Severe patient sedation state evaluation method based on multi-modal network model

Abstract

The application discloses a critical patient sedation state evaluation method based on a multi-modal network model, relates to the technical field of medical monitoring and signal processing, and comprises the following steps: analyzing a plurality of channel physiological signals of a patient, constructing a dynamic time-varying function network, and performing multi-scale community detection to identify stable function modules in the network and dynamic connection modes thereof. The network topology mode is fused with instantaneous phase information of the signals to form a fusion feature space-time atlas. After the atlas is subjected to deep feature extraction and structured coding, a sequence modeling network is used to capture time sequence dependence, and finally a quantitative sedation evaluation signal is output by a decoder. Through dynamic modular structure analysis of the brain function network and cross-level feature fusion, the method realizes more continuous and accurate objective evaluation of the sedation state.

Description

Technical Field

[0001] This invention belongs to the field of medical monitoring and signal processing technology, specifically a method for assessing the sedation state of critically ill patients based on a multimodal network model. Background Technology

[0002] In intensive care clinical practice, accurate assessment of a patient's sedation state is crucial. Current assessment methods primarily rely on intermittent scoring using subjective scales by healthcare professionals, or fixed-time-window analysis based on physiological signals such as electroencephalograms (EEGs). The latter often generates quantitative indicators by calculating statistical characteristics such as signal spectrum and entropy values, or by constructing static functional connectivity networks. These techniques simplify brain activity into a homogenized overall state or isolated local oscillations, failing to effectively characterize the dynamic coordination and reorganization of whole-brain functions under sedation.

[0003] The shortcomings of existing technical solutions lie in their lack of dynamism and integration. Traditional network analysis methods are mostly limited to the temporal tracking of global topological indicators, failing to analyze the dynamic evolution and interactions of substructures within the network. Furthermore, the oscillatory characteristics of physiological signals and macroscopic network topological features are often extracted and modeled separately. This fragmented approach ignores the intrinsic mechanisms of cross-scale coupling in neural activity, resulting in limited ability of assessment models to distinguish subtle changes in sedation depth. Consequently, the robustness and continuity of assessment results are insufficient to meet the needs of precision medicine. Therefore, there is an urgent need for an objective assessment method capable of characterizing the dynamic modular reorganization of brain function and achieving deep integration of microscopic oscillations and macroscopic network features. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] Therefore, this invention proposes a method for assessing the sedation status of critically ill patients based on a multimodal network model, including:

[0006] The multi-channel physiological monitoring signals of critically ill patients were acquired and decomposed to generate multiple intrinsic modal components and a residual trend term.

[0007] Key mode components with coherent oscillation characteristics are selected from multiple intrinsic mode components, and the dynamic mutual information matrix among the key mode components is calculated.

[0008] Based on a dynamic mutual information matrix, a dynamic time-varying functional network is constructed, which includes a set of node connection weights that evolve over time.

[0009] Multi-scale community detection is performed on dynamic time-varying functional networks to identify stable functional modules and inter-module connection patterns in the network.

[0010] By integrating the connection mode of the fusion stable functional module with the instantaneous phase information of the key modal components, a fusion feature spatiotemporal map is generated.

[0011] The fused spatiotemporal map of features is input into a cascaded convolution module to extract deep spatiotemporal invariant features, which have translation and scale invariance.

[0012] By performing structured sparse coding on deep spatiotemporal invariant features, a high-dimensional sparse feature representation is obtained, while preserving the structural relationship of features in the spatiotemporal dimension.

[0013] High-dimensional sparse feature representation and residual trend term are introduced, and sequence dependency modeling is performed through gated recurrent unit network to output temporal dependency feature vector;

[0014] A state decoder is driven by a time-dependent feature vector, which outputs a quantitative assessment signal characterizing the patient's sedation state.

[0015] Furthermore, the acquisition of multi-channel physiological monitoring signals from critically ill patients, after decomposition, generates multiple intrinsic modal components and a residual trend term, including:

[0016] Acquire multi-channel physiological monitoring signals from critically ill patients to form a raw monitoring dataset;

[0017] Adaptive mode decomposition is performed on the original monitoring dataset to generate multiple intrinsic mode components and a residual trend term, specifically including:

[0018] A mode decomposition algorithm with adaptive stopping conditions is used to iteratively decompose the signal of each channel in the original monitoring dataset;

[0019] In each iterative decomposition process, the local extreme points of the signal are calculated, and the upper and lower envelopes of the signal are constructed by interpolation methods respectively;

[0020] The local mean curve of the signal is calculated based on the upper and lower envelopes of the signal, and the candidate component is obtained by subtracting the local mean curve from the original signal.

[0021] Determine whether the candidate component satisfies the preset intrinsic mode conditions. The intrinsic mode conditions include that the difference between the number of extreme points and the number of zero crossings in the entire data segment is less than a threshold, and that the local mean defined by the upper and lower envelopes of the signal is zero.

[0022] Candidate components that satisfy the intrinsic mode conditions are identified as intrinsic mode components. The intrinsic mode components are separated from the original signal, and the iterative decomposition process is repeated on the remaining signal until the remaining signal becomes a monotonic function or only one extreme point remains. At this point, the remaining signal is identified as the residual trend term.

[0023] Furthermore, the step of selecting key mode components with coherent oscillation characteristics from multiple intrinsic mode components and calculating the dynamic mutual information matrix among the key mode components includes:

[0024] Calculate the Hilbert spectrum for each intrinsic mode component to obtain the instantaneous frequency and instantaneous amplitude of the component;

[0025] The center oscillation frequency of each intrinsic mode component is determined based on the statistical distribution of instantaneous frequency, and intrinsic mode components whose center oscillation frequency is within the preset physiological frequency range are selected as candidate key components.

