A multi-organ dysfunction comprehensive early warning system and method for trauma critical patients
By employing a multimodal fusion prediction method combining large language models and GRU-D temporal missing dynamic modeling, this approach addresses the issues of insufficient utilization of unstructured text and inadequate handling of missing values in structured data in existing technologies. It enables accurate early warning of MODS risk in critically injured trauma patients, providing reliable support for early clinical intervention.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- INST OF MEDICAL SUPPORT TECH OF ACAD OF SYST ENG OF ACAD OF MILITARY SCI
- Filing Date
- 2026-04-08
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies cannot effectively utilize unstructured clinical text information, lack the ability to dynamically model structured time-series data, and have simple multimodal fusion methods, lacking forward-looking risk prediction, resulting in insufficient early warning of multiple organ dysfunction syndrome (MODS) in trauma patients with severe illness.
A multimodal fusion prediction method based on semantic modeling of large language model combined with GRU-D time-series missing value dynamic modeling is adopted to process unstructured clinical text and structured time-series data. By dynamically compensating for missing values through GRU-D model, multimodal feature collaborative modeling is realized, and a risk warning of MODS for the next 24 hours is output.
It enables accurate prediction of MODS risk in critically injured trauma patients within the next 24 hours, providing reliable decision support for early clinical intervention and stratified management, and improving the accuracy and stability of prediction.
Smart Images
Figure CN122392932A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of multimodal medical data processing and injury risk prediction technology, and in particular relates to a comprehensive early warning system and method for multiple organ dysfunction in patients with severe trauma. Background Technology
[0002] Trauma is a major global public health problem. Severe trauma patients admitted to the intensive care unit (ICU) are prone to developing multiple organ dysfunction syndrome (MODS) due to hemorrhagic shock, tissue hypoxia, and inflammatory cascade reactions. MODS manifests as a persistent systemic inflammatory response under severe trauma or other stress, leading to dysfunction or failure of two or more organ systems. Progression to multiple organ failure (MOF) results in extremely high mortality, and there is a lack of specific treatments. Clinical data show a significant positive correlation between the number of organ dysfunctions and patient mortality; therefore, early and accurate warning and timely intervention for MODS are crucial for improving the success rate of treatment for critically injured trauma patients.
[0003] Currently, commonly used clinical tools for assessing multiple organ dysfunction syndrome (MODS) include Sequential Organ Failure Assessment (SOFA), Logical Organ Dysfunction Scale (LODS), Marshall Multiorgan Dysfunction Scale (Marshall), and Quick SOFA (QSOFA). Among these, SOFA is recognized as the gold standard for MODS identification. However, traditional scoring systems rely solely on immediate static values of physiological parameters, failing to incorporate dynamic changes and statistical distribution information over time. Consequently, they cannot achieve dynamic monitoring and accurate early warning, making it difficult to meet the needs of early intervention.
[0004] With the widespread adoption of electronic medical records and continuous physiological monitoring technologies, multi-source clinical data provides a foundation for intelligent prediction of critical illnesses, and medical artificial intelligence is widely used in ICU patient prediction. However, existing AI models have significant shortcomings: First, the prediction of MODS is often based on a single complication or single organ dysfunction, or only on whether MODS has occurred. It does not build a multi-organ, multi-label classification and early warning framework, and cannot reflect the pathological characteristics of multiple organs deteriorating in parallel. Second, the lack of integration of unstructured data such as admission text information and the failure to systematically process missing electronic medical record data and differences in physiological parameter sampling intervals resulted in incomplete input information and data quality defects affecting prediction accuracy. Third, existing multi-organ prediction studies are few in number and the technology is not mature, and they are not specifically adapted to the group of patients with severe trauma. Fourth, there is a lack of application of multi-label classification frameworks in the comprehensive early warning of MODS in critically injured trauma patients; Fifth, the model has poor interpretability, and its output is difficult to translate into clinically understandable decision-making criteria, making it difficult to gain clinical trust. Sixth, most existing related patents are general or single-disease risk prediction frameworks, which do not construct organ-level multi-output models for trauma critical care MODS, do not make full use of data missing features and admission text information, lack continuous dynamic modeling and interpretable analysis, and cannot meet the needs of trauma critical care patients for accurate early warning of MODS.
[0005] To address the shortcomings of existing injury warning technologies for trauma and critically ill patients, there is an urgent need for a multimodal intelligent early warning system and method capable of simultaneously processing unstructured clinical text and structured time-series data, possessing dynamic modeling capabilities for missing values, and predicting the risk of MODS (Multiple Occurrence of Diseases) in trauma patients within the next 24 hours. This would solve the problem that existing traditional scoring systems rely on static physiological parameter thresholds for risk classification, only reflecting the patient's current state and failing to depict the continuous temporal changes in physiological indicators, resulting in limited early warning capabilities. While existing intelligent prediction models incorporate multi-source medical data, they primarily focus on modeling structured numerical data, underutilizing unstructured clinical text information in electronic medical records and failing to fully extract semantic information about disease progression and risk warnings contained in doctors' records. Furthermore, in the process of structured time-series modeling, clinical data commonly suffers from missing values and irregular sampling issues, making it difficult for traditional interpolation or simple completion methods to accurately reflect the impact of time intervals on disease assessment. In addition, multimodal fusion methods often employ simple feature splicing, lacking collaborative modeling mechanisms tailored to the characteristics of different data modalities, resulting in insufficient predictive stability and generalization ability. Summary of the Invention
[0006] This invention addresses the shortcomings of existing technologies, such as insufficient utilization of unstructured clinical text, inadequate handling of missing values in structured time-series data, simplistic multimodal fusion methods, and a lack of forward-looking risk prediction capabilities. It provides a multimodal intelligent early warning system and method for predicting the risk of MODS (Multiple Organ Dysfunction Syndrome) in the next 24 hours for critically injured trauma patients. By employing a technical solution of "semantic modeling based on a large language model pre-trained in a public medical database + dynamic modeling of missing values in time-series data using GRU-D + post-modal fusion prediction," the system fully leverages the dynamic evolution characteristics of semantic information in clinical text and structured physiological indicators to achieve accurate prediction of the risk of MODS in critically injured trauma patients within the next 24 hours, providing reliable decision support for early clinical intervention and stratified management.
