Method and system for monitoring signs of elderly patients based on multi-modal perception

Abstract

The application provides a kind of old patient sign monitoring method and system based on multi-modal perception, it is related to intelligent health monitoring technical field, including: acquisition and fusion physiological and behavioral data, construct unified representation and modal correlation matrix.Baseline features are calculated and tensor fusion is carried out, abnormal fragments are identified and dominant influence factors are traced back by encoding reconstruction.Risk level is mapped by deducing data evolution trajectory and calculating deviation from health benchmark.Simulate the timeliness of intervention scheme, determine the target scheme and generate monitoring report.The application realizes the deep correlation analysis and early warning of abnormality of multi-modal data, can automatically determine the risk dominant factor and generate individualized intervention scheme, improves the accuracy and initiative of old health monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent health monitoring technology, and in particular to a method and system for monitoring the vital signs of elderly patients based on multimodal perception. Background Technology

[0002] In the field of elderly health monitoring, existing technologies typically collect heart rate, blood oxygen saturation, or body movement data through wearable devices, aiming to identify obvious abnormal events, such as excessively high heart rate or prolonged periods of stillness, by setting fixed thresholds or simple trend analysis.

[0003] However, conventional monitoring methods rely on single-modal data, making it difficult to comprehensively and accurately reflect the complex overall health status of elderly patients. Changes in the vital signs of elderly patients are often the result of multiple coupled factors. An abnormality in a single signal source may be caused by multiple reasons, or the overall condition may already indicate risk even when multiple signals have not triggered an alarm individually. Conventional technical means cannot effectively capture the deep correlations and synergistic change patterns between different physiological and behavioral signals, resulting in insufficient early identification of potential risks and frequent false alarms and missed alarms.

[0004] When a monitoring system identifies anomalies, existing intervention decision-making logic typically employs a pre-defined static rule base, directly mapping specific abnormal signals to fixed intervention recommendations. This lacks in-depth analysis of the dynamic evolution of abnormal events and the dominant factors, and fails to consider individual differences and specific contexts. Consequently, the generated intervention plans may lack specificity, making it difficult to effectively block risks in their early stages, and also unable to predict the potential subsequent impacts of the intervention measures themselves. Summary of the Invention

[0005] This invention provides a method and system for monitoring the vital signs of elderly patients based on multimodal perception, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of this invention provides a method for monitoring vital signs in elderly patients based on multimodal sensing, comprising: Physiological signal data and behavioral state data of elderly patients are collected and feature extraction is performed to obtain physiological feature vectors and behavioral feature vectors. After aligning the physiological feature vectors and behavioral feature vectors, multi-layer nonlinear transformation is performed and projected onto a common semantic space to obtain a unified representation vector and construct a modal correlation matrix. The unified representation vector is segmented by a sliding time window and its statistical distribution characteristics are calculated to obtain baseline features. The baseline features are then fused with the modal correlation matrix using tensor fusion to obtain fused state features. The fused state features are encoded, compressed, decoded, and reconstructed, and the reconstruction error is calculated to identify abnormal state segments. The abnormal state segments are then backtracked to extract precursor change patterns and combined with the modal correlation matrix to determine the dominant influencing factors. Based on the dominant influencing factor, a subset of data is extracted from the unified characterization vector and iteratively extrapolated over a continuous time scale to determine the evolution trajectory. The deviation between the evolution trajectory and the preset health benchmark template is calculated, and a graded mapping is performed based on the deviation to obtain the risk level. Based on the risk level and the dominant influencing factor, a set of candidate intervention schemes is determined by querying a preset intervention strategy library. For each candidate intervention scheme, a multi-step state transition simulation is performed and a timeliness score is calculated. Based on the timeliness score, a target intervention scheme is determined and a monitoring report is generated by combining the abnormal state fragments.

[0007] In one alternative implementation, Physiological signal data and behavioral state data of elderly patients are collected and feature extracted to obtain physiological feature vectors and behavioral feature vectors. These physiological and behavioral feature vectors are then aligned, subjected to multi-level nonlinear transformation, and projected onto a common semantic space to obtain a unified representation vector. A modal correlation matrix is ​​then constructed, including: Physiological signal data and behavioral state data of elderly patients are acquired. The physiological signal data is decomposed into time-frequency data to extract frequency domain energy distribution features. The mean and variance of the physiological signal data are calculated to obtain time domain statistical features, which are then combined with the frequency domain energy distribution features to obtain a physiological feature vector. The behavioral state data is analyzed to extract joint angle sequences. The position change rate of the behavioral state data is calculated to obtain motion trajectory features, which are then combined with the joint angle sequences to obtain a behavioral feature vector. The timestamp information of the physiological feature vector and the behavioral feature vector is extracted. The time offset is calculated based on the timestamp information and linear interpolation and feature alignment are performed. The aligned physiological feature vector and behavioral feature vector are subjected to nonlinear transformation to obtain physiological intermediate representation and behavioral intermediate representation respectively. The physiological intermediate representation and the behavioral intermediate representation are concatenated and subjected to nonlinear transformation to obtain cross-modal intermediate representation. The cross-modal intermediate representation is subjected to nonlinear transformation and normalized after hyperbolic tangent activation to obtain a unified representation vector. The Pearson correlation coefficients between the physiological and behavioral feature components of the unified representation vector at different times are calculated and arranged according to modality type and time order to obtain the correlation matrix. The correlation matrix is ​​then thresholded to obtain the modality correlation matrix.

[0008] In one alternative implementation, The baseline features are obtained by segmenting the unified representation vector into sliding time windows and calculating its statistical distribution characteristics. The fused state features are then obtained by performing tensor fusion operations on the baseline features and the modal correlation matrix. The unified representation vector is divided into multiple time window segments according to the preset window length and sliding step size. The central trend index and dispersion index are calculated for the unified representation vector in each time window segment to obtain the distribution concentration feature. The distribution symmetry index and distribution kurtosis index are calculated for the unified representation vector in each time window segment to obtain the distribution morphology feature. The distribution concentration feature and the distribution morphology feature are spliced ​​together in the order of the time window to obtain the baseline feature. The baseline features are arranged into a baseline feature tensor according to the time window and feature dimension. The modal correlation matrix is ​​expanded into a correlation tensor according to the modal type and time order. Tensor contraction operation is performed on the baseline feature tensor and the correlation tensor in the time dimension to obtain a time fusion intermediate tensor. Tensor expansion is performed on the time fusion intermediate tensor in the feature dimension to obtain an expanded feature matrix. The expanded feature matrix is ​​decomposed to extract the dominant feature components. The dominant feature components are reconstructed and combined with the baseline feature tensor to calculate the enhanced baseline tensor. The enhanced baseline tensor and the associated tensor are then subjected to tensor outer product operation to obtain the fused state features.

[0009] In one alternative implementation, After encoding, compressing, decoding, and reconstructing the fused state features, the reconstruction error is calculated to identify anomalous state segments. Backtracking analysis is performed on these anomalous state segments to extract precursor change patterns, and the dominant influencing factors are determined by combining these patterns with the modal correlation matrix. The fusion state features are subjected to multi-level nonlinear dimensionality reduction transformation and the mean vector and variance vector in the latent space are calculated. Based on the mean vector and variance vector, reparameterized sampling is performed to obtain a compressed coding representation. The divergence measure between the compressed coding representation and the preset prior distribution is calculated and multi-level nonlinear dimensionality increase transformation is performed to obtain the reconstructed fusion state features. The reconstruction loss value between the reconstructed fusion state features and the fusion state features is calculated and a reconstruction error sequence is constructed. The time period in the reconstruction error sequence where the reconstruction loss value exceeds a preset threshold is marked and clustered and merged to obtain abnormal state segments. Extract the abnormal start time from the abnormal state segment and backtrack to determine the backtracking time window. Extract the fusion state features within the backtracking time window and calculate the feature differences between adjacent time steps to construct a change amplitude sequence. Smooth the change amplitude sequence and identify local maxima as key transition nodes. Extract the fusion state features and corresponding change directions at the time corresponding to the key transition nodes to obtain the precursor change pattern. The Pearson correlation coefficient corresponding to the anomaly initiation time is extracted from the modal correlation matrix, and the modal volatility is calculated. The feature component with the largest modal volatility is selected, and the corresponding modal type and feature dimension information are extracted to obtain the dominant influencing factor.

[0010] In one alternative implementation, Based on the dominant influencing factors, a subset of data is extracted from the unified characterization vector, and the evolutionary trajectory is determined through iterative deduction on a continuous time scale. The deviation between the evolutionary trajectory and the preset health benchmark template is calculated, and a graded mapping is performed based on the deviation to obtain the risk level, including: Based on the dominant influencing factor, time series data corresponding to the feature dimensions are extracted from the unified characterization vector to obtain an initial data subset. The initial data subset is then decomposed into a time series to obtain trend components and calculate the rate of change. Based on the rate of change, abrupt change points are identified and time periods are divided at the abrupt change points to determine the stable evolution stage, thus obtaining the data subset. The data subset is discretized and sampled according to the time step to obtain the initial time state vector. The state transition rule is determined based on the dominant influencing factor, and the initial time state vector is iteratively transformed. The iterative deduction is repeated on a continuous time scale to obtain the predicted state vector sequence and smooth it to obtain the evolution trajectory. A baseline state vector sequence is extracted from a preset health baseline template. The Euclidean distance between the predicted state vector in the evolution trajectory and the baseline state vector in the baseline state vector sequence is calculated to obtain a time step deviation sequence. The time step deviation sequence is then weighted and summed to obtain a comprehensive deviation. Based on the comprehensive deviation and a preset multi-level threshold interval, interval matching is performed to map to a discrete risk level. The risk level is determined based on the discrete risk level and the number of rate of change mutation points.

