Method and system for state assessment based on risk indicators and multi-dimensional biomarkers

Abstract

This invention relates to a state assessment method and system based on RISK indicators and multidimensional biomarkers. The method includes: acquiring four-modal time-series data, including four types of data: organism response, organ damage, vital signs, and organ function; processing the first modal data using an event-triggered first encoder, updating the output only when data arrives, to obtain a first feature vector; using the first feature vector as a constraint, processing the second modal data using a second encoder, with the update amplitude decreasing as the output increases and degradation requiring multiple confirmations, to obtain a second feature vector; using the second feature vector as a constraint, performing multi-scale statistical deviation detection on the third modal data using a statistical state estimator, to obtain a third feature vector; using the third feature vector as a constraint, performing time-series trend analysis on the fourth modal data using a trend state estimator, to obtain a fourth feature vector; inputting the four feature vectors into four scoring heads to obtain four state scores and determining the state assessment result.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a state assessment method and system based on RISK indicators and multidimensional biomarkers. Background Technology

[0002] In the field of data processing technology, the comprehensive evaluation of multi-source time-series data is a technical direction with broad application needs. Taking the monitoring scenario as an example, clinical testing and monitoring equipment generates various types of time-series data, such as, but not limited to, biomarker data obtained through blood tests, organ damage marker data obtained through blood or body fluid tests, vital sign parameter data obtained through bedside monitoring equipment, and organ function parameter data obtained through regular testing.

[0003] Different types of time-series data exhibit inherent temporal relationships. Specifically, changes in biomarkers typically occur earliest, reflecting the initial response at the system level; changes in organ damage markers follow closely, reflecting the degree of disturbance to individual organ systems at the cellular level; changes in vital signs parameters occur after changes in the above two types of markers, reflecting the disturbed state of system homeostasis; and changes in organ function parameters typically occur last, reflecting changes in the overall functional level of the system. These temporal relationships constitute the inherent temporal structure among the four types of data. Theoretically, if these four types of data can be systematically integrated and evaluated according to their temporal structure, a more comprehensive evaluation result should be obtained than if only a single type or partial type of data is relied upon.

[0004] However, most existing technologies rely on data processing solutions for status assessment based on a single type of data. For example, patent application CN121964149A discloses a health status assessment method that acquires a user's physiological characteristic data within a preset period, constructs tensor data containing individual, characteristic, and time dimensions, uses frequency domain transformation and singular value decomposition to complete missing data, and inputs it into a health status assessment model to obtain the assessment result. This solution only uses physiological characteristic data for assessment and does not introduce other types of data. Patent application CN120884270A discloses a non-contact microwave monitoring method and system for the entire process management of physiological parameters. It extracts human vital sign parameters for assessment by collecting microwave signals of user physiological parameters. This solution also only assesses human vital sign parameters, reflecting a single-dimensional state information. Patent application CN121737286A discloses a method, system and related biomarkers for evaluating the efficacy of treatment for a disease. The evaluation result is determined by obtaining the test substance in a first state and a second state, detecting the biomarkers therein and comparing them. This scheme only uses biomarker data for evaluation and also only reflects a single dimension of state information.

[0005] Existing technologies also include a few schemes that perform fusion evaluation based on partial types of data. For example, patent application CN120526971A discloses a method for identifying abnormal medical information, which acquires real-time patient vital sign data and real-time equipment operation status data, obtains basic patient information data and medical resource information data based on the hospital's information management system, and constructs a comprehensive objective function to train a fusion prediction model. Although this scheme integrates multiple types of information, it does not involve biomarker data and organ damage marker data, and the information dimensions covered are still limited. Patent application CN121938636A discloses an ICU multi-organ failure risk prediction and prevention system, which constructs a cross-organ temporal causal intensity tracking module to calculate the causal influence intensity and effect delay time between organ function indicators within a sliding time window. This scheme only evaluates based on organ function indicators and does not include biomarker data and organ damage marker data in the evaluation system, thus only covering partial dimensions of information.

[0006] Therefore, existing technologies employ assessment methods based on single or partial types of data. Their results only reflect one or a partial aspect of the target object's state, lacking complete information and failing to comprehensively reflect the full picture. Thus, there is a need in this field for a technical solution that can systematically integrate four types of data—biomarker data, organ damage marker data, vital sign parameter data, and organ function parameter data (each corresponding to different state dimensions)—to achieve a comprehensive assessment of the target object's state. Summary of the Invention

[0007] In view of the shortcomings of the prior art, the purpose of the invention is to provide a state assessment method and system based on RISK indicators and multidimensional biomarkers, so as to solve the technical problem in the prior art that the assessment information is incomplete and cannot fully reflect the overall state of the target object because only a single type or partial type of data is used for assessment.

[0008] In a first aspect, the present invention proposes a state assessment method based on RISK indicators and multidimensional biomarkers, comprising: acquiring four-modal time-series data of a target object, wherein the four-modal time-series data includes: first time-series data related to bodily response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function; processing the first time-series data using a first encoder to output a first feature vector; the first encoder updates its output only when data arrives, and keeps its output unchanged when no data arrives; using the first feature vector as a context constraint, processing the second time-series data using a second encoder to output a second feature vector; and the update magnitude of the second encoder... The value decreases as the output value increases, and the downgrade update requires multiple consecutive detection confirmations; using the second feature vector as a context constraint, a statistical state estimator is used to process the third time series data, outputting a third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales; using the third feature vector as a context constraint, a trend state estimator is used to process the fourth time series data, outputting a fourth feature vector; the trend state estimator determines the fourth feature vector based on time series trend analysis; the first to fourth feature vectors are respectively input into the first to fourth scoring heads to obtain four state scoring values, and the state evaluation result of the target object is determined based on the four state scoring values.

[0009] According to a second aspect of the present disclosure, a storage medium is provided, the storage medium including a stored program, wherein, when the program is executed, a processor performs the method described in any of the above embodiments.

[0010] According to a third aspect of the present disclosure, a state assessment system based on RISK indicators and multidimensional biomarkers is provided, comprising: a data acquisition module for acquiring four-modal time-series data of a target object, the four-modal time-series data including: first time-series data related to bodily response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function; a first processing module for processing the first time-series data using a first encoder and outputting a first feature vector; the first encoder updates its output only when data arrives, and keeps its output unchanged when no data arrives; a second processing module for processing the second time-series data using a second encoder with the first feature vector as a context constraint and outputting a second feature vector; the second encoder updates its output only when data arrives, and keeps its output unchanged when no data arrives; and a second processing module for processing the second time-series data using a second encoder with the first feature vector as a context constraint and outputting a second feature vector. The new amplitude decreases as the output value increases, and the downgrade update requires confirmation through multiple consecutive checks; the third processing module is used to process the third time series data using a statistical state estimator with the second feature vector as a context constraint, and outputs a third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales; the fourth processing module is used to process the fourth time series data using a trend state estimator with the third feature vector as a context constraint, and outputs a fourth feature vector; the trend state estimator determines the fourth feature vector based on time series trend analysis; the state evaluation module is used to input the first to fourth feature vectors into the first to fourth scoring heads respectively to obtain four state score values, and determine the state evaluation result of the target object based on the four state score values.

[0011] According to a fourth aspect of the present disclosure, a state assessment system based on RISK metrics and multidimensional biomarkers 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 four-modal time-series data of a target object, the four-modal time-series data including: first time-series data related to bodily response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function; processing the first time-series data using a first encoder to output a first feature vector; the first encoder updating its output only when data arrives, and maintaining its output unchanged when no data arrives; and processing the second time-series data using a second encoder with the first feature vector as a context constraint. The second encoder outputs a second feature vector; the update magnitude of the second encoder decreases as the output value increases, and the downgrade update needs to be confirmed by multiple consecutive detections; using the second feature vector as a context constraint, a statistical state estimator is used to process the third time series data, and outputs a third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales; using the third feature vector as a context constraint, a trend state estimator is used to process the fourth time series data, and outputs a fourth feature vector; the trend state estimator determines the fourth feature vector based on time series trend analysis; the first to fourth feature vectors are respectively input into the first to fourth scoring heads to obtain four state scoring values, and the state evaluation result of the target object is determined based on the four state scoring values.

[0012] This application acquires four-modal time-series data of the target object, covering information from four different dimensions: bodily response, organ damage, vital signs, and organ function, providing a complete data foundation for comprehensively assessing the target object's state. Through the event-triggered mechanism of the first encoder, the output is updated only when data arrives, remaining unchanged when no data arrives, matching the encoder's behavior with the sparse and abrupt changes in biomarker data sampling. Through the asymmetric update mechanism of the second encoder, the update amplitude is suppressed when the output increases, and multiple consecutive checks are required to confirm the output during degradation, matching the encoder's update behavior with the asymmetric changes of organ damage markers, which show rapid increases and slow decreases. A statistical state estimator processes densely sampled vital sign data based on statistical deviation detection at multiple time scales, effectively extracting the key information of deviation from steady state while suppressing the interference of normal physiological fluctuations on the assessment results. A trend state estimator, based on time... Sequential trend analysis processes organ function parameter data, extracting trend features such as the slope of change and cumulative deviation, accurately reflecting the cumulative change characteristics of organ function parameters. Through hierarchical contextual passing—processing the second time-series data with the first feature vector as contextual constraint, the third time-series data with the second feature vector as contextual constraint, and the fourth time-series data with the third feature vector as contextual constraint—the inherent sequential relationships between the four data types are embedded as structural constraints into the evaluation process. This ensures that the processing of each subsequent dimension is conditional on the evaluation results of the previous dimension, structurally guaranteeing that the evaluation results conform to the inherent sequential relationships between the four data types. By mapping each feature vector to four independent state score values ​​through the first to fourth scoring heads, and determining the state evaluation result of the target object based on these four state score values, a multi-dimensional and structured evaluation result output is achieved. This ensures that the evaluation results possess overall integrity while retaining fine-grained information for each dimension. Thus, a comprehensive and accurate evaluation of the target object's state is realized. Attached Figure Description

[0013] 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.

