Sepsis dynamic risk prediction model and device based on deep learning

Abstract

This invention relates to the field of medical information technology, and particularly to a deep learning-based dynamic risk prediction model and device for sepsis. The model includes: acquiring multi-source time-series data of sepsis patients within a preset time window, including physiological indicator sequences, laboratory indicator sequences, and treatment intervention sequences; constructing intervention background features based on the treatment intervention sequences, generating corresponding dynamic risk reference benchmarks, and performing reference calibration processing on the physiological indicators to obtain calibrated physiological state features; further constructing response relationship features and compensatory maintenance features between intervention changes and physiological changes, and performing mild anomaly identification and time-dimensional cumulative analysis on the calibrated physiological state features to form risk accumulation features and risk change trend features; based on this, the multi-type features are fused and input into a deep learning time-series prediction network to output the sepsis deterioration risk value and risk change trend within a future time period, and to generate risk level determination results and early warning information.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a deep learning-based dynamic risk prediction model and device for sepsis. Background Technology

[0002] Sepsis is a systemic inflammatory response syndrome caused by infection, characterized by rapid disease progression and highly variable clinical outcomes. To achieve early identification and dynamic monitoring, risk prediction models based on machine learning or deep learning have been increasingly adopted in recent years to jointly model patients' physiological indicators, laboratory test results, and treatment intervention information. Existing technologies typically construct multivariate time series prediction models to analyze monitoring data within a certain time window and output a risk value or warning level for sepsis exacerbation in the future. In some protocols, treatment intervention variables are also included as input features to enhance the model's ability to express clinical status. During the continuous treatment intervention for sepsis patients, the values ​​of physiological indicators are often directly affected by external support measures, such as vasopressors, mechanical ventilation, or fluid resuscitation, which may maintain the indicators in a relatively stable range. Existing risk prediction models are usually based on the original physiological indicators or simply integrate intervention variables, making it difficult to distinguish between the natural stable state of physiological indicators and the state maintained by intervention. As a result, it is difficult to accurately depict the evolution trend of real pathological risk under the condition of intervention masking. Summary of the Invention

[0003] To overcome the above shortcomings, this invention provides a deep learning-based dynamic risk prediction model and device for sepsis, aiming to improve the problem of accurately depicting the evolution trend of real pathological risk under intervention masking conditions.

[0004] In a first aspect, the present invention provides the following technical solution: a sepsis dynamic risk prediction model based on deep learning, comprising the following steps: S1. Acquire multi-source time series data of sepsis patients within a preset time window, wherein the multi-source time series data includes physiological index sequences, laboratory index sequences, and treatment intervention sequences; S2. Construct intervention background features based on treatment intervention sequences, characterize the intervention intensity, duration and trend of intervention at the current time point, and generate corresponding dynamic risk reference benchmarks based on intervention background features; S3. The physiological indicator sequence is reference-calibrated with the dynamic risk reference benchmark to obtain calibrated physiological state characteristics that reflect the degree of deviation of the physiological indicators from the intervention background. S4. Based on the response relationship between the treatment intervention sequence and the physiological indicator sequence, construct the compensatory maintenance characteristics; S5. Perform mild abnormality identification and time-dimensional cumulative analysis on the calibrated physiological state characteristics to construct risk accumulation characteristics and risk change trend characteristics; S6. The calibration physiological state characteristics, compensatory maintenance characteristics and risk accumulation characteristics are fused and input into the deep learning time series prediction network to output the sepsis deterioration risk value and risk change trend in the preset future time period. S7. Based on the risk value, compensation maintenance characteristics, and risk change trends, generate risk level determination results and early warning information.

[0005] By adopting the above technical solution, it is possible to construct intervention background features and generate dynamic risk reference benchmarks, transforming physiological indicators from the original observation value space into a calibrated deviation representation relative to intervention conditions, and realizing structured modeling of the patient's physiological state under the background of treatment intervention. On this basis, by combining the modeling of the response relationship between intervention changes and physiological changes and the time accumulation analysis of mild abnormal states, risk prediction is based on the dynamic state evolution process under the constraints of intervention conditions. It can conditionally calibrate the physiological state under the masking condition of continuous intervention and characterize the risk evolution trend in the time dimension, thereby improving the problem in the existing technology that it is difficult to distinguish between the natural stable state and the intervention maintenance state and difficult to accurately characterize the evolution trend of real pathological risk.

[0006] Preferably, in S1, acquiring multi-source time-series data of sepsis patients within a preset time window includes: Set a preset time window length and construct a sliding time window with the current time point as the end time; Physiological index data, laboratory test data, and treatment intervention data are collected within a sliding time window, and corresponding timestamp information is recorded for each data point. Based on timestamp information, multi-source time series data with different sampling frequencies are reconstructed using a unified time axis, mapping various types of data to a unified time scale. Time alignment processing is performed on the mapped multi-source time series data to make various types of data correspond to each other at the same time scale; Missing time nodes in multi-source time series data are imputed; The processed multi-source time series data are numerically normalized, and a multi-dimensional time series input matrix is ​​constructed.

[0007] Preferably, in S2, the construction of intervention background features based on the treatment intervention sequence includes: Intensity extraction is performed on the treatment intervention sequence to obtain the current dose value or support intensity value of various intervention measures within the time window; The duration of treatment intervention sequences was statistically analyzed to calculate the duration and cumulative exposure of various intervention measures within the time window. We performed trend analysis on the treatment intervention sequence to extract the rate of change and fluctuation characteristics of the intervention intensity within the time window; The intensity features, duration features, and trend features are concatenated to construct the intervention background feature vector.

[0008] Preferably, in S2, the generation of the corresponding dynamic risk reference benchmark based on the intervention background characteristics includes: The intervention background feature vector is input into the reference benchmark generation network, which is a multi-layer neural network structure. The reference benchmark generation network performs feature mapping processing on the intervention background feature vector and outputs reference benchmark values ​​corresponding to each key physiological indicator in the physiological indicator sequence. According to the categories of physiological indicators, corresponding dynamic risk reference sequences are constructed so that the dynamic risk reference sequences are aligned with the physiological indicator sequences in the time dimension.

[0009] Preferably, in S3, the reference calibration process of comparing the physiological indicator sequence with the dynamic risk reference baseline includes: Align the physiological indicator sequence with the corresponding dynamic risk reference baseline sequence in the time dimension, so that each key physiological indicator can establish a correspondence with the corresponding reference baseline value at the same time scale. For each key physiological indicator, the difference characteristics between the physiological indicator value and the corresponding reference value are calculated at each time point; The difference features are normalized by amplitude or proportionally mapped to obtain the calibration deviation features. The calibration deviation features corresponding to each key physiological indicator are combined in chronological order to construct a calibration physiological state feature sequence.