[0011] In one alternative implementation, The initial data subset is decomposed into a time series to obtain trend components and calculate the rate of change. Based on the rate of change, abrupt change points are identified, and time periods are divided at these abrupt change points to determine stable evolution stages. The resulting data subset includes: The initial data subset is sorted by time dimension to construct a time index sequence and segmented by sliding window to obtain multiple local time windows. The data in each local time window is fitted with a polynomial to obtain a local trend curve and connected in segments to obtain a global trend component. Based on the initial data subset and the global trend component, the detrended residual sequence is solved and frequency domain transformation is performed to identify the dominant frequency component. The dominant frequency component is inversely transformed to reconstruct the periodic component. The global trend component is subjected to first-order difference operation to obtain a trend change rate sequence. The trend change rate sequence is subjected to second-order difference to obtain an acceleration sequence. A set of candidate mutation points is determined by combining the acceleration sequence with a preset mutation threshold. The amplitude change of the periodic component in the time window before and after each candidate mutation point in the candidate mutation point set is extracted and the amplitude jump ratio is calculated. The candidate mutation points are screened based on the amplitude jump ratio to obtain the rate of change mutation point. The abrupt change point of the rate of change is projected onto the time index sequence to determine the time position, and the initial data subset is divided into multiple time periods based on the time position. The variance of the trend change rate sequence is calculated for each time period to obtain the time period stability index. Based on the time period stability index, a set of stable time periods is determined, and the corresponding data is extracted from the initial data subset and spliced ​​to obtain the data subset.

[0012] In one alternative implementation, Based on the risk level and the dominant influencing factor, a set of candidate intervention schemes is determined by querying a preset intervention strategy library. For each candidate intervention scheme, a multi-step state transition simulation is performed, and a timeliness score is calculated. Based on the timeliness score, a target intervention scheme is determined, and a monitoring report is generated by combining the abnormal state fragments, including: Intervention strategy entries are extracted from the intervention strategy library and the matching conditions are parsed. The risk level is matched with the risk range in the matching conditions to obtain the risk matching degree. The dominant influencing factor is matched with the influencing factor type in the matching conditions to obtain the factor matching degree. The comprehensive matching degree is calculated based on the risk matching degree and the factor matching degree, and the candidate intervention scheme set is obtained by threshold screening based on the comprehensive matching degree. Intervention parameters for each candidate intervention scheme in the candidate intervention scheme set are extracted and a state update rule is constructed. The predicted state vector at the current moment in the evolution trajectory is used as the simulation starting state, and the state update rule is applied to perform multi-step iterations to obtain a multi-step simulated state sequence. The deviation change trend between the multi-step simulated state sequence and the corresponding moment benchmark state vector in the health benchmark template is calculated to obtain a deviation improvement curve. The area under the deviation improvement curve is calculated, and the timeliness is discounted by combining the response delay in the intervention parameters to obtain a timeliness score. The candidate intervention plan with the highest timeliness score is selected from the set of candidate intervention plans as the target intervention plan. Intervention measures are extracted from the target intervention plan and combined with the abnormal state fragment to generate a monitoring report.

[0013] A second aspect of this invention provides a multimodal sensing-based vital sign monitoring system for elderly patients, comprising: The data acquisition unit is used to collect physiological signal data and behavioral state data of elderly patients and perform feature extraction to obtain physiological feature vectors and behavioral feature vectors. After aligning the physiological feature vectors and behavioral feature vectors, multi-layer nonlinear transformation is performed and projected onto a common semantic space to obtain a unified representation vector and construct a modal correlation matrix. An anomaly identification unit is used to perform sliding time window segmentation on the unified representation vector and calculate statistical distribution characteristics to obtain baseline features, perform tensor fusion operation on the baseline features and the modal correlation matrix to obtain fused state features, encode and compress the fused state features, decode and reconstruct them, calculate reconstruction error to identify anomalous state segments, perform backtracking analysis on the anomalous state segments to extract precursor change patterns, and combine them with the modal correlation matrix to determine the dominant influencing factor. The risk grading unit is used to extract a subset of data from the unified characterization vector based on the dominant influencing factor and perform iterative deduction on a continuous time scale to determine the evolution trajectory, calculate the deviation between the evolution trajectory and the preset health benchmark template, and perform a graded mapping based on the deviation to obtain the risk level. The intervention execution unit is used to query a preset intervention strategy library to determine a set of candidate intervention schemes based on the risk level and the dominant influencing factor, perform multi-step state transition simulation for each candidate intervention scheme and calculate the timeliness score, determine the target intervention scheme based on the timeliness score and generate a monitoring report in combination with the abnormal state fragment.

[0014] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] In this invention, by fusing multimodal data of physiological signals and behavioral states, a unified semantic representation is constructed and a modal correlation matrix is ​​established. This effectively captures the inherent connections and synergistic change patterns between different information sources, overcoming the limitations of a single data source and significantly improving the completeness and reliability of state perception. By identifying abnormal segments through an encoding reconstruction mechanism and combining retrospective analysis and dominant factor determination, the starting point and key triggers of state deterioration can be quickly located, greatly improving the timeliness of anomaly detection and the accuracy of attribution analysis, providing a solid basis for early warning. Based on the extraction of data subsets from dominant factors for continuous time extrapolation, the dynamic evolution trajectory of health status can be simulated, realizing the quantitative assessment and forward-looking prediction of potential risk development trends. This transforms risk warning from post-event alerts to pre-event predictions, enhancing the initiative of health management. By performing multi-step state transition simulation and timeliness scoring on candidate solutions, the most efficient target intervention plan can be selected, and a monitoring report containing anomaly analysis and response suggestions can be generated, significantly improving the intelligence level and intervention efficiency of elderly health monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for monitoring vital signs in elderly patients based on multimodal sensing, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the risk level analysis of a multimodal sensing-based method for monitoring vital signs in elderly patients, as described in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the method for monitoring vital signs in elderly patients based on multimodal sensing, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Physiological signal data and behavioral state data of elderly patients are collected and feature extraction is performed to obtain physiological feature vectors and behavioral feature vectors. After aligning the physiological feature vectors and behavioral feature vectors, multi-layer nonlinear transformation is performed and projected onto a common semantic space to obtain a unified representation vector and construct a modal correlation matrix. The unified representation vector is segmented by a sliding time window and its statistical distribution characteristics are calculated to obtain baseline features. The baseline features are then fused with the modal correlation matrix using tensor fusion to obtain fused state features. The fused state features are encoded, compressed, decoded, and reconstructed, and the reconstruction error is calculated to identify abnormal state segments. The abnormal state segments are then backtracked to extract precursor change patterns and combined with the modal correlation matrix to determine the dominant influencing factors. Based on the dominant influencing factor, a subset of data is extracted from the unified characterization vector and iteratively extrapolated over a continuous time scale to determine the evolution trajectory. The deviation between the evolution trajectory and the preset health benchmark template is calculated, and a graded mapping is performed based on the deviation to obtain the risk level. Based on the risk level and the dominant influencing factor, a set of candidate intervention schemes is determined by querying a preset intervention strategy library. For each candidate intervention scheme, a multi-step state transition simulation is performed and a timeliness score is calculated. Based on the timeliness score, a target intervention scheme is determined and a monitoring report is generated by combining the abnormal state fragments.

[0021] In one alternative implementation, Physiological signal data and behavioral state data of elderly patients are collected and feature extracted to obtain physiological feature vectors and behavioral feature vectors. These physiological and behavioral feature vectors are then aligned, subjected to multi-level nonlinear transformation, and projected onto a common semantic space to obtain a unified representation vector. A modal correlation matrix is ​​then constructed, including: Physiological signal data and behavioral state data of elderly patients are acquired. The physiological signal data is decomposed into time-frequency data to extract frequency domain energy distribution features. The mean and variance of the physiological signal data are calculated to obtain time domain statistical features, which are then combined with the frequency domain energy distribution features to obtain a physiological feature vector. The behavioral state data is analyzed to extract joint angle sequences. The position change rate of the behavioral state data is calculated to obtain motion trajectory features, which are then combined with the joint angle sequences to obtain a behavioral feature vector. The timestamp information of the physiological feature vector and the behavioral feature vector is extracted. The time offset is calculated based on the timestamp information and linear interpolation and feature alignment are performed. The aligned physiological feature vector and behavioral feature vector are subjected to nonlinear transformation to obtain physiological intermediate representation and behavioral intermediate representation respectively. The physiological intermediate representation and the behavioral intermediate representation are concatenated and subjected to nonlinear transformation to obtain cross-modal intermediate representation. The cross-modal intermediate representation is subjected to nonlinear transformation and normalized after hyperbolic tangent activation to obtain a unified representation vector. The Pearson correlation coefficients between the physiological and behavioral feature components of the unified representation vector at different times are calculated and arranged according to modality type and time order to obtain the correlation matrix. The correlation matrix is ​​then thresholded to obtain the modality correlation matrix.

