Time series state evaluation method and system based on hour theory and deep learning

Abstract

The present application relates to a kind of time series state evaluation method and system based on HOUR theory and deep learning, the method includes: obtaining the time series detection data of multiple biomarkers and inputting time series coding module, to adaptively update hidden state according to the time interval between each detection time point, output hidden state sequence and input to deep learning module, with the help of neural network and the half-life prior defined by HOUR theory Continuous time evolution, output the expected concentration time series estimate of each biomarker in future preset time window;With multiple biomarkers as node, with the physiological interaction between multiple biomarkers as edge, according to the edge type defined by HOUR theory, construct heterogeneous graph, and the expected concentration time series estimate is as the current node feature of each node, update the feature representation of each node by attention mechanism and output dynamic threshold matrix;Based on expected concentration time series estimate and dynamic threshold matrix, generate corresponding state evaluation result.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a time-series state evaluation method and system based on HOUR theory and deep learning. Background Technology

[0002] In the field of bioinformatics, the comprehensive processing and analysis of time-series data of multiple biomarkers is a technical direction with broad application needs. Taking clinical monitoring scenarios as an example, detection generates various types of time-series biomarker data, such as, but not limited to, inflammation-related biomarkers reflecting the level of inflammatory response, such as C-reactive protein, procalcitonin, interleukin-6, interleukin-8, interleukin-10, and serum amyloid A; coagulation-related biomarkers reflecting coagulation function, such as D-dimer, fibrinogen, and international normalized ratio; immune-related biomarkers reflecting immune cell function, such as lymphocyte count and neutrophil count; and damage-related biomarkers reflecting the degree of tissue damage, such as ferritin and lactate dehydrogenase.

[0003] In the field of critical care medicine, the HOUR (Host / Organ Unregulated Response) theory serves as a theoretical framework for understanding the multi-system dysregulation response in critically ill patients. This theory posits that when faced with severe trauma (such as infection or injury), the initial protective response may become uncontrolled due to sustained stimulation or inappropriate intervention, leading to a cascade of dysregulation across multiple systems, including the nervous, endocrine, inflammatory, immune, coagulation, and metabolic bioenergetics systems. Notably, inflammation, immunity, and coagulation do not occur sequentially but rather are activated synchronously, forming a positive feedback network that amplifies each other, constituting a complex coupling structure among multiple biomarkers.

[0004] The aforementioned different types of biomarkers are not independent of each other, but rather exhibit inherent physiological correlations. Taking inflammation as an example, changes in inflammation-related biomarkers can be accompanied by corresponding changes in coagulation-related biomarkers, changes in immune-related biomarkers can be accompanied by corresponding changes in inflammation-related biomarkers, and changes in tissue damage-related biomarkers can be accompanied by further changes in inflammation-related biomarkers, forming a multi-axis cascade effect. These correlations constitute a complex coupling structure among the various biomarker data. At the data processing level, if the aforementioned multi-biomarker data can be systematically fused and analyzed according to their correlations, it should be possible to extract more comprehensive feature information from the data than relying solely on a single biomarker or analyzing multiple biomarkers in isolation.

[0005] Furthermore, the aforementioned biomarkers also exhibit inherent sequential relationships in their temporal responses. For example, interleukin-6 rapidly increases within 1-2 hours after stress stimulation, earlier than C-reactive protein and procalcitonin; procalcitonin increases within 2-4 hours, reflecting a specific response to pathogen infection; serum amyloid A rapidly increases within 3-6 hours, with its early sensitivity superior to C-reactive protein; C-reactive protein begins to increase at 6-8 hours and peaks at 24-48 hours. These temporal dynamics constitute another important structural information between the biomarkers, in addition to their interactions, providing a physiological basis for subsequent continuous temporal evolution and prediction using deep learning models.

[0006] However, most existing technologies rely on fixed thresholds or single biomarkers for data processing and status assessment. For example, patent application CN113096792A discloses a machine learning-based intelligent health monitoring and early warning method and system. While this method uses threshold parameters for early warning, its warning logic still relies on a single fixed threshold, failing to adapt to individual differences. Furthermore, it processes each biomarker independently, neglecting to utilize correlation information and temporal dynamic trends between biomarkers. Patent application CN115547502A discloses a risk prediction device based on time-series data, using a recurrent neural network to process clinical time-series data. This approach assumes the input data is an equally spaced sampling sequence, failing to effectively handle the irregular sampling and missing value problems commonly found in clinical settings. Moreover, it only models each biomarker independently, neglecting the coupling relationships between biomarkers.

[0007] The common drawbacks of the above-mentioned approaches are: the fixed threshold method lacks individualized adaptability and fails to utilize temporal dynamic trend information; traditional time-series models cannot effectively handle irregular sampling data; and none of them utilize the correlation information and coupling structure between biomarkers. These shortcomings result in data processing results that can only reflect one or a partial aspect of the target object's state, failing to generate comprehensive state information that fully reflects the target object.

[0008] Therefore, there is an urgent need for a data processing solution that can perform a comprehensive status assessment of target objects based on time-series detection data of multiple biomarkers. Summary of the Invention

[0009] In view of the shortcomings of the prior art, the purpose of the invention is to provide a time-series state evaluation method and system based on HOUR theory and deep learning.

[0010] A first aspect of the present invention proposes a temporal state assessment method based on HOUR theory and deep learning, comprising: acquiring temporal detection data of multiple biomarkers of a target object; the temporal detection data including detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; inputting the temporal detection data into a temporal encoding module, the temporal encoding module adaptively updating the hidden state according to the time interval between each detection time point, and outputting a hidden state sequence composed of the hidden states corresponding to each detection time point; inputting the hidden state sequence into a first deep learning module, the first deep learning module performing continuous temporal evolution of the hidden state based on a neural network and predefined prior terms, and outputting a temporal estimate of the expected concentration of each biomarker within a future preset time window. The prior terms are defined based on the half-life information of various biomarkers in the HOUR theory. A heterogeneous graph is constructed using the various biomarkers as nodes and the physiological interactions between them as edges, according to the edge types defined in the HOUR theory. The expected concentration time-series estimate is used as the current node feature of each node. The current node features of each node are input into a second deep learning module, which updates the feature representation of each node through an attention mechanism and outputs a dynamic threshold matrix based on the updated node features. The elements of the dynamic threshold matrix are the probability distributions of each biomarker pair falling into a preset grading interval. The preset grading interval is used to indicate the risk level. Based on the expected concentration time-series estimate and the dynamic threshold matrix, a corresponding state assessment result is generated.

[0011] According to a second aspect of the present disclosure, a storage medium is provided, the storage medium including a stored program, wherein the method described above is executed by a processor when the program is running.

[0012] According to a third aspect of the present disclosure, a time-series state assessment system based on HOUR theory and deep learning is provided, comprising: a detection data acquisition module for acquiring time-series detection data of multiple biomarkers of a target object; the time-series detection data including detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; a hidden state sequence construction module for inputting the time-series detection data into a time-series encoding module, the time-series encoding module adaptively updating the hidden states according to the time interval between each detection time point, and outputting a hidden state sequence composed of the hidden states corresponding to each detection time point; and a future concentration prediction module for inputting the hidden state sequence into a first deep learning module, the first deep learning module performing continuous time evolution of the hidden states based on a neural network and predefined prior terms, and outputting the expected concentration of each biomarker within a future preset time window. The system comprises: a time-series estimate; prior terms defined based on the half-life information of various biomarkers in the HOUR theory; a heterogeneous graph construction module, used to construct a heterogeneous graph with the various biomarkers as nodes and the physiological interactions between the various biomarkers as edges, according to the edge types defined in the HOUR theory, and using the expected concentration time-series estimate as the current node feature of each node; a dynamic threshold matrix construction module, used to input the current node features of each node into a second deep learning module, the second deep learning module updating the feature representation of each node through an attention mechanism, and outputting a dynamic threshold matrix based on the updated node features; the elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset grading interval; the preset grading interval is used to indicate the risk level; and a state assessment module, used to generate corresponding state assessment results based on the expected concentration time-series estimate and the dynamic threshold matrix.

[0013] According to a fourth aspect of the present disclosure, a temporal state assessment system based on HOUR theory and deep learning is provided, comprising: a processor; and a memory connected to the processor, configured to provide the processor with instructions for processing the following steps: acquiring temporal detection data of multiple biomarkers of a target object; the temporal detection data including detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; inputting the temporal detection data to a temporal encoding module, the temporal encoding module adaptively updating the hidden state according to the time interval between each detection time point, and outputting a sequence of hidden states composed of the hidden states corresponding to each detection time point; and inputting the hidden state sequence to a first deep learning module, the first deep learning module performing continuous-time evolution of the hidden state based on a neural network and predefined prior terms, and outputting a sequence of hidden states. The expected concentration time-series estimate of biomarkers within a preset future time window; the prior terms are defined based on the half-life information of each biomarker in the HOUR theory; using the multiple biomarkers as nodes and the physiological interactions between the multiple biomarkers as edges, a heterogeneous graph is constructed according to the edge type defined by the HOUR theory, and the expected concentration time-series estimate is used as the current node feature of each node; the current node features of each node are input into a second deep learning module, the second deep learning module updates the feature representation of each node through an attention mechanism, and outputs a dynamic threshold matrix based on the updated node features; the elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset classification interval; the preset classification interval is used to indicate the risk level; based on the expected concentration time-series estimate and the dynamic threshold matrix, a corresponding state assessment result is generated.

[0014] This application provides a raw data foundation for adaptive encoding processing of irregularly sampled data by acquiring time-series detection data of multiple biomarkers of the target object. The time-series encoding module adaptively updates the hidden state based on the time interval between each detection time point, providing an adaptive encoding foundation for irregularly sampled multi-biomarker time-series data without interpolation preprocessing. A first deep learning module performs continuous-time evolution of the hidden state based on a neural network and predefined prior terms, where the prior terms are defined based on the half-life information of each biomarker in the HOUR theory. This integrates known physiological prior knowledge into the continuous-time prediction process, improving the predictive rationality in sparse data regions. Furthermore, a heterogeneous graph is constructed using the multiple biomarkers as nodes and the physiological interactions between them as edges, based on the edge types defined in the HOUR theory, without relying on external databases. This approach enables the modeling of cross-biomarker coupling relationships to possess both physiological interpretability and ease of deployment. By updating the feature representations of each node through the attention mechanism of the second deep learning module and outputting a dynamic threshold matrix based on the updated node features, the grading thresholds for each biomarker can be dynamically and adaptively adjusted according to individual differences and time series, and the grading results are output in the form of a probability distribution, thus replacing fixed thresholds. By generating corresponding state assessment results based on the expected concentration time series estimate and the dynamic threshold matrix, the assessment results integrate future trend predictions and cross-biomarker coupling probability information, providing a comprehensive quantitative basis for the integrated state assessment of multi-biomarker time series data. This achieves a comprehensive state assessment of the target object based on multi-biomarker time series data, overcoming the technical deficiency in existing technologies where data processing results can only reflect one or a partial aspect of the target object's state. Attached Figure Description

[0015] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. It is obvious that the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings.

