Multidimensional physiological parameter prediction methods, storage media and electronic devices

Abstract

This application provides a method, storage medium, and electronic device for predicting multidimensional physiological parameters, belonging to the field of data processing technology. The method includes: acquiring a data stream from a user being tested; generating a corresponding high-dimensional input vector based on the data stream, wherein the high-dimensional input vector includes the most recent observation value at each observation time, data freshness and data missingness indicators, trend feature data corresponding to the most recent observation value, and parameter linkage feature data; inputting the high-dimensional input vector into a pre-constructed asymmetric neural frequent differential interaction network model, and processing it through a cascaded variable selection layer, a continuous-time evolution encoder, a prototype semantic interaction layer, and an asymmetric security-gated decoder in the asymmetric neural frequent differential interaction network model, outputting the final predicted values ​​of preset multidimensional physiological parameters. This application can improve the accuracy of predicting the physiological parameters of the user being tested.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for predicting multidimensional physiological parameters, a storage medium, and an electronic device. Background Technology

[0002] In critical clinical settings such as intensive care unit (ICU), emergency care, and perioperative period, patients' hemodynamic status changes rapidly and complexly. Real-time and accurate prediction of future trends and risk thresholds of core physiological parameters such as mean arterial pressure (MAP) and heart rate (HR) is a key means to achieve early disease intervention and reduce the incidence of adverse events. With the development of medical informatics and artificial intelligence technologies, physiological parameter prediction technology based on monitoring data has gradually become an important direction for clinical decision support. Its core requirement is to reliably predict the minimum values ​​of parameters and risk levels within multiple future time windows using collected discrete physiological sign data.

[0003] Currently, physiological parameter prediction technologies are mainly divided into two categories: traditional statistical models and deep learning models. Traditional statistical models, such as autoregressive integral moving average models and linear regression models, achieve prediction by analyzing the statistical regularities of historical data. However, these models are difficult to handle discrete data with different update frequencies and irregular sampling intervals. They usually need to use interpolation alignment to convert non-uniform data into equal time interval sequences. This process not only introduces spurious trends that do not match the patient's true physiological state, but also loses the data freshness, a key information reflecting the timeliness of the parameters, resulting in a large deviation between the prediction results and clinical reality.

[0004] In recent years, deep learning models such as Long Short-Term Memory (LSTM) networks and Transformer models have been widely used in the field of physiological parameter prediction due to their powerful temporal feature extraction capabilities. However, existing deep learning solutions still face several technical challenges in clinical application: 1. Existing deep learning models generally use mean squared error (MSE) as the loss function, which cannot effectively distinguish between predicted values ​​higher than and lower than the true values, leading to potential underreporting of risks; 2. In actual clinical applications, models need to output prediction results for multiple time windows such as 5 minutes, 10 minutes, and 15 minutes, but existing methods still exhibit a contradiction where the accuracy of predictions for longer time windows is higher than that for shorter time windows.

[0005] In conclusion, current physiological parameter prediction technologies, whether traditional statistical models or deep learning models, still need improvement in terms of prediction safety and reliability. Summary of the Invention

[0006] The purpose of this application is to provide a method, storage medium, and electronic device for predicting multidimensional physiological parameters to solve at least one of the above-mentioned technical problems.

[0007] In its first aspect, this application provides a method for predicting multidimensional physiological parameters, including:

[0008] Obtain the data stream of the user under test {(x i , t i )},t i For the actual arrival timestamp, x i Let represent the observation vector of the i-th dimension;

[0009] Based on the data stream, a corresponding high-dimensional input vector X(t) is generated. The high-dimensional input vector includes the most recent observation value at each observation time, data freshness and data missing indication values, change trend feature data corresponding to the most recent observation value, and parameter linkage feature data.

[0010] The high-dimensional input vector is input into a pre-constructed asymmetric neuromorphic interaction network model. After processing by the cascaded variable selection layer, continuous-time evolution encoder, prototype semantic interaction layer and asymmetric security-gated decoder in the asymmetric neuromorphic interaction network model, the final predicted value of the preset multidimensional physiological parameters is output.

[0011] Optionally, the processing of the variable selection layer includes: dynamically assigning weights and filtering noise on the high-dimensional input vector based on a gated residual network to form weighted features of multidimensional physiological parameters.

[0012] Optionally, the processing of the continuous-time evolution encoder includes: using a Jump-Flow mechanism to adapt to irregular time intervals, instantly updating the hidden state of the network through a time-series update unit when the observed data arrives, and simulating the continuous evolution of the hidden state through a differential equation network during the observation blank period to obtain the final hidden state;

[0013] Optionally, the processing of the prototype semantic interaction layer includes: calculating the similarity distribution between the final hidden state and the preset prototype library through an attention mechanism, generating semantic enhancement features by combining the similarity distribution, and mapping the corresponding risk coefficient.

[0014] Optionally, the processing procedure of the asymmetric security gating decoder includes: performing regression calculation on the semantic enhancement features to obtain a basic predicted value, correcting the basic predicted value based on a negative security bias term controlled by the risk coefficient, and using the predicted minimum value of the multidimensional physiological parameters as the final predicted value.

[0015] Optionally, the step of dynamically assigning weights and filtering noise on the high-dimensional input vector based on a gated residual network to form weighted features of multidimensional physiological parameters includes: performing feature transformation on the high-dimensional input vector, generating gated weights for feature channels adapted to multidimensional physiological parameters in parallel, and suppressing feature channels with severe missing or high noise levels by weighted fusion of the gated weights and the transformed features to achieve noise filtering and form weighted features of multidimensional physiological parameters.

[0016] Optionally, the time-series update unit is a recurrent neural network unit, the differential equation network is a neural ordinary differential equation network, and the continuous evolution result of the hidden state during the observation blank time period is integrally calculated by a numerical solver.

[0017] Optionally, the similarity distribution is obtained by calculating the correlation between the final hidden state and the prototype matrix using the Softmax function; the semantic enhancement feature is generated by superimposing the weighted sum of the prototypes in the prototype matrix using the similarity distribution with the final hidden state.

[0018] Optionally, the negative safety bias term is generated by a preset nonlinear function. The output of the nonlinear function is always positive. After sign inversion, the negative safety bias term is formed, so that the safety correction can only lower the basic prediction value.

[0019] Optionally, the feature transformation is achieved through linear layer mapping to obtain mapped features; the gate weights are generated by calculating the Sigmoid function; the weighted fusion includes element-wise multiplication of the mapped features and the gate weights to generate the weighted features.

[0020] Optionally, the recurrent neural network unit is a gated recurrent GRU unit; the numerical solver is a Runge-Kutta solver or an Euler solver; the step of instantaneously updating the network hidden state through a time-series update unit at the arrival of observed data, and simulating the continuous evolution of the hidden state through a differential equation network during the observation gap period to obtain the final hidden state includes: instantaneously updating the network hidden state through a GRU unit combined with the weighted features at the arrival of observed data; simulating the continuous dynamic evolution of the hidden state through a neural differential equation network during the observation gap period, and calculating the final hidden state through a numerical solver.

[0021] Optionally, the nonlinear function is a Softplus function; the correction process is to superimpose the base prediction value with a safety correction term, wherein the safety correction term is the product of a trainable scaling parameter, a risk coefficient, and a negative safety bias term.

[0022] Optionally, the method further includes: constructing a variable selection layer, a continuous-time evolutionary encoder, a prototype semantic interaction layer, and an asymmetric security-gated decoder in sequence to form an asymmetric neural frequent differential interaction network to be trained;

[0023] The pre-set training data is input into the asymmetric neural frequent differential interaction network to be trained for iterative training. A composite loss function including asymmetric safe regression loss, multi-time window consistency constraint loss, and prototype orthogonal regularization term is used to train and optimize the asymmetric neural frequent differential interaction network to be trained, so as to obtain the trained asymmetric neural frequent differential interaction network model.

[0024] Optionally, the composite loss function is a weighted sum of asymmetric safe regression loss, multi-time-window consistency constraint loss, and prototype orthogonal regularization term.

[0025] Optionally, the multidimensional physiological parameters include at least one of mean arterial pressure, heart rate, blood oxygen saturation, and respiratory rate; the trend characteristic data includes raw observations, a first-order trend, and a second-order trend, wherein the first-order trend is the rate of change of the raw observations calculated through a sliding time window, and the second-order trend is the rate of change of the first-order trend; the parameter linkage characteristic data is the Pearson correlation coefficient or the proportion of changes in the same direction between different physiological parameters within the sliding time window.