[0022] Physiological signal data is collected through wearable devices deployed on elderly patients, including electrocardiogram (ECG) sensors, blood oxygen sensors, and body temperature sensors. The sampling frequency is set between 250Hz and 500Hz to ensure signal quality. Behavioral status data is collected via a depth camera and inertial measurement unit (IMU) installed in the monitoring room. The depth camera captures the patient's limb movement trajectories at a rate of 30 frames per second, while the IMU records acceleration and angular velocity information at a frequency of 100Hz. The physiological signal data includes ECG waveforms, blood oxygen saturation curves, and body temperature change sequences, while the behavioral status data includes skeletal joint coordinate sequences and motion acceleration vectors.

[0023] Preprocessing was performed on the acquired physiological signal data to eliminate baseline drift and high-frequency noise. A bandpass filter was applied to the ECG signal to retain the effective components in the 0.5Hz to 40Hz frequency band. The filter order was set to 4th order to balance transition band steepness and phase distortion control. A short-time Fourier transform was performed on the filtered ECG signal with a 2-second time window and a 50% window overlap rate, resulting in a time-spectrum matrix. Five sub-bands were divided in the frequency domain: 0.5-4Hz (extremely low frequency), 4-8Hz (low frequency), 8-15Hz (mid frequency), 15-30Hz (high frequency), and 30-40Hz (extremely high frequency). The energy integral value within each band was calculated as the frequency domain energy distribution feature. The frequency domain energy distribution feature forms a five-dimensional vector, with each dimension representing the energy proportion of the corresponding frequency band. The mean and variance of the original ECG signal were calculated in the time domain within each 2-second window. The mean reflects the DC bias level of the signal, and the variance represents the degree of signal fluctuation. By concatenating the frequency domain energy distribution characteristics with the time domain statistical characteristics, a seven-dimensional physiological feature vector is obtained, which contains five frequency band energy values, a mean parameter, and a variance parameter.

[0024] For behavioral state data, human skeleton extraction is performed using depth images captured by a depth camera. A joint detection algorithm is used to identify the 3D coordinates of fifteen key points, including the shoulder, elbow, hip, knee, and ankle joints. Joint angles are calculated based on the vector relationships between adjacent joints; for example, the elbow flexion angle is determined by the vector angle formed by the shoulder, elbow, and wrist joints. For the lower limb chain formed by the hip, knee, and ankle joints, the knee flexion angle and hip flexion-extension angle are calculated. All joint angles are arranged in chronological order to form a joint angle sequence, with twelve angle parameters at each time step. Simultaneously, the rate of position change is calculated based on the changes in joint coordinates between consecutive frames. The torso center point is selected as a reference point, and the displacement vector of this point between adjacent frames is calculated, divided by the inter-frame time interval to obtain the instantaneous velocity vector. Moving averages are calculated for the horizontal and vertical components of the velocity vector, with an average window length of ten frames, resulting in smoothed motion trajectory features. The motion trajectory features are two-dimensional vectors, representing the horizontal and vertical motion rates, respectively. The joint angle sequence is combined with the motion trajectory features to form a fourteen-dimensional behavioral feature vector.

[0025] Because physiological signal data and behavioral state data originate from different sensor systems, their timestamps inherently deviate. Timestamp information for each data sample is extracted, with a required accuracy at the millisecond level. The time offset between the corresponding timestamps of the physiological feature vector and the behavioral feature vector is calculated; alignment is performed when the offset exceeds 20 milliseconds. A linear interpolation method is used to fill in missing feature values ​​on the time axis; the interpolation formula is as follows: ,in and For adjacent known times, and For the corresponding eigenvalues, The point at which interpolation is needed is defined. Interpolation processes ensure precise alignment of the physiological and behavioral feature vectors on the time axis, with alignment errors controlled within 5 milliseconds.

[0026] A nonlinear transformation is performed on the aligned physiological feature vectors, using a two-layer fully connected network structure. The first layer maps the seven-dimensional physiological feature vectors to a thirty-two-dimensional space, with the mapping relationship as follows: ,in For physiological feature vectors, This is the weight matrix. For bias vectors, The ReLU activation function is used. The second layer maps the 32-dimensional vector to a 64-dimensional space, with the mapping relationship as follows: This yields a 64-dimensional physiological intermediate representation. The same nonlinear transformation process is applied to the behavioral feature vectors: the first layer maps the 14-dimensional behavioral feature vectors to 32 dimensions, and the second layer maps them to 64 dimensions, resulting in a 64-dimensional behavioral intermediate representation.

[0027] Physiological and behavioral intermediate representations are concatenated according to feature dimensions to form a 128-dimensional cross-modal intermediate representation. This representation undergoes further nonlinear transformation, compressing the 128-dimensional vector to 32 dimensions through a single fully connected network layer. The transformation relationship is as follows: ,in This is the intermodal intermediate representation after splicing. and The transformation parameters are used. Hyperbolic tangent activation is applied to the transformation result to compress the values ​​to between -1 and 1. L2 normalization is then performed, resulting in a 32-dimensional unified representation vector.

[0028] To construct the modal correlation matrix, unified representation vectors were collected at least one hundred time points over a continuous time period. For each unified representation vector, its corresponding physiological and behavioral feature components were extracted. The physiological feature components consist of the first sixteen dimensions of the unified representation vector, and the behavioral feature components consist of the last sixteen dimensions. The Pearson correlation coefficient between the physiological and behavioral feature components at different time points was calculated using the following formula: ,in For the first The physiological characteristic component in the first The value at each moment, For the first The behavioral feature component in the first The value at each moment, and These are the time averages of the corresponding components. The total number of moments is given. The correlation coefficients between all physiological and behavioral feature components are arranged according to modal type, with physiological feature components arranged in the order of frequency domain energy and time domain statistics, and behavioral feature components arranged in the order of joint angle and movement speed, forming a 16x16 correlation matrix.

[0029] The correlation matrix is ​​thresholded to highlight significant associations. A correlation coefficient threshold of 0.3 is set, and elements with absolute values ​​less than 0.3 are set to zero, while elements with absolute values ​​greater than or equal to 0.3 retain their original values, resulting in a sparse modal correlation matrix. The positions of non-zero elements in this matrix indicate significant temporal synergistic relationships between different modal features. The sign of the element values ​​reflects the direction of synergy, with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation. The modal correlation matrix provides a quantitative description of cross-modal dependencies for subsequent fusion of state features, enabling the system to identify the intrinsic connection between physiological state changes and behavioral performance.

[0030] In one alternative implementation, The baseline features are obtained by segmenting the unified representation vector into sliding time windows and calculating its statistical distribution characteristics. The fused state features are then obtained by performing tensor fusion operations on the baseline features and the modal correlation matrix. The unified representation vector is divided into multiple time window segments according to the preset window length and sliding step size. The central trend index and dispersion index are calculated for the unified representation vector in each time window segment to obtain the distribution concentration feature. The distribution symmetry index and distribution kurtosis index are calculated for the unified representation vector in each time window segment to obtain the distribution morphology feature. The distribution concentration feature and the distribution morphology feature are spliced ​​together in the order of the time window to obtain the baseline feature. The baseline features are arranged into a baseline feature tensor according to the time window and feature dimension. The modal correlation matrix is ​​expanded into a correlation tensor according to the modal type and time order. Tensor contraction operation is performed on the baseline feature tensor and the correlation tensor in the time dimension to obtain a time fusion intermediate tensor. Tensor expansion is performed on the time fusion intermediate tensor in the feature dimension to obtain an expanded feature matrix. The expanded feature matrix is ​​decomposed to extract the dominant feature components. The dominant feature components are reconstructed and combined with the baseline feature tensor to calculate the enhanced baseline tensor. The enhanced baseline tensor and the associated tensor are then subjected to tensor outer product operation to obtain the fused state features.

[0031] After obtaining the unified representation vector, it needs to undergo temporal structuring to extract stable baseline features. A preset window length is set to... There are sampling points, and the sliding step size is . Each sampling point, starting from the beginning of the unified representation vector sequence, is sequentially truncated at intervals of length [length missing]. The fragment. Assume the total length of the unified representation vector sequence is... The number of time window segments that can be obtained is Each time window segment is denoted as ,in This fragment contains information from time [time]. At that time A unified representation vector.

[0032] For each time window segment The unified representation vector within the vector is first calculated using the central tendency index and the dispersion index. The central tendency index is expressed as the arithmetic mean. With median A dual metric is used: the arithmetic mean is calculated as the average of all vector elements within the segment, and the median is calculated as the middle value of the vector elements within the segment after numerical sorting. The standard deviation is used as the dispersion index. Interquartile range To represent these values, the standard deviation is calculated as the square root of the sum of squares of the deviations of each vector element from the mean within a segment, and the interquartile range is calculated as the difference between the third quartile and the first quartile. These four indicators are then concatenated to form a four-dimensional vector. This reflects the concentrated distribution characteristic of this time window.

[0033] Further calculations were performed on the distribution symmetry and kurtosis indices to characterize the distribution's morphological features. Distribution symmetry was assessed using the skewness coefficient. For quantification, the skewness coefficient is calculated as the cube of the deviation of each vector element within a segment from the mean, divided by the cube of the standard deviation. An absolute value close to zero indicates a symmetrical distribution, a positive value indicates right skewness, and a negative value indicates left skewness. Kujicicity is represented by the kurtosis coefficient. For characterization, the kurtosis coefficient is calculated as the fourth power of the deviation of each vector element within a segment from the mean, divided by the fourth power of the standard deviation, and then subtracted by 3. A positive value indicates a more angular distribution, while a negative value indicates a more flattened distribution. A distribution range index is also introduced. This is calculated as the difference between the maximum and minimum values ​​of the vector elements within the segment. The skewness coefficient, kurtosis coefficient, and distribution range index are then concatenated to form a three-dimensional vector. This serves as the characteristic of the distribution pattern within that time window.

