A method for predicting fall risk of the elderly based on multi-signal features

Abstract

The application belongs to the technical field of fall risk prediction, and relates to a fall risk prediction method for the elderly based on multiple signal features, comprising the following steps: collecting surface electromyography signals, inertial signals and plantar pressure signals of a human body in real time, and performing time-frequency analysis on the surface electromyography signals to obtain a time-frequency spectrum of the surface electromyography signals; extracting feature vectors of corresponding signals in the same time period according to the signals and the time-frequency spectrum; obtaining a feature matrix by weighted fusion of the feature vectors; and inputting the feature matrix into a long short-term memory network model for processing, and outputting a fall risk probability.The scheme of the application can accurately output a fall risk probability by fusing multiple signal features, and can realize early prediction and real-time alarm, thereby preventing the elderly from falling.

Description

Technical Field

[0001] This invention belongs to the field of fall risk prediction technology, and relates to a method for predicting fall risk in the elderly based on multiple signal features. Background Technology

[0002] With the increasing aging of the global population, falls among the elderly are becoming a serious public health challenge. Falls not only cause direct physical injury to the elderly but also lead to long-term consequences such as fear and limited mobility. Therefore, developing a scientific and accurate fall risk prediction system for the elderly has become an urgent issue to be addressed.

[0003] In existing technologies, fall risk prediction schemes for the elderly typically involve attaching inertial detection elements, such as six-axis sensors, to the joints of the body to acquire inertial signals. Fall events are then identified by detecting acceleration or angular velocity within these signals. However, this fall risk prediction scheme, which relies solely on a single inertial signal, cannot comprehensively reflect the various physiological changes that occur when a fall happens, such as changes in electromyography (EMG) signals and plantar pressure signals. Therefore, the prediction results are inaccurate. Furthermore, this method primarily detects falls during the actual event and cannot provide advance prediction or real-time warnings. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting fall risk in the elderly based on multiple signal features, so as to solve the technical problems of existing solutions being inaccurate in predicting fall risk in the elderly and unable to achieve early prediction and real-time warning.

[0005] To achieve the above objectives, the present invention provides a specific technical solution for a fall risk prediction method for the elderly based on multiple signal features, as follows: A method for predicting fall risk in older adults based on multiple signal features includes the following steps: Real-time acquisition of surface electromyography (EMG), inertial signals, and plantar pressure signals of the human body, and time-frequency analysis of the surface EMG signals to obtain the time spectrum of the surface EMG signals; Based on the surface electromyography (SEMG) signal, inertial signal, plantar pressure signal, and the time spectrum of the SEMG signal, feature vectors within the same time period are extracted; the feature vectors are then weighted and fused to obtain a feature matrix; wherein, the feature vectors include the mean value of acceleration and angular velocity, the root mean square value of acceleration, and the variance of angular velocity extracted from the inertial signal; the plantar pressure distribution characteristics and plantar center of gravity trajectory change characteristics extracted from the plantar pressure signal; the root mean square value, mean absolute value, zero crossover rate, and waveform length extracted from the SEMG signal; and the instantaneous frequency entropy and multi-joint coordination entropy extracted from the time spectrum of the SEMG signal. The feature matrix is ​​input into a long short-term memory network model for processing, and the probability of falling is output.

[0006] Furthermore, the first Weight of each signal for: ; In the formula, For the first The signal-to-noise ratio of each signal. For the first The cross-correlation coefficient of a signal is used to characterize the consistency of the trend of change of the signal with other signals.

[0007] Furthermore, the cross-correlation coefficient is the mean of the Pearson correlation coefficients between the corresponding signal and the other signals.

[0008] Furthermore, the time spectrum of the surface electromyography signal is obtained by short-time Fourier transform.

[0009] Furthermore, the surface electromyography (EMG) signal is acquired by an EMG sensor, which is attached to the superficial muscle surface area of ​​the lower limb; the inertial signal is acquired by a six-axis sensor, which is attached to the swing segment of the upper limb, the mid-thigh and mid-calf of the lower limb; and the plantar pressure signal is acquired by a plantar pressure array sensor, which is placed under the insole.