[0026] Calculate the amplitude envelope signal of all candidate key components within the sliding time window, and calculate the dynamic mutual information value between each pair of candidate key components based on the amplitude envelope signal;

[0027] The calculated dynamic mutual information values ​​are aligned with the center point of the time window to construct a three-dimensional tensor dynamic mutual information matrix, where the three dimensions of the three-dimensional tensor correspond to the candidate key component number, the candidate key component number, and the time window index, respectively.

[0028] Furthermore, the dynamic time-varying functional network is constructed based on the dynamic mutual information matrix. This dynamic time-varying functional network includes a set of node connection weights that evolve over time, comprising:

[0029] Each candidate key component in the dynamic mutual information matrix is ​​mapped to a node in a dynamic time-varying functional network.

[0030] For each time window in the dynamic mutual information matrix, the mutual information value matrix corresponding to the time window is thresholded, and connections with mutual information values ​​exceeding the statistical significance threshold are retained.

[0031] The retained connections are used as valid connection edges of the dynamic time-varying functional network within the time window, and the corresponding mutual information values ​​are normalized and used as the weights of the valid connection edges.

[0032] By integrating all valid connection edges and their weights within the time window, a set of node connection weights that evolves over time in a dynamic time-varying functional network is formed. The set of node connection weights is organized in the form of a four-dimensional tensor, with the four dimensions corresponding to the source node number, the target node number, the connection edge type, and the time window index, respectively.

[0033] Furthermore, the multi-scale community detection of the dynamic time-varying functional network to identify stable functional modules and inter-module connection patterns in the network includes:

[0034] Cluster analysis was performed on the set of node connection weights in a dynamic time-varying functional network under multiple different connection strength thresholds.

[0035] At each connection strength threshold, nodes are divided into different communities using a modularity optimization algorithm, with each community corresponding to a potential functional module;

[0036] By comparing the community division results under different connection strength thresholds, communities that maintain stable member composition under multiple thresholds are identified, and these communities are determined as stable functional modules.

[0037] By analyzing the average connection strength of nodes within each stable functional module and the average connection strength of nodes between different stable functional modules, the internal connection patterns and inter-module connection patterns of the stable functional modules can be obtained.

[0038] Furthermore, the connection mode of the fusion stabilization functional module and the instantaneous phase information of the key modal components are used to generate a fusion feature spatiotemporal map, including:

[0039] Extract the analytical signals of the key modal components and calculate the instantaneous phase angle of the analytical signals;

[0040] The instantaneous phase angles of all key modal components belonging to the same stable functional module are weighted and averaged to obtain the representative phase timing signal of each stable functional module.

[0041] Construct a module-level connection matrix based on the internal connection patterns and inter-module connection patterns of the stable functional modules;

[0042] The representative phase timing signal of each stable functional module is subjected to tensor product operation with the connection matrix at the stable functional module level, and then expanded along the time dimension to generate a three-dimensional fusion feature spatiotemporal map, where the three dimensions correspond to the spatial module dimension, the spatial module dimension and the time point, respectively.

[0043] Furthermore, the step of inputting the fused spatiotemporal feature map into the cascaded convolution module to extract deep spatiotemporally invariant features includes:

[0044] The cascaded convolutional module is composed of multiple convolutional layers and pooling layers cascaded alternately.

[0045] The fused spatiotemporal feature map is taken as input and passed sequentially through the convolutional and pooling layers in the cascaded convolutional module.

[0046] In each convolutional layer, a 3D convolutional kernel is used to perform a convolution operation on the input feature map to extract local spatiotemporal features;

[0047] In each pooling layer, the feature map output by the convolutional layer is downsampled to enhance the translation and scale invariance of the features;

[0048] After the last layer of the cascaded convolution module, a feature map with a higher level of abstraction is output, which is the deep spatiotemporal invariant feature.

[0049] Furthermore, the structured sparse coding of deep spatiotemporal invariant features to obtain high-dimensional sparse feature representations includes:

[0050] Pre-define an overcomplete atom dictionary, where each atom is a basis vector with the same spatiotemporal dimension as the deep spatiotemporal invariant features;

[0051] Using deep spatiotemporal invariant features as the input signal, a structured sparse coding algorithm is used to find a linear combination of atoms in the overcomplete atom dictionary to approximate the input signal;

[0052] During the encoding process, a group sparsity constraint is introduced to ensure that the selected atoms have continuity in the spatiotemporal structure, thereby ensuring that the feature representation after sparse encoding can preserve the spatiotemporal structure relationship of the input features.

[0053] The resulting sparse coefficient vector is used as a high-dimensional sparse feature representation, in which most coefficients are zero and only a few coefficients are non-zero.

[0054] Furthermore, the introduction of high-dimensional sparse feature representation and residual trend term, followed by sequence dependency modeling through a gated recurrent unit network, outputs a temporal dependency feature vector, including:

[0055] The high-dimensional sparse feature representation is segmented along the time dimension to form a series of feature fragment sequences arranged in chronological order.

[0056] The feature fragments of each time step are concatenated with the residual trend term sampled values ​​at the corresponding time points to form the joint input vector of the time step;

[0057] The joint input vector sequence is fed into the gated recurrent unit network in sequence. The gated recurrent unit network captures the time dependencies in the joint input vector sequence through its update gate and reset gate mechanism.

[0058] The hidden state vector of the last time step of the gated recurrent unit network is used as the context summary of the entire sequence, and the context summary vector is the temporal dependent feature vector.

[0059] Furthermore, the use of time-dependent feature vectors to drive a state decoder, the state decoder outputting a quantitative assessment signal characterizing the patient's sedation state, including:

[0060] The state decoder consists of a fully connected feedforward neural network;

[0061] The temporally dependent feature vector is used as the input to the state decoder and then fed into the first layer of the fully connected feedforward neural network.

