Fall risk early warning system and method

Abstract

This application provides a fall risk warning system and method, relating to the field of computer technology. The system includes: an EEG acquisition module for acquiring a user's EEG signals; a motion state monitoring module for acquiring motion state data of the user's target nodes; a multimodal fusion analysis module for receiving the EEG signals and the motion state data; decoding the EEG signals to obtain a predicted motion vector; the predicted motion vector representing the user's current motion intention; determining the user's actual motion state vector based on the motion state data; calculating a dynamic mismatch index between the predicted motion vector and the actual motion state vector; generating a risk assessment result based on the dynamic mismatch index; and a warning execution module for executing a warning action corresponding to the risk level indicated by the risk assessment result. This system can provide timely and accurate fall risk warnings.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a fall risk early warning system and method. Background Technology

[0002] Falls are a major cause of disability and impairment among the elderly. Current fall monitoring technologies mainly fall into two categories: post-fall detection and long-term risk assessment. Post-fall detection methods capture motion characteristics at the time of a fall using inertial sensors or vision devices, triggering a response only after the fall has been completed or the body has tilted. Long-term risk assessment methods predict chronic fall tendencies through gait parameter statistical modeling, but cannot address sudden, situational, instantaneous instability. Summary of the Invention

[0003] In view of this, this application provides a fall risk early warning system and method, which can provide timely and accurate fall risk warnings. The specific solution is as follows: A fall risk warning system includes: The EEG acquisition module is used to acquire the user's EEG signals; The motion status monitoring module is used to collect motion status data of the user's target node; A multimodal fusion analysis module is used to receive the EEG signals and the motion state data; decode the EEG signals to obtain the expected motion vector; the expected motion vector represents the user's current motion intention; determine the user's actual motion state vector based on the motion state data; calculate the dynamic mismatch index between the expected motion vector and the actual motion state vector; and generate a risk assessment result based on the dynamic mismatch index. The early warning execution module is used to execute the early warning action corresponding to the risk level indicated by the risk assessment result.

[0004] Optionally, the multimodal fusion analysis module in the above system includes: The feature extraction submodule is used to extract features from the EEG signal based on a temporal feature extraction algorithm to obtain neural activity features; An execution submodule is used to input the extracted neural activity features into a first decoder to obtain the expected motion vector containing at least one of motion direction features, velocity estimates, and intensity estimates.

[0005] Optionally, in the aforementioned system, the multimodal fusion analysis module is specifically used for: The motion state data is processed based on the attitude calculation algorithm to obtain the actual motion state vector containing at least one of the following: current torso tilt angle, center of gravity projection trajectory, and gait cycle parameters.

[0006] Optionally, in the above system, the multimodal fusion analysis module is used for: The expected motion vector sequence in the first time series and the actual motion state vector sequence in the second time series are input into a deep neural network model, so that the alignment network layer in the deep neural network model performs temporal feature alignment and mapping on the expected motion vector sequence and the actual motion state vector sequence; the aligned feature map is input into a multi-head attention mechanism layer to calculate the spatial domain and channel domain correlation weights between cross-modal features; the weighted fused feature vector is input into a multilayer perceptron, and the multilayer perceptron outputs the dynamic mismatch index at the current time.

[0007] Optionally, in the above system, the multimodal fusion analysis module is used for: When the deep neural network model calculates the dynamic mismatch index, it uses the user's historical baseline feature parameters and environmental context feature parameters as auxiliary inputs to correct the dynamic mismatch index.

[0008] Optionally, in the aforementioned system, the early warning execution module includes: The first early warning execution submodule is used to trigger an audio-visual prompt action for the user when the dynamic mismatch index exceeds a preset first risk threshold but does not exceed a second risk threshold. The second early warning execution submodule is used to send an early warning notification to the associated terminal when the dynamic mismatch index exceeds a preset second risk threshold but does not exceed a third risk threshold. The third early warning execution submodule is used to respond to the dynamic mismatch index exceeding the preset third risk threshold and to link auxiliary equipment to perform physical intervention actions.

