Patient-oriented multi-chronic disease prediction method and device based on hierarchical clustering coding

Abstract

The invention discloses a patient-oriented multi-chronic disease prediction method and device based on hierarchical clustering coding. The method comprises the following steps: constructing a double-tower neural network model comprising a user expert tower for patient information modeling and a chronic disease expert tower for multi-task prediction, and carrying out multi-chronic disease combined prediction training; using the trained user expert tower to encode the patient information into hierarchical embedded representation, and performing hierarchical clustering to obtain hierarchical clustering embedded representation of the patient; and inputting the hierarchical clustering embedding representation of the patient into a task weighting network of a multi-task prediction model so as to carry out weighting adjustment on the multi-task prediction model and then carry out multi-chronic disease prediction. According to the method, the constitutive property and interpretability of patient characterization can be improved, the dynamic adaptive capacity of patient embedding is enhanced, the personalized modeling capacity in multi-task prediction is improved, the overall performance of a multi-task model is optimized, and the method is suitable for various medical health intelligent prediction scenes and has important theoretical value and practical application prospects.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of intelligent medical treatment, and particularly relates to a patient-oriented multi-chronic disease prediction method and device based on hierarchical clustering coding. BACKGROUND

[0002] With the improvement of computing power, the accumulation of large-scale data, and the continuous innovation of algorithms, deep learning technology has made a great leap, and is widely used in image recognition, speech recognition, natural language processing, recommendation system and other fields. Deep learning automatically extracts features through multi-layer neural networks, gradually learns the potential structure of data from low-level to high-level semantic representation, and has obvious advantages in feature extraction, generalization ability and other aspects compared with traditional machine learning methods.

[0003] Among various models of deep learning, the encoder, as one of the core structures of representation learning, plays an important role. According to whether it depends on label information, the encoder can be roughly divided into three categories: supervised, unsupervised and self-supervised. The supervised encoder relies on manually annotated data, and guides the model to learn the potential representation through the loss function of a specific task (such as classification, regression), so that the encoder output has strong discriminability. This method has fast convergence speed, clear goal, and can obtain superior performance in the scene with sufficient high-quality annotation, but it is strongly dependent on labels, and its generalization and migration ability is limited, especially in the cross-task or cross-domain scene, there is a problem of insufficient applicability. The unsupervised encoder gets rid of the dependence on labels, and usually takes reconstruction of input as the training target, such as autoencoder or variational autoencoder, etc., and learns the potential representation by minimizing the error between the input and the reconstruction. This method is suitable for dimensionality reduction, denoising, data compression and other tasks, but its features lack task-driven nature, and the semantic expression ability is weak, which is difficult to be directly applied to complex discrimination tasks. To make up for the shortcomings of the above two, the self-supervised encoder becomes a representation learning method that combines the flexibility of unsupervised and the effect of supervised. Its core idea is to automatically construct supervision signals from unlabeled data by designing "pre-training tasks" such as masking prediction, context modeling, contrastive learning, etc., to promote the encoder to learn a general representation with transferability.

[0004] Despite the wide applicability and strong modeling power of supervised, unsupervised, and self-supervised encoders in various tasks and data conditions, they still face many challenges and limitations in practical applications. First, these encoders mostly rely on dense representations in high-dimensional continuous vector spaces, and their training objectives often only focus on single-task loss or reconstruction accuracy, lacking the ability to model hierarchical structures or group attributes between samples. For example, in medical, recommendation system, or user profiling fields, different samples may have obvious categories, subgroups, or implicit structures (such as age groups, behavior patterns), but traditional encoders fail to explicitly model these potential hierarchical relationships, resulting in generated feature vectors that cannot reflect the heterogeneity in the data, limiting the scalability and generalization ability of the model.

[0005] In addition, such encoders generally have poor interpretability. Since the model learns features in high-dimensional space through nonlinear neural networks, its internal representation often lacks explicit semantics or human-understandable structure, which is not conducive to model diagnosis, debugging, and application in safety-sensitive scenarios (such as medical diagnosis or financial risk control). Especially in self-supervised and unsupervised methods, encoder training does not rely on label signals, making it more difficult to establish a direct correspondence between the output representation and the specific task goal or human intuition. Even under supervised training, the learned representation, although optimal for the task, is essentially a "black box" mapping process that lacks explicit feature selection or hierarchical logic. Therefore, in the context of increasing demand for interpretability and structured modeling, traditional encoders face the urgent need to evolve towards structure-aware, causally-guided, and hierarchical modeling, becoming one of the frontiers of current research.