[0026] Optionally, after outputting the final predicted value of the preset multidimensional physiological parameters, the method further includes: obtaining an individualized threshold that matches the user being tested; and determining whether to issue an alarm based on the individualized threshold and the final predicted value.

[0027] In a second aspect, this application provides a computer-readable storage medium storing executable instructions that, when executed by a processor, cause the processor to perform the method described in any embodiment of this application.

[0028] In a third aspect, this application provides an electronic device comprising: one or more processors; and a memory for storing one or more programs, which, when executed by the one or more processors, cause the one or more processors to perform the method as described in any embodiment of this application.

[0029] The multidimensional physiological parameter prediction method, storage medium, and electronic device in this application expand the original data stream to form a more dimensional high-dimensional input vector as the input feature for physiological parameter prediction, providing the model with a comprehensive and dynamic description of physiological state. Furthermore, in the asymmetric neural frequent differential interaction network model adopted, through the collaborative processing of four cascaded sub-modules, feature optimization, state evolution, semantic matching, and security correction are performed on the high-dimensional input vector, fully exploring the temporal patterns and risk modes of physiological data, and finally outputting physiological parameter prediction values ​​that combine security, accuracy, and logical consistency. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0031] Figure 1 This is a flowchart illustrating a multidimensional physiological parameter prediction method in one embodiment;

[0032] Figure 2 This is a schematic diagram of the structure of an electronic device in one embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0035] For example, the terms "first," "second," etc., used in this application may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from another element.

[0036] For example, the terms "comprising" or "including" used in this application indicate the presence of features, steps, operations and / or components, but do not exclude the presence or addition of one or more other features, steps, operations or components.

[0037] This application provides a method for predicting multidimensional physiological parameters, such as... Figure 1 As shown, the method includes:

[0038] Step 110: Obtain the data stream of the user being tested.

[0039] In this embodiment, the data stream {(x i , t i} represents a collection of time-series data related to the physiological state of the user being monitored, collected in real time from medical monitoring equipment. Data stream {(x i , t i In )}, t i x represents the actual arrival timestamp of the data (i.e., the time point when the monitoring device completes a physiological parameter measurement and outputs the data). iLet t represent the observation vector of the i-th dimension (containing the measurement value of at least one physiological parameter). The physiological parameters may include at least one or more of the following: mean arterial pressure (MAP), heart rate (HR), oxygen saturation (SpO2), respiratory rate (RR), body temperature, central venous pressure (CVP), blood glucose (GLU), lactate (Lac), and pulse rate (PR). For example, if the multidimensional physiological parameters include mean arterial pressure (MAP), heart rate (HR), oxygen saturation (SpO2), and respiratory rate (RR), and the user being measured is a critically ill ICU patient, and the monitoring equipment outputs data every 15-60 seconds, then at a certain time t... i =2024-05-20 10:00:00, corresponding to x i =[85mmHg, 78 bpm, 96%, 18 bpm], corresponding to the measured values ​​of mean arterial pressure (MAP), heart rate (HR), oxygen saturation (SpO2), and respiratory rate (RR), respectively. Understandably, the specific physiological parameters collected can be determined based on the actual situation, and will not be elaborated upon here.

[0040] This embodiment collects time-series physiological parameter data of the tested users, retains the original timestamps and multidimensional observation information of the data, and does not perform any form of interpolation or preprocessing to ensure the authenticity and integrity of the data.

[0041] Optionally, the data collection source is commonly used medical monitoring equipment in clinical settings (such as ICU intensive care monitors, perioperative vital signs monitoring equipment, emergency portable monitors, etc.). The core physiological parameters collected include, but are not limited to, mean arterial pressure (MAP), heart rate (HR), oxygen saturation (SpO2), respiratory rate (RR), and other indicators closely related to hemodynamic status. During the collection process, the generation time of each data point (i.e., the actual arrival timestamp t) is recorded in real time. i And encapsulate all physiological parameter measurements at the same timestamp into a multidimensional observation vector x. i This forms the original data stream {(x)} i , t i It is understandable that the timing and frequency of data collection for different physiological parameters may not be the same.

[0042] Construction of observation vectors: For each timestamp t i If a certain physiological parameter is effectively measured, its measured value is included in x. i If the measurement fails due to equipment malfunction, user posture interference, or other reasons, then this parameter will be in x. i The value is temporarily marked as blank and will be marked with a missing indicator value in the subsequent feature engineering stage.

[0043] Continuing with the example of multidimensional physiological parameters including mean arterial pressure (MAP), heart rate (HR), oxygen saturation (SpO2), and respiratory rate (RR), a sample of the raw data stream of a user being tested is: {(x1=[85, 78, 96, 18], t1=2024-05-20 10:00:00), (x2=[82, 80, 95, 19], t2=2024-05-20 10:00:45), (x3=[79,82, , 20], t3=2024-05-20 10:01:30) (SpO2 measurement failed, temporarily recorded as blank)}.

[0044] Step 120: Generate a corresponding high-dimensional input vector X(t) based on the data stream. The high-dimensional input vector includes the most recent observation value at each observation time, data freshness and data missing indication value, change trend feature data and parameter linkage feature data corresponding to the most recent observation value.

[0045] The high-dimensional input vector X(t) refers to a high-dimensional feature vector formed by fusing multiple types of effective information based on data streams. Its core function is to provide the model with a comprehensive and dynamic description of the physiological state, where t is the target time for the current feature construction. Data freshness refers to the time interval between the current time t and the most recent effective observation of the corresponding physiological parameter, used to quantify the timeliness of the data. The data missing indicator value is used to mark whether there is a valid observation of the corresponding physiological parameter at the current time t, and takes only 0 or 1, where 0 indicates no valid observation and 1 indicates a valid observation. The trend feature data are features extracted based on the most recent observation sequence that reflect the dynamic change law of the physiological parameter, including the original observation value, the first-order trend (velocity), and the second-order trend (acceleration). The first-order trend is the rate of change of the original observation value calculated through a sliding time window, and the second-order trend is the rate of change of the first-order trend. The parameter linkage feature data reflects the correlation or synergistic change law between different physiological parameters, used to capture the physiological significance of multi-parameter interaction. Specifically, the parameter linkage feature data is the Pearson correlation coefficient or the proportion of unidirectional changes between different physiological parameters within the sliding time window.

[0046] For example, the current time t = 2024-05-20 10:01:30, and the most recent valid observation time t of the MAP parameter. j_last =2024-05-20 10:00:45, then the data freshness Δt of this parameter is... j (t)=tt j_last =45s; Due to a temporary malfunction of the monitoring equipment, no SpO2 data was collected at t=2024-05-20 10:02:00. Therefore, the missing SpO2 indicator value m j(t)=0, the other parameters were collected normally, and the missing indicator values ​​were all 1;

[0047] The formula for calculating the first-order trend s(t) is s(t) = (v(t) - v(tw)) / w; the formula for calculating the second-order trend a(t) is a(t) = (s(t) - s(tw)) / w. Here, w represents the sliding time window size, v(t) represents the observed value at time t, v(tw) represents the observed value at time tw, s(tw) represents the first-order trend at time tw, and a(t) represents the second-order trend at time t. For example, if the last five observations of the MAP parameter are 85 mmHg, 82 mmHg, 79 mmHg, 76 mmHg, and 73 mmHg, the corresponding first-order and second-order trend data can be calculated using a sliding window. Taking the Pearson correlation coefficient as an example, by calculating the Pearson correlation coefficient between MAP and HR within the sliding window, the parameter linkage characteristic data of MAP and HR is -0.85, indicating a strong negative correlation between the two.

[0048] In this embodiment, by fusing three types of information—state features (recent observations, data freshness, and missing data), trend features, and parameter linkage features—a high-dimensional input vector that can comprehensively reflect the physiological state is constructed, providing the model with rich decision-making basis.

[0049] Specifically, firstly, the state features (most recent observation, data freshness, and missing value) of each physiological parameter at time t are extracted based on the original data stream; secondly, the trend features of each parameter are calculated using a sliding time window; then, the parameter linkage features between different parameters are calculated; finally, all features are concatenated according to preset rules to form a high-dimensional input vector X(t). No interpolation or alignment is performed throughout the process, preserving the discrete characteristics and timeliness information of the original data.

