Fall risk assessment device

Abstract

A human balance sensor for assessing a user's risk of falling includes a transparent glass plate, a sheet of emulsion positioned on a top surface of the glass plate, a light source to project light into an edge of the glass plate, and a high resolution camera positioned below the glass plate to capture light diffused from the glass plate when a user's foot exerts pressure on the glass plate. Based on the principle of frustrated total internal reflection (FTIR), when a user stands on the glass plate with a foot, the total internal reflection condition at the location of pressure due to the foot's pressure is eliminated, and diffused light passes through a bottom surface of the glass plate and forms a haptic image of the foot's contact area, which can be analyzed over time to determine the user's balance ability, and thus the user's risk of falling.

Description

[0001] This international patent application claims the benefit of U.S. Provisional Patent Application No. 63 / 210,596, filed June 15, 2021, the entire contents of which are incorporated herein by reference for all purposes. Technical Field

[0002] This invention relates to assessing the risk of falls in older adults, and more specifically, to an apparatus that allows a user to step on it and measure the dynamic force distribution on the user's feet to calculate the user's risk of falling. Background Technology

[0003] Falls pose a major threat to the health and independent living of older adults. It is estimated that 10% of falls among older adults are associated with fractures, and some can result in head injuries and death. Falls and related injuries (such as hip fractures) are risk factors for placement in nursing homes [MT1997]. Even minor falls can lead to significant impairment of mobility and daily living activities in older adults. They can trigger a negative domino effect, leading to complications such as pneumonia, thromboembolism, loss of autonomy, disability, anxiety, and depression, impairing individual quality of life and placing a burden on families. Falls in older adults are costly for the healthcare system because they often require accident and emergency services, as well as long-term hospitalizations, procedures, surgeries, and rehabilitation services. The societal burden of falls will increase as the population ages.

[0004] Since an average of 20% of the older population suffers accidental falls each year, significant harm could be avoided if at least 10% of the older population were alerted to the risk of an impending fall so they could take appropriate preventative measures. In particular, serious injuries such as fractures, head injuries, and death could be reduced from 10% to 3% through appropriate preventative measures. However, current fall and balance assessment tools require a clinician to be present to perform the tests and interpret the results. Because individuals lack practical and objective ways to assess their daily fall risk, they may underestimate or overestimate it. While underestimation can lead to unsafe behavior and increased falls, overestimation is also problematic because it can generate unfounded fear of falls and their downstream effects, such as limited physical mobility, social isolation, and functional loss.

[0005] Balance ability assessment relies on specific procedures or methods that can quantitatively or qualitatively analyze human balance ability. Currently, there are various methods for assessing human balance ability. These methods can be divided into three categories: observational methods, scale methods, and balance testing device methods.

[0006] The simplest and most commonly used methods are observational methods, such as the Romberg test [FB1982, YA2011], the Single-Leg Standing Test (OLST) [TM2009], and the Postural Stress Test [JC1990]. In the Romberg test, the subject closes their eyes, stands on both feet, and raises their arms forward. The assessor then evaluates balance based on the degree of swaying. Similar to the Romberg test, in the OLST, the subject stands on one leg. The Postural Stress Test is clinically applicable and used to obtain quantitative measurements. In this method, an instability force is applied to the subject's lower back. Balance is assessed based on the subject's ability to maintain an upright posture.

[0007] More refined methods include scale-based approaches, such as the Berg Balance Test [SM2008], the Tinetti Test [SK2006], and the Timed Stand-Up-Walk Test (TUG) [TS2002]. The Tinetti Test has also been widely used for assessing balance ability and predicting falls in older adults. In this approach, the assessor scores the subject's performance on a range of different tasks.

[0008] The first balance testing device method was introduced by Yuriy V. Terekhov [YT1976] in 1976 and is known as the stability measurement method. This measures the mechanical oscillations of the subject's center of gravity and converts them into electronic signals. A computer is then used to analyze the frequency, amplitude, and duration of the oscillations to assess the subject's balance ability. Over the years, this method has been improved and developed into different versions, but the basic principle remains unchanged; these versions all consist of a pressure testing plate, a computer, and specialized analysis software (see Figure 1).

[0009] Balance testing devices have even been modified for recreational use. For example, the balance board [RC2010] (see Figures 2A and 2B) uses Bluetooth technology and contains four pressure sensors, one at each corner, to measure the center of pressure under each foot. Similar to the balance board are the Intec Action Board and GameOn compatible balance boards.

[0010] Fall risk assessment is used to determine whether a subject's fall risk is low, moderate, or high. It is primarily performed on older adults and typically involves an initial screening followed by completion of a set of tasks known as fall assessment tools. The initial screening includes a series of questions about the subject's overall health, as well as whether they have a history of falls or problems with balance, standing, or walking; while the fall assessment tools test the subject's strength, balance, and gait.

[0011] Initial screening questions include: “Have you fallen in the past year?”; “Do you feel unsteady when standing or walking?”; and “Are you worried about falling?”. Many questionnaires are available for screening, such as the Patient Fall Questionnaire [NR1984] and the Fall Assessment Questionnaire [LR1993].

[0012] Fall assessment tools include the aforementioned TUG test [TS2002], the 30-second chair stand test [KJ2015], and the 4-stage balance test [JG2017]. In the TUG test, the subject starts in a chair, stands up, and walks approximately 10 feet at a normal pace while the healthcare provider examines the subject's gait. The 30-second chair stand test examines strength and balance. First, the subject sits in a chair with their arms crossed over their chest. They then repeat standing and sitting for 30 seconds, while the healthcare provider counts the number of repetitions. The 4-stage balance test examines the subject's ability to maintain balance. The subject stands in four different positions, holding each position for 10 seconds. In the first position, the subject stands with both feet together. In the second position, the subject moves one foot forward halfway. In the third position, the subject moves one foot completely in front of the other so that the toes touch the heel of the other foot. In the fourth position, the subject stands on only one foot. There are many other similar fall assessment tools, such as the Berg Balance Test [KB1989], the Fall Screening Test for Older Adults [JC1998], the Dynamic Gait Index [SW2000], and the Tinetti Performance Orientation Motor Test [MT1986].

