Intensive care state real-time prediction method and system based on mask dynamic composition

Abstract

The invention discloses an intensive care state real-time prediction method and system based on mask dynamic composition. The method directly faces the challenge that clinical time sequence data is extremely sparse, and a monitoring missing mode is creatively converted into valuable prediction features. A monitoring mode dynamic relation graph is constructed to encode the clinical cooperative monitoring intensity between the indexes; comprehensively modeling time sequence dependence in combination with a time sequence attention mechanism and multi-level local feature extraction; and finally, through adaptive multi-channel fusion, accurate and end-to-end prediction of the future multi-index clinical trajectory of the patient is realized. Compared with a traditional filling or special architecture method, the method has the advantages that knowledge hidden in clinical monitoring behaviors can be more effectively mined, more excellent prediction performance can be achieved on an intensive care state data set with extremely sparse data, and a new technical path is provided for development of an intelligent clinical early warning system.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of smart healthcare and clinical artificial intelligence, and particularly to a method and system for real-time prediction of intensive care status based on mask dynamic mapping. Addressing the widespread problem of extreme data sparsity in electronic medical records caused by selective clinical monitoring, this invention provides an end-to-end solution capable of high-precision prediction directly utilizing monitoring behavior patterns. Background Technology

[0002] In the intensive care unit, accurately predicting the future trajectory of patients' vital signs and laboratory indicators is crucial for early warning and precise intervention. However, clinical monitoring data has two inherent characteristics: irregular sampling and extreme sparsity. Measurement behavior depends on the patient's real-time condition and clinical judgment, resulting in non-uniform distribution of data over time; at the same time, many laboratory tests are only developed when a specific pathology is suspected, leading to extremely high missing data points. This "non-random missing" characteristic poses a significant challenge to traditional time series forecasting methods.

[0003] Current mainstream methods can be broadly categorized into two types. The first is the "fill-in-the-gap" paradigm, such as mean imputation, forward imputation, or RNN-based interpolation models (e.g., BRITS). These methods introduce errors during the imputation phase and completely discard any clinical intent information that the missing pattern itself might contain. The second is the "dedicated architecture" paradigm, such as GRU-D, which models the process of information decaying over time through learnable decay mechanisms, or neural ODEs, which model sequences as continuous dynamic systems. While these methods can handle irregular intervals, their design intent is to treat missing data as "defects" requiring compensation, rather than interpretable "signals."

[0004] Therefore, existing technologies have failed to fully explore and utilize the rich domain knowledge encoded in clinical monitoring behaviors, resulting in performance bottlenecks in predictive scenarios with extremely sparse data in intensive care settings. A novel framework capable of directly transforming monitoring patterns into predictive features is urgently needed. Summary of the Invention

[0005] This invention provides a method and system for real-time prediction of intensive care status based on mask dynamic graphing, which can solve the problem that existing technologies cannot effectively utilize clinical monitoring behavioral information, resulting in poor prediction performance of sparse data.

[0006] The first aspect of this invention provides a method for real-time prediction of intensive care status based on mask dynamic mapping, comprising:

[0007] Multivariate time-series monitoring data of patients in the intensive care unit were collected, and a sparse input containing a numerical matrix and a monitoring identifier matrix was constructed. The numerical matrix was filled with missing values ​​and standardized. The processed numerical matrix was then combined with trainable indicator feature codes to generate a high-dimensional representation vector that integrates time-series location information. Global temporal dependency encoding is performed on the high-dimensional representation vector to extract the global temporal dependency features of each indicator; Using the monitoring identifier matrix as the core input, the monitoring frequency of each clinical indicator is calculated; a dynamic, parameter-free relational graph adjacency matrix is ​​constructed based on the monitoring frequency between indicators to quantify the collaborative monitoring strength between indicators; the relational graph adjacency matrix is ​​used to perform multi-order feature diffusion on the initial representation to extract the correlation features between indicators. The standardized time series data is segmented into overlapping local windows; using a parallel structure composed of independent channel convolution and encoder / decoder, multi-level feature extraction and dimensionality reduction are performed on the local windows to capture multi-granular time series patterns from instantaneous changes to stage trends, and summarized into local pattern representations for each indicator. Global temporal dependency features, inter-indicator correlation features, and local pattern representations are integrated through a trainable adaptive weighted fusion gate; the fused representations are input into a prediction network, which outputs the predicted sequence of all clinical indicators within a preset future time period.

[0008] Optionally, constructing a sparse input comprising a numerical matrix and a monitoring identifier matrix includes: Define sparse time-series data containing N clinical indicators, with an input observation window length of L, a numerical matrix denoted as S, and a corresponding binary monitoring identifier matrix denoted as O. In the monitoring identifier matrix, an element of 1 indicates that the corresponding indicator has a monitoring record at the corresponding time, and an element of 0 indicates that there is no monitoring record. The original numerical matrix is ​​filled with gaps, and the filled numerical matrix is ​​then subjected to min-max normalization. The standardized scalar values ​​are converted into numerical embeddings through nonlinear mapping. The numerical embeddings are added to the trainable indicator identity codes and then subjected to layer standardization to generate basic time series representations. The sinusoidal positional encoding is superimposed on the basic temporal representation of each index to obtain the final input representation tensor.