[0014] 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 state assessment method based on RISK indicators and multidimensional biomarkers according to Embodiment 1 of this application; Figure 3This is a schematic diagram of the framework of the state assessment method based on RISK indicators and multidimensional biomarkers according to Embodiment 1 of this application; Figure 4 This is a schematic diagram of the state assessment system based on RISK indicators and multidimensional biomarkers according to Embodiment 2 of this application; Figure 5 This is a schematic diagram of the state assessment system based on RISK indicators and multidimensional biomarkers according to Embodiment 3 of this application. Detailed Implementation

[0015] 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.

[0016] 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.

[0017] Terminology Explanation: RISK: A four-dimensional quantitative index system used in this application to evaluate the state of the target object, wherein: R stands for Response, which represents the organism's response / host dysregulation response and is used to reflect the initial response changes of the target object at the system level; I stands for Organ Injury, which indicates organ damage and is used to reflect the degree of damage to various organ systems of the target object at the cellular level. S stands for Changes of Vital Signs, which indicates changes in vital signs and is used to reflect the real-time steady-state state of the target object's circulatory, respiratory, and other systems. K stands for "Killing and Killed" Organ Appears, indicating organ failure and reflecting the cumulative trend of changes in the functional levels of various organs in the target subject.

[0018] Example 1

[0019] According to this embodiment, a method embodiment for state assessment based on RISK indicators and multidimensional biomarkers 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. Also, 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.

[0020] 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 state assessment method based on RISK metrics and multidimensional biomarkers 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.

[0021] 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).

[0022] 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 state assessment method based on RISK indicators and multidimensional biomarkers 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 state assessment method based on RISK indicators and multidimensional biomarkers of the aforementioned application. 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.

[0023] 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.

[0024] 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.

[0025] 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.

[0026] It should be noted that the state assessment method based on RISK indicators and multidimensional biomarkers provided in this application belongs to computer-implemented information processing technology. Its direct output results are four state scores, a total score, and a hierarchical identifier, rather than a definitive disease diagnosis conclusion. Essentially, this application achieves a multidimensional quantitative assessment of the target object's state through a series of information processing steps, including event-triggered encoding, memory-lock encoding, multi-scale statistical deviation detection, and time-series trend analysis of the four-modal time-series data. Each modal data is mapped into a feature vector, which is then mapped into four independent state scores via a scoring head, and the final assessment result is obtained by combining these features. (The encoder and scoring head are trained using machine learning techniques.) The entire process involves only numerical calculations and feature transformations of the multimodal time-series data, without direct contact with the human body or any invasive procedures. Furthermore, the final assessment result is intended to provide structured state codes 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.

[0027] Under the above operating environment, according to the first aspect of this embodiment, a state assessment method based on RISK indicators and multidimensional biomarkers is provided. Figure 2 A flowchart illustrating the method is shown. Figure 3 A schematic diagram of the framework of this method is shown, for reference. Figure 2 and Figure 3 As shown, the method includes: S1: Obtain four-modal time-series data of the target object, the four-modal time-series data including: first time-series data related to body response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function.

[0028] In this embodiment, the four-modal time-series data originates from multiple types of time-series data generated during the detection and monitoring of the target object. The target object is the object from which the data is collected. The four types of time-series data come from different data sources: the first and second time-series data are obtained through detection data generated after detecting samples of the target object; the third time-series data are obtained through real-time monitoring data collected by bedside monitoring equipment; and the fourth time-series data are obtained through detection result data generated after periodic detection of the target object. Each modal time-series data is recorded according to its respective sampling timestamp.

[0029] Specifically, the indicators included in the first time-series data are related to the body's response, reflecting the initial response changes of the target object at the system level; these indicators typically show the earliest changes in time. The indicators included in the second time-series data are related to organ damage, reflecting the degree of disturbance to various organ systems of the target object at the cellular level; these indicators typically appear after changes in indicators related to the body's response. The indicators included in the third time-series data are related to vital signs, reflecting the real-time status of the target object's circulatory, respiratory, and other systems; this data is continuously collected through monitoring equipment and typically appears after changes in indicators related to the body's response and organ damage. The indicators included in the fourth time-series data are related to organ function, reflecting the overall functional level of various organs of the target object; these indicators typically appear last in time.

[0030] The four types of time-series data mentioned above correspond to the four dimensions of the RISK index: data related to bodily response corresponds to the R dimension, data related to organ damage corresponds to the I dimension, data related to vital signs corresponds to the S dimension, and data related to organ function corresponds to the K dimension. The RISK index is a set of multi-dimensional quantitative indicators used to assess the state of a target object, composed of the indicators from the above four dimensions. By mapping the four types of time-series data to their corresponding RISK dimensions, a multi-dimensional quantitative assessment of the target object's state can be achieved. The four types of time-series data exhibit an inherent temporal correlation: changes in bodily response data occur earliest, followed by changes in organ damage data; changes in vital signs data occur after the changes in the first two types of data; and changes in organ function data occur last. This temporal correlation is determined by the inherent temporal structure of the target object's state evolution.

[0031] Therefore, by acquiring four-modal time-series data and establishing its correspondence with the four dimensions of RISK, a complete data foundation is provided for subsequent quantitative assessment of the target object's state from four dimensions: bodily response, organ damage, vital signs, and organ function.

[0032] S2: The first encoder is used to process the first time series data and output the first feature vector; the first encoder only updates the output when data arrives, and keeps the output unchanged when no data arrives.

[0033] In this embodiment, the sampling interval of the first time-series data is typically 6 to 24 hours, with long periods of no data between adjacent detection times, and the numerical changes often exhibit abrupt rather than smooth gradual changes. In this case, interpolation or state speculation during the no-data intervals will introduce unrealistic intermediate values, leading to distorted feature extraction. To address this issue, the first encoder employs an event-triggered operating mode: it only performs a forward calculation and updates its output when new data is detected arriving in the first time-series data; when no new data arrives, the first encoder does not perform calculations, and its output remains unchanged from the previous time-series value. The first encoder processes the first time-series data and outputs a first feature vector, which serves as the feature representation of the R dimension (organism response) in the RISK metric.

[0034] Therefore, based on the characteristics of sparse sampling (sampling interval is usually 6 to 24 hours) and abrupt changes of the first time series data, the first encoder was designed in a targeted manner, so that the output first feature vector remains unchanged when no data arrives, avoiding the non-real information introduced by interpolation filling in the no-data interval in the prior art, and ensuring the authenticity of the extraction of body response dimension features.

[0035] S3: Using the first feature vector as a context constraint, the second encoder is used to process the second time series data and output the second feature vector; the update amplitude of the second encoder decreases as the output value increases, and the downgrade update needs to be confirmed by multiple consecutive detections.

[0036] In this embodiment, the sampling interval of the second time-series data is typically 12 to 24 hours, and its index value exhibits asymmetric characteristics of rapid increase and slow decrease, meaning that the index value does not quickly return to its original level after increasing. If the encoder allows the output value to decrease rapidly in sync with the decrease in a single detection value after increasing, it will lead to frequent misjudgments.

[0037] Furthermore, as mentioned earlier, the four types of time-series data exhibit an inherent temporal correlation: changes in data related to bodily responses precede changes in data related to organ damage. That is, in the RISK index, changes in the R dimension precede changes in the I dimension. Therefore, when assessing the I dimension state, incorporating the R dimension assessment results as contextual information into the encoding process ensures that the I dimension state assessment is based on the R dimension assessment results, thus maintaining causal consistency and logical order among the four RISK dimensions. Without this contextual constraint, encoding solely based on the second time-series data itself cannot guarantee alignment between the encoded results and the causal relationship of the R dimension, potentially leading to the learning of feature combinations that violate the aforementioned temporal correlation.

[0038] To address the aforementioned issues, the second encoder uses the first feature vector as a context constraint and employs an asymmetric update mechanism: the update magnitude of the second encoder is determined by its current output value; when the current output value is high, the update magnitude automatically decreases (i.e., it becomes less sensitive to new input changes), and when the current output value is low, the update magnitude automatically increases (i.e., it becomes more sensitive to new input changes). Furthermore, for degradation updates (i.e., output values ​​changing from high to low), the second encoder does not respond immediately upon a single detection of a decrease in value; instead, it requires multiple consecutive detections for confirmation before performing a degradation update. The second encoder processes the second time-series data and outputs a second feature vector, which serves as the feature representation for dimension I (organ damage) in the RISK metric.

[0039] Therefore, based on the asymmetric variation characteristics of the second time series data, the second encoder was specifically designed to suppress the update amplitude when the output value increases and to require multiple confirmations when the output value decreases. This improves the stability of feature extraction in the organ damage dimension and avoids misjudgment caused by fluctuations in a single detection value. At the same time, by introducing the first feature vector as a context constraint, the feature extraction in the I dimension is based on the evaluation results in the R dimension, maintaining the causal consistency and logical order among the four dimensions of RISK and avoiding the learning of feature combinations that violate the inherent sequential relationship between the four types of data.