[0050] Specifically, regarding state characteristics: regarding the most recent observation x j_last (t), for each physiological parameter j, at time t, find the measurement value of the most recent valid observation in the raw data stream. If a valid observation exists, it is used directly; if all historical observations are invalid, it is marked as the default value (e.g., 0, used in conjunction with the missing data indicator). Regarding data freshness Δt j (t), through the formula Δt j (t)=tt j_last Calculate, where t j_last Let Δtj(t) be the timestamp of the most recent valid observation for parameter j. If there are no valid observations, then Δtj(t) is the difference between the current time and the data acquisition start time. Regarding the missing indicator value m... j (t), if parameter j has a valid most recent observation at time t, then m j (t)=1; otherwise mj (t)=0.

[0051] Regarding the characteristics of the changing trends: First, a sliding time window can be set. For example, based on the update frequency of clinical data, the window size w can be selected from 3 to 10 observation periods, and the window sliding step is 1 observation period. Based on this sliding time window, the corresponding first-order trend and second-order trend can be calculated respectively.

[0052] For parameter linkage characteristics: calculate the Pearson correlation coefficient of any two parameters (such as MAP and HR, SpO2 and RR) within the sliding window, and use the calculated Pearson correlation coefficient as the corresponding parameter linkage characteristic; or calculate the proportion of changes in the same direction, such as the proportion of observation periods in which two parameters increase or decrease simultaneously within the statistical window to the total number of periods, as a supplementary linkage characteristic.

[0053] High-dimensional input vector concatenation: For each physiological parameter j, first concatenate its state features (x... j_last (t), Δt j (t), m j (t) and trend characteristics (v) j (t), a j (t) is concatenated to form the feature vector of the parameter; then the feature vectors of all parameters are concatenated with the linkage features (Pearson correlation coefficient, same direction change ratio) of all parameter pairs to form the final high-dimensional input vector X(t).

[0054] For example, for four physiological parameters (e.g., MAP, HR, SpO2, RR), if each individual physiological parameter extracts five core features to form an independent sub-vector, then the feature sub-vector dimension for these four physiological parameters is 4 × 5 = 20 dimensions. Any two physiological parameters constitute one parameter pair, resulting in six parameter pairs. Each pair of linked features has two dimensions, so the linked feature dimension is 12 dimensions. Therefore, the total dimension of X(t) is 20 + 12 = 32 dimensions. Understandably, if the physiological parameters contain more or fewer dimensions, the corresponding feature sub-vector dimension, the number of parameter pairs, and the total dimension of the final X(t) will adaptively change. For example, taking seven physiological parameters (e.g., MAP, HR, SpO2, RR, systolic blood pressure SBP, diastolic blood pressure DBP, and oxygen perfusion index PI), the corresponding feature sub-vector dimension is 7 × 5 = 35 dimensions, the linked feature dimension is 42 dimensions, and the total dimension of the final X(t) is 35 + 42 = 77 dimensions.

[0055] Step 130: Input the high-dimensional input vector into the pre-constructed asymmetric neural frequent differential interaction network model. After processing by the cascaded variable selection layer, continuous-time evolution encoder, prototype semantic interaction layer and asymmetric security gated decoder in the asymmetric neural frequent differential interaction network model, the final predicted value of the preset multidimensional physiological parameters is output.

[0056] In this embodiment, the preset asymmetric neural frequent differential interaction network model is a deep learning model designed to process discrete physiological data at different frequencies. Through the collaborative work of four cascaded sub-modules (variable selection layer, continuous-time evolution encoder, prototype semantic interaction layer, and asymmetric security-gated decoder), it achieves highly secure and logically consistent physiological parameter prediction. The final predicted value refers to the predicted value (e.g., minimum value) of multidimensional physiological parameters output by the model within a preset future time window (e.g., 5 min, 10 min, 15 min). This value is a conservative prediction result after safety correction and is used to assess the risk of the tested user's future physiological state. For example, if the model outputs a predicted minimum MAP value of 70 mmHg for the next 5 minutes, 65 mmHg for the next 10 minutes, and 62 mmHg for the next 15 minutes, it indicates that the tested user has a cumulative risk of hypotension.

[0057] This embodiment utilizes the collaborative processing of four cascaded sub-modules to perform feature optimization, state evolution, semantic matching, and security correction on high-dimensional input vectors. This fully explores the temporal patterns and risk models of physiological data, ultimately outputting physiological parameter predictions that combine safety, accuracy, and logical consistency.

[0058] The variable selection layer dynamically evaluates the effectiveness of each feature channel based on a gated residual network, suppressing features with severe missing values ​​or high noise levels, thus improving the model's anti-interference ability. The continuous-time evolution encoder uses a jump-flow mechanism to adapt to irregular time intervals in the data, eliminating the need for interpolation alignment and realistically restoring the temporal evolution process of physiological states. The prototype semantic interaction layer, based on a pre-set library of typical clinical risk patterns, uses an attention mechanism to achieve semantic matching between the current user's state and historical risk patterns, enhancing model interpretability and generating risk coefficients. The asymmetric security-gated decoder combines basic regression prediction with security bias correction to achieve a clinical safety orientation that prioritizes avoiding false alarms and preventing missed alarms, outputting the final predicted value.

[0059] Specifically, firstly, a variable selection layer dynamically assigns weights and filters noise from the high-dimensional input vector to retain effective features; secondly, a continuous-time evolution encoder adapts to irregular time intervals to simulate the continuous evolution of physiological states and obtain the final hidden state; then, a prototype semantic interaction layer retrieves the similarity between the current state and typical clinical risk patterns to generate semantically enhanced features and risk coefficients; finally, an asymmetric security-gated decoder performs security correction on the basic predicted values ​​and outputs the final predicted values ​​of multidimensional physiological parameters (such as the minimum values ​​of multiple future time windows).

[0060] In this embodiment, the original data stream with real timestamps is directly obtained without interpolation alignment, thus avoiding the false trends introduced by traditional interpolation methods. Based on the original data stream, the most recent observation, data freshness, data missing indication, trend feature data, and parameter linkage feature data are generated and then concatenated to form a high-dimensional input vector X(t), providing the model with a comprehensive and dynamic description of the physiological state. After processing by the cascaded variable selection layer, continuous-time evolution encoder, prototype semantic interaction layer, and asymmetric security gated decoder in the model, the final output prediction value can reduce the false negative rate of malicious events and improve the consistency of multi-time-window prediction logic.

[0061] In one embodiment, the processing of the variable selection layer includes: dynamically assigning weights and filtering noise on the high-dimensional input vector based on a gated residual network to form weighted features of multidimensional physiological parameters.

[0062] This embodiment utilizes the adaptive feature selection capability of gated residual networks to dynamically evaluate the weights of each feature channel in the high-dimensional input vector, thereby enhancing the contribution of effective features, suppressing noise and interference from missing features, generating high-quality weighted features, and providing accurate input for subsequent network processing.

[0063] Specifically, feature transformation is performed on the high-dimensional input vector, and gate weights for feature channels adapted to multi-dimensional physiological parameters are generated in parallel. By weighted fusion of the gate weights and the transformed features, feature channels with severe missing features or high noise are suppressed, noise filtering is achieved, and weighted features of multi-dimensional physiological parameters are formed.

[0064] The high-dimensional input vector is transformed and mapped to a feature space that adapts to the gating mechanism. Gating weights corresponding to each physiological parameter feature channel are generated in parallel to quantify the effectiveness of each channel. By weighted fusion of the gating weights and the transformed features, feature channels with severe missing features or high noise are significantly suppressed, while the core information of effective features is retained, ultimately forming a weighted feature of multidimensional physiological parameters.

[0065] The feature transformation is achieved through linear layer mapping to obtain the mapped features; the gate weights are generated by calculating the Sigmoid function; and the weighted fusion includes element-wise multiplication of the mapped features and the gate weights to generate the weighted features.

[0066] The gated residual network achieves its function through a four-step process: feature transformation, weight generation, weighted fusion, and residual connection. First, the high-dimensional input vector is linearly transformed and mapped to a suitable feature space. Second, the gate weights of each feature channel are generated through the sigmoid activation function to quantify the effectiveness of the features. Then, the transformed features are multiplied element-wise with the gate weights to filter out invalid information. Finally, the core information of the original features is preserved through residual connections to avoid information loss during the feature transformation process.

[0067] The gated residual network structure parameters include the input dimension, hidden layer dimension, and residual connection units. The input dimension is consistent with the dimension of the high-dimensional input vector X(t) (let's say D dimensions, such as 32 dimensions in the previous example); the hidden layer dimension can be set to 2D dimensions (such as 64 dimensions), using the ReLU activation function; the residual connection unit directly adds the input vector X(t) to the hidden layer output, with the formula Res(X(t))=ReLU(Wr×X(t)+br)+X(t), where Wr is the residual layer weight matrix, br is the residual layer bias vector, and the initialization uses a He normal distribution.