[0009] Optionally, based on multi-order feature diffusion of the initial representation using the adjacency matrix of the relational graph, the correlation features between indicators are extracted, including: The total number of times each indicator is monitored within the entire observation window is calculated based on the monitoring identifier matrix. The total number of monitoring times reflects the clinical attention density to the indicator. An adjacency matrix is ​​constructed using the total number of monitoring sessions to visualize the relationships between indicators, and the numerical values ​​of the correlation strength between indicators are calculated. The adjacency matrix is ​​row normalized to ensure the stability of subsequent feature diffusion.

[0010] Optionally, multi-order feature diffusion of the initial representation using the adjacency matrix of the relational graph includes: The initial node feature tensor obtained by the linear transformation of the input representation is used to propagate K-order features using the normalized adjacency matrix; Each propagation step multiplies the normalized adjacency matrix with the current-order feature tensor to obtain the next-order feature tensor. The final-order feature tensor is used as the index representation of the graph-enhanced relation graph, which integrates the correlation information between indicators based on the monitoring frequency.

[0011] Optionally, multi-level feature extraction and dimensionality reduction of the local window can be performed using a parallel structure composed of independent channel convolutions and encoders / decoders, including: The standardized input is divided into a series of overlapping local windows; Each window is fed into multiple parallel processing channels, and each processing channel first processes each metric using independent one-dimensional convolution to extract local patterns. Each processing channel then uses a lightweight symmetric codec to compress the convolutional output, learning a compact representation; The outputs of all processing channels are averaged to obtain a window-level representation, and the window-level representation is mapped to the local temporal pattern features of each index through inverse window operation and various summarization strategies.

[0012] Alternatively, integration via a trainable adaptive weighted fusion gate includes: For each metric, the temporal dependency features, inter-metric correlation features, and local pattern features are calculated using independently learned fusion gate weights. The fusion gate weights are obtained by multiplying the learnable parameters with the corresponding features and then activating them using the sigmoid function; The features of the three channels are multiplied element-wise with their respective fusion gate weights and then added together to obtain the fused representation. If external temporal context information is provided, the fused representation is modulated using a context modulation layer to generate a context-aware representation.

[0013] Optionally, the output of predicted sequences for all clinical indicators within a preset future time period includes: The fused representation is input into a two-layer feedforward prediction network, which directly outputs a matrix of predicted values ​​for all N indicators within the next H time points. The objective function is defined as calculating the mean square error only at locations with actual monitoring records during the future prediction period; The objective function ensures that model optimization is based only on monitoring points where true values ​​are available, while inferring the complete prediction space through a representation learning mechanism.

[0014] A second aspect of the present invention provides a real-time prediction system for intensive care status based on mask dynamic graphing, comprising: The data acquisition module is used to collect multivariate time-series monitoring data of patients in the intensive care unit and construct a sparse input containing a numerical matrix and a monitoring identifier matrix. The representation generation module is used to fill missing values ​​and standardize the numerical matrix, and combine the processed numerical matrix with trainable index feature encoding to generate a high-dimensional representation vector that integrates temporal location information. The temporal coding module is used to perform global temporal dependency coding on high-dimensional representation vectors and extract the global temporal dependency features of each indicator. The relational graph construction module is used to calculate the monitoring frequency of each clinical indicator with the monitoring identifier matrix as the core input, and to construct a dynamic relational graph adjacency matrix without training parameters based on the monitoring frequency between indicators, so as to quantify the collaborative monitoring strength between indicators. The feature diffusion module is used to perform multi-stage feature diffusion on the initial representation using the adjacency matrix of the relation graph, and to extract the correlation features between indicators. The local pattern extraction module is used to divide the standardized time series data into overlapping local windows. It uses a parallel structure composed of independent channel convolution and encoder / decoder to perform multi-level feature extraction and dimensionality reduction on the local windows, capture multi-granular time series patterns, and summarize the local pattern representation of each indicator. The adaptive fusion module is used to integrate global temporal dependency features, inter-indicator correlation features, and local pattern representations through a trainable adaptive weighted fusion gate. The prediction output module is used to input the fused representations into the prediction network and directly output the prediction sequence of all clinical indicators within a preset time period.

[0015] A third aspect of the present invention provides a real-time prediction device for intensive care status based on mask dynamic mapping, comprising: One or more processors; A memory on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the real-time prediction method for intensive care status based on dynamic mapping of masks as described above.

[0016] A fourth aspect of the present invention provides a computer storage medium for storing a program, which, when executed, is used to implement the real-time prediction method for intensive care status based on dynamic masking as described in any of the preceding claims.

[0017] Compared with the prior art, the present invention has the following advantages: This invention is the first to use a clinical monitoring identifier matrix as the core modeling object. By constructing a dynamic relationship graph through monitoring frequency, the "monitoring pattern" is transformed into valuable features, and the diagnosis and treatment logic and patient status correlation hidden in clinical behavior are discovered.

[0018] This invention employs a parameter-free dynamic graph construction method, which generates an adjacency matrix based solely on the monitoring frequency of each patient. This avoids the risk of overfitting when learning graph structures on extremely sparse data, and is computationally efficient and robust.

[0019] This invention constructs a multi-channel adaptive fusion architecture that integrates global temporal dependency, dynamic relationship graph diffusion, and local pattern capture. Through intelligent integration via learnable fusion gates, it comprehensively characterizes the multi-scale features of time series data, significantly improving the accuracy of multi-indicator collaborative prediction. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for real-time prediction of intensive care status based on dynamic masking graphs, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of an embodiment of the intensive care status monitoring and prediction analysis system provided in this invention. Figure 3 A schematic diagram of the structure of a real-time prediction system for critical care status based on dynamic masking graphs provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention. Detailed Implementation

[0022] This invention provides a method and system for real-time prediction of intensive care status based on mask dynamic graphing, which can solve the problem of poor prediction performance of sparse data caused by the inability of existing technologies to effectively utilize clinical monitoring behavioral information. This invention proposes an end-to-end prediction framework called MPNet. The core idea of ​​this invention is to perform a "perspective shift": redefining the "missing patterns" of clinical monitoring as "information patterns" that reflect the diagnostic and treatment approach and the patient's condition.

