Multi-label chronic disease prediction method, device and system

Abstract

The invention discloses a multi-label chronic disease prediction method, device and system. The method comprises the following steps: acquiring electronic health record data of a plurality of to-be-predicted patients, and constructing corresponding medical history feature representations and patient feature representations; calculating the medical history similarity between the to-be-predicted objects according to the medical history feature representation, and constructing a patient similarity relation graph based on a similarity threshold value and a neighbor screening strategy; performing feature construction on edges in the patient similarity relation graph, and screening and filtering the edges through an edge classification model to obtain an optimized patient relation graph; and inputting the optimized patient relation graph and the corresponding patient features into a graph neural network model, and outputting a multi-label chronic disease prediction result corresponding to the to-be-predicted object. Through the patient relation modeling and edge filtering mechanism based on the medical history similarity, the influence of noise correlation on the prediction result is effectively reduced, and the accuracy and stability of multi-label chronic disease prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of medical and health data analysis and disease risk prediction technology, specifically to a multi-label chronic disease prediction method, device, and system. Background Technology

[0002] In recent years, Electronic Health Records (EHRs) have been widely used in medical institutions, accumulating a large amount of structured and semi-structured data containing information on diagnosis, medication, laboratory tests, and hospitalization processes. Disease risk prediction based on EHRs can provide data support for early screening, follow-up management, and decision support for chronic diseases, and therefore has received continuous attention from researchers and industry.

[0003] Chronic diseases are characterized by long course, numerous influencing factors, and widespread comorbidity. In real-world clinical scenarios, a single patient may simultaneously or sequentially develop multiple chronic disease risks, often presenting prediction tasks as multi-label risk prediction problems. Compared to single-disease prediction, multi-label prediction requires addressing issues such as correlations between labels, uneven data distribution, and individual patient differences, significantly impacting the accuracy and stability of multi-label disease prediction and posing challenges to the construction and practical application of prediction models. Summary of the Invention

[0004] To address the shortcomings of existing technologies in multi-label chronic disease risk prediction, this invention proposes a multi-label chronic disease prediction method, device, and system. By constructing a relationship structure among patients and filtering and optimizing these relationships, the invention achieves joint prediction of the risks of multiple chronic diseases, thereby improving the accuracy and stability of the prediction results.

[0005] Firstly, this invention proposes a multi-label chronic disease prediction method. The method includes:

[0006] Obtain electronic health record data of multiple subjects to be predicted, and construct corresponding medical history feature representations and patient feature representations based on the electronic health record data;

[0007] The medical history similarity S between any two objects to be predicted is calculated based on the medical history feature representation. Based on the medical history similarity S, a patient similarity graph G is constructed, where node V represents the object to be predicted and edge E represents the medical history association between any two objects to be predicted.

[0008] The patient similarity graph G is filtered and screened using the full data of the electronic health record data of the object to be predicted, resulting in an optimized patient similarity graph G'. The full data includes medical history data and various test data.

[0009] Based on the optimized patient similarity graph G', the patient feature representation is input into the graph neural network model, and the multi-label chronic disease prediction result of the object to be predicted is output.

[0010] Secondly, the present invention proposes a multi-label chronic disease prediction device, which is used to implement the multi-label chronic disease prediction method based on patient similarity graphs as described in the first aspect of the present invention, characterized in that the device comprises:

[0011] The data acquisition module is used to acquire electronic health record data of multiple subjects to be predicted, and to construct medical history feature representations and patient feature representations;

[0012] The mapping module is used to calculate the medical history similarity between the objects to be predicted based on the medical history feature representation, and to construct a patient similarity relationship graph;

[0013] The edge filtering module is used to construct edge features from the edges in the patient similarity graph and filter them based on the edge classification model to obtain an optimized patient relationship graph.

[0014] The prediction module is used to input the optimized patient relationship graph and the corresponding patient feature representation into the graph neural network model and output the multi-label chronic disease prediction results.

[0015] Thirdly, the present invention also proposes a multi-label chronic disease prediction system, which includes a processor, a communication interface, and a memory, wherein the processor, the communication interface, and the memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to invoke the program instructions to execute the multi-label chronic disease prediction method based on patient similarity graphs as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By constructing patient similarity relationships based on medical history information and screening and optimizing these relationships, this invention effectively reduces the impact of noise associations on prediction results; this invention can achieve joint prediction of the risk of multiple chronic diseases in scenarios with multiple coexisting diseases, improving the accuracy and stability of prediction results and demonstrating strong practicality; this invention is applicable to chronic disease risk assessment and auxiliary diagnosis scenarios and has good application prospects. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating an application scenario of an embodiment of this application;

[0018] Figure 2 This is a schematic flowchart illustrating the steps of the multi-label chronic disease prediction method in this application.