[0013] Fall assessment also employs scales, such as the Gait Disorder Rating Scale (GMRS) [LW1990, JV1996] and the Morse Fall Scale [JM1989]. For example, the GMRS includes variables designed to describe the gait of test subjects associated with increased fall risk, such as foot and arm movements, alertness, shuffling, staggering, swaying, the percentage of time spent in the sway phase of the gait cycle, foot contact, hip range of motion, knee range of motion, elbow extension, shoulder extension, shoulder abduction, arm-heel strike synchronization, head forward, shoulder elevation, and upper trunk forward flexion.

[0014] Among various methods for assessing human balance ability, observational and scale-based methods require assessors, making the assessment subjective. Similarly, fall assessment tools require healthcare providers to administer the assessment, implying that the results are also subjective. Therefore, balance testing devices are more objective because they do not require assessors or healthcare providers. While most balance testing devices rely on electronic force sensor arrays (see Figure 3), their structure results in very low resolution of the obtained force distribution. Furthermore, the equipment is expensive. Therefore, a method that can objectively assess human balance ability with high accuracy and low cost is currently lacking. Summary of the Invention

[0015] In one aspect of the present invention, a human balance sensor for assessing a user's fall risk is provided, comprising:

[0016] A transparent glass plate having a flat upper and lower surface, and the refractive index of the transparent glass plate being greater than that of air;

[0017] A latex sheet is positioned on the top surface of the glass plate, and during operation, the feet of a standing user are placed on the latex sheet.

[0018] A light source, positioned to direct light from the edge of the glass plate into the glass plate;

[0019] A high-resolution camera, positioned below the lower surface of the glass plate, is provided to capture light diffused from the glass plate when a user's foot applies pressure to it; and

[0020] Therefore, based on the principle of Frustrated Total Internal Reflection (FTIR), when a user stands on the glass plate with their foot: (a) the latex sheet presses against the upper surface of the glass plate, (b) the total internal reflection condition at the pressure point of the foot is eliminated, and (c) the diffuse reflection of light passes through the bottom surface of the glass plate and is focused onto the image plane of the camera, so as to form a tactile image of the contact area of ​​the foot with different pixel intensities based on the different pressures of the foot at different positions on the glass plate.

[0021] In a preferred embodiment, the light emitted by the light source is invisible, or more preferably infrared light. The light source can be any light-emitting device capable of emitting invisible light (i.e., invisible to the human eye) at a fixed wavelength, particularly infrared light at a fixed wavelength. A light source is advantageous for evaluation because it minimizes noise detected throughout the process, thereby improving measurement accuracy.

[0022] In another embodiment, the human balance sensor further includes a waveguide or optical waveguide to block unwanted light from entering the transparent glass plate or minimize unwanted effects caused by noise. This can improve the accuracy of the assessment.

[0023] In another aspect of the invention, a human balance sensor for assessing a user's fall risk is provided, comprising:

[0024] shell;

[0025] Two transparent glass plates, each having a flat upper and lower surface, and having a refractive index greater than that of air, are positioned side by side on the top of the housing and spaced apart from each other approximately the distance between the feet of a standing human body.

[0026] Latex sheet, a latex sheet is positioned on the top surface of each of the glass plates, and during operation, the feet of a standing user are placed on the corresponding latex sheet;

[0027] A light source, the light source being positioned to direct light from the edge of the glass plate into each of the glass plates;

[0028] A high-resolution camera, positioned below the lower surface of the glass plate, is provided to capture light diffused from the glass plate when pressure is applied to it; and

[0029] Therefore, based on the principle of suppressed total internal reflection (FTIR), when a user stands on the glass plate with their foot: (a) the latex sheet presses against the upper surface of the corresponding glass plate, (b) the total internal reflection condition at the pressure point of the foot is eliminated, and (c) the diffuse reflection of light passes through the bottom surface of the glass plate and is focused onto the image plane of the camera, so as to form a tactile image of the contact area of ​​the foot with different pixel intensities based on the different pressures of the foot at different positions on the glass plate.

[0030] Similarly, human balance sensors in this regard may also include invisible light sources and optical waveguides as described above.

[0031] In another aspect, the present invention relates to a method for assessing a user's risk of falling using a human balance sensor as described herein.

[0032] Compared to RP Bettes and T. Duckworth's "A device for measuring plantar pressure under the sole of the foot," the hardware design has some key differences:

[0033] (1) Light source: In the prior art, visible light sources are used. However, the present invention uses infrared LEDs with a fixed wavelength. This can significantly reduce noise in total internal reflection (TIR) ​​and improve measurement accuracy.

[0034] (2) Glass waveguide: The glass used in the prior art is conventional glass. However, the present invention uses a glass waveguide that only allows infrared wavelengths of light to pass through and form a TIR. It can completely block light from the environment, making the measurement more accurate. Measurements can even be performed without a soft surface layer.

[0035] (3) Mirror and camera: Since the present invention uses infrared light of a fixed wavelength as a light source, both the mirror and the camera are used at the same wavelength. It can reduce noise by blocking ambient light and improve accuracy by using light of a fixed wavelength.

[0036] To achieve objectivity in assessing human balance, this invention employs a comprehensive physical system called a "balance sensor" for balance evaluation. This system has a dedicated sensing unit for collecting force distribution information under the feet. The sensing unit does not rely on an electronic force sensor array but operates based on the principle of suppressed total internal reflection (FTIR) optics. This invention is widely used in the development of tactile sensors for robots [HM1992, NN1990, SB1988_1, SB1988_2], extending the principle far beyond robotics and utilizing abundant available tactile information applied to the study of human balance. Although the structure of the FTIR-based sensing unit is much simpler than that of an electronic force sensor array, it is more sensitive and achieves higher force distribution resolution. Furthermore, its manufacturing cost is very low.