[0016] Figure 1 This is a hardware structure block diagram of a computing device for implementing the method described in Embodiment 1 of this disclosure; Figure 2 This is a flowchart of the time-series state evaluation method based on HOUR theory and deep learning according to Embodiment 1 of this application; Figure 3 This is a schematic diagram of a time-series state evaluation system based on HOUR theory and deep learning, as described in Embodiment 2 of this application. Figure 4This is a schematic diagram of a time-series state evaluation system based on HOUR theory and deep learning, as described in Embodiment 3 of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this disclosure.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0019] Terminology Explanation: 1. HOUR Theory: HOUR stands for Host / Organ Unregulated Response. It refers to the pathophysiological framework of the body's initial protective response to severe shocks (such as infection or trauma) which becomes uncontrolled due to sustained stimulation or inappropriate intervention, leading to multi-organ dysfunction. This theory encompasses dysregulation responses in multiple systems, including the nervous, endocrine, inflammatory, immune, coagulation, bioenergetics, and metabolic systems. Among these, the inflammatory and immune responses are activated synchronously, driving the coagulation response. These three components form a mutually amplifying positive feedback network, forming the core theoretical framework for understanding the complex coupling relationships between multiple biomarkers.

[0020] In this invention, the HOUR theory provides a theoretical basis for the physiological interactions between various biomarkers, and is used to guide the definition of edge types in heterogeneous graphs and the selection of half-life information in prior terms.

[0021] 2. Hidden State: This refers to the hidden state vector output by the timing coding module at each detection time point during the processing of timing detection data. The hidden state is a compressed representation of historical timing information up to the current time point, containing the temporal dynamic information of previous detection values. The hidden states at each detection time point are arranged in chronological order to form a hidden state sequence.

[0022] Example 1

[0023] According to this embodiment, a method embodiment of a time-series state evaluation method based on HOUR theory and deep learning is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] The method embodiments provided in this example can be executed on a server or similar computing device. Figure 1 A hardware block diagram of a computing device for implementing a time-series state evaluation method based on HOUR theory and deep learning is shown. Figure 1 As shown, a computing device may include one or more processors (processors may include, but are not limited to, microprocessors such as MCUs or programmable logic devices such as FPGAs), a memory for storing data, a transmission device for communication functions, and an input / output interface. The memory, transmission device, and input / output interface are connected to the processor via a bus. In addition, it may also include a display, keyboard, and cursor control device connected to the input / output interface. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, a computing device may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0025] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element in a computing device. As involved in the embodiments of this disclosure, the data processing circuits serve as processor control (e.g., selection of a variable resistor termination path connected to an interface).

[0026] The memory can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the time-series state evaluation method based on HOUR theory and deep learning in the embodiments of this disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned application's time-series state evaluation method based on HOUR theory and deep learning. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the computing device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0027] The transmission device is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the computing device's communication provider. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0028] The display can be, for example, a touchscreen liquid crystal display (LCD), which allows users to interact with the user interface of the computing device.

[0029] It should be noted here that, in some optional embodiments, the above... Figure 1 The computing device shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computing devices.

[0030] It should be noted that the temporal state assessment method based on HOUR theory and deep learning provided in this application belongs to computer-implemented information processing technology. Its direct output is numerical assessment data characterizing the state of multiple biomarkers of the target object, rather than a definitive disease diagnosis conclusion. Essentially, this application achieves quantitative assessment and feature extraction of the target object's state through a series of information processing steps, including temporal encoding, continuous temporal evolution, heterogeneous graph attention processing, and weighted fusion (where the temporal encoding module, the first deep learning module, and the second deep learning module all employ deep learning models trained on labeled temporal data). The entire process involves only the processing and analysis of the acquired temporal detection data, without direct contact with the human body or any invasive procedures. Furthermore, the final assessment result aims to provide structured comprehensive state assessment data for subsequent information processing equipment, rather than replacing the final judgment of a professional. Therefore, it does not fall under the definition of a disease diagnosis method in the sense of patent law.

[0031] Under the above operating environment, according to the first aspect of this embodiment, a time-series state evaluation method based on HOUR theory and deep learning is provided. Figure 2 A flowchart illustrating the method is shown below. (Refer to...) Figure 2 As shown, the method includes: S1: Obtain time-series detection data of multiple biomarkers of the target object; the time-series detection data includes detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; The target population consists of individuals undergoing clinical monitoring. The biomarkers are selected from multiple of the following: biomarkers reflecting the level of inflammatory response (e.g., C-reactive protein, procalcitonin, interleukin-6, interleukin-8, interleukin-10, serum amyloid A, etc.), biomarkers reflecting coagulation function status (e.g., D-dimer, fibrinogen, international normalized ratio, etc.), biomarkers reflecting immune cell function (e.g., lymphocyte count, neutrophil count, etc.), and biomarkers reflecting the degree of tissue damage (e.g., ferritin, lactate dehydrogenase, etc.).

[0032] Time-series detection data can be obtained by establishing a communication connection with a clinical testing system (such as a laboratory information management system or a hospital information system) through a data interface, and extracting the detection records of the target object within a preset historical time period from the clinical testing system; alternatively, it can be obtained by reading pre-stored detection record files of the target object from a storage medium. The detection records contain the timestamps of each test and the detection values ​​obtained for each test.

[0033] Due to constraints imposed by various factors such as testing costs, timing, and differences in clinical medical orders in actual clinical testing scenarios, not all items of the aforementioned biomarkers are tested at every testing time point. For example, at one testing time point, only inflammation-related biomarkers may be tested, while coagulation-related biomarkers may not be tested; at another testing time point, only coagulation-related biomarkers may be tested, while inflammation-related biomarkers may not be tested; or at a certain testing time point, some inflammation-related biomarkers and some coagulation-related biomarkers may be tested simultaneously, but other biomarkers may remain undetected. Therefore, the structure of the time-series testing data is as follows: there are multiple testing time points, each testing time point corresponds to a set of test values, and each set of test values ​​contains the values ​​of the portion of biomarkers actually tested at that testing time point. That is, the time-series testing data includes the test values ​​obtained by testing at least one of the aforementioned biomarkers at multiple testing time points, meaning that the set of test values ​​at each testing time point is a subset of the complete set of the aforementioned biomarkers.

[0034] Since the types of markers detected at different detection time points in the time-series detection data are not entirely the same, and the time intervals between adjacent detection time points may be inconsistent, the time-series detection data is irregularly sampled multivariate time series data containing missing values. Furthermore, the acquired time-series detection data also includes timestamp information for each detection time point, enabling subsequent steps to determine the actual time intervals between each detection time point based on the timestamp information.

[0035] Therefore, by acquiring time-series detection data of multiple biomarkers in the above data structure, a raw data foundation is provided for the adaptive encoding processing of subsequent irregular sampling data. At the same time, the types and values ​​of biomarkers actually detected at each detection time point, as well as the timestamp information of each detection time point, are preserved, so that the raw data naturally contains sampling information with inconsistent detection time points of different biomarkers.

[0036] S2: Input the timing detection data into the timing encoding module. The timing encoding module adaptively updates the hidden state according to the time interval between each detection time point and outputs a sequence of hidden states composed of the hidden states corresponding to each detection time point. In this embodiment, the time-series detection data obtained in step S1 is provided as input to the time-series encoding module. The time-series encoding module employs a recurrent neural network structure capable of sensing time intervals. Unlike traditional recurrent neural networks, when updating the hidden state at each detection time point, the time-series encoding module actively acquires the actual time interval between the current detection time point and the previous detection time point, and adaptively adjusts the degree of retention of historical information based on the magnitude of this time interval.

[0037] Specifically, for the first detection time point in the time-series detection data, the time-series encoding module encodes the detection value at that time point to generate the hidden state corresponding to that detection time point. For each subsequent detection time point in the time-series detection data, the time-series encoding module first uses the timestamp information to calculate the actual time interval between the current detection time point and the previous detection time point. Subsequently, the time-series encoding module processes the hidden state corresponding to the previous detection time point according to the size of this time interval: when the time interval is large, the contribution of historical information in the hidden state of the previous detection time point to the current moment is reduced; when the time interval is small, this contribution is relatively preserved. After completing the adaptive processing of the hidden state of the previous moment, the time-series encoding module further combines the detection value of the current detection time point and generates the hidden state of the current detection time point according to the update mechanism of the recurrent neural network.

[0038] For the detection value at the current detection time point, since the types of markers detected at different detection time points are different, the time-series coding module will process it accordingly based on the actual marker items contained in the detection value: for markers that have been detected, their detection values ​​are used as input; for markers that have not been detected, adaptive estimation is performed based on the time interval since they were last detected.

[0039] The timing encoding module performs the above operations sequentially for each detection time point in the timing detection data, thereby generating a corresponding hidden state for each of the multiple detection time points. The hidden states corresponding to each detection time point are arranged in chronological order to form a hidden state sequence.

[0040] Thus, by adaptively updating the hidden state according to the time interval between each detection time point through the time-series coding module, an adaptive coding foundation without interpolation preprocessing is provided for irregularly sampled multi-biomarker time-series data. This enables the time-series coding module to directly process time-series data sampled at unequal intervals without having to convert the data into an equally spaced sequence through interpolation or other means in the data preprocessing stage. At the same time, the output sequence and the input sequence maintain a one-to-one correspondence in the time dimension.

[0041] S3: The hidden state sequence is input into the first deep learning module. The first deep learning module performs continuous time evolution on the hidden state based on the neural network and predefined prior terms, and outputs the expected concentration time series estimate of each biomarker within a future preset time window; the prior terms are defined based on the half-life information of each biomarker in the HOUR theory. In this embodiment, the hidden state sequence output in S2 is provided as input to the first deep learning module. The last hidden state in the hidden state sequence has been fused with historical information from all detection time points in the temporal detection data through the cyclic update mechanism of the temporal coding module, and therefore the last hidden state is used as the initial state for continuous temporal evolution.

