Multi-source data-based sepsis early risk prediction system and method

Abstract

The invention relates to the field of intelligent prediction, and particularly discloses a sepsis early-stage risk prediction system and method based on multi-source data, and the method comprises the steps: carrying out the time decay embedding of historical information based on an observation time interval after the historical data of a patient is encoded; performing feature attenuation and filling on the current incomplete data by using the mask vector; dynamically enhancing internal association of different data sources through a multi-modal attention mechanism; finally, a loop unit state is updated based on the fused vector to generate a risk score. Therefore, the problem of irregular clinical data sampling can be effectively solved, dynamic association among heterogeneous data is deeply mined, more accurate and timely sepsis risk prediction is realized, powerful decision support is provided for clinical early intervention, and the method can be deployed on an intelligent chip.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent prediction, and more specifically, to a sepsis early risk prediction system and method based on multi-source data. BACKGROUND

[0002] Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection, with high morbidity and mortality, and has become one of the main causes of patient death in intensive care units. Clinical practice shows that early identification and timely intervention of sepsis is the key to improving patient prognosis and reducing mortality. However, the clinical manifestations of sepsis are complex and diverse, and the early symptoms are not specific. Medical staff need to analyze and diagnose the dynamic data from multiple sources such as vital signs and laboratory tests, which undoubtedly increases the difficulty of early and accurate judgment. Therefore, it is of great practical significance to build a system that can automatically and accurately predict the risk of sepsis occurrence to assist clinical decision-making and gain valuable treatment time.

[0003] To achieve early prediction of sepsis, existing technologies have attempted to use machine learning methods to model patient electronic health record data. These methods have made some progress, but there are still significant technical bottlenecks in dealing with the complex characteristics of clinical data. First, the collection of clinical data has natural irregularity and asynchronicity in time. For example, vital sign monitoring may be interrupted due to patient movement, while laboratory tests are performed on demand according to the judgment of the doctor, with intervals ranging from a few hours to several days. Traditional time series models usually assume that the data sampling interval is fixed, which is seriously inconsistent with the actual situation. In order to deal with data missing, existing models often use simple methods such as forward filling, but this distorts the true trajectory of the dynamic evolution of the patient's state. More importantly, these models cannot effectively learn and utilize the clinical information contained in the irregular sampling interval itself, and it is difficult to distinguish the essential difference between "not measured" and "measured value normal", which limits the accuracy and reliability of the model's prediction, and makes it difficult to meet the clinical needs of early and accurate warning.

[0004] Therefore, an optimized sepsis early risk prediction system based on multi-source data is expected. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. Embodiments of the present application provide a sepsis early risk prediction system and method based on multi-source data, which first encodes the historical vital signs and test data of a patient, then performs time decay embedding on the historical information based on the time interval between the current observation and the last time to simulate the natural forgetting process of information over time. Meanwhile, the system uses a mask vector to perform feature decay and filling on the incomplete input data, and dynamically captures and enhances the internal correlation between different data sources through a multi-modal attention mechanism. Finally, the input vector fused with multiple information is used to update the recurrent unit state to generate the final risk score. In this way, the information distortion problem caused by irregular clinical data sampling can be effectively overcome, the complex dynamic correlation between heterogeneous data can be deeply mined, and finally more accurate and timely sepsis risk prediction can be realized, providing strong decision support for early clinical intervention.

[0006] According to an aspect of the present application, a sepsis early risk prediction system based on multi-source data is provided, which comprises: a data preprocessing module for preprocessing and time aligning historical raw vital sign data and historical raw test data of a target patient object to obtain a historical data matrix; a sequence encoding module for sequence encoding the historical data matrix to obtain a hidden state at the last time; a target patient object information acquisition module for acquiring a current time data vector, a current time interval vector and a current time mask vector of the target patient object; a time decay embedding module for time decay embedding the hidden state at the last time based on the current time interval vector to obtain a decayed hidden state; a feature decay and filling module for feature decay and filling the current time data vector based on the current time mask vector to obtain a decayed input; a multi-modal attention enhancement module for multi-modal attention enhancement of the decayed input and the decayed hidden state to obtain an attention enhanced input; a recurrent unit state updating module for recurrent unit state updating of the attention enhanced input and the decayed hidden state to obtain a final hidden state at the current time; a sepsis risk score module for generating a sepsis risk score based on the final hidden state at the current time.

[0007] According to another aspect of the present application, a sepsis early risk prediction method based on multi-source data is provided, which comprises: preprocessing and time aligning historical raw vital sign data and historical raw test data of a target patient object to obtain a historical data matrix; sequence-encoding the historical data matrix to obtain a hidden state of a previous time instant; obtaining a current time instant data vector, a current time instant time interval vector and a current time instant mask vector of a target patient subject; based on the current time instant time interval vector, performing time decay embedding on the hidden state of the previous time instant to obtain a decayed hidden state; based on the current time instant mask vector, performing feature decay and padding on the current time instant data vector to obtain a decayed input; performing multi-modal attention enhancement on the decayed input and the decayed hidden state to obtain an attention-enhanced input; performing recurrent unit state update on the attention-enhanced input and the decayed hidden state to obtain a final hidden state of the current time instant; based on the final hidden state of the current time instant, generating a sepsis risk score.