[0009] Optionally, the motion state data in the above-described system includes at least one of three-dimensional acceleration data, three-dimensional angular velocity data, plantar pressure distribution data, and electromyographic signal data.

[0010] A fall risk warning method includes: Acquire the user's brainwave signals and motion state data; The EEG signal is decoded to obtain the expected motion vector; the expected motion vector represents the user's current motion intention. The user's actual motion state vector is determined based on the motion state data; Calculate the dynamic mismatch index between the expected motion vector and the actual motion state vector; Risk assessment results are generated based on the dynamic mismatch index. Execute the early warning action corresponding to the risk level indicated by the risk assessment results.

[0011] Optionally, in the above method, decoding the EEG signal to obtain the expected motion vector includes: The EEG signal is used to extract features based on a time-series feature extraction algorithm to obtain neural activity features; The extracted neural activity features are input into a first decoder to obtain the expected motion vector containing at least one of motion direction features, velocity estimates, and intensity estimates.

[0012] Optionally, in the above method, calculating the dynamic mismatch index between the expected motion vector and the actual motion state vector includes: The expected motion vector sequence in the first time series and the actual motion state vector sequence in the second time series are input into a deep neural network model, so that the alignment network layer in the deep neural network model performs temporal feature alignment and mapping on the expected motion vector sequence and the actual motion state vector sequence; the aligned feature map is input into a multi-head attention mechanism layer to calculate the spatial domain and channel domain correlation weights between cross-modal features; the weighted fused feature vector is input into a multilayer perceptron, and the multilayer perceptron outputs the dynamic mismatch index at the current time.

[0013] Based on the above, this application provides a fall risk warning system and method. The system includes: an EEG acquisition module for acquiring a user's EEG signals; a motion state monitoring module for acquiring motion state data of the user's target nodes; a multimodal fusion analysis module for receiving the EEG signals and the motion state data; decoding the EEG signals to obtain a predicted motion vector; the predicted motion vector representing the user's current motion intention; determining the user's actual motion state vector based on the motion state data; calculating a dynamic mismatch index between the predicted motion vector and the actual motion state vector; generating a risk assessment result based on the dynamic mismatch index; and a warning execution module for executing a warning action corresponding to the risk level indicated by the risk assessment result. This system can provide timely and accurate fall risk warnings. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0015] Figure 1A schematic diagram of a fall risk warning system provided in this application; Figure 2 A schematic diagram of the structure of a multimodal fusion analysis module provided in this application; Figure 3 This application provides a schematic diagram of the structure of an early warning execution module; Figure 4 A flowchart of a fall risk warning method provided in this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0018] This application provides a fall risk warning system, the structural diagram of which is shown below. Figure 1 As shown, it specifically includes: The EEG acquisition module 101 is used to acquire the user's EEG signals; wherein, EEG signals refer to the neurophysiological signals related to motor preparation, motor imagery, or motor execution detected from the user's scalp surface or cortical region. EEG signals appear at the stage when the user has a clear intention to move but before the limbs have made significant displacement, and therefore can be used as a neurophysiological basis for inferring motor intention.

[0019] Optionally, the EEG acquisition module 101 can be configured as a wearable headband device containing multiple electrode channels to achieve continuous signal acquisition in a non-invasive manner. During the acquisition process, the raw signal can be amplified and converted from analog to digital to obtain a discrete time series that can be used for subsequent analysis. Suppression of power line interference and electromyography artifacts can be performed simultaneously with the acquisition to preserve the rhythmic components and preparatory potential components associated with motor intention.

[0020] The motion state monitoring module 102 is used to collect motion state data of the user's target nodes. The target nodes refer to pre-selected anatomical locations on the user's body that reflect overall motion posture and stability, such as one or more of the head, trunk, bilateral shoulder joints, hip joints, knee joints, and ankle joints. The motion state data characterizes the kinematic features of the corresponding nodes in three-dimensional space and may encompass one or more combinations of three-dimensional acceleration data, three-dimensional angular velocity data, plantar pressure distribution data, and electromyographic signal data.