[0062] Each layer of a fully connected feedforward neural network performs a linear transformation on the input and passes it through a non-linear activation function, ultimately producing a multi-dimensional vector at the output layer;

[0063] The multidimensional vector is standardized and mapped to constrain its numerical range to a preset evaluation scale, thereby generating a continuous quantitative evaluation signal for characterizing the depth of sedation in patients.

[0064] Compared with the prior art, the beneficial effects of the present invention are:

[0065] By implementing multi-scale community detection on dynamic time-varying functional networks, the system can identify relatively stable functional modules in the network topology over time and simultaneously analyze the dynamic evolution of the connection weights between these modules. This achieves a shift from tracking the overall state of the network to characterizing the dynamic reorganization of its internal modular structure. The features extracted by this method can more precisely reflect the spatiotemporal specificity of the brain's functional integration and separation processes under sedation, thereby improving the sensitivity and discriminative ability of the assessment model to continuous and subtle changes in sedation depth.

[0066] The instantaneous phase information of key modal components is fused with the connection patterns of the aforementioned stable functional modules to generate a fused feature spatiotemporal map. This process achieves cross-scale correlation between the microscopic oscillation characteristics of the signal and the macroscopic topological characteristics of the network at the data representation level. The generated features simultaneously include the synchronous oscillation dynamics information of local neural clusters and the functional interaction pattern information of these clusters in the global network. This deep fusion of features enhances the model's ability to represent complex neurophysiological patterns, making the evaluation basis more comprehensive and helping to improve the accuracy of state discrimination.

[0067] By performing structured sparse encoding on deep spatiotemporal invariant features, a high-dimensional sparse feature representation is obtained while preserving the structural correlations of the original features in the temporal and spatial dimensions. This operation not only achieves feature dimensionality reduction and redundancy removal but also maintains the structured information of the feature tensor. This enables subsequent sequence modeling networks to more effectively utilize the structured prior knowledge of the features, thereby more accurately capturing the temporal dependencies of sedation state evolution. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating the steps of the method for assessing the sedation status of critically ill patients based on a multimodal network model, as described in this invention.

[0069] Figure 2 A flowchart for screening key modal components and calculating the dynamic mutual information matrix;

[0070] Figure 3 A flowchart for constructing a dynamic time-varying functional network;

[0071] Figure 4 Heatmap of dynamic mutual information matrix for key modal components;

[0072] Figure 5 A time-series visualization of the activity intensity of dynamic time-varying functional network modules and the connectivity between modules. Detailed Implementation

[0073] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] See Figure 1 Multi-channel physiological monitoring signals from critically ill patients, such as electroencephalograms (EEGs), electrocardiograms (ECGs), and electromyograms (EMGs), are acquired and processed by adaptive mode decomposition to generate multiple intrinsic mode components (IMCs) and a residual trend term. Key mode components with coherent oscillation characteristics are selected from these IMCs, and the dynamic mutual information matrix between these key mode components is calculated. Based on this dynamic mutual information matrix, a dynamic time-varying functional network (DTNF) is constructed, containing a set of node connection weights that evolve over time. Multi-scale community detection is performed on the DTNF to identify stable functional modules and their inter-module connection patterns. Then, the connection patterns of stable functional modules are fused with the instantaneous phase information of the key mode components to generate a fused feature spatiotemporal map. This fused feature spatiotemporal map is input into a cascaded convolutional module to extract deep spatiotemporally invariant features, which are translation- and scale-invariant. Structured sparse encoding is applied to these deep spatiotemporally invariant features to obtain a high-dimensional sparse feature representation while preserving the structural relationships of the features in the spatiotemporal dimension. A high-dimensional sparse feature representation and the residual trend term obtained from the aforementioned decomposition are introduced, and sequence dependency modeling is performed through a gated recurrent unit network, outputting a temporally dependent feature vector. This temporally dependent feature vector drives a state decoder, which outputs a quantitative assessment signal characterizing the patient's sedation state.

[0075] See Figure 2In one embodiment of the present invention, the method for assessing the sedation state of critically ill patients based on a multimodal network model involves acquiring multi-channel physiological monitoring signals from critically ill patients. These signals include electroencephalogram (EEG), electrocardiogram (ECG), and electromyogram (EMG) signals. After acquiring these signals, an original monitoring dataset is formed. Adaptive mode decomposition (AMD) is performed on the original monitoring dataset to generate multiple intrinsic mode components and a residual trend term. An AMD algorithm with adaptive stopping conditions is used to iteratively decompose each channel signal in the original monitoring dataset. During each iteration, local extrema of the signal are calculated, and upper and lower envelopes of the signal are constructed using interpolation. The local mean curve of the signal is calculated based on the upper and lower envelopes. Subtracting the local mean curve from the original signal yields candidate components. It is then determined whether the candidate components meet preset intrinsic mode conditions. These intrinsic mode conditions include a difference between the number of extrema and the number of zero-crossing points within the entire data segment being less than a threshold, and a local mean defined by the upper and lower envelopes of the signal being zero. The mathematical expression of the intrinsic mode conditions is: for candidate components... The number of extreme points is The number of zero points is Set the threshold to The conditions are And for all time points Local mean :

[0076]

[0077] in: It is an upper envelope function. It is the lower envelope function. Candidate components that satisfy the intrinsic mode condition are identified as intrinsic mode components. Intrinsic mode components are separated from the original signal, and the remaining signal is iteratively decomposed repeatedly until the remaining signal becomes a monotonic function or only one extreme point remains. At this point, the remaining signal is identified as the residual trend term.

