Neurological nursing risk dynamic early warning system based on multi-modal data

Abstract

This invention discloses a dynamic early warning system for neurological nursing risks based on multimodal data, belonging to the field of medical information processing and risk early warning technology. The system includes: a data acquisition module, which collects EEG, EMG, vital sign time series data, and historical nursing record text to construct a multimodal raw data stream; a tensor decomposition module, which performs joint sparse tensor decomposition to extract core risk feature tensors and separate abnormal residual components; a risk quantification module, which performs tensor condensation of the core risk feature tensor and the risk level quantization matrix to generate multidimensional risk propensity coefficients; a probability prediction module, which inputs the multidimensional risk propensity coefficients into an attention-based sequence prediction network and outputs a predicted sequence of the probability of occurrence of each risk category in the next nursing cycle; and an early warning triggering module, which performs threshold segmentation on the predicted sequences and triggers graded risk early warning signals.

Description

Technical Field

[0001] This invention relates to the field of medical information processing and risk warning technology, specifically a dynamic early warning system for neurological nursing risks based on multimodal data. Background Technology

[0002] Neurological patients often experience rapid changes in their condition and are frequently accompanied by complex neurological dysfunctions. Nursing care requires dynamic monitoring of various risks, including falls, seizures, and changes in intracranial pressure. Existing nursing risk warning methods are mostly based on threshold alarms using single-modal data, such as alerts issued solely based on abnormal fluctuations in heart rate or blood oxygen saturation. This approach fails to integrate information from electroencephalogram (EEG), electromyography (EMG), and the disease progression information implicit in written nursing records, making it difficult to comprehensively depict changes in the patient's neurological state. This results in a high rate of missed or false alarms.

[0003] In terms of data fusion and feature extraction, existing technologies attempt to perform dimensionality reduction or shallow feature analysis after splicing multi-source data, but they lack effective methods for modeling high-order interaction relationships between different modalities. In anomaly detection, traditional methods typically only identify outliers in the overall data, failing to distinguish whether the anomaly is caused by a sudden deterioration in the patient's physiological state or by improper nursing procedures or timing deviations. This coarse-grained anomaly identification makes early warning signals lack attributability, hindering clinicians from quickly locating the source of risk and implementing targeted interventions.

[0004] At the risk prediction level, existing solutions mostly employ static classification models or simple time-series extrapolation methods, which can only output risk scores for a single moment or a single risk type. They fail to provide structured predictions of the probability of occurrence and dynamic evolution of multiple risk categories over future care cycles. This lack of predictive capability makes nursing interventions unproactive and hinders early intervention during the risk accumulation phase.

[0005] To address the above issues, the current technical requirements include: how to jointly extract risk features reflecting the patient's current neurological state from multimodal heterogeneous data and simultaneously separate abnormal patterns that can indicate deviations in nursing operations; and how to dynamically predict the probability of occurrence of various risks in the next nursing cycle based on multidimensional risk propensity information, thereby supporting graded early warning. Summary of the Invention

[0006] This invention provides a dynamic early warning system for neurological nursing risks based on multimodal data. The purpose is to overcome the shortcomings of existing technologies, such as insufficient multimodal information fusion, inability to distinguish abnormal sources, and insufficient dynamic risk prediction capabilities, and to achieve early identification of potential nursing risks for neurological patients, correlation and location of operational deviations, and hierarchical and temporal early warning.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] This invention provides a dynamic early warning system for neurological nursing risks based on multimodal data, including a data acquisition module, a tensor decomposition module, a risk quantification module, a probability prediction module, and an early warning triggering module. The data acquisition module acquires EEG signals, EMG signals, vital sign time-series data, and historical nursing record text of neurological patients, and constructs a multimodal raw data stream. In constructing the multimodal raw data stream, the EEG signal within the same time window is decomposed into multiple frequency band sub-signals. The power spectral density of each frequency band sub-signal is aligned with the root mean square amplitude of the EMG signal according to the timestamp to form a first modal tensor slice. The vital sign time-series data of heart rate, respiratory rate, blood oxygen saturation, and blood pressure are normalized and stacked to form a second modal tensor slice. The historical nursing record text is converted into a semantic feature vector sequence through a word embedding model and compressed into a fixed-length third modal tensor slice. The above three modal tensor slices are concatenated at the same time dimension to generate the multimodal raw data stream. By using the above methods, heterogeneous multi-source data can be unified into a regular tensor structure, providing a consistent data foundation for subsequent joint analysis.

[0009] The tensor decomposition module performs joint sparse tensor decomposition on the multimodal raw data stream, extracts the core risk feature tensor representing the patient's current neurological state, and simultaneously separates the anomalous residual components used to identify nursing operation deviations. Preferably, the multimodal raw data stream is represented as a third-order observation tensor, with dimensions corresponding to time, modality type, and feature channel, respectively; a low-rank tensor is introduced to approximate the main structure of the observation tensor, and a sparse tensor is introduced to capture outliers and mutation points; the low-rank tensor and sparse tensor are iteratively updated using the alternating direction multiplier method until the reconstruction error is less than a preset threshold, and the converged low-rank tensor is used as the core risk feature tensor, and the converged sparse tensor is used as the anomalous residual components. When extracting the core risk feature tensor, Tucker decomposition is performed on the low-rank tensor along the modality type dimension to obtain the kernel tensor and three factor matrices corresponding to time, modality type, and feature channel. The elements with the largest absolute values ​​in the kernel tensor are retained, while the rest are set to zero to obtain a sparsified kernel tensor. This sparsified kernel tensor is then reconstructed and multiplied with the three factor matrices to generate a denoised risk feature tensor. The last time slice is extracted from this risk feature tensor as the core risk feature tensor. This process effectively suppresses cross-modal noise interference and highlights structured features highly correlated with the patient's neurological state.