[0021] Optionally, the three-dimensional acceleration data represents the linear acceleration change of the node along three orthogonal axes, the three-dimensional angular velocity data represents the rotational angular rate of the node about three orthogonal axes, the plantar pressure distribution data reflects the magnitude and distribution of pressure between different areas of the foot and the supporting surface, and the electromyographic signal data reflects the action potential firing of the motor units of the target muscle group. In implementation, the motion state monitoring module 102 can be composed of distributed sensing units.

[0022] The multimodal fusion analysis module 103 is used to receive EEG signals and motion state data; decode the EEG signals to obtain the expected motion vector; the expected motion vector represents the user's current motion intention; determine the user's actual motion state vector based on the motion state data; calculate the dynamic mismatch index between the expected motion vector and the actual motion state vector; and generate a risk assessment result based on the dynamic mismatch index.

[0023] Specifically, the expected motion vector is a multidimensional feature vector extracted from electroencephalogram (EEG) signals. Its dimensions may include at least one of the following: an estimate of the direction of movement, an estimate of the speed of movement, and an estimate of the intensity of movement. The expected motion vector reflects the instructions issued by the user's central nervous system to the motor system before a fall occurs.

[0024] The actual motion state vector is a multi-dimensional feature vector calculated from motion state data, used to describe the user's actual execution of commands at the current moment. The actual motion state vector may include at least one of the following: current trunk tilt angle, center of gravity projection trajectory, gait cycle parameters, and interlimb coordination parameters.

[0025] The Dynamic Mismatch Index (VMI) is a real-time calculated numerical metric used to measure the instantaneous deviation between the expected motion vector and the actual motion vector. When the VMI exceeds a predetermined range, it indicates an unexpected deviation between the movement intention and the body's execution capability, suggesting a risk of fall.

[0026] The early warning execution module 104 is used to execute early warning actions corresponding to the risk level indicated by the risk assessment results.

[0027] The early warning execution module 104 triggers different levels of intervention measures based on the numerical range of the dynamic mismatch index.

[0028] The system provided in this embodiment, by simultaneously collecting EEG signals and motor state data, quantifies the instantaneous mismatch between motor intention and actual execution state. This allows for the identification of fall risks before significant bodily imbalance occurs, providing a window for proactive intervention. It enables timely and accurate fall risk warnings.

[0029] In one embodiment provided in this application, based on the above-described solution, optionally, a multimodal fusion analysis module, such as... Figure 2 As shown, it includes: The feature extraction submodule 201 is used to extract features from the electroencephalogram signal based on a temporal feature extraction algorithm to obtain neural activity features. Optionally, the temporal feature extraction algorithm can employ a co-spatial pattern algorithm. The co-spatial pattern algorithm constructs a set of spatial filters by jointly diagonalizing the covariance matrices of two or more classes of EEG signals, maximizing the variance difference of the filtered signals. For the motor intent decoding task, EEG signal segments corresponding to different movement directions or different movement patterns can be treated as different categories, and a set of spatial filters with discriminative capabilities can be trained. The spatially filtered signal is the time series of neural activity features.

[0030] Optionally, the temporal feature extraction algorithm can also employ the filter bank common-space pattern algorithm. The filter bank common-space pattern algorithm introduces a parallel processing structure of multiple sub-frequency bands based on the common-space pattern. In implementation, the original EEG signal can be divided into several frequency bands, a common-space pattern filter can be constructed for each frequency band, and the variance features of that frequency band can be extracted. Finally, the features extracted from each frequency band are concatenated to form a neural activity feature vector containing multi-frequency band information.