[0040] S4: Using the second feature vector as a context constraint, a statistical state estimator is used to process the third time series data and output a third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales.

[0041] In this embodiment, the sampling frequency of the third time series data is relatively high, typically on the order of seconds or minutes. The data volume is dense and fluctuates continuously, with many fluctuations being physiological fluctuations within the normal range and not representing actual deviations from the state. If point-by-point sequence modeling is directly performed on the densely sampled time series data, not only will the computational load be large, but it is also easy to misjudge normal fluctuations as state changes.

[0042] Furthermore, as mentioned earlier, the four types of time-series data have an inherent sequential relationship: changes in organ injury-related data precede changes in vital sign-related data. That is, in the RISK index, changes in dimension I precede changes in dimension S. Therefore, when assessing the state of dimension S, incorporating the assessment results of dimension I as contextual information into the statistical deviation detection process ensures that the state assessment of dimension S is based on the assessment results of dimension I, thus maintaining causal consistency and logical order among the four dimensions of RISK. Without this contextual constraint, the judgment threshold used for statistical deviation detection will be fixed and cannot be dynamically adjusted according to the state of the upstream dimension I. This could lead to oversensitivity to normal vital sign fluctuations before organ injury occurs, or the use of excessively high judgment thresholds after organ injury has occurred, failing to identify abnormal deviations in vital signs in a timely manner.

[0043] To address the aforementioned issues, the statistical state estimator uses the second feature vector as a contextual constraint and employs a multi-timescale statistical deviation detection approach: the second feature vector is introduced as contextual information into the statistical deviation detection process, detecting the degree of deviation of the current data from the statistical distribution at each preset time scale, and determining the third feature vector based on the detection results. The statistical state estimator processes the third time-series data and outputs the third feature vector, which serves as the feature representation of the S dimension (vital signs) in the RISK metric.

[0044] Therefore, based on the dense sampling (second-level or minute-level) and continuous fluctuation characteristics of the third time series data, a targeted design for the statistical state estimator was carried out using multi-time-scale statistical deviation detection. This effectively extracted the key information of deviation from steady state while suppressing the interference of normal physiological fluctuations on the evaluation results. By introducing a second feature vector as a contextual constraint, the feature extraction of the S dimension is based on the evaluation results of the I dimension, maintaining the causal consistency and logical order among the four dimensions of RISK, and avoiding oversensitivity or response lag caused by using a fixed judgment threshold.

[0045] S5: Using the third feature vector as a context constraint, the trend state estimator is used to process the fourth time series data and output the fourth feature vector; the trend state estimator determines the fourth feature vector based on time series trend analysis.

[0046] In this embodiment, the sampling interval for the fourth time-series data is typically 1 to 6 hours. The changes in its index values ​​are the result of continuous accumulation over multiple moments. The instantaneous value at a single moment cannot accurately reflect the actual direction of state evolution; the trend of change is a better indicator of the true direction of state evolution than the instantaneous value. If only the instantaneous value of a single detection is considered and its trend over time is ignored, the cumulative change characteristics of this type of data cannot be accurately captured.

[0047] Furthermore, as mentioned earlier, the four types of time-series data have an inherent sequential relationship: changes in vital sign-related data precede changes in organ function-related data. That is, in the RISK index, changes in the S dimension precede changes in the K dimension. Therefore, when assessing the K dimension state, incorporating the S dimension assessment results as contextual information into the trend analysis process ensures that the K dimension state assessment is based on the S dimension assessment results, thus maintaining causal consistency and logical order among the four RISK dimensions. Without this contextual constraint, the decision threshold used in trend analysis will be fixed and cannot be dynamically adjusted according to the upstream S dimension state. This could lead to oversensitivity to subtle trends in organ function indicators when vital signs are still stable, or the use of excessively high decision thresholds when vital signs are already significantly disordered, failing to identify changes in organ function trends in a timely manner.

[0048] To address the aforementioned issues, the trend state estimator uses the third eigenvector as a contextual constraint and employs a time-series trend analysis approach: the third eigenvector is introduced as contextual information into the trend analysis process; trend analysis is performed on the fourth time-series data within a preset time window to extract the direction of change and cumulative degree of each indicator; and the fourth eigenvector is determined based on the trend analysis results. The trend state estimator processes the fourth time-series data and outputs the fourth eigenvector, which serves as the feature representation of the K dimension (organ function) in the RISK index.

[0049] Therefore, based on the sampling interval (usually 1 to 6 hours) and cumulative change characteristics of the fourth time series data, a trend state estimator was specifically designed using time series trend analysis. By extracting trend features such as the slope of change and the cumulative deviation, the cumulative change characteristics of the organ function dimension data were accurately reflected. By introducing a third feature vector as a contextual constraint, the feature extraction of the K dimension was based on the evaluation results of the S dimension, maintaining the causal consistency and logical order among the four dimensions of RISK, and avoiding oversensitivity or response lag caused by using a fixed judgment threshold.

[0050] S6: Input the first to fourth feature vectors into the first to fourth scoring heads respectively to obtain four state scoring values, and determine the state evaluation result of the target object based on the four state scoring values.

[0051] In this embodiment, the first, second, third, and fourth scoring heads receive the first, second, third, and fourth feature vectors, respectively, and map each of them to a state score value, thus obtaining four independent state score values. These four score values ​​correspond to the R, I, S, and K dimensions of the RISK index, respectively quantifying the target object's state in four dimensions: bodily response, organ damage, vital signs, and organ function.

[0052] After obtaining four state scores, the state assessment result of the target object is determined based on these scores. This assessment result includes at least the score values ​​of each dimension and their comprehensive information, thereby achieving a multi-dimensional and structured assessment result output. This ensures that the assessment result is holistic while retaining fine-grained information for each dimension, enabling a multi-dimensional quantitative assessment of the target object's state.

[0053] As described in the background section, the evaluation methods used in the prior art are all based on evaluation schemes that use a single type of data or a partial type of data. Their evaluation results can only reflect one aspect or a partial aspect of the target object's state. The information dimensions are incomplete and cannot fully reflect the overall state of the target object.

[0054] In view of this, this application acquires four-modal time-series data of the target object, covering information from four different dimensions: bodily response, organ damage, vital signs, and organ function, providing a complete data foundation for comprehensively assessing the state of the target object; through the event-triggered mechanism of the first encoder, the output is updated only when data arrives and remains unchanged when no data arrives, matching the encoder's behavior with the sparse and abrupt changes in biomarker data sampling; through the asymmetric update mechanism of the second encoder, the update amplitude is suppressed when the output increases and multiple consecutive checks are required to confirm when it decreases, matching the encoder's update behavior with the data characteristics of organ damage markers, which show rapid increases and slow decreases; through a statistical state estimator based on statistical deviation detection at multiple time scales, the densely sampled vital sign data is processed, effectively extracting the key information of "deviation from steady state" while suppressing the interference of normal physiological fluctuations on the assessment results; through a trend state estimator... By processing organ function parameter data through time-series trend analysis, trend features such as the slope of change and cumulative deviation are extracted, accurately reflecting the cumulative change characteristics of organ function parameters. Through hierarchical contextual passing—processing the second time-series data with the first feature vector as contextual constraint, the third time-series data with the second feature vector as contextual constraint, and the fourth time-series data with the third feature vector as contextual constraint—the inherent sequential relationship between the four types of data is embedded as a structural constraint in the evaluation process. This ensures that the processing of each subsequent dimension is conditional on the evaluation results of the previous dimension, structurally guaranteeing that the evaluation results conform to the inherent sequential relationship between the four types of data. By mapping each feature vector to four independent state score values ​​through the first to fourth scoring heads, and determining the state evaluation result of the target object based on these four state score values, a multi-dimensional and structured evaluation result output is achieved. This ensures that the evaluation result retains fine-grained information of each dimension while possessing overall comprehensiveness. Thus, a comprehensive and accurate evaluation of the target object's state is realized.

[0055] Optionally, the first time-series data includes time-series detection values ​​of leukocytes, lymphocytes, procalcitonin, interleukin-6, D-dimer, serum amyloid A, interleukin-8, interleukin-10, fibrinogen, ferritin, and lactate dehydrogenase; the first encoder is a fully connected neural network employing an event-triggered mechanism; and the operation of processing the first time-series data with the first encoder to output a first feature vector includes: calculating the proportion of indicators with detected values ​​at the current time to all indicators; when the proportion is not lower than a preset threshold, inputting the values ​​of indicators with detected values ​​at the current time and the latest historical values ​​of missing indicators into the first encoder to obtain the first feature vector; when the proportion is lower than the preset threshold, keeping the first feature vector output at the previous time unchanged, and outputting a preset confidence level label.

[0056] In this embodiment, the first time-series data originates from the detection data generated after blood testing of the target subject, including time-series detection values ​​of 11 indicators: leukocytes, lymphocytes, procalcitonin, interleukin-6, D-dimer, serum amyloid A, interleukin-8, interleukin-10, fibrinogen, ferritin, and lactate dehydrogenase. These indicators relate to multiple pathophysiological pathways, such as inflammatory responses (e.g., indicators corresponding to interleukin-6, procalcitonin, and C-reactive protein), immune status (e.g., lymphocytes, interleukin-8, and interleukin-10), and coagulation function (e.g., D-dimer and fibrinogen), collectively constituting a comprehensive quantitative description of the body's response state.