[0078] In some embodiments, the Hilbert spectrum of each intrinsic modal component is calculated to obtain the instantaneous frequency and instantaneous amplitude of the component. The central oscillation frequency of each intrinsic modal component is determined based on the statistical distribution of the instantaneous frequency, and intrinsic modal components whose central oscillation frequencies fall within a preset physiological frequency band are selected as candidate key components. The preset physiological frequency band is set to 0.5 Hz to 30 Hz, corresponding to the common physiological oscillation frequency bands of EEG. The amplitude envelope signal of all candidate key components within a sliding time window is calculated, and the dynamic mutual information value between each pair of candidate key components is calculated based on the amplitude envelope signal. The calculated dynamic mutual information values ​​are aligned with the center point of the time window to construct a three-dimensional tensor dynamic mutual information matrix. The three dimensions of the three-dimensional tensor correspond to the candidate key component number, the candidate key component number, and the time window index, respectively. Optionally, when selecting key modal components with coherent oscillation characteristics, the matching degree between the central oscillation frequency of different intrinsic modal components and the preset physiological frequency band is compared. For example, intrinsic modal components whose central oscillation frequencies fall outside the preset physiological frequency band are excluded. When calculating the dynamic mutual information matrix, the length of the sliding time window is adjusted according to the signal sampling rate. In the specific implementation, the length of the sliding time window is set to 2 seconds with an overlap ratio of 50%. It can be understood that the adaptive mode decomposition algorithm can effectively handle nonlinear and non-stationary signals. In the specific implementation, empirical mode decomposition is used as the basis of the adaptive mode decomposition algorithm. The selection of intrinsic mode components is based on the statistical characteristics of the center oscillation frequency, and the construction of the dynamic mutual information matrix reflects the statistical dependence of candidate key components over time.

[0079] In practical implementation, data comparison is reflected in the fact that the number of intrinsic modal components obtained after decomposing signals from different channels may vary. For example, EEG signals may generate more high-frequency intrinsic modal components, while ECG signals may generate more low-frequency intrinsic modal components. When screening key modal components, only components with central oscillation frequencies within a preset physiological frequency band are retained. The calculation of the dynamic mutual information matrix is ​​based on the amplitude envelope signal capturing the nonlinear correlation between components. For example, for the alpha and beta frequency band components in an EEG signal, the dynamic mutual information value may change over time to reflect fluctuations in the sedation state. In some embodiments, the preset physiological frequency band range can be further subdivided into delta, theta, alpha, and beta bands, corresponding to different physiological states. Optionally, when calculating the dynamic mutual information value, the amplitude envelope signal within a sliding window is used to obtain the mutual information value through a histogram estimation method. In practical implementation, the three-dimensional tensor structure of the dynamic mutual information matrix allows for tracking changes in connectivity patterns along the time dimension in subsequent analysis, and data comparison can be achieved by comparing the distribution of mutual information values ​​under different time windows.

[0080] See Figure 3In one embodiment of the present invention, a dynamic time-varying functional network (DTNF) is constructed based on a dynamic mutual information matrix. The DTNF includes a set of node connection weights that evolve over time. Each candidate key component in the dynamic mutual information matrix is ​​mapped to a node in the DTNF. For each time window in the dynamic mutual information matrix, the mutual information value matrix corresponding to the time window is thresholded, retaining connections whose mutual information values ​​exceed a statistical significance threshold. The statistical significance threshold is obtained by performing a permutation test on the distribution of mutual information values ​​within each time window. For example, in a specific implementation, the zero distribution is calculated by randomly shuffling the time series of the amplitude envelope signals of the candidate key components and setting the significance level to 0.05. The retained connections are used as valid connection edges of the DTNF within the time window, and the corresponding normalized mutual information values ​​are used as the weights of the valid connection edges. The normalization method uses min-max normalization to map the mutual information values ​​to the interval between 0 and 1. The effective connection edges and their weights within all time windows are integrated to form a set of node connection weights that evolve over time in a dynamic time-varying functional network. The set of node connection weights is organized in the form of a four-dimensional tensor, with the four dimensions corresponding to the source node number, target node number, connection edge type, and time window index, respectively.

[0081] In some embodiments, cluster analysis is performed on the node connection weight set of the dynamic time-varying functional network under multiple different connection strength thresholds. At each connection strength threshold, nodes are divided into different communities using a modularity optimization algorithm, with each community corresponding to a potential functional module. The objective function for modularity optimization is to maximize the modularity value. :

[0082]

[0083] in: It is a node With nodes Connection weights between them It is a node The sum of all connection weights, It is half the sum of the weights of all connections in the network. Represents a node The community to which it belongs At the node and nodes A value of 1 indicates membership in the same community, while 0 indicates membership in the same community. By comparing community partitioning results under different connection strength thresholds, communities with stable member composition across multiple thresholds are identified and defined as stable functional modules. The average connection strength of nodes within each stable functional module and the average connection strength between nodes in different stable functional modules are analyzed to obtain the internal connection patterns and inter-module connection patterns of stable functional modules. Optionally, the connection strength threshold is gradually adjusted downwards from the upper quartile of the normalized weight range, generating multiple network snapshots with varying sparsity for multi-scale community detection. It can be understood that the construction of a dynamic time-varying functional network maps time-varying statistical dependencies to a time-varying network topology. Multi-scale community detection identifies stable functional modules unaffected by changes in connection sparsity by analyzing the robustness of the community structure under different connection strength thresholds.