[0010] Furthermore, the acquired surface electromyography signals are subjected to bandpass filtering and baseline correction; the acquired inertial signals are subjected to low-pass filtering and smoothing; and the acquired plantar pressure signals are subjected to normalization.

[0011] Furthermore, when 0 ≤ probability of falling < 0.5, it is a normal state and no alarm is triggered; when 0.5 ≤ probability of falling < 0.7, it is a mild risk and a risk warning is issued, but no alarm is triggered; when 0.7 ≤ probability of falling < 0.8, it is a moderate risk and an enhanced warning is issued, but no alarm is triggered; when 0.8 ≤ probability of falling ≤ 1, it is a high risk, an audible and visual alarm is output, and data is automatically recorded.

[0012] Furthermore, the Long Short-Term Memory (LSTM) network model consists of an input layer, two memory unit layers, and a fully connected output layer; the input layer has a feature dimension of 60; each layer has 128 hidden units, and the Dropout ratio is 0.3.

[0013] Furthermore, the fully connected output layer employs the Sigmoid activation function.

[0014] Furthermore, the training phase of the Long Short-Term Memory network model uses binary cross-entropy as the loss function, the optimization algorithm is the Adam optimization algorithm, and 5-fold cross-validation is used.

[0015] The fall risk prediction method for the elderly based on multi-signal features of the present invention has the following advantages: By collecting surface electromyography (EMG), inertial, and plantar pressure signals during the movement of elderly individuals, and after signal synchronization and time-frequency feature extraction, feature vectors for each signal are obtained. The corresponding feature vectors are then weighted and fused using the weights of each signal to generate a feature matrix. This feature matrix is ​​input into a long short-term memory (LSTM) network model to calculate the probability of fall risk and output an early warning, thereby enabling early prediction of fall risk in the elderly. This invention, by fusing EMG, inertial, and plantar pressure signals, comprehensively reflects various physiological changes in the human body during a fall, making the output fall risk probability more accurate. Furthermore, by introducing EMG and plantar pressure signals, prediction and real-time alarms can be provided before a fall occurs, thus preventing falls in the elderly. Attached Figure Description

[0016] Figure 1 This is a flowchart of the fall risk prediction method for the elderly based on multiple signal features according to the present invention. Detailed Implementation

[0017] The technical solutions of this application will now be described clearly and in detail with reference to the accompanying drawings. In the description of the embodiments of this application, unless otherwise stated, "multiple" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0018] like Figure 1 As shown, this invention provides a method for predicting fall risk in the elderly based on multiple signal features, comprising the following steps: Step S1: Real-time acquisition of surface electromyography (EMG) signals, inertial signals, and plantar pressure signals of the human body, and time-frequency analysis of the surface EMG signals to obtain the time spectrum of the surface EMG signals; Step S2: Extract feature vectors within the same time period based on the surface electromyography (SEMG) signal, inertial signal, plantar pressure signal, and the time spectrum of the SEMG signal; use the feature vectors to perform weighted fusion to obtain a feature matrix; wherein, the feature vectors include the mean value of acceleration and angular velocity, the root mean square value of acceleration and the variance of angular velocity extracted from the inertial signal, the plantar pressure distribution features and plantar center of gravity trajectory change features extracted from the plantar pressure signal, the root mean square value, mean absolute value, zero crossover rate and waveform length extracted from the SEMG signal, and the instantaneous frequency entropy and multi-joint coordination entropy extracted from the time spectrum of the SEMG signal; Step S3: Input the feature matrix into the long short-term memory network model for processing, and output the probability of falling.

[0019] The purpose of this invention in collecting surface electromyography (EMG) signals from the human body is that, due to decreased nerve conduction velocity, slowed muscle response, and diminished proprioception, elderly individuals often exhibit abnormal EMG activity and micro-vibration characteristics 1 to 3 seconds before a fall. These changes manifest as short-term spectral drift, nonlinear entropy fluctuations, and other physiological signal abnormalities, which are important precursors to fall risk. Therefore, this invention extracts the feature vector of the surface EMG signal, as well as the corresponding instantaneous frequency entropy and multi-joint coordination entropy extracted from the time-spectrum.