[0010] Preferably, in S4, the construction of the compensatory maintenance feature includes: Temporal difference processing is performed on the treatment intervention sequence to extract the variation characteristics of intervention intensity between adjacent time nodes; Temporal difference processing is performed on the physiological indicator sequence to extract the variation characteristics of the physiological indicators between adjacent time nodes; Based on the temporal correspondence between intervention change characteristics and physiological indicator change characteristics, a response relationship feature between intervention change and physiological indicator change is constructed. Extract physiological fluctuation features that reflect autonomic regulatory activities, including heart rate change features, respiratory rate change features, or blood pressure fluctuation features; The response relationship features and physiological fluctuation features are concatenated to construct a compensatory maintenance feature sequence, and its time dimension is kept consistent with the calibration physiological state feature sequence.

[0011] Preferably, in S5, the identification of mild abnormalities and the time-dimensional cumulative analysis of the calibrated physiological state characteristics include: For each key physiological indicator in the calibration physiological state characteristics, a threshold for mild abnormality intervals is preset; Within the time window, determine whether the calibration deviation of each key physiological indicator is within the corresponding mild abnormality range at each time node, and mark the mild abnormality state. The duration and frequency of mild abnormal states are statistically analyzed over time to construct a cumulative feature of mild abnormalities for a single indicator. The mild abnormality accumulation characteristics of multiple key physiological indicators are combined and processed to construct multi-indicator superimposed accumulation characteristics; Based on the cumulative characteristics of mild anomalies in single indicators and the cumulative characteristics of multiple indicators, the cumulative change trend characteristics are extracted in the time dimension to form a risk accumulation feature sequence.

[0012] Preferably, in S6, the fusion of calibration physiological state characteristics, compensatory maintenance characteristics, and risk accumulation characteristics includes: The features of calibrated physiological state, compensatory maintenance and risk accumulation are respectively encoded, and the features from different sources are mapped to a unified feature representation space. The encoded features are aligned according to the time dimension to construct a multi-channel time series feature tensor; Channel-weighted processing is performed on the multi-channel time series feature tensor to generate fusion weights; The features of each channel are weighted and aggregated according to the fusion weight to form a unified fusion feature sequence; The fused feature sequence is used as input to a deep learning temporal prediction network.

[0013] Preferably, in S6, the deep learning temporal prediction network includes: A deep neural network structure containing a time modeling layer is constructed, wherein the time modeling layer is used to model the temporal dependencies of the fused feature sequence; The time modeling layer includes at least one of a recurrent neural network structure, a gated recurrent unit structure, or a self-attention structure. A feature mapping layer is set after the time modeling layer to perform non-linear mapping processing on the sequence features after time modeling; Set up an output layer to perform regression or classification processing on the mapped features and output the risk value of sepsis exacerbation and the trend of risk change within a preset future time period. A deep learning time series prediction network was trained using historical labeled data and then used for risk prediction after training.

[0014] Secondly, the present invention provides the following technical solution: a sepsis dynamic risk prediction device based on deep learning, the device comprising: The data acquisition module is used to acquire multi-source time series data of sepsis patients within a preset time window. The multi-source time series data includes physiological index sequences, laboratory index sequences, and treatment intervention sequences. The intervention background construction module is used to construct intervention background features based on the treatment intervention sequence, and to characterize the intervention intensity, intervention duration and intervention change trend at the current time point. At the same time, it generates a corresponding dynamic risk reference benchmark based on the intervention background features. The reference calibration module is used to perform reference calibration processing on the physiological indicator sequence and the dynamic risk reference benchmark to obtain calibrated physiological state characteristics. A compensatory maintenance feature construction module is used to construct compensatory maintenance features based on the response relationship between the treatment intervention sequence and the physiological indicator sequence. The risk accumulation analysis module is used to identify mild abnormalities and perform time-dimensional accumulation analysis on the calibrated physiological state characteristics, and to construct risk accumulation characteristics and risk change trend characteristics. The feature fusion and prediction module is used to fuse the calibration physiological state features, the compensatory maintenance features and the risk accumulation features, and input them into a deep learning time series prediction network to output the sepsis exacerbation risk value and risk change trend within a preset future time period. The risk assessment module is used to generate risk level assessment results and early warning information based on the risk value, the compensation maintenance characteristics, and the risk change trend.

[0015] The present invention has the following beneficial effects: 1. In this invention, by constructing intervention background features and generating dynamic risk reference benchmarks, reference calibration modeling of physiological indicators under treatment intervention conditions is realized, which solves the problem that it is difficult to distinguish between natural stable state and intervention maintenance state when risk judgment is made based solely on original physiological indicators under continuous intervention maintenance state.

[0016] 2. In this invention, by performing time difference processing on the treatment intervention sequence and the physiological indicator sequence and constructing the response relationship features between intervention changes and physiological changes, and by combining the self-regulation fluctuation features, a compensatory maintenance feature sequence is formed, thereby realizing the modeling of the degree of intervention dependence and physiological response pattern. This solves the problem in the prior art that the risk characterization is incomplete because it is based only on the state value without considering how the intervention maintains the state.

[0017] 3. In this invention, by identifying mild abnormality intervals in the characteristics of calibrated physiological states and performing duration statistics, frequency statistics, and multi-indicator superposition analysis in the time dimension, a structured model of the accumulation process of mild abnormal states is realized, which solves the problem that it is difficult to identify the gradual evolution process of risk when relying solely on severe abnormalities at a single time point for risk judgment. Attached Figure Description

[0018] Figure 1 This is a flowchart of the sepsis dynamic risk prediction model based on deep learning proposed in this invention; Figure 2 This is an architectural diagram of the sepsis dynamic risk prediction device based on deep learning proposed in this invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0020] Example 1: In a first embodiment of the present invention, the present invention provides a sepsis dynamic risk prediction model and device based on deep learning, such as... Figure 1 As shown, it includes the following steps: S1. Acquire multi-source time series data of sepsis patients within a preset time window, wherein the multi-source time series data includes physiological index sequences, laboratory index sequences, and treatment intervention sequences; Furthermore, in S1, acquiring multi-source time-series data of sepsis patients within a preset time window includes: Set a preset time window length and construct a sliding time window with the current time point as the end time; Physiological index data, laboratory test data, and treatment intervention data are collected within a sliding time window, and corresponding timestamp information is recorded for each data point. Based on timestamp information, multi-source time series data with different sampling frequencies are reconstructed using a unified time axis, mapping various types of data to a unified time scale. Time alignment processing is performed on the mapped multi-source time series data to make various types of data correspond to each other at the same time scale; Missing time nodes in multi-source time series data are imputed; The processed multi-source time series data are numerically normalized, and a multi-dimensional time series input matrix is ​​constructed.