[0006] To break through the limitations of single-task modeling, the field of deep learning has further developed multi-task learning, which uses the collaborative training of multiple related tasks to improve model generalization ability, aiming to promote mutual learning between multiple tasks through shared representations, avoid bias from overfitting a single task, and improve overall performance. Typical multi-task learning frameworks include hard parameter sharing, where multiple tasks share a portion of the neural network structure (such as a shared encoder), with different task heads branching out at the output layer; and soft parameter sharing, where each task is assigned an independent model, but guided to be close in the hidden space through regularization terms or attention mechanisms. In addition, there are task relationship learning based on graph structure modeling, dynamic task weighting, and meta-learning-based enhanced multi-task learning strategies.

[0007] Although multi-task learning has many advantages, it also faces a series of challenges. First, task conflict is the most common problem. Different tasks may have inconsistent directions during optimization, leading to mutual interference in shared parameter updates, even negative transfer, and thus performance degradation of some tasks. Second, task weight setting is difficult. Different tasks have different loss scales, learning difficulties, and training speeds, so a careful task weighting strategy or dynamic scheduling mechanism needs to be designed. Third, the shared structure design is complex. Hard sharing may ignore the individual differences between tasks, while soft sharing significantly increases the computational and parameter overhead. In addition, multi-task learning models have poor interpretability, making it difficult to understand the specific dependency and coordination mechanism between tasks, especially when there is no clear prior task relationship. In summary, multi-task learning has significant potential in improving model efficiency and enhancing generalization ability, but the difficulties in task coordination and structure design also prompt researchers to continuously explore more intelligent task modeling and optimization strategies, such as adversarial learning, gradient projection, and multi-task attention mechanism, which have become one of the important research hotspots in current deep learning. SUMMARY

[0008] In view of the above, the purpose of the present application is to provide a patient-oriented multi-chronic disease prediction method and device based on hierarchical clustering coding, which is based on a patient-oriented, multi-layer abstract, quantifiable, and trainable representation learning mechanism, and effectively integrates it into a multi-task learning framework, which can balance the structure of the representation and the performance of the prediction, improve the structure and interpretability of patient representation, enhance the dynamic adaptation ability of patient embedding, improve the individual modeling ability in multi-task prediction, and optimize the overall performance of the multi-task model, suitable for various medical and health intelligent prediction scenarios, and has important theoretical value and practical application prospects.

[0009] To achieve the above-mentioned purpose of the application, the technical solutions provided by the present application are as follows: In a first aspect, the present application provides a patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding, comprising the following steps: A double-tower neural network model including a user expert tower for patient information modeling and a chronic disease expert tower for multi-task prediction is constructed, and multi-chronic disease joint prediction training is performed; The trained user expert tower is used to encode patient information into hierarchical embedding representation, and all patients' sub-vectors at the same level are clustered to obtain clustering centers at each level. Each patient's sub-vector at each level in the hierarchical embedding representation is mapped to the corresponding clustering center number to obtain a patient hierarchical clustering embedding representation; The patient hierarchical clustering embedding representation is input into the task weighting network of the multi-task prediction model to perform task-by-task weighting adjustment on the intermediate representation output by the multi-task prediction model, and the multi-task prediction model is used for multi-chronic disease prediction.

[0010] Preferably, the construction includes a double tower neural network model of a user expert tower for patient information modeling and a chronic disease expert tower for multi-task prediction, and multi-chronic disease joint prediction training is performed, including: A user expert tower for patient information modeling is constructed, which receives user feature input and extracts patient vector representation through an encoder; A chronic disease expert tower for multi-task prediction is constructed, which receives chronic disease feature input and extracts chronic disease vector representation through an encoder; The patient vector representation output by the user expert tower is segmented according to a preset dimension segmentation manner to obtain a patient hierarchical vector representation composed of multiple sub-vectors at different levels, and multi-chronic disease joint prediction training is performed based on the patient hierarchical vector representation and the chronic disease vector representation.