[0019] Figure 3This is a schematic diagram of the processing flow of the multi-label chronic disease prediction method according to an embodiment of this application;

[0020] Figure 4 This is a schematic diagram of graph construction and edge filtering optimization based on medical history similarity in an embodiment of this application;

[0021] Figure 5 This is a schematic diagram of the structure of the convolutional neural network in the embodiments of this application;

[0022] Figure 6 This is a schematic diagram of the structure of the multi-label chronic disease prediction device according to an embodiment of this application;

[0023] Figure 7 This is a schematic diagram of the structure of the multi-label chronic disease prediction system according to an embodiment of this application. Detailed Implementation

[0024] The terms "first," "second," "third," "fourth," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be used interchangeably where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application.

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0026] Figure 1 This is a schematic diagram illustrating an application scenario of an embodiment of the present invention. The present invention can be applied to diagnostic and treatment assistance systems in various hospitals, remote diagnostic and treatment assistance systems, big data diagnostic and treatment assistance systems, etc. Figure 1 In this process, the server can obtain the electronic health record data of the object to be predicted; calculate the similarity between patients based on the medical history data in the electronic health record data, and construct a patient similarity graph accordingly; construct edge features for the edges in the patient similarity graph and input them into an edge filtering model for filtering to obtain an optimized patient relationship graph; input the optimized patient relationship graph and the corresponding patient node features into a graph convolutional neural network model, and output the multi-label prediction results of the object to be predicted in chronic obstructive pulmonary disease, cardiovascular disease, and diabetes.

[0027] This invention proposes a multi-label chronic disease prediction method, such as... Figure 1 , 2 As shown, the method includes:

[0028] S101: Obtain electronic health record data of multiple subjects to be predicted, and construct corresponding medical history feature representations and patient feature representations based on the electronic health record data.

[0029] Electronic health record data of multiple individuals to be predicted are acquired, the data are categorized by data type, and corresponding medical data features are extracted. The medical history features specifically represent the time-related risk characteristics of the disease evolution status of the individuals to be predicted at different medical treatment times, while the patient features specifically represent the stable risk characteristics of the individual attributes and overall health status of the individuals to be predicted.

[0030] Specifically, based on historical medical record data reflecting the occurrence, development, and changes of the disease over time, a medical history feature representation is constructed. This medical history feature representation is used in subsequent steps to analyze and model the process of disease risk changes over time for the subject to be predicted. After processing, time-related risk features are obtained to characterize the disease evolution state at different follow-up times.

[0031] Specifically, a patient feature representation is constructed based on patient information data and various test data. The test data is screened, filtered, and denoised during the construction process. In at least one embodiment, the screening process is based on medical guidelines and / or clinical expert experience to select test features that are helpful for the diagnosis and risk assessment of chronic diseases. The patient feature representation is used in subsequent steps to analyze and model the overall risk level and individual differences of the subject to be predicted. After processing, stable risk features are obtained to characterize the individual attributes and overall health status of the subject to be predicted.

[0032] In this embodiment of the invention, the electronic health record data is generally referred to as medical data. The medical data originates from the MIMIC-IV database (version 2.0), which contains patient medical records from the Intensive Care Unit of Beth Israel Deaconess Medical Center in Boston, USA, from 2008 to 2019. Patient diagnostic information is encoded using the International Classification of Diseases, Ninth Revision (ICD-9) and Tenth Revision (ICD-10). ICU patient records diagnosed with chronic obstructive pulmonary disease, cardiovascular disease, and diabetes are selected from the database as the target dataset.

[0033] In this embodiment of the invention, the medical data includes at least medical history data and laboratory test data. The medical history data is used to characterize the patient's past or current medical history, preferably represented in discrete feature form, for subsequent construction of a patient relationship graph based on medical history similarity. The laboratory test data is used to characterize the patient's test results during hospitalization, preferably in the form of continuous clinical features, for subsequent graph structure optimization and multi-label disease prediction.