[0037] Since the sensing unit of the balance sensor is based on optical principles, the final signal collection device is a camera. The force distribution underfoot is recorded in the image; therefore, information on changes in force distribution is encoded into a video format (see [link to video]). Figure 4 When subjects attempt to maintain their balance, the force distribution under their feet changes over time, but the changes are very subtle. A highly sensitive and high-resolution balance sensor can not only detect these minute changes, but the device can also analyze the changes based on recorded video to assess a person's balance ability.

[0038] Once the collected data is in video format, it can be analyzed using advanced computer vision processing and AI technologies, greatly enhancing the balance sensor's ability to obtain human balance information. This is another major advantage compared to traditional balance testing devices. If the data analysis results need to be discrete, an algorithm is used to map the raw video data to these discrete results, which is a classification solution. One approach is to manually extract certain features from the raw video and then set up a rule-based algorithm to classify different videos, or to train a machine learning model to classify different videos. Another approach is to simply use the raw video data to train a deep learning algorithm, such as a 3D convolutional neural network (CNN), to generate a classification model. If the data analysis results are continuous, a functional relationship is established between the raw video data and the final continuous results. In this case, it is a regression process. As with the previous approach, features can be extracted manually and a regression model trained accordingly, or a deep learning method can be used to perform the regression. However, in both cases, it is best to compress the raw video data first to extract useful information and discard redundant information before analysis. This is because there is always a large amount of video data, and this data allows for a reduction in model size and improved processing efficiency. In terms of data compression, there are several available methods, such as compression sensing and automatic encoding.

[0039] Alternatively, another option is to analyze data from balance sensors. Since the relationship between pressure values ​​and pixel intensity is constant, this pressure-pixel relationship can be calibrated experimentally, converting the raw image into realistic pressure distribution information. This calibration has been completed. [SW2019, SW2020] Using real pressure distribution variation information, dynamic analysis of the human body can be performed. Specifically, a dynamic model can be built for specific human movements during testing. This model includes multiple differential equations associated with the pressure distribution variation process under the feet. Using this information, a series of boundary value conditions can be set according to the physical characteristics of the human body, allowing the solution of differential equations to obtain a detailed body movement process. A novel differential equation solving algorithm based on the Generative Adversarial Triangle (GAT) model has been proposed, which can solve nonlinear differential equations under any feasible boundary value conditions. Using the detailed body movement process, further balance ability assessment or fall assessment can be achieved.

[0040] There are two possible application scenarios. (1) To qualitatively identify whether a subject is normal, sick, or intoxicated. For example, it may be necessary to conduct a physical examination on a patient to identify whether the subject has a specific neurological disease, or to use a physical examination to identify intoxicated drivers. (2) To quantitatively score a subject's balance ability or likelihood of falling. For example, as part of athlete selection, pilot selection, or during fall assessment of elderly patients. Attached Figure Description

[0041] The foregoing and other objects and advantages of the invention will become more apparent when considered in conjunction with the following detailed description and accompanying drawings, in which similar reference numerals denote similar elements in the various views, and in the drawings:

[0042] Figure 1 is a perspective view of a typical balance tester operated by a user in the prior art;

[0043] Figure 2A is a photograph of the top surface of the balance plate in the prior art, and Figure 2B is a photograph of the bottom surface of the balance plate in the prior art.

[0044] Figure 3 is a photograph of a step force detector in the prior art;

[0045] Figure 4 A series of camera images showing force distribution information;

[0046] Figure 5 This is a schematic illustration of the optical principle of suppressed total internal reflection (FTIR) utilized in this invention;

[0047] Figure 6 This is a schematic diagram of the balance sensor according to the present invention;

[0048] Figure 7A This is an illustration of a user standing on the balance sensor of this invention. Figure 7B It is a simplified line drawing model of the user's human body from behind. Figure 7C It is a simplified line drawing model of the user's human body side view;

[0049] Figure 8 This is an example of the pressure distribution under the feet of a user standing on the balance sensor of this invention;

[0050] Figure 9 This is a diagram illustrating a neural network structure that can be used in this invention;

[0051] Figure 10 This is a flowchart of a Generative Adversarial Tri-model (GAT) model used to solve differential equations.

[0052] Figure 11 It is a graph of the regression model in fall assessment software;

[0053] Figure 12 This is a photograph of the prototype balance sensor according to the present invention;

[0054] Figure 13 This is an illustration of a graphical user interface for the balance sensor system of the present invention; and

[0055] Figures 14A to 14JThis is when the subject is in a normal state ( Figures 14A to 14E ) and when they consume large amounts of alcohol ( Figures 14F to 14J The graph shows the center of pressure (COP) of five test subjects measured on the balance sensor of the present invention within a 10-second cycle. Detailed Implementation

[0056] The main components of the balance sensor of this invention are based on the principle of suppressed total internal reflection (FTIR), such as... Figure 5 As shown. This sensor mainly consists of a high-resolution camera 10, an LED light source 12, and a thick, transparent glass plate 14 with flat upper and lower surfaces. On the top surface of the glass plate, there is a latex sheet 13. The LED light source 12 directs light from the edge of the glass plate 14 into the glass plate. Since the refractive index of glass is greater than that of air, if nothing is in contact with the glass surface, all the light will be reflected back into the glass plate, and the camera 10 will not be able to capture any light. However, when a user stands on the glass plate and places their foot 15 on the latex sheet 13, the latex sheet will be pressed against the upper surface of the glass plate. At the contact area, total internal reflection is disrupted, and diffuse reflection of light occurs instead. A portion of the diffuse light 17 is captured by the camera 10 and focused onto the camera's image plane. Therefore, a tactile image of the contact area will be formed with different pixel intensities.