[0021] Specifically, the acquisition and construction of multi-source time series data first revolves around the time window mechanism. During system operation, the current prediction time is used as the time reference point, and a preset time interval is traced back to form a sliding time window. The time window length can be set according to the clinical monitoring frequency; for example, if major physiological indicators are collected at the minute level, the time window can be set at the hour level. The time window length is denoted as... Its unit can be hours or minutes; the current time is recorded as The time window coverage range is The system updates the current time point and rebuilds the window before each risk prediction, thereby achieving dynamic scrolling processing in the time dimension. After the window is constructed, the system collects physiological indicator data, laboratory test data, and treatment intervention data within the corresponding time interval from the clinical information system, monitoring equipment interface, and laboratory data interface. Physiological indicator data includes continuous monitoring data such as blood pressure, heart rate, respiratory rate, and blood oxygen saturation; laboratory test data includes discrete test data such as lactate, white blood cell count, and creatinine; treatment intervention data includes vasopressor dosage, fluid infusion rate, and ventilator support parameters. Each data record is accompanied by a collection timestamp, denoted as [timestamp name missing]. ,in Indicates the first Data records; different data sources may have different sampling frequencies, for example, physiological indicators are sampled at high frequency, while laboratory tests are sampled at low frequency; Because the sampling frequencies of different data sources are inconsistent, a unified timeline reconstruction is required. In implementation, the system first establishes a unified timescale sequence, denoted as... ,in , This indicates the number of uniform time scales within the window; the uniform time scale interval can be a fixed time interval, such as 5 minutes or 10 minutes; subsequently, the original data records are sorted according to timestamps. Mapped to the most recent unified time scale The mapping function can be expressed as: ; in This represents the absolute value of the time difference; through this mapping process, various types of data are reconstructed into the same time coordinate system, thereby eliminating the time misalignment caused by differences in sampling frequency; After completing the timeline mapping, the system further performs time alignment processing; for each unified time scale... Construct corresponding multi-source data vectors; if a certain data source is on the time scale If no observation exists, the location is marked as missing; through this process, various types of data are made to correspond to each other at the same time scale, thus forming a time series data matrix with a consistent structure. For missing time point data, the system performs missing value imputation. The imputation method can be selected according to the data type. For example, for continuous monitoring data, linear interpolation can be used for imputation. The linear interpolation formula is: ; in Indicates on the time scale The padding value at the location, and These represent the nearest non-missing time scales before and after the missing node, respectively. and For corresponding observation values; for cases with long intervals or where interpolation is not possible, a forward filling method can be used, that is, filling with the most recent valid observation value; for low-frequency data such as laboratory tests, a constant value can be maintained between two test times and mapped to each unified time scale; After handling missing values, the system performs numerical normalization on all numerical features. Normalization can be achieved using standardization methods, converting the original values ​​into a zero-mean, unit-variance form. The standardization formula is: ; in These are the original eigenvalues. This is the mean of the feature in the training data. The standard deviation of this feature in the training data. These are the standardized feature values; during the model deployment phase, the mean and standard deviation use the statistical parameters saved during the training phase. After time window construction, unified time axis reconstruction, time alignment, missing value imputation, and normalization, the system arranges the multi-source feature vectors corresponding to each time scale in chronological order to construct a multi-dimensional time series input matrix; this matrix can be represented as: ; in Indicates on the time scale The multidimensional feature vector at the location contains normalized values ​​of physiological indicators, laboratory indicators, and treatment intervention indicators; this matrix serves as the basic input structure for subsequent intervention background construction, dynamic risk reference generation, and deep learning time series prediction. Through the above processing steps, multi-source time series data are uniformly organized into a dynamic input data structure with consistent structure, time alignment, and numerical normalization, providing a complete data foundation for subsequent model operation.

[0022] S2. Construct intervention background features based on treatment intervention sequences, characterize the intervention intensity, duration and trend of intervention at the current time point, and generate corresponding dynamic risk reference benchmarks based on intervention background features; Furthermore, in S2, the construction of intervention background features based on the treatment intervention sequence includes: Intensity extraction is performed on the treatment intervention sequence to obtain the current dose value or support intensity value of various intervention measures within the time window; The duration of treatment intervention sequences was statistically analyzed to calculate the duration and cumulative exposure of various intervention measures within the time window. We performed trend analysis on the treatment intervention sequence to extract the rate of change and fluctuation characteristics of the intervention intensity within the time window; The intensity features, duration features, and trend features are concatenated to construct the intervention background feature vector.

[0023] Furthermore, in S2, the generation of corresponding dynamic risk reference benchmarks based on intervention background characteristics includes: The intervention background feature vector is input into the reference benchmark generation network, which is a multi-layer neural network structure. The reference benchmark generation network performs feature mapping processing on the intervention background feature vector and outputs reference benchmark values ​​corresponding to each key physiological indicator in the physiological indicator sequence. According to the categories of physiological indicators, corresponding dynamic risk reference sequences are constructed so that the dynamic risk reference sequences are aligned with the physiological indicator sequences in the time dimension.

[0024] Specifically, the construction of intervention background features is carried out using the treatment intervention sequence within a time window as input; for each type of intervention, at a unified time scale... Extract the corresponding intervention intensity value, denoted as ,in Indicates the first Intervention-like channels, , This represents the number of time scales within the time window; the intervention intensity feature corresponding to the current time point can be directly taken. This is used to characterize the current level of intervention; Duration characteristics are obtained by measuring the time span during which the statistical intervention is in an effective state; when If the intervention is considered effective, then the duration of the intervention is determined. It can be represented as: ; in This is an indicator function; it takes a value of 1 if the condition is true, and 0 otherwise. To standardize time intervals; cumulative exposure levels It can be represented as: ; in Characterizes the cumulative level of intervention intensity within a time window; The trend of change is obtained by the difference between adjacent time scales; intervention rate characteristics It can be represented as: ; Used to characterize the direction and magnitude of changes in the current intervention intensity; fluctuation characteristics The variance of the intervention intensity within the time window can be used to represent the intensity. ; in The mean intervention intensity within the time window; The current intensity of various intervention channels Duration Cumulative exposure rate of change and fluctuation characteristics Intervention background feature vector is formed by concatenating the feature dimensions. This vector serves as the input for the reference benchmark generation stage; The dynamic risk benchmark is generated through a multi-layer neural network; let the parameters of the benchmark generation network be the weight matrix. With bias vector The input is the intervention background feature vector. Then the output vector It can be represented as: ; in It is a non-linear activation function. , The number of key physiological indicators. For the first The network structure may include multiple fully connected layers, with non-linear activation functions used between layers to achieve feature mapping; the network parameters are obtained through training with historical data, and the mapping relationship between the intervention background feature vector and the observed values ​​of physiological indicators at the corresponding time scale is used as the learning basis during the training process. To construct a dynamic risk reference series, each time scale is analyzed within a time window. Repeat the intervention background feature calculation and reference baseline generation process to obtain the sequence: ; in Indicates time scale The reference vectors for each key physiological indicator are defined; the reference sequence shares the same time index set with the physiological indicator sequence in the time dimension, thus forming a one-to-one correspondence and providing basic input for subsequent reference calibration processing. Through the above process, the intensity, duration, and trend of intervention are systematically encoded into intervention background feature vectors, and mapped into dynamic risk reference benchmarks that change over time through neural networks, thereby realizing a structured association between intervention conditions and physiological indicator reference spaces.