[0007] To address the aforementioned technical problems, a first aspect of the present invention discloses a comprehensive early warning system for multiple organ dysfunction in critically ill trauma patients, the system comprising: Multi-source clinical data acquisition subsystem, data preprocessing subsystem, multimodal fusion prediction subsystem, and multi-organ dysfunction risk assessment subsystem; The multi-source clinical data acquisition subsystem is used to acquire unstructured clinical text data and structured time-series clinical data of trauma patients after admission. The data preprocessing subsystem is used to clean and standardize the unstructured clinical text data, and to perform missing data marking, time interval calculation and standardization on the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data. The multimodal fusion prediction subsystem is used to extract features from and fuse preprocessed text data and preprocessed structured clinical data based on a multimodal prediction model to obtain risk prediction results for each organ or system in the next 24 hours. The Multi-Organ Dysfunction Risk Assessment Subsystem is used to perform threshold assessment on the risk prediction results of each organ or system obtained in the next 24 hours to obtain MODS risk warning information for the next 24 hours. The multi-source clinical data acquisition subsystem, the data preprocessing subsystem, the multimodal fusion prediction subsystem, and the multi-organ dysfunction risk assessment subsystem are sequentially connected.
[0008] As an optional implementation, in a first aspect of the present invention, the multimodal fusion prediction subsystem includes: The module includes a text feature extraction module, a temporal feature modeling module, a feature fusion module, and a model update module. The text feature extraction module is used to extract semantic features from the preprocessed text data based on a large language model pre-trained on a public medical database to obtain a text feature vector. The time-series feature modeling module is used to dynamically model the preprocessed structured time-series clinical data based on the GRU-D (Gated-Recurrent-Unit-with-Decay) model to obtain time-series feature vectors. The feature fusion module is used to perform post-fusion processing on the text feature vector and the temporal feature vector to obtain the risk probability of each organ or system. The model update module is used to update the parameters of the multimodal prediction model based on real label data.
[0009] A second aspect of this invention discloses a comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients, the method comprising: S1, acquire unstructured clinical text data and structured time-series clinical data of trauma patients after admission; S2, preprocess the unstructured clinical text data and the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data; S3, process the preprocessed text data and the preprocessed structured clinical data to obtain the risk probability of each organ or system; S4, calibrate and threshold the risk probability of each organ or system to obtain risk warning information for each organ or system in the next 24 hours.
[0010] As an optional implementation, in a second aspect of the present invention, the preprocessing of the unstructured clinical text data and the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data includes: S21, the unstructured clinical text data is segmented, denoised, and standardized to obtain preprocessed text data; S22, perform missing value labeling, time interval calculation and numerical standardization on the structured time series clinical data to obtain preprocessed structured clinical data; The preprocessed structured clinical data includes missing marker vectors, time interval vectors, and global mean values of indicators.
[0011] As an optional implementation, in a second aspect of the present invention, the processing of the preprocessed text data and the preprocessed structured clinical data to obtain the risk probability of each organ or system includes: S31, Based on a large language model pre-trained from a public medical database, extract text features from the pre-processed text data to obtain a text feature vector; The text feature extraction expression is: , , in, The text feature vector is defined as follows: Pooling() is used for pooling operations; H is the hidden representation of the preprocessed text data; LLM() is the large language model. The preprocessed text data; S32, Based on the GRU-D model, time-series features are extracted from the preprocessed structured clinical data to obtain a time-series feature vector; S33, the text feature vector and the time sequence feature vector are fused to obtain the risk probability of each organ or system.
[0012] As an optional implementation, in a second aspect of the present invention, the GRU-D model includes: an input layer, an attenuation factor calculation layer, a missing compensation layer, a GRU core layer, and an output layer. The input layer is used to receive structured time-series clinical data, missing marker vectors, and sampling interval vectors; The attenuation factor calculation layer is used to generate attenuation coefficients based on time intervals. and hidden state decay coefficient ; The missing data compensation layer is used to adaptively decay and fill missing data in the structured time-series clinical data to obtain compensated structured time-series clinical data. The GRU core layer is used to complete temporal feature encoding through update gates, reset gates, and candidate hidden states to obtain temporal feature vectors. ; The output layer is used to output the temporal feature vector of fixed dimension. ; The input layer, the attenuation factor calculation layer, the missing compensation layer, the GRU core layer, and the output layer are sequentially connected.