[0031] Optionally, the temporal feature extraction algorithm can also employ a time-frequency analysis method based on wavelet transform. Wavelet transform obtains the energy distribution of the signal at different time scales and frequency resolutions by performing an inner product operation on the signal with a set of scaled and translated wavelet basis functions. In practice, discrete wavelet transforms can be performed on each channel of a continuous EEG signal to obtain wavelet coefficient sequences at each scale, and the energy or entropy values ​​of the wavelet coefficients within a specific scale range can be extracted as neural activity features.

[0032] In practice, the extracted neural activity features can be feature vectors from a single sampling moment or aggregated representations of features from multiple sampling points within a sliding time window. When using a sliding time window, the window length can be determined based on the non-stationary characteristics of the EEG signal and real-time requirements, and the window sliding step size can be set to one-quarter to one-half of the window length. Feature aggregation methods within the window can include taking the mean, taking the maximum value, or extracting temporal dependency representations through a recursive network structure.

[0033] The execution submodule 202 is used to input the extracted neural activity features into the first decoder to obtain the expected motion vector containing at least one of motion direction features, velocity estimates, and intensity estimates.

[0034] In this embodiment, the first decoder is a pre-trained classification or regression model. Its input is neural activity features, and its output is the numerical values ​​or category labels of the parameters of each dimension in the expected motion vector. The training data for the first decoder can be obtained through offline calibration experiments. In the calibration experiments, the user performs motion tasks of different directions, speeds, or intensities according to preset instructions, and the EEG signals are recorded simultaneously and the corresponding ground truth values ​​of the motion parameters are labeled to obtain a training dataset. Based on this training dataset, the first decoder learns the mapping function between neural activity features and motion intention parameters.

[0035] When the first decoder is a classification model, its output is discrete motion direction features. Taking motion direction feature decoding as an example, the classification model can output probability distributions corresponding to direction categories such as "forward," "backward," "left," "right," and "stop," and use the direction information corresponding to the category with the highest probability value as the motion direction feature component in the expected motion vector. The network structure of the classification model can be a linear discriminant analysis classifier, a support vector machine, or a multilayer perceptron.

[0036] When the first decoder is a regression model, its output is continuous numerical values ​​of motion parameters. Taking velocity estimation decoding as an example, the regression model can directly output the velocity scalar value or the values ​​of each component of the velocity vector; taking intensity estimation decoding as an example, the regression model can directly output numerical values ​​representing the intensity of the motion. The network structure of the regression model can be a ridge regression model, support vector regression, or a fully connected regression layer based on a deep neural network.

[0037] In scenarios where at least two parameters among motion direction features, velocity estimates, and intensity estimates are decoded simultaneously, the first decoder can be designed as a multi-task learning structure, sharing the underlying neural activity feature extraction network part, and setting independent output branches at the output end for motion direction feature classification, velocity estimate regression, and intensity estimate regression, respectively.

[0038] In one embodiment provided in this application, based on the above-described solution, optionally, the multimodal fusion analysis module is specifically used for: The motion state data is processed based on the attitude calculation algorithm to obtain the actual motion state vector containing at least one of the following parameters: current torso tilt angle, center of gravity projection trajectory, and gait cycle parameter.

[0039] In this embodiment, the actual motion state vector represents the user's body's actual execution state of neural commands at the current moment. After receiving motion state data from the motion state monitoring module, the multimodal fusion analysis module performs fusion processing on the motion state data based on the posture calculation algorithm to generate an actual motion state vector containing at least one of the following parameters: current torso tilt angle, center of gravity projection trajectory, and gait cycle parameter.

[0040] The attitude estimation algorithm fuses one or more of the following data collected at the target node: 3D acceleration data, 3D angular velocity data, and magnetometer data, to estimate the 3D orientation of each target node in the global coordinate system. Optionally, the attitude estimation algorithm can employ a quaternion-based extended Kalman filter algorithm. This algorithm represents the attitude of the target node as a quaternion, uses angular velocity measurements as the prediction basis for system state transitions, and uses the deviations between acceleration measurements and the gravity reference direction, and the deviations between magnetometer measurements and the geomagnetic reference direction, as observation update quantities. Through a prediction-update iteration process, it suppresses gyroscope integral drift and compensates for the interference of motion acceleration on attitude estimation.