[0034] Each time window Distribution concentration characteristics Distribution morphology characteristics By concatenating the components, a seven-dimensional feature vector is obtained. All according to the time window order. The feature vectors of each time window are concatenated vertically to form a dimension of... The baseline feature matrix is ​​a matrix in which each row corresponds to the statistical distribution characteristics of a time window, and each column corresponds to the evolution of a specific statistical indicator over time.

[0035] To achieve deep fusion of baseline features and modal correlation matrices, both need to be converted into tensor form for computation. The baseline feature matrix is ​​then processed according to the time window dimension. Feature Dimension N The implicit modal dimensions are reconstructed to form the baseline feature tensor. ,in This indicates the number of modes. Baseline features are repeated for each modal channel to maintain the consistency of the tensor structure. The modal correlation matrix is ​​then sorted according to the number of modal types. With the number of time windows To expand, the original modal correlation matrix has the following dimensions: This represents the correlation strength between different modes, achieved by replicating the matrix over time. Next, a correlation tensor is formed. This ensures that each time window corresponds to a set of modal association information.

[0036] Baseline feature tensor in the time dimension With correlation tensor Tensor shrinkage is performed. This operation is carried out along the time dimension, aligning the time dimension of the baseline feature tensor with the time dimension of the associated tensor for each time window position. Extracting slices of baseline feature tensors Slices with associated tensors Matrix multiplication is performed, and a weighted summation is applied along the modal dimension to obtain the fusion result for that time window position. The fusion results for all time window positions are then stacked along the time dimension to form a temporal fusion intermediate tensor. This tensor contains both statistical information about baseline features and interactive information about modal associations.

[0037] Tensor unfolding is performed on the temporal fusion intermediate tensor along the feature dimension, transforming the original three-dimensional tensor... Rearrange along the feature dimension and modality dimension, and rearrange the feature dimension... N Modal dimension Merge and flatten to form an expanded feature matrix. Each row of the matrix corresponds to the full-modal fusion features of a time window, and each column corresponds to the value of a specific feature in a certain modality. The matrix structure facilitates subsequent decomposition and reconstruction operations.

[0038] Expand the characteristic matrix Matrix decomposition is performed to extract dominant feature components. Singular value decomposition is used to decompose the expanded feature matrix into... ,in It is a left singular matrix. It is a singular value diagonal matrix. This is a right singular matrix. Select the first [value] from the singular value diagonal matrix. There are 100 maximal singular values ​​and their corresponding singular vectors, where The value of is determined based on the cumulative contribution rate threshold, typically set to 85% to 95%. The extracted dominant feature components are represented as follows: This component retains the most significant change patterns in the original expanded feature matrix while removing redundant information and noise interference.

[0039] For dominant feature components Dimensional reconstruction is performed to restore it from a two-dimensional matrix form to a three-dimensional tensor form, resulting in the dominant feature tensor. The structure of this tensor is similar to that of the original baseline feature tensor. Maintain consistency. Calculate a weighted combination of the dominant feature tensor and the baseline feature tensor, using adaptive weighting coefficients. Adjustments were made to enhance the baseline tensor. Calculated as The weighting coefficient The proportion of variance explained by the dominant feature component is dynamically determined; the higher the proportion of explained variance, the better. Larger values ​​ensure that the contribution of dominant features is fully reflected. The augmented baseline tensor retains the statistical stability of the original baseline features while incorporating the dominant change patterns after dimensionality reduction and refinement, resulting in a more robust feature representation. With correlation tensor Perform tensor outer product operations to obtain the fused state features.

[0040] In one alternative implementation, After encoding, compressing, decoding, and reconstructing the fused state features, the reconstruction error is calculated to identify anomalous state segments. Backtracking analysis is performed on these anomalous state segments to extract precursor change patterns, and the dominant influencing factors are determined by combining these patterns with the modal correlation matrix. The fusion state features are subjected to multi-level nonlinear dimensionality reduction transformation and the mean vector and variance vector in the latent space are calculated. Based on the mean vector and variance vector, reparameterized sampling is performed to obtain a compressed coding representation. The divergence measure between the compressed coding representation and the preset prior distribution is calculated and multi-level nonlinear dimensionality increase transformation is performed to obtain the reconstructed fusion state features. The reconstruction loss value between the reconstructed fusion state features and the fusion state features is calculated and a reconstruction error sequence is constructed. The time period in the reconstruction error sequence where the reconstruction loss value exceeds a preset threshold is marked and clustered and merged to obtain abnormal state segments. Extract the abnormal start time from the abnormal state segment and backtrack to determine the backtracking time window. Extract the fusion state features within the backtracking time window and calculate the feature differences between adjacent time steps to construct a change amplitude sequence. Smooth the change amplitude sequence and identify local maxima as key transition nodes. Extract the fusion state features and corresponding change directions at the time corresponding to the key transition nodes to obtain the precursor change pattern. The Pearson correlation coefficient corresponding to the anomaly initiation time is extracted from the modal correlation matrix, and the modal volatility is calculated. The feature component with the largest modal volatility is selected, and the corresponding modal type and feature dimension information are extracted to obtain the dominant influencing factor.

[0041] The process of encoding, compressing, and reconstructing the fused state features employs a variational autoencoder architecture to achieve compressed representation of the feature space. The fused state features are input into an encoder network consisting of three fully connected layers with 512, 256, and 128 nodes per layer, respectively. The activation function used is a LeakyReLU function with a leakage coefficient of 0.2. During encoding, the output of the last fully connected layer splits into two branches, which respectively calculate the mean vector of the latent space. Sum of variance vectors The mean vector and variance vector are both 64-dimensional, and together they describe the probability distribution characteristics in the latent space. To ensure the numerical stability of the variance vector, an exponential transformation is applied to the output. , where logvar is the logarithmic variance directly output by the network.

[0042] When performing reparameterization sampling in the latent space, random noise vectors are sampled from a standard normal distribution. ,in It is a 64-dimensional identity matrix. Through transformation... The compressed coding representation is obtained, where This represents element-wise multiplication. This reparameterization technique makes the sampling process differentiable, facilitating backpropagation optimization. The KL divergence between the compressed encoded representation and the preset prior distribution is calculated; the prior distribution is chosen as a standard normal distribution. The formula for calculating divergence is: The divergence value serves as a regularization term to constrain the distribution pattern of the latent space.

[0043] Compressed encoding representation The input decoder network is used for dimensionality reconstruction. The decoder also employs a three-layer fully connected structure with 128, 256, and 512 nodes respectively. Finally, it maps back to the original dimension of the fused state features through a linear layer. The activation functions of each layer in the decoder are consistent with those in the encoder, while the output layer does not use an activation function to preserve the complete numerical range of the features. This yields the reconstructed fused state features. Then, its characteristics compared with the original fusion state are calculated. The reconstruction loss between time steps is expressed as mean squared error. The reconstruction loss values ​​for all time steps are arranged in chronological order to construct a reconstruction error sequence. .

[0044] Anomaly detection is performed by setting a dynamic threshold for the reconstruction error sequence. The threshold is calculated using a sliding window statistical method. A window length of 50 time steps is selected, and the median reconstruction loss within the window is calculated. Interquartile range The threshold is defined as When the reconstruction loss value at a certain time step At that time step, the time step is marked as an outlier. To avoid interference from isolated noise points, morphological closing operations are performed on the marking results. First, a dilation operation is performed to connect adjacent outliers, with the dilation kernel size set to 3 time steps. Then, an erosion operation is performed to remove fine burrs, with the erosion kernel also set to 3 time steps. Connectivity analysis is performed on the processed outliers to group temporally continuous outliers into the same outlier state segment, and the start time of each segment is recorded. and end time If the time interval between two segments is less than 10 time steps, they are merged into a single complete segment.

[0045] For each identified abnormal state segment, extract its start time. An analysis window is established by backtracking backwards. The length of the backtracking time window is set to twice the duration of the anomalous segment, but not less than 30 time steps and not more than 100 time steps. From time point... arrive Extract the fusion state feature sequence within this time period ,in The length of the backtracking window is given. The Euclidean distance between feature vectors at adjacent time steps is calculated to construct a sequence of change magnitudes. The change magnitude sequence quantifies the rate of feature evolution over time.

[0046] A Gaussian smoothing filter was applied to the amplitude variation sequence to eliminate high-frequency noise. The standard deviation of the Gaussian kernel was set to 2.0, and the kernel window width was 9 sampling points. The smoothed sequence is denoted as... Identify local maxima points on this sequence, i.e., satisfy... and The location of each key transition point is determined by selecting points with magnitudes exceeding the sequence mean plus one standard deviation. For each identified key transition point, its corresponding time index is recorded. Extract the fusion state features at that moment. and feature change vector Principal component analysis is used to extract the direction of change vectors, and the first three principal components are retained as the main directions of change. These direction vectors, together with the eigenvalues ​​at the corresponding time points, constitute a description of the precursor change pattern.