[0084] In practical implementation, data comparison reveals that dynamically time-varying functional networks constructed under different time windows may have different connection densities and topologies. For example, during deeper sedation of patients, the network may exhibit a stronger modular structure while overall connection strength decreases. In the example scenario, the node connection weight set is a four-dimensional tensor, whose time window index dimension records the dynamic evolution of the network over time, allowing for comparative analysis between time points. In some embodiments, the identification of stable functional modules is achieved by comparing, for example, the membership composition of the community at connection strength thresholds of 0.3, 0.4, and 0.5. At least 80% of the member nodes of a stable functional module remain consistent across these three thresholds. Regarding data comparison, the differences in average connection strength within different stable functional modules can be compared. For example, the connection strength within a functional module related to sensory processing weakens as sedation deepens, while the connection strength change pattern within another functional module related to the default mode network may differ. In practical implementation, the four-dimensional tensor structure of the node connection weight set allows for simultaneous analysis of inter-module connection patterns in both time and space dimensions. Comparing changes in inter-module connection strength across different time intervals can reveal the process of brain functional network reorganization.

[0085] In one embodiment of the present invention, the connection mode of the fusion stable functional module and the instantaneous phase information of the key modal components are used to generate a fused feature spatiotemporal map. The analytic signals of the key modal components are extracted and the instantaneous phase angles of the analytic signals are calculated. In a specific implementation, the analytic signals and instantaneous phase angles of the key modal components are obtained through Hilbert transform. :

[0086]

[0087] in: It is the kth critical mode component. This is its Hilbert transform. The representative phase-time signal of each stable functional module is obtained by weighted averaging of the instantaneous phase angles of all key modal components belonging to the same stable functional module. The weights of the averaging are allocated based on the connection strength or amplitude energy of the key modal components within the corresponding stable functional module. A module-level connection matrix is ​​constructed based on the internal and inter-module connection modes of the stable functional modules. The module-level connection matrix is ​​a symmetric matrix, where the diagonal elements represent the average internal connection strength of the stable functional module, and the off-diagonal elements represent the average connection strength between different stable functional modules.

[0088] In some embodiments, the representative phase time-series signal of each stable functional module is subjected to a tensor product operation with the connection matrix at the stable functional module level, and then expanded along the time dimension to generate a three-dimensional fusion feature spatiotemporal map. The three dimensions of the three-dimensional fusion feature spatiotemporal map correspond to the spatial module dimension, the spatial module dimension, and the time point, respectively. In a specific implementation, for a system containing M stable functional modules, the representative phase time-series signal consists of M time series, and the module-level connection matrix is ​​an M×M matrix. The representative phase time-series signal vector is calculated for each time point t. With connection matrix The tensor product yields an M×M matrix. ,in The Hadamard product represents element-wise multiplication, covering all time points. Stacked in chronological order to form a three-dimensional fusion feature spatiotemporal map Optionally, when calculating the representative phase time-series signal using weighted averaging, the weights can be set to the average amplitude envelope values ​​calculated in the screening step for the corresponding key modal components. It can be understood that the fused feature spatiotemporal map simultaneously encodes the structural connection strength information between functional modules and the collective phase dynamic information of oscillations within the modules, providing a clearly defined spatiotemporal structured input for subsequent feature extraction.

[0089] In practical implementation, data comparison is reflected in the fact that representative phase-series signals from different stable functional modules may exhibit different oscillation patterns. For example, the phase dynamics of functional modules from the sensory processing cortex may be dominated by alpha band oscillations, while functional modules from the prefrontal cortex may exhibit more theta band oscillations. In the example scenario, the module-level connectivity matrix reflects the interaction strength between different brain functional systems. For instance, the connectivity strength between the sensorimotor network and the attention network is high in a conscious state, but may weaken under deep sedation. In some embodiments, the generated three-dimensional fusion feature spatiotemporal map can be viewed as a series of functional connectivity maps evolving over time, with each time point corresponding to a two-dimensional slice that is a phase-weighted functional connectivity matrix. Optionally, when constructing the module-level connectivity matrix, the average connectivity strength values ​​of internal connectivity patterns and inter-module connectivity patterns can be z-score normalized to eliminate baseline differences in the inherent connectivity strength between different modules. Data comparison can be achieved by observing the time-varying sequence of specific elements in the fusion feature spatiotemporal map, which simultaneously captures the structural connectivity strength and phase synchronization dynamics between the two modules. In practice, the fusion operation combines the periodic variable phase with the non-periodic variable connection strength in tensor operations. In the example, the Hadamard product operation ensures that the connection matrix only plays a spatial modulation role on the phase tensor without introducing new temporal dynamics.

[0090] In one embodiment of the present invention, the cascaded convolutional module is composed of multiple convolutional layers and pooling layers cascaded alternately. The fused spatiotemporal feature map is taken as input and passed sequentially through the convolutional and pooling layers in the cascaded convolutional module. In each convolutional layer, a three-dimensional convolutional kernel is used to perform convolution operations on the input feature map to extract local spatiotemporal features. In each pooling layer, the feature map output by the convolutional layer is downsampled to enhance the translation and scale invariance of the features. After passing through the last layer of the cascaded convolutional module, a feature map with a higher level of abstraction is output. This feature map is the deep spatiotemporally invariant feature. In specific implementations, the size, stride, and padding parameters of the three-dimensional convolutional kernel need to be set according to the dimension of the fused spatiotemporal feature map. For example, for a spatial module of the fused spatiotemporal feature map with a dimension of M and a number of time points of T, the convolutional kernel can be set to 3×3×3 in both the spatial and temporal dimensions, the stride can be set to 1×1×1, and the padding method can be set to the same padding to maintain the feature map size. In practice, pooling operations typically employ max pooling or average pooling, with a pooling window size set to 2×2×2 and a stride of 2, thereby performing downsampling in both spatial and temporal dimensions. The specific number of layers and parameter configuration of the cascaded convolutional module can be adjusted according to computational resources and task complexity; see Table 1 for an exemplary cascaded convolutional module structure.