[0010] When separating outlier residual components, the coordinates of all non-zero elements in the sparse tensor are traversed, and the corresponding original data entries in the multimodal raw data stream are retrieved. The timestamp of each original data entry is then matched with the execution time of the operation in the pre-stored nursing operation record table. If the timestamp falls within the execution time window of any nursing operation, the non-zero element is marked as a nursing operation-related anomaly; if it falls outside the window, it is marked as a spontaneous physiological anomaly. The set of all marked non-zero elements and their coordinates is taken as the outlier residual component. This not only enables outlier detection but also automatically distinguishes whether the source of the anomaly is a deviation in nursing operation or a sudden change in the patient's physiological state, providing a clear direction for the traceability and improvement of nursing quality.

[0011] The risk quantification module performs tensor contraction operations on the core risk feature tensor and a pre-defined risk level quantification matrix to generate a multi-dimensional risk propensity coefficient for each patient. Specifically, it obtains a pre-constructed risk level quantification matrix, where rows correspond to different risk categories and columns correspond to channels of the core risk feature tensor. The core risk feature tensor is expanded along the channel dimension into a risk feature vector. The tensor contraction inner product between the risk feature vector and each row of the risk level quantification matrix is ​​calculated to obtain the raw score for each risk category, which is then normalized to form the multi-dimensional risk propensity coefficient. When generating the multi-dimensional risk propensity coefficient for each patient, each coefficient is compared with the corresponding risk category threshold to determine whether it is in an active state. For active risk categories, the coefficient values ​​are retained, and the corresponding EEG characteristic frequency bands and EMG characteristic patterns are recorded. For inactive risk categories, the coefficients are reset to zero, ultimately obtaining a sparse risk vector composed of non-zero coefficients. This process, while quantifying risk propensity, eliminates insignificant background risk perturbations, making the output more focused on risk dimensions that actually show a worsening trend.

[0012] The probability prediction module is used to input multi-dimensional risk propensity coefficients into an attention-based sequence-to-sequence prediction network, outputting a predicted sequence of the occurrence probability of each risk category in the next nursing cycle. The multi-dimensional risk propensity coefficients from the current time and several consecutive previous time steps are arranged chronologically to form an input feature sequence, which is then fed into an encoder network composed of multiple stacked attention layers. Each attention layer outputs a weighted sequence of intermediate hidden states. The output of the encoder at the last time step is taken as the context semantic vector. This context semantic vector is fed into a decoder network, which generates prediction results step-by-step in an autoregressive manner. At each time step, the decoder generates an original output vector with dimensions equal to the total number of risk categories, which is then processed... The activation function is converted into a probability distribution vector, where each element represents the probability of occurrence of the corresponding risk category at that time step. The index of the risk category corresponding to the maximum probability at that time step is recorded as the predicted risk label. The probability distribution vectors of all time steps are stacked in chronological order to form a probability prediction sequence. By dynamically weighting historical risk tendencies through an attention mechanism, the temporal dependence of risk evolution can be effectively captured, improving the ability to detect progressive neurological deterioration events in advance.

[0013] The early warning triggering module is used to threshold-cut the probability prediction sequence and trigger corresponding risk warning signals. It extracts the probability distribution vector for each time step and determines whether its maximum value is greater than a preset first warning threshold. If it is greater than the first warning threshold, an immediate risk warning signal is output, and the risk category and time step position corresponding to the maximum value are recorded. If it is less than or equal to the first warning threshold, the mean vector of the probability distribution vectors for three consecutive time steps is further calculated, and it is determined whether the maximum value of this mean vector is greater than a preset second warning threshold. If so, a trend risk warning signal is output, where the second warning threshold is less than the first warning threshold. This dual-layer threshold discrimination mechanism can provide immediate alerts for sudden high-risk events and issue warnings for low-amplitude but continuously rising risk trends. It reduces false alarm rates while avoiding missed alarms caused by excessively high single threshold settings, enabling nursing staff to adjust nursing plans and resource allocation in a timely manner according to different warning levels.

[0014] The beneficial effects of this invention are:

[0015] By combining sparse tensor decomposition, the multimodal raw data stream is decomposed into low-rank tensors and sparse tensors. The low-rank tensor serves as the core tensor for risk characteristics, representing the main structure of the patient's neurological state, while the sparse tensor captures outliers and mutation points. Based on this, the coordinates of all non-zero elements in the sparse tensor are traversed, and the original data entries in the multimodal raw data stream are checked against their timestamps. These timestamps are matched with the operation execution times in the nursing operation record table. Entries falling within the nursing operation execution time window are marked as nursing operation-related anomalies, while those falling outside the window are marked as spontaneous physiological anomalies. This yields anomaly residual components that distinguish the source of the anomalies. This approach simultaneously extracts the overall risk characteristics reflecting the patient's neurological state and the fine-grained anomaly information directly related to nursing operation deviations. This allows the warning signals to clearly indicate whether the risk originates from physiological changes in the patient or deviations in the nursing process, providing interpretable decision-making basis for clinicians.