[0034] In some embodiments, the medical historical data and the laboratory test data are feature-filtered based on evidence-based medicine principles to form a unified feature set. After data acquisition, the medical data can be preprocessed, including but not limited to deleting missing value samples and adjusting label distribution, to ensure the stability of subsequent model training and prediction.

[0035] like Figure 3 As shown, the medical history data of each patient is vectorized to obtain the first... The patient's medical history characteristics are represented as a binary vector (0 or 1), specifically as follows:

[0036] ,

[0037] in, This indicates whether the patient has a past or present medical history as described in item (i). , 𝑘 ∈[1,m].

[0038] The aforementioned medical history characteristics can stably reflect the patient's long-term comorbidity and are suitable as a priori basis for constructing patient relationships.

[0039] Through step 101, the present invention completes the acquisition and data type classification of the medical data of the object to be predicted, providing basic data support for subsequent patient graph structure construction and prediction model calculation.

[0040] In this embodiment of the invention, in order to characterize the co-occurrence effect of chronic diseases from the patient's perspective and to explore the potential associations between patients, this embodiment of the invention quantifies the similarity between patient nodes based on the medical historical data obtained in step 101, and constructs a patient graph structure accordingly.

[0041] S102: Calculate the medical history similarity S between any two objects to be predicted based on the medical history feature representation, and construct a patient similarity graph G based on the medical history similarity S, G=(V,E), where node V corresponds to the object to be predicted, and edge E represents the medical history association between the objects to be predicted.

[0042] First, calculate the similarity S of medical histories between any two patients (patient X and patient Y). ij This forms a similarity matrix, i.e. Used to measure the similarity in the composition of medical history among different patients.

[0043] In the context of chronic disease diagnosis and risk assessment, patient medical history is typically coded and recorded according to a medical disease classification system. Different diseases exhibit a clear hierarchical structure and co-occurrence relationships. Clinical diagnosis focuses more on the presence and combination patterns of diseases rather than continuous numerical differences. Therefore, from the perspective of medical cognition and clinical practice, similarity modeling based on the "co-occurrence" of medical history items is more in line with the logical characteristics of disease diagnosis.

[0044] On the other hand, from the perspective of data characteristics, medical history features are usually represented in a high-dimensional, discrete, and sparse binary form, with a large number of common missing terms among different patients. If a similarity measure based on vector distance is directly used, it is easily affected by the common missing terms, thus introducing similarity relationships that have no medical significance.

[0045] Based on the dual considerations of the above-mentioned medical diagnostic logic and data distribution characteristics, this invention uses Jaccard similarity to model the similarity of patients' medical history features. By characterizing the co-occurrence ratio of medical history items, it effectively reflects the degree of similarity of patients at the disease composition level.

[0046] The similarity S ij Jaccard similarity can be used, and its calculation formula is as follows:

[0047] ,

[0048] in, This indicates the number of common occurrences of certain medical history between the 𝑖-th patient and the 𝑗-th patient. This represents the number of items appearing in a certain medical history for the nth patient and the xith patient.

[0049] By adopting the above similarity modeling method, the association between patient nodes with similar disease structural features can be highlighted in the patient similarity graph, while reducing the interference caused by missing or incomplete medical history records. This results in the construction of a structurally stable and semantically consistent patient relationship graph, providing a reliable foundation for subsequent feature propagation and disease risk prediction based on graph structure.

[0050] It should be noted that, without departing from the concept of this invention, the similarity can also be measured using cosine similarity, Dice coefficient, or similarity based on weighted medical history items to adapt to different medical history coding systems and sparsity conditions. However, given that the medical history data is usually binary sparsely distributed, using Jaccard similarity can more reasonably characterize the co-occurrence intensity of medical history, reduce the interference caused by common missing items, and thus improve the stability and interpretability of similarity measurement.

[0051] Secondly, based on the similarity of medical history S ijA similarity threshold T is set for filtering, and an initial patient similarity graph is constructed.

[0052] like Figure 4 As shown, in this embodiment of the invention, a similarity threshold T is set to sparsify the similarity matrix S, thereby reducing the number of similarity matrices that satisfy S. ij <T entries are set to zero; further, the maximum number of neighbors for each patient node is set to Only retain each patient with the highest similarity to their previous patient. The edges corresponding to each neighbor are used to construct an initial patient similarity graph, namely G=(V,E), where V is the set of vertices, representing the patient nodes corresponding to the object to be predicted; and E is the set of edges, representing the similarity relationship edges established between patient nodes based on similarity.