[0008] According to still another aspect of the present application, there is provided a computing device comprising a processor and a memory having stored therein a computer program which, when executed by the processor, implements the method for early risk prediction of sepsis based on multi-source data as described above. In an embodiment, the processor and the memory can be integrated within the same smart chip.

[0009] Compared with the prior art, the system and method for early risk prediction of sepsis based on multi-source data provided by the present application firstly encode the historical vital signs and test data of a patient, then perform time decay embedding on the historical information based on the current observation and the time interval of the previous time instant to simulate the natural forgetting process of information over time, meanwhile, the system utilizes a mask vector to perform feature decay and padding on the current incomplete input data, and dynamically captures and enhances the internal correlation between different data sources through a multi-modal attention mechanism, finally, updates the recurrent unit state based on the input vector that integrates multi-party information to generate a final risk score. In this way, the information distortion problem caused by irregular clinical data sampling can be effectively overcome, the complex dynamic correlation between heterogeneous data can be deeply mined, and finally more accurate and timely sepsis risk prediction is realized, which provides strong decision support for early clinical intervention. In particular, when the method is implemented through a smart chip, it can run on edge devices in a low-power and high-efficiency manner, providing a powerful intelligent core for real-time clinical monitoring. BRIEF DESCRIPTION OF DRAWINGS

[0010] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description thereof taken in conjunction with the accompanying drawings in which:

[0011] Figure 1 a block diagram of a sepsis early risk prediction system based on multi-source data according to an embodiment of the present application; Figure 2 a data flow diagram of a sepsis early risk prediction system based on multi-source data according to an embodiment of the present application; Figure 3 a flowchart of a sepsis early risk prediction method based on multi-source data according to an embodiment of the present application. DETAILED DESCRIPTION

[0012] Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. It should be understood that the exemplary embodiments described herein are only a part of embodiments of the present application and are thus not to be used to limit the present application, and that the present application is not limited by the exemplary embodiments described herein.

[0013] As used in the present application and in the claims, the indefinite articles "a", "an", and "the" and / or "at least one" do not exclude a plurality. The mere fact that a measure has a plurality should not rule out that at least one of these measures can be used. The phrases "at least one of" followed by a list of two or more items, cover all of the individual items included in the list, as well as any of the individual items, or combination of the individual items, in the list. The phrases "at least one of" followed by a list of two or more items, cover all of the individual items included in the list, as well as any of the individual items, or combination of the individual items, in the list.

[0014] Although the present application makes various references to certain modules in the system according to embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or server. The modules are merely illustrative, and different aspects of the system and method can use different modules.

[0015] Flowcharts have been used herein to illustrate the operations performed by the system according to embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in the exact order shown. Rather, various steps can be handled in reverse order, or simultaneously, as desired. Other operations can also be added to, or removed from, these processes, or one or more steps can be removed from these processes.

[0016] In the technical solution of the present application, a sepsis early risk prediction system based on multi-source data is proposed. Figure 1 a flowchart of a sepsis early risk prediction system based on multi-source data according to an embodiment of the present application.Figure 2 FIG. 1 is a schematic diagram of data flow of a sepsis early risk prediction system based on multi-source data according to an embodiment of the present application. As shown in FIG. 1, the sepsis early risk prediction system based on multi-source data according to an embodiment of the present application comprises a data preprocessing module 310, a sequence encoding module 320, a target patient object information acquisition module 330, a time decay embedding module 340, a feature decay and padding module 350, a multi-modal attention enhancement module 360, a recurrent unit state update module 370, and a sepsis risk score module 380. Figure 1 Figure 2 As shown in FIG. 3, the sepsis early risk prediction system based on multi-source data according to an embodiment of the present application comprises a data preprocessing module 310, a sequence encoding module 320, a target patient object information acquisition module 330, a time decay embedding module 340, a feature decay and padding module 350, a multi-modal attention enhancement module 360, a recurrent unit state update module 370, and a sepsis risk score module 380.

[0017] In particular, the data preprocessing module 310 is configured to preprocess and time-align the historical raw vital sign data and the historical raw test data of the target patient object to obtain a historical data matrix. It should be understood that the raw vital sign data (such as heart rate, blood pressure) is usually recorded continuously or at high frequency, while the test data (such as blood routine, biochemical indicators) is obtained at discrete and irregular time points based on clinical needs. This heterogeneity and irregularity makes it impossible to directly fuse and analyze the two types of data. Therefore, in the technical solution of the present application, the preprocessing and time alignment are performed to convert these multi-source, heterogeneous raw data into a structured data form that is regular, synchronized, and can be used for subsequent sequence encoding and model training, i.e., a historical data matrix, thereby ensuring the input quality of the subsequent risk prediction model and the accuracy and reliability of the final prediction result.

