patient, and when the data processing module pre-processes the image data, a deep learning-based image segmentation algorithm is used to extract lung key region features.

10. An interpretable ordinal classification system for predicting the prognosis of an AECOPD patient for performing an interpretable ordinal classifier for predicting the prognosis of an AECOPD patient according to any one of claims 1-9, characterized in that, Comprise: A data processing unit for acquiring nonlinear data or medical data; A model unit connected with the data processing unit, comprising a model construction module and a model training module; A prediction output unit connected with the model unit and capable of outputting a prediction result.

Images