[0020] Among them, the root mean square value extracted from surface electromyography (SEMG) signals within the same time period is used to reflect the overall activation intensity of the muscle and is an important indicator for evaluating the level of muscle contraction; the average absolute value extracted from SEMG signals is obtained by taking the absolute value of the amplitude of the SEMG signals and averaging it, and is used to describe the average level of muscle activity, which can reflect the changes in the degree of muscle exertion; the zero-crossing rate extracted from SEMG signals is obtained by counting the number of times the SEMG signals cross zero points, and is used to reflect the changes in the frequency components of the signal, which is closely related to the neuromuscular control state; the waveform length extracted from SEMG signals is obtained by accumulating the amplitude changes of adjacent sampling points of the SEMG signals, and is used to measure the signal complexity and the degree of instability of muscle activity.

[0021] Instantaneous frequency entropy originates from the spectral energy distribution obtained after performing a short-time Fourier transform (STFT) on the target motion signal. Within each sliding time window, the energy of the signal at different frequency components exhibits certain distribution characteristics. This energy distribution reflects the frequency domain structure of the signal at the current moment. To characterize the complexity of this frequency domain structure, the spectral energy distribution is transformed into a probability distribution, and its information entropy is calculated, defined as instantaneous frequency entropy. Instantaneous frequency entropy is calculated based on the time-frequency spectrum of surface electromyography (EMG) signals. It is used to quantify the complexity and disorder of EMG signals in the time-frequency dimension, and can sensitively reflect pre-fall warning signs such as muscle tremors and neurological control disorders. Instantaneous frequency entropy is used to characterize the dispersion of the signal's frequency domain distribution within the current time period. The higher the entropy value, the more dispersed the spectral distribution and the more complex the signal structure.

[0022] The calculation process for instantaneous frequency entropy is as follows: (1) Normalize the energy of all frequencies in the time spectrum of the surface electromyography signal and construct a probability distribution: ; In the formula, For the first Energy of each frequency component For the first The probability distribution of each frequency component.

[0023] (2) Calculate the information entropy based on the probability distribution: ; In the formula, For the first The probability distribution of each frequency component It represents the instantaneous frequency entropy.

[0024] Multi-joint coordination entropy originates from the joint state distribution of motion features of multiple joints. Within the same time window, motion parameters (such as angle, angular velocity, or acceleration) of multiple joints are extracted to form a multi-dimensional feature vector, which reflects the coordinated motion state of multiple joints. To quantify the distribution of different coordinated states, state statistics are performed on the multi-dimensional feature space to construct a probability distribution, and information entropy is calculated, defined as multi-joint coordination entropy. Multi-joint coordination entropy is calculated by combining the electromyographic time-frequency features of related muscle groups of multiple joints, and is used to describe the level of coordinated control of different joint muscle groups during movement. Increased coordination entropy indicates a decline in joint coordination ability and is an important high-order discriminant feature for fall risk in the elderly.

[0025] The calculation process for the multi-joint coordination entropy is as follows: (1) Constructing multi-joint feature vectors: ; In the formula, For the first The motion parameters of each joint are defined as angle, angular velocity, or acceleration.

[0026] (2) Perform state statistics on the multi-joint feature vectors to obtain the frequency of occurrence of each state; (3) Constructing the state probability distribution: ; In the formula, For the first The frequency of occurrence of this state For the first The probability distribution of a state.

[0027] (4) Calculate the multi-joint coordination entropy: ; In the formula, For the first The probability distribution of the states. This refers to the multi-joint coordination entropy.

[0028] The purpose of this invention in collecting plantar pressure signals is to detect changes in both the plantar pressure distribution and the trajectory of the foot's center of gravity before a fall. Extracting these feature vectors allows for early prediction of fall risk. Specifically, the plantar pressure distribution feature is based on pressure values ​​collected from multiple pressure sensing areas on the sole of the foot. The pressure percentage and distribution in each area are calculated to reflect the uniformity of force on the sole, thus determining the stability of the support surface. The trajectory of the foot's center of gravity is determined by calculating the position and trajectory of the plantar pressure center based on the changes in plantar pressure distribution over time. This characterizes the degree of shift in the body's center of gravity during movement. Increased fluctuations in the center of gravity trajectory typically indicate a decline in balance control.