[0057] The first encoder is a fully connected neural network employing an event-triggered mechanism. The fully connected neural network includes an input layer, at least one hidden layer, and an output layer; the number of neurons in the input layer corresponds to the preset maximum number of indicators in the first time-series data; the hidden layer uses the ReLU activation function, and the output layer uses linear activation; the output layer dimension is a preset feature dimension.

[0058] The first encoder performs the following operations upon event triggering: First, it calculates the proportion of indicators with detected values ​​at the current moment out of all indicators. This proportion reflects the information completeness of the first time-series data at the current moment. Then, it compares this proportion with a preset threshold. When the proportion is not lower than the preset threshold, it indicates that a sufficient number of indicators have provided new detected values ​​at the current moment, and the data completeness meets the processing requirements. In this case, the values ​​of indicators with detected values ​​at the current moment and the latest historical values ​​of missing indicators are input into a fully connected neural network for a forward propagation calculation to obtain the first feature vector at the current moment. When the proportion is lower than the preset threshold, it indicates that only a few indicators have provided new detected values ​​at the current moment, and the data completeness is insufficient to support reliable feature extraction. In this case, the first encoder keeps the first feature vector output from the previous moment unchanged to avoid deriving unreliable feature representations based on incomplete data, and outputs a preset confidence level label to indicate that the feature vector output at the current moment is a continuation value from the previous moment, and its reliability is lower than that of the output at a normal update moment.

[0059] Thus, by using the above 11 indicators to comprehensively cover the body's response state from multiple pathways such as inflammation, immunity, and coagulation, a complete data foundation is provided for the quantitative assessment of the R dimension. Through the data integrity gating mechanism, the first feature vector is updated only when the proportion of effective data meets the requirements, and historical values ​​are maintained and confidence labels are output when the proportion of data is insufficient. This avoids feature extraction based on a small amount of unreliable information when data is severely missing, and ensures the reliability and stability of R dimension feature extraction.

[0060] Optionally, the second time-series data includes time-series detection values ​​of troponin, soluble advanced glycation end product receptor, KL-6, neuron-specific enolase, neurofilament light chain, kidney injury molecule-1, IGFBP-7, citrulline, intestinal fatty acid-binding protein, alanine aminotransferase (ALT), and aspartate aminotransferase (AST); the second encoder is a gated loop unit with a continuous strength lock; the first feature vector serves as the initial hidden state of the gated loop unit and is input into the gated loop unit after being concatenated with the second time-series data at each time step; the hidden state update of the gated loop unit is constrained by the continuous strength lock, and the locking strength of the continuous strength lock is... The hidden state update of the gated loop unit increases as the current output score increases, satisfying the following: ; in, The hidden state is calculated by the standard gated loop unit. This is the hidden state from the previous moment. This represents the hidden state at the current moment; The conditions for the second encoder to downgrade and update are: the current detection value is lower than a preset threshold, and the values ​​of at least two consecutive previous detections are lower than the preset threshold, and the time interval between two adjacent detections is not less than the preset minimum confirmation interval.

[0061] In this embodiment, the second time-series data originates from detection data generated after blood or body fluid testing of the target subject, including time-series detection values ​​of 11 indicators: troponin, soluble advanced glycation end product receptor (AGER), KL-6, neuron-specific enolase, neurofilament light chain, kidney injury molecule-1 (RIM-1), IGFBP-7, citrulline, intestinal fatty acid-binding protein (IFB), alanine aminotransferase (ALT), and aspartate aminotransferase (AST). These indicators correspond to the injury states of multiple organ systems: troponin corresponds to cardiovascular system injury; soluble AGER and KL-6 correspond to alveolar injury; AGER and neurofilament light chain correspond to nervous system injury; IGFBP-7 and AST correspond to renal tubular injury; citrulline and IGFBP-7 correspond to intestinal barrier injury; and ALT and AST correspond to hepatobiliary injury. These indicators collectively constitute a comprehensive quantitative description of the degree of injury to multiple organ systems.

[0062] The second encoder is a gated loop unit with a continuous strength lock. The gated loop unit is configured according to a standard structure with an update gate, a reset gate, and a candidate hidden state calculation unit, and its hidden state dimension is the same as the dimension of the first feature vector.

[0063] The first feature vector is passed to the second encoder as a context constraint as follows: the first feature vector serves as the initial hidden state of the gated recurrent unit, establishing the initial state of the second encoder based on the R-dimensional evaluation result; simultaneously, at each time step, the first feature vector is concatenated with the second time-series data and input into the gated recurrent unit, ensuring that the R-dimensional evaluation result participates as conditional information in the I-dimensional encoding process at each time step. Through this dual injection method, a unidirectional causal constraint is achieved from the organism's response dimension to the organ damage dimension.

[0064] The second encoder constrains its hidden state updates using a continuous-strength lock. The locking strength of the continuous-strength lock... The value is a continuous value within the interval [0,1], which increases as the current output score increases. That is, the higher the current output score, the stronger the locking strength and the smaller the update magnitude. Specifically, the hidden state update of the gated loop unit satisfies: , The hidden state is calculated by the standard gated loop unit. This is the hidden state from the previous moment. This represents the hidden state at the current moment. When When = 0, the hidden state is completely updated; when When =1, the hidden state is completely locked and not updated; when 0 < When the value is less than 1, the hidden state is partially updated, and the locking strength gradually increases as the output score increases.

[0065] The conditions for the second encoder to downgrade and update are: the current detection value is lower than a preset threshold, and the values ​​of at least two consecutive previous detections are both lower than the preset threshold, with the time interval between two adjacent detections not less than a preset minimum confirmation interval. Downgrade updating is allowed when these conditions are met. The basis for these conditions is that an increase in organ damage markers often reflects existing tissue damage, and their decline requires a certain amount of time. A single decrease in the detection value is insufficient to confirm an actual improvement in the damage state; multiple detections are needed for confirmation before it can be considered that the damage state has indeed improved.

[0066] Thus, by covering multiple organ systems such as the cardiovascular, alveoli, nervous, renal tubules, intestines, and hepatobiliary systems with the aforementioned 11 indicators, a comprehensive data foundation is provided for quantitative assessment of dimension I. The dual injection method, using the first feature vector as the initial hidden state and concatenating it step-by-step over time, embeds the causal order constraint from the organism's response dimension to the organ damage dimension into the encoding process of the second encoder. Through a continuous intensity locking mechanism, the locking intensity gradually increases as the output score increases to suppress the update amplitude, matching the asymmetric change characteristics of organ damage markers. The condition setting requiring multiple consecutive tests for confirmation before downgrade updates avoids misjudging improvement due to a single decrease in detection value, ensuring the stability and reliability of dimension I feature extraction.

[0067] Optionally, the third time-series data includes time-series records of heart rate, mean arterial pressure, respiratory rate, body temperature, and consciousness status score; and, using the second feature vector as a context constraint, the operation of processing the third time-series data with a statistical state estimator to output a third feature vector includes: mapping the second feature vector to a first weighting coefficient, and correcting a preset baseline deviation judgment threshold according to the first weighting coefficient to obtain an actual deviation judgment threshold; calculating the mean and standard deviation of each indicator in the third time-series data at each time scale at at least two preset time scales; calculating the deviation of the value of the third time-series data at the current moment relative to the mean and standard deviation at each time scale; comparing the deviation of each indicator in the third time-series data with the actual deviation judgment threshold to determine whether each indicator is in an abnormal state at the current moment; and determining the third feature vector based on the accumulation of abnormal states within a preset time window; wherein, the third feature vector includes: the deviation of each indicator in the third time-series data at the current moment, the abnormal state marker at the current moment, and the accumulated information of abnormal states within the preset time window.

[0068] In this embodiment, the third time-series data originates from monitoring data generated after continuous monitoring of the target subject via bedside monitoring equipment. This data includes time-series records of five indicators: heart rate, mean arterial pressure, respiratory rate, body temperature, and consciousness status score. These indicators respectively reflect the target subject's circulatory status (heart rate, mean arterial pressure), respiratory function (respiratory rate), metabolic and inflammatory status (body temperature), and nervous system function (consciousness status score), collectively constituting a comprehensive quantitative description of the vital signs.

[0069] The statistical state estimator is an information processing module based on statistical computation. The statistical state estimator uses the second feature vector output by the second encoder as a context constraint, maps the second feature vector to a first weighting coefficient, and corrects the preset benchmark deviation judgment threshold according to the first weighting coefficient to obtain the actual deviation judgment threshold. That is, when the assessment result of the organ damage dimension is high, the deviation judgment threshold of the vital signs dimension is correspondingly reduced, making the statistical state estimator more sensitive to the detection of deviations in vital signs; conversely, when the assessment result of the organ damage dimension is low, the judgment threshold remains at a high level, making the statistical state estimator insensitive to normal fluctuations in vital signs.

[0070] The statistical state estimator calculates the statistical distribution parameters of each indicator in the third time series data at at least two preset time scales. These preset time scales include at least two of the following: short-scale (e.g., 5 minutes), medium-scale (e.g., 30 minutes), and long-scale (e.g., 2 hours). At each time scale, the mean and standard deviation of each indicator within the current time window are calculated. The mean reflects the central trend of the indicator at the current time scale, and the standard deviation reflects the fluctuation range of the indicator at the current time scale. Then, the deviation of the current value of each indicator from the mean and standard deviation at each time scale is calculated, i.e., the ratio of the difference between the current value and the mean to the standard deviation. This deviation reflects the degree of deviation of the current data from its historical statistical distribution. The calculated deviation is compared with the actual deviation judgment threshold. When the deviation exceeds the actual deviation judgment threshold, the indicator is determined to be in an abnormal state at the current time; otherwise, it is determined to be in a normal state. Based on the comparison results, an abnormal state label is generated for each indicator at the current time.