[0016] A sequence-to-sequence prediction network based on an attention mechanism is used to dynamically model multi-dimensional risk propensity coefficients. The risk propensity coefficients from the current and previous consecutive time steps are arranged as an input feature sequence. This sequence passes through an encoder network composed of multiple stacked attention layers to capture long-range dependencies between different time steps and between different risk categories, generating a contextual semantic vector. The decoder network then generates predictions for multiple future time steps in an autoregressive manner. Each time step outputs a probability distribution vector covering all risk categories, ultimately forming a probability prediction sequence for each risk category within the next nursing cycle. This approach fully reveals the dynamic evolution trajectory of risk events within future time windows and the potential for multiple risks to coexist. This transforms early warning from a static judgment at a single moment into a continuous characterization of risk development trends within subsequent nursing cycles, supporting multi-level threshold segmentation on the probability sequence to trigger immediate risk warnings and trend risk warnings respectively. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the structure of a dynamic early warning system for neurological nursing risks based on multimodal data;

[0019] Figure 2 This is a flowchart of the extraction process for the core tensor of risk characteristics and the abnormal residual components;

[0020] Figure 3 This is a flowchart of risk quantification and processing;

[0021] Figure 4 This is a flowchart of sequence-to-sequence risk probability prediction and early warning triggering based on the attention mechanism. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See Figure 1This invention provides a dynamic early warning system for neurological nursing risks based on multimodal data, comprising: a data acquisition module for collecting EEG signals, EMG signals, vital sign time-series data, and historical nursing record text of neurological patients to construct a multimodal raw data stream; a tensor decomposition module for performing joint sparse tensor decomposition on the multimodal raw data stream to extract a core risk feature tensor characterizing the patient's current neurological state and simultaneously separating abnormal residual components used to identify nursing operation deviations; a risk quantization module for performing tensor shrinkage operations based on the core risk feature tensor and a preset risk level quantization matrix to generate a multidimensional risk propensity coefficient for each patient; a probability prediction module for inputting the multidimensional risk propensity coefficient into an attention-based sequence-to-sequence prediction network to output a probability prediction sequence for each risk category in the next nursing cycle; and an early warning triggering module for thresholding the probability prediction sequence to trigger a risk early warning signal of the corresponding level.

[0024] In practice, the data acquisition module acquires EEG signals from neurology patients through a multi-channel EEG acquisition device, acquires EMG signals through EMG acquisition electrodes, acquires vital signs time-series data such as heart rate, respiratory rate, blood oxygen saturation, and blood pressure through a bedside monitor, and extracts historical nursing record texts from the hospital information system.

[0025] The process of constructing the multimodal raw data stream includes: decomposing the EEG signals acquired within the same time window into multiple frequency band sub-signals. Specifically, the EEG signal decomposition involves using a set of zero-phase bandpass filters with passband ranges set to 0.5Hz to 4Hz, 4Hz to 8Hz, 8Hz to 13Hz, 13Hz to 30Hz, and 30Hz to 50Hz, respectively, to obtain delta, theta, alpha, beta, and gamma frequency band sub-signals. For each frequency band sub-signal, the power spectral density is calculated using the Welch method. The Welch method uses a Hanning window as the window function, with a window length of 256 sampling points and an overlap of 128 sampling points, resulting in the power spectral density vector for each frequency band sub-signal. Simultaneously, the root mean square amplitude of the EMG signals within the same time window is calculated using the following formula:

[0026] in, The root mean square amplitude of the electromyographic signal. This indicates the total number of electromyographic signal sampling points within the time window. Indicates the first time within the time window The amplitude of each electromyographic signal sampling point.

[0027] The power spectral density vector of each frequency band sub-signal is aligned with the root mean square amplitude of the electromyography (EMG) signal according to the timestamps recorded by the acquisition device. The alignment process uses the common timestamp sequence of both EEG and EMG signals as a reference, and a timestamp deviation of less than 5 milliseconds is considered aligned. The power spectral density vectors of all aligned frequency band sub-signals are then concatenated with the root mean square amplitude of the EMG signal to form a first-mode tensor slice. The rows of the first-mode tensor slice correspond to different frequency band features and EMG features, while the columns correspond to time points within the time window.

[0028] The processing method for vital sign time-series data involves extracting heart rate, respiratory rate, blood oxygen saturation, and blood pressure sequences within the same time window. For each value in the heart rate sequence, a min-max normalization method is applied. The normalized heart rate value is the original heart rate value minus the minimum heart rate value, then divided by the difference between the maximum and minimum heart rate values. The same normalization operation is performed on the respiratory rate, blood oxygen saturation, and blood pressure sequences. The normalized heart rate, respiratory rate, blood oxygen saturation, and blood pressure sequences are then stacked vertically to form a second modality tensor slice. The rows of the second modality tensor slice correspond to the four dimensions of heart rate, respiratory rate, blood oxygen saturation, and blood pressure, while the columns correspond to the time points within the time window.

[0029] The processing method for historical nursing records involves collecting all nursing records from the patient within the most recent six months prior to this admission. Each record is segmented, stop words are removed, and a pre-trained continuous bag-of-words model is used to convert each word into a semantic feature vector. The vector dimension of the continuous bag-of-words model is set to 200 dimensions. The arithmetic mean of the semantic feature vectors of all words in a nursing record is calculated to obtain the text representation vector of that record. The arithmetic mean of the text representation vectors of all historical nursing records is then calculated again to obtain a 200-dimensional global historical semantic vector. This global historical semantic vector is then copied and expanded along the time dimension. The expansion length is the same as the number of time points in the first and second modal tensor slices, forming a third modal tensor slice. The rows of the third modal tensor slice correspond to each dimension of the semantic feature vector, and the columns correspond to each time point within the time window, with each column having identical values.