[0041] Optionally, the attitude estimation algorithm can also employ a complementary filtering algorithm. The complementary filtering algorithm leverages the complementary characteristics of the gyroscope's fast dynamic response and the accelerometer and magnetometer's good low-frequency stability. It weights and fuses the high-frequency attitude changes obtained from gyroscope integration with the low-frequency attitude references calculated by the accelerometer and magnetometer, thereby obtaining attitude estimation results with lower computational overhead.

[0042] After obtaining the three-dimensional orientation of each target node in the global coordinate system, the multimodal fusion analysis module calculates the current torso tilt angle. The torso tilt angle characterizes the degree of deviation of the torso's longitudinal axis relative to the vertical line of gravity. In practice, the attitude quaternion of the target node corresponding to the torso can be converted into a rotation matrix. The component corresponding to the direction of gravity in this matrix is ​​extracted, and the angle between this component and the vertical unit vector is calculated. The resulting angle is the current torso tilt angle. Continuous monitoring of the torso tilt angle can reflect the stability changes of the torso segment during the user's dynamic balance maintenance.

[0043] The center of gravity projection trajectory is calculated by fusing kinematic data from each target node and human segment parameter models. Specifically, based on the three-dimensional position of each target node and the pre-calibrated positions and mass ratios of the center of gravity of each human segment, the instantaneous position of the center of gravity can be calculated segment by segment, and the summation yields the three-dimensional coordinates of the whole body's center of gravity in the global coordinate system. The three-dimensional coordinates of the whole body's center of gravity are projected onto the horizontal support surface to obtain the center of gravity projection point. The trajectory formed by the change of the center of gravity projection point over time is the center of gravity projection trajectory. The distribution range, movement speed, and relative positional relationship of the center of gravity projection trajectory with the support surface boundary constitute important criteria for assessing the body's stability. When the center of gravity projection point exceeds the foot support surface boundary and fails to retract in time, it indicates that the body has entered an unstable state that cannot be recovered spontaneously.

[0044] Gait cycle parameters are obtained by analyzing the temporal patterns of motion data from lower limb target nodes. Optionally, gait events, such as heel strike and toe-off events, can be identified using angular velocity signals or plantar pressure distribution data from foot target nodes. The gait cycle duration is calculated based on the time interval between adjacent heel strike events on the same side; the ratio of the duration of the double support phase to the duration of the single support phase is calculated based on the time difference between the left and right heel strike events and the gait cycle duration; and the stride length is calculated based on the displacement amplitude of the foot target nodes within the gait cycle. Real-time changes in gait cycle parameters can reflect the instantaneous fluctuations in gait rhythm and symmetry.

[0045] The actual motion state vector is composed of at least one combination of the current trunk tilt angle, the center of gravity projection trajectory, and the gait cycle parameters mentioned above. In practice, the actual motion state vector can be organized as a multi-dimensional feature vector, where each dimension corresponds to the instantaneous value or short-term statistical value of the kinematic parameters. For example, one dimension corresponds to the trunk tilt angle amplitude, two dimensions correspond to the shortest distance from the center of gravity projection point to the support surface boundary, and three dimensions correspond to the current gait phase index or the proportion of the dual support phase.

[0046] In one embodiment provided in this application, based on the above-described solution, optionally, the multimodal fusion analysis module is used for: The expected motion vector sequence in the first time series and the actual motion state vector sequence in the second time series are input into a deep neural network model. The alignment network layer in the deep neural network model performs temporal feature alignment and mapping between the expected motion vector sequence and the actual motion state vector sequence. The aligned feature map is input into a multi-head attention mechanism layer to calculate the spatial and channel domain correlation weights between cross-modal features. The weighted fused feature vector is input into a multilayer perceptron, which outputs the dynamic mismatch index at the current time.

[0047] Deep neural network models include alignment network layers, multi-head attention mechanism layers, and multilayer perceptrons.

