Method and device for anti-falling early warning and storage medium

Abstract

The invention discloses an anti-falling early warning method and device and a storage medium. The method comprises the following steps: at least collecting surface myoelectricity detection signals of lower limbs of a detected person at a plurality of target muscles in a natural walking process and pressure detection signals of soles of the detected person at a plurality of target detection points; respectively performing frequency domain analysis and time domain analysis on the basis of the surface myoelectricity detection signal so as to correspondingly obtain frequency domain characteristics and time domain characteristics; extracting plantar pressure features based on the pressure detection signal; performing feature processing operation on the frequency domain feature, the time domain feature and the plantar pressure feature to obtain a target feature result; and inputting the target feature result into an anti-falling early warning model to carry out anti-falling prediction so as to obtain an anti-falling early warning result. By means of the scheme, the accuracy of anti-falling prediction can be improved, anti-falling early warning is effectively achieved, and falling events are avoided.

Description

Technical Field

[0001] The present application generally relates to the field of fall prevention warning technology. More specifically, the present application relates to a method, device and computer-readable storage medium for fall prevention warning. Background Art

[0002] The aging of the population, especially the increasing degree of aging, has brought about disability, resulting in an increase in the number and proportion of elderly people who cannot take care of themselves. Accompanying the aging of the population are a series of problems such as the decline in motor system function caused by physiological degeneration and pathological changes in the neuromuscular system of the elderly. In particular, the demand for real-time warning of falls caused by neuromuscular system lesions in elderly people who are not in hospital is becoming increasingly prominent. In the process of population aging, the increase in falls caused by neuromuscular system degeneration and lesions will further increase the burden of medical resource allocation. For families and even society, the time cost of caring for the elderly will also increase dramatically, which poses a huge challenge to the allocation of medical resources and the efficiency of medical treatment.

[0003] In view of this, there is an urgent need to provide a solution for anti-fall warning, so as to improve the accuracy of anti-fall prediction, effectively implement anti-fall warning, and avoid fall events. Summary of the invention

[0004] In order to at least solve one or more of the technical problems mentioned above, the present application proposes solutions for anti-fall warning in multiple aspects.

[0005] In a first aspect, the present application provides a method for anti-fall warning, comprising: at least collecting surface electromyography detection signals of a subject's lower limbs at multiple target muscles during natural walking and pressure detection signals of the subject's soles at multiple target detection points; performing frequency domain analysis and time domain analysis based on the surface electromyography detection signals to obtain frequency domain features and time domain features respectively; extracting plantar pressure features based on the pressure detection signals; performing feature processing operations on the frequency domain features, the time domain features and the plantar pressure features to obtain target feature results; and inputting the target feature results into an anti-fall warning model to perform anti-fall prediction to obtain an anti-fall warning result.

[0006] In some embodiments, the frequency domain features include at least the median frequency and / or the average power frequency, the time domain features include at least the integrated electromyographic value per unit time; and the plantar pressure features include at least the foot center of gravity offset and / or the foot center of gravity trajectory curve.

[0007] In some embodiments, the frequency-domain features are obtained by the following operations: performing frequency-domain preprocessing on the surface electromyogram detection signal to obtain a frequency-domain signal; calculating a power spectral density function based on the frequency-domain signal; and calculating the median frequency and / or the mean power frequency according to the power spectral density function.

[0008] In some embodiments, the time-domain features are obtained by the following operations: performing time-domain preprocessing on the surface electromyogram detection signal to obtain a preprocessed time-domain signal; calculating the total area under the time-domain signal; and dividing the total area by the total time to obtain the integrated electromyogram value per unit time.

[0009] In some embodiments, the time-domain preprocessing includes one or more of filtering, rectifying, or smoothing.

[0010] In some embodiments, extracting plantar pressure features based on the pressure detection signal includes: processing the pressure detection signal into a set of pressure detection time series points; calculating the pressure cumulative impulse according to the set of pressure detection time series points; calculating the degree of foot center of gravity offset based on the pressure cumulative impulse; and / or calculating the foot center of gravity trajectory curve according to the set of pressure detection time series points.

[0011] In some embodiments, the feature processing operations include feature sampling and / or feature dimensionality reduction.

[0012] In some embodiments, the feature sampling is achieved by the following operations: calculating the nearest neighbor features corresponding to a small number of categories in the frequency-domain features, the time-domain features, and the plantar pressure features; and interpolating between the corresponding features and the nearest neighbor features to generate new features to achieve the feature sampling.

[0013] In some embodiments, the fall prevention warning model includes a support vector machine classifier, a random forest classifier, and a gradient boosting classifier.