[0047] Locate the anomaly initiation time from the modal correlation matrix. The corresponding matrix slices. The modal correlation matrix has the following dimensions: ,in For time steps, This represents the total dimension of the modal features. Extraction... Moment Correlation coefficient matrix The elements of the matrix Indicates the first Dimensional features and the first The Pearson correlation coefficient between features. The modal volatility of each feature dimension is calculated, defined as the variance of the absolute values ​​of the correlation coefficients between that dimension and all other dimensions. ,in For the first The mean of the absolute values ​​of the correlation coefficients between a feature dimension and other dimensions. High modal volatility indicates that the correlation strength between this feature dimension and other dimensions varies drastically, and is more likely to be the dominant factor causing anomalies.

[0048] Sort the modal volatility of all feature dimensions in descending order and select the dimension index with the highest volatility. Based on the structure of the feature vectors, the indices are mapped back to specific modality types and feature dimensions. The first 32 dimensions of the physiological feature vector correspond to heart rate variability indicators, dimensions 33-64 correspond to blood oxygen saturation-related features, dimensions 65-96 correspond to blood pressure fluctuation features, and dimensions 97-128 correspond to respiratory pattern features. The first 40 dimensions of the behavioral feature vector represent limb activity intensity, dimensions 41-80 represent posture transition frequency, dimensions 81-120 represent sleep quality parameters, and dimensions 121-160 represent daily behavioral regularity indicators. This is achieved through index values... Determine the modal category to which it belongs, for example if This pinpoints the 13th dimension of blood oxygen saturation within the physiological modality. Further analysis extracts the numerical change curve of this dimension within the backtracking window, calculating the standard deviation of its first difference as a quantitative indicator of fluctuation severity. The triplet of modality type, specific feature dimension, and fluctuation severity serves as a complete description of the dominant influencing factors, providing targeted support for subsequent trajectory extrapolation and intervention decisions. The entire anomaly identification and factor localization process forms a closed-loop analysis chain, enabling precise control over changes in vital signs in elderly patients, from data-driven anomaly detection to mechanistic-level cause tracing.

[0049] Figure 2This is a flowchart illustrating the risk level analysis of a multimodal sensing-based method for monitoring vital signs in elderly patients, as described in an embodiment of the present invention.

[0050] In one alternative implementation, Based on the dominant influencing factors, a subset of data is extracted from the unified characterization vector, and the evolutionary trajectory is determined through iterative deduction on a continuous time scale. The deviation between the evolutionary trajectory and the preset health benchmark template is calculated, and a graded mapping is performed based on the deviation to obtain the risk level, including: Based on the dominant influencing factor, time series data corresponding to the feature dimensions are extracted from the unified characterization vector to obtain an initial data subset. The initial data subset is then decomposed into a time series to obtain trend components and calculate the rate of change. Based on the rate of change, abrupt change points are identified and time periods are divided at the abrupt change points to determine the stable evolution stage, thus obtaining the data subset. The data subset is discretized and sampled according to the time step to obtain the initial time state vector. The state transition rule is determined based on the dominant influencing factor, and the initial time state vector is iteratively transformed. The iterative deduction is repeated on a continuous time scale to obtain the predicted state vector sequence and smooth it to obtain the evolution trajectory. A baseline state vector sequence is extracted from a preset health baseline template. The Euclidean distance between the predicted state vector in the evolution trajectory and the baseline state vector in the baseline state vector sequence is calculated to obtain a time step deviation sequence. The time step deviation sequence is then weighted and summed to obtain a comprehensive deviation. Based on the comprehensive deviation and a preset multi-level threshold interval, interval matching is performed to map to a discrete risk level. The risk level is determined based on the discrete risk level and the number of rate of change mutation points.

[0051] After identifying the dominant influencing factor, it is necessary to extract the feature dimension data related to that factor from the unified representation vector. The unified representation vector is typically a high-dimensional vector containing multiple feature dimensions from physiological signals and behavioral states. The dominant influencing factor indicates the feature dimension number or combination of dimensions that contributes most to the abnormal state. Through indexing operations, the value of the corresponding dimension in the unified representation vector is located, and the values ​​of that dimension at all time points are extracted along the time axis to form a time series data. For example, if the dominant influencing factor indicates that heart rate is the main influencing factor, then the numerical sequence of the corresponding heart rate dimension throughout the entire monitoring period is extracted. This sequence is arranged in chronological order to form an initial data subset.

[0052] The initial data subset typically contains various components such as trend changes, periodic fluctuations, and random noise. Time series decomposition techniques are used to separate it into trend components, seasonal components, and residual components. The trend component reflects the long-term direction of data change and can be extracted using moving averages or locally weighted regression smoothing. In practice, a sliding window width is set, and a weighted average is calculated for the data points within the window. Weights can be applied using triangular or Gaussian kernel functions, ensuring that the center points of the window have larger weights and the edge points have smaller weights. The sliding window moves gradually along the time axis, and the weighted average calculated at each position constitutes the trend component sequence. Subtracting the trend and seasonal components from the original data yields the residuals, which mainly contain noise and abnormal fluctuations.

[0053] The rate of change is calculated based on the trend component, defined as the difference between the trend values ​​at adjacent time points divided by the time interval. Let the trend component at time point... The value is Then the time point arrive rate of change ,in The sampling time interval. Rate of change sequence. The rate of change of the trend component is described across different time periods. Abrupt changes in the rate of change are identified by analyzing the rate of change series. The identification of abrupt changes employs the second-order difference method of the rate of change, calculating the difference between adjacent rates of change. When the absolute value of the difference exceeds the preset threshold At that time, it was considered that at a certain point in time There is a sudden change in the rate of change. This threshold is determined based on the statistical characteristics of historical data and can be set to twice the standard deviation of the rate of change fluctuation.

[0054] After identifying all abrupt changes in the rate of change, the entire time period was divided according to the location of these abrupt changes. The time interval between two adjacent abrupt changes was defined as a stable evolutionary stage, within which the rate of change remained relatively stable without drastic trend reversals. The initial data subset was divided into multiple stable evolutionary stages, with data within each stage exhibiting similar evolutionary characteristics. Stages that were too short were removed, and stages whose duration exceeded a minimum threshold were retained, resulting in a final filtered data subset. This data subset contains several stable evolutionary stages, within which the data trend is continuous and the rate of change is stable.

[0055] Discretize the data subset according to a fixed time step. Select data points. Set the initial time to... At this moment, the values ​​of each feature dimension in the data subset are extracted to form the initial state vector. The initial state vector is a multi-dimensional vector, and the number of dimensions depends on the number of features involved in the dominant influencing factor. For example, if the dominant influencing factor involves two features, heart rate and blood oxygen, then the state vector is a two-dimensional vector, with components representing the initial heart rate value and blood oxygen value, respectively.

[0056] The state transition rule is determined based on the characteristics and historical evolution patterns of the dominant influencing factor. The changes in each stable evolution stage within the data subset are analyzed to extract the mathematical relationships of the state transitions. If the dominant influencing factor exhibits a linear trend, the state transition rule can be expressed as follows: ,in This is a trend coefficient vector, obtained by fitting historical data using the least squares method. If the dominant influencing factor exhibits nonlinear evolution, the state transition can be described using a polynomial or exponential function. For example, for exponential decay characteristics, the state transition rule is as follows: ,in This is the attenuation coefficient.

[0057] Based on the state transition rule, the initial state vector is iteratively transformed. From the initial state... Starting from this point, calculate the state at the next time step using the state transition rules. , and then from calculate The process is iterated sequentially. The number of iterations is determined by the prediction duration and time step; if it is necessary to predict the future duration... The number of iterations is During the iteration process, the state vector obtained from each transformation is recorded to form a sequence of predicted state vectors. This sequence describes the evolution of states over a continuous time scale.

[0058] The predicted state vector sequence may contain fluctuations due to accumulated errors from iterations. To obtain a smooth evolution trajectory, the sequence is smoothed. Moving average filtering or Savitzky-Golay filtering can be used to smooth the sequence. Moving average filtering calculates the average of each point and its neighbors as the smoothed value; the window width is determined based on the sequence length and the desired smoothness. Savitzky-Golay filtering fits a polynomial within a local window, replacing the original value with the polynomial's value at the center point, thus preserving the characteristic peaks of the data while smoothing. The resulting smoothed sequence is the evolution trajectory, which is continuous and smooth, reflecting the temporal evolution trend of the states.

[0059] The health baseline template is pre-established based on historical data from a large number of healthy elderly patients. The template contains a standard value sequence for each feature dimension under the same time scale, representing the health status. The feature dimensions corresponding to the dominant influencing factors are extracted from the template to obtain the baseline state vector sequence. The length of the baseline state vector sequence is the same as that of the predicted state vector sequence, with one baseline state vector corresponding to each time step.

[0060] Calculate the Euclidean distance between the predicted state vector and the baseline state vector in the evolution trajectory. At time step... At this point, the Euclidean distance is defined as ,in The number of dimensions of the state vector. and The predicted state vector and the baseline state vector are respectively in the th... The components of the dimension. Calculate the Euclidean distance for all time steps to obtain the time step deviation sequence. .

[0061] The weighted sum of the time-step deviation sequences yields the comprehensive deviation. The weighting coefficients are determined based on the importance of each time step; closer time steps have larger weights, while longer time steps have smaller weights, reflecting the greater impact of recent conditions on risk assessment. Let the first... The weight of the time step is ,satisfy The overall deviation is calculated as follows: The weights can be applied using an exponential decay function, for example... ,in The normalization coefficient is... Control the decay rate.