[0053] The sparsification process can be expressed as:

[0054] ,

[0055] in, Let T represent the similarity matrix after sparse processing, and let T represent the similarity threshold.

[0056] In the patient similarity graph structure G, for each node (the i-th node), select... The corresponding neighbor set N K (i), thus obtaining the edge set E, E={(i,j)|j∈N K (i)}.

[0057] By using the threshold constraints and neighbor number constraints mentioned above, the embodiments of the present invention can suppress weak associations between patients with low similarity, avoid noise propagation and computational burden caused by excessively dense connections, and at the same time ensure the sparsity and representativeness of the patient graph structure, so that subsequent graph convolution propagation is mainly carried out in the patient group with "similar co-occurrence of medical history", thereby providing a more reliable prior structure for graph structure optimization and prediction.

[0058] In this embodiment of the invention, although the initial patient medical history similarity graph structure is constructed from medical historical data and has a clear co-occurrence prior, there may still be erroneous connections introduced by sparsity of medical history, differences in coding granularity, or accidental co-occurrence. Therefore, this embodiment of the invention further introduces full data to optimize the edge filtering of the patient similarity graph.

[0059] S103: Filter and screen the patient similarity graph G using the full data of the electronic health record data of the subject to be tested, and obtain the optimized patient similarity graph G', G'=(V,E'). The full data includes medical history data and various test data.

[0060] The patient's electronic health record data includes a variety of examination and testing information, such as medical history data and test data; among which, the test data are usually continuous clinical values, which can supplement and reflect the real-time physiological status during hospitalization.

[0061] By using the method of "first mapping with medical history and then correcting with full features", this invention not only preserves the long-term prior of the comorbidity structure of chronic diseases, but also uses the detection data to verify the local edge relationships, thereby reducing the interference of noisy edges on downstream predictions.

[0062] like Figure 4 As shown, in some embodiments, the full features of patient X's electronic health record data are represented as x. i , This includes binary characteristics of medical history and continuous characteristics of detection; the continuous characteristics of detection may consist of statistical quantities of laboratory indicators during hospitalization or effective detection items after screening.

[0063] To distinguish candidate edges (the connection between patient X and patient Y), embodiments of the present invention construct edge feature vectors. The edge feature vector is then input into an edge classifier for evaluation. It can be obtained in any of the following ways:

[0064] ,

[0065] Where, x i x represents the full characteristics of the electronic health record data of the nth patient. j represents the full characteristics of the electronic health record data of the nth patient, and ⊙ represents element-wise multiplication.

[0066] Subsequently, the edge retention probability or edge class label is output by the edge classifier f(.):

[0067] p ij =f(y ij ),

[0068] Where, p ij p represents the probability of retaining the edge output by the edge classifier. ij ∈ (0,1), This represents the edge feature vector.

[0069] The edge classifier is a multilayer perceptron (MLP), and the output can be mapped to a probability value using the sigmoid function.

[0070] Furthermore, in this embodiment of the invention, an edge filtering threshold is set, when p ij Delete edges when < 1 When p ijPreserve edges when ≥ 1 Thus, the optimized edge set E' is obtained, forming the optimized patient similarity graph G', G'=(V,E').

[0071] By using edge filtering based on full features, this invention can remove weakly related edges in the patient graph that are inconsistent with the features of the detection data, reduce the negative impact of irrelevant neighbors in the graph convolution aggregation process, and avoid noise being amplified in multi-layer propagation. At the same time, the optimized patient similarity graph G' is closer to the relationship between patients with similar clinical conditions in reality, so that the subsequent graph neural network can perform representation learning on a relatively reliable graph topology, thereby obtaining a more stable multi-label prediction output.

[0072] S104. Based on the optimized patient similarity graph G', input the patient feature representation into the graph neural network model and output the multi-label chronic disease prediction result of the object to be predicted.

[0073] In embodiments of the present invention, such as Figure 5 As shown, the optimized patient similarity graph G' is compared with the full patient feature matrix. A common input graph convolutional network (GCN) is used for multi-label chronic disease prediction.

[0074] Let A be the adjacency matrix of the graph structure G′, and I be the identity matrix. In this embodiment of the invention, the adjacency matrix is ​​subjected to self-loop addition and normalization to obtain a normalized adjacency matrix. The specific technical formula is as follows:

[0075] ,

[0076] Where D is the degree matrix of (A+I).

[0077] Normalization can suppress the differences in aggregation scale caused by nodes of different degrees, avoid the disproportionate impact of highly connected nodes on the propagation results, and thus improve training stability.