[0042] The first deep learning module employs a continuous-time neural network model based on ordinary differential equations, capable of transforming discrete latent state sequences into dynamic processes in the continuous-time domain. This module constructs a dynamic equation describing the rate of change of latent variable states in the continuous-time domain. This dynamic equation is determined by two parts of information: one part comes from the neural network, used to learn complex dynamic patterns of biomarker concentration changes from training data; the other part comes from predefined prior terms. These prior terms are defined based on the half-life information of various biomarkers in the HOUR theory, which provides information on the elimination half-life of various biomarkers under normal physiological conditions. For example, the half-life of C-reactive protein is approximately 19 hours, procalcitonin's half-life is approximately 20 to 24 hours, and fibrinogen's half-life is approximately 100 hours. These prior terms transform this half-life information into constraints on the evolution of latent variable states, ensuring that the evolution of latent variable states does not deviate from known elimination dynamics during time intervals lacking observational data.

[0043] Furthermore, the HOUR theory provides information on the temporal response sequence of various biomarkers under stress. For example, interleukin-6 rises rapidly within 1-2 hours after stress stimulation, earlier than C-reactive protein and procalcitonin; serum amyloid A rises rapidly within 3-6 hours, with its early sensitivity being superior to C-reactive protein; C-reactive protein begins to rise at 6-8 hours and peaks at 24-48 hours. Although these temporal dynamic values ​​are not directly used as calculation parameters for prior terms, they provide a physiologically reasonable benchmark for verifying the continuous temporal evolution of the first deep learning module in this application, i.e., the predicted concentration change time window should be consistent with the aforementioned known temporal dynamic laws.

[0044] Starting from the initial state, within the time interval from the current detection time to the end of a future preset time window, the first deep learning module uses an ordinary differential equation solver to numerically integrate the aforementioned dynamic equations, obtaining the latent variable states at each time point within the future preset time window. Then, a mapping layer projects these latent variable states from the latent variable space back into the original data space, obtaining the expected concentration estimates of each biomarker at each time point within the future preset time window, i.e., the expected concentration time-series estimates. The length of the future preset time window can be preset according to the actual application scenario.

[0045] Thus, the first deep learning module performs continuous-time evolution of the hidden state based on neural networks and predefined prior terms. The prior terms are defined based on the half-life information of various biomarkers in the HOUR theory, which integrates known physiological prior knowledge into the continuous-time prediction process, improves the prediction rationality in sparse data regions, and enables the output of predicted values ​​at any future time point, breaking through the limitation of discrete time steps.

[0046] S4: Using the multiple biomarkers as nodes and the physiological interactions between the multiple biomarkers as edges, define the edge types according to the HOUR theory, construct a heterogeneous graph, and use the expected concentration time series estimate as the current node feature of each node; In this embodiment, after obtaining the expected concentration time-series estimates of each biomarker, in order to explicitly model the mutual influence between different biomarkers, it is necessary to construct the physiological interactions between the multiple biomarkers into a heterogeneous graph structure.

[0047] Specifically, each of the multiple biomarkers is defined as a node, with the number of nodes matching the number of biomarkers. Each node uniquely corresponds to one biomarker, forming a node set. Edges between nodes represent physiological interactions between biomarkers. For any two biomarkers, if a physiological interaction exists between them, an edge is established between the nodes corresponding to those two biomarkers; otherwise, no edge is established. The determination of these physiological interactions is based on the HOUR theory. The HOUR theory reveals the cascade coupling relationship between inflammatory responses, immune function, and the coagulation system. This theoretical framework specifically includes: changes in inflammatory response-related biomarkers can be accompanied by corresponding changes in coagulation-related biomarkers; changes in immune-related biomarkers can be accompanied by corresponding changes in inflammatory biomarkers; and changes in tissue damage-related biomarkers can be accompanied by further changes in inflammatory biomarkers.

[0048] Based on this theoretical framework, this step further assigns edge type identifiers to the aforementioned physiological interactions. Specifically, according to the differences in the types of interactions in the HOUR theory, the physiological interactions between the multiple biomarkers are divided into various edge types. Different edge types correspond to different types of physiological regulatory relationships, such as edge types reflecting the regulatory relationship between inflammation-related markers and immune-related markers, edge types reflecting the regulatory relationship between inflammation-related markers and coagulation-related markers, edge types reflecting the regulatory relationship between immune-related markers and coagulation-related markers, and edge types reflecting the regulatory relationship between tissue damage-related markers and other types of markers, etc. After the above division, each edge in the heterogeneous graph has a clear type identifier, used to indicate the specific category of the physiological interaction corresponding to that edge.

[0049] After the edge types are defined, a heterogeneous graph is constructed using the node set and the edge set defined according to the edge types. Each node in the node set of the heterogeneous graph has a node type identifier, and each edge in the edge set has an edge type identifier. It should be noted that the node and edge types of the heterogeneous graph are entirely predefined based on the HOUR theory. The entire construction process does not rely on any external bioinformatics database and does not require querying external data sources at runtime to obtain graph structure information.

[0050] After the heterogeneous graph is constructed, the expected concentration time-series estimate output in step S3 is used as the current node feature of each node. Specifically, for each node in the node set, the expected concentration time-series estimate of its corresponding biomarker is used as the current node feature vector of that node. That is, the expected concentration time-series estimate includes the expected concentration estimate of the biomarker at each time point within a future preset time window. Therefore, the dimension of this feature vector is related to the number of time points within the future preset time window.

[0051] Therefore, by using the multiple biomarkers as nodes and the physiological interactions between them as edges, and constructing a heterogeneous graph based on the edge types defined by the HOUR theory, without relying on external databases, the modeling of cross-marker coupling relationships is both physiologically interpretable and easy to deploy. At the same time, it connects the originally isolated biomarkers into a whole through edges with clear physiological meanings, providing a structured data foundation for subsequent graph attention processing.

[0052] S5: Input the current node features of each node into the second deep learning module. The second deep learning module updates the feature representation of each node through an attention mechanism and outputs a dynamic threshold matrix based on the updated node features. The elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset grading interval. The preset grading interval is used to indicate the risk level. In this embodiment, after the heterogeneous graph is constructed, the current node features of each node are input into the second deep learning module. The second deep learning module is a graph neural network based on an attention mechanism, used to perform message passing on the heterogeneous graph to update the feature representations of each node.

[0053] For each node in the heterogeneous graph, the second deep learning module calculates the attention weights of each neighboring node based on the node's own features and the features of its neighboring nodes. These attention weights reflect the importance of each neighboring node to the node. Then, the second deep learning module performs weighted aggregation of the features of each neighboring node based on these attention weights to obtain the updated feature representation of the node. Since edges in the heterogeneous graph have different types, and the neighboring nodes connected to different types of edges influence the node in different ways, the second deep learning module performs attention calculation and feature aggregation separately when processing different types of edges, enabling each node's feature representation to obtain information from different types of neighboring nodes.

[0054] After the above message transmission, the feature representation of each node is updated to a new feature representation that incorporates the information of its neighboring nodes. That is, the feature of each node not only includes the information of its own markers, but also the information of other markers with which it has physiological interaction.

[0055] After obtaining the updated feature representations of each node, the second deep learning module outputs a dynamic threshold matrix based on these features. Specifically, for each biomarker, the second deep learning module outputs the probability that the biomarker falls into each of K preset grading intervals based on the updated features of the corresponding node. The preset grading intervals are pre-divided intervals based on the concentration levels of each biomarker, used to indicate risk levels; the boundaries of the grading intervals for different biomarkers can be set independently. Based on the above probability distributions of each biomarker, the second deep learning module further generates a dynamic threshold matrix, the elements of which are the probability distributions of each biomarker falling into each preset grading interval.

[0056] Thus, the feature representation of each node is updated through the attention mechanism of the second deep learning module, and a dynamic threshold matrix is ​​output based on the updated node features. This enables the grading threshold of each biomarker to be dynamically and adaptively adjusted according to individual differences and time sequence, and the grading results are output in the form of a probability distribution, thereby replacing the fixed threshold. At the same time, the updated node features contain coupling information across biomarkers.

[0057] S6: Based on the expected concentration time series estimate and the dynamic threshold matrix, generate the corresponding state assessment result.

[0058] In this embodiment, after obtaining the expected concentration time-series estimate and the dynamic threshold matrix, these two pieces of information are integrated to generate the final state assessment result. The expected concentration time-series estimate includes predicted concentration changes of each biomarker within a preset time window, reflecting the future trends of each biomarker. The dynamic threshold matrix includes the joint probability distribution information of each biomarker falling into each preset grading interval, reflecting the current cross-marker coupling grading state.

[0059] Specifically, firstly, trend feature information is extracted from the expected concentration time-series estimates to characterize the concentration change trend of each biomarker within a preset future time window. Simultaneously, current state feature information is extracted from the dynamic threshold matrix to characterize the coupling hierarchy state of each biomarker at the current moment. The trend feature information and the current state feature information are then combined to obtain comprehensive evaluation information. This comprehensive evaluation information includes both future trend information and current cross-biomarker coupling hierarchy probability information.

[0060] After obtaining the comprehensive assessment information, the expected concentration time-series estimate, the probability distribution information in the dynamic threshold matrix, the comprehensive assessment information, and the probability distribution of each biomarker falling into each preset grade interval are structured and encapsulated to form the state assessment result. The state assessment result is multi-dimensional numerical assessment data, which does not contain any diagnostic conclusions or treatment recommendations. Its data format is a structured data format, which can be read and applied by downstream equipment or professionals.

[0061] Therefore, by generating corresponding state assessment results based on the expected concentration time series estimates and the dynamic threshold matrix, the assessment results integrate future trend predictions and cross-biomarker coupling probability information, providing a comprehensive quantitative basis for the integrated state assessment of multi-biomarker time series data.

[0062] As described in the background section, most existing technologies rely on fixed thresholds or single biomarkers for data processing and state assessment. However, fixed threshold methods lack individualized adaptability and fail to utilize time-series dynamic trend information; traditional time-series models cannot effectively handle irregularly sampled data; and they fail to utilize the correlation information and coupling structure between biomarkers. These shortcomings result in data processing results that only reflect one or a partial aspect of the target object's state, failing to generate comprehensive state information that fully reflects the target object.