[0014] In some embodiments, inputting the target feature result into the fall prevention warning model for fall prevention prediction to obtain a fall prevention warning result includes: inputting the target feature result into the support vector machine classifier, the random forest classifier, and the gradient boosting classifier respectively for fall prevention prediction to obtain their respective fall prevention warning results; and obtaining the final fall prevention warning result according to the weighted sum of their respective fall prevention warning results.

[0015] In some embodiments, it further includes: obtaining the basic information of the subject; and inputting the basic information and the target feature result into the fall prevention warning model for fall prevention prediction to obtain a fall prevention warning result.

[0016] In a second aspect, the present application provides a device for fall prevention warning, including: a processor; and a memory, in which program instructions for fall prevention warning are stored. When the program instructions are executed by the processor, the device implements one or more embodiments in the foregoing first aspect.

[0017] In a third aspect, the present application provides a computer-readable storage medium, on which computer-readable instructions for fall prevention warning are stored. When the computer-readable instructions are executed by one or more processors, one or more embodiments in the foregoing first aspect are implemented.

[0018] Through the solution for fall prevention warning provided as above, embodiments of the present application collect surface electromyography detection signals and pressure detection signals of the subject during natural walking, and correspondingly extract frequency domain features, time domain features, and plantar pressure features, so as to better reflect the overall state change of the subject during natural walking. Among them, the frequency domain features and time domain features can more accurately reflect the muscle fatigue degree and cumulative impulse contribution rate. Combined with stable plantar pressure features, and after feature processing, input into the fall prevention warning model for fall prevention prediction, a more accurate fall prevention warning result can be obtained. Further, embodiments of the present application also perform feature sampling and feature dimensionality reduction processing to improve the performance and stability of the fall prevention warning model and improve the accuracy of the fall prevention warning result. Description of the Drawings

[0019] By referring to the accompanying drawings and reading the following detailed description, the above and other purposes, features, and advantages of the exemplary embodiments of the present application will become easy to understand. In the drawings, several embodiments of the present application are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts, where:

[0020] Figure 1A is an exemplary schematic diagram showing the acquisition of pressure detection signals;

[0021] Figure 1B is an exemplary schematic diagram showing the acquisition of surface electromyography detection signals;

[0022] Figure 2 is an exemplary flowchart showing the method for fall prevention warning according to an embodiment of the present application;

[0023] Figure 3 is an exemplary flowchart showing the overall process for fall prevention warning according to an embodiment of the present application;

[0024] Figure 4 is an exemplary schematic diagram showing the comparison between the fall prevention warning method and the single classifier prediction method according to an embodiment of the present application;

[0025] Figure 5 is an exemplary structural block diagram showing a device for anti-fall warning according to an embodiment of the present application. Detailed implementation manners

[0026] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts fall within the protection scope of the present application.

[0027] It should be understood that the terms "including" and "comprising" used in the specification and claims of the present application indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their combinations.

[0028] It should also be understood that the terms used in the specification of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. As used in the specification and claims of the present application, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include the plural forms. It should be further understood that the term "and / or" used in the specification and claims of the present application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0029] As used in this specification and the claims, the term "if" can be interpreted as "when" or "once" or "in response to determining" or "in response to detecting" depending on the context. Similarly, the phrase "if determined" or "if [the described condition or event] is detected" can be interpreted as meaning "once determined" or "in response to determining" or "once [the described condition or event] is detected" or "in response to detecting [the described condition or event]" depending on the context.

[0030] Figure 1A is an exemplary schematic diagram showing the acquisition of pressure detection signals. As Figure 1A shown in Figure (a), there are multiple target detection points on the sole of the subject during natural walking, including the big toe 101 of the foot, the first metatarsal bone 102, the second and third metatarsal bones 103, the fourth and fifth metatarsal bones 104, the inner side of the arch of the foot 105, the inner side of the heel 106, the outer side of the heel 107, and the back side of the heel 108. As Figure 1AAs shown in Figure (b), the plantar pressure acquisition device is provided with a plurality of pressure sensors 1110 at the aforementioned multiple target detection points. In one implementation scenario, the plurality of pressure sensors 1110 can be communicatively connected to the data processing unit 1111 by wire or wirelessly. For example, in the figure, it shows that the plurality of pressure sensors 1110 and the data processing unit 1111 are connected by wire. In this scenario, by setting a corresponding plurality of pressure sensors at the aforementioned multiple target detection points, each pressure sensor senses the pressure of the foot of the subject to be examined at each target detection point, resulting in a corresponding resistance change, so as to detect the pressure detection signals of the subject's plantar at the multiple target detection points.