[0048] The first time sequence refers to a segment of time taken backward from the current moment, containing the expected motion vector of multiple consecutive sampling points. The second time sequence refers to another segment of time taken backward from the current moment, containing the actual motion state vector of multiple consecutive sampling points. The lengths of the first and second time sequences can be the same or different. In practice, the lengths of the first and second time sequences can be set based on the statistical characteristics of the response delay of the neuromuscular system. Specifically, there is an inherent transmission and execution delay from the generation of motor intention in the cerebral cortex to the corresponding muscle activation and limb movement performance, which can be on the order of tens to hundreds of milliseconds. By setting a certain time offset between the first and second time sequences, or by automatically learning the alignment mapping relationship between the two time sequences through a deep neural network model, the model can correctly correlate the neural command issued at a specific moment with the observed body response after the delay, avoiding false instantaneous deviations between the expected motion vector and the actual motion state vector caused by time misalignment.

[0049] The alignment network layer receives the expected motion vector sequence in the first time series and the actual motion state vector sequence in the second time series, and performs temporal feature alignment and mapping on the two sequences. Optionally, the alignment network layer can employ a dynamic temporal warping layer to obtain the aligned feature sequence by calculating the warped path between the two sequences. Alternatively, the alignment network layer can also employ a temporal convolutional layer or attention alignment layer based on learnable parameters, learning the temporal offset correction parameters between the sequences through training. The alignment network layer outputs an aligned feature map, which contains the correspondence between the two modality data in a unified temporal dimension.

[0050] The multi-head attention mechanism layer receives the feature maps output by the alignment network layer and calculates the spatial and channel domain correlation weights between cross-modal features. The spatial domain correlation weights characterize the degree of association between each feature dimension of the expected motion vector and each feature dimension of the actual motion state vector at different time positions; the channel domain correlation weights characterize the differences in the contribution of different feature dimension combinations to the mismatch calculation. The multi-head attention mechanism executes multiple sets of attention calculations in parallel, with each set focusing on different subspace representations of the input features. The calculation results from each set are concatenated and linearly transformed to form a weighted and fused feature vector.

[0051] A multilayer perceptron receives fused feature vectors and outputs the dynamic mismatch index at the current time. The multilayer perceptron consists of multiple stacked fully connected layers, with nonlinear activation functions that can be set between each layer. The output layer of the multilayer perceptron can be configured as a single node, outputting a scalar value as the dynamic mismatch index; or it can be configured as multiple nodes, each outputting different dimensions of the mismatch, which are then weighted or summed to obtain the comprehensive mismatch index.

[0052] In practice, training data for the deep neural network model can be collected through controlled experiments. Specifically, EEG signals and motion data of subjects are recorded simultaneously during normal walking and simulated imbalance states, and the fall risk level or mismatch ground truth value is labeled at each moment. The model optimizes its parameters with the goal of minimizing the loss function between the predicted mismatch index and the labeled ground truth value. During training, the parameters of the alignment network layer, the multi-head attention mechanism layer, and the multilayer perceptron can be jointly updated.

[0053] In one embodiment provided in this application, based on the above-described solution, optionally, the multimodal fusion analysis module is used for: When deep neural network models calculate the dynamic mismatch index, they use the user's historical baseline feature parameters and environmental context feature parameters as auxiliary inputs to correct the dynamic mismatch index.

[0054] Historical baseline feature parameters refer to the statistical measures of the matching relationship between the expected motion vector and the actual motion state vector accumulated by the user in daily activities such as walking and standing, such as the mean and variance of the mismatch degree under different motion modes. Environmental context feature parameters refer to information reflecting the complexity of the current environment obtained through additional sensors or external data interfaces, such as the estimated value of the ground friction coefficient, lighting conditions, and the presence of obstacles. The model concatenates or weights the auxiliary input with the fused feature vector, and then outputs a corrected dynamic mismatch degree index by the multilayer perceptron to improve the specificity of risk assessment.

