Continuous monitoring method and system for multi-modal physiological parameter fusion

Abstract

The invention relates to the technical field of artificial intelligence monitoring, in particular to a continuous monitoring method and system for multi-modal physiological parameter fusion, and the method comprises the steps: carrying out the continuous collection of physiological electric signals of a target individual in a plurality of behavior states; performing period segmentation on the multi-lead electric signal according to the heartbeat rhythm, and standardizing the multi-lead electric signal into characteristic fragments with the same time length and channel number; performing two-stage weighting processing on each feature fragment to obtain high-weight feature representation; performing feature alignment on the high-weight feature representation according to the corresponding behavior state, and executing fusion operation; and inputting the fusion feature representation into a convolutional neural network, performing individual physiological state recognition, and periodically judging whether the target individual has a physiological abnormality risk at present. According to the method and the device, the problem of signal distortion caused by neglecting multi-behavior state switching in the prior art is effectively solved, the feature adaptability and the cross-state recognition robustness in a dynamic scene are remarkably improved, and continuous and accurate physiological anomaly risk tracking is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence monitoring, in particular to a multi-modal physiological parameter fusion continuous monitoring method and system. BACKGROUND

[0002] With the rapid development of wearable devices, edge computing and artificial intelligence, intelligent monitoring of physiological signals has been widely used in health management, chronic disease early warning, sports rehabilitation and other fields. Electrocardiogram (ECG) as one of the core physiological indicators can reflect the changes of the electrical activity of the heart, and it will show significant differences in rhythm, morphology, noise interference and other aspects under different behavior states.

[0003] In the prior art, in the continuous monitoring scene, the common method usually extracts and models features based on a single physiological signal segment in a static scene, which cannot adapt to the non-steady-state characteristics of signals caused by dynamic behavior switching such as walking and running, ignores the fact that people exist in real scenes with multiple behavior state switching such as sitting, walking, running, going up and down stairs, and the signal feature changes greatly under different behavior states, with many interference factors, which poses a challenge to model recognition. In addition, some solutions use static statistical features such as heart rate variability, QRS interval, or use frequency domain features as input, but it is difficult to realize the signal adaptation ability of the model under behavior state changes, resulting in a decrease in cross-state recognition accuracy. SUMMARY

[0004] The present application provides a multi-modal physiological parameter fusion continuous monitoring method and system, thereby effectively solving the problems pointed out in the background art.

[0005] In order to achieve the above purpose, the technical solution adopted by the present application is: The multi-modal physiological parameter fusion continuous monitoring method comprises: continuously collecting physiological electrical signals of a target individual under multiple behavior states, the physiological electrical signals being multi-lead electrical signals; periodically dividing the multi-lead electrical signals according to the heartbeat rhythm, and standardizing them into feature segments with the same time length and channel number; performing two-level weighted processing on each of the feature segments to obtain a high-weight feature representation; aligning the high-weight feature representation according to the corresponding behavior state, and performing fusion operation to obtain a fusion feature representation containing multi-behavior state information; inputting the fusion feature representation into a convolutional neural network composed of channel-by-channel convolution and cross-channel convolution to identify the individual physiological state; based on the output of the convolutional neural network, periodically judging whether the target individual currently has a physiological abnormal risk, and continuously updating the output result.

[0006] Further, the high-weight feature representation is aligned with the corresponding behavior state, and a fusion operation is performed to obtain a fusion feature representation containing multi-behavior state information, including: The high-weight feature representation is divided into several feature groups according to the corresponding behavior state, and the high-weight features are synchronized and aligned on the cycle time axis in each feature group; Global statistical information and local time sequence change information of the fusion feature representation in the cycle time dimension and the feature dimension are extracted respectively; Based on the global statistical information and the local time sequence change information, the weight of the fusion stage is calculated to obtain the fusion feature representation containing multi-behavior state information.

[0007] Further, based on the global statistical information and the local time sequence change information, the weight of the fusion stage is calculated, including: The feature mean vector and the feature standard deviation vector on the global statistical information are first spliced with the vectorized form of the feature correlation matrix in the feature dimension to obtain a first splicing result; The time sequence change amount sequence and the local variance sequence in the local time sequence change information are second spliced to obtain a second splicing result; The first splicing result and the second splicing result are fused to generate a joint feature vector; The joint feature vector is subjected to a nonlinear transformation to output an initial weight value; The initial weight value is normalized to generate a fusion weight coefficient corresponding to each feature group.

[0008] Further, the first weighting of the two-level weighting includes: The lead channel feature response of the multi-lead electrical signal under each behavior state is subjected to average pooling and maximum pooling operations to extract statistical features for representing lead response intensity; According to the statistical features, a first weighting coefficient of each lead channel is calculated; The first weighting coefficient is multiplied by the feature corresponding to each lead channel to obtain a lead feature representation subjected to channel weighting processing.