[0071] Furthermore, the statistical state estimator accumulates abnormal state information for each indicator within a preset time window. Specifically, the statistical state estimator accumulates information such as the duration and percentage of abnormal states for each indicator within the preset time window prior to the current moment. This is because a single deviation may be caused by transient disturbances, but a sustained deviation more reliably reflects a substantial state deviation. By accumulating abnormal state information, it is possible to distinguish between transient fluctuations and sustained deviations.

[0072] The statistical state estimator comprehensively determines and outputs a third feature vector based on the deviation of each indicator at the current moment, the abnormal state markers of each indicator, and the cumulative information of abnormal states within a preset time window. This third feature vector simultaneously includes the instantaneous deviation information at the current moment, the abnormal state judgment result, and the cumulative state information within the time window.

[0073] Thus, by comprehensively covering S-dimensional information through multi-dimensional vital sign parameters such as heart rate, mean arterial pressure, respiratory rate, body temperature, and consciousness status score, the system maps the second feature vector to weighting coefficients and corrects the deviation judgment threshold, enabling deviation detection in the vital sign dimension to be based on the assessment results of the organ damage dimension, thereby realizing the causal constraint transfer from the organ damage dimension to the vital sign dimension. Through multi-timescale statistical deviation detection, the statistical distribution of data is calculated and the degree of deviation is detected at multiple time scales, effectively extracting the deviation information of vital signs relative to their own steady-state distribution, while suppressing the interference of normal physiological fluctuations on the assessment results. By comparing the deviation degree with the threshold to generate abnormal state markers and accumulating them within the time window, the system effectively distinguishes between instantaneous deviations and continuous abnormalities, avoiding misjudgments caused by single instantaneous fluctuations.

[0074] Optionally, the fourth time-series data includes time-series records of oxygenation index, serum creatinine, bilirubin, urine output, and Glasgow Coma Scale score; and, using the third feature vector as a context constraint, the operation of processing the fourth time-series data with a trend state estimator to output the fourth feature vector includes: mapping the third feature vector to a second weighting coefficient, and correcting a preset baseline trend judgment threshold according to the second weighting coefficient to obtain an actual trend judgment threshold; fitting a linear trend to each indicator in the fourth time-series data within a preset time window to obtain the slope of change of each indicator; and calculating the slope of change of each indicator in the fourth time-series data. The cumulative deviation relative to the respective baseline values ​​within the preset time window; the trend state of each indicator is determined based on the change slope and cumulative deviation of each indicator in the fourth time series data; the trend state of each indicator is compared with the actual trend judgment threshold to determine whether the current moment is a trend abnormality state; the fourth feature vector is determined based on the accumulation of trend abnormality states within the preset time window; wherein, the fourth feature vector includes: the change slope of each indicator in the fourth time series data, the cumulative deviation of each indicator, the trend abnormality state marker at the current moment, and the cumulative information of trend abnormality states within the preset time window.

[0075] In this embodiment, the fourth time-series data originates from the test results generated after periodic monitoring of the target subject, including time-series records of five indicators: oxygenation index, serum creatinine, bilirubin, urine output, and Glasgow Coma Scale score. These indicators reflect the functional levels of various organ systems in the target subject: the oxygenation index reflects the oxygenation function of the lungs; serum creatinine and urine output reflect the filtration and excretion functions of the kidneys; bilirubin reflects the metabolic function of the liver; and the Glasgow Coma Scale score reflects the functional state of the nervous system. Together, these indicators constitute a comprehensive quantitative description of the functional levels of multiple organ systems.

[0076] The trend state estimator is a numerical computation-based information processing module. The trend state estimator uses the third feature vector output by the statistical state estimator as a context constraint, maps the third feature vector to a second weighting coefficient, and corrects the preset baseline trend judgment threshold based on this second weighting coefficient to obtain the actual trend judgment threshold. That is, when the assessment result of the vital signs dimension is high, the trend judgment threshold of the organ function dimension decreases accordingly, making the trend state estimator more sensitive to the changing trends of organ function indicators; conversely, when the assessment result of the vital signs dimension is low, the judgment threshold remains at a high level, making the trend state estimator insensitive to slight fluctuations in organ function indicators.

[0077] The trend state estimator performs linear trend fitting on each indicator in the fourth time series data within a preset time window to obtain the slope of change for each indicator. Specifically, within the preset time window (e.g., 6 to 12 hours), least-squares linear fitting is performed on multiple time series detection values ​​for each indicator. The slope obtained from the fitting reflects the direction and rate of change of the indicator within the time window: a slope greater than 0 indicates that the indicator is on an upward trend, a slope less than 0 indicates that the indicator is on a downward trend, and the larger the absolute value of the slope, the faster the rate of change.

[0078] Simultaneously, the trend state estimator calculates the cumulative deviation of each indicator in the fourth time series data relative to its respective baseline value within a preset time window. The baseline value is a reference value measured for the indicator during the system initialization phase or the historical stable phase. The cumulative deviation is the sum of the deviations of the detected values ​​at each time point within the time window from the baseline value, reflecting the overall degree to which the indicator deviates from the baseline level within the time window.

[0079] The trend state estimator comprehensively determines the trend state of each indicator based on its slope and cumulative deviation. Specifically, the slope reflects the current instantaneous direction and rate of change, while the cumulative deviation reflects the overall degree of deviation over a historical period. Combining these two factors allows for a comprehensive assessment of the evolution of each indicator: when neither the slope nor the cumulative deviation exceeds a threshold, the trend is considered normal; when either the slope or the cumulative deviation exceeds the actual trend determination threshold, the trend is considered abnormal. By comparing the trend state of each indicator with the actual trend determination threshold, it is determined whether the current moment is an abnormal trend state.

[0080] Furthermore, the trend state estimator accumulates trend anomaly information for each indicator within a preset time window, including the duration and percentage of anomalies. By accumulating trend anomaly information, it can distinguish between instantaneous fluctuations and continuously deteriorating trends.

[0081] The trend state estimator comprehensively determines and outputs a fourth feature vector based on the slope of change of each indicator, the cumulative deviation of each indicator, the trend anomaly status marker at the current moment, and the cumulative trend anomaly status information within a preset time window. This fourth feature vector simultaneously contains trend change information at the current moment, cumulative deviation information, and cumulative trend anomaly information within the time window.

[0082] Thus, by comprehensively covering K-dimensional information through multi-organ function indicators such as oxygenation index, serum creatinine, bilirubin, urine output, and Glasgow Coma Scale score, the trend analysis of organ function dimension is based on the assessment results of vital signs dimension by mapping the third feature vector to weight coefficients and correcting the trend judgment threshold, realizing the causal constraint transmission from vital signs dimension to organ function dimension. By performing linear trend fitting on each indicator within a preset time window to obtain the slope of change, and combining it with the cumulative deviation for comprehensive judgment, the cumulative change characteristics of organ function indicators are accurately captured. By generating trend anomaly markers through threshold comparison and accumulating them within the time window, the effective identification of continuously deteriorating trends is realized, avoiding misjudgments caused by single fluctuations.

[0083] Optionally, the four state score values ​​are all continuous values ​​within a preset numerical range; and the operation of determining the state evaluation result of the target object based on the four state score values ​​includes: adding the four state score values ​​to obtain a total score; determining the corresponding hierarchical identifier according to the preset score range into which the total score falls; and outputting the four state score values, the total score, and the hierarchical identifier as the state evaluation result of the target object.

[0084] In this embodiment, the four state scores are the scores output by the first, second, third, and fourth scoring heads, respectively, corresponding to the assessment results of the body response dimension, organ damage dimension, vital signs dimension, and organ function dimension in the RISK index. Each score is a continuous value within a preset numerical range. The preset numerical range is a pre-defined range of scores set by the system, such as 0 to 3. Each score takes a continuous value within this range, which can accurately reflect the degree of difference in the state of each dimension.

[0085] After obtaining the four status scores, the four status scores are added together to obtain the total score. Since each score is a continuous value within a preset range, the total score is also a continuous value, and its range is the sum of the ranges of each score. The total score reflects the overall status level of the target object in four dimensions.

[0086] Based on the preset scoring intervals into which the total score falls, a corresponding stratification identifier is determined. The preset scoring intervals are multiple consecutive numerical intervals pre-defined by the system, each interval corresponding to a stratification identifier. Different stratification identifiers are used to distinguish the overall level of the target object's state. For example, when the total score falls into the first interval, it corresponds to the first stratification identifier; when it falls into the second interval, it corresponds to the second stratification identifier, and so on. The stratification identifier is a discrete level marker used to characterize the severity level of the target object's overall state.

[0087] Finally, the four state scores, the total score, and the hierarchical identifier are combined to form the state evaluation result of the target object and output. That is, the output evaluation result includes both the independent scores for each dimension (the four state scores) and the comprehensive total score and hierarchical identifier. The four state scores provide fine-grained information for each dimension, the total score provides a comprehensive quantitative level, and the hierarchical identifier provides a graded comprehensive judgment result.

[0088] Thus, by outputting four independent state scores, a total score, and a hierarchical identifier, a multi-dimensional and structured evaluation result output is achieved. This ensures that the evaluation result is holistic while retaining fine-grained information for each dimension, overcoming the shortcomings of existing technologies that output information with only one dimension.