[0030] The first modality tensor slice, the second modality tensor slice, and the third modality tensor slice are concatenated along the same time dimension. The concatenation is performed along the feature dimension direction, and the concatenation generates a multimodal raw data stream. The multimodal raw data stream is a third-order tensor with three dimensions: time, modality type, and feature channel, respectively.

[0031] In practice, the tensor decomposition module receives the multimodal raw data stream output by the data acquisition module and performs joint sparse tensor decomposition.

[0032] The multimodal raw data stream is represented as a third-order observation tensor, denoted as . ,in This indicates the size of the time dimension, corresponding to the number of timestamps in the original multimodal data stream. This indicates the size of the modality type dimension, with a value of 3, corresponding to the EEG / EMG fusion modality, vital signs modality, and nursing record text modality, respectively. This indicates the size of the feature channel dimension, corresponding to the total number of feature channels across all modalities.

[0033] Introduce a low-rank tensor denoted as low-rank tensor Dimensions and third-order observation tensor The dimensions are the same, for low-rank tensor Used for approximating third-order observation tensors The main structure reflects the stable and regular components in the multimodal raw data stream. A sparse tensor, denoted as [missing information], is also introduced. sparse tensor The dimension is also the same as the third-order observation tensor. The dimensions are the same, for sparse tensor Used to capture third-order observation tensors Outliers and mutation points in sparse tensors are represented by sparse tensors. The non-zero elements in.

[0034] The joint sparse tensor decomposition problem is formalized as solving the following optimization problem, the objective of which is to optimize the low-rank tensors. Low-rank constraints and sparse tensors Minimize the sum of sparsity constraints while ensuring a low-rank tensor. With sparse tensors The sum and the third-order observation tensor The reconstruction error is less than a preset threshold. The formula for calculating the reconstruction error is:

[0035]

[0036] in, Represents the reconstruction error function. Represents the third-order observation tensor. This represents a low-rank tensor. Represents a sparse tensor. Describe the Frobenius norm. This represents the square of the Frobenius norm.

[0037] The above optimization problem is solved using the alternating direction multiplier method. The execution process of the alternating direction multiplier method includes: initializing the low-rank tensor. Initialize sparse tensors as zero tensors Initialize the Lagrange multiplier tensor to zero. Initialize the penalty coefficient as a zero tensor. The value of is 1.0, the maximum number of initial iterations is 1000, and the initial convergence threshold is . In each iteration, the sparse tensor is fixed. and Lagrange multiplier tensor Update low-rank tensors The update method is to calculate the third-order observation tensor. Subtract sparse tensors Add Lagrange multiplier tensors Divide by penalty coefficient The result is then subjected to singular value decomposition, and a soft thresholding operation is performed on the singular value vectors obtained from the singular value decomposition. The threshold for the soft thresholding operation is set as a penalty coefficient. The reciprocal of the result of the soft thresholding operation is used to reconstruct the result into a low-rank tensor. Then fix the low-rank tensor. and Lagrange multiplier tensor Update sparse tensors The update method is to calculate the third-order observation tensor. Subtract low-rank tensors Add Lagrange multiplier tensors Divide by penalty coefficient As a result, a soft thresholding operation is performed on each element in the calculation result, with the threshold for the soft thresholding operation also set as a penalty coefficient. The reciprocal of the product yields the sparse tensor. Then update the Lagrange multiplier tensor. The update method is the Lagrange multiplier tensor. Add penalty coefficient Multiply by the third-order observation tensor Subtract low-rank tensors Subtract the sparse tensor The difference. Update the penalty coefficient. The current penalty coefficient Multiply by an increment factor, set to 1.1. Calculate the reconstruction error for the current iteration. If the reconstruction error is less than the preset threshold If the number of iterations reaches the maximum number of iterations (1000), then the iteration will stop.

[0038] The low-rank tensor after iterative convergence This serves as the core tensor for risk characteristics. The low-rank tensor at this point... It preserves the main structure of the original multimodal data stream, removes noise and abnormal interference, and can characterize the stable features of the patient's current neurological state. The sparse tensor after iterative convergence... As an anomalous residual component. Sparse tensor Non-zero elements in the data indicate outliers or mutations at the corresponding time, modality type, and feature channel location.

[0039] In specific implementation, please refer to Figure 2 After the tensor decomposition module completes the joint sparse tensor decomposition, it processes the low-rank tensor and sparse tensor respectively to extract the core tensor of risk features and the abnormal residual components.

[0040] The process of extracting the core risk feature tensor to characterize the patient's current neurological state involves performing Tucker decomposition along the modality type dimension on the iteratively converged low-rank tensor. The low-rank tensor is represented as... The modality type dimension is the second dimension, and its size is [missing value]. Tucker decomposition transforms low-rank tensors Decomposed into a kernel tensor and three factor matrices, the decomposition form is as follows: The kernel tensor is denoted as... ,in Denotes the decomposition rank in the time dimension. The value can be the size of the time dimension. The product of the product with the preset compression ratio of 0.5 and rounded down. The decomposition rank of the modality type dimension. The value can be the modal type dimension. , Represents the decomposition rank of the feature channel dimension. The value is the size of the feature channel dimension. The product of the product with the preset compression ratio of 0.3 and rounded down. The first factor matrix is ​​denoted as... The first factor matrix corresponds to the time dimension, and each column of the first factor matrix represents a latent pattern in the time dimension. The second factor matrix is ​​denoted as... The second factor matrix corresponds to the modality type dimension, and each column of the second factor matrix represents a latent pattern in the modality type dimension. The third factor matrix is ​​denoted as... The third factor matrix corresponds to the feature channel dimension, and each column of the third factor matrix represents a latent pattern in the feature channel dimension.