[0013] As an optional implementation, in a second aspect of the present invention, the step of extracting time-series features from the preprocessed structured clinical data based on the GRU-D model to obtain a time-series feature vector includes: S321, Obtain the attenuation factor ; The expression for the attenuation factor is: , in, The attenuation factor; For time intervals; This is the first learnable parameter; This is the second learnable parameter; S322, Based on the GRU-D model and the preprocessed structured clinical data, the structured clinical data is compensated to obtain compensated structured clinical data; The compensation processing expression is: ， in, For the aforementioned compensating structured time-series clinical data; The structured time-series clinical data; The missing marker vector; The global mean of the aforementioned indicator; The attenuation factor for the input structured time-series clinical data; For Hadamah accumulation; S323, Based on the GRU-D model, the hidden state of the previous time step is decayed to obtain the decayed hidden state; The attenuation processing expression is: ， in, This is the hidden state after decay; This is the hidden state from the previous moment; The hidden state decay coefficient; S324, Based on the GRU-D model, perform gating calculation on the compensated structured clinical data and the attenuated hidden state to obtain the hidden state at the current moment; S325, Based on the GRU-D model, perform global pooling on the hidden states at all time points to obtain temporal feature vectors.
[0014] As an optional implementation, in a second aspect of the present invention, the fusion processing of the text feature vector and the temporal feature vector to obtain the risk probability of each organ or system includes: S331, regarding the text feature vector and the time-series feature vector Fully connected mapping fusion processing is performed to obtain text temporal fusion features; The fully connected mapping fusion processing expression is: , in, The text temporal fusion feature; The text feature vector and the temporal feature vector are concatenated along the channel dimension. It is a non-linear activation function; Weights for the fusion layer; For the fusion layer bias; S332, The text temporal fusion features are weighted to obtain the risk probability of each organ or system; The weighted processing expression is: ， in, Probability of MODS risk for each organ system; Use the Sigmoid activation function; These are the output layer weights; This is the output layer bias.
[0015] A third aspect of this invention discloses a comprehensive early warning device for multiple organ dysfunction in critically ill trauma patients, the device comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the program code stored in the memory to execute some or all of the steps in the comprehensive early warning method for multiple organ dysfunction in trauma patients disclosed in the second aspect of the present invention.
[0016] The fourth aspect of the present invention discloses a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, which, when invoked, execute some or all of the steps in the comprehensive early warning method for multiple organ dysfunction in trauma patients disclosed in the second aspect of the present invention.
[0017] Compared with the prior art, the present invention has the following beneficial effects: First, this invention introduces a large language model to perform semantic modeling on clinical texts, which can fully explore the implicit risk information contained in unstructured data such as doctors' medical records and diagnostic descriptions, make up for the shortcomings of traditional methods that rely solely on structured data, and improve the dimension of information utilization. Second, this invention uses the GRU-D model to dynamically model structured time-series data, and uses a time decay mechanism to characterize the impact of missing values and irregular sampling on disease assessment, thereby improving the model's adaptability to the distribution of real clinical data and its predictive stability. Third, this invention achieves collaborative modeling of text semantic features and physiological temporal features through a multimodal post-fusion mechanism, thereby improving the accuracy and generalization ability of risk prediction. Fourth, this invention can output the probability of a patient's MODS risk occurring within the next 24 hours, enabling prospective risk prediction and providing decision support for early clinical intervention, resource allocation, and tiered management. Therefore, this invention can achieve more accurate and stable early risk warnings in various trauma and critical care monitoring scenarios. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0019] Figure 1 This is a schematic diagram illustrating an application scenario of the comprehensive early warning device for multiple organ dysfunction in critically ill trauma patients disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of the composition of the comprehensive early warning system for multiple organ dysfunction in critically injured trauma patients disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the comprehensive early warning method for multiple organ dysfunction in critically injured trauma patients disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of another comprehensive early warning device for multiple organ dysfunction in critically ill trauma patients disclosed in an embodiment of the present invention.
[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.
[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0023] It should be noted that since the methods in this application are executed in computer devices, the processing objects of each computer device exist in the form of data or information; if time, quantity, location, etc. are mentioned in subsequent embodiments, they all refer to the existence of corresponding data so that the computer devices can process them.
[0024] In addition, the artificial intelligence-related technologies involved in this application include natural language processing, machine learning / deep learning, etc., which are used for feature extraction, fusion modeling and risk prediction of clinical text and structured time series data.
[0025] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating the application scenario of the comprehensive early warning device for multiple organ dysfunction in trauma patients provided in this application within an intensive care unit. The intensive care unit may include a computer device 100, which integrates the comprehensive early warning device for multiple organ dysfunction in trauma patients.
[0026] In this embodiment, the computer device 100 can be a standalone server, a server network or server cluster, or a cloud server. Those skilled in the art will understand that... Figure 1 The application environment given is merely an example and does not constitute a limitation on the application scenarios of the solution in this application.
[0027] In addition, such as Figure 1 As shown, the intensive care system may also include a memory 200 for storing unstructured clinical text data, structured time-series clinical data, model parameters, and early warning result data.
[0028] This invention addresses the shortcomings of existing technologies, such as insufficient utilization of unstructured clinical text, inadequate handling of missing values and irregular sampling in structured temporal data, simplistic multimodal fusion methods, and lack of forward-looking risk prediction capabilities for future time windows. It provides a comprehensive early warning system and method for multi-organ dysfunction (MODS) in critically injured trauma patients based on multimodal data fusion. By employing a core scheme of "semantic modeling based on a large language model pre-trained in a public medical database + dynamic modeling of missing temporal values using GRU-D + post-fusion risk prediction + threshold / adaptive threshold determination," it can predict and warn of the risk of MODS in critically injured trauma patients within the next 24 hours, providing a basis for early clinical intervention and stratified management.