[0011] Preferably, the patient information is encoded into hierarchical embedding representation using the trained user expert tower, and the sub-vectors at the same level of all patients are clustered to obtain clustering centers at each level, including: The hierarchical embedding representation of all patients is extracted using the trained user expert tower, the sub-vectors of each patient at the same level are extracted to obtain a sub-vector set, and then the sub-vector set at each level is clustered to obtain a plurality of clustering centers corresponding to each level.

[0012] Preferably, the sub-vector at each level of the hierarchical embedding representation of each patient is mapped to the corresponding clustering center number to obtain a patient hierarchical clustering embedding representation, including: The sub-vector at each level of the hierarchical embedding representation of the patient to be processed is matched with the clustering center at the corresponding level, and the sub-vector at each level is mapped to the unique identification number of the matching clustering center, thereby forming a patient hierarchical clustering embedding representation composed of clustering center numbers at each level.

[0013] Preferably, the clustering center number at a high level represents a coarse-grained global health status feature, and the clustering center number at a low level represents a fine-grained local health signal feature.

[0014] Preferably, the patient hierarchical clustering embedding representation is input into a task weighting network of the multi-task prediction model to perform task-by-task weighting adjustment on the intermediate representation output by the multi-task prediction model, including: The patient hierarchical clustering embedding representation is provided as an additional input feature to the task weighting network; A set of task-specific weight vectors is dynamically generated based on the patient hierarchical clustering embedding representation through the task weighting network; The task-specific weight vectors are applied to the intermediate representation output by the multi-task prediction model for task-by-task weighting adjustment, and the weights of each sub-expert network in the multi-task prediction model are adjusted.

[0015] Preferably, the multi-task prediction model comprises a mixture-of-experts model MOE, a multi-gate mixture-of-experts model MMOE, or a progressive multi-task learning model PLE.

[0016] In a second aspect, the embodiments of the present application further provide a patient-oriented multi-chronic disease prediction device based on hierarchical clustering coding, which is implemented by using the patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding described above, and comprises a double-tower hierarchical representation training module, a hierarchical clustering coding module, and a weight adjustment multi-task prediction module. The double-tower hierarchical representation training module is configured to construct a double-tower neural network model comprising a user expert tower for modeling patient information and a chronic disease expert tower for multi-task prediction, and perform multi-chronic disease joint prediction training. The hierarchical clustering coding module is configured to encode patient information into hierarchical embedding representation by using the trained user expert tower, cluster sub-vectors at the same level for all patients to obtain clustering centers of each level, and map each level sub-vector in the hierarchical embedding representation of each patient to the corresponding clustering center number to obtain patient hierarchical clustering embedding representation. The weight adjustment multi-task prediction module is configured to input the patient hierarchical clustering embedding representation into a task weighting network of a multi-task prediction model, to perform task-by-task weighting adjustment on the intermediate representation output by the multi-task prediction model, and to perform multi-chronic disease prediction by using the multi-task prediction model after the weighting adjustment.

[0017] In a third aspect, the embodiments of the present application further provide an electronic device comprising a memory and one or more processors, wherein the memory is configured to store a computer program, and the processor is configured to implement the patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding described above when the computer program is executed.

[0018] In a fourth aspect, the embodiments of the present application further provide a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program is configured to implement the patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding described above when the computer program is executed by a computer.

[0019] Compared with the prior art, the present application has at least the following beneficial effects: (1) Improve the structure and interpretability of patient representation: the present application divides the deep representation vector of the patient into multiple levels, and combines the clustering coding means to construct a hierarchical quantization representation with semantic hierarchy, so that the model has the ability to express from coarse to fine when abstracting the health status of the patient, which is helpful for subsequent diagnosis and model interpretability analysis.

[0020] (2) Enhance the dynamic adaptability of patient embedding: the encoding method based on hierarchical clustering proposed in the present application no longer takes the patient embedding representation as a static input, but trains a learnable quantization representation that can be updated synchronously during multi-task training, improving the adaptability of the representation to the task target and avoiding the problem of inconsistency between the sample space and the task space.

[0021] (3) Improve the individual modeling capability in multi-task prediction: the present application inputs the patient hierarchical clustering embedding representation into the task weighting network of the multi-task prediction model to dynamically generate task weights, thereby realizing individualized prediction adjustment strategies, improving the response capability of the model to patient heterogeneity, and effectively alleviating the task conflict and negative transfer problem.