[0009] Further, the second weighting of the two-level weighting includes: The channel-weighted feature is subjected to a global pooling operation in the original time sequence dimension to extract statistical descriptions reflecting the response change trend in different time sequence segments under the behavior state; According to the statistical description, a second weighting coefficient related to the time position is generated; The second weighting coefficient is multiplied point by point with the feature at each time position to obtain a time-weighted processed time sequence feature representation.

[0010] Further, the convolutional neural network comprises: a depthwise one-dimensional convolutional layer for extracting local time sequence features within each lead channel of the fused feature representation; a channel mapping one-dimensional convolutional layer for establishing feature correlation information between different lead channels based on the output of the depthwise one-dimensional convolutional layer; an output layer for generating an identification result of an individual physiological state.

[0011] Further, the multi-lead electrocardio signal is periodically segmented according to the heartbeat rhythm, comprising: obtaining original data of the multi-lead electrocardio signal within a continuous time period and analyzing rhythm change characteristics under different behavior states; determining a period boundary in the continuous signal according to the rhythm change characteristics, and dividing the period boundary into a plurality of independent period segments.

[0012] Further, the periodicity determines whether the target individual currently has a physiological abnormality risk, and continuously updates the output result, comprising: obtaining a cumulative probability intensity of an abnormal state within a continuous monitoring window based on the state probability output by the convolutional neural network; establishing an adaptive reference according to the historical state data of the target individual, and triggering an abnormality determination when the cumulative probability intensity deviates from the reference beyond a dynamic tolerance range; corresponding each determination result to a time stamp to construct a continuously updated risk state sequence, and refreshing the risk state sequence each time a new time window is analyzed.

[0013] A multi-modal physiological parameter fusion continuous monitoring system, the system comprising: a signal acquisition module for continuously acquiring physiological electrocardio signals of a target individual under a plurality of behavior states, the physiological electrocardio signals being multi-lead electrocardio signals; a period segmentation module for periodically segmenting the multi-lead electrocardio signals according to the heartbeat rhythm and standardizing them into feature segments with the same time length and number of channels; a feature weighting module for performing two-level weighting processing on each of the feature segments to obtain a high-weight feature representation; a feature fusion module for performing feature alignment on the high-weight feature representation according to the corresponding behavior states and performing fusion operation to obtain a fused feature representation containing multi-behavior state information; a state recognition module, which inputs the fused feature representation into a convolutional neural network composed of a channel-by-channel convolution and a cross-channel convolution, to recognize an individual physiological state; a result output module, which periodically determines whether the target individual currently has a risk of physiological abnormality based on the output of the convolutional neural network and continuously updates the output result.

[0014] Further, the feature fusion module comprises: a feature alignment unit, which divides the high-weight feature representation into a plurality of feature groups according to the corresponding behavior states and synchronously aligns the high-weight features on a cycle time axis in each feature group; an information extraction unit, which extracts global statistical information and local time-series change information of the fused feature representation on a cycle time dimension and a feature dimension, respectively; a weight calculation unit, which calculates the weight of a fusion stage based on the global statistical information and the local time-series change information to obtain the fused feature representation containing multi-behavior state information.

[0015] The technical scheme of the present application can achieve the following technical effects: The technical scheme effectively solves the signal distortion problem caused by ignoring multi-behavior state switching in the prior art, significantly improves the feature adaptability and cross-state recognition robustness in a dynamic scene, and realizes continuous and accurate physiological abnormality risk tracking. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical scheme in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0017] Figure 1 Flowchart of the multi-modal physiological parameter fusion continuous monitoring method; Figure 2 Flowchart of obtaining the fused feature representation containing multi-behavior state information; Figure 3 Flowchart of calculating the weight of the fusion stage; Figure 4 Flowchart of the first weighting; Figure 5 Flowchart of the second weighting; Figure 6 Flowchart of the cycle division of the multi-lead electrical signal; Figure 7A flowchart for periodically determining whether a target individual currently has a risk of physiological abnormality and continuously updating the output results. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0020] Example 1 like Figure 1 As shown, the present invention provides a continuous monitoring method for multimodal physiological parameter fusion, the method comprising: S1: Continuously collect physiological electrical signals of the target individual in multiple behavioral states. The physiological electrical signals are multi-lead electrical signals. Specifically, the acquisition method continuously acquires the ECG signals of the target individual, using a multi-lead wearable ECG acquisition system to meet the needs of long-term monitoring under different behavioral states. The system arranges multiple electrodes on the chest and limbs of the target individual according to standard ECG lead specifications to form a twelve-lead acquisition layout, which can comprehensively capture the electrical activity of the heart in multiple directions and positions. The electrode material is preferably a low-impedance, flexible, skin-friendly conductive material to reduce contact impedance changes and artifacts generated during exercise. Each lead electrode is connected to the portable acquisition unit through a flexible connector to ensure stable signal contact quality in different exercise states such as sitting still, walking, jogging, and going up and down stairs. During the acquisition process, the system continuously records the ECG signals of each lead at a high sampling rate, and simultaneously records the behavioral state information of the target individual, thereby providing data annotation for subsequent cross-state analysis. Preferably, the acquisition system has the ability to operate continuously for a long time, and can still obtain ECG waveforms with high signal-to-noise ratio and no obvious drift in a dynamic environment, thereby accurately capturing the electrical activity of the heart.