[0029] Example 1 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a comprehensive early warning system for multiple organ dysfunction in critically ill trauma patients, as disclosed in an embodiment of the present invention. Figure 2 The described comprehensive early warning system for multiple organ dysfunction in critically ill trauma patients is applied to intensive care systems, such as local servers or cloud servers used in intensive care systems; however, this invention does not limit its application. Figure 2 As shown, this comprehensive early warning system for multiple organ dysfunction in critically ill trauma patients includes: Multi-source clinical data acquisition subsystem 101, data preprocessing subsystem 102, multimodal fusion prediction subsystem 103, and multi-organ dysfunction risk assessment subsystem 104; The multi-source clinical data acquisition subsystem 101 is used to acquire unstructured clinical text data and structured time-series clinical data of trauma patients after admission. It should be noted that the multi-source clinical data acquisition subsystem obtains unstructured clinical text data and structured time-series clinical data through hospital information systems, intensive care systems or manual input. It should be noted that the unstructured clinical text data includes admission text information; the admission text information includes chief complaint, description of injury, mechanism of trauma, past medical history and preliminary diagnosis; It should be noted that the structured time-series clinical data includes non-invasive physiological parameters, vital sign records, and routine laboratory indicators; the non-invasive physiological parameters include age, sex, body mass index, Glasgow Coma Scale, heart rate, respiratory rate, body temperature, non-invasive systolic blood pressure, non-invasive diastolic blood pressure, non-invasive mean arterial pressure, blood oxygen saturation, inhaled oxygen concentration, and urine output. The data preprocessing subsystem 102 is used to clean and standardize the unstructured clinical text data, and to perform missing data marking, time interval calculation and standardization on the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data. The multimodal fusion prediction subsystem 103 is used to extract features and fuse preprocessed text data and preprocessed structured clinical data based on a multimodal prediction model to obtain risk prediction results for each organ or system in the next 24 hours. The multi-organ dysfunction risk assessment subsystem 104 is used to perform threshold assessment processing on the risk prediction results of each organ or system obtained in the next 24 hours to obtain MODS risk warning information for the next 24 hours. The multi-source clinical data acquisition subsystem 101, the data preprocessing subsystem 102, the multimodal fusion prediction subsystem 103, and the multi-organ dysfunction risk assessment subsystem 104 are sequentially connected in data. As can be seen, the comprehensive early warning system for multi-organ dysfunction in critically injured patients based on multimodal data fusion, as described in the embodiments of the present invention, accurately captures the temporal precursor features and individual differences of injury deterioration by employing a comprehensive early warning system composed of a "multi-source clinical data acquisition subsystem, a data preprocessing subsystem, a multimodal fusion prediction subsystem, and a multi-organ dysfunction risk assessment subsystem." This achieves accurate and real-time early warning of general injury deterioration risks, providing a reliable basis for clinical intervention in multiple scenarios.
[0030] Optionally, the multimodal fusion prediction subsystem includes: The module includes a text feature extraction module, a temporal feature modeling module, a feature fusion module, and a model update module. The text feature extraction module is used to extract semantic features from the preprocessed text data based on a large language model pre-trained on a public medical database to obtain a text feature vector. The time-series feature modeling module is used to dynamically model the preprocessed structured time-series clinical data based on the GRU-D model to obtain time-series feature vectors. The feature fusion module is used to perform post-fusion processing on the text feature vector and the temporal feature vector to obtain the MODS risk probability of each organ system. The model update module is used to update the parameters of the multimodal prediction model based on real label data; As can be seen, the comprehensive early warning system for multi-organ dysfunction in critically injured patients based on multimodal data fusion described in this embodiment of the invention accurately captures the temporal precursor features and individual differences of injury deterioration through a multimodal fusion prediction subsystem composed of a text feature extraction module, a temporal feature modeling module, a feature fusion module, and a model update module. This enables accurate and real-time early warning of general injury deterioration risk and provides a reliable basis for clinical intervention in multiple scenarios.
[0031] Example 2 Please see Figure 3 , Figure 3 This is a schematic diagram of the architecture of a comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients disclosed in an embodiment of the present invention. Figure 3 The described comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients is applied to intensive care systems, such as local servers or cloud servers used in intensive care systems; however, this invention does not limit the application to such systems. Figure 3 As shown, this comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients includes: S1, acquire unstructured clinical text data and structured time-series clinical data of trauma patients after admission; It should be noted that the multi-source clinical data acquisition subsystem is used to acquire and output the unstructured clinical text data and the structured time-series clinical data to the data preprocessing subsystem. S2, preprocess the unstructured clinical text data and the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data; It should be noted that the data preprocessing subsystem is used to preprocess the unstructured clinical text data and the structured time-series clinical data to obtain and output preprocessed text data and preprocessed structured clinical data to the multimodal fusion prediction subsystem. S3, process the preprocessed text data and the preprocessed structured clinical data to obtain the risk probability of each organ or system; It should be noted that the multimodal fusion prediction subsystem is used to process the preprocessed text data and the preprocessed structured clinical data to obtain and output the risk probability of each organ or system to the multi-organ dysfunction risk assessment subsystem. It should be noted that the multimodal fusion prediction subsystem processes the current parameter status based on the data acquired by the multi-source clinical data acquisition subsystem within the patient's learning window to determine whether the current data is in a missing state and the sampling interval between this data and the previous real data. t, and calculate the original risk probability of dysfunction of each organ system in the future prediction window, where: the patient's current time t is the prediction time, 0-t is the learning window, and t~t+24h is the prediction window; S4, calibrate and threshold the risk probability of each organ or system to obtain MODS risk warning information for the next 24 hours; It should be noted that the threshold determination process refers to determining the threshold range to which the MODS risk probability of each organ system belongs, and determining the MODS risk warning information for the next 24 hours based on the threshold range. It should be noted that in this embodiment of the invention, the MODS risk probability of each organ system is corrected by combining the local medical environment, ICU treatment capacity and doctor experience to obtain the corrected risk prediction probability; based on the threshold range, a graded judgment is made. When the corrected risk of dysfunction of one or more organs exceeds the preset threshold, the system determines that the patient has a high risk of MODS within the prediction window and outputs the corresponding organ-level risk warning information to medical staff to guide the clinical implementation of targeted intervention in advance. It should be noted that the Multi-Organ Dysfunction Risk Assessment Subsystem is used to perform threshold assessment on the MODS risk probability of each organ system, and to obtain and output MODS risk warning information for the next 24 hours to the user. As can be seen, the comprehensive early warning method for multi-organ dysfunction (MODS) in critically injured trauma patients based on multimodal data fusion described in the embodiments of the present invention, by adopting the technical solution of "semantic modeling of a large language model pre-trained on a public medical database + dynamic modeling of GRU-D time-series missing data + multimodal post-fusion prediction", fully explores the dynamic evolution characteristics of clinical text semantic information and structured physiological indicators, and achieves accurate prediction of the risk of MODS in various organs or systems of critically injured trauma patients in the next 24 hours, providing reliable decision support for early clinical intervention and hierarchical management.