[0025] S3. The physiological indicator sequence is reference-calibrated with the dynamic risk reference benchmark to obtain calibrated physiological state characteristics that reflect the degree of deviation of the physiological indicators from the intervention background. Furthermore, in S3, the reference calibration process of comparing the physiological indicator sequence with the dynamic risk reference benchmark includes: Align the physiological indicator sequence with the corresponding dynamic risk reference baseline sequence in the time dimension, so that each key physiological indicator can establish a correspondence with the corresponding reference baseline value at the same time scale. For each key physiological indicator, the difference characteristics between the physiological indicator value and the corresponding reference value are calculated at each time point; The difference features are normalized by amplitude or proportionally mapped to obtain the calibration deviation features. The calibration deviation features corresponding to each key physiological indicator are combined in chronological order to construct a calibration physiological state feature sequence.

[0026] Specifically, the calibration process is based on a unified time scale; after the aforementioned time window and unified time axis are constructed, the physiological indicator sequence and the dynamic risk reference baseline sequence share the same time scale set. ,in Let the first Key physiological indicators on the time scale The observed value at point is denoted as The dynamic risk reference value at the corresponding time scale is denoted as By matching time indexes, a one-to-one correspondence is established, thereby ensuring that each observation has a unique reference benchmark. After aligning the time dimension, the differential characteristics are calculated for each key physiological indicator at each time point; the differential characteristics can be expressed in the form of differences as follows: ; in Indicates the first A physiological indicator on a time scale The deviation from the dynamic risk reference benchmark; this deviation characterizes the degree of change in the current physiological state relative to the expected state under intervention conditions; To avoid the impact of differences in the units of measurement of different indicators on subsequent modeling, the differences are normalized or scaled. In one implementation, a standardized scaled mapping method can be used, which divides the differences by the standard deviation of the indicators obtained during the training phase. ,get: ; in For the first The standard deviation of each physiological indicator in the training data This represents the normalized calibration deviation feature; alternatively, a proportional mapping approach can be used. ; in To avoid tiny constants with a denominator of zero, It indicates the proportional relationship between physiological indicators and reference standards; two mapping methods can be selected according to the type of indicator. After normalization or scaling, the calibration deviation features of each key physiological indicator obtained at the same time scale are concatenated according to the indicator dimension to form a time-scale feature vector. Then, the feature vectors of all time scales are arranged in chronological order to construct a calibrated physiological state feature sequence: ; in To calibrate the physiological state characteristic sequence, serving as the input basis for subsequent construction of compensatory maintenance characteristics and risk accumulation analysis; Through the above time alignment, difference calculation and normalization process, the original observation values ​​of physiological indicators are converted into a deviation representation relative to the dynamic reference benchmark of the intervention background, so that the subsequent modeling process can run under the same scale and reference conditions, thereby forming a structurally consistent calibration physiological state feature sequence.

[0027] S4. Based on the response relationship between the treatment intervention sequence and the physiological indicator sequence, construct the compensatory maintenance characteristics; Furthermore, in S4, the construction of the compensatory maintenance feature includes: Temporal difference processing is performed on the treatment intervention sequence to extract the variation characteristics of intervention intensity between adjacent time nodes; Temporal difference processing is performed on the physiological indicator sequence to extract the variation characteristics of the physiological indicators between adjacent time nodes; Based on the temporal correspondence between intervention change characteristics and physiological indicator change characteristics, a response relationship feature between intervention change and physiological indicator change is constructed. Extract physiological fluctuation features that reflect autonomic regulatory activities, including heart rate change features, respiratory rate change features, or blood pressure fluctuation features; The response relationship features and physiological fluctuation features are concatenated to construct a compensatory maintenance feature sequence, and its time dimension is kept consistent with the calibration physiological state feature sequence.

[0028] Specifically, the construction of compensatory maintenance characteristics involves placing the treatment intervention sequence and the physiological indicator sequence on a unified time scale. The corresponding relationships are based on this; both the treatment intervention sequence and the physiological indicator sequence have already undergone time axis reconstruction and alignment in the aforementioned steps, therefore, at each time scale... Intervention intensity values ​​and physiological index values ​​can be obtained from the above; let the first... Intervention-like channels on the time scale The strength is , No. Key physiological indicators on the time scale The observed value is ,in , To intervene in the number of channels, , The number of key physiological indicators; during model operation, the system uses continuous time scales within the time window as the sequence unit, calculates the compensatory maintenance features corresponding to each time scale in sequence, and forms a serialized input; In the intervention change feature extraction stage, time difference processing is performed on the intervention intensity sequence; the time difference is implemented using an adjacent time scale difference method, defining the first... Intervention-like channels on the time scale The amount of change due to intervention is: ; in , The number of time scales within the time window; the above difference components are used to characterize the adjustment magnitude and direction of intervention intensity between adjacent time nodes; in the physiological indicator change feature extraction stage, the same adjacent difference method is used for the physiological indicator sequence, defining the first... A physiological indicator on a time scale The change is: ; in Used to characterize the changes in physiological indicators between adjacent time points; the above difference processing belongs to the discrete time difference algorithm, which is implemented by subtracting adjacent elements in the time index in the time-aligned array or matrix to obtain the change sequence; In the response relationship feature construction stage, response representations are extracted based on the temporal correspondence between intervention change features and physiological indicator change features; in one implementation, the intervention change vector at each time scale can be... With physiological change vector Combine them to form a response relationship feature vector. The response relationship feature vector can be constructed using either element-wise multiplication or concatenation. The element-wise multiplication method can be represented as: ; in The matrix representing the outer product of intervention-induced changes and physiological changes. This represents a vectorization operator that expands a matrix into vectors in a preset order; this outer product construction is used to explicitly express the coupling relationship between interventional changes and physiological changes at the same time scale; in another implementation, a response hysteresis window may be optionally introduced. and Pairing is performed to form hysteresis response characteristics, where For the lagging step size, It can be selected as 1 or several time scales; In the stage of extracting physiological fluctuation characteristics of autonomic regulation, fluctuation representations are constructed based on indicators that can reflect the strength of autonomic regulation, such as heart rate, respiratory rate, or blood pressure; assuming the heart rate sequence is... The respiratory rate sequence is The mean arterial pressure sequence is In one implementation, the adjacent differences of the above indicators can be calculated as change characteristics, for example... In another implementation, the local fluctuation amplitude can be calculated within a time window, for example, within a time window of length [length missing]. The mean square deviation is calculated within a local sub-window as the variability, and the blood pressure variability is defined. for: ; in Indicated by The local mean at the termination point. The length of the local sub-window; the heart rate variation features, respiratory rate variation features, and blood pressure fluctuation features obtained through the above method together constitute the physiological fluctuation feature vector of autonomic regulation activities. ; In the compensatory maintenance feature sequence construction stage, the response relationship feature vector at each time scale is... With physiological fluctuation feature vector The features are concatenated along the feature dimension to form a compensatory maintenance feature vector. The compensatory maintenance characteristic sequence is then obtained by arranging the sequences in chronological order. ; The sequence begins at The index range is calculated using the matching difference, and can be adjusted as needed. Features are filled with zeros or copied. Features are processed to maintain consistent length; compensatory maintenance keeps the feature sequence and the calibration physiological state feature sequence consistent in time index, so that the input can be aligned according to the time scale in the subsequent feature fusion stage.