[0078] Graph convolutional networks learn patient representations using a layer-by-layer neighborhood aggregation approach, and their propagation computation can be represented as:

[0079] ,

[0080] in, For the first Layer node representation, For the first The trainable parameter matrix of the layer, This is the activation function.

[0081] In some embodiments, two or more layers of graph convolution can be set to take into account both local neighborhood information and higher-order relational information; and Dropout or regularization terms can be introduced between layers to reduce the risk of overfitting.

[0082] The patient, Gao Wei, ultimately stated... The output layer is mapped to multi-label prediction probabilities. :

[0083]

[0084] in, , To predict the number of disease labels, This represents the output layer weight parameter matrix that maps the high-dimensional representation of the patient to the disease prediction space. This represents the bias parameter vector corresponding to the output layer. In this embodiment of the invention, the disease labels include three categories: chronic obstructive pulmonary disease, cardiovascular disease, and diabetes, and comorbidities are allowed; therefore, a multi-label output format is adopted.

[0085] In some embodiments, the present invention employs binary cross-entropy loss to train the model:

[0086]

[0087] in, For real labels, , To predict probabilities, Indicates the number of training samples. This indicates the total number of predicted disease labels.

[0088] It should be noted that, in the data preprocessing stage, the present invention can remove data containing missing values ​​to reduce the bias introduced by imputation; at the same time, it can use sample removal to balance the multi-label distribution to reduce the bias caused by class imbalance to training and improve generalization ability.

[0089] Through the above steps, this invention establishes a unified framework of "patient graph prior construction based on medical history similarity + edge filtering optimization based on full data + graph convolution multi-label prediction based on optimized graph structure". This framework can effectively integrate co-occurrence associations among patients and individual detection features under clinical data conditions where comorbidities are common and data noise is significant, to achieve joint prediction of the risk of multiple chronic diseases, and provide usable decision support outputs for clinical screening and risk stratification.

[0090] This invention also provides a multi-label chronic disease prediction device, which is used to implement the multi-label chronic disease prediction method as described in the above embodiments. Figure 6As shown, the multi-label chronic disease prediction device 200 includes a data acquisition module 201, a medical history data graph structure construction module 202, a medical history and test data optimized graph structure module 203, and a graph convolutional neural network prediction module 204.

[0091] The data acquisition module 201 is used to acquire electronic health record data of the object to be predicted; the electronic health record data includes at least medical history data and test data, wherein the medical history data is used to characterize the past or current disease history information of the object to be predicted, and the test data is used to characterize the test results and related clinical measurement information of the object to be predicted during hospitalization or medical treatment; the data acquisition module 201 is also used to perform structured processing and sample screening on the electronic health record data to obtain input data that can be used for mapping and prediction.

[0092] The medical history data graph structure construction module 202 is used to calculate the similarity between patients pairwise based on medical history data, and construct the initial graph structure of the patient graph accordingly. In some embodiments, the medical history data graph structure construction module 202 can calculate the similarity matrix based on the co-occurrence relationship of medical history, and perform sparsification processing on the similarity matrix through similarity threshold and neighbor number constraint, thereby obtaining the initial patient graph structure G=(V,E), where V is the set of patient nodes and E is the set of patient edges.

[0093] The medical history and test data optimization graph structure module 203 is used to further optimize the graph structure by integrating medical history data and test data based on the initial patient graph structure. In some embodiments, the medical history and test data optimization graph structure module 203 can construct edge features for candidate edges in the graph and filter or weight candidate edges through an edge classifier, thereby suppressing noise propagation introduced by low-confidence connections and obtaining an optimized patient graph structure. Figure 6 As shown, the medical history data graph structure construction module 202 and the medical history and test data graph structure optimization module 203 can be executed in parallel or serially, and the optimized patient graph structure can be used as input for subsequent graph convolution prediction.

[0094] The graph convolutional neural network prediction module 204 is used to input the optimized patient graph structure and the node features corresponding to the patient nodes into the graph convolutional neural network model for representation learning and information propagation, and output the prediction results of the subject to be predicted on multiple chronic disease labels; in some embodiments, the prediction results are multi-label prediction results, which can be used to characterize the risk or diagnostic tendency of the subject to be predicted to have one or more diseases such as chronic obstructive pulmonary disease, cardiovascular disease, and diabetes.