[0021] S2: Multi-lead electrical signals are segmented into cycles according to the heart rhythm and standardized into feature segments with the same time length and number of channels; Specifically, in order to overcome the problems of rhythm instability, morphological fluctuation and noise interference of electrocardiogram signals affected by behavior state changes in dynamic monitoring scenarios, the application introduces a cycle segmentation strategy based on heartbeat rhythm in the processing stage of multi-lead electrocardiogram signals. The strategy divides the continuous electrocardiogram waveform into cycle segments that are independent of each other and have complete physiological significance by taking the heartbeat as the cycle boundary, so that the different segments have consistency in rhythm structure. Compared with direct segmentation according to a fixed time window, this method can better adapt to the cycle length difference caused by heart rate changes in different behavior states, avoid feature loss or aliasing caused by window and heartbeat cycle misalignment, and further adjust the multi-lead data after cycle segmentation to the same time length and channel number through unified standardization processing, thereby ensuring the comparability and fusion of cross-cycle and cross-lead data in subsequent analysis.

[0022] S3: Two-level weighted processing is performed on each feature segment to obtain a high-weight feature representation; Specifically, the application introduces a two-level weighted processing mechanism to obtain a high-weight feature representation based on the feature segments after cycle segmentation and standardization, in the order of channel weighting first and then time weighting. The first level of weighting addresses the importance differences of different lead channels by analyzing the response intensity of each channel in a specific behavior state, giving higher weights to channels that contribute more to the identification task, and reducing the influence of channels with more noise or poorer stability, thereby completing feature selection in the spatial dimension. The second level of weighting is performed on the features after the first level of weighting, addressing the differences in the time dimension by evaluating the response trend at different time positions within a heartbeat cycle, highlighting the key time periods closely related to the physiological state and reducing the interference of irrelevant time segments. This sequential weighting from channel to time enables the final high-weight feature representation to have both spatial selectivity and temporal attention, retaining the global structure of multi-lead signals and significantly improving the signal-to-noise ratio of features related to the target task.

[0023] S4: Aligning the high-weight feature representation according to the corresponding behavior state and performing fusion operation to obtain a fusion feature representation containing multi-behavior state information; Specifically, the core of this step is to map the high-weight features obtained under different behavior states to a unified reference framework, making them comparable across states, and retaining the key information and differences contained in each state in the fusion result. Since electrocardiogram signals exhibit significant changes in rhythm, amplitude and waveform morphology under states such as sitting, walking, running and climbing stairs, directly mixing these features can lead to feature misalignment and dilution of important information.

[0024] S5: input the fusion feature representation into a convolutional neural network composed of channel-wise convolution and cross-channel convolution for individual physiological state recognition; Specifically, first, local feature patterns are extracted in the time series range of each lead by channel-wise convolution, fully retaining the independence and detail information of each lead feature; then, cross-channel convolution is used to establish association between different leads, integrating spatial correlation and complementary information into a unified feature representation; finally, the integrated features are classified by the output layer to obtain the individual's current physiological state recognition result. This "first channel-wise and then cross-channel" phased processing method not only reduces the redundant calculation caused by traditional fully connected convolution, but also avoids excessive mixing of features between channels, enabling accurate capture of local temporal changes and cross-channel associations in multi-state fusion features, thereby significantly improving computational efficiency and real-time processing capability of the system while ensuring recognition accuracy. In this embodiment, a lightweight CNN model is preferably used, which is not only to reduce computational complexity and power consumption, but also based on the special needs of the present application in channel connection mode and convolution strategy. Unlike traditional convolution which directly performs full connection calculation between all channels, the lightweight CNN separates channel-wise convolution from cross-channel convolution, thereby preserving the independent feature details of each lead while establishing a targeted association relationship between leads. This separated channel connection method reduces unnecessary cross-channel calculation overhead and avoids excessive mixing of channel features by full connection convolution, thereby being more conducive to preserving the spatial structure features of multi-lead ECG signals. In terms of convolution strategy, the lightweight CNN uses a phased processing method of first channel-wise convolution to extract local patterns in the time dimension, and then cross-channel convolution to integrate information between leads, which is different from traditional one-time convolution. This can more accurately capture local changes and cross-channel correlations of ECG signals in multi-state fusion features. This structure not only improves recognition accuracy, but also ensures the real-time running capability of the model in resource-limited environments such as wearable devices, enabling stable and efficient physiological state recognition in dynamic and multi-state ECG monitoring.

[0025] S6: based on the output of the convolutional neural network, periodically determine whether the target individual currently has a physiological abnormality risk, and continuously update the output result.