[0029] S5. Perform mild abnormality identification and time-dimensional cumulative analysis on the calibrated physiological state characteristics to construct risk accumulation characteristics and risk change trend characteristics; Furthermore, in S5, the identification of mild abnormalities and the time-dimensional cumulative analysis of the calibrated physiological state characteristics include: For each key physiological indicator in the calibration physiological state characteristics, a threshold for mild abnormality intervals is preset; Within the time window, determine whether the calibration deviation of each key physiological indicator is within the corresponding mild abnormality range at each time node, and mark the mild abnormality state. The duration and frequency of mild abnormal states are statistically analyzed over time to construct a cumulative feature of mild abnormalities for a single indicator. The mild abnormality accumulation characteristics of multiple key physiological indicators are combined and processed to construct multi-indicator superimposed accumulation characteristics; Based on the cumulative characteristics of mild anomalies in single indicators and the cumulative characteristics of multiple indicators, the cumulative change trend characteristics are extracted in the time dimension to form a risk accumulation feature sequence.

[0030] Specifically, the identification of mild anomalies and the cumulative analysis over time are carried out using calibrated physiological state feature sequences as input; the calibrated physiological state feature sequences are analyzed on a unified time scale. The following is represented as ,in For time scale The multidimensional feature vector at point 1; let the first... Key physiological indicators on the time scale The calibration deviation characteristics are ,in , The system constructs risk accumulation features in the order of "threshold for mild abnormality interval - state marking - time accumulation - trend extraction" during operation, and outputs the results in sequence to the subsequent feature fusion stage. The threshold for mild abnormality intervals can be determined jointly by training data statistics and clinical rules; in one implementation, a lower threshold is configured for each key physiological indicator. With upper threshold To define the mildly abnormal range of this indicator. The threshold can be selected from quantile statistics during the training phase, manually configured safety boundaries, or parameters from previous clinical pathways in hospitals, and is fixed into a configuration file or parameter table when the model is deployed. To avoid differences in scales of different indicators, the threshold for the mild abnormality interval is kept in the same representation space as the calibration deviation feature. For example, when the calibration deviation feature is a standardized deviation value, the threshold is also configured using a standardized scale. Perform minor anomaly status checks at each time point within the time window; for each time scale... With each key physiological indicator Mildly anomalous marker variables are generated based on the relationship between calibration deviation characteristics and threshold ranges. The marking rule can be expressed as: ; in This is an indicator function; it takes a value of 1 if the condition is true, and a value of 0 otherwise. Indicates the first Each indicator on the time scale Is it in a slightly abnormal state? When using proportional mapping deviation features, the threshold interval can be changed to a proportional threshold interval and a labeled variable can be generated according to the same judgment rule. In the stage of constructing the cumulative features of single-indicator mild anomalies, mild anomaly markers are aggregated in the time dimension; duration statistics are performed using a unified time scale interval. The duration of a mild abnormality in a single indicator is measured in time units. It can be represented as: ; Frequency statistics can be performed by counting the number of transitions from non-abnormal to abnormal, thus defining the frequency. for: ; Among the symbols The logical AND is represented; in an alternative approach, a continuous segment length statistic can be introduced to characterize the segment distribution of mild anomalies. The continuous segment statistic is calculated based on the maximum continuous length of adjacent 1 values ​​in the labeled sequence; the above duration and frequency of occurrence together constitute the cumulative feature of mild anomalies in a single index. In the stage of constructing multi-indicator cumulative features, the mild abnormality cumulative features of multiple key physiological indicators are combined. The combination method can be to sum the abnormality markers at the same time scale to obtain the multi-indicator superimposed markers. Its expression is: ; in Indicates time scale The number of indicators simultaneously in a mildly abnormal state; furthermore, it can be used to... Aggregation is performed along the time dimension to obtain the superimposed duration. and superposition frequency The duration of the superimposed layer can be achieved using... Formal calculation, where The threshold value represents the sum of all possible values. When all indicators are simultaneously at a slightly abnormal level, the total value is accumulated. The value can be selected as an integer of 2 or greater and used as a parameter configuration; the superposition frequency can be determined using the same transfer counting method. The sequence was obtained through statistical analysis; In the risk change trend feature extraction stage, the cumulative process is transformed into a serialized feature that can be updated over time. In one implementation, a prefix accumulation method is used to construct an accumulated sequence that increases with the time scale, defining a single-indicator prefix accumulation. for: ; And define multi-indicator overlay prefix accumulation for: ; in and A recursive sequence is formed on a time scale; based on this, the cumulative trend characteristics can be obtained through adjacent differences, and the trend characteristics are defined. for And also construct a system for the cumulative sum of multiple indicators. The cumulative prefix values ​​of each indicator are concatenated with the trend characteristics along the dimensions to obtain a time-scale risk accumulation feature vector. ; Arrange the risk accumulation feature vectors corresponding to each time scale in chronological order to form a risk accumulation feature sequence. This sequence shares the same set of time indices with the calibration physiological state feature sequence, thus allowing the input to be aligned on the time scale in the subsequent feature fusion stage.