[0095] The multi-label chronic disease prediction device 200 can be integrated into a computer device, which can be a terminal or a server. The terminal can be a smartphone, tablet, laptop, personal computer, or other terminal device. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services. The terminal and the server can be directly or indirectly connected through wired or wireless communication, which is not limited herein.

[0096] This invention also provides a multi-label chronic disease prediction system 300, with reference to... Figure 7 As shown, this system is used to implement the multi-label chronic disease prediction method as described in the above embodiments. Figure 6 As shown, the multi-label chronic disease prediction system 300 includes at least one processor 301, at least one memory 302, and a communication bus 303; the communication bus 303 is used to realize the connection and communication between the processor 301 and the memory 302.

[0097] The memory 302 is used to store program instructions and data; the processor 301 is used to call and execute the program instructions stored in the memory 302 to implement the steps corresponding to the above-mentioned multi-label chronic disease prediction method, including acquiring and processing electronic health record data, constructing an initial patient graph structure based on medical history data, optimizing the graph structure by fusing medical history data and detection data, and outputting multi-label chronic disease prediction results based on graph convolutional neural network.

[0098] Those skilled in the art will understand that Figure 7 The multi-label chronic disease prediction system 300 shown does not constitute a limitation on the system. It may include more or fewer components than shown, or combine certain components or use different component arrangements. For example, the system 300 may also include network interfaces, input / output interfaces, external storage devices, etc., to complete functions such as data access, model deployment and result output.

[0099] In one embodiment of the present invention, the sample data used in the present invention comprises a total of 8645 records, each record corresponding to a patient's hospitalization feature vector; the features consist of a medical history data field and laboratory test, vital sign, and other detection data fields, wherein the medical history data field is a binary discrete feature, and the detection data field is a continuous clinical feature. The continuous clinical features undergo missing value removal and numerical normalization processing before being input into the model; the prediction output is a three-column multi-label result, corresponding to chronic obstructive pulmonary disease, cardiovascular disease, and diabetes, respectively.

[0100] In one embodiment of the present invention, the graph convolutional neural network adopts a three-layer graph convolutional structure with a hidden layer dimension of 128 and a Dropout value of 0.5. The optimization algorithm used is Adam, with a learning rate of 0.005, a weight decay of 5 × 10^-4, and a maximum training epoch of 500 epochs. The dataset is divided into a training set, a validation set, and a test set, with the test set accounting for 20% and the remaining samples allocated as a validation set for model parameter selection and performance evaluation.

[0101] The following are the specific experimental results, comparing the evaluation parameters of this invention with other algorithms and their combinations. As shown in Table 1:

[0102] Table 1. Accuracy comparison of the six time accumulation operators designed in this invention on the longitudinal multi-task model.

[0103] algorithm accuracy Accuracy Recall rate F1 score Macro average F1 Micro-average F1 CC 0.7848 0.9412 0.9166 0.9274 0.9248 0.9294 BR 0.8224 0.9482 0.9397 0.9436 0.9400 0.9444 RAkEL-SVM 0.6680 0.8977 0.8736 0.8792 0.8812 0.8804 CLR 0.7854 0.9445 0.9153 0.9284 0.9260 0.9303 COCOA 0.7739 0.9176 0.9282 0.9228 0.9228 0.9267 GCLE 0.7976 0.9381 0.9220 0.9332 0.9308 0.9324 MLP 0.8207 0.9460 0.9391 0.9423 0.9385 0.9433 MLLEM 0.7283 0.9295 0.8894 0.9060 0.9067 0.9080 XGBoost 0.7293 0.9295 0.8894 0.9060 0.9067 0.9080 DeepFM 0.7929 0.9389 0.9102 0.9243 0.9217 0.9252 SNNs 0.8282 0.9528 0.9306 0.9417 0.9329 0.9432 Transformer 0.8246 0.9529 0.9291 0.9408 0.9287 0.9399 CMLP 0.8207 0.9474 0.9350 0.9413 0.9413 0.9474 This invention 0.8456 0.9585 0.9414 0.9493 0.9458 0.9504

[0104] The algorithms listed in Table 1 include: RAKEL-SVM (a multi-label classification method based on a random k-label set using support vector machines), CLR (a classifier chain method), COCOA (a collaborative optimization multi-label classification method), GCLE (a graph-constrained label embedding method), MLLEM (a multi-label logical embedding model), SNNs (a Siamese neural network model), and CMLP (a conditional multilayer perceptron model). In Table 1, the macro-average F1 score refers to the average of the F1 scores calculated for each label, used to measure the overall discriminative balance of the model across different disease labels; the micro-average F1 score refers to the uniform calculation of the F1 score across all label samples, focusing more on reflecting the overall predictive performance of the model.