[0031] Figure 1B is an exemplary schematic diagram showing the acquisition of surface electromyography detection signals. As Figure 1B As shown in Figure (a), the multiple target muscles of the lower limb include the rectus femoris 109, vastus lateralis 110, vastus medialis 111, biceps femoris 112, semitendinosus 113, peroneus longus 114, tibialis anterior 115, lateral gastrocnemius 116, medial gastrocnemius 117, and soleus 118. As Figure 1B As shown in Figure (b), the surface electromyography acquisition device is provided with a plurality of surface induction electrodes (such as shown by the multiple small rectangular frames in the figure) 1112 at the aforementioned multiple target muscles, and each surface induction electrode 1112 can be correspondingly pasted at each target muscle. In one implementation scenario, the plurality of surface induction electrodes 1112 can be communicatively connected to the data processing unit 1113 by wire or wirelessly. For example, in the figure, it shows that the plurality of surface induction electrodes 1112 and the data processing unit 1113 are connected by wire. In this scenario, by correspondingly setting a plurality of surface induction electrodes at the aforementioned multiple target muscles to sense the surface electrophysiological signals generated by the lower limb neuromuscles of the subject, so as to acquire the surface electromyography detection signals of the subject's both lower limbs at the multiple target muscles.

[0032] The following will describe in detail the specific implementation manners of the present application with reference to the accompanying drawings.

[0033] Figure 2 is an exemplary flowchart showing the method 200 for fall prevention warning according to an embodiment of the present application. As Figure 2 As shown therein, at step S201, at least the surface electromyography detection signals of the lower limb of the subject at multiple target muscles and the pressure detection signals of the subject's plantar at multiple target detection points are acquired during natural walking. In some embodiments, the surface electromyography detection signals and the pressure detection signals can be respectively based on the above Figure 1A and Figure 1BIt is obtained by collecting in the described manner. In particular, the embodiments of the present application collect the surface electromyography detection signals and pressure detection signals of the subject during the entire natural walking process, so that the features extracted subsequently can better reflect the overall state changes of the subject during the natural walking process.

[0034] Next, at step S202, frequency-domain analysis and time-domain analysis are respectively performed on the surface electromyography detection signals to obtain frequency-domain features and time-domain features correspondingly. In some embodiments, the frequency-domain features may at least include median frequency and / or mean power frequency, the time-domain features may at least include the integrated electromyogram value per unit time, and the plantar pressure features may at least include the degree of foot center of gravity offset and / or the foot center of gravity trajectory curve.

[0035] It can be understood that the original surface electromyography detection signals refer to the electrical signal data collected from the muscle surface electrodes without being processed or processed, recorded in the form of a time series, which reflects the changes in muscle electrical activity. These electrical signal data include signals from one or more channels, and each channel corresponds to an electrode placed on the muscle surface, recording the weak electrical signals caused by muscle activity. The original surface electromyography detection signals often contain a lot of noise. Therefore, in some embodiments, before performing time-domain analysis and frequency-domain analysis on the surface electromyography detection signals, initial preprocessing can also be performed on the surface electromyography detection signals. As an example, Matlab can be used, for example, to extract the effective information hidden in the surface electromyography detection signals and remove the unnecessary parts and interferences. Then, time-domain analysis and frequency-domain analysis can be performed on the surface electromyography detection signals after initial preprocessing.

[0036] In some embodiments, frequency-domain preprocessing can be performed on the surface electromyography detection signals to obtain frequency-domain signals, and then the power spectral density function is calculated based on the frequency-domain signals to calculate the median frequency and / or mean power frequency according to the power spectral density function. In some implementation scenarios, the surface electromyography detection signals can be frequency-domain preprocessed by, for example, fast Fourier transform to obtain frequency-domain signals. In some implementation scenarios, the power spectral density function PSD(f) can be calculated based on the frequency-domain signal f by, for example, the periodogram method or the autocorrelation function, and then the median frequency and / or mean power frequency are calculated according to the power spectral density function.

[0037] In some implementation scenarios, the median frequency MF can be calculated based on the following formula:

[0038]

[0039] In some other implementation scenarios, the mean power frequency MPF can be calculated based on the following formula:

[0040]

[0041] It can be understood that the median frequency is the mathematical average of the spectral curve, which can reflect the muscle fatigue state and the change of muscle strength, and has strong signal aliasing ability, and can effectively resist noise interference. The average power frequency is the frequency that divides the spectral area into two equal parts, and is used to evaluate the muscle fatigue state and the characteristics of muscle activity.

[0042] In some embodiments, the surface electromyogram detection signal can be preprocessed in the time domain to obtain the preprocessed time-domain signal, then the total area under the time-domain signal is calculated, and then the total area is divided by the total time to obtain the integrated electromyogram value per unit time. In some embodiments, the aforementioned time-domain preprocessing may include one or more of filtering processing, rectification processing, or smoothing processing.