[0031] S6. The calibration physiological state characteristics, compensatory maintenance characteristics and risk accumulation characteristics are fused and input into the deep learning time series prediction network to output the sepsis deterioration risk value and risk change trend in the preset future time period. Furthermore, in S6, the fusion of calibration physiological state characteristics, compensatory maintenance characteristics, and risk accumulation characteristics includes: The features of calibrated physiological state, compensatory maintenance and risk accumulation are respectively encoded, and the features from different sources are mapped to a unified feature representation space. The encoded features are aligned according to the time dimension to construct a multi-channel time series feature tensor; Channel-weighted processing is performed on the multi-channel time series feature tensor to generate fusion weights; The features of each channel are weighted and aggregated according to the fusion weight to form a unified fusion feature sequence; The fused feature sequence is used as input to a deep learning temporal prediction network.

[0032] Furthermore, in S6, the deep learning temporal prediction network includes: A deep neural network structure containing a time modeling layer is constructed, wherein the time modeling layer is used to model the temporal dependencies of the fused feature sequence; The time modeling layer includes at least one of a recurrent neural network structure, a gated recurrent unit structure, or a self-attention structure. A feature mapping layer is set after the time modeling layer to perform non-linear mapping processing on the sequence features after time modeling; Set up an output layer to perform regression or classification processing on the mapped features and output the risk value of sepsis exacerbation and the trend of risk change within a preset future time period. A deep learning time series prediction network was trained using historical labeled data and then used for risk prediction after training.

[0033] Specifically, the feature fusion and deep learning temporal prediction network operates by taking the calibration physiological state feature sequence, the compensation maintenance feature sequence, and the risk accumulation feature sequence within the same time window as input; the calibration physiological state feature sequence is denoted as... The compensatory maintenance characteristic sequence is denoted as The risk accumulation characteristic sequence is denoted as ;in To standardize the time scale, This represents the number of time scale divisions within the time window. These are feature vectors from different channels at the same time scale. The system reads the three types of features in time scale order during each prediction and enters the encoding and fusion process. After obtaining a unified input, it is sent to the time series prediction network to complete the risk output. Feature encoding is used to map features from different sources to a feature representation space of uniform dimension. In one implementation, an independent encoder network is configured for each type of feature, and the encoder network can be implemented using a multi-layer fully connected structure. Let the encoder for calibrating physiological states be... The compensatory encoder is The risk accumulation encoder is Then on the time scale The encoding results at each position are represented as follows: The encoder output dimension is uniformly set to... ,in The preset feature dimensions are used; the encoder network parameters are updated during the training phase through error backpropagation. After encoding, the various encoded features are aligned along the time dimension and a multi-channel time series feature tensor is constructed. In one possible representation, the three types of encoded sequences are stacked along the channel dimension to form a tensor. The first dimension corresponds to the time scale, the second dimension corresponds to the channel category, and the third dimension corresponds to the unified feature dimension; this tensor guarantees that at any time scale... There exists a set of parallel channel coding vectors. This provides a calculation object for subsequent channel weighting; Channel weighting is used to generate fusion weights and perform weighted aggregation; in one implementation, a channel weight generation network is set up. Its input is a time scale. The channel encoded vector concatenation result The output is a channel weight vector. To ensure the weights have normalizable properties, the weight vector can be obtained using the Softmax function, expressed as: ; in For the weight generation network in the channel The output value on , To correspond to the normalized weights; after obtaining the weights, the channel encoding vectors are weighted and aggregated to form a fused feature vector. Its expression is: ; Arrange the fused feature vectors at each time scale in chronological order to obtain the fused feature sequence. This fused feature sequence serves as the input to a deep learning temporal prediction network; Deep learning time series prediction networks consist of a time modeling layer, a feature mapping layer, and an output layer. The time modeling layer is used to model the temporal dependencies of the fused feature sequences, and can be selected from at least one of the following structures: recurrent neural network, gated recurrent unit, or self-attention structure. Taking a gated recurrent unit as an example, the time modeling layer is defined on a time scale. Receive input And output the hidden state. ,in The sequence representation is modeled over time; the feature mapping layer is used to perform non-linear mapping on the hidden states to obtain the mapped features. The mapping process can be implemented using fully connected layers and activation functions; the output layer is used to convert the mapped features into risk outputs, which may include the risk value of sepsis worsening and the risk change trend within a preset future time period; in one implementation, the risk value output is denoted as... The trend output is denoted as The output layer is implemented using either a regression head or a classification head, where the regression head outputs a continuous risk score and the classification head outputs the risk category probability. Network training utilizes historical labeled data, which includes time-series samples corresponding to the fused input and outcome labels corresponding to those samples. Outcome labels can be selected as tags indicating whether events such as shock, organ function deterioration, or death will occur within a future time period. During training, the parameters of the encoder network, weight generation network, and time-series prediction network are updated by minimizing the loss function between the predicted output and the labels. The loss function can be a weighted combination of hazard value prediction loss and trend prediction loss. Hazard value prediction loss can use cross-entropy loss or mean squared error loss, while trend prediction loss can use mean squared error loss. After training, the network parameters are solidified and used in the inference phase to encode, fuse, and predict the time series inputs of new patients from multiple sources, outputting the hazard value and risk change trend for the corresponding future time period.

[0034] S7. Based on the risk value, compensation maintenance characteristics, and risk change trends, generate risk level determination results and early warning information.

[0035] Specifically, risk level determination and early warning generation are performed after the deep learning time series prediction network outputs its data; the deep learning time series prediction network outputs the risk value and risk change trend for a preset future time period at the current prediction time; the risk value is denoted as... ,in For the current predicted time, For continuous risk scoring or risk probability; the trend of risk change is denoted as... , It can be used to score the trend of risk changes over time or to provide a sequential description of risk at multiple future points in time; the compensation maintenance feature comes from the aforementioned compensation maintenance feature sequence, and the compensation maintenance feature vector corresponding to the current prediction time is denoted as... ,in It can be constructed by splicing together response relationship features and physiological fluctuation features; the system reads during runtime. , and And then proceed to the risk level assessment process; Risk level determination is achieved using a threshold segmentation method; in one implementation, a risk level set is pre-defined. ,in The risk level can be selected as low risk, medium risk, high risk, or more levels; a risk threshold sequence can also be configured. ,satisfy ;when When performing continuous risk scoring, the risk level determination function can be expressed as: ; in This represents the current risk level assessment result; when When dealing with risk category probability vectors, one can choose to use the category corresponding to the highest probability as the level output, or combine the probability with a threshold to obtain the level; threshold It can be determined based on validation set statistics during the training phase and loaded as a configuration parameter during the deployment phase; When risk change trends are used in the judgment, the system... Extract the trend determination quantity; in one implementation, the trend determination quantity is denoted as... and configure trend thresholds ;when When the trend marker variable s(t0) is set to 1, it is set to 0 otherwise. The expression is as follows: When risk trends are output as risk sequences at multiple future points in time, the maximum value, slope, or aggregated value of adjacent differences in the future risk sequences can be selected as the output. Trend threshold Configure it as part of the parameter table; When the compensatory maintenance feature is involved in the determination, the system... Provide a compensation determination quantity; in one implementation, set a compensation determination function. Map the compensatory maintenance feature vector to a scalar compensatory score ;function The compensation threshold can be either a linear transformation plus a threshold comparison or a trainable mapping implemented by a small multilayer perceptron; the compensation threshold is denoted as... And define the compensation marker variable. Compensation threshold It can be statistically derived from the training phase or set by clinical rules; when a trainable mapping is used, the training samples can come from historical data on the combination of intervention changes and physiological fluctuations related to compensation, and the mapping parameters are updated through supervised learning and solidified at deployment. The early warning information is generated based on the risk level assessment and combined with trend marker variables and compensation marker variables to execute rule-based triggering; in one implementation, the early warning triggering function is denoted as... ,when An early warning status is output when a preset level is reached or when a combination of trend and compensation conditions is met; the early warning trigger logic can be expressed as: ; in The set of levels that trigger an alert, symbol Indicates logical OR; when The system generates early warning information and outputs it to the display terminal, nursing station system, or message push interface. The early warning information may include status fields such as the current risk level, risk value, trend marker, and compensation marker, and is recorded in the log along with the patient identifier and timestamp. The system repeats the above judgment and triggering process at subsequent time scales, thereby realizing continuous risk level updates and early warning output.