[0018] ​In implementation, first, the system cleans the collected historical raw data. In a clinical environment, raw data often contains obvious outliers or errors due to equipment failure, human recording errors or signal interference, such as body temperature exceeding the normal physiological range, instantaneous impossible spikes in heart rate, etc. The system identifies these data points by setting reasonable physiological threshold or using statistical methods (such as IQR-based outlier detection) and processes them, which can be direct rejection or interpolation using adjacent valid observations to ensure the basic validity of the data. Subsequently, the system standardizes the cleaned data. Due to the huge difference in the value range and dimension of different clinical indicators (such as heart rate in "times / minute" and white blood cell count in "10^9 / L"), if not processed, the features with larger value range will dominate the model training, affecting the convergence speed and learning effect of the model. Therefore, the embodiment adopts Z-score standardization method to convert the value of each feature to a distribution with mean value of 0 and standard deviation of 1, thereby eliminating the scale difference between different features. Through this operation, all features are placed in a unified scale space. Further, since the collection time points of clinical data are asynchronous and irregular, for example, vital signs data may be minute or hour level, while test data may be day level. In order to build a unified time series model, the system first defines a fixed time window, for example, discretizes the time axis into continuous time steps with 1 hour as a unit. Then, all standardized observation values are classified into the corresponding time steps according to their time stamps. If a time step contains multiple observation values of a certain feature, mean, maximum or last value aggregation strategy can be used for merging; if there is no observation value in a time step, the position is explicitly identified as missing. Through the above process, patient data from different sources and different frequencies are integrated into a unified dimension and time reference, finally forming a TxD-dimensional historical data matrix, where T represents the total number of time steps, and D represents the total number of selected clinical features. Each element in this matrix represents the standardized state of a certain clinical indicator of a specific patient at a specific time point, laying a solid foundation for subsequent sequence encoding and dynamic risk assessment.

[0019] In particular, the sequence encoding module 320 is configured to sequence encode the historical data matrix to obtain the hidden state of the previous time. It should be understood that the evolution of the patient's health condition is a dynamic process with time dependence, and the current state is deeply influenced by the historical state. Therefore, in the technical solution of the present application, the sequence encoding is performed on the historical data matrix to extract and compress the discrete and multi-dimensional physiological state sequence of the patient in the past period of time (i.e., the historical time window) into a fixed-dimensional dense vector containing rich time sequence dynamic information. That is, the long-term dependence, trend change and interaction between features in the data are captured by the deep learning model, which provides the indispensable historical context for subsequent judgment and update based on the current new information.

[0020] In a specific implementation, the sequence encoding module can be implemented by using the architecture of a recurrent neural network, such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU). Specifically, the module starts from the first time step (t = 1) of the sequence and processes the data vector at each time step. At each time step t, the calculation unit of the recurrent neural network receives two inputs: the data vector of the current time step (i.e., the t-th row of the historical data matrix), and the hidden state passed from the previous time step; then, through an internal complex gating mechanism (such as the reset gate and update gate in GRU), the calculation unit selectively fuses and updates the current new information with the historical memory to generate the hidden state of the current time step.

[0021] In particular, the target patient object information acquisition module 330 is configured to acquire the current time data vector, the current time interval vector and the current time mask vector of the target patient object. It should be understood that in the actual clinical environment, the measurement of various vital signs and test indicators is not continuous, and the acquisition frequency is different, resulting in irregular characteristics of the data in the time dimension, and some data may be missing at any time. It is far from enough to only acquire the observation value at the current time. Therefore, in the technical solution of the present application, the current time data vector, the current time interval vector and the current time mask vector of the target patient object are acquired; the current time data vector provides the numerical information of the patient's indicators and is the basis for model judgment; the current time interval vector quantifies the freshness or obsolescence of each data, solves the problem of non-uniform time points of data measurement, and enables the model to evaluate the timeliness of historical information; the current time mask vector explicitly identifies the authenticity of the current time, and provides a basis for subsequent targeted data filling and feature attenuation. The synergistic effect of the three vectors ensures the integrity and accuracy of the model input information, which is a key prerequisite for realizing accurate and reliable early risk prediction of sepsis.

[0022] In implementation, the procedure is executed by a target patient object information acquisition module, which connects in real time with the electronic health record (EHR) system, the monitoring device data interface, and the laboratory information system (LIS) of the hospital to ensure the timeliness and accuracy of the data. The execution process includes: First, the system determines the time node t of the current analysis, and then the system queries all data sources related to the target patient to collect all original vital sign data and original test data at the time node t and before. Based on these collected data, the system generates three vectors in parallel. Specifically, the process of generating the current time data vector is as follows: the system creates a d-dimensional vector according to the preset feature list (for example, including heart rate, respiratory rate, body temperature, etc. d features). The system will traverse this feature list, and for each feature in the list, it will retrieve the patient's medical record to find the exact measurement value at the current time t. If found, the measurement value is filled into the corresponding position of the vector. If the measurement value of the feature at time t cannot be found, the system will trace back in time to find the last valid measurement record of the feature, and fill the value of the record as the current value into the vector. This method is called the last observation carry forward. In this way, the system ensures that the data vector is complete at any time; The process of generating the current time interval vector is closely related to the generation of the data vector. The system also traverses the d preset features. For each feature, the system records the time stamp of its last measurement. Then, by calculating the difference between the current time t and the time stamp, the specific time interval is obtained. This difference is filled into the corresponding dimension of the time interval vector. If a feature has a measurement value at the current time t, the time stamp is equal to t, and the time interval is 0. The time interval values of all features together constitute the current time interval vector; The process of generating the current time mask vector is an identification process. After creating a d-dimensional zero vector, the system will traverse all features again. It will check whether each feature has real, newly collected observation data at the current time t. If so, the value of the corresponding dimension in the mask vector is modified from 0 to 1. If there is no new observation data at the current time t (i.e., the value in the data vector is carried over from the past), the value of the corresponding dimension remains 0. This finally generated binary vector clearly indicates the new and old situation of the data.