[0063] In view of this, this application provides a raw data foundation for adaptive encoding processing of irregularly sampled data by acquiring time-series detection data of multiple biomarkers of the target object; by using a time-series encoding module to adaptively update the hidden state according to the time interval between each detection time point, it provides an adaptive encoding foundation for irregularly sampled multi-biomarker time-series data without interpolation preprocessing; by using a first deep learning module to continuously evolve the hidden state based on a neural network and predefined prior terms, and the prior terms are defined based on the half-life information of each biomarker in the HOUR theory, the continuous-time prediction process incorporates known physiological prior knowledge, improving the prediction rationality in sparse data regions; by constructing a heterogeneous graph based on the multiple biomarkers as nodes and the physiological interactions between the multiple biomarkers as edges, according to the edge type defined by the HOUR theory, and without relying on external... The database enables the modeling of cross-biomarker coupling relationships to be both physiologically interpretable and easy to deploy. The attention mechanism of the second deep learning module updates the feature representations of each node, and outputs a dynamic threshold matrix based on the updated node features. This allows the grading thresholds for each biomarker to be dynamically and adaptively adjusted according to individual differences and time series, and the grading results are output in the form of a probability distribution, thus replacing fixed thresholds. By generating corresponding state assessment results based on the expected concentration time series estimates and the dynamic threshold matrix, the assessment results integrate future trend predictions and cross-biomarker coupling probability information, providing a comprehensive quantitative basis for the integrated state assessment of multi-biomarker time series data. This achieves a comprehensive state assessment of the target object based on multi-biomarker time series data, overcoming the technical deficiency in existing technologies where data processing results can only reflect one or a partial aspect of the target object's state.

[0064] Optionally, before inputting the time-series detection data into the time-series coding module, a preprocessing step is further included: constructing a mask matrix and a time interval matrix based on the time-series detection data; the mask matrix is ​​used to indicate whether each biomarker is observed at each detection time point, and the time interval matrix is ​​used to indicate the length of time during which the detection values ​​of each biomarker have been continuously missing up to each detection time point.

[0065] In this embodiment, the time-series detection data is an irregularly sampled multivariate time series data with two significant characteristics. First, the types of biomarkers detected at different detection time points are not entirely the same; that is, the set of detection values ​​at each detection time point is a subset of the complete set of multiple biomarkers, rather than all items being detected. Second, the time intervals between adjacent detection time points may be inconsistent; that is, the detection time points are not equally spaced on the time axis. These two data characteristics result in a large number of missing values ​​in the time-series detection data. If a biomarker is not detected at a certain detection time point, it is considered that the detection value of that biomarker at that time point is missing.

[0066] Since the missing values ​​in the time-series detection data are not randomly distributed but are related to the actual detection behavior (e.g., only some detection items were ordered in a certain detection), the missing values ​​themselves contain meaningful sampling pattern information. Simultaneously, the continuous missing time length of each biomarker since its last detection reflects the duration for which the biomarker has not been sampled; this duration information is crucial for determining the degree of uncertainty in historical information. Based on the above analysis, this embodiment preprocesses the time-series detection data before inputting it into the time-series coding module to construct a mask matrix and a time interval matrix.

[0067] The rows of the mask matrix correspond to the plurality of detection time points, and the columns of the mask matrix correspond to the plurality of biomarkers. For each element in the mask matrix, if the biomarker corresponding to the column of the element is observed at the detection time point corresponding to the row of the element (i.e., a detection value exists), the value of the element is set to 1; if the biomarker is not observed at the detection time point (i.e., a detection value is missing), the value of the element is set to 0. Thus, a position in the mask matrix with a value of 1 indicates that the corresponding biomarker was observed at the detection time point, and a position with a value of 0 indicates that the corresponding biomarker was not observed at the detection time point. That is, the mask matrix can be used to indicate whether each biomarker was observed at each detection time point.

[0068] The rows of the time interval matrix correspond to the plurality of detection time points, and the columns of the time interval matrix correspond to the plurality of biomarkers. For each element in the time interval matrix, the value of the element represents the length of time during which the biomarker corresponding to the column of that element has been continuously missing up to the detection time point corresponding to the row of that element. Specifically, for a biomarker at a certain detection time point, if the biomarker is observed at that detection time point, the length of continuous missing time corresponding to that detection time point is 0; if the biomarker is not observed at that detection time point, the length of continuous missing time is the time interval from the detection time point where the biomarker was last observed to the current detection time point; if the biomarker has never been observed, the length of continuous missing time is the time interval from the first detection time point to the current detection time point. Thus, the time interval matrix can be used to indicate the length of time during which the detection values ​​of each biomarker have been continuously missing up to each detection time point.

[0069] Therefore, by constructing the mask matrix and the time interval matrix, the missing patterns and irregular sampling information in the time-series detection data are extracted into a structured matrix form, providing structured information input for the subsequent time-series coding module regarding which biomarkers were observed, which were not observed, and how long each biomarker was continuously missing at each detection time point.

[0070] Optionally, the operation of inputting the time-series detection data into a time-series encoding module, wherein the time-series encoding module adaptively updates the hidden states according to the time interval between each detection time point and outputs a sequence of hidden states composed of the hidden states corresponding to each detection time point, includes: inputting the time-series detection data, the mask matrix, and the time interval matrix into the time-series encoding module; the time-series encoding module is a time decay network based on gated recurrent units; determining whether each biomarker is observed at the current detection time point according to the mask matrix; for unobserved biomarkers, calculating an input decay factor according to the corresponding continuous missing time length in the time interval matrix, and adaptively filling the detection value of the biomarker according to the input decay factor; for observed biomarkers... For biomarkers, the corresponding detection values ​​are used directly. A hidden state decay factor is calculated based on the consecutive missing time lengths corresponding to each biomarker in the time interval matrix, and the hidden state at the previous detection time point is decayed according to the hidden state decay factor. The detection values ​​of each biomarker at the current detection time point and the hidden state at the previous detection time point after hidden state decay are input to a gated loop unit, which updates the hidden state at the current detection time point. The detection values ​​of each biomarker at the current detection time point are the actual detection values ​​for observed biomarkers and the adaptively filled detection values ​​for unobserved biomarkers. The multiple detection time points are traversed, and a hidden state sequence consisting of the hidden states corresponding to each detection time point is output.

[0071] In this embodiment, the time-series coding module can be implemented using a time-decay network structure based on a gated recurrent unit. This network structure is not designed by this application, but is an existing network structure in the prior art. It is an improved solution based on the standard gated recurrent unit, addressing the common problems of missing values ​​and irregular sampling in multivariate time series. This application directly adopts the aforementioned existing network structure as an optional implementation of the time-series coding module.

[0072] Specifically, the temporal coding module processes the temporal detection data sequentially according to the detection time points. For each detection time point, the temporal coding module first obtains the row corresponding to the current detection time point in the mask matrix, and determines the observation status of each biomarker at the current detection time point based on the values ​​of each element in that row. For biomarkers not observed at the current detection time point, the temporal coding module calculates an input attenuation factor based on the element values ​​in the time interval matrix corresponding to the biomarker and the current detection time point, i.e., the length of time the biomarker has been continuously missing up to the current detection time point. The value of the input attenuation factor decreases as the length of the continuous missing time increases, thereby reducing the impact of the biomarker's fill value on the current time when the continuous missing time is long.

[0073] Then, the temporal encoding module multiplies the input decay factor by a preset fill value to achieve adaptive filling of missing detection values. For biomarkers observed at the current detection time point, the temporal encoding module directly uses the corresponding actual detection value. Simultaneously, the temporal encoding module calculates the hidden state decay factor based on the time interval matrix, obtains the time interval between the current detection time point and the previous detection time point, and calculates the hidden state decay factor based on this time interval. The value of the hidden state decay factor decreases as the time interval increases, thereby reducing the contribution of historical information in the hidden state of the previous detection time point to the current moment when the interval between adjacent detection time points is long.

[0074] Then, the temporal encoding module multiplies the hidden state decay factor element-wise with the hidden state of the previous detection time point to obtain the hidden state of the previous detection time point after hidden state decay. After completing the above input decay processing and hidden state decay processing, the temporal encoding module inputs the detection values ​​of each biomarker at the current detection time point and the hidden state of the previous detection time point after hidden state decay to the gated loop unit. The detection values ​​of each biomarker at the current detection time point are the actual detection values ​​for observed biomarkers and the adaptively filled detection values ​​for unobserved biomarkers. The gated loop unit, according to the calculation mechanism of its update and reset gates, nonlinearly combines the input value of the current detection time point with the decayed hidden state of the previous detection time point to calculate and output the hidden state of the current detection time point. The temporal encoding module performs the above operations sequentially for each of the multiple detection time points. After traversal, the temporal encoding module outputs a hidden state sequence composed of the hidden states corresponding to each detection time point arranged in chronological order.

[0075] It should be noted that although the temporal coding network structure used in this step and its application in clinical assessment scenarios are known technologies, the technical contribution of this application does not lie in the independent design of this network structure or its conventional application, but in the organic connection of this network structure with the subsequent first deep learning module (neural frequent differential equation network) and second deep learning module (heterogeneous graph attention network) to construct a complete data processing flow of "temporal coding → continuous time evolution → cross-marker coupling modeling → dynamic threshold output". Specifically, the temporal coding module plays a front-end coding role in the overall process of this application, that is, converting the original irregular sampling temporal data into a hidden state sequence to provide the initial state for the subsequent continuous time evolution. In this overall process, the hidden state sequence output by the temporal coding module is not the final processing result of this application, but serves as intermediate data for the first deep learning module to perform continuous time evolution, and then is fused with the dynamic threshold matrix output by the second deep learning module to form the final state assessment result.

[0076] Therefore, by introducing the mask matrix and the time interval matrix, the time-series coding module can explicitly perceive which biomarkers were observed, which were not observed, and for how long each biomarker has been continuously missing at each detection time point. Through the input decay mechanism, the time-series coding module can adaptively fill in the missing detection values, causing the filled values ​​to decay towards the global mean over time. Through the hidden state decay mechanism, the time-series coding module can adaptively adjust the degree of retention of historical information according to the time interval. Thus, the time-series coding module can directly process irregularly sampled time-series data without converting the data into an equally spaced sequence through interpolation or other means in the data preprocessing stage, avoiding the information distortion that may be introduced by the above preprocessing operations.

[0077] Optionally, the first deep learning module performs continuous-time evolution of the hidden states based on a neural network and predefined prior terms, and outputs the expected concentration time-series estimates of each biomarker within a future preset time window. This includes: taking the last hidden state in the hidden state sequence as the initial state; constructing a differential equation, which includes a neural network part and a prior term part, wherein the neural network part takes latent variables, time, and the static features of the target object as input, and the prior term part includes an elimination rate constant term and a generation rate constant term, wherein the elimination rate constant term is calculated based on the half-life of each biomarker, and the generation rate constant term is predefined according to the HOUR theory; starting from the initial state, numerically integrating the differential equation using an adaptive step-size ordinary differential equation solver, and outputting the latent variable states at each time point within the future preset time window; mapping the latent variable states at each time point within the future preset time window back to the original data space to obtain the expected concentration time-series estimates of each biomarker within the future preset time window.