[0032] Optionally, the preprocessing of the unstructured clinical text data and the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data includes: S21, the unstructured clinical text data is segmented, denoised, and standardized to obtain preprocessed text data; S22, perform missing value labeling, time interval calculation and numerical standardization on the structured time series clinical data to obtain preprocessed structured clinical data; The preprocessed structured clinical data includes missing marker vectors, time interval vectors, and global mean values of indicators; It should be noted that the missing marker vector ,when This indicates that the data exists. Indicates missing data; the time interval vector This is used to represent the time difference between the current moment and the last valid observation; the global mean of the index , used for filling in missing data; As can be seen, the multimodal data fusion-based comprehensive early warning method for multi-organ dysfunction in trauma patients described in this embodiment of the invention preprocesses the unstructured clinical text data and the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data. This provides data support for fully mining the dynamic evolution characteristics of clinical text semantic information and structured physiological indicators, and for accurately predicting the risk of MODS in trauma patients in the next 24 hours. It also provides reliable decision support for early clinical intervention and hierarchical management.
[0033] Optionally, the processing of the preprocessed text data and the preprocessed structured clinical data to obtain the risk probability of each organ or system includes: S31, Based on a large language model pre-trained from a public medical database, extract text features from the pre-processed text data to obtain a text feature vector; The text feature extraction expression is: , , in, The text feature vector is defined as follows: Pooling() is used for pooling operations; H is the hidden representation of the preprocessed text data; LLM() is the large language model. The preprocessed text data; It should be noted that, through the text feature extraction, unstructured admission text is transformed into high-dimensional semantic features, which fully preserves key information such as the patient's trauma background, underlying diseases, and initial injury, thus supplementing the information gaps of static physiological parameters from the source and improving the model's ability to perceive trauma causes and basic conditions. S32, Based on the GRU-D model, time-series features are extracted from the preprocessed structured clinical data to obtain a time-series feature vector; It should be noted that by extracting temporal features, missing data and irregular sampling are transformed into learnable features, avoiding information bias caused by simple interpolation / deletion; by modeling the dynamic evolution of physiological indicators of trauma patients through time decay, the robustness and predictive stability of features under high missing data and high noise ICU data are improved. S33, the text feature vector and the time sequence feature vector are fused to obtain the risk probability of each organ or system; As can be seen, the comprehensive early warning method for multi-organ dysfunction (MODS) in critically injured trauma patients based on multimodal data fusion described in this embodiment of the invention processes the preprocessed text data and the preprocessed structured clinical data to obtain the risk probability of MODS in each organ system, thereby achieving accurate prediction of the risk of MODS in critically injured trauma patients in the next 24 hours and providing reliable decision support for early clinical intervention and hierarchical management.
[0034] Optionally, the GRU-D model includes: an input layer, a decay factor calculation layer, a missing compensation layer, a GRU core layer, and an output layer; The input layer is used to receive structured time-series clinical data, missing marker vectors, and sampling interval vectors; The attenuation factor calculation layer is used to generate attenuation coefficients based on time intervals. and hidden state decay coefficient ; The missing data compensation layer is used to adaptively decay and fill missing data in the structured time-series clinical data to obtain compensated structured time-series clinical data. The GRU core layer is used to complete temporal feature encoding through update gates, reset gates, and candidate hidden states to obtain temporal feature vectors. ; The output layer is used to output the temporal feature vector of fixed dimension. ; The input layer, the attenuation factor calculation layer, the missing compensation layer, the GRU core layer, and the output layer are sequentially connected by data. It should be noted that the GRU-D model is a recurrent neural network model specifically designed for time series data with irregular sampling and high missing rates. As can be seen, the comprehensive early warning method for multi-organ dysfunction in trauma patients based on multimodal data fusion described in the embodiments of the present invention utilizes the GRU-D model to extract temporal features from the preprocessed structured clinical data to obtain temporal feature vectors. This lays the foundation for accurate prediction of the risk of MODS in trauma patients in the next 24 hours and provides reliable decision support for early clinical intervention and hierarchical management.