[0023] In particular, the time-decay embedding module 340 is configured to perform time-decay embedding on the hidden state of the previous time point based on the current time interval vector to obtain the decayed hidden state. It should be understood that in clinical practice, the measurement frequency of different physiological indicators varies greatly. For example, heart rate can be continuously monitored, while blood test results can be updated only once every few hours or even a day. This means that between any two consecutive observation time points, there is a time interval of varying length. If the model does not take into account these time intervals, it will incorrectly treat obsolete data from hours or even days ago as equal to fresh data just measured, which seriously violates the clinical diagnosis and treatment logic. Because the patient's condition is dynamic, the value of the information will decay over time. Therefore, the present application introduces a time decay mechanism. The hidden state of the previous time point is a condensed representation of all historical information up to the previous time point, which contains the memory of the patient's state. However, each dimension of this memory is associated with a specific clinical feature, and its timeliness is different. By using the current time interval vector, this step can calculate a specific decay weight for each dimension of the hidden state. For those features that have just been updated (small time interval), their historical information in the hidden state should be retained more; while for those features that have not been updated for a long time (large time interval), their corresponding historical information should be greatly decayed to reduce their interference with the current state judgment. This is particularly important for diseases such as sepsis, which develop rapidly, and can effectively prevent the model from being misled by outdated normal indicators, thereby missing the best opportunity for early warning.

[0024] In a specific implementation, first, the system performs linear transformation of the decay parameter. In this process, the system uses two learnable parameters, i.e., a weight matrix and a bias vector, to perform a linear mapping of the current time interval vector. Here, the weight matrix and the bias vector are automatically learned by the model during the training process through the backpropagation algorithm, so that the model can autonomously determine the decay rate of different clinical features over time; for example, the model can learn that the timeliness of blood lactate value is very strong, and its decay rate should be higher than that of body temperature and other relatively slow-changing indicators. Second, the system applies a non-negative activation function. The result of linear transformation is input into a max function to ensure that the decay rate is always non-negative. In clinical logic, the passage of time only makes information more obsolete, and its impact should be decayed, not enhanced, so the decay rate cannot be negative. This step ensures the stability of the calculation and the interpretability of the results; Then, the system computes an exponential decay factor. The non-negative result from the previous step is taken as the power of a natural exponential function. This exponential function is the core of the decay effect. Since its input is non-positive, its output value range is between (0, 1]. When the time interval is 0, the input of the exponential function is 0, and the output is 1, indicating no decay. As the time interval increases, the input of the exponential function becomes more negative, and its output value quickly approaches 0, indicating stronger decay. Finally, the system performs element-wise decay application. The decay factor computed in the previous step is a vector with the same dimension as the hidden state, and the system element-wise multiplies this decay factor vector with the hidden state vector at the previous time step. This means that each element in the hidden state vector at the previous time step is scaled by its corresponding decay factor. After this operation, the final output, i.e., the decayed hidden state, is obtained.

[0025] More specifically, based on the time interval vector at the current time step, the hidden state at the previous time step is time-decay embedded as follows: , where, is the time interval vector at the current time step, and are learnable parameters for controlling the decay rate, is the hidden state at the previous time step, is the max function.

[0026] In particular, the feature attenuation and padding module 350 is configured to perform feature attenuation and padding on the current time data vector based on the current time mask vector to obtain the attenuated input. It should be understood that in clinical monitoring and diagnosis practice, data acquisition at any time point is incomplete. For example, vital signs can be continuously or high-frequency monitored, while blood biochemical tests and other indicators are usually performed at intervals of several hours or even longer. This results in the current time data vector directly obtained actually being a mixture containing part of the real recent observation values and part of the obsolete values carried over from historical records. If these actual missing values (i.e., obsolete values) are not effectively processed, the model can make inaccurate judgments based on outdated information, especially in critical conditions such as sepsis, which can lead to early warning delay or failure. Therefore, in the technical solution of the present application, the current time data vector is subjected to feature attenuation and padding based on the current time mask vector. This mechanism introduces the core idea of feature attenuation, that is, according to the time span of data missing (calculated by the input attenuation rate of the time interval vector), the weights of the above two information sources in the fusion process are dynamically adjusted. The longer the time span of data missing, the more the credibility of the personalized estimation is attenuated, and the model will rely more on the robust global mean. In this way, the rationality and robustness of the padding value are ensured, and the attenuated input generated finally not only retains all the real observation information without loss, but also gives the most comprehensive estimate of the missing information under the current conditions, thereby providing an information complete and high-quality input for the subsequent module.

[0027] In a specific implementation, first, based on the current time mask vector, the observation part vector is separated from the current time data vector; then, based on the input attenuation rate, the observation part vector is subjected to content padding to obtain the observation part padding vector; and then, the observation part padding vector and the observation part vector are fused to obtain the attenuated input.