[0043] In some implementation scenarios, a Butterworth band-pass filter of, for example, the 8th order can be used to filter the surface electromyogram detection signal. Preferably, the lower cut-off frequency can be set to 10 Hz and the upper cut-off frequency to 500 Hz. Since the positive and negative amplitudes of the surface electromyogram detection signal are roughly equal, if directly used for subsequent calculations, it often leads to the average value approaching zero. Therefore, in the embodiments of the present application, the surface electromyogram detection signal can also be rectified so that the amplitudes of the surface electromyogram detection signal are all positive, thereby being able to more accurately reflect the characteristics and changes of muscle activity. In addition, during the generation of the surface electromyogram detection signal, there are many non-renewable signals generated along with the change of muscle action potentials. Therefore, in the embodiments of the present application, for example, root mean square smoothing processing can also be adopted to obtain a more stable and smooth signal curve, which is convenient for subsequent accurate calculation of the integrated electromyogram value per unit time.

[0044] In some implementation scenarios, the root mean square smoothing RMS can be calculated by the following formula:

[0045]

[0046] Among them, N represents the length of the smoothing window, and f(i) represents the amplitude of the surface electromyogram signal at the i-th point in the smoothing window. Based on this, the preprocessed time-domain signal can be obtained.

[0047] Furthermore, the total area under the time-domain signal is calculated, and then the total area is divided by the total time to obtain the integrated electromyogram value per unit time. In some implementation scenarios, the total area iEMG under the time-domain signal can be calculated based on the following formula:

[0048]

[0049] Among them, Z represents the number of time-domain signals, Δt represents the time difference between adjacent signals, and x(i) represents the amplitude of the i-th time-domain signal point. Then, the integral EMG value per unit time is obtained by dividing the total area by the total time, and the contribution rate of each muscle is obtained. That is, the percentage of each muscle in the total integral EMG value per unit time of all muscles.

[0050] Further, at step S203, plantar pressure features are extracted based on the pressure detection signal. In some embodiments, the pressure detection signal can be processed into a set of pressure detection time series points; the pressure cumulative impulse is calculated according to the set of pressure detection time series points, and the foot center of gravity offset degree is calculated based on the pressure cumulative impulse and / or the foot center of gravity trajectory curve is calculated according to the set of pressure detection time series points.

[0051] Specifically, in an implementation scenario, when processing the pressure detection signal into a set of pressure detection time series points, the aforementioned digital voltage signal can be calibrated according to a programmable pressure tester to obtain a calibrated digital voltage signal. Then, the pressure detection signal can be processed into a set of pressure detection time series points, for example, by discrete linear interpolation or high-order polynomial interpolation, to calculate the foot center of gravity offset degree and / or the foot center of gravity trajectory curve based on the set of pressure detection time series points.

[0052] In one embodiment, the pressure cumulative impulse can be calculated according to the set of pressure detection time series points to calculate the foot center of gravity offset degree based on the pressure cumulative impulse; and / or the foot center of gravity trajectory curve is calculated according to the set of pressure detection time series points. More specifically, the aforementioned foot center of gravity offset degree includes the left and right foot center of gravity offset degrees and the front and rear sole center of gravity offset degrees, and the aforementioned pressure cumulative impulse includes the left foot pressure cumulative impulse, the right foot pressure cumulative impulse, the front sole pressure cumulative impulse, and the rear sole pressure cumulative impulse. Among them, the left and right foot center of gravity offset degrees are calculated based on the left foot pressure cumulative impulse and the right foot pressure cumulative impulse, and the front and rear sole center of gravity offset degrees are calculated based on the front sole pressure cumulative impulse and the rear sole pressure cumulative impulse.

[0053] In an implementation scenario, the left and right foot center of gravity offset degrees can be calculated based on the following formula:

[0054]

[0055] Among them, X LR represents the left and right foot center of gravity offset degrees, represents the left foot pressure cumulative impulse, represents the right foot pressure cumulative impulse, represents solving the norm.

[0056] In another implementation scenario, the front and rear sole center of gravity offset degrees can be calculated based on the following formula:

[0057]

[0058] Among them, X FB represents the center of gravity offset degree between the forefoot and the rear foot, represents the cumulative impulse of the forefoot pressure, represents the cumulative impulse of the rear foot pressure, represents the solution norm.

[0059] For calculating the center of gravity trajectory curve, the center of gravity trajectory curve can be calculated based on the following formula:

[0060]

[0061] Among them, C(k) represents the center of gravity trajectory curve, represents the pressure detection time series point set, ε(i) represents the weight value at each target detection point of the plantar pressure, ω(k, j) represents the pressure detection time series weight value, i represents the serial number of each target detection point, k represents the pressure detection time serial number, and j represents the discrete point serial number of the center of gravity trajectory curve.