[0041] kernel tensor The elements in the kernel tensor reflect the interaction strength between potential patterns in different dimensions. The elements with the largest absolute values ​​are retained, and the number of retained elements depends on the kernel tensor. The total number of elements and the preset sparsity are determined. The preset sparsity is set to 0.15, based on a 10-fold cross-validation search within the range of 0.05 to 0.5 with a step size of 0.05. 0.15 represents a trade-off between reconstruction error and sparsity. The kernel tensor... Sort all elements in the kernel tensor by absolute value from largest to smallest, retain the top 15% of elements after sorting, and set the remaining 85% of elements to zero, forming a sparse kernel tensor, denoted as . .

[0042] sparsify kernel tensor With the first factor matrix The second factor matrix and the third factor matrix Perform refactored multiplication; the refactored multiplication operation is as follows: Generate a denoised risk feature tensor, denoted as The denoised risk feature tensor Residual noise introduced by low-contribution components was removed from the low-rank tensor. From the denoised risk feature tensor... Extract the last time slice, which corresponds to the end of the time dimension and represents the feature state at the current moment. The extraction operation uses a fixed time dimension index. By retaining all values ​​of the modality type dimension and feature channel dimension, a two-dimensional matrix is ​​obtained, which is then used as the core tensor of the risk features.

[0043] The process of simultaneously separating anomalous residual components used to identify deviations in nursing procedures includes: traversing sparse tensors The system retrieves the position coordinates of all non-zero elements, presented as triples containing time, modality type, and feature channel coordinates. For each non-zero element, the system looks up the corresponding original data entry in the multimodal raw data stream based on its time and modality type coordinates. Each original data entry is a specific multimodal data record. The timestamp of the original data entry is then matched with the operation execution time in a pre-stored nursing operation record table, which is located in the hospital information system and contains a nursing operation name field and an operation execution time field.

[0044] If the timestamp of the original data entry falls within the execution time window of any nursing operation, the non-zero element is marked as a nursing operation-related anomaly. The execution time window is defined as the time interval from 10 minutes before the operation execution time to 30 minutes after the operation execution time, and the execution time window is set based on the timeliness research results of the impact of clinical nursing operations on patients' physiological indicators. If the timestamp of the original data entry falls outside the execution time window of all nursing operations, the non-zero element is marked as a spontaneous physiological anomaly. The set of all marked non-zero elements and their position coordinates is taken as the anomalous residual component. Nursing operation-related anomalies in the anomalous residual component indicate that the anomaly is caused by the nursing operation, while spontaneous physiological anomalies indicate that the anomaly is caused by the patient's own physiological changes.

[0045] See Figure 3 The process of performing tensor condensation operation based on the core risk characteristic tensor and the preset risk level quantization matrix includes: obtaining the pre-constructed risk level quantization matrix, denoted as [missing information]. The row indices of the risk level quantification matrix correspond to different risk categories, with a total of [number missing] risk categories. kind, The value is 7, representing the risks of falls, cerebral hemorrhage, seizures, aspiration, pressure sores, deep vein thrombosis, and delirium, respectively. The column indices of the risk level quantification matrix correspond to the channels of the risk feature core tensor, with a total of [number missing] channels. , The value of is consistent with the dimension of the feature channels of the core risk feature tensor. Each element in the risk level quantification matrix Indicates the first Risk category in The quantization weights on each feature channel are real numbers ranging from 0 to 1. The risk level quantization matrix is ​​pre-constructed as follows: historical neurology patient data samples are collected, each data sample is manually labeled with a risk category, a linear discriminant analysis algorithm is used to solve for the projected weight vector that maximizes the separability of risk categories, and the obtained projected weight vectors are arranged by category to form the risk level quantization matrix.

[0046] Expand the core risk feature tensor along the channel dimension into a risk feature vector. The core risk feature tensor is of dimension 1. A two-dimensional matrix, Indicates the size of the modality type dimension. This indicates the size of the feature channel dimension. The unfolding operation concatenates the core risk feature tensor column-wise into a one-dimensional vector with a length of [value missing]. , denoted as risk feature vector .

[0047] Calculate the risk feature vector Risk level quantification matrix The tensor shrinking inner product between each row yields the raw score for each risk category. Each row of the risk level quantization matrix represents a weight vector for a risk category, denoted as . ,in For risk category indexing, The value range is 1 to Integers. The formula for calculating the tensor contraction inner product is:

[0048] in, Indicates the first The raw scores corresponding to each risk category Indicates the size of the modality type dimension. Indicates the size of the feature channel dimension. This represents the total length of the risk feature vector. Represents risk feature vector In the The element value at each position, Represents a risk level quantification matrix The Middle Line 1 The element value of the column represents the first element. The risk category is in the first Quantization weights at each position.

[0049] The raw scores for each risk category are normalized to obtain a multi-dimensional risk propensity coefficient. The normalization process employs... Function for all The original fraction vector composed of the original fractions is transformed, and the result of the transformation is the first... The value of the risk category is the number of... The risk propensity coefficient for each risk category. The multidimensional risk propensity coefficient is a... A dimensional vector, where each element has a value between 0 and 1, and the sum of all elements equals 1.

[0050] The process of generating multidimensional risk propensity coefficients for each patient includes comparing each coefficient in the multidimensional risk propensity coefficients with its corresponding risk category threshold. The risk category threshold is a pre-defined discrimination boundary for each risk category. The risk category threshold is set as follows: Historical neurology patient data samples are collected and the actual occurrence of each risk category is labeled; the receiver operating characteristic (ROC) curve for the risk propensity coefficient is calculated for each risk category; and the risk propensity coefficient corresponding to the maximum Youden's index on the ROC curve is selected as the risk category threshold for that risk category. The risk category threshold for each risk category is denoted as .