[0078] In this embodiment, the first deep learning module is implemented using a neural network of ordinary differential equations. This network describes the rate of change of latent variable states in the continuous time domain, so as to transform the sequence of latent states at discrete time points into a dynamic process in the continuous time domain, thereby breaking through the limitation of discrete time step and realizing the prediction of any future time point.

[0079] Since the last hidden state in the hidden state sequence has incorporated the historical information of all detection time points in the time-series detection data through the cyclic update mechanism of the aforementioned time-series coding module, the first deep learning module uses the last hidden state in the hidden state sequence as the initial state for continuous time evolution.

[0080] To perform continuous-time evolution, the first deep learning module first constructs a differential equation describing the rate of change of the latent variable state in the continuous-time domain. Unlike general neural network differential equations, this application does not rely solely on a purely data-driven neural network in the differential equation, but also introduces prior terms predefined based on the half-life information of various biomarkers in HOUR theory. These prior terms are specifically constructed for the half-life information in HOUR theory, rather than being a general design. HOUR theory provides the elimination half-life of various biomarkers under normal physiological conditions; for example, the half-life of C-reactive protein is approximately 19 hours, the half-life of procalcitonin is approximately 20 to 24 hours, and the half-life of fibrinogen is approximately 100 hours. These half-lives reflect the rate at which each biomarker is cleared in vivo and are the direct basis for the design of the prior terms in this scheme.

[0081] The differential equation consists of two parts: a neural network part and a prior term part. The neural network part takes the latent variable, the current time, and the static features of the target object as inputs, and calculates the rate of change of the latent variable state through forward propagation of the neural network. This neural network part adopts a multilayer perceptron structure, and its learnable parameters are obtained through training data. It is used to learn the complex dynamic patterns of biomarker concentration changes from the data, compensating for the inability of pure prior models to adapt to individual differences.

[0082] The prior terms include an elimination rate constant and a generation rate constant. The elimination rate constant is calculated based on the half-life of each biomarker: for each biomarker, the elimination rate constant is equal to the logarithm of the natural constant 2 divided by the half-life of that biomarker. Biomarkers with shorter half-lives have larger elimination rate constants, indicating faster elimination in vivo; biomarkers with longer half-lives have smaller elimination rate constants, indicating slower elimination in vivo. The generation rate constant is predefined according to the HOUR theory and describes the baseline generation rate of the biomarker. For example, for C-reactive protein, under healthy physiological conditions, its serum concentration remains at a relatively stable baseline level of approximately 1 mg / L. This corresponds to an elimination half-life of approximately 19 hours for C-reactive protein (and an elimination rate constant of approximately 0.0365 h). -1 Its base generation rate can be set as the product of the elimination rate constant and the baseline concentration, which is approximately 0.0365 mg·L⁻¹. -1 ·h -1 Similarly, for procalcitonin, its healthy baseline concentration is much lower than that of C-reactive protein (approximately 0.05 μg / L), and the corresponding basal production rate is also correspondingly lower. The aforementioned production rate constant term, in conjunction with the elimination rate constant term in the differential equation, causes the latent variable state to tend towards this baseline level under undisturbed conditions, thus providing a physiological reference point for the model during time intervals lacking observational data. The production rate constants of different biomarkers can be independently set based on their respective theoretical half-life and baseline concentration.

[0083] The prior term applies a physiological constraint to the evolution of the latent variable state. As the latent variable state evolves over time, the elimination rate constant continuously exerts a constraint tending towards elimination, ensuring that the evolution of the latent variable state does not deviate from the known elimination dynamics during time intervals lacking observational data. In other words, the prior term ensures that the model's predictions remain physiologically reasonable in data-sparse regions.

[0084] After the differential equation is constructed, the first deep learning module, starting from the initial state, uses an adaptive step-size ordinary differential equation solver to numerically integrate the differential equation within a preset future time window. The adaptive step-size ordinary differential equation solver can automatically adjust the integration step size according to the behavior of the equation itself while ensuring solution accuracy, adapting to different time scale requirements. Through the above numerical integration, the first deep learning module outputs the latent variable states at each time point within the preset future time window.

[0085] After obtaining the latent variable states at each time point within a future preset time window, the first deep learning module projects these latent variable states from the latent variable space back to the original data space through a mapping layer, thereby obtaining the expected concentration estimates of each biomarker at each time point within the future preset time window. These expected concentration estimates constitute the expected concentration time-series estimates. The length of the future preset time window can be preset according to the actual application scenario.

[0086] Thus, through the continuous-time evolution of the first deep learning module, the ability to output predicted values ​​at any future time point is achieved, overcoming the limitation of discrete time steps. Simultaneously, by introducing prior terms based on HOUR theory half-life information, the continuous-time prediction process incorporates known physiological prior knowledge, ensuring that the prediction results remain physiologically plausible even in data-sparse regions or time intervals without observational data. The obtained time-series estimates of the expected concentration will be provided to subsequent heterogeneous graph construction steps as the current node features of each node.

[0087] It should be noted that the neural network part of the above differential equation has learnable parameters, while the elimination rate constant and generation rate constant in the prior term part are fixed terms predefined based on HOUR theory and do not participate in training. The training process of the first deep learning module is as follows: A retrospective dataset containing multiple samples is obtained, each sample including historical time-series detection data of the target object and its corresponding future actual detection value. During training, for each sample, the first deep learning module takes the last hidden state in the hidden state sequence as the initial state, integrates the differential equation forward on the time axis starting from the initial state, outputs the latent variable state at each time point within a future preset time window, and obtains the predicted concentration estimate of each biomarker through a mapping layer. The predicted concentration estimate is compared with the future actual detection value, and the difference between the two is calculated as the loss function value. Then, the first deep learning module uses the adjoint sensitivity method to calculate the gradient of the loss function value with respect to the learnable parameters, and updates the parameters of the neural network part according to the gradient. Through multiple iterations of training, the neural network part gradually learns a dynamic pattern that matches the HOUR theory prior term. The loss function uses root mean square error or mean absolute error to measure the difference between the predicted value and the true value.

[0088] Optionally, the second deep learning module is a heterogeneous graph attention network; and the operation of inputting the current node features of each node into the second deep learning module, updating the feature representation of each node through an attention mechanism, and outputting a dynamic threshold matrix based on the updated node features includes: inputting the current node features of each node into the heterogeneous graph attention network; calculating the attention weight between each node and its neighboring nodes for each edge type through a node-level attention mechanism, and aggregating the features of each neighboring node according to the attention weight to obtain the aggregated features of each node under that edge type; calculating the importance weight of each edge type through a semantic-level attention mechanism, and weighting and fusing the aggregated features of each node under different edge types according to the importance weight to obtain the updated feature representation of each node; and outputting a dynamic threshold matrix based on the updated feature representation of each node.

[0089] In this embodiment, the second deep learning module is implemented using a heterogeneous graph attention network. This network is configured to perform message passing based on the heterogeneous graph constructed in step S4 to update the feature representations of each node. Unlike general graph neural networks that treat different neighbors equally or rely on predefined aggregation functions, the heterogeneous graph attention network adaptively learns the importance weights of each neighbor node to the current node through an attention mechanism, enabling the model to focus on neighbor nodes with greater information content. Furthermore, unlike general heterogeneous graph attention networks, the heterogeneous graph node types and edge types in this application are not statistically learned from data, nor are they obtained from external bioinformatics databases. Instead, message passing is performed based on the edge types predefined in step S4 according to HOUR theory. That is, the attention mechanism operates under the guidance of the physiological interaction relationships defined by HOUR theory, which is a product of the fusion of HOUR theory and the heterogeneous graph attention network.

[0090] The current node features of each node in the heterogeneous graph constructed in step S4 are input into the heterogeneous graph attention network. The heterogeneous graph attention network first executes a node-level attention mechanism. For each edge type, the heterogeneous graph attention network independently calculates the attention weights between each node and its neighboring nodes under that edge type. Specifically, for a target node, the heterogeneous graph attention network transforms and concatenates the features of the target node and the features of each of its neighboring nodes, obtains the attention scores of each neighboring node after activation function processing, and then obtains the attention weights through normalization. These attention weights reflect the importance of the corresponding neighboring nodes to the target node; the larger the weight, the stronger the physiological interaction between the neighboring node and the target node, and the greater its influence on the target node. Subsequently, the heterogeneous graph attention network performs a weighted summation of the features of each neighboring node according to the attention weights to obtain the aggregated features of the target node under that edge type. For each node, the above process is executed independently for each edge type, thereby obtaining multiple aggregated features of the node under each edge type. For example, if a node is connected to both inflammation-immunity edges and immunity-coagulation edges, then after node-level attention is completed, the node will obtain two aggregated features, which correspond to the information transmitted by the two types of edges respectively.

[0091] After node-level attention is completed, the heterogeneous graph attention network further executes a semantic-level attention mechanism. Since the physiological interactions carried by different edge types have varying importance to the state assessment task—for example, the information conveyed by the inflammation-coagulation edge, "inflammatory factors affect coagulation function," may be more critical than that of the inflammation-immunity edge—the network needs to learn the importance weights for each edge type. Specifically, the heterogeneous graph attention network performs a nonlinear transformation on the aggregated features of each node under various edge types, performs a dot product operation with the learnable semantic-level attention vector, and then normalizes the results to obtain the importance weights for each edge type. Subsequently, the network weights and fuses the aggregated features of each node under different edge types according to these importance weights to obtain the updated feature representation of each node. This semantic-level attention mechanism enables the network to adaptively determine the contribution of different physiological interactions to state assessment.

[0092] After the node-level attention and semantic-level attention processes described above, the feature representation of each node is updated to a new feature representation that incorporates information from its different types of neighboring nodes. That is, the feature of each node not only includes information about its own markers, but also information about other markers with which it has physiological interactions.

[0093] After obtaining the updated feature representations of each node, the heterogeneous graph attention network outputs a dynamic threshold matrix based on these features. Specifically, for each biomarker, the heterogeneous graph attention network maps the node features to a K-dimensional space through a fully connected layer based on the updated features of the node corresponding to that biomarker, and then normalizes them using a softmax function to output the probability distribution of the biomarker falling into each of the K preset classification intervals. The preset classification intervals are intervals pre-divided based on the concentration levels of each biomarker to indicate the risk level, and the boundaries of the classification intervals for different biomarkers can be set independently.