[0089] Optionally, the first encoder, the second encoder, and the first to fourth scoring heads are trained through the following steps: A training dataset is obtained, comprising multiple samples, each sample including the four-modal time-series data and corresponding four scoring value labels; based on the training dataset, a phased training strategy is adopted, sequentially training each encoder and its corresponding scoring head according to the causal order of the four-modal time-series data, and using the output features of the previous stage as the context constraint for the next stage of training, wherein: in the first stage, loss is calculated only for the data corresponding to the time when the first time-series data in each sample has a detection value, and the time interval without a detection value is not included in the loss calculation, training the first encoder and the first scoring head to obtain the trained first encoder and the first scoring head; in the second stage, the feature vector output by the first encoder obtained in the first stage is used as the context constraint to train the second encoder and the second scoring head; the loss function in the second stage includes a first loss term and a second loss term, the first loss term being... The first loss term constrains the output fluctuation of the second encoder during periods of high output values, and the second loss term constrains the degradation behavior of the second encoder when the condition of multiple consecutive detection confirmations is not met. In the third stage, the feature vector output by the second encoder obtained in the second stage is used as a context constraint to train the third scoring head, and the feature vector output by the statistical state estimator is used as a context constraint to train the fourth scoring head. In the fourth stage, based on the completion of the above three stages, all trainable parameters are jointly fine-tuned end-to-end. The loss function in the fourth stage includes the following four terms: a score prediction loss term, used to constrain the difference between the predicted score value and the corresponding labeled value of each scoring head; a causal consistency loss term, used to constrain the temporal order and directional consistency of score changes along the direction from the first mode to the fourth mode; an order relation loss term, used to constrain the order relation between the four score values ​​at the same time; and a total score consistency loss term, used to constrain the consistency between the sum of the four score values ​​and the preset total score.

[0090] In this embodiment, a training dataset is first obtained. This training dataset contains multiple samples, each including four-modal temporal data of the target object and four corresponding rating labels. The four rating labels are the labeled rating values ​​corresponding to the first, second, third, and fourth temporal data, respectively, serving as supervision signals during the training of the four rating heads.

[0091] After acquiring the training dataset, a phased training strategy is employed to train the first encoder, the second encoder, and the first through fourth scoring heads. This phased training strategy follows the causal order of the four modal temporal data, namely, the inherent sequential relationship between data related to bodily responses, organ damage, vital signs, and organ function. Each encoder and its corresponding scoring head are trained sequentially, with the output features of the previous stage serving as the contextual constraint for the next stage of training.

[0092] The first stage trains the first encoder and the first scoring head. In this stage, the loss is calculated only for the time intervals in the first time series data where a detection value arrives; time intervals without detection values ​​are not included in the loss calculation. Specifically, for each sample, detection values ​​arrive at some times and do not arrive at others in the first time series data. At times where a detection value arrives, the first encoder performs forward propagation calculation based on the input data, outputting a first feature vector. The first scoring head outputs a predicted score value based on the first feature vector. This predicted score value is compared with the corresponding labeled score value, the loss is calculated, and the parameters of the first encoder and the first scoring head are updated through backpropagation. During time intervals without detection values, the first encoder and the first scoring head do not perform forward propagation calculations, and these time intervals are not included in the loss calculation. The loss function for the first stage is as follows. The expression is: ; Where N is the number of training samples, i is the sample index, and t is the time index. Let be the set of times when the first time series data in the i-th sample has a detected value. The number of moments in this set. Let be the predicted score value of the first scorer at time t. Let be the labeled score value corresponding to time t. Through the above training method, the first encoder learns to accurately extract features when data arrives and to maintain the same output when no data arrives, consistent with the working method of the inference stage.

[0093] The second stage trains the second encoder and the second scoring head. This stage uses the feature vector output by the first encoder obtained in the first stage as context constraints. Specifically, when training the second encoder, the feature vector output by the first encoder is used as the initial hidden state of the gated recurrent unit, and it is concatenated with the second time-series data at each time step and input together. The loss function of the second stage includes the base loss for score prediction, a first loss term, and a second loss term, with a total loss of [missing information]. The expression is: ; in, , , These are the preset weighting coefficients for each loss item.

[0094] The scoring predicts the basic loss term. The expression used to constrain the difference between the predicted score value and the corresponding labeled value of the second scoring head is: ; in, Let be the predicted score value of the second scorer at time t. The labeled score value corresponding to time t. Let be the total number of time points for the i-th sample.

[0095] First loss item This is used to constrain the output fluctuation of the second encoder during periods of high output value. When the current output score of the second encoder is high, the output change in adjacent time steps should be constrained to suppress unnecessary fluctuations. The expression is: ; in, This is a preset adjustment coefficient used to control the degree to which the output value amplifies the fluctuation penalty. The weight term is monotonically increased as the output score increases. Its physical meaning is: when the output score is high, the penalty for fluctuations increases exponentially with the increase of the output value, thereby suppressing unstable fluctuations during periods of high output value.

[0096] Second loss item This is used to constrain the degradation behavior of the second encoder when the consecutive multiple detection confirmation condition is not met. If the second encoder prematurely performs a degradation update when the detection value is below a preset threshold but the consecutive multiple detection confirmation condition has not yet been met, a penalty is applied. Its expression is: ; in, To determine the extent of the downgrade, This is an indicator function; it takes a value of 1 when the confirmation condition for multiple consecutive detections is not met at time t, and a value of 0 otherwise. The physical meaning of this loss term is: to penalize the degradation magnitude only when the confirmation condition is not met. When the confirmation condition is met, this loss term is 0, and degradation behavior is not constrained. In this way, the second loss term guides the model to maintain the current output value when the confirmation condition is not met, suppressing premature degradation behavior.

[0097] Through the two loss terms mentioned above, the second encoder learns to suppress the update amplitude when the output value is high and to wait for multiple confirmations before performing the downgrade update during the downgrade process, which is consistent with the asymmetric update mechanism in the inference phase.

[0098] The third stage trains the third and fourth scoreheads. In this stage, the third scorehead is trained using the feature vector output by the second encoder (trained in the second stage) as context constraints, and the fourth scorehead is trained using the feature vector output by the statistical state estimator as context constraints. In this stage, the parameters of the second encoder and the statistical state estimator remain unchanged; only the parameters of the third and fourth scoreheads are updated. The total loss for the third stage is... The expression is: ; in, and These are the predicted scores for the third and fourth score heads at time t, respectively. and These are the corresponding labeled score values. Through the above loss, each score head learns to accurately map its corresponding feature vector to the score value.

[0099] The fourth stage is the end-to-end joint fine-tuning stage. Building upon the previous three stages, end-to-end joint fine-tuning is performed on all trainable parameters (i.e., all parameters of the first encoder, second encoder, first scorehead, second scorehead, third scorehead, and fourth scorehead). The loss function in the fourth stage includes the following four terms: The first item is the scoring prediction loss term. This is used to constrain the difference between the predicted score value and the corresponding labeled value for each scorehead. Its expression is: ; Where N is the number of training samples; i is the sample index (i=1,2,...,N); t is the time index; and m is the score head index (m=1,2,3,4, corresponding to the first to fourth score heads respectively). Let be the predicted score value output by the m-th score head at time t for the i-th sample; Let be the labeled score value of the m-th score head in the i-th sample at time t.

[0100] The second item is the causal consistency loss item. This is used to constrain the temporal order and directional consistency of score changes along the first to fourth modes. Specifically, the causal consistency loss term includes two terms: a temporal order constraint term and a directional consistency constraint term, and its expression is: ; Timing sequence constraints The expression used to constrain upstream score changes to precede downstream score changes in time is: ; in, Let be the change in rating of the m-th rating head at time t. This is a preset time delay parameter (reflecting the inherent time interval between the four types of data). The physical meaning of this constraint is: the change in downstream scores should not exceed the change in upstream scores in advance. The change over time, i.e. Should be in What happened next.

[0101] Directional consistency constraint To constrain upstream score changes to be in the same direction (co-directional changes), its expression is: ; in, This is a sign function. When the upstream score change and the downstream score change are in opposite directions, the product is negative. It produces a positive loss value; when the directions are consistent, the product is positive and the loss value is 0. This constraint guides the model to learn a pattern where the upstream and downstream ratings change in the same direction.

[0102] The third item is the order relation loss term. This is used to constrain the order relationship among four rating values ​​at the same time, and its expression is: ; in, The preset marginal threshold corresponding to the m-th rating head ( , , All are greater than 0). This loss constraint satisfies the following four score values ​​at the same time. , , The order relationship is defined as follows: the score value of the organism response dimension is no lower than the score value of the organ damage dimension plus a marginal threshold; the score value of the organ damage dimension is no lower than the score value of the vital signs dimension plus a marginal threshold; and the score value of the vital signs dimension is no lower than the score value of the organ function dimension plus a marginal threshold. This loss ensures that the score values ​​output by each scoring head at the same time conform to the order relationship between the dimensions.

[0103] The fourth item is the overall score consistency loss item. This is used to constrain the consistency between the sum of the four rating values ​​and the preset total rating, and its expression is: ; in, Let be the total score of the i-th sample at time t. This loss term ensures that the sum of the four scores remains consistent with the preset total score.

[0104] The total loss function in the fourth stage The weighted sum of the above four losses: ; in, , , , These are the preset weight coefficients for each loss term, used to balance the contribution of each loss term in joint training.