[0105] For the multi-label chronic disease collaborative prediction dataset constructed in this invention, the prediction model based on historical similarity edge filtering and graph convolution proposed in this invention achieves an accuracy of 84.56% on the test set, which is the best performance among all compared models, followed by SNNs models. Compared with the suboptimal model, this invention achieves better results in terms of accuracy, precision, recall, F1 score, macro-average F1, and micro-average F1. Specifically, the accuracy is improved by 1.74 percentage points, the precision by 0.56 percentage points, the F1 score by 0.57 percentage points, the macro-average F1 by 0.45 percentage points, and the micro-average F1 by 0.30 percentage points, indicating that this invention has higher overall discriminative ability and multi-label consistency in multi-label diagnostic prediction tasks.

[0106] Ablation experiments verified:

[0107] This ablation experiment was used to verify the impact of different similarity measurement methods on the quality of the graph structure and subsequent prediction performance during the construction of a patient similarity graph based on patient medical history characteristics. The specific results are shown in Table 2.

[0108] Table 2 Comparison of Ablation Experiment Results

[0109] method accuracy Accuracy Recall rate F1 score Macro average F1 Micro-average F1 Jaccard (Historical Data) 0.8485 0.8524 0.8485 0.8475 0.8440 0.8485 Euclidean (All Features) 0.8213 0.8284 0.8213 0.8212 0.8160 0.8213 Cosine (All Features) 0.8323 0.8425 0.8323 0.8325 0.8206 0.8323 Mixed (Jaccard + Euclidean) 0.8230 0.8285 0.8230 0.8215 0.8149 0.8230 Mixed (Jaccard + Cosine) 0.7085 0.7151 0.7085 0.7060 0.6870 0.7085

[0110] In Table 2, Jaccard (Historical Data) is a Jaccard similarity graph based on historical co-occurrence, which is an important technical means of the graphing method used in this invention; Euclidean (All Features) is a Euclidean distance graph based on all features; Cosine (All Features) is a cosine similarity graph based on all features; Mixed (Jaccard + Euclidean) and Mixed (Jaccard + Cosine) are also mixed similarities. In the ablation experiments, the evaluation metrics include Validation Accuracy, Test Accuracy, Precision, Recall, F1, Macro F1, Micro F1, and Weighted F1, used to comprehensively measure the method's effectiveness from both overall performance and category perspectives.

[0111] As shown in Table 2, Jaccard (Historical Data) achieved the best performance across all evaluation metrics. In patient association modeling based primarily on medical history, Jaccard similarity, based on the co-occurrence ratio of medical history items, can more effectively construct patient relationships, thus providing a higher-quality graph structure foundation for subsequent feature propagation and prediction in graph models.

[0112] In real-world medical scenarios, the data distribution and feature fields differ among different medical institutions, causing the optimal configuration of traditional machine learning models or different network structures to fluctuate with data changes. Based on experimental results, this invention, by introducing a similarity graph based on historical medical data and performing edge filtering, effectively reduces the interference of noisy connections on graph propagation. This allows the model to more fully utilize stable comorbidity associations among patients, resulting in more reliable comprehensive performance in multi-label predictions of chronic obstructive pulmonary disease, cardiovascular disease, and diabetes. In summary, this invention outperforms traditional single-model methods and methods without historical similarity edge filtering mechanisms in terms of prediction accuracy and the ability to represent complex comorbidity relationships, verifying the effectiveness and superiority of the proposed solution.

[0113] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include ROM, RAM, disk, or optical disk, etc.

[0114] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-label chronic disease prediction method, characterized in that, include: Obtain electronic health record data of multiple subjects to be predicted, and construct corresponding medical history feature representations and patient feature representations based on the electronic health record data; The medical history similarity S between any two objects to be predicted is calculated based on the medical history feature representation. Based on the medical history similarity S, a patient similarity graph G is constructed, where node V represents the object to be predicted and edge E represents the medical history association between any two objects to be predicted. The patient similarity graph G is filtered and screened using the full data of the electronic health record data of the object to be predicted, resulting in an optimized patient similarity graph G'. The full data includes medical history data and various test data. Based on the optimized patient similarity graph G', the patient feature representation is input into the graph neural network model, and the multi-label chronic disease prediction result of the object to be predicted is output.