[0022] (4) Optimize the overall performance of the multi-task model: the hierarchical quantization mechanism based on hierarchical clustering encoding introduced in the present application not only improves the representation capability and interpretability, but also realizes effective collaboration between multiple chronic disease prediction tasks through the structured encoding and feature-guided weighting mechanism, thereby improving the overall generalization capability and prediction accuracy of the model. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0024] Figure 1 is a flowchart of the patient-oriented multi-chronic disease prediction method based on hierarchical clustering encoding provided by the embodiment of the present application; Figure 2 is a structural diagram of a double-tower neural network model provided by the embodiment of the present application; Figure 3 is a flowchart of hierarchical clustering encoding provided by the embodiment of the present application; Figure 4 is a mapping relationship diagram of user encoding and user features provided by the embodiment of the present application; Figure 5 is a usage manner diagram of patient hierarchical clustering embedding representation in the MOE model provided by the embodiment of the present application; Figure 6 is a usage manner diagram of patient hierarchical clustering embedding representation in the MMOE model provided by the embodiment of the present application; Figure 7 is a usage manner diagram of patient hierarchical clustering embedding representation in the PLE model provided by the embodiment of the present application; Figure 8FIG. 1 is an AUC improvement range diagram of chronic diseases of different sample proportions provided by an embodiment of the present application; Figure 9 FIG. 2 is a structural diagram of a patient-oriented multi-chronic disease prediction device based on hierarchical clustering coding provided by an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the protection scope of the present application.

[0026] The inventive concept of the present application is that, in order to solve the problems of poor encoder interpretability, lack of hierarchical structure modeling capability, patient representation being not updatable, and insufficient utilization of individual features in multi-task prediction in the prior art, an embodiment of the present application provides a patient-oriented multi-chronic disease prediction method and device based on hierarchical clustering coding. The structured modeling of patient information and task adaptation are realized through double-tower hierarchical representation training. The explainability enhancement and dynamic updating of patient representation are realized through hierarchical clustering coding. The precise modeling of individual features and task association is realized through weight adjustment multi-task prediction. Finally, the multi-level abstract expression of patient features is realized through the above method, and is effectively introduced into the multi-task prediction model to improve the prediction accuracy and model explainability.

[0027] As shown in FIG. 1, an embodiment provides a patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding, which comprises the following steps: Figure 1 S1, a double-tower neural network model including a user expert tower for patient information modeling and a chronic disease expert tower for multi-task prediction is constructed, and multi-chronic disease joint prediction training is performed.

[0028] In the embodiment, as shown in FIG. 2, a double-tower neural network model is constructed, mainly including a user expert tower and a chronic disease expert tower. Figure 2

[0029] The user expert tower is used for patient information modeling. After receiving user feature input, it performs feature extraction through an encoder to obtain a patient vector representation (embedding). In the training stage, the patient vector representation is segmented according to a preset dimension segmentation manner to obtain a patient hierarchical vector representation composed of multiple sub-vectors of different levels. For example, if a three-level structure is set, a 128-dimensional embedding is divided into: 1-15 dimensions for the first layer (coarse granularity), 16-48 dimensions for the second layer (medium granularity), and 49-128 dimensions for the third layer (fine granularity). Each layer represents health information features of different granularities. Finally, the encoding representation of the patient is composed of multiple sub-vectors of different levels, which has a hierarchical structure. ​​

[0030] The chronic disease expert tower is used for multi-task prediction, which receives chronic disease feature input, extracts features through an encoder to obtain a chronic disease vector representation, and jointly predicts multiple chronic diseases (such as hypertension, diabetes, hyperlipidemia, coronary heart disease, etc.) for each patient. During training for joint prediction, label values indicating whether a patient has a disease are obtained based on the combined patient hierarchical vector representation and the chronic disease vector representation as a supervision signal, a threshold is determined according to the similarity of the supervision signal, and the threshold is fed back to the iterative update of the model parameters.

[0031] S2, using the trained user expert tower to encode patient information into a hierarchical embedding representation, clustering the sub-vectors of all patients at the same level to obtain clustering centers at each level, and mapping each layer of the hierarchical embedding representation of each patient to the corresponding clustering center number to obtain a patient hierarchical clustering embedding representation.