[0023] The technical solution of this invention is as follows: First, sparse clinical time-series data is standardized and jointly embedded to construct a unified input. The core innovation lies in the introduction of a dynamic relationship graph module for monitoring patterns. This module parses the unique monitoring identifier matrix of each patient, dynamically constructs a graph structure reflecting the correlation of monitoring intensity between indicators, and encodes this clinical behavior pattern into the indicator representation through a feature diffusion process. Simultaneously, a temporal attention encoder is used to capture global temporal dependence, and a multi-level local convolution module is designed to extract local morphological features at different time scales. Finally, an adaptive multi-channel fusion gate dynamically integrates information from the above three parallel channels, directly outputting a prediction of future multi-indicator clinical sequences.

[0024] This invention eliminates the need for explicit data imputation and directly operates on the original sparse observations. By actively learning and utilizing the hidden patterns in the monitored behavior, it significantly improves the accuracy and reliability of predicting the clinical trajectory of patients in intensive care.

[0025] See Figure 1 This figure is a flowchart illustrating a real-time prediction method for critical care status based on mask dynamic graphing provided by an embodiment of the present invention. It shows the end-to-end processing flow of the MPNet framework. First, the input sparse clinical time-series numerical matrix and monitoring identifier matrix are filled with gaps, standardized, and embedded to generate a high-dimensional representation tensor that integrates numerical values, indicator attributes, and temporal location information. Then, long-range temporal dependency features, inter-indicator collaborative monitoring correlation features, and multi-scale local fluctuation features are extracted through three parallel channels: a global temporal Transformer encoder, a parameter-free monitoring mode dynamic relationship graph, and a multi-level local convolutional encoder-decoder. Subsequently, a trainable adaptive fusion gate is used to assign weights to the features of the three channels and perform weighted fusion. Finally, the input is fed into the prediction network to output the future multi-indicator sequence. The loss function only calculates the mean square error of the actual monitoring location to ensure the effectiveness of the optimization.

[0026] The real-time prediction method for critical care status based on dynamic mapping of masks (referred to as MPNet) provided in this embodiment of the invention can be implemented, for example, through the following steps S101-108.

[0027] S101: Collect multivariate time-series monitoring data of patients in the intensive care unit and construct a sparse input containing a numerical matrix and a monitoring label matrix.

[0028] In this embodiment of the invention, sparse time-series data containing N clinical indicators is defined, the input observation window length is L, the numerical matrix is ​​denoted as S, and the corresponding binary monitoring identifier matrix is ​​denoted as O. An element of 1 in the monitoring identifier matrix indicates that the corresponding indicator has a monitoring record at the corresponding time, and an element of 0 indicates that there is no monitoring record. The original numerical matrix is ​​filled with gaps, and the filled numerical matrix is ​​then subjected to min-max normalization. The normalized scalar values ​​are converted into numerical embeddings through nonlinear mapping. The numerical embeddings are added to the trainable indicator identity codes and then subjected to layer normalization to generate basic time-series representations. A sinusoidal position code is superimposed on the basic time-series representation of each indicator to obtain the final input representation tensor.

[0029] Specifically, multivariate clinical time-series data of patients are collected from bedside monitors, laboratory information systems, and electronic medical record systems within the monitoring window of the intensive care unit. Each data record includes the following: (1) Given a containing Sparse time-series data of 1 clinical indicator, with an input observation window length of 1 Numerical matrix (Represents an L x N real number matrix). Records Several clinical indicators (such as heart rate, systolic blood pressure, blood lactate, etc.) in continuous A point in time (e.g., one point per hour), The measured value is recorded on the ), and the unmeasured position is recorded as an invalid value (such as NaN).

[0030] (2) Binary monitoring and identification matrix (Represents an L x N matrix with values ​​of 0 or 1): and Isomorphic binary matrices, when When there is a valid measurement at the corresponding position, it is recorded as Indicators At any moment If there are monitoring records, then the value is 0; otherwise, the value is 0. Indexed by time point; For indexing indicators.

[0031] (3) Optional time series features: such as hour, day of the week, whether it is night, etc., can form a time series feature matrix. ,in Dimensions are optional features.

[0032] S102: Fill in missing values ​​and standardize the numerical matrix, and combine the processed numerical matrix with trainable index feature encoding to generate a high-dimensional representation vector that integrates temporal location information.

[0033] In this embodiment of the invention, the following preprocessing steps are performed on each patient sample: Step 1: Filling gaps, filling the numerical matrix Invalid values ​​in the result are filled with 0, resulting in ,in This represents matrix multiplication.

[0034] Step 2: For each clinical indicator, calculate its minimum and maximum values ​​based on the training set, and use min-max normalization to... The values ​​in the linear mapping to Interval.

[0035] Step 3: Dataset partitioning: Ensure that each patient appears only in one of the training set, validation set, or test set to prevent data leakage.

[0036] For the padded and standardized matrix Embedded encoding is performed: (1) Numerical embedding: for each position scalar value Generate through learnable linear transformation plus nonlinear activation dimensional vector : ; in, and These are the learnable weight matrix and bias vector, respectively; The hyperbolic tangent function maps the input to the interval [-1, 1].