[0029] The purpose of this invention in collecting human inertial signals is that when a person falls, their inertial signals, such as acceleration and angular velocity, undergo abnormal changes. Therefore, analyzing these signals can be used to predict fall risk. Specifically, the mean values ​​of acceleration and angular velocity within the same time period can be calculated from the inertial signals, reflecting their average levels. The root mean square value of acceleration is calculated from the acceleration signal, reflecting the overall energy level of acceleration amplitude during lower limb movement. A larger root mean square value indicates more pronounced gait vibration and lower movement stability. The variance of angular velocity is calculated from the angular velocity signal, characterizing the degree of fluctuation during joint rotation. An increased variance in angular velocity usually indicates unstable joint control or abnormal swaying, and is an important indicator of fall risk.

[0030] The aforementioned feature vectors represent the movement status of the elderly from the perspectives of neuromuscular control, mechanics, and kinematics, respectively. The feature extraction process is based on signal statistical analysis and time-frequency analysis methods within the same time period. The resulting features together constitute a multidimensional feature vector for fall risk assessment.

[0031] The feature matrix is ​​a two-dimensional data structure composed of multiple weighted feature vectors, which is used to comprehensively represent the motion state information of multiple signals in the same time period.

[0032] The process of the fall risk prediction method for the elderly based on multi-signal features in this invention is as follows: First, surface electromyography (EMG), inertial signals, and plantar pressure signals of the human body are collected in real time, and time-frequency analysis is performed on the EMG signals to obtain the corresponding time spectrum. Second, within the same time period, the mean value of acceleration and angular velocity, the root mean square value of acceleration, and the variance of angular velocity are extracted from the inertial signals. The plantar pressure distribution features and the plantar center of gravity trajectory change features are extracted from the plantar pressure signals. The root mean square value, mean absolute value, zero crossover rate, and waveform length are extracted from the EMG signals. The instantaneous frequency entropy and multi-joint coordination entropy are extracted from the time spectrum of the EMG signals. Then, the feature vectors extracted above are weighted and fused using the weights of their corresponding signals to obtain a feature matrix. Finally, the feature matrix is ​​input into a long short-term memory network model for processing to output the fall risk probability.

[0033] This invention presents a multi-signal feature-based method for predicting fall risk in the elderly. It collects surface electromyography (EMG), inertial, and plantar pressure signals during the elderly's movement. After signal synchronization and time-frequency feature extraction, feature vectors for each signal are obtained. The corresponding feature vectors are then weighted and fused using the weights of each signal to generate a feature matrix. This feature matrix is ​​input into a long short-term memory (LSTM) network model to calculate the probability of fall risk and output an early warning, thus enabling early prediction of fall risk in the elderly. This invention, by fusing EMG, inertial, and plantar pressure signals, comprehensively reflects various physiological changes in the body during a fall, resulting in a more accurate fall risk probability. Furthermore, by introducing EMG and plantar pressure signals, prediction and real-time warnings can be provided before a fall occurs, thus preventing falls in the elderly. Experiments have shown that this invention can identify risks approximately 0.8 seconds before a fall occurs, with an accuracy rate of 97%, significantly superior to existing detection methods.

[0034] In some embodiments of the present invention, the first Weight of each signal for: ; In the formula, For the first The signal-to-noise ratio of each signal. For the first The cross-correlation coefficient of a signal is used to characterize the consistency of the trend of change of the signal with other signals.

[0035] In this embodiment, the signal-to-noise ratio (SNR) is a core indicator for measuring signal quality, and it is characterized by the ratio of the useful signal strength to the background noise strength. A higher SNR indicates a stronger useful signal, and a higher value represents a cleaner signal, better system performance, and less interference from background noise.

[0036] The cross-correlation coefficient measures the consistency of the changing trends of a Type I signal with other signals, serving as an indicator of the coordinated changes among multiple signals. This is because during human movement, surface electromyography (EMG), inertial signals, and plantar pressure signals exhibit interconnected responses to changes in posture stability. When the change in a particular signal shows high consistency with other signals, it indicates that the signal's representation of the current motion state is more reliable. Therefore, the cross-correlation coefficient describes the degree of coordination among signals in responding to changes in the overall motion state within the same time period; a higher correlation indicates a stronger ability of the signal to reflect changes in posture stability.