[0062] After obtaining the above frequency domain features, time domain features, and plantar pressure features, at step S204, feature processing operations are performed on the frequency domain features, time domain features, and plantar pressure features to obtain the target feature results. In some embodiments, the foregoing feature processing operations may include, for example, feature sampling and / or feature dimensionality reduction. For feature sampling, in some embodiments, the nearest neighbor features corresponding to a small number of categories in the frequency domain features, time domain features, and plantar pressure features can be calculated to interpolate new features between the corresponding features (i.e., the frequency domain features, time domain features, and plantar pressure features) and the nearest neighbor features to achieve feature sampling.

[0063] Specifically, for the corresponding feature x of each minority class, first, multiple nearest neighbor features that meet the preset value can be calculated, for example, by Euclidean distance or similarity measure. The nearest neighbor features are those whose Euclidean distance meets the distance threshold or whose correlation meets the correlation threshold. Then, the target neighbor feature x nn can be selected from multiple nearest neighbor features by the sampling ratio, so that the target neighbor feature is close to the minority class samples in the feature space. Further, the following formula is used to interpolate new features x new between the corresponding feature and the nearest neighbor feature:

[0064] x new = x + λ(x nn - x) (8)

[0065] Among them, λ represents a random number between 0 and 1, which can control the random position of the generated new feature between the original corresponding feature and its nearest neighbor feature. Based on this, while increasing the number of samples of the minority class, the diversity of the samples can be further improved. In some implementation scenarios, the sample data can also be scaled proportionally through, for example, standardization operations, avoiding the unreasonable excessive influence of some feature data with too large thresholds on the prediction results, which helps to improve the performance and stability of the algorithm.

[0066] For feature dimensionality reduction, in some embodiments, it can be achieved by means of artificial feature selection, principal component analysis data dimensionality reduction, linear discriminant analysis, etc. Among them, based on artificial feature selection, summary features can be deleted from the dataset to reduce the dimensionality of the data. As an example, based on physiological knowledge, summary features such as "left leg" and "right leg" do not provide additional information for the model because this information has already been included by more specific features such as "left thigh" and "left calf", and thus can be deleted through artificial feature selection.

[0067] When adopting principal component analysis data dimensionality reduction, first, the covariance matrix of the features can be calculated and the eigenvalues and eigenvectors of the covariance matrix can be extracted. Then, the eigenvector with the largest eigenvalue is selected as the principal component. In some embodiments, the number of principal components can be determined according to the cumulative contribution rate of the eigenvalues, that is, the first k eigenvalues are selected so that their cumulative contribution rate reaches the target percentage of the total variation (for example, 90%). Finally, the original feature data is multiplied by the selected eigenvector matrix to realize the conversion of the original feature data into a new feature space composed of principal components, and the dimensionality-reduced features are obtained.

[0068] When performing linear discriminant analysis, for example, the within-class scatter matrix and the between-class scatter matrix can be used. Among them, the within-class scatter matrix S W is a matrix that measures the difference between the feature data within the same class and the mean of that class. For each class c, its within-class scatter matrix is defined as:

[0069]

[0070] where x represents the sample points (i.e., feature data) in class c, and m c represents the sample mean in class c. Further, the global within-class scatter matrix is the sum of the within-class scatter matrices of all classes:

[0071]

[0072] In addition, for the between-class scatter matrix S B which focuses on the difference between the means of data of different classes, the global between-class scatter matrix can be expressed as:

[0073] SB = ∑ c N c (m c - m)(m c - m) T (11)

[0074] where N c represents the number of samples in class c, and m represents the overall mean of all samples. By determining a linear projection ω such that the between-class scatter after projection is maximized and the within-class scatter is minimized, that is, maximizing the following ratio:

[0075]

[0076] In some implementation scenarios, through the Lagrange multiplier method, it can be proven that ω is the eigenvector corresponding to the largest eigenvalue of the matrix S W -1 S B After determining the optimal projection direction, the original feature data of the embodiments of the present application can be projected onto a new low-dimensional space through ω to achieve feature dimensionality reduction. Preferably, the embodiments of the present application adopt linear discriminant analysis for feature dimensionality reduction.

[0077] Based on the obtained target feature results above, at step S205, the target feature results are input into the anti-fall warning model for anti-fall prediction to obtain the anti-fall warning results. In some embodiments, the anti-fall warning model may include a support vector machine classifier, a random forest classifier, and a gradient boosting classifier. Specifically, in some embodiments, by inputting the target feature results into the support vector machine classifier, the random forest classifier, and the gradient boosting classifier respectively for anti-fall prediction to obtain their respective anti-fall warning results, and then obtaining the final anti-fall warning result according to the weighted sum of their respective anti-fall warning results.