[0037] (2) Indicator identity embedding: for each clinical indicator Assign a learnable one 3D embedding vector .

[0038] (3) Combination and Standardization: The numerical embedding and the indicator identity embedding are added together, and layer standardization is performed: ; in, Standardized operations at the presentation layer.

[0039] (4) Location coding: Time series of each indicator (This represents stacking the vectors from L time points into a matrix), superimposed with sinusoidal position codes. This yields the final input representation. ,in, , For the representation dimension.

[0040] S103: Perform global temporal dependency encoding on the high-dimensional representation vector to extract the global temporal dependency features of each indicator.

[0041] In embodiments of the present invention, such as Figure 1 As shown, Channel 1 is a global time-dependent encoding channel. This channel is based on the Transformer encoder structure and independently processes the time series sequence of each metric. Transformer encoding layer: will Enter one Layer Transformer encoder (e.g.) Each layer contains a multi-head self-attention mechanism and a feedforward network, and employs residual connections and layer normalization.

[0042] Temporal aggregation: sequences output by the Transformer Global average pooling is performed along the time dimension to obtain the global time-series features of each indicator. ,in, Indicates the time sequence.

[0043] S104: Using the monitoring identifier matrix as the core input, calculate the monitoring frequency of each clinical indicator, and construct a dynamic adjacency matrix of the relationship graph without training parameters based on the monitoring frequency between indicators to quantify the collaborative monitoring intensity between indicators.

[0044] In this embodiment of the invention, the total number of times each indicator is monitored within the entire observation window is calculated based on the monitoring identifier matrix. The total number of monitoring times reflects the clinical attention density to the indicator. An adjacency matrix of the relationship graph between indicators is constructed using the total number of monitoring times, and the correlation strength between indicators is calculated. The adjacency matrix is ​​row normalized to ensure the stability of subsequent feature diffusion.

[0045] Specifically, such as Figure 1 As shown, Channel 2 is the Dynamic Relationship Graph of Monitoring Modes (MAGraph). The core of this channel is to dynamically construct the graph structure from the monitoring identifier matrix without training parameters. First, the monitoring frequency is calculated based on the monitoring identifier matrix. Calculate each indicator Total number of monitoring times within the observation window: ; in, For time point indexes within the observation window; monitoring frequency This reflects the level of attention paid to this indicator.

[0046] Then, an adjacency matrix is ​​constructed, and the correlation strength between indicators is calculated based on the monitoring frequency to form the adjacency matrix. Its elements Indicates from indicators To the target The correlation strength number is calculated using the following formula: ; in, and Sub-tables represent indicators and Total number of monitoring sessions; Indicates all indicators The maximum value of the product of the number of monitoring sessions; It is a very small positive constant used to maintain numerical stability and prevent division by zero. Then... Perform row normalization to obtain This is to ensure the stability of feature diffusion.

[0047] S105: Use the adjacency matrix of the relation graph to perform multi-level feature diffusion on the initial representation and extract the correlation features between indicators.

[0048] In this embodiment of the invention, the initial node feature tensor obtained by linear transformation of the input representation is used to perform K-order feature propagation using a normalized adjacency matrix; each propagation multiplies the normalized adjacency matrix with the current-order feature tensor to obtain the next-order feature tensor; the final-order feature tensor is used as the indicator representation of the relationship graph enhancement, which integrates the correlation information between indicators based on the monitoring frequency.

[0049] Specifically, feature diffusion is the process of representing the input. Initial node features are obtained through linear transformation. (express Each indicator (Dimensional features). Perform Step (such as) Image dissemination: , Ultimately, the indicator features enhanced by the relationship graph are obtained. ,in, Indicates the first Post-propagation index eigenvectors, It represents the relationship; this representation integrates information between indicators based on monitoring frequency.

[0050] S106: The standardized time series data is divided into overlapping local windows. The local windows are then subjected to multi-level feature extraction and dimensionality reduction using a parallel structure composed of independent channel convolution and encoder / decoder. This process captures multi-granular temporal patterns and summarizes them into local pattern representations for each indicator.

[0051] In this embodiment of the invention, the standardized input is divided into a series of overlapping local windows; each window is fed into multiple parallel processing channels, and each processing channel first processes each index using independent one-dimensional convolution to extract local patterns; each processing channel then uses a lightweight symmetric codec to compress the convolution output to learn compact representations; the outputs of all processing channels are averaged to obtain window-level representations, and the window-level representations are mapped to the local temporal pattern features of each index through inverse window operations and multiple summarization strategies.

[0052] Specifically, such as Figure 1 As shown, channel three is a multi-level local pattern capture channel, which is designed to extract fine-grained short-term temporal patterns.

[0053] First, overlapping local windows are divided, and the standardized input is then processed. Divide the space along the time dimension into a series of overlapping local windows. Let the window length be... (like ), sliding step size (like ),get window ,in For window indexing.

[0054] Then perform multi-channel independent convolution, using One (e.g.) Parallel depthwise separable one-dimensional convolutional layers independently process each window, extracting local features: ; in, This represents the activation function of the linear rectifier unit; This represents a one-dimensional convolution operation performed independently on each index channel.

[0055] Next, symmetric encoder-decoder compression is performed: each convolutional output is compressed through a lightweight symmetric autoencoder to learn compact representations. : , in, , These are the weight matrix and bias vector of the encoder, respectively; , These are the weight matrix and bias vector of the decoder, respectively; It is the hyperbolic tangent activation function.