[0057] The varying pixel intensities originate from the different diffuse light intensities at each point of contact. These varying diffuse light intensities are solely caused by different contact pressures, as the surface properties of the latex and glass are uniform throughout. Therefore, the tactile image captured by the camera is actually a force distribution image of the pressure exerted underfoot. Furthermore, since the camera can record video at high frame rates, pressure distribution information can be recorded over time at a high frame rate. Figure 6A schematic diagram of the balance sensor device of the present invention is shown. The balance sensor device includes a sensing unit and a microprocessor 20. The sensing unit is generally represented as a box 22, which may have a black interior and two glass plates 14 on the top. Above the two glass plates are two disposable latex sheets. Both glass plates are surrounded by LED light strips 12', which may emit red light, for example. At the bottom inside the box is a camera 10. However, since the peak tremor frequency that the human limbs can reach is only about 10 Hz [JM1997], according to Shannon's sampling theorem, in order to retain all the information contained in the original human movement, the camera's frame rate is 30 fps. The resolution is 1920x1440, which is much higher than that of existing balance testing devices currently available on the market and can meet the requirements for assessing human dynamic balance. The LED lights and camera are operated by the microprocessor (microcomputer) 20 according to the FTIR principle. In addition, the fall assessment result, i.e., the probability of the user falling, which can be calculated in the microprocessor, is displayed on a display 29 positioned on the top surface of the box. The camera can be wired to microprocessor 20 or another remote computing device, or it can be wirelessly connected so that images generated by the camera can be transmitted to a remote display, such as a mobile device (e.g., iPhone) 25. Battery 27 powers the camera, lights, and microprocessor. As shown in Figure 7, a coordinate system model can be constructed for a standing person to describe their dynamic balance. Figure 7A The human body shown can be used as follows Figure 7B The simulation uses three rigid rods hinged together. When viewed from behind, the torso and arms are considered as single rods that remain upright at all times. The two legs are considered as... xz Two rods swinging in a plane. Two legs in... xz The rotation in the plane remains the same. Viewing the model from the right ( Figure 7C The whole body can yz The model swings in a plane. The rotation of the torso and both legs remains constant. The mass of the human body is taken as m, with the torso at 3m / 5 and each leg at m / 5. The height is h, with each leg and torso at h / 2. The entire model has two degrees of freedom (DoF). The first degree of freedom is... θ 1 ,express yz Rotation in a plane, with the positive direction being counterclockwise. The second degree of freedom is... θ 2 , indicating that the two legs are xz In a plane, rotation is also counterclockwise in the positive direction. The torso remains upright at all times. Figure 8 The image shows the pressure distribution under the user's feet as measured by a balance sensor. p(x,y)A dynamic model of a standing human body can be constructed using the Lagrange equation, as shown in Equation 1. Here... (x 0 ,y 0 ) These are the coordinates of the midpoints of the two ankles. (COP x COP y ) These are the coordinates of the center of pressure (COP) under the feet. Equation 2 shows... (COP x COP y ) The calculation method. (x 0 ,y 0 ) The value is available (COP x COP y ) The average value over a relatively long period of time is approximated.

[0058] (Formula 1)

[0059] (Formula 2)

[0060] Equation 1 is a nonlinear ordinary differential equation without an analytical solution. Even if only a numerical solution is needed, the equation still lacks initial conditions. However, Equation 1 can be solved using other boundary conditions. Since the user or tester will not fall during the experiment, therefore... θ 1 and θ 2 It must oscillate around 0 at all times. (It has an angle.) and It must also oscillate around 0 at all times. Because θ 1 , θ 2 , , None of them diverge, therefore their integrals over the entire experimental period (0, T) are considered to be 0. In this way, the boundary conditions can be obtained, as shown in Equation 3.

[0061] (Formula 3)

[0062] This invention employs a novel method to solve ordinary differential equations, namely the so-called Generative Adversarial Triangle (GAT) model. The GAT method combines analytical methods with neural networks to numerically solve nonlinear ordinary differential equations with the following non-initial conditions (e.g., Equation 3):

[0063] The first equation 1 is transformed into four first-order differential equations, as shown in equation 4, where , , , Specifically, four neural networks are used to represent... , , , The neural networks described have the same network structure, such as... Figure 9 As shown. The number of hidden nodes is equal to the camera's T * frame rate. This network can reproduce any numerical solution of Equation 4 using a simple procedure. The loss function is the mean squared residual of Equation 4 at all discrete numerical points, where the derivative is approximated using the Euler or Runge-Kutta method.

[0064] (Formula 4)

[0065] Figure 10 The flowchart of the GAT model is shown. As a first step 30, the neural network is initialized randomly or using an approximate solution. Then, the GAT model is trained (step 32) until convergence to obtain the numerical solution of Equation 4. Specifically, the neural network is trained using either the Runge-Kutta loss function or the Euler loss function until convergence. To determine this numerical solution, a decision is made at step 34 regarding whether the boundary value condition is met. If it is met, the process ends at step 36. If it is not met, the process proceeds to step 38, where the output of the current neural network is adjusted to meet the boundary value condition and the network parameters are reset so that the output is the adjusted value. A new network is trained at step 32, and the process is repeated until the boundary value condition is met.

[0066] Furthermore, an approximate solution is obtained and then used as the initialization of the GAT model. This method allows the HAN model to converge faster and better. Specifically, for Equation 4, the nonlinear terms in the equation are first discarded, transforming Equation 4 into the linear differential equation Equation 5. Since Equation 5 is linear, its numerical solution can be obtained using the finite difference method with the help of the boundary condition Equation 3. The numerical solution of Equation 5 is then used as the initialization of the HAN model. This significantly accelerates the convergence of the HAN model. The multiple differential equations solved using this method are related to the changing pressure distribution under the user's feet.