[0094] Based on the aforementioned marginal probability distributions of each biomarker, the heterogeneous graph attention network further generates a joint probability distribution of each biomarker pair falling into each preset hierarchical interval, forming a dynamic threshold matrix. Since the features of each node dynamically change with different input data, the dynamic threshold matrix also dynamically changes with different target objects and different detection time points.

[0095] The learnable parameters in the aforementioned heterogeneous graph attention network (including the transformation matrix and attention vector in node-level attention, the semantic vector in semantic-level attention, and the weight matrix and bias terms of the fully connected layer) are learned through training data. During training, the heterogeneous graph attention network uses the expected concentration time-series estimate output in step S3 as the initial features of each node, performs message passing as described above, and outputs a dynamic threshold matrix. The cross-entropy between the output probability distribution and the true classification label is used as the loss function, and the learnable parameters are updated through the backpropagation algorithm.

[0096] Thus, through a node-level attention mechanism, the feature representations of each node can adaptively acquire information from different neighboring nodes, and the magnitude of the attention weights directly indicates the strength of physiological interactions, providing interpretability for the model. Through a semantic-level attention mechanism, the network can adaptively determine the contribution of different physiological interactions in the HOUR theory to state assessment. By outputting the grading results in the form of a probability distribution, this step replaces the fixed-threshold hard-judgment method, allowing the grading thresholds of each biomarker to adaptively adjust according to the differences in input data. The final output dynamic threshold matrix will be provided to subsequent steps as one of the inputs for generating state assessment results.

[0097] It should be noted that the learnable parameters in the aforementioned heterogeneous graph attention network include the transformation matrix and attention vector in node-level attention, the semantic vector in semantic-level attention, and the weight matrix and bias terms of the fully connected layers. The training process of this network is as follows: A retrospective dataset containing multiple samples is obtained, each sample including historical time-series detection data of the target object and its corresponding biomarker classification labels. During training, for each sample, the heterogeneous graph attention network uses the expected concentration time-series estimate output in step S3 as the initial feature of each node, performs node-level attention and semantic-level attention as described above, obtains the updated feature representation of each node, and outputs the probability distribution of each biomarker falling into each preset classification interval based on the updated features. The probability distribution is compared with the classification labels, and the cross-entropy between them is calculated as the loss function value. Then, the heterogeneous graph attention network uses the backpropagation algorithm to calculate the gradient of the loss function value with respect to the learnable parameters, and updates the learnable parameters according to the gradient. Through multiple iterations of training, the heterogeneous graph attention network gradually learns the optimal attention weights and classification parameters under the graph structure defined by HOUR theory.

[0098] Optionally, the operation of outputting a dynamic threshold matrix based on the updated feature representations of each node includes: mapping the updated feature representations of each node to a K-dimensional space through a fully connected layer to obtain the initial scores of each node in each dimension; processing the initial scores of each node using a softmax function to obtain the probability distribution of each biomarker falling into each of the K preset grading intervals; generating a dynamic threshold matrix based on the probability distribution of each biomarker falling into each preset grading interval; the dynamic threshold matrix is ​​a D×D×K×K dimensional tensor, where D is the total number of biomarkers and K is the number of preset grading intervals; the index in the tensor is (i,j,k). i ,k j The element represents the i-th biomarker falling into the k-th element. i There are grading intervals, and the j-th biomarker falls within the k-th interval. j The joint probability of the hierarchical intervals, wherein the joint probability is equal to the i-th biomarker falling into the k-th interval. i The probability of each grading interval and the probability of the j-th biomarker falling into the k-th interval are related. j The product of the probabilities of each graded interval.

[0099] In this embodiment, the process from the updated feature representation of each node to the generation of the dynamic threshold matrix is ​​as follows: First, for each node corresponding to a biomarker, the heterogeneous graph attention network maps the updated feature representation of that node to a K-dimensional space via a fully connected layer, obtaining the initial score of the node in each dimension. The fully connected layer shares weights across all nodes, with its input dimension being the dimension of the updated feature representation of the node, and its output dimension being K. The weight matrix and bias terms of this fully connected layer are learnable parameters. The K dimensions correspond to K preset grading intervals.

[0100] Then, the initial scores of each node in the heterogeneous graph attention network are processed by a softmax function, converting the initial scores in each dimension into probability values, such that the sum of the output probabilities of each dimension is 1. This yields the probability distribution of each biomarker falling into each of the K preset classification intervals. Each dimension of the K-dimensional probability vector corresponding to each biomarker has a value between 0 and 1, and the sum of the values ​​of each dimension is 1. The preset classification intervals are pre-divided intervals based on the concentration levels of each biomarker, used to indicate risk levels. The boundaries of the classification intervals for different biomarkers can be set independently. For example, for C-reactive protein, a low-risk range of less than 10 mg / L, a medium-risk range of 10 to 100 mg / L, and a high-risk range of greater than 100 mg / L can be set; for procalcitonin, a low-risk range of less than 0.5 μg / L, a medium-risk range of 0.5 to 2.0 μg / L, and a high-risk range of greater than 2.0 μg / L can be set; for D-dimer, a low-risk range of less than 0.5 mg / L FEU, a medium-risk range of 0.5 to 5.0 mg / L FEU, and a high-risk range of greater than 5.0 mg / L FEU can be set. The boundary values ​​of the grading ranges for different biomarkers vary due to their respective physiological characteristics and clinical significance; this application allows for the independent setting of grading range boundaries for each biomarker.

[0101] Finally, the heterogeneous graph attention network generates a dynamic threshold matrix based on the probability distribution of each biomarker falling into each preset hierarchical interval. The dynamic threshold matrix is ​​a D×D×K×K dimensional tensor. Specifically, the indices in the tensor are (i,j,k). i ,k j The element represents the i-th biomarker falling into the k-th element. i There are grading intervals, and the j-th biomarker falls within the k-th interval. j The joint probability of the grading intervals. This joint probability is defined as the probability that the i-th biomarker falls into the k-th interval. i The marginal probability of each grading interval and the fact that the j-th biomarker falls into the k-th interval. j The product of the marginal probabilities of each graded interval. This indicates that the i-th biomarker falls into the k-th category. i The probability of each graded interval, in This indicates that the j-th biomarker falls into the k-th category. j The probability of each graded interval is given, and the index in the dynamic threshold matrix is ​​(i,j,k). i ,k j The element value of ) Therefore, the dynamic threshold matrix expands the marginal grading probability information of each individual biomarker into joint grading probability information between all biomarker pairs. The diagonal elements (when i=j) reflect the joint probability of the same biomarker falling into two grading intervals simultaneously, while the off-diagonal elements (when i≠j) reflect the joint grading probability information across different biomarkers. This tensor completely stores the joint probability distribution between all possible biomarker pairs and is a numerical representation of the coupled grading state of multiple biomarkers.

[0102] Thus, through fully connected layer mapping and softmax normalization, the updated features of each node are transformed into edge-level probability distributions with clear probabilistic meaning. By constructing a D×D×K×K dimensional tensor, the edge probability distribution of each individual marker is extended to a joint probability distribution among all marker pairs, fully expressing the coupled hierarchical information across markers. The resulting dynamic threshold matrix will be provided to subsequent steps as one of the inputs for generating the state evaluation results.

[0103] Optionally, the operation of generating a corresponding state assessment result based on the expected concentration time-series estimate and the dynamic threshold matrix includes: extracting the terminal value and average rate of change from the expected concentration time-series estimate, and obtaining a concentration trend feature vector through linear projection; the terminal value is the expected concentration value at the last moment of a future preset time window, and the average rate of change is the average rate of change of the expected concentration value within the future preset time window; marginalizing the dynamic threshold matrix along non-target dimensions, and extracting the marginal probability vector of each biomarker being in the highest grade interval; weighting and fusing the concentration trend feature vector and the marginal probability vector through learnable gating parameters to obtain an abnormal deviation comprehensive index vector; wherein, the weighted fusion operation is: for For each biomarker, the normalized value of the corresponding component in the concentration trend feature vector and the corresponding component in the marginal probability vector are assigned learnable weights and then summed in a weighted manner. A six-dimensional axial deviation vector is generated based on the anomaly deviation comprehensive index vector. This six-dimensional axial deviation vector is obtained by weighting and aggregating the anomaly deviation comprehensive index vector according to the six dimensions of the HOUR theory, including neural, endocrine, inflammatory, immune, coagulation, and metabolic bioenergy dimensions. The expected concentration time-series estimate, the dynamic threshold matrix, the six-dimensional axial deviation vector, and the probability distribution of each biomarker belonging to each preset grade interval are encapsulated to obtain the state assessment result.

[0104] In this embodiment, after obtaining the expected concentration time-series estimate and the dynamic threshold matrix, these two pieces of information are integrated to generate the final state assessment result. The expected concentration time-series estimate includes predicted concentration changes of each biomarker within a preset time window, reflecting the future trends of each biomarker. The dynamic threshold matrix includes the joint probability distribution information of each biomarker falling into each preset grading interval, reflecting the current cross-marker coupling grading state. Specifically, it includes: First, a concentration trend feature vector is extracted from the expected concentration time-series estimates. For each biomarker, the terminal value (i.e., the expected concentration value at the last moment of the future preset time window) and its average rate of change are extracted within a future preset time window, i.e., the average rate of change of the expected concentration value relative to time within the future preset time window. The terminal value and the average rate of change are concatenated into a 2D vector, and then mapped to a D-dimensional vector through a linear projection layer to obtain the concentration trend feature vector.

[0105] Simultaneously, the dynamic threshold matrix is ​​edge-processed along non-target dimensions to extract the edge probabilities of each biomarker falling within the highest-level interval, forming an edge probability vector. Specifically, for each biomarker, the index i of the biomarker and its corresponding highest-level interval index are fixed in the dynamic threshold matrix. (That is, the highest risk level among the preset risk levels) =K), along another marker dimension j and the corresponding hierarchical interval dimension Summing these values ​​yields the marginal probability that the biomarker is in the highest risk range. Iterating through the D biomarkers yields a D-dimensional marginal probability vector. This vector represents the probability that each biomarker is currently at the highest risk level.