[0056] Finally, the features are summarized. The window representation is obtained by averaging the outputs of each channel. The time series was reconstructed using inverse windowing, and local features of each indicator were aggregated using various strategies such as mean pooling and max pooling. ,in Indicates local mode.

[0057] S107: Integrates global temporal dependency features, inter-indicator correlation features, and local pattern representations through a trainable adaptive weighted fusion gate.

[0058] In this embodiment of the invention, for each index, the temporal dependency features, inter-index correlation features, and local pattern features are calculated using independently learned fusion gate weights. The fusion gate weights are obtained by multiplying the learnable parameters with the corresponding features and then activating them using the sigmoid function. The features of the three channels are multiplied element-wise with the corresponding fusion gate weights and then added together to obtain the fused representation. If external temporal context information is provided, the context modulation layer is used to adjust the fused representation to generate a context-aware representation.

[0059] Specifically, adaptive multi-channel fusion, with three channels for each indicator. Output respectively , , Dynamic integration is achieved through independently learnable fusion gates. First, the gate weights are calculated using the following formula: ; in, It is the sigmoid function; , These are learnable parameters; Indicate the feature type, including global time series, relationship graph, and local; Indicators Feature vectors under a certain feature channel (including global time series, relational graph and local).

[0060] Then, a weighted fusion is performed, using the following formula: ; in, This represents element-wise multiplication; Indicators Global temporal features The corresponding fusion gate weights; Indicators Association features The corresponding fusion gate weights; Indicators Local pattern features The corresponding fusion gate weights.

[0061] If external time series features are provided (e.g., hourly, daily cycles) then use a feature linear modulation layer to adjust the fused features to generate a context-aware representation. : ; in , From a small feedforward network It is generated in the middle and used to scale and offset the fused features.

[0062] S108: Input the fused representations into the prediction network and directly output the predicted sequence of all clinical indicators within a preset future time period.

[0063] In this embodiment of the invention, the fused representation is input into a two-layer feedforward prediction network, which directly outputs a matrix of predicted values ​​for all N indicators within the next H time points. The objective function is defined as calculating the mean square error only at locations with real monitoring records during the future prediction period. The objective function ensures that model optimization is based only on monitoring points where real values ​​can be obtained, while inferring the complete prediction space through a representation learning mechanism.

[0064] Specifically, the modulated features are input into a two-layer feedforward prediction network (the middle layer uses GELU activation), which directly outputs the future... a point in time (e.g.) )all Predicted value matrix of each indicator .

[0065] In one implementation of this invention, model training includes the following: Loss function: The masked mean squared error loss is used, and the error is calculated only at locations with actual measurements within the prediction period. ; in, It is a monitoring and identification matrix for future time periods. These are actual monitoring values. It represents the predicted value. This loss ensures that model optimization is based only on monitoring points where true values ​​are available, while simultaneously enabling inference of the complete prediction space through a representation learning mechanism.

[0066] Optimizer: Use the Adam optimizer, with an initial learning rate set to... Batch size is ,Every After each training round, multiply the learning rate by the decay factor. .

[0067] Early stopping strategy: Monitor loss on the validation set; if continuous... The relative decrease in round-verification loss is insufficient. If so, training should be stopped early.

[0068] After training, the model can be deployed in real-world clinical settings (such as...). Figure 2 As shown): Real-time data stream access; the system continuously receives real-time data from ICU bedside devices and automatically constructs past... Hourly numerical matrix and monitoring identifier matrix. Online prediction: Input the data into the MPNet model, perform forward propagation, and obtain future... The system predicts the trajectory of each indicator for each hour. Based on the predicted trajectory, it calculates risk scores for key indicators (e.g., an increasing lactate level indicates a risk of insufficient tissue perfusion). When these scores exceed preset thresholds, it sends warning messages to healthcare workers and visualizes the predicted trend compared to historical data.

[0069] Figure 2 This is a schematic diagram of an embodiment of the intensive care unit (ICU) status monitoring and prediction analysis system, demonstrating the practical application scenario of the MPNet framework. It implements a complete process from clinical data input to early warning decision output, with the specific steps as follows: (a) Data Input Layer: Connects to real-time monitoring equipment in the ICU, collects sparse time-series data including multiple indicators such as heart rate, blood pressure, and blood oxygen analysis, and automatically generates a numerical matrix and a monitoring identifier matrix. (b) Core Processing Layer: Deployment... Figure 1 The MPNet model performs embedding, multi-channel feature extraction, and adaptive fusion operations on the input data, outputting future trend curves for each indicator. (c) Decision output layer: Calculates the abnormal risk level of indicators based on the prediction curves. For example, when the predicted lactate level continues to rise, it is marked as "high risk of septic shock". It pushes early warning information on disease deterioration and key indicator trends to clinicians to assist in timely intervention.

[0070] Beneficial effects: a) The "monitoring model is information" mechanism uncovers the hidden value of clinical behavior.

[0071] Existing technologies such as GRU-D and BRITS either fill in missing data as noise or indirectly process time intervals through attenuation functions, both ignoring the rich information inherent in the decision-making process of "which indicators clinicians choose to monitor." This invention is the first to propose using a binary monitoring identifier matrix as the core modeling object. By constructing a dynamic relationship graph of monitoring patterns (MAGraph), it explicitly quantifies and utilizes the synergistic monitoring intensity between indicators, thereby capturing the implicit correlation between clinical diagnosis and treatment logic and patient status that traditional methods cannot reach.