[0067] (Formula 5)

[0068] Numerous measurements based on different coordinates of the center of pressure (COP) can be used to assess human balance or for fall assessment. Time-domain “distance” measurements [TP1996] include the average distance of the COP from the origin, the root mean square distance of the COP from the origin, the total length of the COP path, and the average velocity of the COP [MG1990], etc. Time-domain “area” measurements include the area of ​​a 95% confidence circle (the area of ​​a circle with a radius equal to one side), the 95% confidence limit of an RD time series, the area of ​​a 95% confidence ellipse (expected to enclose approximately 95% of the points on the COP path), and so on. There are also time-domain “mixed” measurements. For example, the sway area estimate estimates the area enclosed by the COP path per unit time [AH1980]. The average frequency is the rotational frequency of the COP as it travels around a circle with a radius equal to the average distance, expressed in revolutions per second or Hz [FH1989]. Fractal dimension is a unitless measure of the degree to which a curve fills the space it encompasses.

[0069] In addition to time-domain measurements, frequency-domain measurements are also possible. Various qualitative and quantitative methods have been used to characterize the frequency distribution of COP shift [ID1983, TP1993], such as power spectral moment, total power, 50% power frequency, 95% power frequency, centroid frequency, and frequency dispersion. There are also some statistical measurements, such as Romberg ratio and Riley phase plane parameters.

[0070] It is worth noting that in 1981, the International Society of Posturography (ISP) recommended the use of two COP-based measurements in its recommendations on standardizing force-platform-based postural stability assessments [ID1983]: the average velocity of the COP and the root mean square distance of the COP from the origin.

[0071] Since COP can be calculated based on the pressure distribution under the user's feet obtained through the balance sensor of this invention, all the aforementioned COP-based measurements can be employed in the application of the balance sensor. Furthermore, the pressure distribution provides richer information than a single COP location. Regarding the pressure distribution under the user's feet, footprint analysis can be used. Footprint analysis is a functional diagnostic tool that provides accurate and reliable information for foot function analysis and foot pathology diagnosis. Foot deformities and functional impairments can be detected during the analysis of barefoot pressure distribution. This additional pathological information will greatly facilitate balance assessment and fall assessment.

[0072] Furthermore, in the aforementioned COP-based measurements and assessments, COP can be replaced by the center of gravity (COG). In this way, a series of COG-based measurements can be created. Moreover, since the movement of COG represents the actual physical movement of the human body and COP can be considered as control of the body to maintain balance, the body's balance control ability can be analyzed by comparing changes in COP and COG. These changes are direct indicators of the body's balance ability and fall tendency. Therefore, a more accurate assessment can be obtained.

[0073] A fall assessment software can be developed by integrating balance ability measurements (COP), footprint analysis, and COG measurements. The core of the software is a regression model generated through machine learning, which outputs the fall probability of the test subject. This regression model is fused through two parts. One part is based on a support vector machine (SVM). The results of COP-based measurements, footprint analysis, and COG-based measurements are extracted and fed into this SVM. This SVM outputs the fall probability of the user or test subject. The other part is based on the use of a deep convolutional neural network (DNN), which directly takes video data from a balance sensor as input and outputs another fall probability for the test subject. The weighted average of the two fall probabilities is then obtained from the SVM and the deep neural network and used as the final evaluation result of the fall assessment software.

[0074] Figure 11 A diagram of the regression model is shown. At step 40, balance sensor video data is obtained. This balance sensor video data is used to determine COP-based measurements at 41, footprint analysis at 43, and COG-based measurements at 45, and is also passed to a convolutional neural network 42. The feature outputs of COP, footprint analysis, and COG are combined in a support vector machine 46, whose output is fall probability 1. The output of the convolutional neural network 42 is fall probability 2. Fall probabilities 1 and 2 are combined in a fusion machine 44, whose output becomes the fall assessment result 48. All parameters in the support vector machine 46, convolutional neural network 42, and fusion weights 44 are trainable. To obtain this model, human experimental data was collected and labeled with data from normal individuals, the elderly, and patients whose balance abilities are affected by disease for training and testing.

[0075] The evaluation results are displayed on screen 29 on the balance sensor and / or via voice prompts from a speaker (not shown). Additionally, the results can be transmitted via WiFi or Bluetooth to mobile device 25 and / or other computers (not shown) for display and recording.

[0076] The rectangular box 22 of the sensing unit can be, for example, approximately 60 × 43 × 10 cm.3 ,like Figure 12 As shown. Each of the two glass plates 14 measures approximately 36 × 18 × 1 cm. 3 Two disposable latex sheets in Figure 12 The image shows the test subject positioned above two glass plates. The test subject's feet are shown on each latex sheet.

[0077] Figure 13 The graphical user interface (GUI) of the balance sensor is shown. This GUI is primarily used for sensor programming and maintenance and can run on mobile devices or PCs connected to the balance sensor via WiFi. The interface can also be used to display a live stream of images captured by camera 10. Figure 13 The image shows a tactile image of the test subject's foot. Due to the uneven distribution of pressure under the user's or test subject's feet, the pixel intensity of the image varies. On the tactile image, three white dots represent, from left to right, the pressure on the left foot, the entire pressure distribution, and the pseudo-center of pressure on the right foot. The coordinates of these centers are shown in the upper right corner of the GUI. The position and coordinates of these centers also differ in the live video stream.

[0078] exist Figure 13 On the right side of the GUI, besides the coordinates of the three centers, there are buttons for controlling the camera. These buttons allow you to manually record video by controlling the start and stop times. Alternatively, the video duration can be arbitrarily set to a fixed value, and video data collection can be initiated. Additionally, the recorded video can be downloaded from the camera to a computer for further research. At the bottom of the GUI, there are several entries for entering personal information about the user or test subject, such as age, gender, height, and weight. After clicking the "Collect Data" button, the video will be automatically recorded and downloaded to the computer. All test subject information will be recorded in a separate CSV file.

[0079] The procedure to open the balance sensor is as follows:

[0080] i. Place two clean, disposable latex sheets on the glass plate on top of the sensor.