[0106] Then, the concentration trend feature vector and the marginal probability vector are weighted and fused using a learnable gating parameter to obtain an anomaly deviation comprehensive index vector. Specifically, for each biomarker, the normalized value of the corresponding component in the concentration trend feature vector and the corresponding component in the marginal probability vector are assigned learnable weights and then weighted and summed. The normalized value is obtained by mapping the components in the concentration trend feature vector to the interval between 0 and 1. The learnable gating parameter is a D-dimensional vector, where each dimension controls the relative contribution of the concentration trend feature and marginal probability feature of the corresponding biomarker during fusion. The weighted fusion operation is as follows: for the i-th biomarker, its anomaly deviation comprehensive index is equal to the value of the sigmoid function applied to the i-th dimension of the gating parameter multiplied by the normalized value of the i-th component of the concentration trend feature vector, plus 1 and minus the value of the sigmoid function applied to the i-th dimension of the gating parameter multiplied by the value of the i-th component of the marginal probability vector. The resulting anomaly deviation comprehensive index vector fuses the future trend information of each biomarker and the current cross-marker coupling hierarchical probability information.

[0107] Further, a six-dimensional axial deviation vector is generated based on the anomaly deviation comprehensive index vector. This six-dimensional axial deviation vector is obtained by weighting and aggregating the anomaly deviation comprehensive index vector according to the six dimensions of the HOUR theory. Specifically, each of the multiple biomarkers is pre-assigned to one of the six dimensions according to the HOUR theory: neural dimension, endocrine dimension, inflammation dimension, immune dimension, coagulation dimension, and metabolic bioenergy dimension. The corresponding components of the biomarkers belonging to the same HOUR dimension in the anomaly deviation comprehensive index vector are weighted and aggregated to obtain the axial deviation value for that HOUR dimension. Traversing the six dimensions yields the six-dimensional axial deviation vector, which reflects the comprehensive deviation of the target object from the six system axes within the HOUR theory framework.

[0108] After obtaining the comprehensive evaluation vector, the expected concentration time-series estimate, the dynamic threshold matrix, the six-dimensional axial deviation vector, and the probability distribution of each biomarker falling into each preset grade interval are structured and encapsulated to form the state evaluation result. Specifically, the encapsulation adopts a structured data format and includes the following fields: the expected concentration time-series estimate field, storing the predicted concentration curve of each biomarker within a future preset time window; the dynamic threshold matrix field, storing the joint probability distribution of each biomarker falling into each preset grade interval; the six-dimensional axial deviation vector field, storing the comprehensive deviation score of the target object in the six HOUR dimensions; and the marginal probability distribution field of each biomarker, storing the probability distribution of each biomarker falling into each preset grade interval. This state evaluation result is multi-dimensional numerical evaluation data, does not contain any diagnostic conclusions or treatment recommendations, and can be read and applied by downstream devices or professionals.

[0109] Therefore, by generating corresponding state assessment results based on the expected concentration time-series estimates and the dynamic threshold matrix, the assessment results integrate future trend predictions and cross-biomarker coupling probability information. Specifically, the concentration trend feature vector characterizes the future direction and magnitude of change for each biomarker, while the marginal probability vector characterizes the probability that each biomarker is currently at its highest risk level. These two are adaptively fused using learnable gating parameters, allowing the fusion weights to be dynamically adjusted according to data characteristics. The six-dimensional axial deviation vector aggregates the abnormal deviation information of each biomarker according to the six dimensions of the HOUR theory, providing comprehensive deviation information at the system axis level and offering a comprehensive quantitative basis for the integrated state assessment of multi-biomarker time-series data.

[0110] It should be noted that the learnable gating parameters involved in the weighted fusion process described above, as well as the aggregation weights involved in the calculation of the six-dimensional axial deviation vector, are all learnable parameters. The training process is as follows: A retrospective dataset containing multiple samples is obtained, each sample including historical time-series detection data of the target object and its corresponding actual state assessment label. During training, for each sample, features are extracted from the expected concentration time-series estimate and the dynamic threshold matrix in the manner described above. After weighted fusion and axial aggregation, a six-dimensional axial deviation vector is obtained. The difference between this vector and the actual state assessment label is used as the loss function value, and the learnable parameters are updated using the backpropagation algorithm. Through multiple iterations of training, the gating parameters gradually learn the optimal weight allocation of concentration trend features and marginal probability features during fusion.

[0111] In summary, this application has the following beneficial effects: (1) The time-series coding module adaptively updates the hidden state based on the time interval between each detection time point, providing an adaptive coding foundation for irregularly sampled multi-biomarker time-series data without interpolation preprocessing. Furthermore, by introducing a mask matrix and a time interval matrix, the time-series coding module explicitly perceives which biomarkers are observed, which are not observed, and how long each biomarker has been continuously missing at each detection time point. This enables it to directly process time-series data sampled at unequal intervals, avoiding information distortion that may be introduced by interpolation preprocessing.

[0112] (2) The latent state is continuously evolved in time by the first deep learning module based on a neural network and predefined prior terms. The prior terms are defined based on the half-life information of various biomarkers in the HOUR theory, which integrates known physiological prior knowledge into the continuous-time prediction process. Furthermore, the prior terms include elimination rate constant and generation rate constant. The elimination rate constant is calculated based on the half-life of various biomarkers, ensuring that the evolution of the latent variable state does not deviate from the known elimination dynamics in time intervals lacking observation data. This improves the predictive rationality in sparse data regions and enables the output of predicted values ​​at any future time point, breaking through the limitation of discrete time steps.

[0113] (3) By constructing a heterogeneous graph using the aforementioned biomarkers as nodes and the physiological interactions between them as edges, and defining edge types according to the HOUR theory, the modeling of cross-marker coupling relationships is both physiologically interpretable and easy to deploy. The edge types of the heterogeneous graph are entirely based on the predefined HOUR theory, and the entire construction process does not rely on any external bioinformatics database, allowing it to run stably in different deployment environments. Simultaneously, the expected concentration time-series estimates are used as the current node features of each node, integrating future trend information into the initial features of each node, providing a structured data foundation for subsequent graph attention processing.

[0114] (4) The feature representations of each node are updated through the attention mechanism of the second deep learning module, and a dynamic threshold matrix is ​​output based on the updated node features. This allows the grading thresholds of each biomarker to be dynamically and adaptively adjusted according to individual differences and time sequence, and the grading results are output in the form of a probability distribution. Through the node-level attention mechanism, the feature representations of each node can adaptively obtain information from different neighboring nodes; through the semantic-level attention mechanism, the network can adaptively determine the contribution of different physiological interactions in the HOUR theory to the state assessment. The dynamic threshold matrix is ​​a D×D×K×K dimensional tensor, which completely stores the joint probability distribution between all possible biomarker pairs, realizing the replacement of the fixed threshold hard judgment method, and at the same time enabling the grading results to reflect the uncertainty in the state assessment.

[0115] (5) By generating corresponding state assessment results based on the expected concentration time-series estimates and the dynamic threshold matrix, the state assessment results integrate future trend prediction and cross-biomarker coupling probability information. Furthermore, a concentration trend feature vector is obtained by extracting the terminal value and average rate of change from the expected concentration time-series estimates, and an edge probability vector of each biomarker being in the highest-level interval is extracted from the dynamic threshold matrix. These two are adaptively fused using learnable gating parameters to obtain an abnormal deviation comprehensive index vector, which in turn generates a six-dimensional axial deviation vector. The final state assessment results are multi-dimensional numerical assessment data, containing information on both the dynamic trend dimension and the coupling risk dimension, providing a comprehensive quantitative basis for the integrated state assessment of multi-biomarker time-series data.

[0116] In addition, refer to Figure 1 As shown, according to a second aspect of this embodiment, a storage medium is provided. The storage medium includes a stored program, wherein, when the program is executed, a processor performs any of the methods described above.

[0117] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0118] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0119] Example 2

[0120] Figure 3 A temporal state evaluation system based on HOUR theory and deep learning according to this embodiment is shown, which corresponds to the method described in Embodiment 1. (Reference) Figure 3As shown, the system includes: a detection data acquisition module 310, used to acquire time-series detection data of multiple biomarkers of a target object; the time-series detection data includes detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; a hidden state sequence construction module 320, used to input the time-series detection data to a time-series encoding module, the time-series encoding module adaptively updates the hidden state according to the time interval between each detection time point, and outputs a hidden state sequence composed of the hidden states corresponding to each detection time point; and a future concentration prediction module 330, used to input the hidden state sequence to a first deep learning module, the first deep learning module performing continuous time evolution of the hidden state based on a neural network and predefined prior terms, and outputting the expected concentration time-series estimate of each biomarker within a future preset time window; the prior terms are based on HOUR theory. The study defines the half-life information of various biomarkers; a heterogeneous graph construction module 340 is used to construct a heterogeneous graph with the various biomarkers as nodes and the physiological interactions between the various biomarkers as edges, according to the edge type defined by the HOUR theory, and uses the expected concentration time series estimate as the current node feature of each node; a dynamic threshold matrix construction module 350 is used to input the current node features of each node into a second deep learning module, the second deep learning module updates the feature representation of each node through an attention mechanism, and outputs a dynamic threshold matrix based on the updated node features; the elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset grading interval; the preset grading interval is used to indicate the risk level; a state assessment module 360 ​​is used to generate a corresponding state assessment result based on the expected concentration time series estimate and the dynamic threshold matrix.

[0121] It should be noted that the time-series state evaluation system based on HOUR theory and deep learning provided in this embodiment can realize all the functions and steps in the above method embodiments, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated here.

[0122] Example 3

[0123] Figure 4 A temporal state evaluation system based on HOUR theory and deep learning according to this embodiment is shown, which corresponds to the method described in Embodiment 1. (Reference) Figure 4As shown, the system includes: a processor 410; and a memory 420 connected to the processor 410, used to provide the processor 410 with instructions to process the following steps: acquiring time-series detection data of multiple biomarkers of a target object; the time-series detection data includes detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; inputting the time-series detection data to a time-series encoding module, the time-series encoding module adaptively updating the hidden state according to the time interval between each detection time point, and outputting a hidden state sequence composed of the hidden states corresponding to each detection time point; inputting the hidden state sequence to a first deep learning module, the first deep learning module performing continuous time evolution of the hidden state based on a neural network and predefined prior terms, and outputting the biomarkers within a future preset time window. The expected concentration time-series estimate is obtained; the prior term is defined based on the half-life information of various biomarkers in the HOUR theory; a heterogeneous graph is constructed with the various biomarkers as nodes and the physiological interactions between the various biomarkers as edges, according to the edge type defined by the HOUR theory, and the expected concentration time-series estimate is used as the current node feature of each node; the current node feature of each node is input into the second deep learning module, the second deep learning module updates the feature representation of each node through an attention mechanism, and outputs a dynamic threshold matrix based on the updated node features; the elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset classification interval; the preset classification interval is used to indicate the risk level; based on the expected concentration time-series estimate and the dynamic threshold matrix, the corresponding state assessment result is generated.