[0072] b) A parameterless dynamic graph construction method to overcome the learning challenge in extremely sparse scenarios.

[0073] Faced with critical care status data exhibiting extremely high detection miss rates, traditional gradient-based graph neural networks struggle to learn stable and reliable indicator relationship structures. This invention proposes a dynamic relationship graph module for monitoring patterns, which generates an adjacency matrix dynamically based solely on the monitoring frequency of each patient instance, requiring no trainable parameters. This method is not only computationally efficient but also fundamentally avoids the risk of overfitting on sparse and noisy data, providing a highly robust solution for indicator relationship modeling and maintaining stable performance gains even on the sparsest test set.

[0074] c) Construct a multi-channel adaptive fusion architecture to achieve spatiotemporal dependency modeling.

[0075] This invention integrates three major information processing channels: global temporal dependency encoding, dynamic relationship graph diffusion, and multi-level local pattern capture, and intelligently integrates them through a learnable adaptive fusion gate. This architecture can simultaneously capture long-term temporal evolution patterns, clinically based indicator co-occurrence patterns, and fine-grained short-term fluctuation characteristics, providing comprehensive, multi-scale representation learning capabilities for complex and incomplete intensive care clinical time-series data, thereby achieving more reliable and accurate multi-indicator collaborative prediction.

[0076] Based on the methods provided in the above embodiments, this invention also provides a real-time prediction system for intensive care status based on dynamic masking graphs. The following describes the real-time prediction system for intensive care status based on dynamic masking graphs in conjunction with the accompanying drawings.

[0077] See Figure 3 The figure is a schematic diagram of the structure of a real-time prediction system for critical care status based on dynamic mapping of a mask, provided by an embodiment of the present invention.

[0078] The real-time prediction system 300 for critical care status based on mask dynamic graphing provided in this embodiment of the invention includes: a data acquisition module 301, a representation generation module 302, a temporal coding module 303, a relationship graph construction module 304, a feature diffusion module 305, a local pattern extraction module 306, an adaptive fusion module 307, and a prediction output module 308.

[0079] The data acquisition module 301 is used to collect multivariate time-series monitoring data of patients in the intensive care unit and construct a sparse input containing a numerical matrix and a monitoring identifier matrix. The representation generation module 302 is used to fill missing values ​​and standardize the numerical matrix, and combine the processed numerical matrix with trainable index feature encoding to generate a high-dimensional representation vector that integrates temporal location information. The temporal coding module 303 is used to perform global temporal dependency coding on the high-dimensional representation vector and extract the global temporal dependency features of each indicator. The relationship graph construction module 304 is used to calculate the monitoring frequency of each clinical indicator with the monitoring identifier matrix as the core input, and construct a dynamic relationship graph adjacency matrix without training parameters based on the monitoring frequency between indicators to quantify the collaborative monitoring intensity between indicators. The feature diffusion module 305 is used to perform multi-stage feature diffusion on the initial representation using the adjacency matrix of the relation graph to extract the correlation features between indicators. The local pattern extraction module 306 is used to divide the standardized time series data into overlapping local windows, and use a parallel structure composed of independent channel convolution and encoder-decoder to perform multi-level feature extraction and dimensionality reduction on the local windows, capture multi-granular time series patterns and summarize the local pattern representation of each indicator. The adaptive fusion module 307 is used to integrate global temporal dependency features, inter-indicator correlation features, and local pattern representations through a trainable adaptive weighted fusion gate. The prediction output module 308 is used to input the fused representations into the prediction network and directly output the prediction sequence of all clinical indicators within a preset time period in the future.

[0080] In one possible implementation, the data acquisition module 301 is specifically used for: Define sparse time-series data containing N clinical indicators, with an input observation window length of L, a numerical matrix denoted as S, and a corresponding binary monitoring identifier matrix denoted as O. In the monitoring identifier matrix, an element of 1 indicates that the corresponding indicator has a monitoring record at the corresponding time, and an element of 0 indicates that there is no monitoring record. The original numerical matrix is ​​filled with gaps, and the filled numerical matrix is ​​then subjected to min-max normalization. The standardized scalar values ​​are converted into numerical embeddings through nonlinear mapping. The numerical embeddings are added to the trainable indicator identity codes and then subjected to layer standardization to generate basic time series representations. The sinusoidal positional encoding is superimposed on the basic temporal representation of each index to obtain the final input representation tensor.

[0081] In one possible implementation, the relationship graph construction module 304 is specifically used for: The total number of times each indicator is monitored within the entire observation window is calculated based on the monitoring identifier matrix. The total number of monitoring times reflects the clinical attention density to the indicator. An adjacency matrix is ​​constructed using the total number of monitoring sessions to visualize the relationships between indicators, and the numerical values ​​of the correlation strength between indicators are calculated. The adjacency matrix is ​​row normalized to ensure the stability of subsequent feature diffusion.

[0082] In one possible implementation, the feature diffusion module 305 is specifically used for: The initial node feature tensor obtained by the linear transformation of the input representation is used to propagate K-order features using the normalized adjacency matrix; Each propagation step multiplies the normalized adjacency matrix with the current-order feature tensor to obtain the next-order feature tensor. The final-order feature tensor is used as the index representation of the graph-enhanced relation graph, which integrates the correlation information between indicators based on the monitoring frequency.