[0035] Optionally, the step of extracting time-series features from the preprocessed structured clinical data based on the GRU-D model to obtain a time-series feature vector includes: S321, Obtain the attenuation factor ; The expression for the attenuation factor is: , in, The attenuation factor; For time intervals; This is the first learnable parameter; This is the second learnable parameter; S322, Based on the GRU-D model and the preprocessed structured clinical data, the structured clinical data is compensated to obtain compensated structured clinical data; The compensation processing expression is: ， in, For the aforementioned compensating structured time-series clinical data; The structured time-series clinical data; The missing marker vector; The global mean of the aforementioned indicator; The attenuation factor for the input structured time-series clinical data; For Hadamah accumulation; S323, Based on the GRU-D model, the hidden state of the previous time step is decayed to obtain the decayed hidden state; The attenuation processing expression is: ， in, This is the hidden state after decay; This is the hidden state from the previous moment; The hidden state decay coefficient; S324, Based on the GRU-D model, perform gating calculation on the compensated structured clinical data and the attenuated hidden state to obtain the hidden state at the current moment; S325, Based on the GRU-D model, perform global pooling on the hidden states at all time points to obtain temporal feature vectors; It should be noted that the expression for the time-series feature vector is as follows: ， in, The time-series feature vector; This is structured time-series clinical data; It is a gated cyclic unit model with a decay mechanism; As can be seen, the comprehensive early warning method for multi-organ dysfunction in trauma patients based on multimodal data fusion described in the embodiments of the present invention utilizes the GRU-D model to extract temporal features from the preprocessed structured clinical data to obtain temporal feature vectors. This lays the foundation for accurate prediction of the risk of MODS in trauma patients in the next 24 hours and provides reliable decision support for early clinical intervention and hierarchical management.
[0036] Optionally, the fusion processing of the text feature vector and the temporal feature vector to obtain the risk probability of each organ or system includes: S331, regarding the text feature vector and the time-series feature vector The text temporal fusion features are obtained by performing fusion processing. It should be noted that the fusion process is one or a combination of feature splicing fusion process and weighted summation fully connected mapping fusion process, and the embodiments of the present invention are not limited thereto; Optionally, the fusion process is a feature splicing fusion process; The expression for feature splicing and fusion processing is: , in, The text temporal fusion feature; The text feature vector and the temporal feature vector are concatenated along the channel dimension. Optionally, the fusion process is a weighted summation process; The weighted summation expression is as follows: , in, The text temporal fusion feature; Text feature vectors can be learned to fuse weights; The temporal feature vectors can be used to learn fusion weights; It should be noted that the weighted summation process is a learnable weighted fusion method, and the learnable fusion weights of the text feature vector and the learnable fusion weights of the temporal feature vector are adaptively updated through backpropagation; Optionally, the fusion process is a fully connected mapping fusion process; The fully connected mapping fusion processing expression is: ， in, The text temporal fusion feature; The text feature vector and the temporal feature vector are concatenated along the channel dimension. It is a non-linear activation function; Weights for the fusion layer; For the fusion layer bias; S332, The text temporal fusion features are weighted to obtain the risk probability of each organ or system; The weighted processing expression is: ， in, Probability of MODS risk for each organ system; Use the Sigmoid activation function; These are the output layer weights; For output layer bias; It should be noted that the Sigmoid activation function outputs a probability value between 0 and 1. It should be noted that the organ systems mentioned include the respiratory system, cardiovascular system, kidneys, liver, coagulation system, and central nervous system; It should be noted that, through the text feature vector and the time-series feature vector The fusion and weighted processing achieves deep coupling between textual semantic information and temporal dynamic information, adaptively learns the contribution of features of different modalities, outputs the parallel risk probability of multiple organs, accurately reflects the pathological process of synchronous deterioration and mutual influence of multiple organs after trauma, and supports organ-level hierarchical early warning. As can be seen, the comprehensive early warning method for multi-organ dysfunction (MODS) in critically injured trauma patients based on multimodal data fusion described in the embodiments of the present invention fuses the text feature vector and the temporal feature vector to obtain the MODS risk probability of each organ system, thereby achieving accurate prediction of the risk of MODS in critically injured trauma patients in the next 24 hours and providing reliable decision support for early clinical intervention and hierarchical management.