[0051] The first of the multi-dimensional risk propensity coefficients The coefficient and the first Risk category thresholds Compare and determine the first Is each risk category active? If the first risk category is active... The coefficient is greater than the first. Risk category thresholds Then determine the first Several risk categories are in an active state. For risk categories identified as active, the risk propensity coefficient value for that risk category is recorded. Simultaneously, the corresponding EEG and EMG characteristic frequency bands are retrieved and recorded from the mapping table of EEG characteristic frequency bands and EMG characteristic patterns. EEG characteristic frequency bands include delta, theta, alpha, beta, and gamma bands; EMG characteristic patterns include amplitude abnormality patterns, frequency abnormality patterns, and co-contraction abnormality patterns.

[0052] If the first The coefficient is less than or equal to the first coefficient. Risk category thresholds Then determine the first Each risk category is inactive. For risk categories deemed inactive, the corresponding risk propensity coefficient is reset to zero. All risk propensity coefficients that remain non-zero after the reset are arranged in their original order to form a sparse risk vector, which represents the multi-dimensional risk propensity coefficient for each patient. The sparse risk vector retains only the risk propensity coefficients of active risk categories; the values ​​in the remaining positions are zero.

[0053] In specific implementation, please refer to Figure 4 The probability prediction module receives the multi-dimensional risk propensity coefficient for each patient from the risk quantification module, performs sequence-to-sequence prediction, and outputs the predicted probability sequence for each risk category in the next nursing cycle. The early warning triggering module performs threshold segmentation on the predicted probability sequence and triggers the corresponding level of risk warning signal.

[0054] The process of inputting multidimensional risk propensity coefficients into an attention-based sequence-to-sequence prediction network includes: arranging the multidimensional risk propensity coefficients from the current time step and several consecutive previous time steps in chronological order to form an input feature sequence. The multidimensional risk propensity coefficients are sparse risk vectors with a dimension of 7, corresponding to seven risk categories: fall risk, cerebral hemorrhage risk, epileptic seizure risk, aspiration risk, pressure sore risk, deep vein thrombosis risk, and delirium risk. The current time step is denoted as time step index 0, and the number of consecutive previous time steps is set to 48, corresponding to data from the past 48 hours, collected hourly. The input feature sequence is a sequence of length 48, where each time step is a 7-dimensional multidimensional risk propensity coefficient vector.

[0055] The input feature sequence is fed into the encoder network. The encoder network consists of multiple stacked attention layers, with a total of 6 attention layers. Each attention layer contains a multi-head self-attention sub-layer and a fully connected feedforward sub-layer. The multi-head self-attention sub-layer has 8 attention heads, each with a dimension of 64. The hidden layer dimension of the fully connected feedforward sub-layer is set to 2048, and the activation function is a rectified linear unit. Each attention layer performs the following operations on the input sequence: calculates the attention weight matrix across all time steps in the input sequence. The attention weight matrix is ​​calculated by scaling the product of the transpose of the query matrix and the key matrix using a scaling factor equal to the square root of the attention head dimension. The scaling dot product result is then applied... The function yields attention weights, which are multiplied by the value matrix to obtain a weighted intermediate hidden state sequence. A fully connected feedforward sublayer performs two linear transformations on the vector at each time step of the weighted intermediate hidden state sequence, applying a rectified linear unit activation function between the two transformations. Residual connections and layer normalization are added after each sublayer. Each attention layer outputs a weighted intermediate hidden state sequence with the same length as the input sequence, and the vector dimension at each time step is 512. The next attention layer takes the weighted intermediate hidden state sequence output by the previous attention layer as input and continues to perform the same operation until the last attention layer outputs the final weighted intermediate hidden state sequence.

[0056] The output of the last time step of the encoder network is used as the context semantic vector. The vector dimension of the last time step of the weighted intermediate hidden state sequence output by the encoder network is 512. This 512-dimensional vector is extracted as the context semantic vector, which encodes the global information of all time steps in the input feature sequence.

[0057] The context semantic vector is fed into the decoder network. The decoder network generates predictions time-step by time in an autoregressive manner. The decoder network consists of multiple stacked attention layers, with a total of 6 attention layers. Each attention layer in the decoder network contains a masked multi-head self-attention sublayer, an encoder-decoder attention sublayer, and a fully connected feedforward sublayer. The masked multi-head self-attention sublayer has 8 attention heads, each with a dimension of 64. The masked multi-head self-attention sublayer calculates attention weights for the sequences already generated by the decoder, using a lower triangular mask matrix to mask information from future time steps during the calculation. The encoder-decoder attention sublayer uses the output of the masked multi-head self-attention sublayer as the query matrix and the weighted intermediate hidden state sequence output by the encoder network as the key and value matrices to calculate cross-sequence attention weights. The hidden layer dimension of the fully connected feedforward sublayer is set to 2048, and the activation function is a rectified linear unit. Residual connections and layer normalization operations are added after each sublayer. The output vector of the decoder network at each time step has a dimension of 512.

[0058] The process of outputting the predicted probability sequence of each risk category for the next care cycle includes: the decoder network generating an original output vector at each time step. A fully connected output layer is then connected after the last layer of the decoder network. The weight matrix of the fully connected output layer has dimensions of [dimension missing]. The bias vector dimension is , The total number of risk categories. The value is 7. The fully connected output layer maps the 512-dimensional vector output by the decoder network to an original output vector with a dimension of 7.