[0026] Specifically, by continuously monitoring the physiological state of the target individual, the convolutional neural network will output the individual's physiological state probability value in real time, reflecting the normal and abnormal degree of the current state. In each monitoring period, these probability values are used to assess whether there is an abnormal risk, and potential physiological abnormalities are identified when the pre-set threshold is exceeded. At the same time, as new monitoring data is continuously input, the output result is continuously updated, thereby achieving dynamic risk judgment and adjustment.

[0027] By the present application, the signal distortion problem caused by ignoring multi-behavior state switching in the prior art is effectively solved, the feature adaptability and cross-state recognition robustness in a dynamic scene are significantly improved, and continuous and accurate physiological abnormality risk tracking is realized.

[0028] As a preferred embodiment of the above embodiment, as shown in Figure 2 Step S4, the high-weight feature representation is aligned according to the corresponding behavior state, and a fusion operation is performed to obtain a fusion feature representation containing multi-behavior state information, including: S41: dividing the high-weight feature representation into several feature groups according to the corresponding behavior state, and synchronously aligning the high-weight features on the cycle time axis in each feature group; S42: extracting global statistical information and local time series change information of the fusion feature representation in the cycle time dimension and the feature dimension, respectively; S43: calculating the weight of the fusion stage based on the global statistical information and the local time series change information to obtain the fusion feature representation containing multi-behavior state information.

[0029] Specifically, the cycle time dimension and the cycle time axis are two key concepts for describing the periodic fluctuations in the electrocardiogram signal, and they are closely related and jointly constitute an effective analysis framework for the time series features of the electrocardiogram signal. The cycle time dimension refers to the time range of the electrocardiogram signal in a complete cardiac cycle, from the starting point of one cycle to the starting point of the next cycle. In this dimension, each cycle of the electrocardiogram signal has a fixed time length, and reflects all important changes of the cardiac electrical activity in the cycle. The cycle time axis is a specific form of the cycle time dimension, representing all time points of each cardiac cycle, especially the time distribution of different waveform stages in the cycle. The cycle time axis is a specific time line representing the state of each instant of the signal change with time from one R wave to the next R wave. It provides a unified time frame for each cycle of the electrocardiogram signal, enabling the signal features of each cycle to be compared and analyzed in the same time scale.

[0030] The behavior states can include sitting, walking, running, going up and down stairs, and the like. The electrocardiogram signal features in each state can have obvious changes. In this case, the features in different states are divided into different feature groups respectively to ensure that the physiological information in different states can be correctly distinguished during feature processing. For high-weight features in each feature group, further synchronization alignment is performed according to the cycle time axis to ensure that the time position and signal waveform in each cycle can be correctly aligned, thereby ensuring that the signal features between different behavior states can be accurately fused in the subsequent processing process. Second, the aligned features are further processed, and global statistical information and local time sequence change information of the fusion features in the cycle time dimension and the feature dimension are extracted respectively. The global statistical information reflects the overall trend in the entire electrocardiogram signal cycle, including the mean value, the standard deviation and the like, and is mainly used to describe the global features of the physiological signal in different behavior states. The local time sequence change information focuses on describing the physiological fluctuations at some key moments in the cycle, which usually reflects the physiological state changes of the individual in a specific period. After these information is extracted by a specific algorithm, necessary basis is provided for subsequent weighting and fusion. Finally, based on the global statistical information and the local time sequence change information, the weight of the fusion stage is calculated, and the fusion feature representation containing multi-behavior state information is obtained. The purpose of this process is to reasonably fuse the features in different behavior states through weighting, so as to obtain a more comprehensive and representative feature representation. The weighting process combines the characteristics of each behavior state, and gives different weights according to the importance of physiological information in each state. In this way, the final fusion feature representation can better adapt to the differences in signal features in multiple states, and improve the accuracy and robustness of subsequent individual physiological state recognition and anomaly detection.

[0031] As a preferred embodiment of the above embodiment, as shown in Figure 3 the weight of the fusion stage is calculated based on the global statistical information and the local time sequence change information, including: A10: The feature mean value vector and the feature standard deviation vector on the global statistical information are first spliced with the vectorized form of the feature correlation matrix in the feature dimension to obtain a first splicing result; A20: The time sequence change amount sequence and the local variance sequence in the local time sequence change information are second spliced to obtain a second splicing result; A30: The first splicing result and the second splicing result are fused to generate a joint feature vector; A40: The joint feature vector is subjected to nonlinear transformation to output an initial weight value; A50: The initial weight value is normalized to generate a fusion weight coefficient corresponding to each feature group.