[0078] As an example, the anti-fall warning results of each classifier can be fused through, for example, linear weighted L2 norm to obtain the final anti-fall warning result. In one embodiment, the aforementioned linear weighted L2 norm can be expressed as where is the L2 norm, ω i is the linear weight coefficient, i = 1, 2, 3 are the serial numbers of each classifier, where 1 is the support vector machine classifier, 2 is the random forest classifier, 3 is the gradient boosting classifier, F i (·) is the single classifier function, and l(·) is the dimensionality reduction function.

[0079] More specifically, the calculation method of the linear weight coefficient ω i can be solved inversely through the inverse problem model. The objective function is to minimize the missed diagnosis rate, that is, to maximize the recall rate. The inverse problem inverse model solved is where f -1 (·) = max(recall rate), is a row vector of weight coefficients. In addition, compared with the Boolean output of a single model, the output of the fusion model is a continuous function between 0 and 1, which can not only fuse discrete models into a continuous model, but also further reduce the threshold of the positive determination result of the fusion model. For example, when the output value of the fusion model is greater than or equal to 0.75, it is determined that there is a risk of falling.

[0080] In some implementation scenarios, a radial basis function kernel can be used in the support vector machine classifier to make the feature data easier to separate, which is beneficial to improving the classification accuracy. Among them, the γ parameter can be used to control the influence range of the kernel function, and the specific radial basis function kernel can be expressed by the following formula:

[0081] K(x i , x j ) = exp(-γ||x i - x j || 2 )(13)

[0082] where K(x i , x j ) represents the similarity between two feature data x i and x j calculated based on the kernel function, and γ can be set according to the prediction requirements to adjust the influence range of the kernel function.

[0083] In the random forest classifier, the classification performance can be improved by setting the number of trees and the evaluation criteria for splitting quality (such as Gini impurity), and integrating multiple decision trees. In the gradient boosting classifier, the parameters can be adjusted through the objective optimization function for binary classification problems and the number of trees to better fit the data and improve the performance of the model, so as to obtain more accurate classification results.

[0084] In some embodiments, the embodiments of the present application also obtain the basic information of the subject, input the basic information and the target feature result into the fall prevention warning model for fall prevention prediction, so as to output the fall prevention warning result. In some embodiments, the foregoing basic information may include, but is not limited to, one or more of the age information of the subject, historical fall information (such as the number of falls, fall postures, and reasons, etc.) or disease history information. For example, the foregoing basic information may also include the examination information of each joint (such as the ankle joint, hindfoot, forefoot, knee joint, etc.). Based on this, through the joint determination of the basic information and the target feature result, the accuracy of the determination result can be further improved, and the accuracy of the fall prevention prediction result can be ensured.

[0085] As described above, in the embodiments of the present application, by collecting the surface electromyography detection signals and pressure detection signals of the subject during natural walking, and correspondingly extracting frequency-domain features, time-domain features, and plantar pressure features, the overall state changes of the subject during natural walking can be better reflected. Among them, the surface electromyography time-domain analysis indexes (including contribution rates) can represent physical characteristics such as the force contribution, intensity change, and time change of contraction and relaxation of muscles and the neuromuscular control system during movement; the surface electromyography frequency-domain analysis indexes mainly reflect the fatigue degree of muscles and the neuromuscular system. That is, assuming that the force sharing is the same, different muscle fatigue degrees may also affect the occurrence probability of the fall risk. In addition, the plantar pressure features mainly reflect the body balance during movement. The reason for including the balance indexes is that in addition to muscles and the neuromuscular control system, factors such as bones and soft tissues also have a certain impact on balance.

[0086] Therefore, in the embodiments of the present application, through the frequency-domain features and time-domain features, the fatigue degree and cumulative impulse contribution rate of muscles can be more accurately reflected. Then, combined with the stable plantar pressure features, after feature processing, they are input into the fall prevention warning model for fall prevention prediction, so as to more comprehensively realize the prediction of the fall risk and obtain a more accurate fall prevention warning result. Further, in the embodiments of the present application, feature sampling and feature dimensionality reduction processing are also performed to improve the performance and stability of the fall prevention warning model and improve the accuracy of the fall prevention warning result.

[0087] Figure 3 is an exemplary flowchart showing the overall process for fall prevention warning according to the embodiments of the present application. It should be understood that Figure 3 is the above Figure 2 a specific embodiment of method 200, so the above description about Figure 2 also applies to Figure 3 .