[0028] Specifically, the specific steps of content padding on the observation part vector include: first, based on the current time interval vector, the input attenuation rate is calculated. In this process, the system receives the previously obtained current time interval vector as input, and calculates the input attenuation rate vector element by element through a pre-set, monotonically decreasing function (for example, an exponential attenuation function , wherein is a learnable attenuation coefficient) to calculate the input attenuation rate vector element by element In this way, the larger the time interval, the smaller the corresponding attenuation rate, thereby quantifying the aging degree of information. Next, the decayed hidden state is feature-decoded to obtain the personalized estimate of the current missing feature. In this process, the system takes the decayed hidden state that has already been adjusted for temporal relevance as input from the previous module. This vector is mapped from the high-dimensional hidden space back to the original feature space through a decoder network (e.g., one or more fully connected layers), generating a personalized prediction vector for all current features, i.e., the personalized estimate; Then, based on the input decay rate, the personalized estimate of the current missing feature and the global mean of the current time data vector are weighted and fused to obtain the fused fill-in value. In this process, the personalized estimate of the current missing feature and the global mean of the current time data vector are weighted and fused as follows: , wherein, is the personalized estimate of the current missing feature, is the global mean of the current time data vector, is the input decay rate; the formula is weighted and fused by and This pair of complementary, time-varying weights, element-wise weighted average of the personalized estimate and the global mean, resulting in a more reliable fusion fill-in value vector that integrates both sources of information ; Subsequently, the fused fill-in value is applied with the missing mask to obtain the observed part fill-in vector. In this process, the fused fill-in value is applied with the missing mask as follows: , wherein, is the current time mask vector. It is worth mentioning that this mask is dynamically generated in the information acquisition stage, by systematically checking whether each clinical feature has a new measurement value at the current time to determine its zero or one state.

[0029] In particular, the multi-modal attention enhancement module 360 is configured to perform multi-modal attention enhancement on the decayed input and the decayed hidden state to obtain an attention-enhanced input. It should be understood that early risk prediction of sepsis requires comprehensive analysis of multi-dimensional and heterogeneous data of patients, including physiological signs and laboratory test results, etc. These different modalities of data carry unique but interrelated information. At the same time, the historical health status and physiological change trend of the patient, i.e., the hidden state, are crucial for understanding the current clinical manifestations. Simply concatenating or linearly combining the current decayed input and the decayed hidden state cannot fully capture the complex nonlinear dependency between modalities and their deep interaction with the historical state. In the technical solution of the present application, by introducing a multi-modal attention enhancement mechanism, the system can adaptively learn and allocate the contribution weights of different modalities of features and historical hidden states to the current prediction task, highlight key information, and suppress irrelevant or redundant information, thereby improving the perception ability of the model to the evolution of the disease and the prediction accuracy.

[0030] In specific implementation, first, the decayed input is divided into modalities to obtain a vital sign modality input vector and a test result modality input vector. It should be understood that in early risk prediction of sepsis, the clinical data of a patient usually includes multiple types, such as vital signs (e.g., body temperature, heart rate, respiratory rate, blood pressure) and laboratory test results (e.g., white blood cell count, C-reactive protein, lactic acid). These different modalities of data have different semantics, statistical properties, and measurement frequencies, and they have different indications for the occurrence and development of sepsis. Simply stacking these different modalities of information together as a single input can make it difficult for the model to effectively distinguish and learn the complex relationships within and between modalities. Therefore, in the technical solution of the present application, the decayed input is divided into modalities to classify and separate the features at the current time point mixed together according to their inherent attributes (e.g., physiological measurements or laboratory test values). In this way, the system can design or apply more appropriate processing strategies for each modality, especially in the subsequent cross-modal attention mechanism, which can promote more accurate and targeted feature interaction between different modalities, thereby improving the model's ability to capture the potential information of multi-source data and the final prediction performance.

[0031] In specific implementation, it is assumed that the decayed input is a vector containing multiple features. In system design, it is predefined which features belong to the vital sign modality and which features belong to the test result modality. Therefore, the process of modality division is to segment and intercept or index extract the decayed input vector according to the preset feature grouping rules, so as to construct the vital sign modality input vector and the test result modality input vector respectively. For example, if a decayed input vector contains a total of N features, and it is known that the first v features represent vital sign data and the next r features represent test result data, then the vital sign modality input vector is obtained by intercepting the first v elements of the original vector, and the test result modality input vector is obtained by intercepting from the (v+1)th element to the (v+r)th element. This method ensures that the decayed input can be accurately decomposed into two independent sub-vectors with specific modality semantics, which will be used as independent inputs for subsequent feature interaction based on the cross-modality attention mechanism, ensuring that different modality information can be processed and fused more effectively; Further, the vital sign modality input vector, the test result modality input vector and the decayed hidden state are subjected to feature interaction based on the cross-modality attention mechanism to obtain the attention-enhanced input. It can be understood that in the early risk prediction task of sepsis, vital sign data and laboratory test result data as different modalities each carry unique clinical information, while the decayed hidden state provides key context information on the patient's historical time series. These information are related to each other, but their indications for the current condition are not equivalent. Simply concatenating or averaging these modality information cannot fully capture the complex nonlinear dependence relationship between them, nor can it dynamically highlight the most critical information for prediction. Therefore, in the technical solution of the present application, by introducing the cross-modality attention mechanism, the system can learn and adaptively allocate the contribution weights of different modality features and historical hidden states to the current prediction. Specifically, the hidden state can query the input vectors of different modalities to determine which modalities or which features in the modalities are most important for understanding the condition at the current time. This dynamic and selective information fusion method enables the model to more accurately capture the subtle changes and potential associations of multi-source data before the onset of sepsis, thereby improving the accuracy and robustness of the prediction.