[0032] Specifically, first, the feature mean vector and the feature standard deviation vector are extracted from the global statistical information, the feature mean vector represents the average trend of the electrocardiogram signal in the whole period, and reveals the global change characteristics of the electrocardiogram signal; the feature standard deviation vector describes the fluctuation degree of the signal, which embodies the stability and instability of the signal in the time dimension. On this basis, the feature correlation matrix is calculated, which represents the mutual dependence between different features in the signal. After the mean vector, the standard deviation vector and the feature correlation matrix are vectorized and spliced, the first splicing result is obtained. This step combines the global statistical information such as mean and standard deviation with the correlation between features, providing a global information basis for subsequent weighted fusion; then, the local time sequence change information is extracted, including the time sequence change sequence and the local variance sequence. The time sequence change sequence reflects the fluctuation amplitude of the signal in a certain time period, showing the dynamic fluctuation of the signal; while the local variance sequence is used to describe the fluctuation range of the signal in the local time period, revealing the local stability of the signal. By splicing the time sequence change sequence and the local variance sequence, the second splicing result is obtained. This process fully reflects the local time sequence fluctuation information, providing more detailed time sequence features for subsequent fusion weighting; the first splicing result and the second splicing result are merged through splicing operation to generate a joint feature vector. This joint feature vector integrates the global trend and local fluctuation of the electrocardiogram signal, providing a more comprehensive feature expression. This vector enables the global features and local change features in the signal to be integrated, providing rich information for subsequent weighting and recognition; then, a nonlinear transformation is performed on the joint feature vector. The nonlinear transformation introduces activation functions such as ReLU, Sigmoid, etc., so that the feature vector can be mapped to a nonlinear space to capture more complex relationships between features. This process improves the model's ability to adapt to complex patterns of signals. After nonlinear transformation, the output is the initial weight value, which is the weight initial value calculated according to the joint feature vector; finally, the initial weight value is normalized to enable the weights of all feature groups to be compared under the same standard. The purpose of normalization is to eliminate the influence of different dimensions between different features, ensuring the comparability of the weights of each feature group. After normalization, the final fusion weight coefficient determines the contribution size of each feature group in the final feature representation.

[0033] As a preferred embodiment of the above embodiment, as shown in Figure 4 The first weighting of the two-level weighting includes: B10: performing average pooling and maximum pooling operations on the lead channel feature responses of the multi-lead electrical signal in each behavior state to extract statistical features for representing the lead response intensity; B20: calculating the first weighting coefficient of each lead channel according to the statistical features; B30: multiplying the first weighting coefficient with the features corresponding to each lead channel to obtain the lead channel features subjected to the channel weighting processing.

[0034] Specifically, the lead channel features refer to the relevant information of each lead signal extracted from the multi-lead electrocardio signal, including the time domain variation, amplitude, morphology, etc. of the signal, which reflects the performance of the cardiac electrical activity on each lead. In order to ensure that the key information in the signal is highlighted, a two-level weighting first weighting step is adopted. First, the lead channel features of each lead are extracted from the multi-lead electrocardio signal, which includes the amplitude, waveform morphology, frequency, etc. of the signal. In order to extract the global information and the strongest response of the signal, average pooling and max pooling operations are respectively performed on the signal of each lead channel. The average pooling calculates the average value in each time period, reflecting the overall trend of the lead channel signal; while the max pooling selects the maximum value in each time period, emphasizing the strongest response part of the signal. Through these pooling operations, the extracted statistical features can represent the response intensity of the lead under different behavior states, providing a basic description of the signal for subsequent weighting. Next, according to the extracted statistical features, i.e. the average response intensity and the maximum response intensity of each lead channel, the first weighting coefficient of each lead channel is calculated. This weighting coefficient reflects the importance of each lead channel under a specific behavior state. When the signal response of a certain lead is stronger, its weighting coefficient is larger, indicating that the lead contributes more to the change of the physiological state. Conversely, it is given a lower weighting coefficient. This weighting process ensures that the strongest response of the signal under the behavior state is given more attention, thereby enhancing the influence of the key information. Multiplying the first weighting coefficient calculated previously with the original features of each lead channel, through this weighting processing, the features of each lead will be adjusted according to its response intensity under a specific behavior state, so that in the subsequent analysis, the important lead channel features can be paid more attention to. This process ensures that the most relevant lead signal in the signal occupies a more important position in the final feature representation.

[0035] As a preferred embodiment of the above, as shown in Figure 5 the second weighting of the two-level weighting includes: C10: performing a global pooling operation on the features subjected to the channel weighting in the original time sequence dimension to extract a statistical description reflecting the response change trend in different time sequence segments under the behavior state; C20: generating a second weighting coefficient related to the time position according to the statistical description; C30: multiplying the second weighting coefficient with the features at each time position point by point to obtain the time sequence features subjected to the time weighting processing.

[0036] Specifically, the second weighting step of two-stage weighting is used to further weight the lead features obtained after the first weighting processing, ensuring that the time dimension information of the signal is accurately reflected in the final feature representation, and the second weighting mainly focuses on the changes of the signal on the time axis. By weighting the time dimension, the physiological features of the key time segments under different behavior states are highlighted.