[0032] In the embodiment, the hierarchical embedding representation of all patients is extracted using the trained user expert tower, the corresponding sub-vectors (i.e., embedding at each level) of each patient are extracted at the same level to obtain a sub-vector set, and then the clustering operation is performed on each sub-vector set at each level to obtain a plurality of clustering centers corresponding to each level. The sub-vector at each level of the hierarchical embedding representation of the patient to be processed is matched with the clustering center at the corresponding level, and each sub-vector at each level is mapped to the unique identification number of the matching clustering center, referred to as "hierarchical coding", thereby forming a patient hierarchical clustering embedding representation composed of the hierarchical clustering center numbers. The coding at different levels represents the semantic features of the patient's health information at different granularities. High-level (coarse-grained) is used to capture global and stable health status features, such as overall risk of chronic diseases, basic physiological indicator categories, etc.; low-level (fine-grained) focuses on more detailed and local health signals, such as specific clinical test indicator fluctuations, recent symptom manifestations, or short-term risk factors, etc.

[0033] Through this hierarchical structure, semantic information at different granularities is explicitly organized and expressed in different dimensional subspaces, reducing interference between features and improving the model's ability to perceive different levels of health patterns. Furthermore, based on this hierarchical expression method for clustering coding, the model's ability to model complex patient states can be enhanced, especially for multi-task prediction scenarios that require attention to both long-term trends and short-term fluctuations.

[0034] In the embodiment, as shown in Figure 3 If a four-level structure is used, the four sub-vectors in the hierarchical embedding representation of the patient are mapped to the 3rd, 4th, 1st, and 5th clustering centers through hierarchical clustering, and the hierarchical coding of the patient is (3, 4, 1, 5). This coding form converts high-dimensional continuous vectors into low-dimensional discrete structures while preserving hierarchical semantics. As shown in Figure 4As shown, it can be seen that the more similar the user's coding sequence, the more similar the user's features. Then, the hierarchical coding is discretely embedded mapping to generate a unified length of quantization representation vector, i.e. patient hierarchical clustering embedding representation.

[0035] The patient hierarchical clustering embedding representation has the following characteristics: (1) generated by discrete hierarchical structure coding, with strong interpretability; (2) maintaining semantic correspondence with the original embedding, and can normally participate in training; (3) without inputting in the form of dense features, it can be automatically updated by the model.

[0036] S3, input the patient hierarchical clustering embedding representation into the task weighting network of the multi-task prediction model to perform task-by-task weighting adjustment on the intermediate representation output by the multi-task prediction model, and use the multi-task prediction model after weighting adjustment to perform multi-chronic disease prediction.

[0037] In the embodiment, the patient hierarchical clustering embedding representation is provided as an additional input feature to the task weighting network (including a gate (MLP) and a corresponding weight vector (G)) for adjusting the fusion mode between the shared representation and the task-specific representation in the basic multi-task prediction model. As shown in Figure 5 、 Figure 6 and Figure 7 , respectively, are schematic diagrams of the use of the patient hierarchical clustering embedding representation in the expert mixed model MOE, the multi-gate expert mixed model MMOE and the progressive multi-task learning model PLE. Specifically, a set of task-specific weight vectors is dynamically generated based on the patient hierarchical clustering embedding representation through the task weighting network, and the task-specific weight vectors are applied to the intermediate representation output by the multi-task prediction model for task-by-task weighting adjustment, which regulates the weight of each sub-expert network in the multi-task prediction model, and realizes patient-oriented prediction optimization. By introducing such sample-level dynamic weight mechanism, the modeling ability of the model for different patient groups can be enhanced, and the task conflict problem can be alleviated.

[0038] Further, the multi-task prediction model is trained based on the multi-chronic disease variable joint dataset (Chronic Dataset), integrates biological features, clinical features and demographic features of multiple chronic diseases, and optimizes the model through the approximate gradient descent method, while using multiple indicators such as consistency index and AUC to evaluate the performance of the model. By defining multiple risk indicators (such as baseline risk and relative risk), the model can comprehensively evaluate the risk of patients. This implementation not only fully utilizes the advantages of multi-task learning, but also ensures the accuracy and generalization ability of the model through optimization strategies and evaluation indicators. In the chronic disease prediction scenario, efficient and accurate prediction is realized, providing strong support for clinical decision-making.