[0091] Table 1: Example Structure Table of Cascaded Convolutional Modules

[0092] In some embodiments, the extraction of deep spatiotemporal invariant features relies on the ability of 3D convolution to capture spatiotemporal local patterns. For example, when the convolutional kernel slides on the spatiotemporal map of the fused features, it can simultaneously perceive the spatial connection patterns between adjacent functional modules and the phase change patterns between adjacent time points. After multiple layers of convolution and pooling operations, the size of the output deep spatiotemporal invariant features decreases in both spatial and temporal dimensions, but the number of feature channels increases, making the feature representation more abstract and invariant to small translations and scale changes of the input.

[0093] In practice, structured sparse coding is applied to deep spatiotemporally invariant features to obtain a high-dimensional sparse feature representation. An overcomplete atom dictionary is pre-defined, where each atom is a basis vector with the same spatiotemporal dimension as the deep spatiotemporally invariant features. Using the deep spatiotemporally invariant features as input signals, a structured sparse coding algorithm is used to find a linear combination of atoms in the overcomplete atom dictionary to approximate the input signal. A sparsity constraint is introduced during the encoding process to ensure that the selected atoms have spatiotemporal continuity, thereby ensuring that the sparsely encoded feature representation preserves the spatiotemporal structural relationships of the input features. The final sparse coefficient vector is used as the high-dimensional sparse feature representation, where most coefficients are zero, and only a few are non-zero.

[0094] In some embodiments, an overcomplete atom dictionary can be pre-trained using clustering or dictionary learning algorithms on deep spatiotemporal invariant feature samples in the training dataset. The size of the atom dictionary is typically much larger than the dimension of the deep spatiotemporal invariant features to achieve overcompleteness. Optionally, the atom grouping in the group sparsity constraint is based on the spatiotemporal position of the atoms, for example, grouping adjacent atoms in the spatiotemporal dimension into the same group to promote spatiotemporal continuity of the selected atoms. Optionally, the structured sparse coding algorithm can be solved using iterative optimization algorithms such as proximal gradient descent or alternating direction multiplier method. It can be understood that structured sparse coding, by introducing group sparsity constraints, makes the sparse representation not only have the sparsity of coefficients but also maintain the spatiotemporal structure of coefficients, so that the high-dimensional sparse feature representation can more effectively characterize the structural information in deep spatiotemporal invariant features. In terms of data comparison, after deep spatiotemporal invariant features are structured sparsely encoded, the positions of the non-zero coefficients in the high-dimensional sparse feature representation can reflect significant spatiotemporal patterns in the input features. For example, during changes in sedation state, non-zero coefficients may be concentrated in groups of atoms representing connection patterns of specific functional modules.

[0095] See Figure 4This is a heatmap of the dynamic mutual information matrix of key modal components. This matrix is ​​the core input for constructing a dynamic time-varying functional network, and connections with high mutual information values ​​will be retained as valid edges in the network. The mutual information pattern between key modal components changes at different sedation depths. For example, in deep sedation, the mutual information of some strongly coupled modules may decrease significantly, reflecting the inhibition of brain functional connectivity. Based on this matrix, multi-scale community detection can be further performed to identify stable functional modules and the connection patterns between them, providing a structural basis for the generation of spatiotemporal maps of fused features. The pattern of the mutual information matrix undergoes characteristic changes at different sedation depths. For example, in deep sedation, strongly coupled regions representing key functional modules may significantly decrease or disappear, which can serve as an objective biomarker for assessing sedation depth.

[0096] In one embodiment of the present invention, a high-dimensional sparse feature representation and a residual trend term are introduced. Sequence dependency modeling is performed using a gated recurrent unit network (GRN), outputting a temporally dependent feature vector. The high-dimensional sparse feature representation is segmented along the time dimension, forming a series of feature segments arranged in chronological order. Each segment in the feature segment sequence corresponds to a high-dimensional sparse feature representation within a time window. The feature segment at each time step is concatenated with the sampled value of the residual trend term at the corresponding time point to form a joint input vector for that time step. The sampling of the residual trend term is achieved by directly reading the instantaneous value of the residual trend term signal at a time point aligned with the time window of the high-dimensional sparse feature representation. The joint input vector sequence is sequentially input into the GRN. The GRN captures the temporal dependencies in the joint input vector sequence through its update and reset gate mechanisms. The update of the hidden state of the GRN follows... :

[0097]

[0098] in: It is the output vector of the update gate at time t, which controls the degree to which the hidden state is retained from the previous time step to the current time step. It is the candidate hidden state vector. This represents element-wise multiplication. This is the hidden state vector from the previous time step. The hidden state vector from the last time step of the gated recurrent unit network is used as the context summary of the entire sequence, and the context summary vector is the temporal dependency feature vector.

[0099] In some embodiments, the temporal dimension segmentation of the high-dimensional sparse feature representation is consistent with the sliding time window partitioning used when constructing the dynamic mutual information matrix, with each feature segment corresponding to a high-dimensional sparse feature representation computed within an original time window. The residual trend term is a low-frequency trend component obtained after adaptive mode decomposition, and in specific implementations, the sampling frequency of the residual trend term is aligned with the time window of the high-dimensional sparse feature representation. The dimension of the joint input vector is the sum of the dimension of the high-dimensional sparse feature representation segment and the dimension of the residual trend term sampling value. For example, if the high-dimensional sparse feature representation segment is a 500-dimensional vector and the residual trend term sampling value is a 1-scalar, then the joint input vector is a 501-dimensional vector. Optionally, the gated recurrent unit network can be designed as a multi-layer structure to enhance its ability to model complex temporal dependencies. It can be understood that the update gate and reset gate mechanism of the gated recurrent unit network enables it to selectively remember or forget historical information, thereby effectively modeling long-term and short-term temporal dependencies in the joint input vector sequence and capturing the dynamic patterns of sedation state evolution.