[0068] During the feature transformation process, the input vector X(t) is mapped to the mapped features through a linear layer. (X(t)), the specific formula of which can be: (X(t)) = Wt × Res(X(t)) + bt, where Wt is the feature transformation weight matrix (2D × D dimension), and bt is the feature transformation bias vector (2D × 1 dimension). After mapping, the feature dimension is consistent with the hidden layer dimension (2D dimension). For example, if X(t) is a 32-dimensional vector, Res(X(t)) remains 64-dimensional after ReLU activation. Through Wt (64 × 32) mapping and bt (64 × 1) bias, a 64-dimensional feature is obtained. (X(t)).

[0069] The gating weights are calculated using the Sigmoid function. During gating weight generation, the weight calculation formula is g(X(t)) = σ(Wg × Res(X(t)) + bg), where Wg is the gating weight matrix (2D × D), bg is the gating bias vector (2D × 1), σ is the Sigmoid function, and the output weight vector g(X(t)) has a 2D dimension, with each element ranging from (0,1). bg can be initialized to -1.0 to bias the initial weights below 0.5, avoiding excessive retention of noise features in the initial stage; Wg is initialized using a Xavier uniform distribution. For example, the gating weight corresponding to a noise feature channel in Res(X(t)) is calculated as g = σ(0.2 × Res(X(t))_k - 1.0) = 0.08.

[0070] In the process of generating weighted features, the element-wise multiplication formula can be: Z(t) = (X(t))⊙g(X(t)), where ⊙ denotes element-wise multiplication, Z(t) is 2D, and noisy feature channels are suppressed because their weights are close to 0, while effective feature channels are preserved and enhanced. For example, In (X(t)), the noise eigenvalue is 0.8, corresponding to g=0.08, and after multiplication, it is 0.064 (suppressed); the effective eigenvalue is 0.7, corresponding to g=0.91, and after multiplication, it is 0.637 (preserved), and finally Z(t)=[0.637,0.064, ...].

[0071] In this embodiment, the variable selection layer employs a gated residual network. Through dynamic weight allocation, high weights are assigned to effective features (such as the first-order trend of MAP) in the high-dimensional input vector, while low weights are assigned to noisy features (such as severely missing linkage features), achieving accurate noise filtering. The introduction of residual connections avoids information loss during feature transformation, ensuring the integrity of effective features. The generation of weighted features Z(t) allows subsequent networks to focus only on high-quality features, reducing noise interference, improving the model's accuracy in depicting physiological states, and laying the foundation for subsequent predictions.

[0072] In one embodiment, the processing of the continuous-time evolution encoder includes: using a Jump-Flow mechanism to adapt to irregular time intervals, instantly updating the hidden state of the network through a time-series update unit when the observed data arrives, and simulating the continuous evolution of the hidden state through a differential equation network during the observation blank period to obtain the final hidden state.

[0073] This embodiment utilizes the Jump-Flow dual-mechanism collaboration to adapt to irregular time intervals in physiological data. It eliminates the need for interpolation alignment, rapidly updates the latent state at discrete observation times, and continuously simulates the evolution of physiological states during blank time periods, thus realistically reproducing the temporal trajectory of the physiological state of the tested user.

[0074] The Jump mechanism is responsible for processing discrete observation data. It updates the hidden state of the network instantaneously by combining the current weighted features with a time-series update unit (such as GRU) to capture the state changes brought about by the observation data. The Flow mechanism is responsible for filling the observation gaps. It simulates the continuous dynamic changes of physiological state through differential equation networks (such as neuronormal differential equations ODE) and calculates the integral results through a numerical solver to achieve the smooth evolution of the hidden state. The two work alternately to form a closed loop of "discrete update-continuous evolution" to adapt to the characteristics of discrete data with different frequencies.

[0075] The time-series update unit is a recurrent neural network unit, the differential equation network is a neural ordinary differential equation network, and the continuous evolution result of the hidden state during the observation blank time period is integrally calculated by a numerical solver.

[0076] Recurrent neural network (RNN) units are neural network components capable of processing temporal information and capturing long-term and short-term dependencies in data. They may include GRUs (Gated Recurrent Units) and LSTMs (Long Short-Term Memory networks), used for instantaneous updates of hidden states at discrete observation times. Neuronormal differential equation networks (NDEs) are models that parameterize the rate of change function on the right-hand side of differential equations using neural networks. The integral results are calculated using a numerical solver, achieving a smooth evolution of the hidden state over continuous time. For example, a neuronormal differential equation network f... θ A two-layer MLP structure is adopted, with the hidden state h(t) and the current time t as inputs and dh / dt as outputs to simulate the continuous change trend of MAP over time.

[0077] Specifically, the recurrent neural network unit is a gated recurrent GRU unit; the numerical solver is a Runge-Kutta solver or an Euler solver.

[0078] The method of instantaneously updating the network hidden state through a time-series update unit at the arrival of observed data, and simulating the continuous evolution of the hidden state through a differential equation network during the observation gap period to obtain the final hidden state includes: instantaneously updating the network hidden state through a GRU unit combined with the weighted features at the arrival of observed data; simulating the continuous dynamic evolution of the hidden state through a neural differential equation network during the observation gap period, and calculating the final hidden state through a numerical solver.

[0079] The dimension of the hidden state is set to be consistent with that of the weighted feature Z(t) (2D dimension, such as 64 dimensions); during the initialization of the hidden state, zero initialization is used, i.e., h(0)=[0,0,...,0]. T (64-dimensional), the first update is completed through the Jump mechanism after the first observation data arrives.

[0080] In the Jump process (discrete update), taking the GRU as an example when the time-series update unit is selected, the core formula includes: Reset gate r = σ(Wr × [Z(t)] i h(t) i-1 )]+br); Update gate z=σ(Wz×[Z(t) i h(t) i-1 )]+bz); Candidate hidden state h'=tanh(W×[r⊙h(t i-1 Z(t) i )]+b); Hidden state update h(t) i )=z⊙h(t i-1 )+(1-z)⊙h'. Where Wr, Wz, and W are GRU weight matrices (each with a dimension of 2D×4D), br, bz, and b are bias vectors (each with a dimension of 2D×1), and Z(t i ) represents the current weighted feature, [Z(t) i h(t) i-1 )] represents vector concatenation, and tanh is the hyperbolic tangent activation function.

[0081] For example, t i =10:00:45, Z(t) i ) is a 64-dimensional weighted feature, h(t) i-1 h(10:00:00) is a 64-dimensional hidden state. After GRU calculation, r=[0.12,0.85,...], z=[0.78,0.21,...], h'=[0.32,-0.15,...], and finally h(t) i = [0.28, -0.11, ...] (64 dimensions), completing the instantaneous update of the hidden state.

[0082] During the flow process, a differential equation network f is used. θ Taking a 2-layer MLP (Multilayer Perceptron) implementation as an example, the inputs are the hidden state h(t) and the current time t, and the output is the rate of change of the hidden state dh / dt, with the formula dh / dt=f θ (h(t),t)=ReLU(W1×[h(t);t]+b1)×W2+b2, where W1 (dimension 4D×(2D+1)) and W2 (dimension 2D×4D) are MLP weight matrices, and b1 (4D×1) and b2 (2D×1) are bias vectors.

[0083] The numerical solver can be the Runge-Kutta solver (RK4) to meet the needs of high-precision calculation; when real-time requirements are high, the Euler solver can also be selected to balance accuracy and speed.

[0084] During the integration process, for the blank time period (t) i ,ti+1 The time step is set to Δt_step = 1s (to ensure smooth evolution), and the hidden states are calculated iteratively using the RK4 solver:

[0085] ;

[0086] ;

[0087] ;

[0088] ;

[0089] .

[0090] Where k1 represents the instantaneous rate of change (initial slope) at the initial time ti, and k2 represents the midpoint of time t. i +Δt_step / 2 is the estimated rate of change (first slope at the midpoint); k3 represents the midpoint of time t. i +Δt_step / 2 is the corrected rate of change (second slope at the midpoint); k4 represents the time endpoint t. i +Δt_step is the estimated rate of change (final slope). The above formula is iterated repeatedly until t. i+1 At time t, we obtain h(t) i+1 )_flow, combined with t i+1 Z(t) at time i+1 Updated to the final hidden state h via the Jump mechanism. final =h(t i+1 ).

[0091] For example, the blank time period is from ti=10:00:45 to t i+1 =10:01:30 (Δt=45s), Δt_step=1s, then after 45 iterations, h(t) i The value evolved from [0.28, -0.11, ...] to h(t). i+1 )_flow=[0.25, -0.14, ...], and then combined with Z(t i+1 Updated to h final =[0.23, -0.16, ...].