[0055] In one embodiment provided in this application, based on the above-described solution, optionally, the early warning execution module, such as Figure 3 As shown, it includes: The first early warning execution submodule 301 is used to trigger an audio-visual prompt action for the user when the dynamic mismatch index exceeds a preset first risk threshold but does not exceed a second risk threshold. The second early warning execution submodule 302 is used to send an early warning notification to the associated terminal when the dynamic mismatch index exceeds a preset second risk threshold but does not exceed a third risk threshold. The third early warning execution submodule 303 is used to respond to the dynamic mismatch index exceeding the preset third risk threshold and to link the auxiliary equipment to perform physical intervention actions.

[0056] The first risk threshold corresponds to a mild intention-execution mismatch, such as a slight stumble while walking. In this case, the user is prompted to adjust their gait through audio-visual cues. The second risk threshold corresponds to a moderate mismatch, such as a sudden increase in trunk tilt angle and a significant deviation between the direction of movement and the intended direction. In this case, a notification is sent to nursing staff or the monitoring center terminal for human intervention. The third risk threshold corresponds to a severe mismatch, i.e., the stage where a fall is about to occur or is already occurring. In this case, assistive devices are activated to perform physical intervention actions. Assistive devices include, but are not limited to, airbag belts, ground cushioning devices, or mechanical support arms. The third risk threshold is greater than the second risk threshold, and the second risk threshold is greater than the first risk threshold.

[0057] In one embodiment provided in this application, based on the above-described scheme, optionally, the motion state data includes at least one of three-dimensional acceleration data, three-dimensional angular velocity data, plantar pressure distribution data, and electromyographic signal data.

[0058] and Figure 1 Corresponding to the system, this application also provides a fall risk warning method, the flowchart of which is shown below. Figure 4 As shown, it includes: S401: Acquire the user's EEG signals and motion state data.

[0059] S402: Decode the EEG signal to obtain the expected motion vector; the expected motion vector represents the user's current motion intention.

[0060] S403: Determine the user's actual motion state vector based on the motion state data.

[0061] S404: Calculate the dynamic mismatch index between the expected motion vector and the actual motion state vector.

[0062] S405: Generate risk assessment results based on the dynamic mismatch index.

[0063] S406: Execute the warning action corresponding to the risk level indicated by the risk assessment results.

[0064] In one embodiment provided in this application, based on the above-described solution, optionally, decoding the EEG signal to obtain the expected motion vector includes: The EEG signal is used to extract features based on a time-series feature extraction algorithm to obtain neural activity features; The extracted neural activity features are input into a first decoder to obtain the expected motion vector containing at least one of motion direction features, velocity estimates, and intensity estimates.

[0065] In one embodiment provided in this application, based on the above-described scheme, optionally, calculating the dynamic mismatch index between the expected motion vector and the actual motion state vector includes: The expected motion vector sequence in the first time series and the actual motion state vector sequence in the second time series are input into a deep neural network model, so that the alignment network layer in the deep neural network model performs temporal feature alignment and mapping on the expected motion vector sequence and the actual motion state vector sequence; the aligned feature map is input into a multi-head attention mechanism layer to calculate the spatial domain and channel domain correlation weights between cross-modal features; the weighted fused feature vector is input into a multilayer perceptron, and the multilayer perceptron outputs the dynamic mismatch index at the current time.

[0066] This application also provides an electronic device, the structural schematic diagram of which is shown below. Figure 5 As shown, it specifically includes a memory 501 and one or more instructions 502, wherein one or more instructions 502 are stored in the memory 501 and are configured to be executed by one or more processors 403 to perform the above-mentioned fall risk warning method.