[0037] The original time axis refers to the complete sequence of all time points from the beginning to the end of the signal acquisition process, which is different from the cycle time axis. The original time axis spans multiple cardiac cycles and can contain changes of multiple behavior states. First, on the original time axis, a global pooling operation is performed on each lead signal after the first weighting. The global pooling calculates the statistical quantities of the features within the entire original time axis, such as mean, maximum, minimum, etc., thereby extracting the global change trend and fluctuation of the signal under different behavior states. Through this global pooling operation, the signal response change trend spanning multiple cycles and behavior states can be captured, providing complete time sequence features for subsequent weighting processing. Subsequently, based on the extracted statistical description, the second weighting coefficients are calculated according to the change features of each time position on the original time axis. These weighting coefficients reflect the relative importance of the signal at certain time or time period on the original time axis under different behavior states. If the signal change in a certain period is large, the weighting coefficient of that period will be higher, indicating that the signal in that period has a greater contribution to the overall state change. These weighting coefficients provide the basis for subsequent time weighting processing. Finally, the calculated second weighting coefficients are multiplied point by point with the feature values of each time position to form the weighted time sequence feature representation. This operation adjusts the feature values at each time point according to their corresponding weighting coefficients, thereby highlighting the time periods with important physiological significance on the original time axis. Through this weighting processing, the important information of the signal in the time dimension can be strengthened, providing more accurate input for subsequent feature fusion and physiological state recognition.

[0038] As a preferred embodiment of the above, the convolutional neural network comprises: a depthwise one-dimensional convolutional layer for extracting local time sequence features within each lead channel of the fused feature representation; a channel mapping one-dimensional convolutional layer for establishing feature correlation information between different lead channels based on the output of the depthwise one-dimensional convolutional layer; an output layer for generating the recognition result of the individual physiological state.

[0039] Specifically, first, the fused feature representation is processed by a Depthwise one-dimensional convolution layer. Depthwise convolution is different from traditional full convolution methods. It performs convolution operations independently within each lead channel without cross-computing features between different channels. This convolution method is specifically used to extract local time sequence features within each lead channel, i.e., local change information of each lead in the time dimension. These local time sequence features reflect the physiological state changes of the individual in a specific time period. Through Depthwise convolution, each lead channel can retain its unique time sequence features, reducing the amount of calculation and avoiding unnecessary interference between channels, focusing on signal feature extraction of each lead.

[0040] As a preferred embodiment of the above embodiment, as shown in Figure 6 the periodic segmentation of the multi-lead electrical signal according to the heartbeat rhythm includes: D10: Obtain the original data of the multi-lead electrical signal in a continuous time period, and analyze the rhythm change characteristics under different behavior states; D20: According to the rhythm change characteristics, determine the period boundary in the continuous signal, and divide the period boundary into multiple independent period segments.

[0041] Specifically, in this step, first, signals are extracted from the original data of multi-lead electrical signals, which are usually from the electrical physiological data of multiple lead electrodes in a continuous time period. Different behavioral states during signal acquisition can cause changes in the rhythm of electrocardiogram signals. For example, in a sitting state, the heart rate is slow and the electrocardiogram morphology is stable. When running, the heart rate increases, and the frequency and amplitude of the electrocardiogram waveform change. By analyzing the timing changes of these signals, rhythm change features can be extracted. Rhythm change features refer to the periodic fluctuations and change patterns exhibited by electrocardiogram signals under different behavioral states, which are usually described by timing features such as heart rate, waveform morphology, and frequency. Rhythm change features help identify periodic changes in signals, allowing accurate division of cycle boundaries. Based on the extracted rhythm change features, the periodic fluctuations in the signal are analyzed, and a feature detection algorithm is used to determine the cycle boundaries in the signal. Cycle boundaries refer to the starting and ending points of each cardiac cycle in the electrocardiogram signal, which are the basis for cycle segmentation. By identifying these cycle boundaries, continuous signals can be divided into multiple independent cycle segments, each representing a complete cardiac cycle. In this way, signals can be segmented according to the rhythm of the heartbeat, providing accurate periodic data for subsequent feature extraction, weighting processing, and recognition analysis. Each independent cycle segment corresponds to a complete electrocardiogram cycle in time, and since the rhythm change features are fully considered, cycle segmentation can adapt to changes in electrocardiogram signals under different behavioral states.

[0042] As a preferred embodiment of the above embodiment, as shown in Figure 7 determines whether the target individual currently has a physiological abnormality risk and continuously updates the output result, including: E10: Based on the state probability output by the convolutional neural network, obtain the cumulative probability intensity of the abnormal state within the continuous monitoring window; E20: Establish an adaptive baseline based on the historical state data of the target individual, and trigger an abnormality judgment when the cumulative probability intensity deviates from the baseline beyond a dynamic tolerance range; E30: Correspond the judgment result of each time to a time stamp, and construct a continuously updated risk state sequence, which is refreshed when a new time window analysis is completed.