[0100] In specific implementation, a state decoder is driven by a time-dependent feature vector. The state decoder outputs a quantitative assessment signal characterizing the patient's sedation state. The state decoder is composed of a fully connected feedforward neural network. The time-dependent feature vector is used as input to the state decoder and fed into the first layer of the fully connected feedforward neural network. Each layer of the fully connected feedforward neural network performs a linear transformation on the input and passes it through a non-linear activation function, ultimately generating a multi-dimensional vector at the output layer. The multi-dimensional vector is standardized and mapped to constrain its numerical range to a preset assessment scale, thereby generating a continuous quantitative assessment signal characterizing the patient's sedation depth. In some embodiments, the fully connected feedforward neural network may contain one or more hidden layers, and the dimension of the output layer is consistent with the dimension of the assessment scale, so the output layer can be designed to generate a scalar. Optionally, the standardization mapping can use the sigmoid function to map the values ​​generated by the output layer to the (0,1) interval, and then linearly scale it to the target assessment scale range. The role of the state decoder is to decode abstract time-dependent feature vectors into quantified values ​​of sedation depth with clear clinical significance. Its training objective is to minimize the difference between the predicted quantified assessment signal and the reference value provided by clinical experts or standard instruments. In terms of data comparison, for the same physiological signal, different sedation depths will generate different time-dependent feature vectors through the aforementioned steps, thereby driving the state decoder to output different quantified assessment signal values. For example, when a patient transitions from wakefulness to moderate sedation, the output value of the quantified assessment signal will increase accordingly, which can be verified by comparing it with a reference assessment scale.

[0101] See Figure 5This is a temporal visualization of the activity intensity and inter-module connectivity of a dynamic, time-varying functional network. The activity curves of Module 1 and Module 3 rise synchronously at multiple time points, indicating that they physiologically constitute a highly synergistic functional subsystem. Module 2 shows a negative value at 09:44, which may correspond to the critical time point in clinical practice where sedative drugs take effect and specific brain functional networks are inhibited. Inter-module connectivity significantly increases at the peak activity of Modules 1 and 3, reflecting the reorganization of functional networks at critical time points. The temporal pattern of module activity is an objective biomarker of changes in sedation depth. For example, the synchronous peaks of Modules 1 and 3 may correspond to the moment when a patient transitions from deep sedation to moderate sedation. By observing the rise and fall of module activity, we can intuitively understand how brain functional networks reorganize under different sedation states, which modules are activated, and which are inhibited.

[0102] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for evaluating the sedation state of a critical patient based on a multi-modal network model, characterized in that, include: The multi-channel physiological monitoring signals of critically ill patients were acquired and decomposed to generate multiple intrinsic modal components and a residual trend term. Key mode components with coherent oscillation characteristics are selected from multiple intrinsic mode components, and the dynamic mutual information matrix among the key mode components is calculated. Based on a dynamic mutual information matrix, a dynamic time-varying functional network is constructed, which includes a set of node connection weights that evolve over time. Multi-scale community detection is performed on dynamic time-varying functional networks to identify stable functional modules and inter-module connection patterns in the network. By integrating the connection mode of the fusion stable functional module with the instantaneous phase information of the key modal components, a fusion feature spatiotemporal map is generated. The fused spatiotemporal map of features is input into a cascaded convolution module to extract deep spatiotemporal invariant features, which have translation and scale invariance. By performing structured sparse coding on deep spatiotemporal invariant features, a high-dimensional sparse feature representation is obtained, while preserving the structural relationship of features in the spatiotemporal dimension. High-dimensional sparse feature representation and residual trend term are introduced, and sequence dependency modeling is performed through gated recurrent unit network to output temporal dependency feature vector; A state decoder is driven by a time-dependent feature vector, which outputs a quantitative assessment signal characterizing the patient's sedation state.

2. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The acquired multi-channel physiological monitoring signals of critically ill patients are decomposed to generate multiple intrinsic modal components and a residual trend term, including: Acquire multi-channel physiological monitoring signals from critically ill patients to form a raw monitoring dataset; Adaptive mode decomposition is performed on the original monitoring dataset to generate multiple intrinsic mode components and a residual trend term, specifically including: A mode decomposition algorithm with adaptive stopping conditions is used to iteratively decompose the signal of each channel in the original monitoring dataset; In each iterative decomposition process, the local extreme points of the signal are calculated, and the upper and lower envelopes of the signal are constructed by interpolation methods respectively; The local mean curve of the signal is calculated based on the upper and lower envelopes of the signal, and the candidate component is obtained by subtracting the local mean curve from the original signal. Determine whether the candidate component satisfies the preset intrinsic mode conditions. The intrinsic mode conditions include that the difference between the number of extreme points and the number of zero crossings in the entire data segment is less than a threshold, and that the local mean defined by the upper and lower envelopes of the signal is zero. Candidate components that satisfy the intrinsic mode conditions are identified as intrinsic mode components. The intrinsic mode components are separated from the original signal, and the iterative decomposition process is repeated on the remaining signal until the remaining signal becomes a monotonic function or only one extreme point remains. At this point, the remaining signal is identified as the residual trend term.

3. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The step of selecting key mode components with coherent oscillation characteristics from multiple intrinsic mode components and calculating the dynamic mutual information matrix among the key mode components includes: Calculate the Hilbert spectrum for each intrinsic mode component to obtain the instantaneous frequency and instantaneous amplitude of the component; The center oscillation frequency of each intrinsic mode component is determined based on the statistical distribution of instantaneous frequency, and intrinsic mode components whose center oscillation frequency is within the preset physiological frequency range are selected as candidate key components. Calculate the amplitude envelope signal of all candidate key components within the sliding time window, and calculate the dynamic mutual information value between each pair of candidate key components based on the amplitude envelope signal; The calculated dynamic mutual information values ​​are aligned with the center point of the time window to construct a three-dimensional tensor dynamic mutual information matrix, where the three dimensions of the three-dimensional tensor correspond to the candidate key component number, the candidate key component number, and the time window index, respectively.

4. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The dynamic time-varying functional network is constructed based on a dynamic mutual information matrix. This dynamic time-varying functional network includes a set of node connection weights that evolve over time, including: Each candidate key component in the dynamic mutual information matrix is ​​mapped to a node in a dynamic time-varying functional network. For each time window in the dynamic mutual information matrix, the mutual information value matrix corresponding to the time window is thresholded, and connections with mutual information values ​​exceeding the statistical significance threshold are retained. The retained connections are used as valid connection edges of the dynamic time-varying functional network within the time window, and the corresponding mutual information values ​​are normalized and used as the weights of the valid connection edges. By integrating all valid connection edges and their weights within the time window, a set of node connection weights that evolves over time in a dynamic time-varying functional network is formed. The set of node connection weights is organized in the form of a four-dimensional tensor, with the four dimensions corresponding to the source node number, the target node number, the connection edge type, and the time window index, respectively.

5. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The multi-scale community detection of the dynamic time-varying functional network to identify stable functional modules and inter-module connection patterns includes: Cluster analysis was performed on the set of node connection weights in a dynamic time-varying functional network under multiple different connection strength thresholds. At each connection strength threshold, nodes are divided into different communities using a modularity optimization algorithm, with each community corresponding to a potential functional module; By comparing the community division results under different connection strength thresholds, communities that maintain stable member composition under multiple thresholds are identified, and these communities are determined as stable functional modules. By analyzing the average connection strength of nodes within each stable functional module and the average connection strength of nodes between different stable functional modules, the internal connection patterns and inter-module connection patterns of the stable functional modules can be obtained.

6. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The connection mode and instantaneous phase information of the key modal components of the fusion stabilization functional module are used to generate a fusion feature spatiotemporal map, including: Extract the analytical signals of the key modal components and calculate the instantaneous phase angle of the analytical signals; The instantaneous phase angles of all key modal components belonging to the same stable functional module are weighted and averaged to obtain the representative phase timing signal of each stable functional module. Construct a module-level connection matrix based on the internal connection patterns and inter-module connection patterns of the stable functional modules; The representative phase timing signal of each stable functional module is subjected to tensor product operation with the connection matrix at the stable functional module level, and then expanded along the time dimension to generate a three-dimensional fusion feature spatiotemporal map, where the three dimensions correspond to the spatial module dimension, the spatial module dimension and the time point, respectively.

7. The method for assessing sedation status in critically ill patients based on a multimodal network model according to claim 1, characterized in that, The step of inputting the fused spatiotemporal feature map into the cascaded convolution module to extract deep spatiotemporally invariant features includes: The cascaded convolutional module is composed of multiple convolutional layers and pooling layers cascaded alternately. The fused spatiotemporal feature map is taken as input and passed sequentially through the convolutional and pooling layers in the cascaded convolutional module. In each convolutional layer, a 3D convolutional kernel is used to perform a convolution operation on the input feature map to extract local spatiotemporal features; In each pooling layer, the feature map output by the convolutional layer is downsampled to enhance the translation and scale invariance of the features; After the last layer of the cascaded convolution module, a feature map with a higher level of abstraction is output, which is the deep spatiotemporal invariant feature.

8. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The process of performing structured sparse coding on deep spatiotemporal invariant features to obtain high-dimensional sparse feature representations includes: Pre-define an overcomplete atom dictionary, where each atom is a basis vector with the same spatiotemporal dimension as the deep spatiotemporal invariant features; Using deep spatiotemporal invariant features as the input signal, a structured sparse coding algorithm is used to find a linear combination of atoms in the overcomplete atom dictionary to approximate the input signal; During the encoding process, a group sparsity constraint is introduced to ensure that the selected atoms have continuity in the spatiotemporal structure, thereby ensuring that the feature representation after sparse encoding can preserve the spatiotemporal structure relationship of the input features. The resulting sparse coefficient vector is used as a high-dimensional sparse feature representation, in which most coefficients are zero and only a few coefficients are non-zero.

9. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The introduction of high-dimensional sparse feature representation and residual trend term, followed by sequence dependency modeling through a gated recurrent unit network, outputs a temporal dependency feature vector, including: The high-dimensional sparse feature representation is segmented along the time dimension to form a series of feature fragment sequences arranged in chronological order. The feature fragments of each time step are concatenated with the residual trend term sampled values ​​at the corresponding time points to form the joint input vector of the time step; The joint input vector sequence is fed into the gated recurrent unit network in sequence. The gated recurrent unit network captures the time dependencies in the joint input vector sequence through its update gate and reset gate mechanism. The hidden state vector of the last time step of the gated recurrent unit network is used as the context summary of the entire sequence, and the context summary vector is the temporal dependent feature vector.

10. The method for assessing the sedation status of critically ill patients based on a multimodal network model according to claim 1, characterized in that, The method uses time-dependent feature vectors to drive a state decoder, which outputs a quantitative assessment signal characterizing the patient's sedation state, including: The state decoder consists of a fully connected feedforward neural network; The temporally dependent feature vector is used as the input to the state decoder and then fed into the first layer of the fully connected feedforward neural network. Each layer of a fully connected feedforward neural network performs a linear transformation on the input and passes it through a non-linear activation function, ultimately producing a multi-dimensional vector at the output layer; The multidimensional vector is standardized and mapped to constrain its numerical range to a preset evaluation scale, thereby generating a continuous quantitative evaluation signal for characterizing the depth of sedation in patients.

Images