[0039] In the embodiment, the performance of different models and the improved models using the method of the application is compared, as shown in Table 1, the data set used is a Chronic Dataset containing 180,000 samples extracted from a medical system, and five joint modeling tasks for five chronic diseases (diabetes, hypertension, kidney disease, coronary heart disease and cerebrovascular disease).

[0040] Table 1 AUC improvement of different models using the method of the application As can be seen from Table 1, the patient hierarchical clustering embedding representation obtained by using the hierarchical clustering coding method of the application for model weight adjustment can be widely applied and can achieve an AUC improvement of about 6 thousandths on different multi-task models.

[0041] In the embodiment, for the multi-task prediction scene, the number of training samples of different diseases is usually imbalanced, especially for diseases with less user data, the model often faces challenges such as sparse representation and insufficient learning. The hierarchical quantization coding mechanism of the application divides the patient representation into multiple granularity and semantic levels, so that the coarse-grained coding of the high layer can provide more stable and robust global health status information, effectively supplementing the potential feature missing or incomplete problem in the low-data task. At the same time, the high-layer coding has stronger discriminant ability in the sample-scarce task due to the fewer number of cluster centers and stronger generalization ability. As shown in Table 1, the experimental results also show that the AUC of the model is improved more significantly after introducing the hierarchical structure in the low-resource task (i.e. the disease prediction task with less number of samples). This shows that the hierarchical coding not only provides more rich semantic support, but also enhances the efficiency of feature utilization and expression robustness of the model in the data sparse scene, thereby improving the overall prediction performance. In contrast, the representation without hierarchical structure is more susceptible to noise interference or overfitting in the low-data situation, and it is difficult to effectively capture key discriminant features. Figure 8

[0042] In summary, the patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding provided in the embodiment of the application solves the problems of weak interpretability of traditional encoders, rigid representation and insufficient multi-task coordination ability through the joint design of "hierarchical clustering + coding quantization + multi-task adjustment", and innovatively provides a new path for personalized and structured chronic disease prediction in the medical and health field.

[0043] Based on the same inventive concept, as shown in Figure 9 ​As shown, the embodiment of the present application also provides a patient-oriented multi-chronic disease prediction device 900 based on hierarchical clustering coding, comprising a dual-tower hierarchical representation training module 910, a hierarchical clustering coding module 920, and a weight adjustment multi-task prediction module 930.

[0044] The dual-tower hierarchical representation training module 910 is configured to construct a dual-tower neural network model comprising a user expert tower for patient information modeling and a chronic disease expert tower for multi-task prediction, and perform multi-chronic disease joint prediction training.

[0045] The hierarchical clustering coding module 920 is configured to encode patient information into hierarchical embedding representation using the trained user expert tower, cluster sub-vectors at the same level for all patients to obtain clustering centers at each level, and map each sub-vector at each level in the hierarchical embedding representation of each patient to the corresponding clustering center number to obtain patient hierarchical clustering embedding representation.

[0046] The weight adjustment multi-task prediction module 930 is configured to input the patient hierarchical clustering embedding representation into the task weighting network of the multi-task prediction model, to perform task-by-task weighting adjustment on the intermediate representation output by the multi-task prediction model, and to perform multi-chronic disease prediction using the multi-task prediction model after weighting adjustment.

[0047] Based on the same inventive concept, the embodiment of the present application also provides an electronic device comprising a memory for storing a computer program and one or more processors, wherein the processor is configured to implement the above-mentioned patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding when executing the computer program.

[0048] Based on the same inventive concept, the embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a computer, implements the above-mentioned patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding.

[0049] It should be noted that the patient-oriented multi-chronic disease prediction device, electronic device, and computer-readable storage medium based on hierarchical clustering coding provided by the above-mentioned embodiments all belong to the same inventive concept as the patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding, and the specific implementation process is detailed in the patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding, which will not be repeated here.

[0050] The specific embodiments described above have explained the technical solutions and beneficial effects of the present application. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modifications, supplements, and equivalent replacements made within the principle range of the present application shall be included in the protection scope of the present application.