[0088] As Figure 3 shown, at steps S301 and S302, the surface electromyography detection signals of the lower limbs of the subject at multiple target muscles and the pressure detection signals of the plantar surface of the subject at multiple target detection points are respectively collected during natural walking. The specific collection method can refer to the content described in the above Figure 1A and Figure 1B , and the present application will not elaborate here. Then, at step S303, the surface electromyography detection signals are preprocessed in the frequency domain to obtain frequency-domain signals, and the power spectral density function is calculated. In some embodiments, the surface electromyography detection signals can be preprocessed in the frequency domain by, for example, fast Fourier transform. The power spectral density function is calculated by, for example, the period method or the autocorrelation function. Then, at step S304, the median frequency and / or the mean power frequency can be calculated using the above formulas (1) and (2).

[0089] Further, at step S305, step S306, and step S307, the surface electromyogram detection signal is sequentially subjected to, for example, filtering, rectification, and smoothing processes to obtain a preprocessed time-domain signal. Then, at step S308, by calculating the total area under the time-domain signal, and then dividing the total area by the total time, the integrated electromyogram value per unit time is obtained. The calculation of the total area under the time-domain signal can refer to the above formula (4). Further, at step S309, plantar pressure features are extracted based on the pressure detection signal. The plantar pressure features may include, for example, the degree of foot center of gravity offset and / or the foot center of gravity trajectory curve. In an implementation scenario, the aforementioned plantar pressure features can be calculated based on the above formulas (5)-(7).

[0090] Based on the aforementioned extracted multiple features (median frequency and / or mean power frequency, integrated electromyogram value per unit time, plantar pressure features), at step S310, feature sampling and / or feature dimensionality reduction are performed to obtain a target feature result. For more details on feature sampling and / or feature dimensionality reduction, reference can be made to the above Figure 2 description, which is not elaborated herein in this application. Additionally, at step S311, the basic information of the subject can also be collected. The basic information may include, for example, age information, historical fall information, medical history information, gender, weight, etc. Further, the target feature result and the basic information are respectively input into a support vector machine classifier, a random forest classifier, and a gradient boosting classifier at step S312, step S313, and step S314 for fall prevention prediction to obtain respective fall prevention warning results. Finally, at step S315, the final fall prevention warning result is obtained by calculating the weighted sum of the fall prevention warning results output by the three models. The final fall prevention warning result includes the probability of whether there is a fall risk, and can timely remind the subject to adjust the gait before a fall event occurs, avoiding the occurrence of a fall event.

[0091] Figure 4 is an exemplary diagram showing a comparison between the fall prevention warning method and the single classifier prediction method according to an embodiment of the present application. As Figure 4The accuracy rate, precision rate, recall rate, and F1 score corresponding to the support vector machine classifier combined with dimensionality reduction, the random forest classifier combined with dimensionality reduction, the gradient boosting classifier combined with dimensionality reduction, and the fall prevention warning method of the present application are successively shown. Among them, the fall prevention warning method according to the embodiment of the present application has a relatively high recall rate. It should be understood that considering that the fall risk warning is a conservative medical prevention screening model, and the subsequent treatment is mainly conservative rehabilitation treatment with a relatively small treatment cost. If the fall prevention warning is not timely, it will lead to a sharp increase in the treatment cost. Therefore, the missed diagnosis rate should be reduced as much as possible, and the high recall rate of the fall prevention warning method of the present application can greatly reduce the missed diagnosis rate, sacrificing part of the misdiagnosis rate index fusion strategy. That is, assuming that the risk of missing a patient with a fall risk is much greater than misdiagnosing a patient without a fall risk as having a fall risk, to improve the accuracy of the fall prevention warning.

[0092] Figure 5 FIG. is an exemplary structural block diagram showing a device 500 for fall prevention warning according to an embodiment of the present application. As Figure 5 shown, the device 500 of the present application may include a processor 501 and a memory 502, where the processor 501 and the memory 502 communicate through a bus. The memory 502 stores program instructions for fall prevention warning. When the program instructions are executed by the processor 501, the method steps described above in combination with the accompanying drawings are implemented: at least collecting surface electromyography detection signals of the lower limbs of the subject at multiple target muscles during natural walking and pressure detection signals of the soles of the feet of the subject at multiple target detection points; respectively performing frequency domain analysis and time domain analysis on the surface electromyography detection signals to obtain frequency domain features and time domain features correspondingly; extracting plantar pressure features based on the pressure detection signals; performing feature processing operations on the frequency domain features, the time domain features, and the plantar pressure features to obtain a target feature result; and inputting the target feature result into a fall prevention warning model for fall prevention prediction to obtain a fall prevention warning result.

[0093] According to the above description in combination with the accompanying drawings, those skilled in the art can also understand that the embodiments of the present application can also be implemented through software programs. Therefore, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions for fall prevention warning. When the computer-readable instructions are executed by one or more processors, the method for fall prevention warning described in the present application in combination with the attached Figure 2 drawings is implemented.