[0036] Example 2: In intensive care unit settings, some sepsis patients may have their blood pressure and other physiological indicators maintained within a relatively stable range through external intervention during continuous vasopressor support and fluid resuscitation. However, their actual tissue perfusion status and inflammatory response level continue to deteriorate. Existing risk prediction models are typically based on raw physiological indicators or simply superimposed intervention variables, making it difficult to distinguish between the natural stable state of physiological indicators and the intervention-maintained state. Consequently, under intervention masking conditions, it is difficult to accurately characterize the evolution of the patient's true pathological risk. To address these issues, this invention provides a deep learning-based dynamic risk prediction device for sepsis, the structure of which is as follows: Figure 2 As shown. The specific implementation process of this device is as follows: The data acquisition module is used to acquire multi-source time series data of sepsis patients within a preset time window. The multi-source time series data includes physiological index sequences, laboratory index sequences, and treatment intervention sequences. The intervention background construction module is used to construct intervention background features based on the treatment intervention sequence, and to characterize the intervention intensity, intervention duration and intervention change trend at the current time point. At the same time, it generates a corresponding dynamic risk reference benchmark based on the intervention background features. The reference calibration module is used to perform reference calibration processing on the physiological indicator sequence and the dynamic risk reference benchmark to obtain calibrated physiological state characteristics. A compensatory maintenance feature construction module is used to construct compensatory maintenance features based on the response relationship between the treatment intervention sequence and the physiological indicator sequence. The risk accumulation analysis module is used to identify mild abnormalities and perform time-dimensional accumulation analysis on the calibrated physiological state characteristics, and to construct risk accumulation characteristics and risk change trend characteristics. The feature fusion and prediction module is used to fuse the calibration physiological state features, the compensatory maintenance features and the risk accumulation features, and input them into a deep learning time series prediction network to output the sepsis exacerbation risk value and risk change trend within a preset future time period. The risk assessment module is used to generate risk level assessment results and early warning information based on the risk value, the compensation maintenance characteristics, and the risk change trend.

[0037] Specifically, the device is deployed in the information system environment of the intensive care unit and can optionally establish data interface connections with bedside monitoring equipment, laboratory information systems and medical order execution systems. The device constructs a sliding time window based on the current predicted time, and the data acquisition module collects physiological index data, laboratory test data and treatment intervention data within the time window, forming a multi-dimensional time series input structure under a unified time scale. After receiving the treatment intervention sequence, the intervention background construction module extracts the current intervention intensity, duration within the time window, and cumulative exposure, and calculates the intervention change trend characteristics to construct the intervention background feature vector. The intervention background feature vector is input into the reference benchmark generation network to generate a dynamic risk reference benchmark sequence corresponding to each key physiological indicator, and is aligned with the physiological indicator sequence in the time dimension. The reference calibration module performs difference mapping processing on the physiological indicator sequence and the dynamic risk reference baseline according to the time alignment relationship, forming a calibrated physiological state characteristic sequence. This characteristic sequence reflects the degree of deviation of the physiological state under the constraints of intervention conditions and serves as the input for subsequent analysis modules. The compensatory maintenance feature construction module performs time difference processing on the treatment intervention sequence and the physiological index sequence to construct the response relationship features between intervention changes and physiological changes, and combines them with the autoregulatory fluctuation features to form the compensatory maintenance feature sequence. The risk accumulation analysis module identifies mild abnormalities in the calibrated physiological state characteristics and performs duration statistics and multi-indicator overlay analysis in the time dimension to construct a sequence of risk accumulation characteristics and risk change trend characteristics. The feature fusion and prediction module encodes the calibration physiological state features, compensatory maintenance features, and risk accumulation features respectively, and performs channel-weighted fusion in a unified dimensional space to form a fused feature sequence. The fused feature sequence is input into a deep learning temporal prediction network. The network extracts the time dependency through the time modeling layer, and obtains the sepsis deterioration risk value and risk change trend within a preset time period through the feature mapping layer and the output layer. The risk assessment module classifies risks based on risk values ​​and trends, and executes early warning triggering rules in conjunction with the compensatory maintenance characteristics at the current time scale to generate risk level assessment results and early warning information. The early warning information can be displayed on the monitoring terminal interface or pushed to the device to provide continuous and dynamic alerts on the risk status of sepsis patients.

[0038] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based dynamic risk prediction model for sepsis, characterized in that, Includes the following steps: S1. Acquire multi-source time series data of sepsis patients within a preset time window, wherein the multi-source time series data includes physiological index sequences, laboratory index sequences, and treatment intervention sequences; S2. Construct intervention background features based on treatment intervention sequences, characterize the intervention intensity, duration and trend of intervention at the current time point, and generate corresponding dynamic risk reference benchmarks based on intervention background features; S3. The physiological indicator sequence is reference-calibrated with the dynamic risk reference benchmark to obtain calibrated physiological state characteristics that reflect the degree of deviation of the physiological indicators from the intervention background. S4. Based on the response relationship between the treatment intervention sequence and the physiological indicator sequence, construct the compensatory maintenance characteristics; S5. Perform mild abnormality identification and time-dimensional cumulative analysis on the calibrated physiological state characteristics to construct risk accumulation characteristics and risk change trend characteristics; S6. The calibration physiological state characteristics, compensatory maintenance characteristics and risk accumulation characteristics are fused and input into the deep learning time series prediction network to output the sepsis deterioration risk value and risk change trend in the preset future time period. S7. Based on the risk value, compensation maintenance characteristics, and risk change trends, generate risk level determination results and early warning information.

2. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S1, acquiring multi-source time-series data of sepsis patients within a preset time window includes: Set a preset time window length and construct a sliding time window with the current time point as the end time; Physiological index data, laboratory test data, and treatment intervention data are collected within a sliding time window, and corresponding timestamp information is recorded for each data point. Based on timestamp information, multi-source time series data with different sampling frequencies are reconstructed using a unified time axis, mapping various types of data to a unified time scale. Time alignment processing is performed on the mapped multi-source time series data to make various types of data correspond to each other at the same time scale; Missing time nodes in multi-source time series data are imputed; The processed multi-source time series data are numerically normalized, and a multi-dimensional time series input matrix is ​​constructed.

3. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S2, the construction of intervention background features based on the treatment intervention sequence includes: Intensity extraction is performed on the treatment intervention sequence to obtain the current dose value or support intensity value of various intervention measures within the time window; The duration of treatment intervention sequences was statistically analyzed to calculate the duration and cumulative exposure of various intervention measures within the time window. We performed trend analysis on the treatment intervention sequence to extract the rate of change and fluctuation characteristics of the intervention intensity within the time window; The intensity features, duration features, and trend features are concatenated to construct the intervention background feature vector.

4. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S2, the generation of the corresponding dynamic risk reference benchmark based on the intervention background characteristics includes: The intervention background feature vector is input into the reference benchmark generation network, which is a multi-layer neural network structure. The reference benchmark generation network performs feature mapping processing on the intervention background feature vector and outputs reference benchmark values ​​corresponding to each key physiological indicator in the physiological indicator sequence. According to the categories of physiological indicators, corresponding dynamic risk reference sequences are constructed so that the dynamic risk reference sequences are aligned with the physiological indicator sequences in the time dimension.

5. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S3, the reference calibration process of the physiological indicator sequence with the dynamic risk reference benchmark includes: Align the physiological indicator sequence with the corresponding dynamic risk reference baseline sequence in the time dimension, so that each key physiological indicator can establish a correspondence with the corresponding reference baseline value at the same time scale. For each key physiological indicator, the difference characteristics between the physiological indicator value and the corresponding reference value are calculated at each time point; The difference features are normalized by amplitude or proportionally mapped to obtain the calibration deviation features. The calibration deviation features corresponding to each key physiological indicator are combined in chronological order to construct a calibration physiological state feature sequence.

6. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S4, the constructed compensatory maintenance feature includes: Temporal difference processing is performed on the treatment intervention sequence to extract the variation characteristics of intervention intensity between adjacent time nodes; Temporal difference processing is performed on the physiological indicator sequence to extract the variation characteristics of the physiological indicators between adjacent time nodes; Based on the temporal correspondence between intervention change characteristics and physiological indicator change characteristics, a response relationship feature between intervention change and physiological indicator change is constructed. Extract physiological fluctuation features that reflect autonomic regulatory activities, including heart rate change features, respiratory rate change features, or blood pressure fluctuation features; The response relationship features and physiological fluctuation features are concatenated to construct a compensatory maintenance feature sequence, and its time dimension is kept consistent with the calibration physiological state feature sequence.

7. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S5, the identification of mild abnormalities and the time-dimensional cumulative analysis of the calibrated physiological state characteristics include: For each key physiological indicator in the calibration physiological state characteristics, a threshold for mild abnormality intervals is preset; Within the time window, determine whether the calibration deviation of each key physiological indicator is within the corresponding mild abnormality range at each time node, and mark the mild abnormality state. The duration and frequency of mild abnormal states are statistically analyzed over time to construct a cumulative feature of mild abnormalities for a single indicator. The mild abnormality accumulation characteristics of multiple key physiological indicators are combined and processed to construct multi-indicator superimposed accumulation characteristics; Based on the cumulative characteristics of mild anomalies in single indicators and the cumulative characteristics of multiple indicators, the cumulative change trend characteristics are extracted in the time dimension to form a risk accumulation feature sequence.

8. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S6, the fusion of calibration physiological state characteristics, compensatory maintenance characteristics, and risk accumulation characteristics includes: The features of calibrated physiological state, compensatory maintenance and risk accumulation are respectively encoded, and the features from different sources are mapped to a unified feature representation space. The encoded features are aligned according to the time dimension to construct a multi-channel time series feature tensor; Channel-weighted processing is performed on the multi-channel time series feature tensor to generate fusion weights; The features of each channel are weighted and aggregated according to the fusion weight to form a unified fusion feature sequence; The fused feature sequence is used as input to a deep learning temporal prediction network.

9. The sepsis dynamic risk prediction model based on deep learning according to claim 1, characterized in that, In S6, the deep learning temporal prediction network includes: A deep neural network structure containing a time modeling layer is constructed, wherein the time modeling layer is used to model the temporal dependencies of the fused feature sequence; The time modeling layer includes at least one of a recurrent neural network structure, a gated recurrent unit structure, or a self-attention structure. A feature mapping layer is set after the time modeling layer to perform non-linear mapping processing on the sequence features after time modeling; Set up an output layer to perform regression or classification processing on the mapped features and output the risk value of sepsis exacerbation and the trend of risk change within a preset future time period. A deep learning time series prediction network was trained using historical labeled data and then used for risk prediction after training.

10. A deep learning-based dynamic risk prediction device for sepsis, characterized in that, The apparatus for the deep learning-based sepsis dynamic risk prediction model according to any one of claims 1-9, the apparatus comprising: The data acquisition module is used to acquire multi-source time series data of sepsis patients within a preset time window. The multi-source time series data includes physiological index sequences, laboratory index sequences, and treatment intervention sequences. The intervention background construction module is used to construct intervention background features based on the treatment intervention sequence, and to characterize the intervention intensity, intervention duration and intervention change trend at the current time point. At the same time, it generates a corresponding dynamic risk reference benchmark based on the intervention background features. The reference calibration module is used to perform reference calibration processing on the physiological indicator sequence and the dynamic risk reference benchmark to obtain calibrated physiological state characteristics. A compensatory maintenance feature construction module is used to construct compensatory maintenance features based on the response relationship between the treatment intervention sequence and the physiological indicator sequence. The risk accumulation analysis module is used to identify mild abnormalities and perform time-dimensional accumulation analysis on the calibrated physiological state characteristics, and to construct risk accumulation characteristics and risk change trend characteristics. The feature fusion and prediction module is used to fuse the calibration physiological state features, the compensatory maintenance features and the risk accumulation features, and input them into a deep learning time series prediction network to output the sepsis exacerbation risk value and risk change trend within a preset future time period. The risk assessment module is used to generate risk level assessment results and early warning information based on the risk value, the compensation maintenance characteristics, and the risk change trend.

Images