Claims

1. A patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding, characterized in that, The method comprises the following steps: constructing a double-tower neural network model comprising a user expert tower for patient information modeling and a chronic disease expert tower for multi-task prediction, and performing multi-chronic disease joint prediction training; encoding patient information into hierarchical embedding representations using the trained user expert tower, clustering sub-vectors at the same level for all patients to obtain clustering centers of each layer, and mapping each layer of sub-vectors in the hierarchical embedding representations of each patient to the corresponding clustering center number to obtain patient hierarchical clustering embedding representations; inputting the patient hierarchical clustering embedding representations into the task weighting network of the multi-task prediction model to perform task-by-task weighting adjustment on the intermediate representations output by the multi-task prediction model, and performing multi-chronic disease prediction using the weighted multi-task prediction model.

2. The patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding according to claim 1, characterized in that, The method comprises the following steps: constructing a double-tower neural network model comprising a user expert tower for patient information modeling and a chronic disease expert tower for multi-task prediction, and performing multi-chronic disease joint prediction training, comprising: constructing a user expert tower for patient information modeling, which receives user feature input and extracts patient vector representations through an encoder; constructing a chronic disease expert tower for multi-task prediction, which receives chronic disease feature input and extracts chronic disease vector representations through an encoder; 3.The patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding according to claim 1, wherein, segmenting the patient vector representations output by the user expert tower according to a pre-set dimension segmentation method to obtain patient hierarchical vector representations composed of multiple sub-vectors at different levels, and performing multi-chronic disease joint prediction training based on the patient hierarchical vector representations and the chronic disease vector representations. The method comprises the following steps:

4. The patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding according to claim 1 or 3, characterized in that, extracting hierarchical embedding representations of all patients using the trained user expert tower, extracting corresponding sub-vectors of each patient at the same level to obtain a sub-vector set, and then clustering each hierarchical sub-vector set to obtain a plurality of clustering centers corresponding to each level. The method comprises the following steps:

5. The patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding according to claim 4, characterized in that, matching each layer of sub-vectors in the hierarchical embedding representations of the patient to be processed with the clustering centers of the corresponding level, mapping each hierarchical sub-vector to the unique identification number of the matching clustering center, and forming a patient hierarchical clustering embedding representation composed of hierarchical clustering center numbers.

6. The patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding according to claim 1, wherein, High-level clustering center numbers represent coarse-grained global health status features, and low-level clustering center numbers represent fine-grained local health signal features. The method comprises the following steps: providing the patient hierarchical clustering embedding representations as additional input features to the task weighting network; dynamically generating a set of task-specific weight vectors based on the patient hierarchical clustering embedding representations through the task weighting network; The task-specific weight vector is applied to the intermediate representation output by the multi-task prediction model to perform task-by-task weighting adjustment to regulate the weight of each sub-expert network in the multi-task prediction model.

7. The patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding according to claim 1 or 6, characterized in that, The multi-task prediction model comprises an expert mixture model MOE, a multi-gated expert mixture model MMOE, or a progressive multi-task learning model PLE.

8. A patient-oriented multi-chronic disease prediction device based on hierarchical clustering coding, implemented by using the patient-oriented multi-chronic disease prediction method based on hierarchical clustering coding according to any one of claims 1 to 7. The method comprises: a double-tower hierarchical representation training module, a hierarchical clustering encoding module, and a weight-adjusted multi-task prediction module; The double-tower hierarchical representation training module is configured to construct a double-tower neural network model comprising a user expert tower for modeling patient information and a chronic disease expert tower for multi-task prediction, and perform multi-chronic disease joint prediction training. The hierarchical clustering encoding module is configured to encode patient information into hierarchical embedding representation using the trained user expert tower, cluster sub-vectors at the same level for all patients to obtain clustering centers at each level, and map each layer sub-vector in the hierarchical embedding representation of each patient to the corresponding clustering center number to obtain patient hierarchical clustering embedding representation. The weight-adjusted multi-task prediction module is configured to input the patient hierarchical clustering embedding representation into the task weighting network of the multi-task prediction model to perform task-by-task weighting adjustment on the intermediate representation output by the multi-task prediction model, and perform multi-chronic disease prediction using the multi-task prediction model after the weighting adjustment.

9. An electronic device comprising a memory and one or more processors, the memory to store a computer program, characterized in that, The processor is configured to implement the patient-oriented multi-chronic disease prediction method based on hierarchical clustering encoding according to any one of claims 1-7 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a computer, implements the patient-oriented multi-chronic disease prediction method based on hierarchical clustering encoding according to any one of claims 1-7.

Images