[0081] ii. The sensor is activated by stepping on it. For this purpose, a pressure switch (not shown) is provided below the upper surface.

[0082] iii. The test subject stands still and the measurement begins after a few seconds, with the voice command displayed or from a speaker (not shown) indicating the measurement.

[0083] iv. After a few minutes of measurement, the tester will be notified of the end of the test via visual or audio commands.

[0084] v. The tester can get off the sensor.

[0085] vi. The measurement data will be processed in the onboard microprocessor 20, and the evaluation results will be displayed on the screen 29 on top of the sensor, or via audio prompts.

[0086] vii. Evaluation results and measurement data can be transmitted to mobile device 25 or PC via WiFi.

[0087] viii. After the test is completed, the sensor will automatically turn off.

[0088] Product setup procedure:

[0089] i. Launch the GUI from a mobile device or PC connected to the product.

[0090] ii. Enter the user's personal information, such as the user's name, age, weight, height, etc.

[0091] iii. Set up an evaluation report, which can be a numerical value, a quality rating, or a color indicator in the form of a visual display and / or audio cues.

[0092] iv. Set up data records and transfer data records.

[0093] v. Perform product self-calibration and testing.

[0094] Human balance was measured using the sensor of this invention. Five recruited participants took part in the test. Each participant was measured for 10 seconds using the balance sensor while in a normal state. The change in COP (Coefficient of Performance) of the five participants in the 2D plane over time is shown in the first row of Figure 14. Figures 14A to 14E In contrast, data was collected when test subjects were in an abnormal state after heavy drinking. The changes in COP after drinking are shown in the second row of Figure 14, i.e. Figures 14F to 14J The column shows a comparison of COP changes before and after these test subjects drank alcohol. The red "Var" value represents the variance or mean square of the COP's distance from the origin.

[0095] As shown in Figure 14, for each test subject, the "Var" value increased sharply from its normal state after drinking alcohol, as expected. This clearly indicates a decline in balance ability after drinking. By observing the changes in COP in the 2D plane, it can be directly observed that the range of COP changes increases after drinking alcohol, which means increased swaying of the body. The experimental results show that balance sensors can detect minute changes in human balance ability, providing a basis for assessing fall risk.

[0096] The references cited in this application are incorporated herein by full citation, as shown below:

[0097] [FB1982] FO Black, C. Wall III, HE Rockette Jr. and RJA jo O. Kitch, “Normal subject postural sway during the Romberg test”, Vol. 3, No. 5, pp. 309-318, 1982.

[0098] [YA2011] Y. Agrawal et al., “The modified Romberg Balance Test: normative data in US adults”, Vol. 32, No. 8, p. 1309, 2011.

[0099] [TM2009] T. Michikawa, Y. Nishiwaki, T. Takebayashi and YJJ o. OSToyama, “One-leg standing test for elderly populations”, Vol. 14, No. 5, pp. 675-685, 2009.

[0100] [JC1990] JM Chandler, PW Duncan, and SAJPT Studenski, “Balance performance on the postural stress test: comparison of young adults, healthy elderly, and fallers,” Vol. 70, No. 7, pp. 410-415, June 1990.

[0101] [SM2008] SW Muir, K. Berg, B. Chesworth and MJP t. Speechley, “Use of the Berg BalanceScale for predicting multiple falls in community-dwelling elderly people: a prospective study”, Vol. 88, No. 4, pp. 449-459, 2008.

[0102] [SK2006] S. Köpke and GJZ f. G. u. G. Meyer, “Tinetti test”, Vol. 39, No. 4, pp. 288-291, 2006.

[0103] [NR1984] N. Rainville, “Effect of an implemented fall prevention program on the frequency of patientfalls”, QRB: Quality Review Bulletin, Vol. 10, No. 9, pp. 287-291, 1984.

[0104] [LR1993] LJ Rapport et al., “Predictors of falls among right-hemisphere stroke patients in the therehabilitation setting,” Archives of Physical Medicine and Rehabilitation, Vol. 74, No. 6, pp. 621-626, 1993.

[0105] [TS2002] TM Steffen, TA Hacker, and LJP t. Mollinger, “Age-and gender-related test performance in community-dwelling elderlypeople: Six-Minute Walk Test, Berg Balance Scale, Timed Up & Go Test, and gaitspeeds”, Vol. 82, No. 2, pp. 128-137, 2002.

[0106] [YT1976] YJCMAJ Terekhov, “Stabilometry as a diagnostic tool in clinical medicine”, Vol. 115, No. 7, p. 631, 1976.

[0107] [KJ2015] KL Johansen, RD Stistrup, J. Madsen, CS Schjøtt and A. Vinther, “The timed up and go test and 30 second chair-stand test are reliable for hospitalized patients with stroke,” Physiotherapy, Vol. 101, p. e918, 2015.

[0108] [JG2017] J. Goecke, “Effectiveness of the Stepping on Program in Fall Prevention Measured by the Four Stage Balance Test (FSBT)”, 2017.

[0109] [KB1989] K. Berg, “Balance and its measure in the elderly: a review”, Physiotherapy Canada, Vol. 41, No. 5, pp. 240-246, 1989.

[0110] [JC1998] JG Cwikel, AV Fried, A. Biderman and D. Galinsky, “Validation of afall-risk screening test, the Elderly Fall Screening Test (EFST), for community-dwelling elderly”, Disability and Rehabilitation, Vol. 20, No. 5, pp. 161-167, 1998.

[0111] [SW2000] S. Whitney, M. Hudak and G. Marchetti, “The dynamic gait index relates to self-reported fall history in individuals with vestibular dysfunction,” Journal of Vestibular Research, Vol. 10, No. 2, pp. 99-105, 2000.

[0112] [MT1986] ME Tinetti, TF Williams and R. Mayewski, “Fall risk index for elderly patients based on number of chronic disabilities”, The American Journal of Medicine, Vol. 80, No. 3, pp. 429-434, 1986.