[0032] Specifically, the vital sign modality input vector, the test result modality input vector and the decayed hidden state are subjected to feature interaction based on the cross-modality attention mechanism according to the following formula: , , , wherein, the decayed hidden state, the vital sign modality input vector or the test result modality input vector, the weight matrix, the weight vector.

[0033] Specifically, in this process, first, the decayed hidden state vector is multiplied by a learnable weight matrix, and the vital sign modality input vector or the test result modality input vector is multiplied by another learnable weight matrix, the vectors of different sources are projected into a common feature space, then the two converted vectors are added to fuse their information; Next, the hyperbolic tangent nonlinear activation function is applied to the added result to introduce the nonlinear expression ability of the model, and finally, the nonlinear converted result is inner product with a learnable weight vector to get a scalar value, which is the attention score ; that is, by using a learnable weight matrix to map different modalities and hidden states to the same dimensional space, so that they can be directly compared and interacted. The use of nonlinear activation function makes the model able to capture more complex nonlinear relationships between data. The final inner product operation summarizes this interaction into a single attention score, which intuitively quantifies the relevance or importance between the hidden state and the specific modality input vector; Then, the calculated attention score is then normalized by the Softmax function to obtain the attention weight In this process, for each attention score, first, calculate its exponential value; Then, divide this exponential value by the sum of the exponential values of all related attention scores (i.e. for all vital sign modality input vectors and test result modality input vectors), so that each modality input vector will get a weight value between 0 and 1, and the sum of all these weight values is equal to 1. This step converts the original attention score into a probability distribution, so that the weight of each modality input vector can intuitively represent its relative importance or contribution to the final result; Further, after obtaining the attention weight corresponding to each modality input vector, the vital sign modality input vector is multiplied by its corresponding attention weight, and the test result modality input vector is multiplied by its corresponding attention weight, then the two weighted vectors are added to get the final attention enhanced input . This enhanced input vector will be a highly condensed and targeted input, passed to the subsequent recurrent unit state update module, thereby improving the model's perception and prediction ability of sepsis risk.

[0034] In particular, the recurrent unit state updating module 370 is configured to perform recurrent unit state updating on the attention-enhanced input and the decayed hidden state to obtain a final hidden state at the current time. It should be understood that early risk prediction of sepsis is a typical time series prediction problem, and the clinical state of a patient dynamically changes over time. The condition at the current time is not only affected by the current observation data, but also closely related to the patient's historical physiological trend and medical events. The decayed hidden state summarizes the historical information and considers the time decay effect, while the attention-enhanced input provides key information at the current time after multi-modal fusion. Effective recurrent unit state updating of these two kinds of information can effectively integrate the input information at the current time after multi-modal attention enhancement and the historical context information represented by the decayed hidden state at the previous time, thereby generating a final hidden state that comprehensively reflects the current condition of the patient, and providing a basis for subsequent sepsis risk scoring.

[0035] In a specific implementation, first, the recurrent unit receives the attention-enhanced input at the current time and the decayed hidden state at the previous time as inputs; then, the recurrent unit calculates a series of gate vectors, which are usually generated by an activation function (such as a sigmoid function), and the output value is between 0 and 1, which is used to control the amount of information passing through: Update gate: the attention-enhanced input at the current time and the decayed hidden state at the previous time are linearly transformed (i.e., multiplied by their respective weight matrices and added with a bias), then the results are added, and the result is compressed to 0 to 1 through a Sigmoid activation function. The value of this gate determines how much information of the input at the current time is used to update the hidden state. Through learnable parameters and nonlinear activation, the model can adaptively control the information flow and determine the fusion ratio of new and old information; Reset gate: the attention-enhanced input at the current time and the decayed hidden state at the previous time are linearly transformed, added, and processed through a Sigmoid activation function. The value of this gate determines how much information of the hidden state at the previous time should be forgotten so that it can focus on the current input when calculating the new candidate hidden state. It allows the model to selectively ignore irrelevant historical information, thereby better adapting to new data at the current time; Candidate hidden state: First, the decayed hidden state of the previous time step is multiplied element-wise with the reset gate to selectively retain relevant historical information. Then, this gated historical information is linearly transformed with the attention-enhanced input of the current time step, and the sum is passed through a hyperbolic tangent (tanh) activation function. This result is the new candidate hidden state, representing the new information that can be formed at the current time step without considering the final gating completely. It combines the partially filtered historical context and the current input to generate a preliminary new state representation and increases the model's expressive power through a nonlinear transformation. Further, by combining the update gate and the candidate hidden state, and combining the hidden state of the previous time step, the recurrent unit generates the final hidden state of the current time step. Specifically, the decayed hidden state of the previous time step is multiplied element-wise with the update gate to obtain the old state part that needs to be retained. At the same time, the candidate hidden state is multiplied element-wise with (1 minus the update gate) to obtain the new state part that needs to be introduced. Finally, the two parts are added to form the final hidden state of the current time step. It is worth mentioning that this weighted average or more complex nonlinear combination aims to intelligently integrate new information and old memories. The update gate determines how much old state needs to be retained and how much new state needs to be introduced, so that the recurrent unit can dynamically adjust the degree of dependence on historical information and current information, thereby overcoming the problem of gradient vanishing or explosion in traditional RNNs, effectively processing long sequence data, and capturing long-term dynamic patterns in the evolution of patient conditions.