[0037] Optionally, the calibration and threshold determination of the risk probability of each organ or system to obtain MODS risk warning information for the next 24 hours includes: S41, Receive the MODS risk probability of each organ system; S42, constructing comprehensive clinical calibration factors; The expression for the comprehensive clinical calibration factor is: , in, To integrate clinical calibration factors; For hospital calibration coefficients; This is a correction factor for the severity of the trauma. This is an environmental correction factor; It should be noted that the hospital calibration coefficient mentioned above... The strength of the ICU is determined by its capabilities and resources; the stronger the ICU, the better. The smaller the value, the more conservative the warning; the range is 0.8–1.2; the trauma severity correction coefficient. Determined by the injury mechanism, the more severe the trauma... The larger the value, the more sensitive the warning; the value range is 1.0–1.5; the environmental correction coefficient. The value ranges from 0.9 to 1.1; It should be noted that comprehensive clinical calibration factors are used to make the AI model output adaptable to different hospitals and different trauma groups, avoiding false alarms / missed alarms in different ICUs due to the general model; S43, Based on the comprehensive clinical calibration factor, the MODS risk probability of each organ system is calibrated to obtain the calibrated organ risk probability; The calibration processing expression is: , in, The probability of MODS risk for the i-th organ system. To integrate clinical calibration factors; The default value for calibration bias is 0. Use the Sigmoid function to ensure that the output is between 0 and 1; The calibrated organ risk probability; It should be noted that the MODS risk probabilities of each organ system are calibrated to transform the original AI probabilities into reliable probabilities that conform to local clinical practice, thereby improving the prediction accuracy. S44, determine whether the calibrated organ risk probability is greater than the preset organ threshold, and obtain the probability judgment result; When the probability judgment result is yes, the corresponding organ risk marker is 1; When the probability judgment result is negative, the corresponding organ risk marker is 0; S45, count all the organs with risk markers set to 1 to obtain the number of risky organs. ; S46, Determine the number of organs at risk. The result of determining the number of organs at first risk is obtained by checking whether the number of organs at first risk is greater than or equal to the threshold number of organs at first risk. It should be noted that the threshold for the number of the first risky organs is set to 3; When the result of the first risk organ count determination is yes, obtain the risk organ information, set the warning information to level three warning, and execute S49; It should be noted that the Level 3 warning refers to Critical MODS; the MODS risk warning information for the next 24 hours is a Level 3 warning (Critical MODS). If the result of the first risk organ count determination is negative, execute S47; S47, Determine the number of organs at risk. The result of determining the number of second-risk organs is obtained by checking whether the number of organs is greater than or equal to the threshold for the number of second-risk organs. It should be noted that the threshold for the number of the second risky organs is set to 2; When the result of the second risk organ count determination is yes, obtain the risk organ information, set the warning information to level two warning, and execute S49; It should be noted that the aforementioned Level 2 warning refers to early MODS; If the result of the second risk organ count determination is negative, execute S48; S48, determine the number of organs at risk. The result of determining the number of third-risk organs is obtained by checking whether the number of organs is greater than or equal to the threshold for the number of third-risk organs. It should be noted that the threshold for the number of third risk organs is set to 1; When the result of the determination of the number of third risk organs is yes, the risk organ information is obtained, the warning information is set to a first-level warning, and S49 is executed; It should be noted that the Level 1 warning refers to high risk for a single organ. When the result of the determination of the number of third risk organs is negative, the warning information is set to no warning, and S49 is executed; S49, The risk organ category information is combined with the warning information to obtain MODS risk warning information for the next 24 hours; It should be noted that the combination is based on the order of organ names and warning information; As can be seen, the comprehensive early warning method for multi-organ dysfunction (MODS) in critically injured trauma patients based on multimodal data fusion described in the embodiments of the present invention performs threshold determination processing on the MODS risk probability of each organ system to obtain MODS risk warning information for the next 24 hours. It realizes graded early warning according to the number of risky organs, improves the accuracy of predicting the risk of MODS in critically injured trauma patients in the next 24 hours, and provides reliable decision support for early clinical intervention and hierarchical management.
[0038] Example 3 Please see Figure 4 , Figure 4 This is a schematic diagram of another comprehensive early warning device for multiple organ dysfunction in critically ill trauma patients disclosed in an embodiment of the present invention. Figure 4The described device can be applied to intensive care systems, such as local servers or cloud servers used in intensive care systems, and the embodiments of the present invention are not limited thereto. Figure 4 As shown, the device may include: Memory 201 storing executable program code; Processor 202 coupled to memory 201; The processor 202 calls the executable program code stored in the memory 201 to execute the steps in the comprehensive early warning method for multiple organ dysfunction in trauma patients described in Embodiment 2.
[0039] Example 4 This invention discloses a computer-readable storage medium storing a computer program for electronic data exchange, wherein the computer program causes a computer to perform the steps in the comprehensive early warning method for multiple organ dysfunction in trauma patients described in Embodiment 2.
[0040] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the comprehensive early warning method for multiple organ dysfunction in trauma patients described in Embodiment 2.
[0041] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.
[0042] It should be noted that all calculation expressions or mathematical functions in the embodiments of the present invention have undergone dimensionless processing of the variables involved before calculation.
[0043] It should be noted that in all the calculation expressions or mathematical functions in the embodiments of the present invention, the values of the input independent variables all meet the reasonable requirements of the input value range of the calculation expression or mathematical function, and can ensure that the calculation expression or mathematical function can be calculated smoothly without violating physical laws or mathematical rules.
[0044] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, 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 can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), once programmable read-only memory (OTPROM), electronically erasable rewritable read-only memory (EEPROM), read-only optical disc (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.
[0045] Finally, it should be noted that the comprehensive early warning system and method for multiple organ dysfunction in critically ill trauma patients disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A comprehensive early warning system for multiple organ dysfunction in critically ill trauma patients, characterized in that, The system includes: a multi-source clinical data acquisition subsystem, a data preprocessing subsystem, a multimodal fusion prediction subsystem, and a multi-organ dysfunction risk assessment subsystem; The multi-source clinical data acquisition subsystem is used to acquire unstructured clinical text data and structured time-series clinical data of trauma patients after admission. The data preprocessing subsystem is used to clean and standardize the unstructured clinical text data, and to perform missing data marking, time interval calculation and standardization on the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data. The multimodal fusion prediction subsystem is used to extract features from and fuse preprocessed text data and preprocessed structured clinical data based on a multimodal prediction model to obtain risk prediction results for each organ or system in the next 24 hours. The multi-organ dysfunction risk assessment subsystem is used to verify and threshold the risk prediction results of each organ or system obtained in the next 24 hours to obtain risk warning information of each organ or system in the next 24 hours. The multi-source clinical data acquisition subsystem, the data preprocessing subsystem, the multimodal fusion prediction subsystem, and the multi-organ dysfunction risk assessment subsystem are sequentially connected.