[0092] In this embodiment, the Jump-Flow mechanism of the continuous-time evolution encoder has a clear division of labor: the Jump process captures state abrupt changes brought about by observed data through GRU units, adapting to the instantaneous update characteristics of discrete data; the Flow process simulates the continuous evolution of blank time periods through ODE networks and numerical solvers, restoring the natural change law of physiological state. Compared with traditional interpolation methods, this mechanism does not require mandatory equal intervals between data, thus avoiding the introduction of spurious trends and preserving the temporal information corresponding to data freshness; moreover, the high-precision calculation of the Runge-Kutta solver ensures the accuracy of continuous evolution, making the final hidden state h... final It can accurately reflect the temporal trajectory of the physiological state of the tested user, thus improving the basic accuracy of prediction.

[0093] In one embodiment, the processing of the prototype semantic interaction layer includes: calculating the similarity distribution between the final hidden state and a preset prototype library through an attention mechanism, generating semantic enhancement features by combining the similarity distribution, and mapping the corresponding risk coefficient.

[0094] This embodiment constructs a prototype library of typical clinical risk patterns, uses an attention mechanism to achieve semantic matching between the current physiological state of the tested user and historical typical patterns, generates semantically enhanced features that integrate clinical experience, and quantifies the output risk coefficient, thereby improving both prediction accuracy and model interpretability.

[0095] First, a trainable prototype matrix (prototype library) is constructed to store the feature representations of various typical clinical risk patterns. Then, the correlation between the final hidden state and each prototype is calculated through an attention mechanism to obtain the similarity distribution. Next, the prototype features and the current hidden state are weighted and fused based on the similarity distribution to generate semantically enhanced features. Finally, the similarity distribution and the hidden state are concatenated, and the risk coefficient is obtained through linear mapping to quantify the risk level of the current state.

[0096] Specifically, the similarity distribution is obtained by calculating the correlation between the final hidden state and the prototype matrix using the Softmax function; the semantic enhancement feature is generated by superimposing the weighted sum of the prototypes in the prototype matrix using the similarity distribution with the final hidden state.

[0097] The process of building the prototype library (prototype matrix P) is as follows:

[0098] Wherein, the prototype matrix parameter is: P∈R K×d Where K is the number of prototypes (preferably K=10~20, for example K=15), d is the dimension of the final hidden state hfinal (2D dimension, such as 64 dimensions), and each prototype P k (k=1~15) corresponds to a typical risk pattern.

[0099] During prototype initialization, based on real clinical case data, latent state features of various typical risk patterns (such as acute hypotension, heart rate oscillation, and precursors of respiratory failure) are extracted as initial prototypes, which are then continuously optimized through model training. For example, prototype P1 corresponds to the acute hypotension precipitous drop pattern, with initial values ​​of [0.78, -0.65, 0.52, ...] (64 dimensions); P2 corresponds to the heart rate oscillation worsening pattern, with initial values ​​of [-0.81, 0.43, -0.39, ...] (64 dimensions).

[0100] In the calculation of similarity distribution, h can be calculated using the dot product method. final The correlation with each prototype is used to quantify the probability that the current user being tested belongs to a certain typical risk pattern. The formula is corr_k=(h final ×P k T ) / (d 0.5 (k=1~K), to obtain the K-dimensional correlation vector corr=[corr_1, corr_2, ..., corr_K).

[0101] Normalization: The relevance vector is converted into a similarity distribution using the Softmax function, with the formula α = Softmax(corr) = exp(corr_k) / Σ(exp(corr_m)), (m = 1 ~ K), where α is a K-dimensional vector. k This represents the probability of matching the current state with the k-th prototype, and their sum is 1.

[0102] For example, h final =[0.23, -0.16, ...] (64 dimensions), with corr_1=0.85 for P1, corr_2=0.12 for P2, and all other corr_m values ​​less than 0.3. After Softmax calculation, α=[0.82, 0.03, 0.02, ..., 0.01] (15 dimensions), indicating that the current state has the highest matching degree with the acute hypotension cliff-like drop pattern.

[0103] Semantic enhancement feature h semantic The generation process is as follows:

[0104] First, let's consider the similarity distribution α mentioned above. k With the corresponding prototype P k After performing a weighted summation, we obtain the weighted fusion prototype P. weighted =Σ(αk×Pk)(k=1~K), where P k Let P be the k-th row vector of the prototype matrix P. weighted The dimension of is d-dimensional, which is the comprehensive feature after weighting all prototypes according to similarity.

[0105] Semantic enhancement feature h semantic The calculation formula is h semantic =h final +P weighted By using residual connections, the current state features are superimposed with the prototype fusion features, preserving the uniqueness of the current state while incorporating typical pattern experience.

[0106] In the calculation of the risk coefficient, vector concatenation is first performed, and h is... final (d-dimensional) and α (K-dimensional) are concatenated to form a d+K-dimensional vector, resulting in a concatenated vector concat=[hfinal;α] (e.g., 64+15=79 dimensions). During the linear mapping and activation process, the formula λ... risk =σ(Wr×concat+br), where Wr is the risk mapping weight matrix (dimension 1×(d+K)), br is the risk mapping bias vector (dimension 1×1), and σ is the Sigmoid function, ensuring λ risk ∈[0,1].

[0107] For example, concat is a 79-dimensional vector, Wr concat+br=1.52, λ risk =σ(1.52)=0.82, indicating that the current risk level of the tested user is high risk.

[0108] In this embodiment, the prototype library of the prototype semantic interaction layer is constructed based on real clinical risk patterns. Each prototype corresponds to a clear clinical meaning (such as acute hypotension pattern). The similarity distribution α can intuitively inform medical staff "which typical risk pattern is most similar to the current state of the tested user", thus solving the black box problem of traditional models. In addition, semantic enhancement features h semantic By integrating experience from the current state and typical patterns, the model possesses reasoning abilities similar to a doctor's ability to "associate with historical cases," thus improving prediction accuracy; additionally, the risk coefficient λ... risk By quantifying the correlation strength between similarity distribution and latent states, a clear risk reference is provided for the prediction results, enabling medical staff to quickly determine the reliability and urgency of the prediction results.

[0109] In one embodiment, the processing of the asymmetric security gating decoder includes: performing regression calculation on the semantic enhancement features to obtain a basic predicted value; correcting the basic predicted value based on a negative security bias term controlled by the risk coefficient; and using the predicted minimum value of the multidimensional physiological parameters as the final predicted value.

[0110] Based on the clinical safety requirement of minimizing false alarms and preventing missed reports, this embodiment decodes semantically enhanced features by combining basic regression prediction with negative safety bias correction, outputting conservative and safe predicted minimum values ​​of physiological parameters, thereby reducing the risk of missed reports of serious events from the network output level.

[0111] Optionally, the semantic enhancement features are first fitted using a linear regression layer to obtain the basic predicted values ​​for each future time window; then, a negative safety bias term controlled by the risk coefficient is constructed, with the higher the risk coefficient, the stronger the negative effect of the bias term; finally, the basic predicted values ​​are superimposed with the safety correction term to obtain the final predicted minimum value, ensuring that the prediction results are sufficiently conservative under high-risk conditions and avoiding underestimation of the true value leading to missed reports.

[0112] In one embodiment, the negative safety bias term is generated by a preset nonlinear function. The output of the nonlinear function is always positive. After sign inversion, the negative safety bias term is formed, so that the safety correction can only lower the base prediction value.

[0113] Specifically, the nonlinear function is a Softplus function; the correction process is to superimpose the base prediction value with a safety correction term, which is the product of a trainable scaling parameter, a risk coefficient, and a negative safety bias term.

[0114] In the calculation of the basic forecast value, a linear regression formula is used: y base (h)=Wb×h semantic +bb, where h is the prediction time window (h=5min, 10min, 15min), Wb is the regression weight matrix (dimensional M×d, M is the number of physiological parameters, such as 4-dimensional), bb is the regression bias vector (dimensional M×1), and y base (h) is an M-dimensional baseline prediction vector (corresponding to the predicted values ​​of each physiological parameter). Wb is initialized using a Xavier normal distribution, and bb is initialized to the mean of the clinical normal range of each physiological parameter (e.g., bb for MAP is initialized to 85 mmHg).

[0115] For example, h semantic The matrix is ​​64-dimensional, M=4 (MAP, HR, SpO2, RR), Wb is a 4×64 matrix, and bb is a 4×1 vector. The calculated y... base (5min) = [72mmHg, 84 bpm, 95%, 20 bpm], y base (10min) = [69mmHg, 86 bpm, 94%, 21 bpm].