[0062] Risk levels are categorized based on the overall deviation. Multiple threshold ranges are pre-defined, dividing the range of overall deviation values ​​into several intervals, each corresponding to a discrete risk level. For example, a three-level risk system can be defined: low risk corresponds to an overall deviation less than the threshold. The overall deviation corresponding to medium risk is arrive Between these, the overall deviation corresponding to high risk is greater than [a certain value]. By using interval matching, the calculated overall deviation is mapped to the corresponding discrete risk level.

[0063] The risk level is further adjusted based on the number of rate-of-change abrupt changes. The number of rate-of-change abrupt changes reflects the instability of state evolution; the more abrupt changes, the more drastic and unpredictable the state changes. Let the number of rate-of-change abrupt changes be... ,like Exceeding the preset threshold If so, the discrete risk level will be raised by one level.

[0064] In one alternative implementation, The initial data subset is decomposed into a time series to obtain trend components and calculate the rate of change. Based on the rate of change, abrupt change points are identified, and time periods are divided at these abrupt change points to determine stable evolution stages. The resulting data subset includes: The initial data subset is sorted by time dimension to construct a time index sequence and segmented by sliding window to obtain multiple local time windows. The data in each local time window is fitted with a polynomial to obtain a local trend curve and connected in segments to obtain a global trend component. Based on the initial data subset and the global trend component, the detrended residual sequence is solved and frequency domain transformation is performed to identify the dominant frequency component. The dominant frequency component is inversely transformed to reconstruct the periodic component. The global trend component is subjected to first-order difference operation to obtain a trend change rate sequence. The trend change rate sequence is subjected to second-order difference to obtain an acceleration sequence. A set of candidate mutation points is determined by combining the acceleration sequence with a preset mutation threshold. The amplitude change of the periodic component in the time window before and after each candidate mutation point in the candidate mutation point set is extracted and the amplitude jump ratio is calculated. The candidate mutation points are screened based on the amplitude jump ratio to obtain the rate of change mutation point. The abrupt change point of the rate of change is projected onto the time index sequence to determine the time position, and the initial data subset is divided into multiple time periods based on the time position. The variance of the trend change rate sequence is calculated for each time period to obtain the time period stability index. Based on the time period stability index, a set of stable time periods is determined, and the corresponding data is extracted from the initial data subset and spliced ​​to obtain the data subset.

[0065] The initial data subset extracted from the unified representation vector often contains complex time-varying features and aliased change patterns, requiring fine-grained decomposition to identify stable evolutionary stages. The initial data subset is then sorted in ascending order by collection timestamps to form a time index sequence. ,in Indicates the first Each sampling time, This represents the corresponding data value. This represents the total number of samples in the initial data subset. The sliding window length is set to... Step size is Starting from the beginning of the sequence, the sequence is divided into local time windows with increasing step sizes. Number of windows Satisfying Relationships In physiological signal monitoring of elderly patients, the window length is usually set to cover a time span of three to five physiological cycles. For example, when monitoring heart rate, it can be set to three to five minutes to capture sufficient periodic fluctuation information.

[0066] Regarding the first Local time window The set of data points contained within ,in For window For the corresponding index set, a local trend model is established using a cubic polynomial fitting method. A design matrix is ​​then constructed. , its first Behavior Solving for the coefficient vector using the least squares method This makes the fitting error Minimum. The corresponding local trend curve is represented as follows: To avoid discontinuities at the junctions of adjacent windows, a weighted average fusion strategy is employed at the window boundaries. Fitted values ​​within overlapping regions are weighted according to their distance from the window center, with closer values ​​receiving greater weight. A complete global trend component is formed by connecting all local trend curves segment by segment. .

[0067] By calculating the residual sequence This process yields the fluctuation component after removing the long-term trend. A Fast Fourier Transform is applied to the residual sequence to transform the time-domain signal into the frequency domain, obtaining the spectral amplitude distribution. ,in It is half the sampling frequency. It identifies amplitudes exceeding a preset threshold in the spectrum. The frequency points, and the components corresponding to these frequency points, reflect the main periodic changes in the data. For respiratory rate monitoring in elderly patients, the dominant frequency is typically concentrated in the range of twelve to twenty breaths per minute. Extraction of data meeting the criteria... frequency component set A bandpass filter is constructed to preserve these dominant frequency components. An inverse Fourier transform is then performed on the filtered frequency domain signal to reconstruct the periodic component sequence. .

[0068] After obtaining the global trend component, perform a first-order difference operation on it to calculate the change between adjacent time points. To form a trend change rate sequence This sequence reflects the instantaneous rate of change in vital signs data; in elderly patients experiencing sudden changes in condition, the rate of change exhibits a significant abrupt change. Further calculation of the second difference of the rate of change sequence... , thus obtaining the acceleration sequence This sequence is more sensitive to inflection points in the trend of change. A mutation detection threshold is set. It is usually determined based on two to three times the standard deviation of the acceleration sequence, and will satisfy the conditions. Time Index Add to the candidate mutation point set .

[0069] To improve the accuracy of mutation point recognition, it is necessary to verify by combining the change characteristics of the periodic component. For each index in the candidate mutation point set , extract the time windows of the lengths before and after it , which are respectively denoted as the front window and the rear window . Calculate the average amplitude of the periodic component within the front window , and the average amplitude within the rear window . Define the amplitude jump ratio , where is a very small positive number set to avoid division by zero. Set the ratio threshold , usually taking values between 0.3 and 0.5, and only retain the candidate points that satisfy as the final set of change rate mutation points .

[0070] Project the identified change rate mutation points onto the original time index sequence to obtain the corresponding moment positions . Taking these mutation moments as demarcation points, divide the initial data subset into consecutive time periods , and the time interval corresponding to the rd segment is , where the starting moment is determined by the previous mutation point, the ending moment is determined by the next mutation point, and the boundaries of the first and last segments are the start and end moments of the overall data. For each time period , extract the trend change rate values corresponding to all moments within this segment, and calculate the variance of the change rate within the segment , where is the mean of the change rate within the segment, and is the number of samples within the segment. The variance value reflects the stability of the vital sign data change within this time period, and the smaller the variance, the smoother the change.

[0071] Set the stability determination threshold , which can be determined according to the quantile of the variance distribution of historical health data change rate. Usually, the 75th percentile is selected as the reference value. Mark the time periods that meet the condition as stable time periods to form a set of stable time periods , where is the total number of stable segments. Extract the data points corresponding to the moments of all stable time periods from the initial data subset and splice them in chronological order to form a new data sequence, which is the data subset after screening through the stable evolution stage. In the blood pressure monitoring scenario of elderly patients, stable time periods usually correspond to the periods when patients are in a resting state or a regular activity state, excluding the unstable fluctuation segments caused by interference factors such as sudden changes in body position and emotional fluctuations, thus providing a more reliable data basis for subsequent evolution trajectory deduction.

[0072] In one alternative implementation, Based on the risk level and the dominant influencing factor, a set of candidate intervention schemes is determined by querying a preset intervention strategy library. For each candidate intervention scheme, a multi-step state transition simulation is performed, and a timeliness score is calculated. Based on the timeliness score, a target intervention scheme is determined, and a monitoring report is generated by combining the abnormal state fragments, including: Intervention strategy entries are extracted from the intervention strategy library and the matching conditions are parsed. The risk level is matched with the risk range in the matching conditions to obtain the risk matching degree. The dominant influencing factor is matched with the influencing factor type in the matching conditions to obtain the factor matching degree. The comprehensive matching degree is calculated based on the risk matching degree and the factor matching degree, and the candidate intervention scheme set is obtained by threshold screening based on the comprehensive matching degree. Intervention parameters for each candidate intervention scheme in the candidate intervention scheme set are extracted and a state update rule is constructed. The predicted state vector at the current moment in the evolution trajectory is used as the simulation starting state, and the state update rule is applied to perform multi-step iterations to obtain a multi-step simulated state sequence. The deviation change trend between the multi-step simulated state sequence and the corresponding moment benchmark state vector in the health benchmark template is calculated to obtain a deviation improvement curve. The area under the deviation improvement curve is calculated, and the timeliness is discounted by combining the response delay in the intervention parameters to obtain a timeliness score. The candidate intervention plan with the highest timeliness score is selected from the set of candidate intervention plans as the target intervention plan. Intervention measures are extracted from the target intervention plan and combined with the abnormal state fragment to generate a monitoring report.

[0073] Determining the target intervention plan and generating monitoring reports requires precise retrieval from a pre-defined intervention strategy library. This library stores intervention strategy entries for different health conditions of elderly patients, each entry containing detailed fit criteria fields. These fit criteria fields define risk range parameters and influencing factor type parameters. The risk range parameter is typically expressed as an interval; for example, the fit risk range for a particular strategy entry might be set to... This indicates that the strategy is applicable when the risk level value falls within this range. The influencing factor type parameter is stored as an enumerated set, for example, the applicable factor type for a certain strategy entry is heart rate variability, blood pressure variability, or activity intensity variation.

[0074] When calculating the risk matching score, the current patient's risk level value is compared with the risk range in the strategy entry. A risk matching score of 1.0 is calculated when the risk level value falls entirely within the risk range; otherwise, a score of 1.0 is calculated. At that time, the risk matching degree is calculated by linear decay based on distance, and the specific calculation method is as follows: ,in This indicates the distance between the risk level value and the nearest interval boundary; when the distance exceeds a threshold... At that time, the risk matching degree is directly set to 0. This segmented calculation method ensures that the matching process is both accurate and maintains a certain degree of fault tolerance.