[0036] In particular, the sepsis risk score module 380 is configured to generate a sepsis risk score based on the final hidden state of the current time step. It should be understood that although the previous steps have obtained a highly abstract and informative final hidden state of the current time step through sequence encoding, time decay embedding, feature decay and padding, and multi-modal attention enhancement and recurrent unit state update, this hidden state itself is a vector and does not directly provide clinical risk information for doctors to make decisions. In the technical solution of the present application, the comprehensive understanding of the patient's condition by the model at the current time is quantified into an intuitive and operable risk value for the reference of clinical decision makers, thereby realizing early warning and intervention of sepsis.

[0037] wherein the sepsis risk score is a quantitative indicator, usually a value within a certain range (e.g. between 0 and 1), which represents the model's prediction of the patient's probability or risk level of developing sepsis at the current time step. The higher the score, the greater the risk of the patient developing sepsis.

[0038] In implementation, first, the final hidden state at the current time is mapped to a lower dimensional or specific dimensional space through one or more linear transformation layers to obtain a linearly transformed result vector, and the linear transformation layer is used to learn the feature combination in the hidden state that is most relevant to the prediction of sepsis risk, which allows the model to learn the linear combination features directly related to the final prediction target (sepsis risk) from the highly abstract hidden state, thereby realizing the transition from deep representation to shallow explainable features. Then, the linearly transformed result vector is converted into a final risk score with specific meaning (for example, probability) through an activation function. In the embodiment of the present application, for a binary classification task (such as "at risk of sepsis" or "not at risk of sepsis"), a Sigmoid function can be used, and in this process, the Sigmoid activation function can map any real number value to the interval (0, 1), so that the output can be directly interpreted as a probability value, thereby intuitively representing the risk of sepsis in the patient. Such a probabilistic output is very suitable for clinical risk assessment. Finally, the value processed by the activation function is the sepsis risk score at the current time.

[0039] As described above, the sepsis early risk prediction system based on multi-source data 300 according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with a sepsis early risk prediction algorithm based on multi-source data. In one possible implementation, the sepsis early risk prediction system based on multi-source data 300 according to the embodiment of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the sepsis early risk prediction system based on multi-source data 300 can be a software module in the operating system of the wireless terminal, or can be an application developed for the wireless terminal; of course, the sepsis early risk prediction system based on multi-source data 300 can also be one of the many hardware modules of the wireless terminal.

[0040] Alternatively, in another example, the sepsis early risk prediction system based on multi-source data 300 and the wireless terminal can also be separate devices, and the sepsis early risk prediction system based on multi-source data 300 can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in a conventional data format.

[0041] Further, a sepsis early risk prediction method based on multi-source data is also provided.

[0042] Figure 3 A flowchart of the sepsis early risk prediction method based on multi-source data according to the embodiment of the present application is shown in FIG. 6. As shown in FIG. 6, the sepsis early risk prediction method based on multi-source data according to the embodiment of the present application includes the following steps. Figure 3As shown, the method for early risk prediction of sepsis based on multi-source data according to the embodiments of the present application comprises the following steps: S1, preprocessing and time alignment of historical raw vital sign data and historical raw test data of a target patient object to obtain a historical data matrix; S2, sequence encoding of the historical data matrix to obtain a hidden state at the last time; S3, obtaining a current time data vector, a current time interval vector and a current time mask vector of the target patient object; S4, time decay embedding of the hidden state at the last time based on the current time interval vector to obtain a decayed hidden state; S5, feature decay and padding of the current time data vector based on the current time mask vector to obtain a decayed input; S6, multi-modal attention enhancement of the decayed input and the decayed hidden state to obtain an attention-enhanced input; S7, state update of a recurrent unit of the attention-enhanced input and the decayed hidden state to obtain a final hidden state at the current time; and S8, generating a sepsis risk score based on the final hidden state at the current time.

[0043] In summary, the method for early risk prediction of sepsis based on multi-source data according to the embodiments of the present application is illustrated, which firstly encodes the historical vital signs and test data of a patient, then performs time decay embedding of the historical information based on the current observation and the time interval at the last time to simulate the natural forgetting process of information over time, meanwhile, the system utilizes a mask vector to perform feature decay and padding on the current incomplete input data, and dynamically captures and enhances the internal correlation between different data sources through a multi-modal attention mechanism, finally, the state of the recurrent unit is updated based on the input vector that integrates multiple information to generate a final risk score. In this way, the information distortion problem caused by irregular clinical data sampling can be effectively overcome, the complex dynamic correlation between heterogeneous data can be deeply mined, and finally more accurate and timely sepsis risk prediction is realized, providing strong decision support for early clinical intervention.

[0044] In a preferred embodiment, the method for early risk prediction of sepsis described in the present application can be deployed on a dedicated intelligent chip. The intelligent chip can include an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA) inside, which internally solidifies the hardware logic for executing the method of the present application. Specifically, the data preprocessing, sequence encoding, time decay embedding, feature decay and padding, multi-modal attention enhancement, recurrent unit state update and sepsis risk score steps involved in the method can be realized by independent hardware processing units or pipeline structures designed inside the chip.