[0116] During the generation of the negative safety bias term, the bias term uses the formula δ safety=-Softplus(β), where β is a trainable parameter (dimensional M×1). The Softplus function is defined as Softplus(x)=ln(1+exp(x)), and its output is always positive. Combined with the negative sign, δ safety Always negative, ensuring it only lowers the predicted value. β can be initialized to [1.0, 1.0, 1.0, 1.0]. T (Corresponding to 4 physiological parameters), after calculation by Softplus, Softplus(β) = [1.313, 1.313, 1.313, 1.313]. T Therefore, the initial δ safety =[-1.313mmHg, -1.313 bpm, -1.313%, -1.313 bpm]. β is updated iteratively through backpropagation, making δ... safety Safety requirements adapted to different physiological parameters (such as MAP's delta) safety The absolute value is larger because the risk of low blood pressure is more critical.

[0117] In the calculation of the safety correction term and the final predicted value, the correction term uses the formula: Δy(h)=γ×λ risk ×δ safety , where γ is a trainable scaling parameter (dimensional M×1) used to adjust the strength of the safety bias, initialized to [2.0, 1.5, 1.0, 1.0]. T (MAP has the highest scaling intensity). The final predicted value formula is: y pred (h)=y base (h)+Δy(h) represents the predicted minimum value of each physiological parameter.

[0118] For example, λ risk =0.82 (high risk), γ=[2.0, 1.5, 1.0, 1.0] T δ safety =[-1.313, -1.313, -1.313, -1.313] T The calculated Δy(5min) = [2.0 × 0.82 × (-1.313), 1.5 × 0.82 × (-1.313), ...] = [-2.15 mmHg, -1.61 beats / min, -1.08%, -1.08 beats / min]; the final predicted value y pred (5min) = [72 - 2.15 = 69.85 mmHg, 84 - 1.61 = 82.39 beats / min, 95 - 1.08 = 93.92%, 20 - 1.08 = 18.92 beats / min], y pred(10min) = [69 - 2.15 = 66.85 mmHg, ...], the predicted value is significantly conservative to reduce the risk of underreporting.

[0119] In this embodiment, the negative security bias term δ of the asymmetric security gating decoder safety The Softplus function ensures that the value is always negative, forming a one-way valve constraint. This ensures that safety corrections can only lower the predicted value, structurally preventing underreporting situations where the predicted value is greater than the actual value; the risk coefficient λ risk The combination with the scaling parameter γ enables dynamic matching of the safety correction strength with the risk level. That is, the correction effect is stronger and the predicted value is more conservative under high-risk conditions, while the correction effect is mild under low-risk conditions, balancing accuracy and safety. In addition, different γ and β are set for different physiological parameters, so that the safety mechanism is adapted to the clinical risk weight of each parameter (e.g., the safety correction strength of MAP is higher than that of SpO2), which further improves the targeting and effectiveness of the safety mechanism, and ultimately significantly reduces the underreporting rate of adverse events such as hypotension.

[0120] The multidimensional physiological parameter prediction method in this application uses a continuous-time evolution encoder's flow process to simulate the continuous cumulative evolution of physiological states, allowing the latent states of long time windows to naturally inherit the risk characteristics of short time windows. The correction terms of the asymmetric security-gated decoder exhibit a cumulative effect as the time window lengthens (e.g., if the base prediction value at 10 minutes is lower than that at 5 minutes, after adding the same correction term, the final prediction value still maintains that the long time window is ≤ the short time window). The synergy of these two methods ensures the monotonicity of the predicted minimum value across multiple time windows, avoids the logical contradiction that the risk of long time windows is lower than that of short time windows, enhances the confidence of medical staff in the prediction results, and facilitates practical clinical application.

[0121] In one embodiment, the multidimensional physiological parameter prediction method of this application further includes a model training process, which can be performed before step 110 above. The model training process includes: constructing a variable selection layer, a continuous-time evolutionary encoder, a prototype semantic interaction layer, and an asymmetric security gated decoder in sequence to form an asymmetric neural frequent differential interaction network to be trained; inputting preset training data into the asymmetric neural frequent differential interaction network to be trained for iterative training; using a composite loss function including asymmetric security regression loss, multi-time-window consistency constraint loss, and prototype orthogonal regularization term to train and optimize the asymmetric neural frequent differential interaction network to be trained, and obtaining a trained asymmetric neural frequent differential interaction network model.

[0122] The training data consists of a set of labeled clinical monitoring data used for model training. It includes a time-series physiological parameter data stream of historical test subjects and corresponding real physiological state labels (such as the true values ​​of physiological parameters in future time windows and the occurrence of risk events), forming the basis for the model to learn predictive patterns. The preprocessed training data has the same data format as the high-dimensional input vector mentioned above. For example, the training data includes monitoring data from 10,000 ICU patients. Each case's data covers parameters such as MAP, HR, SpO2, and RR for 24 consecutive hours (sampling interval 15s-60s), labeled with the true minimum values ​​of the parameters for the next 5, 10, and 15 minutes and whether a hypotensive event has occurred.

[0123] Iterative training refers to the process of repeatedly inputting training data into the network to be trained, calculating prediction results through forward propagation, updating network parameters through backpropagation, and gradually reducing the loss function value to bring the model closer to its optimal prediction performance. Electronic devices pre-set conditions for terminating iterations. These conditions can be that the number of iterations reaches a preset threshold, or that the composite loss value calculated using the composite loss function is within a preset numerical range or converges. For example, the number of iterations can be set to 100 rounds, with training data input into the network in batches (batch size 32) in each round. After calculating the composite loss, the parameters are updated using gradient descent until the loss function value converges or the maximum number of iterations is reached.

[0124] In this application, the composite loss function integrates the optimization objective function of multiple loss terms. By weighted combination of different loss terms, it simultaneously constrains multiple performance indicators of the model (such as security, logical consistency, and pattern diversity) to achieve optimal overall performance. Asymmetric security regression loss (L... asym This refers to a loss term designed for medical scenarios where the cost of false alarms outweighs the cost of false alarms. It imposes a higher penalty on false alarm errors where the predicted value is greater than the actual value, forcing the model to learn a safety-oriented prediction logic; multi-time-window consistency constraint loss (L... mono This refers to the loss term used to ensure logical consistency of results across different prediction time windows. It forces the predicted minimum value of a long time window to be no higher than that of a short time window, avoiding the contradiction that the risk of a long time window is lower than that of a short time window; the prototype orthogonal regularization term (L) proto_reg Regularization terms are used to constrain the different typical risk patterns in the prototype library, thereby reducing the correlation between prototypes and ensuring that the model can distinguish between different types of clinical risk patterns.

[0125] In one embodiment, the composite loss function is a weighted sum of asymmetric safe regression loss, multi-time-window consistency constraint loss, and prototype orthogonal regularization term.

[0126] Optionally, the weights of the three loss terms can be designed appropriately according to requirements. For example, if the weights are all 1, then the composite loss function L... total =Lasym +L mono +L proto_reg These correspond to three major constraints: security orientation, logical consistency, and prototype diversity, balancing the core capabilities of the model. For L... asym The penalty weight for missed detection error can be set to w+=5, and the penalty weight for false alarm error can be set to w-=1. When the model misses a low blood pressure, the loss value is 5 times the squared error, and when it is a false alarm, it is 1 times the squared error, guiding the model to prefer false alarms to missed detections. For L mono Three time windows can be set: 5 min, 10 min, and 15 min. If the model predicts a minimum MAP value (70 mmHg) at 15 min that is higher than at 10 min (65 mmHg), a penalty is applied through this loss term to correct the model's prediction logic. For L proto_reg If the features of the two prototypes, acute hypotension and oscillatory deterioration, are too highly correlated (>0.8), the loss term will increase, forcing the model to optimize the prototype vectors during training and reduce the correlation to below 0.3.

[0127] The asymmetric neural frequent differential interaction network model constructed in this application is built by sequentially constructing each sub-module in the cascaded order of "variable selection layer → continuous time evolution encoder → prototype semantic interaction layer → asymmetric security gated decoder". The output of each sub-module serves as the input of the next sub-module, forming a closed loop of data flow. All trainable parameters in the network (such as weight matrix, bias vector, prototype matrix, scaling parameters, etc.) are initialized to ensure that the network to be trained has an optimizable basic structure.