[0075] Calculating factor matching requires semantic comparison between the dominant influence factor and the influence factor types in the fitting criteria. Since the dominant influence factor may contain multiple factor identifiers, set operations are used. Let the set of dominant influence factors be... The set of adaptation factor types for a certain strategy item is: Then the factor matching degree is calculated as follows: When the dominant influencing factor and the matching factor are completely identical, the factor matching degree reaches its maximum value of 1.0; when the two have no overlap, the factor matching degree is 0.

[0076] The calculation of the overall matching degree needs to consider both risk matching degree and factor matching degree. A weighted summation method is used for fusion, and the calculation formula is as follows: ,in and These represent the weighting coefficients for risk matching and factor matching, respectively, and are typically set to... and This reflects the dominant role of risk level in the selection of intervention strategies. After calculating the overall matching degree of each strategy item, a threshold is set. Screening is usually done. A value of 0.5 is used to retain only strategy entries with a comprehensive matching degree greater than this threshold, forming a set of candidate intervention schemes.

[0077] When performing multi-step state transition simulations for each candidate intervention in the candidate intervention set, it is necessary to first analyze its included intervention parameters. Intervention parameters typically include key fields such as intervention intensity, intervention frequency, intervention duration, and response delay. Intervention intensity indicates the degree of influence of the intervention on physiological indicators, usually expressed as a percentage. For example, an intervention intensity of 15% for a blood pressure-lowering intervention means that the intervention can reduce blood pressure by 15%. Intervention frequency describes the frequency of intervention implementation, such as three times daily or once every 6 hours. Intervention duration defines the duration of action of a single intervention; for example, the duration of a drug intervention may be 4 to 6 hours. Response delay represents the time interval from the implementation of the intervention to the production of a noticeable effect; for example, the response delay of some drug interventions is 30 to 60 minutes.

[0078] State update rules are constructed based on these intervention parameters. These rules define how the patient's state vector changes after the intervention is applied. For physiological indicator state components, the state update rules typically employ a linear adjustment model, i.e. ,in Indicates time The Each physiological indicator component Indicates the intensity of intervention. To adjust the coefficients. For behavioral state components, the state update rule may be more complex, requiring consideration of patient compliance and the difficulty of habit change. The update formula is as follows: ,in Indicates the target behavior state value. To change the rate parameter, This is the compliance coefficient.

[0079] In multi-step iterative simulations, the predicted state vector at the current moment in the evolution trajectory is used as the initial state for the simulation. Let the current moment be... The predicted state vector is The simulation step size is Step, usually set The range is 12 to 24, corresponding to the state evolution over the next 12 to 24 hours. In each simulation step... In the process, the state update rule is applied to calculate the state vector at the next time step. Simultaneously, the impact of response delay needs to be considered. When the time difference between the simulation time and the intervention start time is less than the response delay, the intervention effect is discounted proportionally, and the discount factor is calculated as follows: ,in For the time that has already passed, For response delay duration. Through continuous... The iterative calculation of the steps yields the result containing A multi-step simulated state sequence with state vectors.

[0080] To evaluate the effectiveness of the intervention, a multi-step simulated state sequence needs to be compared with the corresponding baseline state vectors in a healthy baseline template. The healthy baseline template stores typical state vectors of healthy elderly patients at different times, reflecting normal physiological rhythms and behavioral patterns. For each time step in the simulated state sequence, the corresponding baseline state vector is extracted and the deviation is calculated. If the intervention is effective, the deviation should gradually decrease as the number of simulated steps increases.

[0081] Connecting the deviations at various time points in chronological order creates a deviation improvement curve. The horizontal axis of the improvement curve represents the simulated time step, and the vertical axis represents the deviation value. An ideal intervention should result in a rapid decline in the improvement curve, eventually stabilizing, indicating that the patient's condition quickly approaches and maintains stability towards a healthy baseline. To quantify the overall effect of the improvement curve, the area under the curve is calculated using the trapezoidal integral method. The formula is as follows: ,in The area under the curve represents the time interval between adjacent moments. The smaller the area under the curve, the lower the overall deviation and the better the intervention effect.

[0082] A timeliness discount mechanism is introduced, adjusting the area under the curve based on response delay. The timeliness discount factor is calculated as follows: ,in This is the discount rate parameter, typically ranging from 0.02 to 0.05. The longer the response delay, the smaller the timeliness discount coefficient. The final timeliness score is calculated as follows: ,in This is to prevent division by zero errors for extremely small positive numbers.

[0083] After calculating the timeliness score for each candidate intervention, the scores are ranked. The candidate intervention with the highest timeliness score is selected as the target intervention. Specific intervention descriptions are extracted from the target intervention, which may include medication adjustment recommendations, lifestyle intervention guidance, or optimization of care procedures. Combined with previously identified abnormal state segments, a structured monitoring report is generated. The monitoring report includes modules such as basic patient information, the time period of the abnormal state, description of abnormal characteristics, analysis of dominant influencing factors, risk level assessment, and recommended intervention, providing comprehensive decision support information for healthcare professionals.

[0084] A second aspect of this invention provides a multimodal sensing-based vital sign monitoring system for elderly patients, comprising: The data acquisition unit is used to collect physiological signal data and behavioral state data of elderly patients and perform feature extraction to obtain physiological feature vectors and behavioral feature vectors. After aligning the physiological feature vectors and behavioral feature vectors, multi-layer nonlinear transformation is performed and projected onto a common semantic space to obtain a unified representation vector and construct a modal correlation matrix. An anomaly identification unit is used to perform sliding time window segmentation on the unified representation vector and calculate statistical distribution characteristics to obtain baseline features, perform tensor fusion operation on the baseline features and the modal correlation matrix to obtain fused state features, encode and compress the fused state features, decode and reconstruct them, calculate reconstruction error to identify anomalous state segments, perform backtracking analysis on the anomalous state segments to extract precursor change patterns, and combine them with the modal correlation matrix to determine the dominant influencing factor. The risk grading unit is used to extract a subset of data from the unified characterization vector based on the dominant influencing factor and perform iterative deduction on a continuous time scale to determine the evolution trajectory, calculate the deviation between the evolution trajectory and the preset health benchmark template, and perform a graded mapping based on the deviation to obtain the risk level. The intervention execution unit is used to query a preset intervention strategy library to determine a set of candidate intervention schemes based on the risk level and the dominant influencing factor, perform multi-step state transition simulation for each candidate intervention scheme and calculate the timeliness score, determine the target intervention scheme based on the timeliness score and generate a monitoring report in combination with the abnormal state fragment.

[0085] A third aspect of the present invention provides an electronic device, comprising: A processor and a memory for storing processor-executable instructions, wherein the processor is configured to invoke instructions stored in the memory to perform the aforementioned method.

[0086] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0087] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring vital signs in elderly patients based on multimodal sensing, characterized in that, include: Physiological signal data and behavioral state data of elderly patients are collected and feature extraction is performed to obtain physiological feature vectors and behavioral feature vectors. After aligning the physiological feature vectors and behavioral feature vectors, multi-level nonlinear transformation is performed and projected onto a common semantic space to obtain a unified representation vector and construct a modal correlation matrix. The unified representation vector is segmented by a sliding time window and its statistical distribution characteristics are calculated to obtain baseline features. The baseline features are then fused with the modal correlation matrix using tensor fusion to obtain fused state features. The fused state features are encoded, compressed, decoded, and reconstructed, and the reconstruction error is calculated to identify abnormal state segments. The abnormal state segments are then backtracked to extract precursor change patterns and combined with the modal correlation matrix to determine the dominant influencing factors. Based on the dominant influencing factor, a subset of data is extracted from the unified characterization vector and iteratively extrapolated over a continuous time scale to determine the evolution trajectory. The deviation between the evolution trajectory and the preset health benchmark template is calculated, and a graded mapping is performed based on the deviation to obtain the risk level. Based on the risk level and the dominant influencing factor, a set of candidate intervention schemes is determined by querying a preset intervention strategy library. For each candidate intervention scheme, a multi-step state transition simulation is performed and a timeliness score is calculated. Based on the timeliness score, a target intervention scheme is determined and a monitoring report is generated by combining the abnormal state fragments.

2. The method according to claim 1, characterized in that, Physiological signal data and behavioral state data of elderly patients are collected and feature extracted to obtain physiological feature vectors and behavioral feature vectors. These physiological and behavioral feature vectors are then aligned, subjected to multi-level nonlinear transformation, and projected onto a common semantic space to obtain a unified representation vector. A modal correlation matrix is ​​then constructed, including: Physiological signal data and behavioral state data of elderly patients are acquired. The physiological signal data is decomposed into time-frequency data to extract frequency domain energy distribution features. The mean and variance of the physiological signal data are calculated to obtain time domain statistical features, which are then combined with the frequency domain energy distribution features to obtain a physiological feature vector. The behavioral state data is analyzed to extract joint angle sequences. The position change rate of the behavioral state data is calculated to obtain motion trajectory features, which are then combined with the joint angle sequences to obtain a behavioral feature vector. The timestamp information of the physiological feature vector and the behavioral feature vector is extracted. The time offset is calculated based on the timestamp information and linear interpolation and feature alignment are performed. The aligned physiological feature vector and behavioral feature vector are subjected to nonlinear transformation to obtain physiological intermediate representation and behavioral intermediate representation respectively. The physiological intermediate representation and the behavioral intermediate representation are concatenated and subjected to nonlinear transformation to obtain cross-modal intermediate representation. The cross-modal intermediate representation is subjected to nonlinear transformation and normalized after hyperbolic tangent activation to obtain a unified representation vector. The Pearson correlation coefficients between the physiological and behavioral feature components of the unified representation vector at different times are calculated and arranged according to modality type and time order to obtain the correlation matrix. The correlation matrix is ​​then thresholded to obtain the modality correlation matrix.