[0045] For example, the recurrent neural network operations in the sequence encoding and recurrent unit state updating steps can be accelerated by designing specialized matrix multiplication and nonlinear activation function computation units on the chip. Similarly, the attention weight computation in the multi-modal attention enhancement step can also be efficiently performed by dedicated hardware logic. By hardwareizing the complex algorithmic model, the smart chip is able to complete real-time, continuous sepsis risk prediction with extremely low latency and power consumption, which is very suitable for integration into medical devices such as bedside monitors, wearable vital sign monitoring devices, or central monitoring stations, thereby realizing efficient intelligent decision support at the edge. The smart chip can provide standard data input interfaces and risk score output interfaces to the outside, facilitating integration with existing medical information systems.

[0046] Embodiments of the present disclosure have been described above, the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes are obvious to those skilled in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles, practical applications, or improvements to the technology in the market of the embodiments, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

Claims

1. A sepsis early risk prediction system based on multi-source data, characterized in that, include: The data preprocessing module is used to preprocess and time-align the historical raw vital signs data and historical raw test data of the target patients to obtain a historical data matrix. The sequence encoding module is used to encode the historical data matrix to obtain the hidden state of the previous time step; The target patient object information acquisition module is used to acquire the current time data vector, the current time interval vector, and the current time mask vector of the target patient object; The time decay embedding module is used to perform time decay embedding on the hidden state of the previous time step based on the time interval vector of the current time step to obtain the decayed hidden state. The feature attenuation and padding module is used to perform feature attenuation and padding on the current time data vector based on the current time mask vector to obtain the attenuated input; The multimodal attention enhancement module is used to perform multimodal attention enhancement on the decayed input and the decayed hidden state to obtain the attention-enhanced input; The recurrent unit state update module is used to update the recurrent unit state of the attention-enhanced input and the decayed hidden state to obtain the final hidden state at the current moment. The sepsis risk scoring module is used to generate a sepsis risk score based on the final hidden state at the current moment.

2. The sepsis early risk prediction system based on multi-source data according to claim 1, characterized in that, The time decay embedding module is used to: embed the hidden state of the previous time step into time decay mode based on the current time interval vector using the following formula: in, This is the time interval vector at the current moment. and These are learnable parameters used to control the decay rate. This is the hidden state from the previous moment. This is the function for finding the maximum value.

3. The sepsis early risk prediction system based on multi-source data according to claim 1, characterized in that, The feature attenuation and filling module includes: The observation vector separation unit is used to separate the observation vector from the current time data vector based on the current time mask vector; The content filling unit is used to fill the content of the observation part vector based on the input decay rate to obtain the observation part filled vector; A fusion unit is used to fuse the observation part fill vector and the observation part vector to obtain the attenuated input.

4. The sepsis early risk prediction system based on multi-source data according to claim 3, characterized in that, The content filling unit includes: The input attenuation rate calculation subunit is used to calculate the input attenuation rate based on the current time interval vector. The feature decoding subunit is used to perform feature decoding on the decayed hidden state to obtain a personalized estimate of the current missing feature; The weighted fusion subunit is used to perform weighted fusion of the personalized estimate of the current missing feature and the global mean of the data vector at the current time step based on the input decay rate to obtain the fused filling value; The missing mask subunit is used to apply a missing mask to the fused padding values ​​to obtain the padding vector of the observed portion.

5. The sepsis early risk prediction system based on multi-source data according to claim 4, characterized in that, The weighted fusion subunit is used to: perform a weighted fusion of the personalized estimate of the current missing feature and the global mean of the data vector at the current time using the following formula: in, This is a personalized estimate of the currently missing features. The global mean of the data vector at the current moment. The input is the attenuation rate.

6. The sepsis early risk prediction system based on multi-source data according to claim 4, characterized in that, The missing mask subunit is used to apply a missing mask to the fused padding value using the following formula: in, This is the mask vector at the current moment.

7. The sepsis early risk prediction system based on multi-source data according to claim 1, characterized in that, The multimodal attention enhancement module includes: The modality partitioning unit is used to partition the attenuated input into modalities to obtain the vital signs modal input vector and the test result modal input vector. The feature interaction unit is used to perform feature interaction based on cross-modal attention mechanism on the vital sign modal input vector, the test result modal input vector and the decayed hidden state to obtain the attention-enhanced input.

8. The sepsis early risk prediction system based on multi-source data according to claim 7, characterized in that, The feature interaction unit is used to perform feature interaction based on a cross-modal attention mechanism on the vital sign modality input vector, the test result modality input vector, and the decayed hidden state using the following formula: in, This is the hidden state after decay. This is either the vital signs modality input vector or the test result modality input vector. This is the weight matrix. This is the weight vector.

9. A method for predicting the early risk of sepsis based on multi-source data, characterized in that, include: The historical raw vital signs data and historical raw laboratory data of the target patients were preprocessed and time-aligned to obtain a historical data matrix; Sequence encoding is performed on the historical data matrix to obtain the hidden state of the previous time step; Obtain the current time data vector, current time interval vector, and current time mask vector of the target patient object; Based on the current time interval vector, the hidden state of the previous time step is embedded with time decay to obtain the decayed hidden state; Based on the mask vector at the current time, feature attenuation and padding are performed on the data vector at the current time to obtain the attenuated input; Multimodal attention enhancement is applied to the decayed input and the decayed hidden state to obtain the attention-enhanced input; The attention-enhanced input and the decayed hidden state are updated using a recurrent unit state to obtain the final hidden state at the current time step. Based on the final hidden state at the current moment, a sepsis risk score is generated.

Images