[0128] The variable selection layer takes preprocessed training data (with a dimension of D=32) as input and outputs a weighted feature Z(t) (also 32-dimensional), which is then connected to the input of the continuous-time evolutionary encoder. The continuous-time evolutionary encoder takes Z(t) and its corresponding timestamps as input and outputs the final hidden state h. final (Dimension d=32), connected to the input port of the prototype semantic interaction layer. The input of the prototype semantic interaction layer is h. final The output is the semantically enhanced feature h. semantic (32 dimensions) and risk coefficient λ risk (1-dimensional), connected to the two input ports of the asymmetric secure gating decoder. The input of the asymmetric secure gating decoder is h. semantic and λ risk The output is the predicted minimum value y_pred for multiple future time windows (dimension M=4, corresponding to 4 types of core physiological parameters), completing the network output.

[0129] Network parameter initialization configuration: For weight matrices (such as Wt and Wg in the variable selection layer, Wr and Wz in the GRU unit, and Wb in the decoder), Xavier normal distribution or He normal distribution is used for initialization to avoid gradient vanishing / exploding due to excessively large or small initial parameters; for bias vectors (such as bt, bg, bb, etc.), zero initialization or small constant initialization is used (such as initializing bg to -1.0 to suppress noise features in the initial stage); for the prototype matrix P (K×d, K=15, d=32), the latent state features can be extracted as initial values ​​based on historical data of 15 typical clinical risk patterns to ensure that the prototype has a preliminary risk pattern discrimination ability; for the trainable scaling parameter γ (4-dimensional), it is initialized to [2.0, 1.5, 1.0, 1.0] to adapt to the safety weights of different physiological parameters (MAP has the highest scaling strength).

[0130] During the iterative training of the model, firstly, complete labeled training data (including time-series physiological data and real labels) is prepared and preprocessed; then, the training data is input into the network to be trained in batches, and the prediction results are obtained through forward propagation; next, the composite loss function value is calculated, and all trainable parameters of the network are updated through the backpropagation algorithm; the above process is repeated until the loss function value converges or the preset number of iterations is reached, thus completing the model training optimization and validation.

[0131] For example, the training data can be divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for parameter updates, the validation set is used to monitor overfitting, and the test set is used for final performance evaluation. The number of iterations (Epoch) is preset to 100 epochs. If the validation set loss does not decrease for 10 consecutive epochs, an early stopping mechanism is triggered to avoid overfitting. The batch size is set to 32 to balance training efficiency and gradient stability. The optimizer can be the Adam optimizer, with an initial learning rate of 1e-4, which decays to 0.5 every 20 epochs (learning rate decay strategy). Gradient clipping can be set with a gradient norm threshold of 5.0 to avoid gradient explosion during training.

[0132] For example, the formula for the composite loss function is L. total =w1×L asym +w2×L mono +w3×L proto_reg Where w1, w2, and w3 take values ​​of 1, 1, and 0.1 respectively, to balance the priority of each loss term.

[0133] ;

[0134] ;

[0135] .

[0136] Indicates the penalty weight for underreporting error. The values ​​represent the false alarm error penalty weights, N is the number of samples in the training batch, H is the number of prediction time windows (i.e., the number of future time windows the model needs to output), M is the number of physiological parameter types to be predicted (i.e., the dimension of the multidimensional physiological parameter), i represents the sample index, h and k are the time window indices, and m is the physiological parameter index. This represents the predicted minimum value of the i-th sample, h-th time window, and m-th physiological parameter output by the model. Let m be the true minimum value of the i-th sample, the h-th time window, and the m-th type of physiological parameter. Let `a` be the predicted minimum value for the i-th sample, the (k+1)-th long time window, and the m-th type of physiological parameter. `max(a,0)` represents the maximum value function, retaining `a` only when `a>0`, otherwise outputting `0`, used to filter for false negative / false alarm errors. `ReLU(x)` is a linear rectified function, outputting `x` when `x>0` and `0` when `x≤0`, penalizing only logical contradictions. `P` represents the prototype matrix (prototype library), storing feature vectors of K typical clinical risk patterns. Let P be the transpose of P. Let be a K-order identity matrix, where K is the number of typical clinical risk patterns in the prototype library, d represents the dimension of the prototype features (consistent with the hidden state dimension of the continuous-time evolutionary encoder output), p is the row index of the prototype matrix, and q is the column index of the prototype matrix. The square of the Frobenius norm is used to quantify the overall deviation between matrices. , indicating that matrix P and The element in the p-th row and q-th column of the product represents the correlation between the p-th prototype and the q-th prototype. Represents the identity matrix The element in the p-th row and q-th column.

[0137] Taking 100 iterations as an example, in rounds 1-20, the initial learning rate is 1e-4, and the model quickly learns the basic prediction patterns. asym The primary focus is on loss optimization, prioritizing the establishment of safety-oriented prediction logic; in rounds 21-60, the learning rate decays to 2.5e-5, L mono and L proto_reg The role of the model gradually becomes apparent. While ensuring safety, the model optimizes the logical consistency of multiple time windows and the prototype discrimination. In rounds 61-100, the learning rate decays to 6.25e-6, the model fine-tunes its parameters, and the validation set loss continues to converge. Finally, when the validation set loss is less than 0.5 and does not decrease for 10 consecutive rounds, training stops.

[0138] After completing iterative training, save the network parameters (weights, biases, prototype matrix, γ, β, etc.) from the last round of training to obtain the trained asymmetric neuromorphic interactive network model.

[0139] The asymmetric neural frequent differential interaction network model constructed in this application features a structured four-tiered sub-module architecture, ensuring functional synergy among the modules. Specifically, a variable selection layer filters noise, a continuous-time evolutionary encoder adapts to heterogeneous frequency data, a prototype semantic interaction layer enhances interpretability, and an asymmetric secure gating decoder ensures security, forming a complete closed loop of "data processing - feature encoding - semantic matching - secure output." In the composite loss function, L... asym The loss term differentiates between false alarm and missed detection errors, forcing the model to learn the logic of preferring false alarms to false alarms during training; L mono The loss term penalizes logical contradictions where long-term window predictions are higher than short-term window predictions, forcing the model to output a monotonically non-increasing predicted minimum, thus completely resolving the issue of lower risk for long-term windows compared to short-term windows in the background technique; during iterative training, L mono With L asym Collaborative optimization ensures that logical consistency is not sacrificed for security, thus improving the logical consistency of the final model's multi-time-window predictions; L proto_reg Orthogonal constraints ensure the distinctiveness of typical risk patterns in the prototype library (such as precipitous drops and volatile deterioration), enabling the model to accurately identify the risk pattern corresponding to the current user being tested. The three loss terms of the composite loss function complement and balance each other, ensuring safety, logical consistency, and interpretability without sacrificing prediction accuracy.

[0140] In one embodiment, after outputting the final predicted value of the preset multidimensional physiological parameters, the multidimensional physiological parameter prediction method further includes: obtaining an individualized threshold that matches the user being tested; and determining whether to issue an alarm based on the individualized threshold and the final predicted value.

[0141] In this embodiment, the individualized threshold refers to a physiological parameter risk assessment standard customized based on the individual static characteristics of the tested user (such as age, underlying diseases, medical history, etc.). This differs from a uniform clinical threshold and is more closely aligned with the physiological differences and health conditions of different users. Static characteristics refer to the relatively stable personal information and health indicators of the tested user during the monitoring period, which do not change frequently over time. These are the core basis for generating the individualized threshold and include demographic characteristics, basic health status, and past medical history. The individualized threshold may not be the same for different types of tested users.

[0142] Specifically, key static features of the tested user can be collected; the static features can be standardized and transformed into computable feature vectors; and then individualized thresholds of various physiological parameters matching the tested user under different prediction time windows can be obtained through rule base matching or related prediction models, forming a threshold set.

[0143] The electronic device pre-constructs a clinical rule base and establishes mapping rules between different static features and corresponding thresholds. Based on the obtained static features, the corresponding individualized threshold can be obtained according to the mapping rules. The prediction model can typically be any suitable model such as a feature-boosting tree model, a random forest model, or a support vector regression model. Based on the obtained static features, they can be used as input to the prediction model, and the corresponding individualized threshold can be output through the calculation of the prediction model.

[0144] The electronic device compares the final predicted values ​​of each physiological parameter in different time windows with the corresponding individualized thresholds, and determines whether to trigger an alarm based on the comparison results, thus ensuring the accuracy and timeliness of the alarm signal.

[0145] Based on preset alarm judgment rules (including single time window rules and multi-time window joint rules), the final predicted value is compared with the corresponding individualized threshold one by one. If any alarm rule is met, it is marked as an alarm candidate. Finally, according to the alarm level classification standard, the alarm level is determined and the corresponding warning signal (such as sound alarm, visual alarm, SMS notification, etc.) is triggered.