2. The comprehensive early warning system for multiple organ dysfunction in critically ill trauma patients according to claim 1, characterized in that, The multimodal fusion prediction subsystem includes: The module includes a text feature extraction module, a temporal feature modeling module, a feature fusion module, and a model update module. The text feature extraction module is used to extract semantic features from the preprocessed text data based on a large language model pre-trained on a public medical database to obtain a text feature vector. The time-series feature modeling module dynamically models the preprocessed structured time-series clinical data based on the GRU-D model to obtain time-series feature vectors. The feature fusion module is used to perform post-fusion processing on the text feature vector and the temporal feature vector to obtain the risk probability of each organ or system. The model update module is used to update the parameters of the multimodal prediction model based on real label data.
3. A comprehensive early warning method for multiple organ dysfunction in critically injured trauma patients, applied to the comprehensive early warning system for multiple organ dysfunction in critically injured trauma patients as described in claim 1 or 2, characterized in that, The method includes: S1, acquire unstructured clinical text data and structured time-series clinical data of trauma patients after admission; S2, preprocess the unstructured clinical text data and the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data; S3, process the preprocessed text data and the preprocessed structured clinical data to obtain the risk probability of each organ or system; S4, calibrate and threshold the risk probability of each organ or system to obtain MODS risk warning information for the next 24 hours.
4. The comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients according to claim 3, characterized in that, The preprocessing of the unstructured clinical text data and the structured time-series clinical data to obtain preprocessed text data and preprocessed structured clinical data includes: S21, the unstructured clinical text data is segmented, denoised, and standardized to obtain preprocessed text data; S22, perform missing value labeling, time interval calculation and numerical standardization on the structured time series clinical data to obtain preprocessed structured clinical data; The preprocessed structured clinical data includes missing marker vectors, time interval vectors, and global mean values of indicators.
5. The comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients according to claim 3, characterized in that, The process of processing the preprocessed text data and the preprocessed structured clinical data to obtain the risk probability of each organ or system includes: S31, Based on a large language model pre-trained from a public medical database, extract text features from the pre-processed text data to obtain a text feature vector; The text feature extraction expression is: ， ， in, The text feature vector; Pooling() To perform pooling operations; H This is a hidden representation of the preprocessed text data; LLM() For the large language model; The preprocessed text data; S32, Based on the GRU-D model, time-series features are extracted from the preprocessed structured clinical data to obtain a time-series feature vector; S33, the text feature vector and the time sequence feature vector are fused to obtain the risk probability of each organ or system.
6. The comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients according to claim 5, characterized in that, The GRU-D model includes: an input layer, an attenuation factor calculation layer, a missing compensation layer, a GRU core layer, and an output layer. The input layer is used to receive structured time-series clinical data, missing marker vectors, and sampling interval vectors; The attenuation factor calculation layer is used to generate attenuation coefficients based on time intervals. and hidden state decay coefficient ; The missing data compensation layer is used to adaptively decay and fill missing data in the structured time-series clinical data to obtain compensated structured time-series clinical data. The GRU core layer is used to complete temporal feature encoding through update gates, reset gates, and candidate hidden states to obtain temporal feature vectors. ; The output layer is used to output the temporal feature vector of fixed dimension. ; The input layer, the attenuation factor calculation layer, the missing compensation layer, the GRU core layer, and the output layer are sequentially connected.
7. The comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients according to claim 5, characterized in that, The GRU-D model is used to extract time-series features from the preprocessed structured clinical data to obtain a time-series feature vector, including: S321, Obtain the attenuation factor ; The expression for the attenuation factor is: , in, The attenuation factor; For time intervals; This is the first learnable parameter; This is the second learnable parameter; S322, Based on the GRU-D model and the preprocessed structured clinical data, the structured clinical data is compensated to obtain compensated structured clinical data; The compensation processing expression is: ， in, For the aforementioned compensating structured time-series clinical data; The structured time-series clinical data; The missing marker vector; The global mean of the aforementioned indicator; The attenuation factor for the input structured time-series clinical data; For Hadamah accumulation; S323, Based on the GRU-D model, the hidden state of the previous time step is decayed to obtain the decayed hidden state; The attenuation processing expression is: ， in, This is the hidden state after decay; This is the hidden state from the previous moment; The hidden state decay coefficient; S324, Based on the GRU-D model, perform gating calculation on the compensated structured clinical data and the attenuated hidden state to obtain the hidden state at the current moment; S325, Based on the GRU-D model, perform global pooling on the hidden states at all time points to obtain temporal feature vectors.
8. The comprehensive early warning method for multiple organ dysfunction in critically ill trauma patients according to claim 5, characterized in that, The process of fusing the text feature vector and the temporal feature vector to obtain the risk probability of each organ or system includes: S331, for the text feature vector and the time-series feature vector Fully connected mapping fusion processing is performed to obtain text temporal fusion features; The fully connected mapping fusion processing expression is: ， in, The text temporal fusion feature; The text feature vector and the temporal feature vector are concatenated along the channel dimension. It is a non-linear activation function; Weights for the fusion layer; For the fusion layer bias; S332, The text temporal fusion features are weighted to obtain the risk probability of each organ or system; The weighted processing expression is: ， in, Probability of MODS risk for each organ system; Use the Sigmoid activation function; These are the output layer weights; This is the output layer bias.
9. A comprehensive early warning device for multiple organ dysfunction in critically ill trauma patients, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the program code to execute the comprehensive early warning method for multiple organ dysfunction in critically injured trauma patients as described in any one of claims 3 to 8.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when invoked, execute the comprehensive early warning method for multiple organ dysfunction in critically injured trauma patients as described in any one of claims 3 to 8.