[0124] It should be noted that the time-series state evaluation system based on HOUR theory and deep learning provided in this embodiment can realize all the functions and steps in the above method embodiments, solve the same technical problems, and achieve the same technical effects. The similarities will not be repeated here.

[0125] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0126] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0127] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The system embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between units or modules, and may be electrical or other forms.

[0128] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0129] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0130] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0131] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A temporal state evaluation method based on HOUR theory and deep learning, characterized in that, include: Acquire time-series detection data of multiple biomarkers for the target object; The time-series detection data includes detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; The timing detection data is input to the timing encoding module, which adaptively updates the hidden state according to the time interval between each detection time point and outputs a sequence of hidden states consisting of the hidden states corresponding to each detection time point. The hidden state sequence is input into the first deep learning module. The first deep learning module performs continuous time evolution on the hidden state based on the neural network and predefined prior terms, and outputs the expected concentration time series estimate of each biomarker within a future preset time window. The prior terms are defined based on the half-life information of each biomarker in the HOUR theory. Using the aforementioned biomarkers as nodes and the physiological interactions between them as edges, a heterogeneous graph is constructed based on the edge type defined by the HOUR theory, and the expected concentration time-series estimate is used as the current node feature of each node. The current node features of each node are input into the second deep learning module. The second deep learning module updates the feature representation of each node through an attention mechanism and outputs a dynamic threshold matrix based on the updated node features. The elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset grading interval. The preset grading interval is used to indicate the risk level. Based on the expected concentration time-series estimate and the dynamic threshold matrix, the corresponding state assessment result is generated.

2. The method according to claim 1, characterized in that, Before inputting the timing detection data into the timing coding module, a preprocessing step is also included: A mask matrix and a time interval matrix are constructed based on the time-series detection data. The mask matrix is ​​used to indicate whether each biomarker has been observed at each detection time point, and the time interval matrix is ​​used to indicate the length of time during which the detection values ​​of each biomarker have been continuously missing up to each detection time point.

3. The method according to claim 2, characterized in that, The operation of inputting the time-series detection data into the time-series encoding module, which adaptively updates the hidden states according to the time interval between each detection time point, and outputs a sequence of hidden states consisting of the hidden states corresponding to each detection time point, includes: The timing detection data, the mask matrix, and the time interval matrix are input into the timing encoding module; the timing encoding module is a time decay network based on a gated cyclic unit. Based on the mask matrix, determine whether each biomarker has been observed at the current detection time point; For unobserved biomarkers, an input decay factor is calculated based on the corresponding consecutive missing time length in the time interval matrix, and the detection value of the biomarker is adaptively filled based on the input decay factor; for observed biomarkers, the corresponding detection value is used directly. The hidden state decay factor is calculated based on the continuous missing time length corresponding to each biomarker in the time interval matrix, and the hidden state at the previous detection time point is decayed based on the hidden state decay factor. The detection values ​​of each biomarker at the current detection time point and the hidden state of the previous detection time point after the hidden state decay are input into the gated loop unit, and the hidden state of the current detection time point is updated through the gated loop unit; wherein, the detection values ​​of each biomarker at the current detection time point are the actual detection values ​​for biomarkers that have been observed, and the detection values ​​for unobserved biomarkers after adaptive filling. Iterate through the multiple detection time points and output a sequence of hidden states consisting of the hidden states corresponding to each detection time point.

4. The method according to claim 1, characterized in that, The first deep learning module, based on a neural network and predefined prior terms, performs continuous-time evolution of the hidden state and outputs the expected concentration time-series estimates of each biomarker within a preset future time window, including: The last hidden state in the hidden state sequence is taken as the initial state; A differential equation is constructed, which includes a neural network part and a prior term part. The neural network part takes latent variables, time and the static features of the target object as inputs. The prior term part includes an elimination rate constant term and a generation rate constant term. The elimination rate constant term is calculated based on the half-life of each biomarker, and the generation rate constant term is predefined according to the HOUR theory. Starting from the initial state, the differential equation is numerically integrated using an adaptive step-size ordinary differential equation solver, and the latent variable states at each time point within a future preset time window are output. By mapping the state of latent variables at each time point within a future preset time window back to the original data space, the expected concentration time series estimates of each biomarker within the future preset time window are obtained.

5. The method according to claim 1, characterized in that, The second deep learning module is a heterogeneous graph attention network; and the operation of inputting the current node features of each node into the second deep learning module, updating the feature representation of each node through an attention mechanism, and outputting a dynamic threshold matrix based on the updated node features includes: The current node features of each node are input into the heterogeneous graph attention network; By using a node-level attention mechanism, for each edge type, the attention weight between each node and its neighboring nodes is calculated, and the features of each neighboring node are aggregated according to the attention weight to obtain the aggregated features of each node under that edge type. The importance weights of various edge types are calculated through a semantic attention mechanism, and the aggregated features of each node under different edge types are weighted and fused according to the importance weights to obtain the updated feature representations of each node. Based on the updated feature representations of each node, a dynamic threshold matrix is ​​output.

6. The method according to claim 5, characterized in that, The operations for outputting a dynamic threshold matrix based on the updated feature representations of each node include: The updated feature representations of each node are mapped to a K-dimensional space through a fully connected layer to obtain the initial scores of each node in each dimension; The initial scores of each node are processed by the softmax function to obtain the probability distribution of each biomarker falling into each of the K preset grading intervals. A dynamic threshold matrix is ​​generated based on the probability distribution of each biomarker falling into each preset grading interval; the dynamic threshold matrix is ​​a D×D×K×K dimensional tensor, where D is the total number of biomarkers and K is the number of preset grading intervals; the tensor has indices (i,j,k). i ,k j The element represents the i-th biomarker falling into the k-th element. i There are grading intervals, and the j-th biomarker falls within the k-th interval. j The joint probability of the hierarchical intervals, wherein the joint probability is equal to the i-th biomarker falling into the k-th interval. i The probability of each grading interval and the probability of the j-th biomarker falling into the k-th interval are related. j The product of the probabilities of each graded interval.

7. The method according to claim 1, characterized in that, The operation of generating corresponding state assessment results based on the expected concentration time-series estimate and the dynamic threshold matrix includes: The terminal value and average rate of change are extracted from the expected concentration time series estimate, and a concentration trend feature vector is obtained by linear projection; the terminal value is the expected concentration value at the last moment of the future preset time window, and the average rate of change is the average rate of change of the expected concentration value within the future preset time window. The dynamic threshold matrix is ​​edged along the non-target dimension, and the edge probability vector of each biomarker in the highest grade interval is extracted; The concentration trend feature vector and the marginal probability vector are weighted and fused using learnable gating parameters to obtain an abnormal deviation comprehensive index vector. The weighted fusion operation is as follows: for each biomarker, the normalized value of the corresponding component in the concentration trend feature vector and the corresponding component in the marginal probability vector are respectively assigned learnable weights and then summed. Based on the abnormal deviation comprehensive index vector, a six-dimensional axial deviation vector is generated; the six-dimensional axial deviation vector is obtained by weighting and aggregating the abnormal deviation comprehensive index vector according to the six dimensions of the HOUR theory, which include the neural dimension, endocrine dimension, inflammation dimension, immune dimension, coagulation dimension, and metabolic bioenergy dimension. The expected concentration time-series estimate, the dynamic threshold matrix, the six-dimensional axial deviation vector, and the probability distribution of each biomarker belonging to each preset grade interval are encapsulated to obtain the state assessment result.

8. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, the method described in any one of claims 1 to 7 is performed by a processor.

9. A time-series state evaluation system based on HOUR theory and deep learning, characterized in that, include: The detection data acquisition module is used to acquire time-series detection data of multiple biomarkers of the target object; The time-series detection data includes detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; The hidden state sequence construction module is used to input the time-series detection data into the time-series coding module. The time-series coding module adaptively updates the hidden state according to the time interval between each detection time point and outputs a hidden state sequence composed of the hidden states corresponding to each detection time point. The future concentration prediction module is used to input the hidden state sequence into the first deep learning module. The first deep learning module performs continuous time evolution on the hidden state based on the neural network and predefined prior terms, and outputs the expected concentration time series estimate of each biomarker within a future preset time window. The prior terms are defined based on the half-life information of each biomarker in the HOUR theory. The heterogeneous graph construction module is used to construct a heterogeneous graph with the multiple biomarkers as nodes and the physiological interactions between the multiple biomarkers as edges, and to define the edge types according to the HOUR theory, and to use the expected concentration time series estimate as the current node feature of each node. A dynamic threshold matrix construction module is used to input the current node features of each node into a second deep learning module. The second deep learning module updates the feature representation of each node through an attention mechanism and outputs a dynamic threshold matrix based on the updated node features. The elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset grading interval. The preset grading interval is used to indicate the risk level. The state assessment module is used to generate corresponding state assessment results based on the expected concentration time-series estimate and the dynamic threshold matrix.

10. A time-series state evaluation system based on HOUR theory and deep learning, characterized in that, include: processor; A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: Acquire time-series detection data of multiple biomarkers of a target object; the time-series detection data includes detection values ​​obtained by detecting at least one of the multiple biomarkers at multiple detection time points; The timing detection data is input to the timing encoding module, which adaptively updates the hidden state according to the time interval between each detection time point and outputs a sequence of hidden states consisting of the hidden states corresponding to each detection time point. The hidden state sequence is input into the first deep learning module. The first deep learning module performs continuous time evolution on the hidden state based on the neural network and predefined prior terms, and outputs the expected concentration time series estimate of each biomarker within a future preset time window. The prior terms are defined based on the half-life information of each biomarker in the HOUR theory. Using the aforementioned biomarkers as nodes and the physiological interactions between them as edges, a heterogeneous graph is constructed based on the edge type defined by the HOUR theory, and the expected concentration time-series estimate is used as the current node feature of each node. The current node features of each node are input into the second deep learning module. The second deep learning module updates the feature representation of each node through an attention mechanism and outputs a dynamic threshold matrix based on the updated node features. The elements of the dynamic threshold matrix are the probability distribution of each biomarker pair falling into a preset grading interval. The preset grading interval is used to indicate the risk level. Based on the expected concentration time-series estimate and the dynamic threshold matrix, the corresponding state assessment result is generated.

Images