[0037] In this embodiment, the weight of the corresponding signal is obtained by the combined effect of signal-to-noise ratio (SNR) and cross-correlation coefficient, which can truly reflect the importance of the signal. That is, the higher the SNR and cross-correlation coefficient, the greater the weight of the corresponding signal, and the higher the accuracy of the final data; conversely, the lower the SNR and cross-correlation coefficient, the smaller the weight of the corresponding signal, and the smaller its impact on the data. When calculating the feature matrix, each feature vector is weighted using the weight of its corresponding signal.

[0038] In some embodiments of the present invention, the cross-correlation coefficient is the mean of the Pearson correlation coefficients between the corresponding signal and the other signals.

[0039] In this embodiment, the Pearson correlation coefficient is an indicator that measures the strength and direction of the linear relationship between two continuous variables. It is a unitless value between -1 and 1. A Pearson correlation coefficient of 1 indicates a perfect positive linear correlation, a Pearson correlation coefficient of -1 indicates a perfect negative linear correlation, and a Pearson correlation coefficient of 0 indicates no linear correlation. The cross-correlation coefficient of the signal can be obtained by calculating the mean of the Pearson correlation coefficients of the corresponding signal with the other signals. Specifically, the first... Cross-correlation coefficients of signals for:

[0040] ; In the formula, For the first A sequence of signals For the first A sequence of signals For the first The mean of a signal sequence, For the first The mean of a signal sequence This represents the total number of signal sequences.

[0041] In some embodiments of the present invention, the time spectrum of the surface electromyography signal is obtained by short-time Fourier transform.

[0042] In this embodiment, the time-frequency analysis step uses Short-Time Fourier Transform (STFT) to perform frame-by-frame processing on the surface electromyography (EMG) signal, obtaining the local spectral energy distribution of the signal over time for subsequent feature extraction. STFT can perform spectral analysis on the signal within a fixed time window, making it suitable for processing continuous time-varying signals during motion processes as described in this application. It can stably reflect the frequency structure changes of EMG and inertial signals at different time intervals.

[0043] Specifically, the expression is as follows: ; In the formula, For window functions, The signal represents the surface electromyography (EMG) signal, τ is the time shift, and ω is the angular frequency. This embodiment employs a Hamming window to reduce spectral leakage.

[0044] In some embodiments of the present invention, surface electromyography (EMG) signals are acquired by a surface EMG sensor, which is attached to the superficial muscle surface area of ​​the lower limb to acquire contraction activity and tremor signals of the lower limb muscle groups involved in gait stability and posture control.

[0045] In some embodiments of the present invention, inertial signals are acquired by a six-axis sensor, which is attached to the body surface area of ​​the upper limb swing segment, the middle section of the thigh and the middle section of the calf, to acquire acceleration and angular velocity signals of the corresponding segments to reflect posture and gait stability.

[0046] In some embodiments of the present invention, plantar pressure signals are collected by a plantar pressure array sensor, which is placed under the insole to collect pressure distribution in multiple areas of the sole, and is used to characterize changes in center of gravity shift and support stability.

[0047] In some embodiments of the present invention, the acquired surface electromyography signals are subjected to bandpass filtering and baseline correction.

[0048] In this embodiment, bandpass filtering is a frequency domain operation that selectively extracts the target frequency band signal by setting upper and lower cutoff frequencies, while attenuating out-of-band noise. Its core function is to separate specific physiological rhythms, suppress high-frequency interference, and remove low-frequency drift. Implementation methods include IIR / FIR digital filters, which are fundamental steps in signal preprocessing and can fundamentally change the spectral composition of the signal.

[0049] Baseline correction is an amplitude-domain operation that aligns the DC component of a signal to zero by subtracting the mean value of a reference time period. This operation preserves the relative shape and spectral characteristics of the signal, performing only a vertical shift, and is a crucial data standardization step before time-domain analysis.

[0050] This embodiment prioritizes bandpass filtering to purify the signal spectrum, followed by baseline correction to establish a unified measurement benchmark. These two processes form a progressive relationship of "spectrum purification → time-domain standardization," jointly ensuring the effectiveness of subsequent feature extraction and statistical analysis.

[0051] In some embodiments of the present invention, the acquired inertial signal is subjected to low-pass filtering and smoothing.