3. The method according to claim 1, characterized in that, The baseline features are obtained by segmenting the unified representation vector into sliding time windows and calculating its statistical distribution characteristics. The fused state features are then obtained by performing tensor fusion operations on the baseline features and the modal correlation matrix. The unified representation vector is divided into multiple time window segments according to the preset window length and sliding step size. The central trend index and dispersion index are calculated for the unified representation vector in each time window segment to obtain the distribution concentration feature. The distribution symmetry index and distribution kurtosis index are calculated for the unified representation vector in each time window segment to obtain the distribution morphology feature. The distribution concentration feature and the distribution morphology feature are spliced ​​together in the order of the time window to obtain the baseline feature. The baseline features are arranged into a baseline feature tensor according to the time window and feature dimension. The modal correlation matrix is ​​expanded into a correlation tensor according to the modal type and time order. Tensor contraction operation is performed on the baseline feature tensor and the correlation tensor in the time dimension to obtain a time fusion intermediate tensor. Tensor expansion is performed on the time fusion intermediate tensor in the feature dimension to obtain an expanded feature matrix. The expanded feature matrix is ​​decomposed to extract the dominant feature components. The dominant feature components are reconstructed and combined with the baseline feature tensor to calculate the enhanced baseline tensor. The enhanced baseline tensor and the associated tensor are then subjected to tensor outer product operation to obtain the fused state features.

4. The method according to claim 1, characterized in that, After encoding, compressing, decoding, and reconstructing the fused state features, the reconstruction error is calculated to identify anomalous state segments. Backtracking analysis is performed on these anomalous state segments to extract precursor change patterns, and the dominant influencing factors are determined by combining these patterns with the modal correlation matrix. The fusion state features are subjected to multi-level nonlinear dimensionality reduction transformation and the mean vector and variance vector in the latent space are calculated. Based on the mean vector and variance vector, reparameterized sampling is performed to obtain a compressed coding representation. The divergence measure between the compressed coding representation and the preset prior distribution is calculated and multi-level nonlinear dimensionality increase transformation is performed to obtain the reconstructed fusion state features. The reconstruction loss value between the reconstructed fusion state features and the fusion state features is calculated and a reconstruction error sequence is constructed. The time period in the reconstruction error sequence where the reconstruction loss value exceeds a preset threshold is marked and clustered and merged to obtain abnormal state segments. Extract the abnormal start time from the abnormal state segment and backtrack to determine the backtracking time window. Extract the fusion state features within the backtracking time window and calculate the feature differences between adjacent time steps to construct a change amplitude sequence. Smooth the change amplitude sequence and identify local maxima as key transition nodes. Extract the fusion state features and corresponding change directions at the time corresponding to the key transition nodes to obtain the precursor change pattern. The Pearson correlation coefficient corresponding to the anomaly initiation time is extracted from the modal correlation matrix, and the modal volatility is calculated. The feature component with the largest modal volatility is selected, and the corresponding modal type and feature dimension information are extracted to obtain the dominant influencing factor.

5. The method according to claim 1, characterized in that, Based on the dominant influencing factors, a subset of data is extracted from the unified characterization vector, and the evolutionary trajectory is determined through iterative deduction on a continuous time scale. The deviation between the evolutionary trajectory and the preset health benchmark template is calculated, and a graded mapping is performed based on the deviation to obtain the risk level, including: Based on the dominant influencing factor, time series data corresponding to the feature dimensions are extracted from the unified characterization vector to obtain an initial data subset. The initial data subset is then decomposed into a time series to obtain trend components and calculate the rate of change. Based on the rate of change, abrupt change points are identified and time periods are divided at the abrupt change points to determine the stable evolution stage, thus obtaining the data subset. The data subset is discretized and sampled according to the time step to obtain the initial time state vector. The state transition rule is determined based on the dominant influencing factor, and the initial time state vector is iteratively transformed. The iterative deduction is repeated on a continuous time scale to obtain the predicted state vector sequence and smooth it to obtain the evolution trajectory. A baseline state vector sequence is extracted from a preset health baseline template. The Euclidean distance between the predicted state vector in the evolution trajectory and the baseline state vector in the baseline state vector sequence is calculated to obtain a time step deviation sequence. The time step deviation sequence is then weighted and summed to obtain a comprehensive deviation. Based on the comprehensive deviation and a preset multi-level threshold interval, interval matching is performed to map to a discrete risk level. The risk level is determined based on the discrete risk level and the number of rate of change mutation points.

6. The method according to claim 5, characterized in that, The initial data subset is decomposed into a time series to obtain trend components and calculate the rate of change. Based on the rate of change, abrupt change points are identified, and time periods are divided at these abrupt change points to determine stable evolution stages. The resulting data subset includes: The initial data subset is sorted by time dimension to construct a time index sequence and segmented by sliding window to obtain multiple local time windows. The data in each local time window is fitted with a polynomial to obtain a local trend curve and connected in segments to obtain a global trend component. Based on the initial data subset and the global trend component, the detrended residual sequence is solved and frequency domain transformation is performed to identify the dominant frequency component. The dominant frequency component is inversely transformed to reconstruct the periodic component. The global trend component is subjected to first-order difference operation to obtain a trend change rate sequence. The trend change rate sequence is subjected to second-order difference to obtain an acceleration sequence. A set of candidate mutation points is determined by combining the acceleration sequence with a preset mutation threshold. The amplitude change of the periodic component in the time window before and after each candidate mutation point in the candidate mutation point set is extracted and the amplitude jump ratio is calculated. The candidate mutation points are screened based on the amplitude jump ratio to obtain the rate of change mutation point. The abrupt change point of the rate of change is projected onto the time index sequence to determine the time position, and the initial data subset is divided into multiple time periods based on the time position. The variance of the trend change rate sequence is calculated for each time period to obtain the time period stability index. Based on the time period stability index, a set of stable time periods is determined, and the corresponding data is extracted from the initial data subset and spliced ​​to obtain the data subset.

7. The method according to claim 1, characterized in that, Based on the risk level and the dominant influencing factor, a set of candidate intervention schemes is determined by querying a preset intervention strategy library. For each candidate intervention scheme, a multi-step state transition simulation is performed, and a timeliness score is calculated. Based on the timeliness score, a target intervention scheme is determined, and a monitoring report is generated by combining the abnormal state fragments, including: Intervention strategy entries are extracted from the intervention strategy library and the matching conditions are parsed. The risk level is matched with the risk range in the matching conditions to obtain the risk matching degree. The dominant influencing factor is matched with the influencing factor type in the matching conditions to obtain the factor matching degree. The comprehensive matching degree is calculated based on the risk matching degree and the factor matching degree, and the candidate intervention scheme set is obtained by threshold screening based on the comprehensive matching degree. Intervention parameters for each candidate intervention scheme in the candidate intervention scheme set are extracted and a state update rule is constructed. The predicted state vector at the current moment in the evolution trajectory is used as the simulation starting state, and the state update rule is applied to perform multi-step iterations to obtain a multi-step simulated state sequence. The deviation change trend between the multi-step simulated state sequence and the corresponding moment benchmark state vector in the health benchmark template is calculated to obtain a deviation improvement curve. The area under the deviation improvement curve is calculated, and the timeliness is discounted by combining the response delay in the intervention parameters to obtain a timeliness score. The candidate intervention plan with the highest timeliness score is selected from the set of candidate intervention plans as the target intervention plan. Intervention measures are extracted from the target intervention plan and combined with the abnormal state fragment to generate a monitoring report.

8. A multimodal sensing-based vital sign monitoring system for elderly patients, used to implement the method of any one of claims 1-7, characterized in that, include: The data acquisition unit is used to collect physiological signal data and behavioral state data of elderly patients and perform feature extraction to obtain physiological feature vectors and behavioral feature vectors. After aligning the physiological feature vectors and behavioral feature vectors, multi-layer nonlinear transformation is performed and projected onto a common semantic space to obtain a unified representation vector and construct a modal correlation matrix. An anomaly identification unit is used to perform sliding time window segmentation on the unified representation vector and calculate statistical distribution characteristics to obtain baseline features, perform tensor fusion operation on the baseline features and the modal correlation matrix to obtain fused state features, encode and compress the fused state features, decode and reconstruct them, calculate reconstruction error to identify anomalous state segments, perform backtracking analysis on the anomalous state segments to extract precursor change patterns, and combine them with the modal correlation matrix to determine the dominant influencing factor. The risk grading unit is used to extract a subset of data from the unified characterization vector based on the dominant influencing factor and perform iterative deduction on a continuous time scale to determine the evolution trajectory, calculate the deviation between the evolution trajectory and the preset health benchmark template, and perform a graded mapping based on the deviation to obtain the risk level. The intervention execution unit is used to query a preset intervention strategy library to determine a set of candidate intervention schemes based on the risk level and the dominant influencing factor, perform multi-step state transition simulation for each candidate intervention scheme and calculate the timeliness score, determine the target intervention scheme based on the timeliness score and generate a monitoring report in combination with the abnormal state fragment.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images