[0059] Apply to the original output vector The activation function transforms it into a probability distribution vector. The formula for the activation function is:

[0060]

[0061] in, Represents the probability distribution vector of the th The element, i.e., the th element The probability of occurrence of each risk category at this time step Represents the first element in the original output vector. One element, This represents the total number of risk categories. The value is 7. Represented by natural constant An exponential function with base 0. Indicates all The probability distribution vector is the sum of the exponential values ​​of the original output values ​​for each risk category. Each element in the vector is a real number between 0 and 1, and the sum of all elements equals 1.

[0062] The risk category index with the highest probability value in the probability distribution vector is recorded as the predicted risk label for that time step. Risk category indices are integers from 1 to 7: index 1 corresponds to fall risk, index 2 to cerebral hemorrhage risk, index 3 to seizure risk, index 4 to aspiration risk, index 5 to pressure sore risk, index 6 to deep vein thrombosis risk, and index 7 to delirium risk. The probability distribution vectors generated from all time steps are stacked in chronological order to form an occurrence probability prediction sequence. The next nursing cycle is 24 hours long, the decoder network generates 24 time steps, and the occurrence probability prediction sequence is a 24x7 matrix, where each row corresponds to the probability distribution vector of one time step, and each column corresponds to a risk category.

[0063] In practice, the training process of the attention-based sequence-to-sequence prediction network includes: collecting multi-dimensional risk propensity coefficient sequence data of historical neurology patients. Each sample contains a continuous 72-hour multi-dimensional risk propensity coefficient sequence as input and a subsequent 24-hour actual risk category label sequence as output. The training set contains no fewer than 10,000 samples. During training, the cross-entropy loss function is used. The cross-entropy loss is calculated between the predicted probability distribution and the true risk category label at each time step of the decoder, and the sum of the cross-entropy losses at all time steps is used as the total loss. The optimizer uses the Adam optimizer, with an initial learning rate of [value missing].

[0064] The batch size is set to 64. The number of training iterations is set to 200. After each training iteration, the loss value on the validation set is calculated. If the validation set loss value does not decrease for 10 consecutive iterations, the learning rate is decayed to 0.5 times the current learning rate. If the validation set loss value does not decrease for 20 consecutive iterations, training is terminated early.

[0065] In practice, the process of thresholding the occurrence probability prediction sequence includes: extracting the probability distribution vector corresponding to each time step in the occurrence probability prediction sequence. The occurrence probability prediction sequence contains 24 time steps, each time step corresponding to a probability distribution vector, and the dimension of the probability distribution vector is 7. For the probability distribution vector of each time step, it is determined whether the maximum value in the probability distribution vector is greater than a preset first warning threshold. The first warning threshold is set to 0.85. The basis for setting the first warning threshold is based on the system's requirement for balancing sensitivity and specificity, by analyzing the distribution of probability values ​​of all risk categories in the historical dataset, and selecting the probability value corresponding to the 85th quantile of the cumulative distribution function.

[0066] If the maximum value in the probability distribution vector is greater than the first warning threshold, an immediate risk warning signal is immediately output, and the risk category and time step position corresponding to the maximum value in the probability distribution vector are recorded. The immediate risk warning signal includes a risk category field, a time step field, and a warning level field, with the warning level field marked as Level 1. If the maximum value in the probability distribution vector is less than or equal to the first warning threshold, the mean vector of the probability distribution vectors of three consecutive time steps in the probability prediction sequence is further calculated. The three consecutive time steps are selected as follows: starting from the current time step, two consecutive time steps are taken, and the probability distribution vectors of the three time steps are calculated element-wise as an arithmetic mean vector. The dimension of the mean vector is 7, and each element in the mean vector represents the average probability of occurrence of the corresponding risk category within the three consecutive time steps.

[0067] Determine if the maximum value in the mean vector exceeds a preset second warning threshold. The second warning threshold is set to 0.60, which is less than the first warning threshold. The second warning threshold is set based on the system's sensitivity requirements for potential trend risks. It is determined by analyzing the continuous upward pattern of risk categories in historical data over time series, selecting the minimum probability value that can detect an upward trend in risk three time steps in advance. If the maximum value in the mean vector exceeds the second warning threshold, a trend risk warning signal is output. This signal includes a risk category field, a time step field, and a warning level field, with the warning level field marked as Level 2.

[0068] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A dynamic early warning system for neurological nursing risks based on multimodal data, characterized in that, include: The data acquisition module is used to collect EEG signals, EMG signals, vital signs time-series data, and historical nursing record texts from neurology patients to construct a multimodal raw data stream; The tensor decomposition module is used to perform joint sparse tensor decomposition on the multimodal raw data stream, extract the risk feature core tensor used to characterize the patient's current neurological state, and simultaneously separate the abnormal residual components used to identify nursing operation deviations. The risk quantification module is used to perform tensor shrinkage operation based on the core risk feature tensor and the preset risk level quantification matrix to generate a multi-dimensional risk propensity coefficient for each patient. The probability prediction module is used to input the multi-dimensional risk propensity coefficient into an attention-based sequence-to-sequence prediction network and output the probability prediction sequence of each risk category in the next nursing cycle. The early warning triggering module is used to perform threshold segmentation on the occurrence probability prediction sequence and trigger a risk warning signal of the corresponding level.

2. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 1, characterized in that, The construction of the multimodal raw data stream includes: The EEG signals collected within the same time window are decomposed into multiple frequency band sub-signals, and the power spectral density of each frequency band sub-signal is timestamped with the root mean square amplitude of the EMG signal to form a first mode tensor slice. The vital signs time series data were normalized according to four dimensions: heart rate, respiratory rate, blood oxygen saturation, and blood pressure, and then stacked into second modality tensor slices. The historical nursing record text is converted into a semantic feature vector sequence through a word embedding model and compressed into a fixed-length third-modality tensor slice; The first modal tensor slice, the second modal tensor slice, and the third modal tensor slice are concatenated along the same time dimension to generate the multimodal raw data stream.

3. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 2, characterized in that, Performing joint sparse tensor decomposition on the multimodal raw data stream includes: The multimodal raw data stream is represented as a third-order observation tensor, where the three dimensions correspond to time, modality type, and feature channel, respectively. A low-rank tensor is introduced to approximate the main structure of the third-order observation tensor, while a sparse tensor is introduced to capture outliers and mutation points in the third-order observation tensor. The low-rank tensor and the sparse tensor are iteratively updated using the alternating direction multiplier method until the reconstruction error is less than a preset threshold. The low-rank tensor after iterative convergence is used as the core tensor of the risk feature, and the sparse tensor after iterative convergence is used as the abnormal residual component.

4. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 3, characterized in that, The extracted risk feature core tensor used to characterize the patient's current neurological state includes: Perform Tucker decomposition on the low-rank tensor along the modality type dimension to obtain the kernel tensor and three factor matrices, where the first factor matrix corresponds to the time dimension, the second factor matrix corresponds to the modality type dimension, and the third factor matrix corresponds to the feature channel dimension. The kernel tensor retains the elements with the largest absolute values ​​and sets the remaining elements to zero to form a sparsified kernel tensor. The sparse kernel tensor is reconstructed and multiplied with the three factor matrices to generate the denoised risk feature tensor; Extract the last time slice from the denoised risk feature tensor as the core risk feature tensor.

5. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 4, characterized in that, The synchronous separation of abnormal residual components used to identify deviations in nursing procedures includes: Traverse the position coordinates of all non-zero elements in the sparse tensor, and look up the corresponding original data entries in the multimodal original data stream based on the position coordinates. The timestamps of the original data entries are matched with the operation execution times in the pre-stored nursing operation record table; If the timestamp of the original data entry falls within the execution time window of any nursing operation, then the non-zero element is marked as a nursing operation association anomaly. If the timestamp of the original data entry falls outside the execution time window of any nursing operation, the non-zero element is marked as a spontaneous physiological abnormality. The set consisting of all marked non-zero elements and their position coordinates is taken as the abnormal residual component.

6. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 1, characterized in that, The tensor reduction operation based on the core risk feature tensor and the preset risk level quantization matrix includes: Obtain a pre-constructed risk level quantification matrix, wherein the row index of the risk level quantification matrix corresponds to different risk categories, and the column index corresponds to each channel of the risk feature core tensor; Expand the core risk feature tensor along the channel dimension into a risk feature vector; Calculate the tensor shrink inner product between the risk feature vector and each row of the risk level quantization matrix to obtain the original score corresponding to each risk category; The original scores corresponding to each risk category are normalized to obtain the multi-dimensional risk propensity coefficient.

7. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 6, characterized in that, The generation of a multi-dimensional risk propensity coefficient for each patient includes: Each coefficient in the multi-dimensional risk propensity coefficient is compared with the corresponding risk category threshold to determine whether each risk category is in an active state. For risk categories that are identified as active, record the value of their risk propensity coefficient, as well as the corresponding EEG characteristic frequency bands and EMG characteristic patterns. For risk categories that are determined to be inactive, their risk propensity coefficient is reset to zero. All risk propensity coefficients that remain non-zero after resetting are combined into a sparse risk vector, which serves as the multidimensional risk propensity coefficient for each patient.

8. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 1, characterized in that, The step of inputting the multi-dimensional risk propensity coefficient into an attention-based sequence-to-sequence prediction network includes: The multi-dimensional risk tendency coefficients of the current time and several consecutive previous times are arranged in chronological order to form an input feature sequence; The input feature sequence is fed into an encoder network, which consists of multiple stacked attention layers, each of which outputs a weighted sequence of intermediate hidden states. The output of the last time step of the encoder network is used as the context semantic vector; The context semantic vector is fed into the decoder network, which generates prediction results step by step in an autoregressive manner.

9. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 8, characterized in that, The output includes the probability prediction sequence for each risk category in the next nursing cycle: The decoder network generates an original output vector at each time step, and the dimension of the original output vector is equal to the total number of risk categories; The original output vector is converted into a probability distribution vector by applying the softmax activation function, where each element represents the probability of occurrence of the corresponding risk category at that time step; Record the risk category index with the highest probability value in the probability distribution vector as the predicted risk label for that time step; The probability distribution vectors generated at all time steps are stacked in chronological order to form the occurrence probability prediction sequence.

10. The dynamic early warning system for neurological nursing risks based on multimodal data according to claim 9, characterized in that, The thresholding of the occurrence probability prediction sequence includes: Extract the probability distribution vector corresponding to each time step in the probability prediction sequence, and determine whether the maximum value in the probability distribution vector is greater than a preset first warning threshold; If the maximum value is greater than the first warning threshold, an immediate risk warning signal is immediately output, and the risk category and time step position corresponding to the maximum value are recorded. If the maximum value is less than or equal to the first warning threshold, then the mean vector of the probability distribution vectors of three consecutive time steps in the occurrence probability prediction sequence is further calculated. Determine whether the maximum value in the mean vector is greater than a preset second warning threshold. If so, output a trend risk warning signal, wherein the second warning threshold is less than the first warning threshold.

Images