[0146] The judgment rules may include any one or more of the following:

[0147] Rule 1: Single time window threshold penetration rule, that is, for any prediction time window h, if the final predicted value of a certain physiological parameter is lower than the lower limit individualized threshold (such as MAP, SpO2) or higher than the upper limit individualized threshold (such as HR, RR), then a single time window alarm for that parameter is triggered.

[0148] Rule 2: Multi-time window continuous penetration rule, that is, if the final predicted value of the same physiological parameter penetrates the corresponding individualized threshold in two or more consecutive time windows, an upgraded alarm is triggered (the alarm level is increased by one level).

[0149] Rule 3: Multi-parameter joint penetration rule, that is, if the final predicted values ​​of two or more core physiological parameters (such as MAP+HR, SpO2+RR) in the same time window all penetrate the threshold, a joint alarm (the highest level alarm) is triggered.

[0150] This application utilizes personalized threshold customization, which is generated based on the static characteristics of the tested user. It fully considers individual differences such as age, underlying diseases, and medical history, avoiding the drawbacks of general thresholds being too lenient for high-risk users and too strict for healthy users. In addition, the alarm rules can cover emergency situations with short-term abnormalities of single parameters, as well as high-risk scenarios with multiple parameters and long-term cumulative abnormalities, so that different levels of risk can be responded to accordingly.

[0151] In one embodiment, a computer-readable storage medium is provided having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the steps in the above method embodiments.

[0152] In one embodiment, an electronic device is also provided, including one or more processors; and a memory storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the steps in the above method embodiments. The electronic device may be a device that deploys the aforementioned large language model and / or knowledge graph, such as a backend server for a platform that provides knowledge-based question answering for enterprise personnel.

[0153] In one embodiment, such as Figure 2 The diagram illustrates the structure of an electronic device used to implement an embodiment of this application. The electronic device includes a central processing unit (CPU) 201, which can perform various appropriate actions and processes based on a program stored in a read-only memory (ROM) 202 or a program loaded from a storage portion 208 into a random access memory (RAM) 203. The RAM 203 also stores various programs and data required for the operation of the electronic device. The CPU 201, ROM 202, and RAM 203 are interconnected via a bus 204. An input / output (I / O) interface 205 is also connected to the bus 204.

[0154] The following components are connected to I / O interface 205: an input section 206 including a keyboard, mouse, etc.; an output section 207 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 208 including a hard disk, etc.; and a communication section 209 including a network interface card such as a LAN card, modem, etc. The communication section 209 performs communication processing via a network such as the Internet. Drive 210 is also connected to I / O interface 205 as needed. Removable media 211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 210 as needed so that computer programs read from them can be installed into storage section 208 as needed.

[0155] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer-readable medium carrying instructions that, in such embodiments, can be downloaded and installed from a network via communication section 209, and / or installed from removable medium 211. When the instructions are executed by central processing unit (CPU) 201, the various method steps described in this application are performed.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0157] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for predicting multidimensional physiological parameters, characterized in that, include: Obtain the data stream of the user under test {(x i , t i )},t i For the actual arrival timestamp, x i To represent the observation vector of the i-th dimension, the multidimensional physiological parameters include at least one of mean arterial pressure, heart rate, blood oxygen saturation, and respiratory rate; Based on the data stream, a corresponding high-dimensional input vector is generated. The high-dimensional input vector includes the most recent observation value at each observation time, data freshness and data missing indication values, trend feature data and parameter linkage feature data corresponding to the most recent observation value, and the trend feature data includes the original observation value, the first-order trend and the second-order trend. The first-order trend is the rate of change of the original observation value calculated by the sliding time window, and the second-order trend is the rate of change of the first-order trend. The high-dimensional input vector is input into a pre-constructed asymmetric neuromorphic interaction network model. After processing by the cascaded variable selection layer, continuous-time evolution encoder, prototype semantic interaction layer and asymmetric security gated decoder in the asymmetric neuromorphic interaction network model, the final predicted value of the preset multidimensional physiological parameters is output. The processing steps of the asymmetric security-gated decoder include: performing regression calculations on the semantic enhancement features output by the prototype semantic interaction layer to obtain a basic predicted value; correcting the basic predicted value based on a negative security bias term controlled by a risk coefficient; and taking the predicted minimum value of the multidimensional physiological parameters as the final predicted value. The negative security bias term is generated by a preset nonlinear function, the output of which is always positive. After sign inversion, the negative security bias term is formed, so that the security correction can only lower the basic predicted value.

2. The method according to claim 1, characterized in that, The processing of the variable selection layer includes: dynamically assigning weights and filtering noise on the high-dimensional input vector based on a gated residual network to form weighted features of multidimensional physiological parameters; The processing of the continuous-time evolution encoder includes: using a Jump-Flow mechanism to adapt to irregular time intervals, instantly updating the hidden state of the network through a time-series update unit when the observed data arrives, and simulating the continuous evolution of the hidden state through a differential equation network during the observation blank period to obtain the final hidden state. The processing steps of the prototype semantic interaction layer include: calculating the similarity distribution between the final hidden state and the preset prototype library through an attention mechanism, generating semantic enhancement features by combining the similarity distribution, and mapping the corresponding risk coefficients.

3. The method according to claim 2, characterized in that, The method of dynamically assigning weights and filtering noise from the high-dimensional input vector based on a gated residual network to form weighted features of multidimensional physiological parameters includes: performing feature transformation on the high-dimensional input vector, generating gated weights for feature channels adapted to multidimensional physiological parameters in parallel, and suppressing feature channels with severe missing or high noise levels by weighted fusion of the gated weights and the transformed features to achieve noise filtering and form weighted features of multidimensional physiological parameters. The time-series update unit is a recurrent neural network unit, and the differential equation network is a neural ordinary differential equation network. The continuous evolution results of the hidden state during the observation blank time period are integrated by a numerical solver. The similarity distribution is obtained by calculating the correlation between the final hidden state and the prototype matrix using the Softmax function; the semantic enhancement feature is generated by superimposing the weighted sum of the prototypes in the prototype matrix using the similarity distribution with the final hidden state.

4. The method according to claim 3, characterized in that, The feature transformation is achieved through linear layer mapping to obtain the mapped features; the gate weights are generated by calculating the Sigmoid function; the weighted fusion includes element-wise multiplication of the mapped features and the gate weights to generate the weighted features. The recurrent neural network unit is a gated recurrent GRU unit; the numerical solver is a Runge-Kutta solver or an Euler solver; the step of instantaneously updating the network hidden state through a time-series update unit at the arrival of observed data, and simulating the continuous evolution of the hidden state through a differential equation network during the observation gap period to obtain the final hidden state includes: instantaneously updating the network hidden state through a GRU unit combined with the weighted features at the arrival of observed data; simulating the continuous dynamic evolution of the hidden state through a neural network of differential equations during the observation gap period, and calculating the final hidden state through a numerical solver; The nonlinear function is a Softplus function; the correction process is to superimpose the base prediction value and the safety correction term, which is the product of the trainable scaling parameter, the risk coefficient, and the negative safety bias term.

5. The method according to claim 1, characterized in that, The method further includes: constructing a variable selection layer, a continuous-time evolutionary encoder, a prototype semantic interaction layer, and an asymmetric security gated decoder in a cascaded manner to form an asymmetric neural frequent differential interaction network to be trained; The pre-set training data is input into the asymmetric neural frequent differential interaction network to be trained for iterative training. A composite loss function including asymmetric safe regression loss, multi-time window consistency constraint loss, and prototype orthogonal regularization term is used to train and optimize the asymmetric neural frequent differential interaction network to be trained, so as to obtain the trained asymmetric neural frequent differential interaction network model.

6. The method according to claim 5, characterized in that, The composite loss function is a weighted sum of asymmetric safe regression loss, multi-time-window consistency constraint loss, and prototype orthogonal regularization term.

7. The method according to claim 1, characterized in that, The parameter linkage feature data refers to the Pearson correlation coefficient or the proportion of changes in the same direction among different physiological parameters within the sliding time window.

8. The method according to any one of claims 1 to 7, characterized in that, After outputting the final predicted values ​​of the preset multidimensional physiological parameters, the following is also included: Obtain individualized thresholds that match the tested user; Whether to issue an alarm is determined based on the individualized threshold and the final predicted value.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores executable instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1 to 8.

10. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to perform the method as described in any one of claims 1 to 8.

Images