[0052] In this embodiment, low-pass filtering is a frequency domain operation that allows low-frequency components in the signal below the cutoff frequency to pass through, while attenuating or eliminating high-frequency components above that frequency. Smoothing is an operation that reduces data fluctuations, highlights main trends, and makes curves or surfaces smoother. Through low-pass filtering and smoothing, the impact of high-frequency jitter and sensor random noise on motion feature extraction can be reduced.

[0053] In some embodiments of the present invention, the acquired plantar pressure signal is normalized.

[0054] In this embodiment, the pressure values ​​of different regions are normalized to reduce the impact of individual weight differences and make the data comparable among different subjects.

[0055] In some embodiments of the present invention, when 0 ≤ fall risk probability < 0.5, it is a normal state and no alarm is triggered; when 0.5 ≤ fall risk probability < 0.7, it is a mild risk, and an interface prompt or vibration is output, but no alarm is triggered; when 0.7 ≤ fall risk probability < 0.8, it is a moderate risk, and enhanced prompts, such as voice prompts and flashing, are output, but no alarm is triggered; when 0.8 ≤ fall risk probability ≤ 1, it is a high risk, and an audible and visual alarm is output and data is automatically recorded for subsequent behavior analysis and model adaptive optimization. This solution can also be wirelessly linked with mobile terminals or rehabilitation monitoring platforms to achieve real-time information push and remote risk monitoring, improving the safety of elderly people during exercise.

[0056] This embodiment establishes a tiered early warning system, classifying early warning information into multiple levels from low to high based on the severity, urgency, and development trend of the risk, and initiating corresponding emergency response actions accordingly. This improves the standardization, scientific nature, and timeliness of emergency decision-making, avoiding "one-size-fits-all" or "overreaction."

[0057] In some embodiments of the present invention, the Long Short-Term Memory (LSTM) network model consists of an input layer, two memory unit layers, and a fully connected output layer; the input layer has a feature dimension of 60; each layer has 128 hidden units, and the Dropout ratio is 0.3.

[0058] In this embodiment, the parameter settings of the Long Short-Term Memory (LSTM) network model are determined based on the composition of multimodal features and the requirements of time series modeling. The input layer feature dimension is 60, which is formed by concatenating statistical features and time-frequency features extracted from multimodal signals within each time window. This dimension can simultaneously include electromyographic signal features, inertial signal features, and plantar pressure features, thereby comprehensively representing human posture stability and neuromuscular control state in a single time step.

[0059] The number of hidden units per layer is set to 128 to ensure that the network has sufficient temporal memory to capture the gradual change trend of motion state within a continuous time window. Falls in the elderly do not happen instantaneously, but rather go through a continuous process from stability to imbalance. 128 hidden units can ensure the model's expressive power while avoiding the instability caused by excessively large parameter size.

[0060] The dropout ratio is set to 0.3 to randomly mask some neuron connections during the training phase, thereby reducing the risk of overfitting the model to specific training samples and improving its generalization ability to different motion forms and individuals.

[0061] By configuring the parameters described above, the network can stably learn the changing patterns of multimodal features over time and output a smooth and continuously changing probability of fall risk, thereby improving the stability and robustness of the prediction.

[0062] In some embodiments of the present invention, the fully connected output layer employs the Sigmoid activation function.

[0063] In this embodiment, the Sigmoid activation function has a good output range and can interpret probabilities. It "squeezes" any real-valued input into a smooth and continuous output between (0, 1). This perfectly bounded range makes it very suitable for the output layer of binary classification problems, where the output value can intuitively interpret the probability of a sample belonging to the positive class or an anomaly. Moreover, the Sigmoid activation function is smooth and continuous throughout.

[0064] In some embodiments of the present invention, the training phase of the long short-term memory network model uses binary cross-entropy as the loss function, the optimization algorithm is the Adam optimization algorithm, and 5-fold cross-validation is used.

[0065] In this embodiment, binary cross-entropy, the Adam optimization algorithm, and the cross-validation strategy are selected, representing an adaptation to the data type and task objective of this invention. Specifically, the prediction objective of this application is the probability of falling, which is essentially a binary classification probability output problem. Therefore, binary cross-entropy is used as the loss function to measure the difference between the predicted risk probability and the actual risk label.