2. The multi-label chronic disease prediction method based on patient similarity graphs according to claim 1, characterized in that, The medical history similarity S between any two individuals to be predicted is calculated based on the aforementioned medical history characteristics. Specifically, the similarity S is calculated as follows: ， in, Indicates two objects to be predicted ( , The similarity of medical history between the two groups Indicates the first The medical history characteristics of the individuals to be predicted are represented. Indicates the first The medical history characteristics of the individuals to be predicted are represented. This indicates finding the intersection. This indicates finding the union of sets.

3. The multi-label chronic disease prediction method according to claim 1 or 2, characterized in that, Based on the aforementioned medical history similarity S, a patient similarity graph G is constructed. The specific process includes: For any two objects to be predicted, the similarity S of their medical history is filtered by a threshold, and candidate edges with a similarity greater than or equal to the preset similarity threshold are retained. For each node corresponding to the object to be predicted, the edges corresponding to the top K neighboring nodes are selected from the candidate edges according to the medical history similarity S from high to low as retained edges, thereby constructing a sparse patient similarity graph, where K is a positive integer.

4. The multi-label chronic disease prediction method according to claim 1, characterized in that, The process of filtering and screening the patient similarity graph G using the full dataset of the electronic health records of the individuals to be predicted includes: For any edge in the patient similarity graph G, obtain the patient feature representations corresponding to the nodes at both ends of the edge, and generate edge features based on the patient feature representations of the nodes at both ends; The edge features are input into the edge classification model, and the classification result or the corresponding confidence score of the edge is output as either a positive edge or a negative edge. The edges are retained or removed based on the classification results or confidence levels. The edges to be removed are removed from the patient similarity graph to obtain the optimized patient relationship graph G'.

5. The multi-label chronic disease prediction method according to claim 4, characterized in that, The edge features are generated by the patient feature representations of the two end nodes through at least one feature combination method, which includes feature concatenation, feature difference, feature product or feature summation.

6. The multi-label chronic disease prediction method according to claim 1, characterized in that, Based on the optimized patient relationship graph G', the patient feature representation is input into the graph neural network model, and the processing includes: Based on the optimized patient relationship graph G', neighborhood information aggregation is performed on the patient feature representation corresponding to each node to obtain an updated feature representation containing information about neighboring patients. The neighborhood information aggregation and feature update process can be represented as follows: ， in, This represents the feature representation matrix corresponding to each node in the l-th layer of the graph neural network. Let represent the updated feature representation matrix obtained at layer l+1 after aggregating neighborhood information. This represents the adjacency matrix or its normalized form constructed from the optimized patient relationship graph G'. This represents the trainable parameter matrix of the l-th layer of the graph neural network. ⋅ The activation function is represented by the feature update process, which is used to fuse the feature information of neighboring patient nodes to obtain a patient feature representation that includes neighborhood structure information.

7. The multi-label chronic disease prediction method according to claim 6, characterized in that, The graph neural network model includes at least one graph convolutional layer or graph attention layer, which is used to integrate the features of the node itself with the features of neighboring nodes to generate node representations for disease prediction.

8. The multi-label chronic disease prediction method according to claim 1, characterized in that, The edge classification model and the graph neural network model constitute a unified chronic disease prediction model. The chronic disease prediction model is trained based on supervised learning. During the training process, the multi-label prediction loss function is minimized according to the real disease labels, and regularization constraints are applied to the parameters of the multi-label chronic disease prediction model.

9. A multi-label chronic disease prediction device, the device being used to implement the multi-label chronic disease prediction method as described in claim 1, characterized in that, The device includes: The data acquisition module is used to acquire electronic health record data of multiple subjects to be predicted, and to construct medical history feature representations and patient feature representations; The mapping module is used to calculate the medical history similarity between the objects to be predicted based on the medical history feature representation, and to construct a patient similarity relationship graph; The edge filtering module is used to construct edge features from the edges in the patient similarity graph and filter them based on the edge classification model to obtain an optimized patient relationship graph. The prediction module is used to input the optimized patient relationship graph and the corresponding patient feature representation into the graph neural network model and output the multi-label chronic disease prediction results.

10. A multi-label chronic disease prediction system, characterized in that, The system includes a processor, a communication interface, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the multi-label chronic disease prediction method as described in any one of claims 1 to 8.

Images