[0083] In one possible implementation, the local pattern extraction module 306 is specifically used for: The standardized input is divided into a series of overlapping local windows; Each window is fed into multiple parallel processing channels, and each processing channel first processes each metric using independent one-dimensional convolution to extract local patterns. Each processing channel then uses a lightweight symmetric codec to compress the convolutional output, learning a compact representation; The outputs of all processing channels are averaged to obtain a window-level representation, and the window-level representation is mapped to the local temporal pattern features of each index through inverse window operation and various summarization strategies.

[0084] In one possible implementation, the adaptive fusion module 307 is specifically used for: For each metric, the temporal dependency features, inter-metric correlation features, and local pattern features are calculated using independently learned fusion gate weights. The fusion gate weights are obtained by multiplying the learnable parameters with the corresponding features and then activating them using the sigmoid function; The features of the three channels are multiplied element-wise with their respective fusion gate weights and then added together to obtain the fused representation. If external temporal context information is provided, the fused representation is modulated using a context modulation layer to generate a context-aware representation.

[0085] In one possible implementation, the prediction output module 308 is specifically used for: The fused representation is input into a two-layer feedforward prediction network, which directly outputs a matrix of predicted values ​​for all N indicators within the next H time points. The objective function is defined as calculating the mean square error only at locations with actual monitoring records during the future prediction period; The objective function ensures that model optimization is based only on monitoring points where true values ​​are available, while inferring the complete prediction space through a representation learning mechanism.

[0086] Since the real-time prediction system for intensive care status based on dynamic masking is a system corresponding to the real-time prediction method for intensive care status based on dynamic masking provided in the above method embodiments, the specific implementation of each module of the real-time prediction system for intensive care status based on dynamic masking is based on the same concept as in the above method embodiments. Therefore, for the specific implementation of each module of the real-time prediction system for intensive care status based on dynamic masking, please refer to the description of the real-time prediction method for intensive care status based on dynamic masking in the above method embodiments, and it will not be repeated here.

[0087] This invention also provides a real-time prediction device for critical care status based on dynamic mask mapping, the device comprising: a processor and a memory; The memory is used to store instructions; The processor is used to execute the instructions in the memory to perform the real-time prediction method for intensive care status based on mask dynamic mapping mentioned in the above embodiments.

[0088] It should be noted that the hardware structure of the real-time prediction device for critical care status based on dynamic masking graphs provided in the embodiments of the present invention can be as follows: Figure 4 The structure shown, Figure 4 This is a schematic diagram of the structure of a device provided in an embodiment of the present invention.

[0089] Please see Figure 4 As shown, device 400 includes: processor 410, communication interface 420, and memory 430. The number of processors 410 in device 400 can be one or more. Figure 4 Taking a processor as an example, in this embodiment of the invention, the processor 410, communication interface 420, and memory 430 can be connected via a bus system or other means. Figure 4 Taking the connection between China and Israel via the bus system 440 as an example.

[0090] Processor 410 may be a central processing unit (CPU), a network processor (NP), or a combination of a CPU and an NP. Processor 410 may further include hardware chips. These hardware chips may be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), generic array logic (GAL), or any combination thereof.

[0091] The memory 430 may include volatile memory, such as random-access memory (RAM); the memory 430 may also include non-volatile memory, such as flash memory, hard disk drive (HDD) or solid-state drive (SSD); the memory 430 may also include a combination of the above types of memory.

[0092] Optionally, the memory 430 stores an operating system and programs, executable modules, or data structures, or subsets thereof, or extended sets thereof. The programs may include various operation instructions for implementing various operations. The operating system may include various system programs for implementing various basic services and handling hardware-based tasks. The processor 410 can read the programs in the memory 430 to implement the real-time prediction method for intensive care status based on mask dynamic mapping provided in this embodiment of the invention.

[0093] The bus system 440 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus system 440 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0094] This invention also provides a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the real-time prediction method for intensive care status based on dynamic masking graphs mentioned in the above embodiments.

[0095] This invention also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the real-time prediction method for intensive care status based on dynamic masking mentioned in the above embodiments.

[0096] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for real-time prediction of intensive care status based on mask dynamic graphing, characterized in that, The method includes: Multivariate time-series monitoring data of patients in the intensive care unit are collected to construct a sparse input containing a numerical matrix and a monitoring identifier matrix; the numerical matrix is ​​filled with missing values ​​and standardized, and the processed numerical matrix is ​​combined with trainable index feature encoding to generate a high-dimensional representation vector that integrates time-series location information. Global temporal dependency encoding is performed on the high-dimensional representation vector to extract the global temporal dependency features of each indicator; Using the monitoring identifier matrix as the core input, the monitoring frequency of each clinical indicator is calculated; a dynamic, parameter-free relational graph adjacency matrix is ​​constructed based on the monitoring frequency between indicators to quantify the collaborative monitoring intensity between indicators; the relational graph adjacency matrix is ​​used to perform multi-order feature diffusion on the initial representation to extract the correlation features between indicators. The standardized time series data is divided into overlapping local windows; using a parallel structure composed of independent channel convolution and encoder / decoder, multi-level feature extraction and dimensionality reduction are performed on the local windows to capture multi-granular time series patterns from instantaneous changes to stage trends, and summarized into local pattern representations for each indicator. The global temporal dependency features, the correlation features between the indicators, and the local pattern representations are integrated through a trainable adaptive weighted fusion gate; the fused representations are input into the prediction network, and the predicted sequence of all clinical indicators within a preset future time period is output.