[0043] Specifically, first, according to the state probability output by the convolutional neural network, the current physiological state of the target individual is evaluated, and the state probability represents the probability of the target individual being in a normal physiological state or an abnormal state at a certain time, which quantifies the duration of the target individual being in an abnormal state at a certain time. In a continuous monitoring window, the cumulative probability intensity of the abnormal state is calculated, which represents the probability value of the target individual being in an abnormal state continuously within the time window, quantifying the duration of the target individual being in an abnormal state at a certain time. By calculating this cumulative probability intensity, the trend of the physiological state of the target individual can be tracked, and potential abnormal patterns in the signal can be identified, providing a basis for subsequent abnormal risk judgment; next, according to the historical state data of the target individual, an adaptive baseline is established, which describes the fluctuation range of the target individual in the normal physiological state, and is dynamically updated over time and with changes in physiological state. When new data is input, the current cumulative probability intensity will be calculated and compared with the historical baseline. If the cumulative probability intensity deviates from the baseline and exceeds the set dynamic tolerance range, it is determined that there is a risk of physiological abnormalities. The dynamic tolerance range can be adjusted adaptively according to the long-term physiological changes of the target individual, ensuring the individualization and accuracy of the abnormal judgment; each time the abnormality is judged, the judgment result is associated with the corresponding time stamp, and the time point of each judgment result is recorded and stored in the risk state sequence, which reflects the physiological risk state of the target individual at different time periods. When each new monitoring window is analyzed, the risk state sequence is refreshed to ensure the timeliness and accuracy of the sequence. In this way, the physiological state of the target individual can be continuously tracked and the risk state can be updated in real time.

[0044] Embodiment Two Based on the same inventive concept as the continuous monitoring method of multi-modal physiological parameters in the foregoing embodiments, the present application also provides a continuous monitoring system of multi-modal physiological parameters, which comprises: A signal acquisition module continuously acquires physiological electrical signals of the target individual in multiple behavior states, and the physiological electrical signals are multi-lead electrical signals. A cycle segmentation module segments the multi-lead electrical signals according to the heartbeat rhythm and standardizes them into feature segments with the same time length and channel number. A feature weighting module performs two-level weighting processing on each feature segment to obtain a high-weight feature representation. A feature fusion module aligns the high-weight feature representation according to the corresponding behavior state and performs fusion operation to obtain a fusion feature representation containing multi-behavior state information. A state recognition module inputs the fusion feature representation into a convolutional neural network composed of channel-by-channel convolution and cross-channel convolution for individual physiological state recognition. The result output module periodically determines whether the target individual currently has a physiological abnormality risk based on the output of the convolutional neural network, and continuously updates the output result.

[0045] The monitoring system in the application can effectively realize the multi-modal physiological parameter fusion continuous monitoring method, and the technical effects are as described above. Here, no longer described.

[0046] As a preferred embodiment of the above-mentioned embodiment, the feature fusion module comprises: The feature alignment unit divides the high-weight feature representation into several feature groups according to the corresponding behavior state, and synchronously aligns the high-weight feature on the cycle time axis in each feature group. The information extraction unit extracts the global statistical information and the local time sequence change information of the fusion feature representation on the cycle time dimension and the feature dimension, respectively. The weight calculation unit calculates the weight of the fusion stage based on the global statistical information and the local time sequence change information, and obtains the fusion feature representation containing multi-behavior state information.

[0047] Similarly, the above-mentioned optimization scheme of the system can also correspondingly realize the optimization effect of the method in Embodiment 1, which will not be described here.

[0048] Although the present application has been described in connection with specific features and embodiments thereof, it will be evident that many modifications and variations can be made thereto without departing from the spirit and scope of the application. Accordingly, it is intended to embrace all modifications and variations within the scope of the application. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and its equivalent technology, the present application is intended to include these modifications and variations.

Claims

1. A continuous monitoring method for multimodal physiological parameter fusion, characterized in that: include: Continuously collecting physiological electrical signals of a target individual under multiple behavioral states, wherein the physiological electrical signals are multi-lead electrical signals; Performing period segmentation on the multi-lead electrical signals according to the heart rhythm, and standardizing the signals into characteristic segments with the same time length and number of channels; Performing two-level weighting processing on each of the feature fragments to obtain a high-weight feature representation; Performing feature alignment on the high-weight feature representations according to the corresponding behavior states, and performing a fusion operation to obtain a fused feature representation containing multiple behavior state information; Inputting the fused feature representation into a convolutional neural network composed of channel-by-channel convolution and cross-channel convolution to perform individual physiological state recognition; Based on the output of the convolutional neural network, it is periodically determined whether the target individual currently has a risk of physiological abnormality, and the output result is continuously updated.

2. The continuous monitoring method of multimodal physiological parameter fusion according to claim 1, characterized in that: Performing feature alignment on the high-weight feature representations according to the corresponding behavior states and performing a fusion operation to obtain a fused feature representation containing multiple behavior state information, including: Dividing the high-weight feature representations into a plurality of feature groups according to the corresponding behavior states, and synchronously aligning the high-weight features on a periodic time axis within each feature group; Extracting global statistical information and local temporal variation information of the fused feature representation in the periodic time dimension and the feature dimension respectively; The weight of the fusion stage is calculated based on the global statistical information and the local temporal change information to obtain the fusion feature representation containing the multi-behavior state information.