[0105] Thus, through the aforementioned phased training strategy, each encoder and its corresponding score head are trained sequentially according to the causal order of the four types of data, and the output features of the previous stage are used as the context constraints for the training of the next stage, aligning the training process with the causal order of the RISK four-dimensional structure. By calculating the loss only when data arrives in the first stage, the training behavior of the first encoder is kept consistent with the event triggering mechanism. Through the first and second loss terms in the second stage, the training process of the second encoder is matched with its asymmetric update mechanism. Through the causal consistency loss term, order relation loss term, and total score consistency loss term in the fourth stage, the causal order constraints, score order relation constraints, and total score constraints among the four types of data are embedded into the model's training process, enabling the trained model to output evaluation results that conform to the inherent constraints of the RISK four-dimensional structure during the inference phase.

[0106] Optionally, the dimensions of the first to fourth feature vectors are determined by the number of neurons in the output layer of each encoder. The number of neurons in the output layer of each encoder is set independently, meaning that the dimensions of each feature vector are independent of each other and can be the same or different.

[0107] Since the input layer dimension of each encoder is limited by the number of indicators in the corresponding time series data, while the output layer dimension is independent of the input layer dimension, even if the number of indicators in each modality is different (for example, the first time series data contains 11 indicators, the second time series data contains 11 indicators, the third time series data contains 6 indicators, and the fourth time series data contains 5 indicators), each encoder can still output feature vectors of the same dimension by designing the number of neurons in the output layer.

[0108] The dimension setting of each feature vector depends on the number of indicators, information richness, and processing requirements of the corresponding time-series data, as well as the processing needs of the subsequent scoring head. For example, when the information content of a certain modality is relatively rich, a higher feature vector dimension can be set to retain more information; when the information content of a certain modality is relatively limited, a lower feature vector dimension can be set to avoid overfitting. Those skilled in the art can set the number of neurons in each encoder output layer according to the actual application scenario and design requirements to determine the dimension of each feature vector.

[0109] As an optional implementation of this embodiment, the first to fourth feature vectors are all 32-dimensional floating-point vectors. In this case, the output dimensions of the first encoder, the second encoder, the statistical state estimator, and the trend state estimator are all 32-dimensional. Specifically, the output layer of the first encoder is 32-dimensional, mapping the 11 input bodily response-related indicators to 32-dimensional feature vectors; the hidden state dimension of the gated recurrent unit of the second encoder is 32-dimensional, mapping the 11 input organ damage-related indicators to 32-dimensional feature vectors; the statistical state estimator outputs 32-dimensional feature vectors to characterize the statistical deviation information of vital sign-related data across multiple time scales; and the trend state estimator outputs 32-dimensional feature vectors to characterize the temporal trend information of organ function-related data.

[0110] Specifically, the first feature vector is a 32-dimensional floating-point vector. The first encoder maps the values, changes, and historical statistical information of each indicator in the first time series data to a 32-dimensional space and outputs a comprehensive feature representation that includes the value of each indicator at the current time, the change of each indicator relative to the previous time, the mean and standard deviation of each indicator within a preset historical time window, and the normalized value of each indicator.

[0111] The second feature vector is a 32-dimensional floating-point vector. The second encoder maps the values, changes and historical statistical information of each indicator in the second time series data to a 32-dimensional space and outputs a comprehensive feature representation that includes the value of each indicator at the current time, the change of each indicator relative to the previous time, the mean and standard deviation of each indicator within a preset historical time window and the normalized value of each indicator.

[0112] The third feature vector is a 32-dimensional floating-point vector. The statistical state estimator maps the statistical deviation information of each indicator in the third time series data at multiple time scales to a 32-dimensional space and outputs a comprehensive feature representation that includes the deviation degree of each indicator at each time scale, the abnormal state mark of each indicator, the abnormal indicator count at the current time, the duration of the abnormal state within the preset time window, and the proportion of the abnormal state within the preset time window.

[0113] The fourth feature vector is a 32-dimensional floating-point vector. The trend state estimator maps the time-series trend information of each indicator in the fourth time-series data to a 32-dimensional space and outputs a comprehensive feature representation that includes the slope of change of each indicator within a preset time window, the cumulative deviation of each indicator from the baseline value within the preset time window, the trend state label of each indicator, the trend abnormal state label at the current moment, and the duration of the trend abnormal state within the preset time window.

[0114] It should be noted that the 32-dimensional feature vectors described above are merely one optional implementation of this invention and do not constitute a limitation on the scope of protection of this invention. Those skilled in the art can adaptively adjust the output dimensions of each encoder and estimator according to actual application scenarios and data characteristics.

[0115] In summary, this application has the following beneficial effects: (1) By acquiring four-modal time-series data of the target object, information covering four different dimensions of the body response, organ damage, vital signs and organ function is provided, which provides a complete data foundation for comprehensively assessing the state of the target object and overcomes the problem of incomplete information dimensions caused by using only a single type or part of the data for assessment in the existing technology.

[0116] (2) Through the event triggering mechanism of the first encoder, the output is updated only when data arrives and the output remains unchanged when no data arrives, so that the behavior of the encoder matches the characteristics of sparse and jump changes in biomarker data sampling, avoiding the non-real information introduced by interpolation filling in the no-data interval, and ensuring the authenticity of the extraction of the body's response dimension features.

[0117] (3) By using the asymmetric update mechanism of the second encoder, the update amplitude is suppressed when the output increases and multiple consecutive detections are required when the output decreases, so that the update behavior of the encoder matches the asymmetric change characteristics of organ damage markers, thereby improving the stability of organ damage dimension feature extraction and avoiding misjudgment caused by fluctuations in single detection values.

[0118] (4) The statistical state estimator is used to process the densely sampled vital signs data based on statistical deviation detection at multiple time scales, effectively extracting key information about deviations from steady state, while suppressing the interference of normal physiological fluctuations on the evaluation results.

[0119] (5) The trend state estimator processes the organ function parameter data based on time-series trend analysis, extracts trend features such as change slope and cumulative deviation, and accurately reflects the cumulative change characteristics of organ function parameters.

[0120] (6) By using the hierarchical context passing of processing the second time series data with the first feature vector as the context constraint, processing the third time series data with the second feature vector as the context constraint, and processing the fourth time series data with the third feature vector as the context constraint, the inherent sequential relationship between the four types of data is embedded into the evaluation process as a structural constraint, so that the processing of the next dimension is based on the evaluation result of the previous dimension, thus ensuring that the evaluation result conforms to the inherent sequential relationship between the four types of data in terms of structure.

[0121] (7) By mapping each feature vector to four independent state score values ​​through the first to fourth score heads, and determining the state evaluation result of the target object based on the four state score values, a multi-dimensional and structured evaluation result output is realized, so that the evaluation result retains fine-grained information of each dimension while having an overall nature.

[0122] Thus, this application achieves a comprehensive and accurate assessment of the state of the target object by differentiating the four-modal time-series data and applying hierarchical context constraints.

[0123] 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.

[0124] 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.

[0125] 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.

[0126] Example 2

[0127] Figure 4 A state assessment system based on RISK indicators and multidimensional biomarkers according to this embodiment is shown, corresponding to the method described in Embodiment 1. Reference Figure 4As shown, the system includes: a data acquisition module 410, used to acquire four-modal time-series data of the target object, the four-modal time-series data including: first time-series data related to bodily response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function; a first processing module 420, used to process the first time-series data using a first encoder and output a first feature vector; the first encoder only updates the output when data arrives, and keeps the output unchanged when no data arrives; a second processing module 430, used with the first feature vector as a context constraint, used a second encoder to process the second time-series data and output a second feature vector; the update amplitude of the second encoder decreases as the output value increases, and downgrade updates require... After multiple consecutive tests and confirmations, the third processing module 440 is used to process the third time-series data using a statistical state estimator with the second feature vector as a context constraint, and output a third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales; the fourth processing module 450 is used to process the fourth time-series data using a trend state estimator with the third feature vector as a context constraint, and output a fourth feature vector; the trend state estimator determines the fourth feature vector based on time-series trend analysis; the state evaluation module 460 is used to input the first to fourth feature vectors into the first to fourth scoring heads respectively to obtain four state score values, and determine the state evaluation result of the target object based on the four state score values.

[0128] It should be noted that the state assessment system based on RISK indicators and multidimensional biomarkers 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.

[0129] Example 3

[0130] Figure 5 A state assessment system based on RISK indicators and multidimensional biomarkers according to this embodiment is shown, corresponding to the method described in Embodiment 1. Reference Figure 5As shown, the system includes: a processor 510; and a memory 520 connected to the processor 510, used to provide the processor 510 with instructions to process the following steps: acquiring four-modal time-series data of a target object, the four-modal time-series data including: first time-series data related to bodily response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function; processing the first time-series data using a first encoder to output a first feature vector; the first encoder only updates its output when data arrives, and keeps its output unchanged when no data arrives; using the first feature vector as a context constraint, processing the second time-series data using a second encoder to output a second feature vector; the second... The encoder's update magnitude decreases as the output value increases, and downgrade updates require multiple consecutive detection confirmations. Using the second feature vector as a context constraint, a statistical state estimator processes the third time-series data and outputs a third feature vector. The statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales. Using the third feature vector as a context constraint, a trend state estimator processes the fourth time-series data and outputs a fourth feature vector. The trend state estimator determines the fourth feature vector based on time-series trend analysis. The first to fourth feature vectors are input into the first to fourth scoring heads respectively to obtain four state score values, and the state evaluation result of the target object is determined based on the four state score values.

[0131] It should be noted that the state assessment system based on RISK indicators and multidimensional biomarkers 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.

[0132] 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.

[0133] 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.

[0134] 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.

[0135] 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.

[0136] 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.

[0137] 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.