[0094] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, 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 computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0095] It should be noted that although the operations of the method of the present application are described in a specific order in the drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be changed in the order of execution. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step for execution, and / or one step can be decomposed into multiple steps for execution.

[0096] It should be understood that when terms such as "first", "second", "third", and "fourth" are used in the claims, the specification, and the drawings of the present application, they are only used to distinguish different objects, rather than to describe a specific order. The terms "including" and "comprising" used in the specification and claims of the present application indicate the presence of the described features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0097] It should also be understood that the terms used in the specification of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. As used in the specification and claims of the present application, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms. It should be further understood that the term "and / or" used in the specification and claims of the present application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0098] Although the embodiments of the present application are as above, the above content is only an example used for easy understanding of the present application, and is not intended to limit the scope and application scenarios of the present application. Any person skilled in the art within the technical field of the present application can make any modifications and changes in the form of implementation and details without departing from the spirit and scope disclosed by the present application. However, the scope of patent protection of the present application shall still be subject to the scope defined by the appended claims.

Claims

1. A method for fall prevention warning, comprising: Collecting at least the surface electromyography detection signals of the lower limbs of the subject at multiple target muscles and the pressure detection signals of the soles of the feet of the subject at multiple target detection points during natural walking; Performing frequency-domain analysis and time-domain analysis on the surface electromyography detection signals respectively to obtain frequency-domain features and time-domain features correspondingly; Extracting plantar pressure features based on the pressure detection signals; Performing feature processing operations on the frequency-domain features, the time-domain features and the plantar pressure features to obtain a target feature result; And Inputting the target feature result into a fall prevention warning model for fall prevention prediction to obtain a fall prevention warning result.

2. The method according to claim 1, wherein the frequency-domain features at least include median frequency and / or mean power frequency, and the time-domain features at least include the integrated electromyography value per unit time; the plantar pressure features at least include the degree of foot center of gravity offset and / or the foot center of gravity trajectory curve.

3. The method according to claim 2, wherein the frequency-domain features are obtained by the following operations: Performing frequency-domain preprocessing on the surface electromyography detection signals to obtain frequency-domain signals; Calculating the power spectral density function based on the frequency-domain signals; and Calculating the median frequency and / or the mean power frequency according to the power spectral density function.

4. The method according to claim 2, wherein the time-domain features are obtained by the following operations: Performing time-domain preprocessing on the surface electromyography detection signals to obtain preprocessed time-domain signals; Calculating the total area under the time-domain signals; and Dividing the total area by the total time to obtain the integrated electromyography value per unit time.

5. The method according to claim 4, wherein the time-domain preprocessing includes one or more of filtering processing, rectification processing or smoothing processing.

6. The method according to claim 2, wherein extracting plantar pressure features based on the pressure detection signals includes: Processing the pressure detection signals into a pressure detection time series point set; Calculating the pressure cumulative impulse according to the pressure detection time series point set; Calculating the degree of foot center of gravity offset based on the pressure cumulative impulse; And / or Calculating the foot center of gravity trajectory curve according to the pressure detection time series point set.

7. The method according to claim 1, wherein the feature processing operations include feature sampling and / or feature dimensionality reduction.

8. The method according to claim 7, wherein the feature sampling is achieved by the following operations: Calculating the nearest neighbor features corresponding to a small number of categories among the frequency-domain features, the time-domain features and the plantar pressure features; and Interpolating between the corresponding features and the nearest neighbor features to generate new features to achieve the feature sampling.

9. The method according to claim 1, wherein the fall prevention warning model includes a support vector machine classifier, a random forest classifier and a gradient boosting classifier.

10. The method according to claim 9, wherein inputting the target feature result into a fall prevention warning model for fall prevention prediction to obtain a fall prevention warning result includes: Input the target feature results into the support vector machine classifier, the random forest classifier, and the gradient boosting classifier respectively for fall prevention prediction, and obtain their respective fall prevention warning results; and Obtain the final fall prevention warning result according to the weighted sum of their respective fall prevention warning results.

11. The method according to claim 1, further comprising: Obtain the basic information of the subject; and Input the basic information and the target feature results into the fall prevention warning model for fall prevention prediction to obtain a fall prevention warning result.

12. A device for fall prevention warning, comprising: A processor; and A memory, which stores program instructions for fall prevention warning. When the program instructions are executed by the processor, the device implements the method according to any one of claims 1-11.

13. A computer-readable storage medium, on which computer-readable instructions for fall prevention warning are stored. When the computer-readable instructions are executed by one or more processors, the method according to any one of claims 1-11 is implemented.

Images