3. The continuous monitoring method of multimodal physiological parameter fusion according to claim 2, characterized in that: Calculating the weight of the fusion stage based on the global statistical information and the local time series change information includes: Performing a first-level splicing on the feature mean vector and the feature standard deviation vector on the global statistical information and the vectorized form of the feature correlation matrix on the feature dimension to obtain a first-level splicing result; Performing second-level splicing on the time series variation sequence and the local variance sequence in the local time series variation information to obtain a second-level splicing result; fusing the first-level splicing result and the second-level splicing result to generate a joint feature vector; Performing a nonlinear transformation on the joint eigenvector and outputting an initial weight value; The initial weight values ​​are normalized to generate fusion weight coefficients corresponding to each feature group.

4. The continuous monitoring method of multimodal physiological parameter fusion according to claim 1, characterized in that: The first weighting of the two-level weighting includes: performing average pooling and maximum pooling operations on the lead channel characteristic responses of the multi-lead electrical signals in each behavioral state to extract statistical features for characterizing the lead response strength; Calculating a first weighting coefficient of each lead channel according to the statistical characteristics; The first weighting coefficient is multiplied by the feature corresponding to each of the lead channels to obtain a lead feature representation after channel weighting processing.

5. The continuous monitoring method of multimodal physiological parameter fusion according to claim 1, characterized in that: The second weighting of the two-level weighting includes: Perform a global pooling operation on the channel-weighted features in the original time series dimension to extract statistical descriptions of the response change trends in different time series segments under the behavioral state; generating a second weighting coefficient associated with the time position according to the statistical description; The second weighting coefficient is multiplied point by point by the feature at each time position to obtain a time series feature representation after time weighting processing.

6. The continuous monitoring method of multimodal physiological parameter fusion according to claim 1, characterized in that: The convolutional neural network comprises: Depthwise one-dimensional convolution layer, used to extract local temporal features in each lead channel represented by the fusion feature; A channel mapping one-dimensional convolutional layer, used to establish feature association information between different lead channels based on the output of the Depthwise one-dimensional convolutional layer; The output layer is used to generate recognition results of individual physiological states.

7. The continuous monitoring method of multimodal physiological parameter fusion according to claim 1, characterized in that: The multi-lead electrical signal is periodically segmented according to the heart rhythm, including: Acquiring raw data of the multi-lead electrical signals in continuous time periods, and analyzing rhythmic change characteristics under different behavioral states; According to the rhythmic change characteristics, period boundaries in the continuous signal are determined, and the period boundaries are divided into multiple independent period segments.

8. The continuous monitoring method of multimodal physiological parameter fusion according to claim 1, characterized in that: Periodically determine whether the target individual currently has a risk of physiological abnormality and continuously update the output results, including: Based on the state probability output by the convolutional neural network, obtaining the cumulative probability intensity of abnormal states in the continuous monitoring window; Establishing an adaptive benchmark based on the historical status data of the target individual, and triggering an abnormality judgment when the cumulative probability intensity deviates from the benchmark beyond a dynamic tolerance range; Each determination result is associated with a timestamp to construct a continuously updated risk status sequence, which is refreshed each time a new time window analysis is completed.

9. A continuous monitoring system integrating multimodal physiological parameters, characterized in that: The system comprises: A signal acquisition module continuously acquires physiological electrical signals of a target individual under multiple behavioral states, wherein the physiological electrical signals are multi-lead electrical signals; a cycle segmentation module, which performs cycle segmentation on the multi-lead electrical signal according to the heart rhythm and standardizes the signal into characteristic segments with the same time length and number of channels; A feature weighting module performs two-level weighting processing on each feature segment to obtain a high-weight feature representation; A feature fusion module aligns the high-weight feature representations according to the corresponding behavior states and performs a fusion operation to obtain a fused feature representation containing multiple behavior state information; A state recognition module inputs the fused feature representation into a convolutional neural network composed of channel-by-channel convolution and cross-channel convolution to perform individual physiological state recognition; The result output module periodically determines whether the target individual currently has a risk of physiological abnormality based on the output of the convolutional neural network, and continuously updates the output result.

10. The continuous monitoring system for multimodal physiological parameter fusion according to claim 9, characterized in that: The feature fusion module includes: a feature alignment unit, which divides the high-weight feature representations into a plurality of feature groups according to the corresponding behavior states, and synchronously aligns the high-weight features on a periodic time axis within each feature group; An information extraction unit extracts global statistical information and local temporal variation information represented by the fused feature in the periodic time dimension and the feature dimension respectively; The weight calculation unit calculates the weight of the fusion stage based on the global statistical information and the local temporal change information to obtain the fusion feature representation containing the multi-behavior state information.

Images