[0138] 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 state assessment method based on RISK indicators and multidimensional biomarkers, characterized in that, include: Acquire four-modal time-series data of the target object, the four-modal time-series data including: first time-series data related to body response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function; The first time-series data is processed by a first encoder to output a first feature vector; the first encoder only updates the output when data arrives, and keeps the output unchanged when no data arrives. Using the first feature vector as a context constraint, the second encoder processes the second time series data and outputs the second feature vector; the update magnitude of the second encoder decreases as the output value increases, and the downgrade update needs to be confirmed by multiple consecutive detections; Using the second feature vector as a context constraint, a statistical state estimator is used to process the third time series data and output a third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales; Using the third feature vector as a context constraint, a trend state estimator is used to process the fourth time series data and output a fourth feature vector; the trend state estimator determines the fourth feature vector based on time series trend analysis. The first to fourth feature vectors are input into the first to fourth scoring heads respectively to obtain four state scoring values, and the state evaluation result of the target object is determined based on the four state scoring values.

2. The method according to claim 1, characterized in that, The first time-series data includes time-series detection values ​​of leukocytes, lymphocytes, procalcitonin, interleukin-6, D-dimer, serum amyloid A, interleukin-8, interleukin-10, fibrinogen, ferritin, and lactate dehydrogenase. The first encoder is a fully connected neural network that employs an event-triggered mechanism; Furthermore, the operation of processing the first time-series data using the first encoder to output the first feature vector includes: Calculate the proportion of indicators that have received a detection value at the current moment out of all indicators; When the proportion is not lower than a preset threshold, the index values ​​with detected values ​​at the current moment and the latest historical values ​​of missing indexes are input together into the first encoder to obtain the first feature vector; when the proportion is lower than the preset threshold, the first feature vector output at the previous moment is kept unchanged, and a preset confidence level label is output.

3. The method according to claim 1, characterized in that, The second time-series data includes time-series detection values ​​of troponin, soluble advanced glycation end product receptor, KL-6, neuron-specific enolase, neurofilament light chain, kidney injury molecule-1, IGFBP-7, citrulline, intestinal fatty acid binding protein, alanine aminotransferase, and aspartate aminotransferase. The second encoder is a gate control loop unit with a continuous strength lock; The first feature vector serves as the initial hidden state of the gated loop unit, and is input into the gated loop unit after being concatenated with the second time-series data at each time step; The hidden state update of the gated loop unit is constrained by a continuous strength lock, wherein the locking strength of the continuous strength lock is... The hidden state update of the gated loop unit increases as the current output score increases, satisfying the following: ； in, The hidden state is calculated by the standard gated loop unit. This is the hidden state from the previous moment. This represents the hidden state at the current moment; The conditions for the second encoder to downgrade and update are: the current detection value is lower than a preset threshold, and the values ​​of at least two consecutive previous detections are lower than the preset threshold, and the time interval between two adjacent detections is not less than the preset minimum confirmation interval.

4. The method according to claim 1, characterized in that, The third time-series data includes time-series records of heart rate, mean arterial pressure, respiratory rate, body temperature, and consciousness status score. Furthermore, the operation of processing the third time-series data using a statistical state estimator with the second feature vector as a context constraint and outputting the third feature vector includes: The second feature vector is mapped to a first weight coefficient, and the preset benchmark deviation judgment threshold is corrected according to the first weight coefficient to obtain the actual deviation judgment threshold. At at least two preset time scales, the mean and standard deviation of each indicator in the third time series data are calculated at each time scale. Calculate the deviation of the value of the third time series data at the current moment from the mean and standard deviation at each time scale; The deviation of each indicator in the third time series data is compared with the actual deviation judgment threshold to determine whether each indicator is in an abnormal state at the current time. The third feature vector is determined based on the accumulation of abnormal states within a preset time window; wherein, the third feature vector includes: the deviation of each indicator in the third time series data at the current moment, the abnormal state marker at the current moment, and the accumulated information of abnormal states within the preset time window.

5. The method according to claim 1, characterized in that, The fourth time-series data includes time-series records of oxygenation index, serum creatinine, bilirubin, urine output, and Glasgow Coma Scale score. Furthermore, the operation of processing the fourth time-series data using a trend state estimator with the third feature vector as a context constraint and outputting the fourth feature vector includes: The third feature vector is mapped to a second weight coefficient, and the preset baseline trend determination threshold is corrected according to the second weight coefficient to obtain the actual trend determination threshold. Within a preset time window, a linear trend is fitted to each indicator in the fourth time series data to obtain the slope of change of each indicator. Calculate the cumulative deviation of each indicator in the fourth time series data relative to its respective baseline value within the preset time window; Based on the slope of change and the cumulative deviation of each indicator in the fourth time series data, the trend status of each indicator is determined; the trend status of each indicator is compared with the actual trend judgment threshold to determine whether the current moment is an abnormal trend state. The fourth feature vector is determined based on the accumulation of abnormal trend states within a preset time window; wherein, the fourth feature vector includes: the slope of change of each indicator in the fourth time series data, the cumulative deviation of each indicator, the abnormal trend state marker at the current moment, and the cumulative information of abnormal trend states within the preset time window.

6. The method according to claim 1, characterized in that, The four state score values ​​are all continuous values ​​within a preset numerical range; and the operation of determining the state assessment result of the target object based on the four state score values ​​includes: The total score is obtained by adding the four status scores together. Based on the preset scoring range into which the total score falls, determine the corresponding stratification identifier; The four state scores, the total score, and the hierarchical identifier are used as the state evaluation results of the target object and output.

7. The method according to claim 1, characterized in that, The first encoder, the second encoder, and the first to fourth scoring heads are obtained through the following steps: Obtain a training dataset, which contains multiple samples, each of which includes the four-modal time-series data and the corresponding four score value labels; Based on the training dataset, a phased training strategy is adopted, sequentially training each encoder and its corresponding scoring head according to the causal order of the four-modal time-series data, and using the output features of the previous stage as the context constraints for the training of the next stage, wherein: In the first stage, the loss is calculated only for the data corresponding to the time when the first time series data in each sample has a detection value. The time interval when no detection value arrives is not included in the loss calculation. The first encoder and the first score head are trained to obtain the trained first encoder and the first score head. In the second stage, the second encoder and the second scoring head are trained using the feature vector output by the first encoder obtained in the first stage as context constraints. The loss function in the second stage includes a first loss term and a second loss term. The first loss term is used to constrain the output fluctuation of the second encoder during periods of high output values, and the second loss term is used to constrain the degradation behavior of the second encoder when the condition of multiple consecutive detection confirmations is not met. In the third stage, the third scoring head is trained using the feature vector output by the second encoder obtained in the second stage as context constraints, and the fourth scoring head is trained using the feature vector output by the statistical state estimator as context constraints. In the fourth stage, based on the completion of the above three stages, end-to-end joint fine-tuning is performed on all trainable parameters. The loss function in the fourth stage includes the following four items: The rating prediction loss term is used to constrain the difference between the predicted rating value and the corresponding labeled value for each rating head; The causal consistency loss term is used to constrain the temporal order and directional consistency of score changes along the first to fourth modes. The order relation loss term is used to constrain the order relation between the four score values ​​at the same time. The overall score consistency loss term is used to constrain the consistency between the sum of the four score values ​​and the preset overall score.

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 state assessment system based on RISK indicators and multidimensional biomarkers, characterized in that, include: The data acquisition module is used to acquire four-modal time-series data of the target object. The four-modal time-series data includes: first time-series data related to body response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function. The first processing module is used to process the first time-series data using a first encoder and output a first feature vector; the first encoder only updates the output when data arrives, and keeps the output unchanged when no data arrives. The second processing module is used to process the second time series data with the first feature vector as a context constraint and the second encoder to output the second feature vector; the update amplitude of the second encoder decreases as the output value increases, and the downgrade update needs to be confirmed by multiple consecutive detections. The third processing module is used to process the third time series data using a statistical state estimator with the second feature vector as a context constraint, and output the third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales; The fourth processing module is used to process the fourth time series data using the third feature vector as a context constraint and a trend state estimator to output the fourth feature vector; the trend state estimator determines the fourth feature vector based on time series trend analysis. The state assessment module is used to input the first to fourth feature vectors into the first to fourth scoring heads respectively to obtain four state score values, and to determine the state assessment result of the target object based on the four state score values.

10. A state assessment system based on RISK indicators and multidimensional biomarkers, characterized in that, include: processor; A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: Acquire four-modal time-series data of the target object, the four-modal time-series data including: first time-series data related to body response, second time-series data related to organ damage, third time-series data related to vital signs, and fourth time-series data related to organ function; The first time-series data is processed by a first encoder to output a first feature vector; the first encoder only updates the output when data arrives, and keeps the output unchanged when no data arrives. Using the first feature vector as a context constraint, the second encoder processes the second time series data and outputs the second feature vector; the update magnitude of the second encoder decreases as the output value increases, and the downgrade update needs to be confirmed by multiple consecutive detections; Using the second feature vector as a context constraint, a statistical state estimator is used to process the third time series data and output a third feature vector; the statistical state estimator determines the third feature vector based on statistical deviation detection at multiple time scales; Using the third feature vector as a context constraint, a trend state estimator is used to process the fourth time series data and output a fourth feature vector; the trend state estimator determines the fourth feature vector based on time series trend analysis. The first to fourth feature vectors are input into the first to fourth scoring heads respectively to obtain four state scoring values, and the state evaluation result of the target object is determined based on the four state scoring values.

Images