[0066] The Adam optimization algorithm is used to improve the convergence stability during multimodal temporal feature training, enabling the network to reach a stable state more quickly when processing high-dimensional temporal inputs composed of electromyography, inertia, and pressure features. Five-fold cross-validation is used to evaluate the model's generalization ability under limited sample size conditions, reduce the impact of individual differences on model training results, and improve the stability of risk prediction results under different motion forms.

[0067] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for predicting fall risk in the elderly based on multi-signal features, characterized in that, Includes the following steps: Real-time acquisition of surface electromyography (EMG), inertial signals, and plantar pressure signals of the human body, and time-frequency analysis of the surface EMG signals to obtain the time spectrum of the surface EMG signals; Based on the time spectrum of surface electromyography (EMG) signals, inertial signals, plantar pressure signals, and surface EMG signals, feature vectors within the same time period are extracted. A feature matrix is ​​obtained by weighted fusion using feature vectors; wherein the feature vectors include the mean of acceleration and angular velocity, the root mean square value of acceleration and the variance of angular velocity extracted from the inertial signal, the plantar pressure distribution features and plantar center of gravity trajectory change features extracted from the plantar pressure signal, the root mean square value, mean absolute value, zero crossover rate and waveform length extracted from the surface electromyography signal, and the instantaneous frequency entropy and multi-joint coordination entropy extracted from the time spectrum of the surface electromyography signal; The feature matrix is ​​input into a long short-term memory network model for processing, and the probability of falling is output.

2. The method for predicting fall risk in the elderly based on multiple signal features according to claim 1, characterized in that, No. Weight of each signal for: ； In the formula, For the first The signal-to-noise ratio of each signal. For the first The cross-correlation coefficient of a signal is used to characterize the consistency of the trend of change of the signal with other signals.

3. The method for predicting fall risk in the elderly based on multiple signal features according to claim 2, characterized in that, The cross-correlation coefficient is the mean of the Pearson correlation coefficients between the corresponding signal and the other signals.

4. The method for predicting fall risk in the elderly based on multiple signal features according to claim 1, characterized in that, The time spectrum of the surface electromyography signal is obtained by short-time Fourier transform.

5. The method for predicting fall risk in the elderly based on multiple signal features according to claim 1, characterized in that, The surface electromyography (EMG) signal is acquired by an EMG sensor, which is attached to the superficial muscle surface area of ​​the lower limb; the inertial signal is acquired by a six-axis sensor, which is attached to the swing segment of the upper limb, the mid-thigh and mid-calf of the lower limb; the plantar pressure signal is acquired by a plantar pressure array sensor, which is placed under the insole.

6. The method for predicting fall risk in the elderly based on multiple signal features according to claim 1, characterized in that, The acquired surface electromyography signals are subjected to bandpass filtering and baseline correction; the acquired inertial signals are subjected to low-pass filtering and smoothing; and the acquired plantar pressure signals are subjected to normalization.

7. The method for predicting fall risk in the elderly based on multiple signal features according to claim 1, characterized in that, When the probability of falling is less than 0.5, it is in normal condition and no alarm is triggered; when the probability of falling is less than 0.7, it is in mild risk and a risk warning is issued, but no alarm is triggered; when the probability of falling is less than 0.8, it is in moderate risk and an enhanced warning is issued, but no alarm is triggered; when the probability of falling is less than 1, it is in high risk, an audible and visual alarm is output, and data is automatically recorded.

8. The method for predicting fall risk in the elderly based on multiple signal features according to claim 1, characterized in that, The Long Short-Term Memory (LSTM) network model consists of an input layer, two memory unit layers, and a fully connected output layer; the input layer has a feature dimension of 60; each layer has 128 hidden units, and the Dropout ratio is 0.

3.

9. The method for predicting fall risk in the elderly based on multiple signal features according to claim 8, characterized in that, The fully connected output layer uses the Sigmoid activation function.

10. The method for predicting fall risk in the elderly based on multiple signal features according to claim 1, characterized in that, The training phase of the Long Short-Term Memory (LSTM) network model uses binary cross-entropy as the loss function, the Adam optimization algorithm as the optimization algorithm, and 5-fold cross-validation.

Images