[0067] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0068] Finally, it should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0069] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0070] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0071] The solution provided in this application has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A fall risk early warning system, characterized in that, include: The EEG acquisition module is used to acquire the user's EEG signals; The motion status monitoring module is used to collect motion status data of the user's target node; A multimodal fusion analysis module is used to receive the electroencephalogram (EEG) signals and the motion state data; The EEG signal is decoded to obtain the expected motion vector; the expected motion vector represents the user's current motion intention. Based on the motion state data, determine the user's actual motion state vector; calculate the dynamic mismatch index between the expected motion vector and the actual motion state vector; Risk assessment results are generated based on the dynamic mismatch index. The early warning execution module is used to execute the early warning action corresponding to the risk level indicated by the risk assessment result.

2. The system of claim 1, wherein, The multimodal fusion analysis module includes: The feature extraction submodule is used to extract features from the EEG signal based on a temporal feature extraction algorithm to obtain neural activity features; An execution submodule is used to input the extracted neural activity features into a first decoder to obtain the expected motion vector containing at least one of motion direction features, velocity estimates, and intensity estimates.

3. The system of claim 1, wherein, The multimodal fusion analysis module is specifically used for: The motion state data is processed based on the attitude calculation algorithm to obtain the actual motion state vector containing at least one of the following: current torso tilt angle, center of gravity projection trajectory, and gait cycle parameters.

4. The system of claim 1, wherein, The multimodal fusion analysis module is used for: The expected motion vector sequence in the first time series and the actual motion state vector sequence in the second time series are input into a deep neural network model, so that the alignment network layer in the deep neural network model performs time dimension feature alignment and mapping on the expected motion vector sequence and the actual motion state vector sequence; The aligned feature map is input to the multi-head attention mechanism layer to calculate the spatial and channel domain correlation weights between cross-modal features; the weighted fused feature vector is input to the multilayer perceptron, which outputs the dynamic mismatch index at the current time.

5. The system of claim 4, wherein, The multimodal fusion analysis module is used for: When the deep neural network model calculates the dynamic mismatch index, it uses the user's historical baseline feature parameters and environmental context feature parameters as auxiliary inputs to correct the dynamic mismatch index.

6. The system of claim 1, wherein, The early warning execution module includes: The first early warning execution submodule is used to trigger an audio-visual prompt action for the user when the dynamic mismatch index exceeds a preset first risk threshold but does not exceed a second risk threshold. The second early warning execution submodule is used to send an early warning notification to the associated terminal when the dynamic mismatch index exceeds a preset second risk threshold but does not exceed a third risk threshold. The third early warning execution submodule is used to respond to the dynamic mismatch index exceeding the preset third risk threshold and to link auxiliary equipment to perform physical intervention actions.

7. The system of claim 1, wherein, The motion state data includes at least one of three-dimensional acceleration data, three-dimensional angular velocity data, plantar pressure distribution data, and electromyographic signal data.

8. A fall risk early warning method, characterized by, include: Acquire the user's brainwave signals and motion state data; The EEG signal is decoded to obtain the expected motion vector; The expected motion vector represents the user's current motion intention; The user's actual motion state vector is determined based on the motion state data; Calculate the dynamic mismatch index between the expected motion vector and the actual motion state vector; Risk assessment results are generated based on the dynamic mismatch index. Execute the warning action corresponding to the risk level indicated by the risk assessment results.

9. The method of claim 8, wherein, Decoding the EEG signal to obtain the expected motion vector includes: The EEG signal is used to extract features based on a time-series feature extraction algorithm to obtain neural activity features; The extracted neural activity features are input into a first decoder to obtain the expected motion vector containing at least one of motion direction features, velocity estimates, and intensity estimates.

10. The method of claim 8, wherein, The calculation of the dynamic mismatch index between the expected motion vector and the actual motion state vector includes: The expected motion vector sequence in the first time series and the actual motion state vector sequence in the second time series are input into a deep neural network model, so that the alignment network layer in the deep neural network model performs temporal feature alignment and mapping on the expected motion vector sequence and the actual motion state vector sequence; the aligned feature map is input into a multi-head attention mechanism layer to calculate the spatial domain and channel domain correlation weights between cross-modal features; the weighted fused feature vector is input into a multilayer perceptron, and the multilayer perceptron outputs the dynamic mismatch index at the current time.

Images