[0113] [LW1990] L. Wolfson, R. Whipple, P. Amerman and JN Tobin, “Gait assessment in the elderly: a gaitabnormality rating scale and its relation to falls,” Journal of Geriatrics, Vol. 45, No. 1, pp. M12-M19, 1990.

[0114] [JV1996] JM VanSwearingen, KA Paschal, P. Bonino and J.-F. Yang, “The modified GaitAbnormality Rating Scale for recognizing the risk of recurrent falls in community-dwelling elderly adults,” Physical Therapy, Vol. 76, No. 9, pp. 994-1002, 1996.

[0115] [JM1989] JM Morse, RM Morse and SJ Tylko, “Development of a scale to identify the fall-prone patient”, Canadian Journal on Aging, Vol. 8, No. 4, pp. 366-377, 1989.

[0116] [HM1992] H. Maekawa, K. Tanie, K. Komoriya, M. Kaneko, C. Horiguchi and T. Sugawara, “Development of a finger-shaped tactile sensor and its evaluation by active touch”, Robotics and Automation, 1992 Conference Proceedings, presented at the 1992 IEEE International Conference on Robotics and Automation, 1992: IEEE, pp. 1327-1334.

[0117] [NN1990] N. Nakao, M. Kaneko, N. Suzuki and K. Tanie, “A finger-shaped tactile sensor using an optical waveguide”, from the Industrial Electronics Society, 1990. IECON'90, 16th Annual IEEE Conference, 1990: IEEE, pp. 300-305.

[0118] [SB1988_1] S. Begej, “Planar and finger-shaped optical tactile sensors for robotic applications”, IEEE Journal on Robotics and Automation, Vol. 4, No. 5, pp. 472-484, 1988, doi: 10.1109 / 56.20431.

[0119] [SB1988_2] S. Begej, “Fingertip-shaped optical tactile sensor for robotic applications,” Robotics and Automation, 1988 Conference Proceedings, presented at the 1988 IEEE International Conference on Robotics and Automation, 1988: IEEE, pp. 1752-1757.

[0120] [SW2019] S. Wang and N. Xi, “Recurrent Transfer Learning by Neural Network Regression for Human Balance Sensor Calibration”, IEEE SENSORS 2019, pp. 1-4.

[0121] [SW2020] S. Wang and N. Xi, “Calibration of Haptic Sensors using Transfer Learning”, IEEE Sensors Journal, 2020.

[0122] [SW2018] S. Wang and N. Xi, “Sensing and Data Analysis for Assessing Human Balance Ability”, presented at the 8th Annual International Conference on Networking Techniques in Automation, Control and Intelligent Systems (CYBER) of IEEE, 2018: IEEE, pp. 807-812.

[0123] [JM1997] J. McAuley, J. Rothwell and C. Marsden, “Frequency peaks of tremor, muscle vibration and electromyographic activity at 10 Hz, 20 Hz and 40 Hz during human finger muscle contraction may reflect rhythmicities of central neural firing,” Experimental Brain Research, Vol. 114, No. 3, pp. 525-541, 1997.

[0124] [TP1996] TE Prieto, JB Myklebust, RG Hoffmann, EG Lovett and BM Myklebust, “Measurements of postural steadiness: differences between healthy young and elderly adults”, IEEE Transactions on Biomedical Engineering, Vol. 43, No. 9, pp. 956-966, 1996.

[0125] [MG1990] M. Granat, C. Kirkwood and B. Andrews, “Problem with the use of total distance travelled and average speed as measures of postural sway”, Medical & Biological Engineering & Computing, Vol. 28, No. 6, pp. 601-602, 1990.

[0126] [AH1980] A. Hufschmidt, J. Dichgans, K.-H. Mauritz and M. Hufschmidt, “Some methods and parameters of bodysway quantification and their neurological applications,” Archives of Psychiatry and Neurological Disorders, Vol. 228, No. 2, pp. 135-150, 1980.

[0127] [FH1989] FB Horak, CL Shupert and A. Mirka, “Components of postural dyscontrol in the elderly: a review”, Neurobiology of aging, Vol. 10, No. 6, pp. 727-738, 1989.

[0128] [ID1983] I. DIRECTIONS, “Standardization in platform stabilometry being a part of posturography,” *Agressologie*, Vol. 24, No. 7, pp. 321-326, 1983. While the invention has been explained with reference to certain embodiments, it should be understood that various modifications will become apparent to those skilled in the art upon reading the specification. Therefore, it should be understood that the invention disclosed herein is intended to cover such modifications falling within the scope of the appended claims.

Claims

1. A human balance sensor for assessing a user's fall risk, comprising: A transparent glass plate having a flat upper and lower surface, and the refractive index of the transparent glass plate being greater than that of air; A latex sheet is positioned on the top surface of the glass plate, and during operation, the feet of a standing user are placed on the latex sheet. A light source, positioned to direct light from the edge of the glass plate into the glass plate; A high-resolution camera is positioned below the lower surface of the glass plate to capture light diffused from the glass plate when the user's foot applies pressure to the glass plate; and Therefore, based on the principle of suppressed total internal reflection (FTIR), when a user stands on the glass plate with their foot: (a) the latex sheet presses against the upper surface of the glass plate, (b) the total internal reflection condition at the pressure point of the foot is eliminated, and (c) diffuse reflection of light passes through the bottom surface of the glass plate and is focused onto the image plane of the high-resolution camera, so as to form a tactile image of the contact area of ​​the foot with different pixel intensities based on the different pressures of the foot at different positions on the glass plate. The high-resolution camera recorded a series of tactile images over a period of time; and The human balance sensor further includes a microprocessor that analyzes changes in the series of tactile images and determines human balance ability based on a regression model that integrates COP measurement, footprint analysis and COG measurement. The regression model is achieved by fusing two parts. The first part, based on COP-based measurements, footprint tracing analysis results, and COG-based measurements extracted from the image, is fed into a support vector machine that outputs the test subject's first fall probability. The second part is based on a deep convolutional neural network, which directly takes video data from the balance sensor as input and outputs the test subject's second fall probability; and The weighted average of the first fall probability and the second fall probability is taken from the support vector machine and the deep neural network, and is used as the final evaluation result of the fall assessment.