2. The method for real-time prediction of intensive care status based on mask dynamic graphing according to claim 1, characterized in that, The sparse inputs for constructing the numerical matrix and the monitoring identifier matrix include: Define sparse time-series data containing N clinical indicators, with an input observation window length of L, a numerical matrix denoted as S, and a corresponding binary monitoring identifier matrix denoted as O. In the monitoring identifier matrix, an element of 1 indicates that the corresponding indicator has a monitoring record at the corresponding time, and an element of 0 indicates that there is no monitoring record. The original numerical matrix is ​​filled with gaps, and the filled numerical matrix is ​​then subjected to min-max normalization. The standardized scalar values ​​are converted into numerical embeddings through nonlinear mapping. The numerical embeddings are added to the trainable indicator identity code and then subjected to layer standardization to generate basic time series representations. The sinusoidal positional encoding is superimposed on the basic temporal representation of each index to obtain the final input representation tensor.

3. The method for real-time prediction of intensive care status based on mask dynamic graphing according to claim 1, characterized in that, The construction of a dynamic relational graph adjacency matrix based on the monitoring frequency between indicators includes: The total number of times each indicator is monitored within the entire observation window is calculated based on the monitoring identifier matrix. The total number of times of monitoring reflects the clinical attention density to the indicator. Using the total number of monitoring sessions, an adjacency matrix is ​​constructed to represent the relationship between indicators, and the correlation strength between indicators is calculated. The adjacency matrix is ​​row normalized to ensure the stability of subsequent feature diffusion.

4. The method for real-time prediction of intensive care status based on mask dynamic graphing according to claim 1, characterized in that, The step of using the adjacency matrix of the relational graph to perform multi-order feature diffusion on the initial representation and extracting the correlation features between indicators includes: The initial node feature tensor obtained by the linear transformation of the input representation is used to propagate K-order features using the normalized adjacency matrix; Each propagation step multiplies the normalized adjacency matrix with the current-order feature tensor to obtain the next-order feature tensor. The final-order feature tensor is used as the index representation of the graph-enhanced relation graph, which integrates the correlation information between indicators based on the monitoring frequency.

5. The method for real-time prediction of intensive care status based on mask dynamic graphing according to claim 1, characterized in that, The multi-level feature extraction and dimensionality reduction of the local window using a parallel structure composed of independent channel convolution and encoder / decoder includes: The standardized input is divided into a series of overlapping local windows; Each window is fed into multiple parallel processing channels, and each processing channel first processes each metric using independent one-dimensional convolution to extract local patterns. Each processing channel then uses a lightweight symmetric codec to compress the convolutional output, learning a compact representation; The outputs of all processing channels are averaged to obtain a window-level representation, which is then mapped to the local temporal pattern features of each index through inverse window operations and various aggregation strategies.

6. The method for real-time prediction of intensive care status based on mask dynamic graphing according to claim 1, characterized in that, The integration via a trainable adaptive weighted fusion gate includes: For each metric, the temporal dependency features, inter-metric correlation features, and local pattern features are calculated using independently learned fusion gate weights. The fusion gate weights are obtained by multiplying learnable parameters with corresponding features and then activating them using the sigmoid function; The features of the three channels are multiplied element-wise with their respective fusion gate weights and then added together to obtain the fused representation. If external temporal context information is provided, the fused representation is modulated using a context modulation layer to generate a context-aware representation.

7. The method for real-time prediction of intensive care status based on mask dynamic graphing according to claim 1, characterized in that, The output includes the predicted sequence of all clinical indicators for a future preset time period, including: The fused representation is input into a two-layer feedforward prediction network, which directly outputs a matrix of predicted values ​​for all N indicators within the next H time points. The objective function is defined as calculating the mean square error only at locations with actual monitoring records during the future prediction period; The objective function ensures that model optimization is based only on monitoring points where true values ​​are available, while inferring the complete prediction space through a representation learning mechanism.

8. A real-time prediction system for intensive care status based on mask dynamic graphing, characterized in that, include: The data acquisition module is used to collect multivariate time-series monitoring data of patients in the intensive care unit and construct a sparse input containing a numerical matrix and a monitoring identifier matrix. The representation generation module is used to fill missing values ​​and standardize the numerical matrix, and combine the processed numerical matrix with trainable index feature encoding to generate a high-dimensional representation vector that integrates temporal location information. The temporal coding module is used to perform global temporal dependency coding on the high-dimensional representation vector and extract the global temporal dependency features of each indicator. The relationship graph construction module is used to calculate the monitoring frequency of each clinical indicator with the monitoring identifier matrix as the core input, and construct a dynamic relationship graph adjacency matrix without training parameters based on the monitoring frequency between indicators to quantify the collaborative monitoring intensity between indicators. The feature diffusion module is used to perform multi-stage feature diffusion on the initial representation using the adjacency matrix of the relation graph to extract the correlation features between indicators; The local pattern extraction module is used to divide the standardized time series data into overlapping local windows, and use a parallel structure composed of independent channel convolution and encoder-decoder to perform multi-level feature extraction and dimensionality reduction on the local windows, capture multi-granular time series patterns and summarize the local pattern representation of each indicator. An adaptive fusion module is used to integrate the global temporal dependency features, the inter-indicator correlation features, and the local pattern representations through a trainable adaptive weighted fusion gate; The prediction output module is used to input the fused representations into the prediction network and directly output the prediction sequence of all clinical indicators within a preset time period.

9. An electronic device, characterized in that, The device includes: a processor and a memory; The memory is used to store instructions; The processor is configured to execute the instructions in the memory to perform the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Including instructions that, when run on a computer, cause the computer to perform the method described in any one of claims 1-7 above.

Images