2. The human body balance sensor according to claim 1, wherein the light source is an LED light source.

3. The human balance sensor according to claim 2, wherein the LED light source is a red LED light strip located around the periphery of the glass.

4. The human balance sensor according to claim 1, wherein the high-resolution camera records a series of tactile images over a period of time; and The human balance sensor further includes a microprocessor that analyzes changes in the series of tactile images and determines the human balance ability based on the changes.

5. A human balance sensor for assessing a user's fall risk, comprising: shell; Two transparent glass plates, each having a flat upper and lower surface, and having a refractive index greater than that of air, are positioned side by side on the top of the housing and spaced apart from each other approximately the distance between the feet of a standing human body. Latex sheet, a latex sheet is positioned on the top surface of each of the glass plates, and during operation, the feet of a standing user are placed on the corresponding latex sheet; A light source, positioned to direct light from the edge of the glass plate into the glass plate; A high-resolution camera is positioned below the lower surface of the glass plate to capture light diffused from the glass plate when pressure is applied to it. and Therefore, based on the principle of suppressed total internal reflection (FTIR), when a user stands on the glass plate with their foot: (a) the latex sheet presses against the upper surface of the corresponding glass plate, (b) the total internal reflection condition at the pressure point of the foot is eliminated, and (c) diffuse reflection of light passes through the bottom surface of the glass plate and is focused onto the image plane of the high-resolution camera, so as to form a tactile image of the contact area of ​​the foot with different pixel intensities based on the different pressures of the foot at different positions on the glass plate; The high-resolution camera recorded a series of tactile images over a period of time; and The human balance sensor further includes a microprocessor that analyzes changes in the series of tactile images and determines human balance ability based on a regression model that integrates COP measurement, footprint analysis and COG measurement. The regression model is achieved by fusing two parts. The first part, based on COP-based measurements, footprint tracing analysis results, and COG-based measurements extracted from the image, is fed into a support vector machine that outputs the test subject's first fall probability. The second part is based on a deep convolutional neural network, which directly takes video data from the balance sensor as input and outputs the test subject's second fall probability; and The weighted average of the first fall probability and the second fall probability is taken from the support vector machine and the deep neural network, and is used as the final evaluation result of the fall assessment.

6. The human balance sensor of claim 5, wherein the high-resolution camera has a frame rate of about 30 fps and a resolution of about 1920x1440.

7. The human balance sensor according to claim 5, wherein the high-resolution camera is capable of wirelessly transmitting images to another computing device.

8. The human balance sensor of claim 5, further comprising a display located on the upper surface of the housing for displaying fall assessment results as a measure of human balance ability.

9. The human balance sensor of claim 5, wherein the microprocessor analyzes the changes in the series of tactile images and determines human balance ability based on measurements of different coordinates of the center of pressure (COP) over time.

10. The human balance sensor of claim 9, wherein the COP measurement comprises at least one of the following: The time-domain "distance" measurement of the average distance of the COP from the origin, the root mean square distance of the COP from the origin, the total length of the COP path, and the average velocity of the COP; Time-domain "area" measurements of the area of ​​a 95% confidence circle, the 95% confidence limit of an RD time series, and the area of ​​a 95% confidence ellipse; A time-domain "hybrid" measurement of wobbling area estimation, average rotation frequency, and fractal dimension; and Frequency domain measurements of power spectral moment, total power, 50% power frequency, 95% power frequency, centroid frequency, and frequency dispersion.

11. The human balance sensor of claim 5, wherein the microprocessor analyzes the changes in the series of tactile images and determines human balance ability based on footprint tracing analysis.

12. The human balance sensor of claim 5, wherein the microprocessor analyzes the changes in the series of tactile images and determines human balance ability based on measurements of a series of center of gravity (COG) coordinates over time.

13. The human balance sensor of claim 5, further comprising a component for manually extracting certain features from the tactile images before the microprocessor analyzes the changes in the series of tactile images.

14. The human balance sensor of claim 5, further comprising a component for training a deep learning algorithm, such as a 3D convolutional neural network (CNN), to generate a classification model prior to the microprocessor analyzing the changes in the series of tactile images.

15. The human balance sensor of claim 13, further comprising, after the manual extraction and before the microprocessor analyzes the changes in the series of tactile images, a component for training a deep learning algorithm, such as a 3D convolutional neural network (CNN), to generate a classification model.

16. The human balance sensor of claim 5, wherein the microprocessor analysis is based on a human body model, the human body model including multiple differential equations associated with changes in pressure distribution under the user's feet, and the analysis is based on the solutions to the equations to obtain a detailed body movement process.

17. The human balance sensor of claim 16, wherein the microprocessor solves the differential equation based on an algorithm derived from the Generative Adversarial Triangle (GAT) model.

18. The human balance sensor according to claim 16, wherein the microprocessor solves the differential equation using the Generative Adversarial Triangle Model (GAT) method, wherein the GAT method combines analytical methods with neural networks to numerically solve nonlinear ordinary differential equations with non-initial conditions as follows: The neural network is initialized randomly or through an approximate solution; Train the model using the Runge-Kutta loss function or the Euler loss function until convergence to obtain a numerical solution; Determine whether convergence has been achieved. If convergence has not been achieved, adjust the current output of the neural network to meet the boundary value conditions and retrain the model. If convergence has been achieved, the process ends with the current result.

19. The human balance sensor of claim 18, wherein the neural network is initialized with an approximate solution, wherein the nonlinear terms in the equations are first discarded; and The solution is determined by using the finite difference method with the aid of the boundary conditions.

Images